UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Improved stove adoption in the Northern Peruvian Andes Agurto Adrianzén, Marcos Miguel 2011

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2011_spring_agurto_marcos.pdf [ 1.38MB ]
Metadata
JSON: 24-1.0071618.json
JSON-LD: 24-1.0071618-ld.json
RDF/XML (Pretty): 24-1.0071618-rdf.xml
RDF/JSON: 24-1.0071618-rdf.json
Turtle: 24-1.0071618-turtle.txt
N-Triples: 24-1.0071618-rdf-ntriples.txt
Original Record: 24-1.0071618-source.json
Full Text
24-1.0071618-fulltext.txt
Citation
24-1.0071618.ris

Full Text

Improved Stove Adoption in the Northern Peruvian Andes by Marcos Miguel Agurto Adrianzén B.Sc. in Economics, Universidad de Piura, Piura, 2001 Master in Rural and Local Development, Instituto de Economía y Geografía - Consejo Superior de Investigaciones Científicas, Madrid, 2003 M.A. in Economics, The University of British Columbia, Vancouver, 2005  A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in The Faculty of Graduate Studies (Economics) THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) March 2011 © Marcos Miguel Agurto Adrianzén 2011  Abstract This dissertation examines outcomes from a development intervention which introduced improved cooking stoves into the rural communities of the Chalaco District, in the Northern Peruvian Andes. The first chapter introduces the dissertation; it presents the intervention’s context and discusses the social capital concept and how it was measured. The second chapter confirms the informational role of village social capital. It explores how bonding social capital and village-level technology usage patterns mutually influence information diffusion during the initial adoption stages of a new cooking device. The results indicate that the effect of village usage patterns on the household’s usage decision is significantly higher in villages with higher levels of bonding social capital, and that the marginal impact of bonding links on the usage decision may be negative if village success in stove usage at initial adoption stages is relatively low. Social capital indicators were collected before the intervention; therefore, reverse causality should not be critical for identification purposes. Village unobservables are not likely to drive the main results; the effect of village usage patterns on the decision to dismantle the improved stove is also increasing in bonding social capital. The third chapter estimates the effect of the improved stove on firewood consumption during the winter season. To identify the impact of stove usage, it exploits random differences in stoves’ material quality. Given this, an indicator of iron frame failure is used as an instrument to predict stove adoption to determine the causal effect of this device. The instrumental variable results indicate that improved stove usage significantly reduces firewood consumption by approximately 40%. The fourth chapter analyses the impact of the new device on health indicators typically affected by indoor air pollution (IAP). To identify the causal impact of improved stove usage, I follow the same identification strategy discussed in chapter three. The results indicate that improved stove usage, with an operative chimney, reduces self-reported respiratory illness and eye discomfort symptoms. These results are only for housewives, who are more likely to be exposed to IAP. No significant health effects were found for housewives using the improved stove without an operative chimney.  ii  Table of Contents Abstract……………….....…………....………………….........………...……………………...ii Table of Contents……....……………………………….........………..………………………iii List of Tables……………………………………...........………..…….........………………….v List of Figures…………………………………………...........………..……………………..viii Acknowledgments....…………………...…………...................…………….…………….…..ix Dedication………....….....................………………………............………………….......……x  1. Introduction.............................................................................................................................1 1.1. The Adoption Decision....................................................................................................1 1.2. Social Capital in the Chalaco District.............................................................................4 1.3. Testing the Role of Social Capital on Stove Usage Decisions........................................9 1.4. Evaluating the Outcomes of the Improved Stove Intervention.....................................11 1.5. A Comprehensive Development Intervention Analysis................................................13  2. Social Capital and Improved Stove Adoption....................................................................14 2.1. Introduction...................................................................................................................14 2.2. Related Literature..........................................................................................................21 2.3. The Intervention............................................................................................................24 2.3.1. Social Learning and Information Diffusion......................................................26 2.4. Data................................................................................................................................27 2.4.1. Village Social Capital Indicators: 2003 Household Survey.............................27 2.4.2. Adoption Patterns: 2004 Stove Monitoring Survey..........................................30 2.5. Empirical Strategy.........................................................................................................35 2.6. Baseline Estimation Results..........................................................................................36 2.6.1. Household Level Determinants of Improved Stove Effective Usage...............47 2.7. Stove Dismantling.........................................................................................................50 2.8. Additional Identification Issues.....................................................................................53 2.8.1. Social Acceptability……………......................................................................53 2.8.2. Unobservable Correlates……….....………………………..........................…55 2.9. Bonding versus Bridging Social Capital.......................................................................59 2.10. Conclusion……….......................................................................................................62  iii  3. Improved Stove Adoption and Firewood Consumption....................................................65 3.1. Introduction...................................................................................................................65 3.2. Related Literature..........................................................................................................70 3.3. Identification Strategy...................................................................................................73 3.4. Data...............................................................................................................................78 3.5. Empirical Estimations...................................................................................................83 3.5.1. Baseline OLS Estimations................................................................................83 3.5.2. Instrumental Variables......................................................................................85 3.5.3. Household Factors............................................................................................90 3.6. Conclusion.....................................................................................................................91  4. Improved Stove Adoption and Health Outcomes..............................................................93 4.1. Introduction...................................................................................................................93 4.2. Data...............................................................................................................................96 4.3. Empirical Methodology and Results.............................................................................99 4.3.1. OLS Estimations.............................................................................................100 4.3.2. Instrumental Variables...................................................................................103 4.3.3. Additional Estimations...................................................................................105 4.4. Conclusion...................................................................................................................110  5. Conclusion............................................................................................................................112  6. Bibliography........................................................................................................................117  Appendix A: Figures and Maps.............................................................................................122  Appendix B: Local Identification in Non-linear in Means Models.....................................125  Appendix C: Additional Tables.............................................................................................129  iv  List of Tables Table 1.1  Trust in the Chalaco District..................................................................................7  Table 2.1  Village social capital indicators...........................................................................29  Table 2.2  Linear correlations between the village social capital indicators........................29  Table 2.3  Improved stove usage patterns at the village level..............................................31  Table 2.4  Main problems encountered by improved stove beneficiaries............................32  Table 2.5  Linear correlations between the household’s effective usage decision and village usage patterns, social capital and geographic characteristics..............................33  Table 2.6  Main household’s characteristics for improved stove users and non users.........34  Table 2.7  Village level determinants of the household’s decision to use the improved stove as the main cooking device..................................................................................38  Table 2.8.A  Village level determinants of the household’s decision to use the improved stove as the main cooking device (allowing for an interaction term between usage patterns and the village bonding social capital indicator)....................................41  Table 2.8.B  Village level determinants of the household’s decision to use the improved stove as the main cooking device (allowing for an interaction term between usage patterns and the village bonding social capital indicator and including additional village controls)...................................................................................................46  Table 2.9  Household level factors affecting the improved stove usage decision................49  Table 2.10  Village factors affecting the household’s likelihood of dismantling the improved stove.....................................................................................................................51  Table 2.11  The effect of social fashion on individual usage decisions (indirect evidence)..54  Table 2.12  The effect of bonding vs. bridging social capital on the household’s likelihood of using the improved stove as the main cooking device.......................................61  Table 3.1  Ex-ante characteristics for beneficiary households using the improved stove without problems and beneficiary households that experienced iron frame problems during the 2004 monitoring visits........................................................77  Table 3.2  Main characteristics for beneficiary households reporting and not reporting iron frame material problems: 2008 Survey................................................................81  Table 3.3  Firewood consumption during the 2008 winter season: improved stove users and non users..............................................................................................................82  Table 3.4  The effect of improved stove usage on monthly firewood consumption............84  v  Table 3.5  The effect of improved stove usage on monthly firewood consumption (user group just includes households that use the improved stove as the only cooking device)..................................................................................................................85  Table 3.6.A  The effect of improved stove usage on monthly firewood consumption. Instrumental variables: second stage regressions.................................................86  Table 3.6.B  The effect of improved stove usage on monthly firewood consumption. Instrumental variables: first stage regressions.....................................................86  Table 3.7.A  The effect of improved stove usage on monthly firewood consumption. Instrumental variables: second stage regressions (users group just includes households that use the improved stove as the only cooking device)..................88  Table 3.7.B  The effect of improved stove usage on monthly firewood consumption. Instrumental variables: first stage regressions (users group just includes households that use the improved stove as the only cooking device).................88  Table 3.8  Natural log of monthly firewood consumption: household level controls..........90  Table 4.1  Incidence of respiratory illnesses and eye discomfort among different groups of household members: 2008 households survey....................................................98  Table 4.2  The effect of improved stove usage with an operative chimney on housewives’ eye discomfort....................................................................................................100  Table 4.3  The effect of improved stove usage with an operative chimney on housewives’ respiratory health...............................................................................................102  Table 4.4  The effect of improved stove usage with an operative chimney on housewives’ eye discomfort: instrumental variables approach...............................................104  Table 4.5  The effect of improved stove usage with an operative chimney on housewives’ respiratory health: instrumental variables approach..........................................104  Table 4.6  The effect of improved stove usage without an operative chimney on housewives’ eye discomfort and respiratory health...........................................106  Table 4.7  The effect of improved stove usage with an operative chimney on eye discomfort symptoms: adult males and children specific group regressions......................107  Table 4.8  The effect of improved stove usage with an operative chimney on respiratory health: adult males and children specific group regressions..............................107  Table 4.9  Housewives’ eye discomfort and respiratory health: individual and household level controls......................................................................................................109  Table 4.10  Adult males’ eye discomfort and respiratory health: individual and household level controls......................................................................................................109  vi  Table C.1  Village level determinants of the household’s decision to use the improved stove as the main cooking device (including additional village controls)...................129  Table C.2  Instrumental variables: first stage regressions for the effect of improved stove usage with an operative chimney on housewives’ eye discomfort symptoms and respiratory health...............................................................................................130  Table C.3  Main characteristics for households reporting and not reporting iron frame material problems: 2008 survey.........................................................................131  Table C.4  Dependent variable: reporting an iron frame material problems. OLS multiple regression results................................................................................................132  vii  List of Figures Figure 1  The improved stove original design.....................................................................122  Figure 2  The traditional firewood cooking technology......................................................122  Figure 3  An improved firewood stove as observed in the summer 2008...........................123  Figure 4  The Chalaco District in the Piura region..............................................................123  Figure 5  The Chalaco District.............................................................................................124  viii  Acknowledgements In first place I would like to especially thank my supervisor Dr. Siwan Anderson for her extraordinary support, patience and encouragement during this research project. I would also like to thank my committee members Dr. David Green and Dr. Patrick Francois for all their help and suggestions. I also want to thank Dr. Ashok Kotwal for his guidance during the initial steps of this research. I am very grateful to Dr. Jorge Viera, Nora Grados, Gabriela Ortega, Fernando Barranzuela, Maria Sofia Dunin-Borkowski, Gonzalo Urday and Andrés Carrasco who introduced me to the Chalaco Program and kindly facilitated me the necessary information which allowed this research to progress. I also specially thank Dr. Antonio Mabres; he always encouraged me to study the development situation of our beloved Piura Region. This research has considerably benefited from the suggestions and comments of my friend Germán Pupato; he is not only a great applied econometrician but also a wonderful person. My friends Cristian Troncoso, Thomas Fujiwara, Anirban Mukherjee, Kim Lehrer, Subrata Sarker, Jonathan Goyette. Javier Torres, David Freeman and Jian Mardukhi were always happy to listen to my ideas, and were awesome at providing feedback and valuable comments. I am totally indebted to the people in the villages in the Chalaco District. They always opened us their doors; they guided us through the mountains in the night; they were always happy to help. I would like to thank the International Development Research Centre for providing financial support for this research. Finally, but most importantly, I want to thank my lovely wife Valeria, her support and her love were my main motivation in this journey.  ix  Dedication  To my mom, Laura, and my dad, José Marcos, with love and gratitude.  To my grandfather José Mercedes, he started it all.  To my grandma, Luz, I would probably not be here if it weren’t for her.  To my sisters Laura and Liliana, I wish I could have been a better brother for you both.  To Guillermo Remicio, thank you for teaching me what friendship is really about.  To the love of my life, Valeria, at every second I love you one billion times more.  To my little princesses, Mariana Celeste and María Valentina, and their promise of life  x  1. Introduction In the fall of 2003, a local NGO, MIRHASPERU, with the financial support of the Spanish International Cooperation Agency, instigated a wide-scale distribution of improved cooking stoves (“cocinas mejoradas”) in the Northern Peruvian Andes. These “more efficient” firewood stoves, with a metallic chimney, were distributed without monetary cost to all households in 37 (of the 39) villages within the Chalaco District1. The main objectives of this development intervention were: (i) to alleviate forest degradation due to firewood extraction for energy purposes; and (ii) to improve respiratory health by reducing exposure to indoor air pollution (IAP) during cooking tasks, particularly among adult women and children. Using first hand data collected by myself in the years 2003 and 2008, the next three chapters in this dissertation empirically explore different features of this particular intervention. Chapter 2 focuses on the role played by village social capital in facilitating social learning and information diffusion during the initial stages of the new stove adoption process. Chapters 3 and 4 attempt to estimate the causal effect of “long term” improved stove usage on firewood consumption and on the incidence of respiratory illnesses and eye discomfort symptoms respectively.  1.1. The Adoption Decision Firewood extraction is one of the main causes of forest degradation in developing countries; it is estimated that approximately three billion people around the world rely on biomass (firewood, charcoal, dung and crop residues) and coal as their main source of domestic energy (Reddy et al (1996)), and that biomass accounts for 50% to 95% of the primary energy consumption in low income countries (Werecko-Brobby et al (1996)). In addition to its impact on forest degradation, several studies in the epidemiological literature indicate that there is a clear connection between biomass and coal usage and the incidence of acute respiratory illnesses and chronic pulmonary diseases due to increased exposure to indoor air pollution (IAP), especially among adult women and infants (Ezzati and Kammen (2002)). In a recent report based in rural Orissa (India), Duflo et al (2008b)  1  The improved stove design distributed in the Chalaco District is shown in figure 1 in Appendix A.  1  show that there is a strong correlation between using a clean fuel2 stove and having better respiratory health, particularly among infants and adult women; which indicates that the use of biomass traditional stoves may be a critical factor behind the incidence of respiratory problems. Furthermore, the World Health Organization has ranked IAP from solid fuels as the 8th most important risk factor for attributable preventable loss of disability-adjusted life years (Diaz et al (2006)). All this suggests that improved “more efficient” firewood stoves with a metallic chimney mechanism, like the ones distributed in the Chalaco district, have the potential to play a critical role in rural areas of the developing world. Not only by alleviating forest degradation (reducing the firewood collection needs of rural households at the margin of forest areas), but also by improving adult women’s and children’s respiratory health conditions (as these stoves have the potential to reduce exposure to IAP during cooking tasks). Given the relevant benefits that can be attained as a result of improved stove usage, beginning in the early 70’s, improved stove dissemination initiatives have been a key component in different environmental and health development interventions in LDC’s. Since then a significant variety of improved “more efficient” stoves designs have been massively distributed. However, the effective usage rates observed after the introduction of these “more efficient” and “cleaner” stoves are, in general, relatively poor (Heltberg et al (2000), Chen et al (2006)). In the specific context of the Chalaco District, by the end of the first year, after its introduction, only 45%3 of the households who freely received the new stove during the distribution stages4 were effectively using it as the main cooking device. In this sense, we can wonder why if this technology appears to have such great benefits on firewood savings and respiratory health (more on the real impacts of these devices in Section 1.4 and Chapters 3 and 4); it was not being effectively used by everyone who received it. Or reformulating the question, we may ask what factors can facilitate achieving higher improved stove’s adoption rates. One possible answer to this situation can be found in the empirical economic development literature that studies social learning during the adoption of agricultural technologies in rural communities, and 2  LPG or electricity stoves MIRHASPERU and Universidad de Piura 2004 Monitoring Report 4 Only 50% of the initial beneficiaries were using their improved stoves as observed in the summer of 2008 3  2  which main results are contained in the research works by Conley and Udry (2010), Bandiera and Rasul (2006), Munshi (2004) and Foster and Rosenzweig (1995). As these papers discuss, the adoption of new agricultural technologies in rural communities is by no means an instantaneous process, and requires continuous experimentation and learning which generates valuable local knowledge that can be shared and diffused among households within different communal and network scopes. That is, the degree of social learning and information diffusion that is present within a given network or community, plays a significant role during the adoption of new technologies. However, despite the results in these papers clearly confirm the presence and relevance of information diffusion and social learning; not much evidence exists on the communal or network social factors that may facilitate or accelerate these social processes. In the specific case of the new stove technology distributed in the Chalaco District, its permanent adoption as the household’s main cooking device was indeed far from being a straightforward process, and qualitative evidence from the intervention area suggest that households’ experimentation and learning by doing played a crucial role during the initial adoption stages. For example, the way of processing the firewood input in this new technology represented a significant departure from the procedure followed with the traditional “open fire” stove. With the new stove, firewood had to be cut and introduced in the combustion box in a very particular manner5 and households needed to master this new way of processing the firewood input in order to achieve the promised efficiency gains in firewood consumption. The qualitative evidence from the intervention area also suggests that the individual experimentation and learning by doing generated local “know how” that was likely to have been diffused among households within the same community. However, this evidence also indicates that while in some villages this local generated information on how to use the new cooking stove (or information related to its real benefits) was intensively diffused among neighbours; in others, this information was  5  Firewood had to be cut in small pieces of a given length and diameter and the stove had to be preheated until reaching the adequate temperature to initiate cooking tasks. Moreover, the type of wood available is not the same in every watershed, and the original firewood processing instructions were based on the type of firewood that is common to villages in the coastal areas of the Piura Region.  3  not shared or was poorly diffused6. The second chapter in this dissertation focuses on this specific issue, and its central hypothesis is that the level of social information (and social learning) available to an individual household within a given village on how to use a new technology (or information related to its real performance) is a joint function of village social capital and village-level usage patterns. More precisely, it is expected that for a given village-level usage pattern, social learning and information diffusion will be stronger in villages with higher levels of social capital. In other words, Chapter 2’s main contribution is to show that the technology usage decision of a given household is more sensitive to its neighbours’ technology usage patterns in rural communities with stronger “within” village (bonding) social capital. In this sense, this chapter not only adds to the existing literature by confirming the presence of social learning and information diffusion during the adoption of a new technology (extending the results to a technology other than an agricultural one); but also provides new evidence on the social factors that facilitate this learning and information diffusion process.  1.2. Social Capital in the Chalaco District In the recent years, the concept of social capital has received special attention in the economic development literature and its benefits, as well as negative effects, for economic development have been particularly emphasized (Knack and Keefer (1997), Narayan et al (1999), Guiso et al (2004), Dasgupta (2005), Francois and Zabojnik (2005), among others). However, there is no consensus on what social capital exactly refers to, and by consequence, measuring this variable and using the obtained indicators in empirical analysis is a complicated as well as a controversial task (as we may not know with exactitude what we are exactly measuring). In many studies in the literature, social capital has been generally associated with trust; for example, for Putnam et al (1993) social capital refers to the “... features of the social organization, such as trust, norms and networks that can improve the efficiency of society by facilitating coordinated actions”; more recently, Bowles and Gintis (2002) affirm that: “Social Capital refers generally to trust, concern for one’s associates, a willingness to live by the norms of one’s 6  This qualitative evidence has been provided by members of the stove distribution team I interviewed in 2008, and was also obtained directly by myself during field visits performed in May 2008.  4  communities and to punish those who do not”. It is not a surprise then that the most used variable in the empirical development literature to account for the degree of social capital that is present within a given society or a community has been some indicator of trust, as it has been the case, for example, in the papers by Narayan et al (1999), Guiso et al (2004) and more recently Wang (2009). Another group of studies in the literature prefer to define social capital in a broader manner, as the “nature and extent of social relationships” (Woolcock (1998)) or simply as “interpersonal relationships” (Dasgupta (2005)). These studies emphasize the multiple dimensionalities of social capital as well as its dynamic and evolving nature. For example Woolcock (1998) identifies two dimensions of social capital at the micro (communal) level: embeddedness or bonding social capital; and autonomy or bridging social capital. The first dimension refers to the nature and extent of social links within the village (or the strength of the intra-communal links), while the second one refers to the nature and extent of the social links of the individuals in the village with agents outside the local community (or the strength of the extra-communal ties). Woolcock (1998) also points out that these dimensions do not necessarily move in the same direction; moreover, an increase in autonomy (bridging social capital) may be accompanied by a decrease in embeddedness (bonding social links). Under this conception of social capital, trust and other social variables such as reciprocity and norms of cooperation, are only viewed as consequences or benefits of social capital; however, Woolcock (1998) clearly states that even though they should not be confused with social capital, these “consequences” can be still used as indicators for the level and different combinations of social capital that is present within a given community. In the second chapter in this dissertation I use village-level trust measures as indicators for the potential level of social capital that is present within a given community in the Chalaco District, and I let the reader decide whether she prefers to identify social capital with village-level trust, or to interpret village-level trust instead as a consequence or benefit of village social capital (that is, as a probably imperfect indicator, but indicator at least, of village social links).  In any case, it is important to highlight that both  5  conceptions of social capital admit the use of village-level trust measures as an indicator of village social capital (and, as I previously mentioned, trust has been the most used indicator of social capital in the empirical economic development literature). Having said this, I must note that in this research I follow Woolcock’s (1998) point of view that there are multiple dimensions of communal social capital (or multiple dimensions of trust, for those who prefer to identify social capital with this variable), and in this thesis I argue that only within village or bonding social capital, measured by village-level trust in local neighbours, is the dimension of social capital that plays a key role facilitating information diffusion and social learning within villages during the adoption process of improved stoves in the Chalaco District7. As mentioned earlier, the technology usage decision of a given household is expected to be more sensitive to its village neighbours’ technology usage patterns in communities with stronger bonding social capital. The main intuition for this is that social links in villages with higher levels of bonding social capital (as measured in this thesis by the village-level trust in local neighbours) are relatively strong, and then, within village-level information will be intensively disseminated in this case (also if trust is relatively higher, this information may appear as more credible). The role of social capital in facilitating information diffusion has been broadly and strongly emphasized in the social capital literature as one of the most important “benefits” (social capital may also have negative effects) associated with this variable (Woolcock (1998), Isham (2002), Dasgupta (2005), etc.). In the context of technology adoption processes in rural communities (e.g. adoption of fertilizers, HYV seeds, telecommunication technologies, etc.), this implies that social capital may have an important role in accelerating social learning and by so increasing the adoption rates of a new technology. The village-level trust measures used in this study were collected by myself in the year 2003, in the months prior to the improved stove distribution and adoption stages in the Chalaco District. In that year, as part of my undergraduate thesis research, I introduced a  7  I will also explore in Chapter 2 how bridging social capital influences adoption decisions; I will use the village-level trust in people from other villages as an indicator for this dimension of social capital.  6  social capital questionnaire 8. In this social capital questionnaire, I asked the interviewed the following trust questions: How much do you think you can trust in: a) your local neighbours, b) your local organizations, c) people from other villages and d) in strangers? The scale of responses goes from 0 to 3 (0=nothing, 1=a little, 2=in a regular degree, 3=a lot). If the questions were not clear enough for the interviewed, the interviewer tried to provide specific examples related to the village social life.  Table 1.1: Trust in the Chalaco District Variable Villages Mean Trust in local neighbours index 26 1.45 Proportion of households that replied 0 to the trust 26 0.27 in local neighbours question Proportion of households that replied 1 to the trust 26 0.17 in local neighbours question Proportion of households that replied 2 to the trust 26 0.43 in local neighbours question Proportion of households that replied 3 to the trust 26 0.14 in local neighbours question Trust in people from other villages index 26 1.47 Proportion of households that replied 0 to the trust 26 0.16 in people from other villages question Proportion of households that replied 1 to the trust 26 0.34 in people from other villages question Proportion of households that replied 2 to the trust 26 0.38 in people from other villages question Proportion of households that replied 3 to the trust 26 0.12 in people from other villages question  S.D 0.40  Min 0.42  Max 2.22  0.16  0  0.68  0.08  0  0.38  0.18  0.10  1  0.12  0  0.48  0.32  0.65  1.92  0.12  0  0.15  0.10  0.15  0.57  0.14  0.13  0.69  0.10  0  0.38  The information in this table is presented at the aggregated village level. On average there are 48 households per village, and on average 21 households per village were randomly interviewed.  Table 1.1 shows descriptive statistics for the variables trust in local neighbours and trust in people from other villages, both of them averaged at the village level for a total of 26 villages in this district (the ones in which improved stoves monitoring visits were performed in 20049). As mentioned before, the village averages for these trust related questions are used in this study as indicators of different dimensions of village social capital10. 8  Survey results were not available during the distribution stages or during performance evaluation visits. As I will discuss in detail in Chapter 2, initially visits to all villages were planned, but due to budgetary, administrative and security reasons, 11 villages were finally not visited. It is important to emphasize that villages in which high usage were expected were not particularly targeted; moreover, at the beginning of the monitoring, special emphasis was set in visiting villages in which low usage rates were expected. 10 I must emphasize that self-reported trust has been one of the most widely used variables in the empirical development literature to account for the degree of trust or social capital that is present within a given 9  7  At this point, the reader may wonder why social capital (as measured by village-level trust), can have a high degree of variation among villages or communities within the same district. Although I do not have a conclusive story that can fully explain the observed variability of the social capital levels in the area (exploring this issue will be a main objective of my future research); different geographical, historical, cultural, and religious factors seem to have played an important role in shaping the social capital structure in the Chalaco District. In terms of geographical factors for example, some villages are relatively more accessible than others; in such cases, higher access to markets and to individuals from other communities, or even strangers, can make households in these communities less likely to depend on their local social connections, which may decrease their incentive to invest in social capital. The geographical conditions (and the weather) have also a strong influence on the types of agricultural crops that are feasible in the different villages and watersheds within the district, and have then an important influence over the type of agricultural organizations (as labor rotating organizations), and hence social relationships, that will be observed at the village level11. It is also known that different watersheds and villages in the area have been influenced by different types of development interventions in the past. For example, at high altitude society or community (see for example Knack and Keefer (1997), La Porta et al (1997), Guiso et al (2004), or more recently Wang (2009)). In the context of this research, it is also important to highlight that trust was measured again in the area in 2008, and that the correlation at the village level between the 2008 and 2003 trust measures is strongly positive and significant; also, the order of the village trust rankings is generally preserved. This suggest that the 2003 responses to the trust questions were not likely to have been significantly affected by whether or not the villages trusted the enumerator or to other factors related to the specific conditions of the 2003 survey collection process. In terms of the validity of self-reported trust as a measure of real trust and by hence social capital, in first place we can refer to Glaeser et al (2000), who performed experiments using the Trust Game among Harvard undergraduates and found that: “Although questions about trusting attitudes do not predict trusting behavior, such questions do appear to predict trustworthiness. An index of an individual’s response to GSS attitudinal trust questions has a 34 percent correlation with the amount of money that the individual himself gives back. While attitudinal trust surveys at best weakly predict any individual’s level of trust, they may be good at predicting the overall level of trustworthiness in society.” (Glaeser et al (2000), page 813). However, in a relatively recent paper that performed the Trust Game in rural communities in Perú, Karlan (2005) indicates that: “The prior literature on the Trust Game claims it measures trust for Player A and trustworthiness for Player B. I find evidence that Player A measures propensity to take risks. I also find evidence to support the social capital or "trust" hypothesis (e.g., both players being indigenous, living near their partner and attending the same church lead to higher passes by Player A). Hence, behavior is determined y both types of traits. This murkiness raises doubts about the ability to use the game as a measure purely of trust” (Karlan (2005), page 1698). 11 Villages in the area are placed among 1000 to 3000 meters above the sea level, and as it is reported in the Chalaco Atlas edited by Universidad de Piura (2006), there is a significant variation in terms of their geographic characteristics among the five watersheds in the district.  8  areas, a watershed management program was promoted by the central government during the 90’s, and different villages within the same watershed were encouraged to work together in the management and conservation of their water resources and irrigation infrastructure. At medium altitude areas, village-level coffee producer’s organizations were promoted by a regional organization “CEPICAFE”, in order to encourage the production of organic coffee. Each of these interventions (of the many that have been observed in the Chalaco District) is likely to have influenced the village social structure in a very specific and different way: while the first one promoted across-village connections, the second one mainly promoted intra-village links. Finally, the penetration of certain religious groups12 at medium and at high altitude areas during the late 90’s is believed to have played an important role reshaping the communal life. For example it is known that there has been a considerable reduction in the level of alcoholism and local violence in villages with relative high presence of these groups, as well as a revival of the importance given to the communal social life. All things considered, the geographic, social, religious and historical factors that influence the social life of the villages within the Chalaco District present a high degree of complexity and variability; and this is very likely to be reflected in a high variability of the types and levels of social capital that will be present in these rural communities.  1.3. Testing the Role of Social Capital on Stove Usage Decisions In order to test the social capital - information diffusion hypothesis, Chapter 2 exploits the village-level measures of social capital that I collected (before the stove intervention in 2003), and detailed data on stove usage patterns that was generated during monitoring visits carried out by the NGO MIRHASPERU and Universidad de Piura in 2004 (these data were collected 8 to 12 months after the distribution of stoves, and are discussed in detail in Chapter 2). The results in this chapter demonstrate that the effect of within village-level adoption patterns on the household adoption likelihood is significantly higher in villages with higher levels of trust in local neighbours (information is more intensively diffused in those communities that are likely to have stronger bonding links). Moreover, the marginal impact of this bonding social capital indicator may be negative if 12  Mainly religious groups from the evangelic church.  9  village-level initial successful adoption levels are relatively low (if a technology is initially perceived as a negative one, the social links are likely to intensively diffuse negative information about the new device). It is also shown in Chapter 2 that only the proportion of users in the village that did not experience problems with their own stoves has a positive impact on the individual usage decisions through its interaction with the social capital indicator; while the reverse is true for the proportion of users experiencing problems with the new technology 13 . In this sense, the results in Chapter 2 are closely related to the results by Conley and Udry (2010), who show that pineapple farmers in Ghana tend to adopt the fertilizer usage levels of those reference neighbours experiencing successful returns. Interestingly the effect that users with problems have on individual usage decisions, through its interaction with the village bonding social capital indicator, is, in absolute size, considerably higher than the analogous interaction effect of users without problems, suggesting that bad news about a new technology may have disastrous effects during the adoption process. Furthermore, it is demonstrated that only the bonding social capital indicator influences the effect that within village usage patterns have on individual usage decisions. On the other hand, the bridging (across villages) social capital indicator only influences the effect that usage patterns in neighbouring villages have on the household’s usage decision. To the best of my knowledge, Isham (2002) is the only previous study in the related literature which has empirically explored the effect that village social capital has on information diffusion in the context of technology adoption/usage decisions in rural communities 14 . However, Isham’s research presents serious limitations. For one, his estimations suffer from reverse causality problems; as individual adoption decisions are likely to have also influenced the equilibrium levels of village social capital. Secondly, Isham’s paper does not properly address the potential presence of unobservable factors 13  Almost 90% of users with problems reported problems with the stove materials. However, there are a few papers in the economic development literature that empirically explore the role of social capital in other local/village contexts: Narayan et al (1998) for example study how social capital influences the economic performance of rural communities; Guiso et al (2004) explore how local social capital affects financial development, and Wang (2009) studies how social capital affects participation in labor rotating participation in rural villages in China. 14  10  simultaneously related to social capital, village adoption patterns, and the household’s usage decisions. Since the measures of social capital used in Chapter 2 were collected before the diffusion and adoption stages, reverse causality should not be a critical issue in terms of identifying the effect of social links. Moreover, in this research I provide additional evidence to rule out the possibility that village unobservables are driving the observed patterns in the stove usage data. More precisely, it is shown that the villagelevel proportion of users without problems also has a negative effect on the decision to uninstall the improved stove among beneficiary non users, mainly through its interaction with the bonding social capital indicator. On the other hand, the village-level proportion of users with problems affects the uninstalling decision in the opposite direction, mainly through its interaction with the village bonding links indicator.  1.4. Evaluating the Outcomes of the Improved Stove Intervention Despite being one of the most promoted strategies to help alleviate forest degradation in the developing world (as previously discussed in Section 1.1), the formal empirical evidence on the impact of firewood improved cooking stoves on household’s firewood extraction and consumption is relatively limited and surprisingly inconclusive (Johnson et al (2010)). The main problem in most of the existing studies (Wallmo et al (1998), Heltberg et al (2000), Chen et al (2006), Masera et al (2008), etc.) is related to the fact that the improved stove’s usage decision is very likely to be correlated with household’s and village unobservables which are also correlated with household’s firewood consumption levels (e.g. female empowerment, pro-forest preferences, ability, etc). Then, unless experimental data or suitable instruments are available, it may not be possible to separate the effect of improved stove usage from the effect of these unobservable correlates. Chapter 3 in this dissertation intends to help filling this gap. In order to identify the causal effect of improved stove usage on firewood consumption the third chapter in this thesis exploits random and ex-ante unobservable differences in stoves’ material quality. Evidence from the 2004 monitoring visits suggests that a proportion of households that decided to adopt the new stove experienced material failures in the stove iron frame (deformations and cracks), and that these problems were  11  not likely to have been systematically caused by households’ stove usage patterns, inadequate installation or maintenance, but rather by poor (exogenous and random) materials quality. Given this, an indicator of iron frame failure can be used as an instrument to predict successful stove adoption to determine the causal effect of the improved stove on households’ firewood consumption. The instrumental variable results in Chapter 3 indicate that successful stove adoption reduces monthly firewood consumption by approximately 40% during the winter season in the area of study; this effect is statistically significant and strongly robust to different model specifications. In recent years, the role of improved firewood stoves in reducing exposure to indoor air pollution (IAP) and improving the respiratory health of adult women and children has been especially emphasized (Duflo et al (2008a) and (2008b)). To my knowledge, the works by Diaz et al (2006) and Smith et al (2009), both based on a experimental design in rural communities in Guatemala (the RESPIRE program) 15 , are, in terms of their identification strategy, the most important evaluation studies on the “short term” health benefits of improved stove usage. Other studies (mainly based on randomized trials) aimed at evaluating the impact of improved stove usage on rural household welfare (as a result of improved health or reduced medical expenditures), are being currently implemented (Duflo et al (2008b)). Using the data I collected in the intervention area in the summer of 2008, Chapter 4 in this dissertation explores how long term improved stove’s usage influences self reported incidences of respiratory illnesses and eye discomfort symptoms. In order to identify the causal health effect of improved stove usage, in Chapter 4 I exploit the same identification strategy followed in Chapter 3. The results indicate that the negative impact of long term improved stove usage, with an operative chimney, on self reported respiratory health and eye discomfort symptoms is statistically significant only among housewives, who are more likely to be exposed to IAP. No effect on respiratory health or eye discomfort symptoms was found on housewives in households using the new stove without an operative chimney. Although the measures of health and 15  Improved stoves with a chimney were randomly distributed among Mayan households in the area.  12  eye discomfort symptoms used in this study are self-reported, the results in this chapter are in line with recent evidence in the literature, and significantly add to the current findings by providing evidence on the health benefits of “long term” improved stove usage with an operative chimney device.  1.5. A Comprehensive Development Intervention Analysis Although each of the next three chapters in this thesis is by itself an independent research essay, taken together they constitute a comprehensive analysis of the improved stove dissemination program implemented in the Chalaco District. This integral research does not only evaluate the long term environmental and health related impacts of the new cooking technology; but also studies the social and individual factors behind the decision to use the new stove as the main cooking device at the very initial adoption stages. In this sense, this thesis provides relevant conclusions applicable to the different stages of improved stove (or any other rural technology) dissemination programs. This research confirms the relevance of information diffusion during the adoption of new cooking technologies, and highlights the importance of having an appropriate understanding of the village social structure, as this structure influences the degree in which local generated information will be shared and diffused. The results indicate that the program members should not expect to rely on learning externalities in villages with poor bonding links. It also points to the relevance of high quality monitoring and extension services; as bad news about the performance of the new technology can have disastrous consequences in terms of the adoption processes (such as the total rejection or abandonment of the technology). This research also develops a singular and interesting methodological strategy to identify the casual effect of technology usage. From the perspective of development policy, it confirms the positive impact the improved stove design distributed in the Chalaco District had on firewood consumption and housewives’ respiratory and eye related health. More generally, this suggests that these cooking devices may indeed have a significant impact on rural households’ welfare, as fewer resources have to be allocated to firewood collection and medical expenses. Moreover, improved health may increase the productivity of female household’s members.  13  2. Social Capital and Improved Stove Adoption 2.1. Introduction When a new technology is introduced in rural communities, not only do individual and household factors (such as ability, wealth, risk aversion, etc.) affect the household’s adoption and usage decisions; but, and perhaps more importantly, village social factors also appear to matter. In recent years, the empirical development literature on technology adoption has highlighted the role of social learning and information diffusion at the village level, mainly in the context of agricultural technologies 16 (Conley and Udry (2010), Bandiera and Rasul (2006), Munshi (2004), Isham (2002), Foster and Rosenzweig (1995), among others). A central result in these studies is that individual decisions are strongly related to network and village adoption and usage patterns. The main empirical challenge has been to demonstrate that a social learning or information diffusion process is indeed what drives the observed correlations. However, although a significant variety of issues related to social learning has been explored in great detail17, not enough attention has been given at understanding how the nature and extent of village social relationships facilitate (or not) social learning and information diffusion; neither at how the initial performance of a new technology influences the type of effects village social links will have on the individual household’s decisions. This chapter aims to help filling this gap and investigates how village social capital and village usage patterns mutually affect the household’s decision to effectively use an improved firewood stove technology at early adoption stages, by influencing the degree of social learning and information diffusion within a given community. In this chapter I focus on the decision to effectively use the new stove as the main cooking device only among beneficiary households; that is, households that received the improved stove during the distribution stages (approximately 85% of all the households in the district18). This improved stove was distributed without monetary cost in the fall of 2003 and stove 16  Where, as broadly documented; experimentation, innovation and social learning play a crucial role. Such as the presence of strategic behaviour at early usage stages (Bandiera and Rasul, 2006) or the impact heterogeneity in household characteristics may have at influencing social learning (Munshi, 2004). 18 Internal NGO reports indicate that by November 2003 close to 95% of the stoves were already installed. 17  14  beneficiaries were not required to immediately abandon their traditional cooking technology (“tulpa”) in order get the new one installed19. Information obtained in the summer of 2008 indicates that approximately 96% of those households that initially asked for an improved stove effectively received one. As it has been extensively documented for the case of new agricultural technologies, being able to effectively use an improved firewood cookstove is by no means an easy process. Moreover, the operation mode of an improved stove differs in many aspects from the operation procedure of traditional “open fire” stoves. For example, in the case of the stove distributed in the Chalaco District, the way of processing the firewood input was significantly different than the one followed with the traditional cooking technology. With the new stove, firewood needed to be cut and introduced in the combustion box in a very specific manner in order to achieve the promised efficiency gains. Also, proper cleaning and maintenance were essential in terms of firewood efficiency and indoor air pollution alleviation. In addition, the stove design did not take into account that at high altitude areas the stove is also expected to work as a heating device, and beneficiaries had to figure out by themselves how to achieve this specific need with the new improved cooking artefact. All this suggests that experimentation and learning by doing were critical for the household in terms of being able to effectively use the new firewood cookstove as the main cooking device. The qualitative evidence from the area of study suggests that while in some villages the “local specific” knowledge generated by individual experimentation and learning by doing was intensively disseminated among village neighbours20; in other communities, it was not shared or it was poorly diffused among households.  19  Although I don’t exactly know the number of households that had the new stove installed without uninstalling their traditional open fire stove; the 2003 stove program guidelines suggest that these probably were the majority of beneficiaries, as the program managers’ idea was to allow for a gradual transition from the old cooking technology to the more efficient improved one. 20 For example, during the stove monitoring visit in 2004, modifications to the dimensions of the combustion box were observed in San Lorenzo village in order to improve efficiency in firewood consumption; and this new information was indeed diffused among beneficiary households within this community.  15  Like other studies in the literature, the empirical results in this chapter confirm that households learn from their local neighbours about a new technology. However, its main contribution is to empirically demonstrate that information diffusion and social learning about a new technology are mutually influenced by village-level technology usage patterns and by village bonding social capital, which is defined as the strength of intracommunal links and is measured in this thesis by the village-level trust in local neighbours. More precisely, this chapter confirms the hypothesis that information about a new technology is more intensively diffused in villages which are likely to have strong levels of bonding social capital. In other words, the individual household’s usage decision appears to be more sensitive to the usage patterns within the village in those communities in which the bonding social capital indicator is relatively high. In addition, the present study shows that the marginal impact of the bonding social capital indicator on the individual usage decision is closely linked to the initial within village usage patterns. If the initial success in improved stove usage at the village level is relatively low (the proportion of beneficiaries using the stove without problems is small, or the proportion of beneficiaries using the technology with some problem or difficulty is large), then the bonding social capital indicator is more likely to negatively influence the individual decision to effectively use the improved stove (probably because in this case the social network spreads negative information about the new device). An important advantage of this research in relation to others in the literature is that the social capital measures used in this chapter were obtained prior to the improved stove’s distribution and adoption stages. Hence, reverse causality should not be a critical issue in terms of identifying the informational effect of village social capital. One of the most important issues in studies which focus on social learning and information diffusion, is to properly define the reference group; that is, the group of neighbours a given household learns from. Some studies attempt to infer the presence of social learning by relating adoption and usage levels at different geographic scales (Rosenzweig and Foster (1995), Isham (2002), Munshi (2004)) to the individual household’s adoption decision; in most the cases the village is used as a proxy for the  16  reference network. In more recent studies the reference group has been self-reported by the household (Conley and Udry (2009), Bandiera and Rasul (2006)). Given the characteristics of the improved stove usage data and this chapter’s focus on the effective usage decision only among the program’s beneficiaries; in this study I follow the first approach and use the total number of beneficiaries in the village as the household’s reference group. If true self-reported reference groups may delineate with more precision the household’s informational network; for the case of the villages analyzed in this chapter, where the average number of households is relatively small (48 on average) and membership is stable in time, the total number of beneficiaries in the village appears as a good proxy for the household’s reference network. However, even if the reference group is well defined, the researcher still has to deal with the main identification issues that are common to studies on social interactions, and have been clearly identified in the seminal work of Manski (1993) and recently discussed by Brock and Durlauf (2001b, 2003, 2007). Probably the most challenging issue in terms of identifying the presence of social learning and information diffusion is the potential presence of correlated unobservables. Adoption or usage decisions may be correlated among households within a village just because they share the same unobservable preferences or characteristics, or because they are subject to the same unobservable shocks, especially in villages with strong bonding social capital. Also, the observed correlation between village/network adoption levels and individual adoption may just reflect a pure imitation process. When the data is non experimental in nature or a suitable instrument is not available, the researcher needs to convincingly argue that no process other than social learning is likely to drive the observed correlation between individual and village/network decisions. One way of doing this, is to provide as much evidence as possible on different data patterns that are more likely to be caused by social learning and not by village unobservables or other social processes (Bandiera et al (2002), Munshi (2004)). When the data at hand contains a high degree of detail on household’s characteristics within different network scopes, it may be feasible to control for those otherwise “confounding” factors (Conley and Udry (2010)).  17  In order to support the hypothesis that information diffusion is the process behind the observed correlation between individual usage decisions, village usage patterns and the village social capital indicator, in first place this research exploits the information in the data to define two types of beneficiary users: those that did not report problems during the 2004 monitoring visits and those that did report problems using the new device; in second place, an interaction term between the village usage patterns and the village bonding social capital indicator (the village-level trust in local neighbours) is introduced in the regressions. The results indicate that only the interaction term between the proportion of users without problems and the bonding social capital indicator has a positive and significant effect on the household’s decisions to use the new stove as the main cooking device; while the reverse is true for the proportion of users with problems. In other words, successful usage rates at the village level positively affect the usage decision through its social capital effect, while village usage levels with problems do exactly the opposite (as in the later case “negative” information about the new stove is being intensively diffused).  In third place, I explore the decision to dismantle the  improved stove among beneficiary non users21; the results indicate that an increase in the village proportion of users that did not report problems with their stoves reduces the likelihood of dismantling the new cooking device mainly through its interaction with the bonding social capital indicator. This last result plays a key role in this research, as it confirms that it is unlikely that correlated unobservables, or pure imitation, are driving the main findings in this study. The results in this chapter are closely related to the findings in the paper by Conley and Udry (2010), who show that pineapple farmers in Ghana tend to adopt the fertilizer usage levels of those reference neighbours experiencing successful returns. In the context of this chapter’s findings, one can question whether the higher impact village usage success levels appear to have in villages with higher levels of trust in local neighbours reflects a situation in which social learning is more intense, or if it just mirrors the fact that in such cases the device appears as more socially acceptable or  21  That is, households who received the stove, had it installed, but by the time of the 2004 monitoring visits were not using it as the main cooking device.  18  fashionable. However, in my opinion such social effects are more relevant in contexts in which the decision of study is related to “requesting the stove” or “having it installed”; in the case of this research all households in the sample have the new stove, and what this study tries to understand is their decision to “effectively use” it as the main cooking device. Then, fashion and acceptability effects are less likely to play a critical role in comparison to learning effects, which facilitate access to key information affecting the decision to effectively use the new device. Section 2.8 in this chapter provides indirect evidence suggesting that fashion or acceptability effects are not likely to drive the main results on usage decisions. This chapter also explores how different dimensions of village social links influence technology usage decisions. The results confirm that only the bonding social capital indicator (village-level trust in local neighbours) has a multiplier effect on how within villages usage patterns affect individual usage decisions; while an indicator for bridging social capital (village-level trust in people from other communities), defined as the strength of the community members’ ties with people from other villages, does not have a significant multiplier effect in this case. The bridging social capital indicator only seems to have a multiplier effect on how usage patterns outside the village (e.g. in the closest neighbour village) affect usage decisions; while in this case the bonding social capital indicator does not play any significant role. The results also show that the bonding social capital indicator has a prevailing effect in the regression in which both measures of social links are simultaneously included. To the best of my knowledge, this is one of the first studies presenting evidence on the different roles played by different dimensions of social capital during technology adoption processes in rural communities. In order to control for village level factors that may be simultaneously correlated with the individual usage decisions and the village-level usage patterns, all the baseline regressions in the estimation section include dummies for village watershed location22, as villages within a given watershed share in general the same geographical and weather characteristics, are likely to have had the same degree of exposure to development 22  There are five watersheds in the Chalaco District.  19  interventions in the past, and have relatively similar levels of access to forest resources. Moreover, it is known that the improved stove intervention was designed, implemented and coordinated at the watershed level. As a robustness check, I additionally control for two key geographical variables available in the data which, as reported by program members, critically influenced stove performance and usage decisions: village altitude and road access. The evidence suggest that the stove design was not well suited to meet the heating needs of households in high altitude areas, and that village accessibility is likely to have influenced NGO effort during the distribution and diffusion stages. As we will see in the estimation section, controlling for these additional geographical factors does not affect the main results in this study. The estimations in this research also provide interesting results on the household-level factors influencing effective usage: wealthier households, households that in the previous year participated in village activities and households that have at least one adult female member seem more likely to use the new improved stove as the main cooking device; on the other hand, households with a higher number of adults are significantly less likely to use it, probably because labor abundance decreases the cost of collecting firewood. This research extends the social learning evidence to a technology other than an agricultural one, and its main contribution is to show that social capital plays a key role in the diffusion of information, a role broadly attributed to this variable in the social capital literature (Dasgupta 2005). Moreover, it also shows that social capital may “negatively” impact individual usage decisions if at initial adoption stages the village-level success rates in improved stove usage are relatively low (or failure rates are relatively high). In addition to this, this study confirms that different dimensions of village social capital have different roles at influencing usage decisions. This chapter develops as follows: Section 2.2 discusses the literature, Section 2.3 explains in detail the improved stove program and the relevance of experimentation and learning by doing, Section 2.4 presents the data, Section 2.5 discusses the baseline empirical equation, Section 2.6 presents the main results, Section 2.7 focuses on the improved stove dismantling decisions, Section 2.8 discusses relevant identification issues, Section 2.9 contrasts the different roles played by bonding and bridging social capital, and finally Section 2.10 concludes.  20  2.2. Related Literature To the best of my knowledge, the paper by Isham (2002) is the only previous empirical study in the economic development literature that focuses on how village social capital affects rural household’s technology usage decisions 23 by influencing information diffusion and social learning. His paper extends the model by Feder and Slade (1984) and uses cross sectional data on fertilizer usage in villages in Tanzania in order to show that two measures potentially linked to social capital: i) the village share of households that report that their local organizations include only members of the same clan and ii) the village share of households that report that members vote and discuss decisions within their local organizations, positively and significantly influence the individual household’s fertilizer usage decision. The main drawback in Isham’s paper is related to the fact that his village social capital indicators are very likely to have been influenced by the households’ fertilizer usage decisions. In other words, households adopting the new agricultural technology may have decided to invest more in their social relations, affecting in this way the equilibrium levels of social capital. If this was the case, his estimations will not capture the causal effect of social capital on the individual decision to use fertilizer. In my research, the social capital indicators were obtained before the improved stove adoption process; then reverse causality should not be a critical issue in terms of identifying the effect of village social links. Moreover, Isham’s paper fails to address the possibility that village unobservable factors may be the ones driving the correlations between individual and village fertilizer usage levels. As discussed earlier, in this research I provide solid evidence to support the hypothesis that social learning is indeed the generating process behind the mains patterns in the data. To some extent, my research is relatively close to the work by Bandiera and Rasul (2006), which studies adoption of sunflower seeds at “early” distribution stages in rural communities in Mozambique. Using cross sectional data, they show that the effect of network adoption on individual adoption decisions is U-shaped; in other words, the 23  There are also a few other papers in the economic development literature that empirically explore the role of social capital in other “local” contexts. Narayan et al. (1998) for example study how social capital influences the economic performance of rural communities. In a relatively recent research paper Guiso et al. (2004) explore how local social capital affects financial development. Wang (2009) studies how social capital affects participation in labor rotating participation in rural villages in China.  21  network adoption effect is decreasing in the number of network adopters, and may at some point become negative. They argue that this result suggests the presence of strategic behaviour at early adoption stages: as others experience is a substitute for the household’s experience, the higher the number of adopters in its network the more likely is the household to postpone adoption and free ride on others’ experimentation. Following Bandiera and Rasul, I also allow for a nonlinear effect of village usage levels on individual decisions; my results show that the individual usage likelihood is indeed decreasing in the village proportion of users without problems and, as in their paper, it may turn out to be negative. Bandiera and Rasul also argue that the strength of social ties matters; they find that the network effect among family and friends is higher than the network effect among members of the same religious group. In this sense my results, which show that social effects are higher in villages with higher levels of trust in local neighbours significantly add to current findings in the literature. In terms of dealing with identification issues, especially those associated with the presence of unobservable correlates, the social learning literature provides a relevant variety of examples on how in the absence of experimental data or a suitable instrumental variable, one still can exploit specific characteristics in the data to support the social learning hypothesis. For example, in the paper just discussed, Bandiera and Rasul (2006) argue that pure imitation or unobservable correlates monotonically related to the number of network adopters and to the household’s likelihood of adoption are not likely to drive the U-shaped effect of network adoption on individual adoption decisions. However, they admit that unobserved heterogeneity may be the process driving the non linear patterns in the data, such as unobserved ability linearly correlated with network size but nonlinearly correlated with individual adoption24. To address this issue, they identify in their data some indicator variables potentially linked to unobserved ability, and include an interaction term between the ability indicator and network adoption; their results confirm that the U-shaped pattern is also present for households with potential higher ability. Finally, they estimate the main regressions excluding 25% of the sampled households 24  While ability may be linearly correlated with network size, it may be non linearly correlated with adoption decisions: e.g. lower ability households may have more difficulties in adopting the technology while households with higher ability may have more outside options and are then also less likely to adopt.  22  with the highest level of ability (as defined by certain ability related variables) and are still able to find a strong U-shaped effect of network usage on household usage decisions. Another interesting example is provided in the research by Munshi (2004) on HYV wheat and rice acreage allocation during the green revolution in India. This paper shows that wheat farmers tend to react to past acreage decisions taken by their village neighbours, while rice farmers do not. As rice crops are more sensitive to farmers’ characteristics, which may be imperfectly observable; in Munshi’s opinion the results support the hypothesis that heterogeneity in population characteristics negatively affects social learning during the adoption of new technologies. In order to confirm that social learning is the process linking village outcomes to individual decisions, Munshi shows that the same patterns are observed in villages where both types of crops are present. Then, in his opinion, it is not likely that unobservable spatial characteristics intrinsically linked to “only wheat” or “only rice” villages are driving the observed results. In the same direction, the paper by Conley and Udry (2010) provides the most recent and original example on how to deal with network unobservables in the context of social learning using observational data. They study how pineapple farmers in Ghana react to news about pineapple productivity related to fertilizer usage by self reported reference neighbours25. The authors show that farmers tend to adopt the fertilizer usage levels of those reference neighbours experiencing successful returns. To isolate the effect of social learning from unobservable network spatial shocks, they exploit the detailed geographical information in their data to construct an index measuring the difference between the household’s past usage level of fertilizer and the current level of fertilizer use by the household’s geographically close reference neighbours, which are likely to be affected by the same unobservable shocks. In the authors’ opinion, this index should control for changes in fertilizer usage only attributable to unobservable spatial correlates, which in the end must allow identifying the effect the proportion of reference neighbours experiencing successful returns has on the household’s current changes in fertilizer usage.  25  In the survey used for their paper, pineapple farmers were asked to identify from a random sample of other farmers those to whom they talked and discussed about farming related issues.  23  This chapter’s empirical approach in dealing with village unobservables is closely related to the approach followed by the previously mentioned papers. Exploiting the unique information in the improved stove usage data, I show that only the proportion of users without problems positively affects individual usage decisions through its interaction with the bonding social capital indicator; while the reverse is true for the proportion of users with problems. More importantly, I also analyze the improved stove dismantling decision among beneficiary non users; the results are consistent with the initial findings and indicate that an increase in the proportion of households using the stove without problems decreases the likelihood of dismantling the new device mainly through its interaction with the bonding links indicator. These results suggest that it is unlikely that correlated unobservables are driving the findings in the improved stove adoption data.  2.3. The Intervention The Program for the Sustainable Development of Mountain Ecosystems in the Chalaco District, Peru, also known as the “Chalaco Program”, was conceived as a comprehensive development strategy and was financed by the Spanish International Cooperation Agency. The main intervention during the program’s first year was the distribution of improved “more efficient” firewood stoves. This strategy was adopted as an immediate response to forest degradation26 in the area as well as a way to improve adult women’s and children’s respiratory health  27  by reducing exposure to IAP (mainly during cooking tasks).  Improved stoves were distributed and installed without monetary cost in 37 of the 39 villages within the Chalaco District in the fall of 200328. With the support of the most representative communal organizations, the NGO (MIRHASPERU) held open meetings in every village in order to explain the improved stove program benefits. Following these meetings, an improved stove was allocated to every household who requested one29. The NGO provided beneficiaries with an iron frame and an aluminium chimney; the 26  These improved stoves were supposed to reduce firewood consumption by at least 40% if used properly. Almost 97% of the households in Chalaco District use firewood for cooking and heating; and firewood scarcity was a critical problem in the area by the time of the improved stove intervention (Ureta (2007)). 27 The stoves were built with an aluminium chimney designed to expel the combustion smoke out of the household dwelling and in this way help reducing the incidence of respiratory illnesses and eye discomfort. 28 Villages in Chalaco District are located in 5 watersheds at altitudes between 1000 m. to 3000 m. 29 During recent field visits in the summer of 2008 I was able to confirm that this was indeed the case.  24  households were supposed to provide the mud bricks for building the combustion box and the stove basement 30 (see figure 1 in Appendix A). Stove installation was also done without monetary cost and it was supported by two village craftsmen selected by the village beneficiaries and trained by the NGO. Beneficiary households were not required to uninstall their traditional stove to get their new improved stove installed. Approximately 85% of all the households in the Chalaco District received an improved stove and by the end of 2003 approximately 95% of all the distributed stoves were installed31. A second stage of stove distribution was originally planned to provide the new device to those households that initially did not ask for one. However, due to administrative reasons as well as other program’s priorities, the second stage was cancelled. The only way a household that initially did not receive an improved stove could have had access to one, was by getting it transferred or sold from another household in the same village or in a close one. The stove monitoring interviews carried during the spring and summer of 2004 indicate these cases were extremely rare; in fact, less than 0.5% of the total beneficiaries visited reported selling or transferring their new stoves to other households. To the best of my knowledge, all the households received the same stove design as well as the same instructions for its installation, usage and maintenance. The improved stove program was coordinated and implemented at the watershed level, and a specific NGO team was allocated to each watershed in the District. Thus, controlling for village watershed location will be important in order to isolate the effect of potential differences in the quality and effort levels of the NGO’s teams allocated to different watersheds. It is important to mention that during the 2004 monitoring visits, some beneficiaries reported material problems with their stoves, mainly deformations in the iron frame and metallic chimney. In the case of iron frame failures, the evidence in the monitoring reports clearly suggests that this were not likely to have been systematically caused by 30  Mud bricks are the most common and accessible material for building, although it requires some labor allocation, I think that providing them should not be excessively costly for the household. 31 2003 MIRHASPERU internal report on improved stove installation.  25  household’s usage patterns, inadequate installation or maintenance; but by lower material quality (see Section 3 in Chapter 3 for more detail on this). Furthermore, the stoves were produced in small local workshops in the main coastal city, and there is no evidence on material quality inspection prior to its distribution. It is also very important to emphasize that there is no evidence on beneficiary households complaining about iron frame’s material quality prior to stove installation and usage; these problems were reported after households made effective use of their new cooking device (Ureta (2007)). Moreover, as NGO members were in charge of stove distribution, it is unlikely that households could have selected their stoves based on observable material characteristics (i.e. thickness of the iron frame). It is also unlikely that some villages may have successfully influenced the NGO in order to obtain stoves of better quality.  2.3.1. Social Learning and Information Diffusion The improved stove technology distributed in the Chalaco District was originally introduced in villages in coastal areas of the Piura Region. The original design was distributed in the villages of the Chalaco District without major modification or adjustment. That is, no special features were introduced ex-ante to adapt the stove to the particular circumstances of these rural communities. For example, the design did not take into account that the stove also performs as a heating device in high altitude areas, where the temperature is relatively colder during the winter months 32 . Also the stove’s combustion box and the operating instructions were designed for the type of firewood that is common in coastal areas and did not take into account the specific varieties of firewood available in the watersheds of the Chalaco District. In addition to this, the way of processing the firewood input in the new stove was quite different from the one that was followed with the traditional open fire cooking technology. Furthermore, it is important to mention that modifications to the combustion box design were not difficult to make, as this was fully made from locally provided mud bricks. Some modifications observed by the responsible NGO include: changes in the combustion box’s internal dimensions; some adjustments to facilitate local foods preparation; and the relocation of 32  The minimum average temperature during the winter season in coastal rural villages is close to 17 Celsius degrees, while at high altitudes villages in the Chalaco District it ranges around 2 to 5 Co degrees.  26  the stove to improve its performance as a heating device. All these suggest that there was plenty of room for experimentation, innovation and learning by doing, and that information on how to use or modify the new device is likely to have been diffused among beneficiaries. Finally, it is important to take into account that villagers are not only likely to communicate each other about how to use the new stove or how to modify it to improve its performance; but also the real benefits they obtained with the device and the perceived quality of the technology. A given household will be more likely to use the stove if others have already experienced real savings in firewood consumption and/or reductions in IAP. On other hand, a given household may be more likely to delay usage decisions if negative news related to the new device are relatively abundant (e.g. deformations in the iron frame).  2.4. Data 2.4.1. Village Social Capital Indicators: 2003 Household Survey From June to August in 2003, a household survey, in which I was directly involved, was executed in all of the villages in the Chalaco District by Universidad de Piura, a local university with a significant experience in development projects in the Piura Region. The NGO-led stove distribution intervention, which took place later in 2003, was completely independent from this household survey and thus did not take into account any of the survey’s results. Likewise, the 2004 stove performance monitoring visits by the NGO were also independent from the earlier household survey’s results. The 2003 survey contains information on members’, dwelling’s and farm’s characteristics, as well as a social capital questionnaire. In the social capital section, the interviewed33 was asked the following trust related questions: How much do you think you can trust in: a) your local neighbours, b) your local organizations, c) people from other villages and d) in strangers? The scale of responses in all cases goes from 0 to 3 (0=nothing, 1=a little, 2=in a regular degree, 3=a lot). If the questions were not clear enough for the interviewed, the interviewer tried to provide specific examples related to the village social life. 33  Interviewers were instructed to apply the social capital questionnaire in half of the cases to the household head and in half of the cases to his (her) spouse. Unfortunately, in many cases they forgot to indicate in the survey document if the person interviewed was a male or female individual.  27  As was discussed in the introductory chapter, some studies in the literature tend to identify social capital with trust; while in others, trust is understood as only a “consequence” of social capital, which is defined in a broader manner as the “nature and extent of village social relationships”, or simply as “interpersonal relationships” (Woolcock (1998), Dasgupta (2005)). For example, for Woolcock (1998) it is important to emphasize that “…trust and norms of reciprocity, fairness, and cooperation are “benefits” that are nurtured in and by particular combinations of social relationships; they are undeniably important for facilitating and reinforcing efficient institutional performance, but they do not exist independently of social relationships” 34. However, Woolcock also adds that although they should not be confused with social capital, these “benefits” or “consequences” can still be used as indicators for the level and different combinations of social capital that is present within a given community. Taking all this into account, in this chapter I use the village average responses to the trust questions in the survey as indicators for the level of social capital that was present in the rural communities in the Chalaco District before the intervention. I let the reader decide whether she prefers to identify social capital with trust, or whether she prefers to refer to trust as only a consequence of social capital and consider this variable as an imperfect indicator, but indicator at least, of village social links. In this chapter, and following Woolcock (1998), I identify two communal dimensions of social capital: embeddedness or bonding social capital and autonomy or bridging social capital. The first dimension refers to the level of social capital within a community (the strength of the intra communal social links); while the second dimension refers to the level of social capital between members of the community and people from other villages (the strength of the social ties between village members and people from neighbouring communities). Table 2.1 describes the relevant survey information on social capital aggregated at the village level for a total of 26 beneficiary villages, which were also visited during the 2004 improved stoves performance monitoring survey 35 . Table 2.2 34  For Dasgupta (2001) good local institutions that clearly define rights and obligations are also a source of communal trust. 35 Villages with higher expected success in stove usage were not systematically targeted during monitoring visits. See section 2.4.2 for a detailed description on how these 26 villages were selected.  28  shows the unconditional linear correlations between the social capital indicators aggregated at the village level.  Table 2.1: Village social capital indicators Variable Villages Mean S.D Households per village 26 48.08 24.05 Sample size per village 26 21.28 8.99 Village proportion of households sampled 26 0.47 0.13 Trust in local neighbours index (t1v) 26 1.45 0.40 Trust in local organization index (t2v) 26 1.89 0.27 Trust in other villages’ people index (t3v) 26 1.47 0.32 Trust in strangers index (t4v) 26 0.60 0.23  Min 19 8 0.19 0.42 1.22 0.65 0.18  Max 126 43 0.74 2.22 2.33 1.92 1.14  The information in this table is presented at the aggregated village level  Table 2.2: Linear correlations between the village social capital indicators t1v t2v t3v Trust in local neighbours index (t1v) 1.00 Trust in local organization index (t2v) 1.00 0.77 *** Trust in people from other villages index (t3v) 0.05 0.22 1.00 Trust in strangers index (t4v) -0.28 0.14 0.34*  t4v  1.00  As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance level.  The village-level trust index in local neighbours (t1v) will be used in this study as an indicator for the village-level of bonding social capital. The index for trust in people from other villages (t3v) will be our indicator for the village-level of bridging social capital. The village index for the level of trust in strangers (t4v) can be interpreted as an indicator of generalized trust in the context of this research. The evidence in Table 2.2 indicates that “trust in local neighbours” (t1v) is positively and significantly correlated with “trust in local organizations” (t2v) (suggesting that probably the last one can also be used as an indicator of village bonding links); while the bridging social capital indicator “trust in people from other villages” (t3v) is positively and significantly correlated with “trust in strangers” (t4v) (both of them measures for the strength of the community members’ links with agents outside the village). Interestingly, the linear correlation coefficient between the bonding social capital indicator (t1v) and the bridging social capital indicator (t3v) is relatively small and not statistically significant. This result suggests that the bonding and bridging dimensions of village social links do not necessarily tend to move in the same direction in a given village at a given period of time.  29  How village social capital influences individual technology usage decisions is the central question this research tries to answer. The main hypothesis is that bonding social capital plays a central role facilitating the diffusion of information: in villages where bonding social capital is strong, information should be more intensively disseminated (Woolcock (1998), Dasgupta (2005)). Furthermore, this research will also demonstrate that the marginal impact of bonding social capital on the individual usage decision will be closely related to the concrete experience that households within a given community are having with the new technology at initial adoption stages: a low initial village-level rate of success (or a high rate of failure) in technology usage may encourage the diffusion of “negative” information about the new technology through the village network.  2.4.2. Adoption Patterns: 2004 Stove Monitoring Survey From April to August in 2004, the NGO (MIRHASPERU) and Universidad de Piura monitored improved stove performance in 26 beneficiary villages. During these visits, interviewers had full access to the kitchen area and were able to confirm the accurate conditions of improved stove usage. Members of Universidad de Piura involved in the monitoring visits reported that visits to all of the villages were originally planned. However, during the initial stages of the monitoring process, special emphasis was set in visiting beneficiaries in villages at high altitudes and less accessible areas, in which relatively low usage rates were expected 36 . It is also important to point out that monitoring visits were not planned in terms of the expected level of village social capital, as the social capital survey’s results were still not available by the time of these visits. Due to budgetary constraints and security issues (in certain areas located relatively close to mining explorations) some beneficiary villages were finally not visited37. As it is shown in Table 2.3, on average 82% of the improved stove beneficiaries were visited in every village. It is important to mention that there is no evidence on households refusing to be interviewed. In most cases uninterviewed households were not at their dwelling units at the time of the visits. Some were out for social visits in neighbouring 36  Then, in any case the sampling procedure should work against our social learning hypothesis. Most of the villages that were not visited are placed at low altitude areas, where the stove is only used for food preparation. In 2008 I was able to confirm that adoption rates were relative high in these villages. 37  30  villages or buying food or tools in the main district town. In other cases, they were still working at their farm plots. In some situations monitors ran out of time during the monitoring visits. Table 2.3 indicates that approximately 45% of the visited beneficiaries per village reported using the improved stove as their main cooking device. Among these users, 28% reported some problem with the new technology (that is, on average 12% of the beneficiaries per village were using the improved stove with some problem), and 72% did not report any problem or complication (that is, on average 33% of the beneficiaries per village were using the stove without any problem). The empirical estimations in this chapter employ the proportion of beneficiaries in the second group (the proportion of beneficiaries using the improved stove without problems) as an indicator of the village level of information on how to properly use the new stove; as well as of the village level of information related to “positive” improved stove performance. In the next sections I simply refer to this group as “the village proportion of users without problems”; on the other hand, I refer to the proportion of beneficiaries using the improved stove with some problem as “the village proportion of users with problems”.  Table 2.3: Improved stove usage patterns at the village level Variable Villages Mean S.D. Number of households per village 26 48.01 24.05 Total number of beneficiary households per village 26 40.71 17.05 Village number of visited beneficiaries per village 26 33.23 14.71 during 2004 stove monitoring visits Village proportion of visited beneficiaries during 2004 26 0.82 0.13 stove monitoring visits Village proportion of visited beneficiaries using the 26 0.45 0.20 improved stove as the main way of preparing food Proportion of users that reported problems using the 26 0.28 0.25 improved stove Proportion of users that did not reported problems using 26 0.72 0.25 the improved stove Village proportion of visited beneficiaries that received 26 0.55 0.19 the improved stove but were not making use of it. Proportion of non users that decided to uninstall the 26 0.32 0.21 improved stove  Min 19 15  Max 126 88  10  76  .50  1  0.06  0.71  0  1  0  1  0.29  0.94  0  0.73  The information in this table is presented at the aggregated village level  In Table 2.4 (below) we can observe that the improved stove users with problems were mainly affected by material deficiencies, particularly deformations in the improved stove’s iron frame and/or chimney. As mentioned before, the evidence from the  31  monitoring visits suggests that there were some differences in improved stoves’ material quality and that these were exogenous to beneficiaries’ and villages’ characteristics (see Section 3 in Chapter 3 for more detail on this issue). We can also see in Table 2.4 that a small proportion of improved stove users with problems reported excessive firewood usage. As the improved stove also has other expected benefits (such as reduced exposure to indoor air pollution); it is possible that these households continued making use of the new stove because the other benefits compensated for their higher firewood consumption levels. Among non users, the principal difficulty reported was the higher firewood consumption by the new stove. Some of them also reported material problems, but the proportion is in this case significantly lower than the observed for users with problems38. Table 2.4: Main problems encountered by improved stove beneficiaries (%) Users that reported problems Material problems (iron frame deformed, chimney broken or both) Stove uses more firewood Non Users Material problems Stove uses more firewood It is not good for heating the house It is hard to get used to it It is dangerous It is time consuming None  90.4 21.9 15.6 34.1 5.6 10.5 2.0 8.1 54.8  The percentages indicate the total proportion of beneficiaries in each category that reported the problem, which explains why in each case the percentages do not add to 100%.  Using the full set of observations in the 2004 monitoring dataset (in total 878 households were visited); Table 2.5 (below) shows the linear correlations between the binary variable representing the household’s decision to use the improved stove as the main cooking device (in other words the usage decision) and some village-level variables of interest. Note that the usage decision is positively and significantly correlated with the village proportion of users without problems, as well as with the village indexes related to bonding social capital: trust in local neighbours and trust in local organizations. On the other hand, the usage decision is not significantly correlated with the village proportion of users with problems and with the village level of trust in strangers. Table 2.5 also shows that effective usage is significantly and negatively correlated with altitude and 38  In some cases affected households were able to overcome this material problem; while in other cases they stopped making use of the new stove technology. More detail on this issue is provided in Chapter 3.  32  significantly and positively correlated with road access. The linear correlation is negative and statistically significant between the household’s usage decision and the household’s village location in the Mijal and Cerro Negro watersheds, and positive and statistically significant between the usage decision and location in the Ñoma watershed39. Table 2.5: Linear correlations between the household’s improved stove usage decision and village usage patterns, social capital and geographic characteristics Household uses the stove as the main cooking device (Yes=1, No=0) Correlation coefficient: Village proportion of users without problems 0.28 *** Village proportion of users with problems 0.04 Village proportion of beneficiaries 0.06 ** Trust in local neighbours index 0.08 *** Trust in local organizations index 0.09 *** Trust in strangers 0.04 Village Altitude -0.22 *** Village is accessible by road (yes=1, no=0) 0.15 ** Household’s village is located in Mijal Watershed (M1=1) -0.22 *** Household’s village is located in Nogal Watershed (M2=1) 0.04 Household’s village is located in Potros Watershed (M3=1) 0.04 Household’s village is located in Noma Watershed (M4=1) 0.22 *** Household’s village is located in Cerro Negro Watershed (M5=1) -0.14 *** N= 878. ***,** and * indicate statistical significance at the 1%, 5% and 10% significance levels  Matching the observations in the 2003 socioeconomic survey and the 2004 stove monitoring data, a total sample of 283 household observations is available for estimation purposes40. Average characteristics for improved stove users and non users are shown in Table 2.6 (below), as well as the p-value for the difference in raw means test for the variables included. As we can observe, the proportion of household heads that have at least secondary education is significantly higher for improved stove users than for non users41. Note also that improved stove users are on average wealthier than non users, but the difference in raw means is not statistically significant. In order to account for household’s involvement in communal activities, entrepreneurship, previous experience with new technologies and preferences for environmental and women related issues, the 39  The Ñoma watershed has relatively good road accessibility during the most part of the year (even in the rainy season). Villages in Mijal and Cerro Negro watersheds are placed at relatively high altitudes and have poor accessibility conditions, especially during the rainy season. 40 For two villages the agricultural section is missing; this leaves us with only 24 villages for estimation. 41 Although it is preferable to control for the educational level of the household head wife, I expect the level of education for the head and his wife to be relatively correlated.  33  following variables are respectively considered: household’s past participation in communal activities, household’s experience with fertilizers and in the elaboration of processed agricultural and animal products (i.e. alcoholic beverages, wheat flour, cheese, etc), and household’s membership in environmental and female based communal organizations. As we can observe in Table 2.6, improved stove users are significantly more involved in communal activities; users and non users are equally likely to use fertilizer or elaborate processed products; and non users have a statistically significant higher degree of participation in environmental groups. Although this last result seems to go in the wrong direction, such groups are relatively abundant at high altitude areas, where the new stove was less likely to meet the household’s heating needs (the initial design did not take into account that at high altitudes the stove is also used as a heating device). Another plausible explanation for this result is that, based on its initial performance, the stove was perceived as a poor technology in terms of firewood consumption; and then households with stronger pro-forest preferences were less likely to effectively use the new improved device.  Table 2.6: Main household level characteristics for improved stove users and non users Users Non users Test of equality N=155 N=128 (p-value) Household head’s sex (male=1, female=0) 0.85 0.91 0.19 51 49 Household head’s age 0.21 (14) (13) 2.72 2.95 Household’s number of adults 0.20 (1.39) (1.61) Adult female in the household (yes-1, no=0) 0.88 0.89 0.73 Household head has secondary education or higher 0.20 0.11 0.03 Household head attended school and has at most primary 0.74 0.78 0.35 education Household head did not attended school 0.06 0.10 0.17 2.66 2.96 Household’s farm size in has. 0.34 (2.72) (3.18) 84 68 Household’s value of farm assets (in Peruvian soles) 0.22 (13.4) (5.6) Household elaborates processed products (yes=1, no=0) 0.55 0.57 0.69 Household uses fertilizer (yes=1, no=0) 0.63 0.68 0.27 Household participated in communal activities during the last 0.50 0.34 0.01 12 months (yes=1, no=0) Household is a member of the local environmental group 0.31 0.54 0.00 (yes=1, no=0) Household’s membership in mothers club (yes=1, no=0) 0.23 0.25 0.61 Standard deviations shown in parenthesis.  34  2.5. Empirical Strategy A given household “i” in a particular village “j” will be more likely to effectively use a new technology if the net economic gains “ a*ij ” it expects from using it are non negative. Suppose that for the case of the improved firewood stove technology introduced in the Chalaco District the expected economic gains are linear in a vector of household level characteristics “ X ij ”, a village level effect “ W j ” and a household specific error term “ uij ”. The reduced form equation is then given by: (1) aij* = α0 + α1 X ij + W j + uij (Where “i” refers to households and “j” to villages) The village effect in (1) is assumed to be function of a non-stochastic component and a village error term. The non-stochastic portion includes a village informational term “ I j ”, which measures the amount of information related to improved stove usage and performance available within the village, and a vector of other village level controls “ Y j ”. The informational term “ I j ” is defined as a non linear function of the technology usage patterns within the village “ APj ” and of the village level of bonding social capital “ SC j ”42; that is I j = I j(APj ,SC j ) . The total village effect is then given by: (2) W j = I j (APj , SC j ) + βY j + e j For now, I prefer not to define a specific functional form for the informational component; in the next sections different model specifications will be empirically tested. In some of them the total proportion of improved stove users will be employed as an indicator of the within village usage patterns; while in others I will distinguish between the proportion of beneficiaries using the new stove with and without problems. Taking (1) and (2) together, the following expression for the net household’s expected gains is obtained: (3) a*i.j = α0 + α1 X ij + I j (APj ,SC j ) + βY j + e j + uij 42  In Section 2.9 I corroborate that only the bonding social capital indicator significantly influence the effect within village usage patterns have on individual usage decisions; while the bridging social capital indicator only influences the effect usage patterns outside the village have on the household’s usage decision.  35  In general, the household’s expected gains “ a*ij ” are unobservable to the econometrician, who is only able to observe the household’s usage decision. In this chapter this decision is represented by the binary variable “ ai.j ”, which takes the value of one if the household uses the new stove as the main cooking device and zero otherwise. Then, the probability with which household “i” in village “j” will use the new technology is given by:   (4) P(a ij = 1 ) = P vij > −{ α0 + α1 X ij + I j (APj ,SC j ) j + βY j }  , where vij = e j + uij    Expression (4) suggests a probit regression to estimate the usage likelihood; however, it is known that the linear probability model is more amenable to the estimation of alternative functional forms for “ I j (.) ” and that the computation of the marginal effects is more transparent when higher order polynomials are fitted onto “ I j (.) 43” (Bandiera and Rasul (2006)). Taking this into account, in the next sections I use the following linear probability model to empirically test the information diffusion hypothesis: (5) ai.j = α0 + α1 X ij + I j (APj ,SC j ) + βY j + vij  2.6. Baseline Estimation Results The empirical specifications in this section have been estimated using a linear probability model; the results for the probit regression estimates (not shown here) are very similar to the ones obtained in the OLS regressions. All the regressions in Table 2.7 control for household’s characteristics 44 , the village total proportion of beneficiaries and include watershed location dummies in order to control for geographic and other characteristics (weather, NGO effort, etc) which tend to be relatively similar among villages within the same watershed. The informational term I j(APj ,SC j ) is defined in column I in Table 2.7  43  It is also important to note that the total usage rate is fairly close to 50% and that in the OLS estimation less than 2% of the predicted estimates lie outside the unit interval. 44 These include: household head´s sex and age, household head´s level of education, household’s number of adults, presence of a female adult member, household’s wealth (value of farm assets), farm size, household’s participation in women and environmental organizations, household’s elaboration of processed products and fertilizer usage and household’s participation in local activities in the previous 12 months  36  as a function of the total proportion of improved stove users ( TPj )45, the square term for this total proportion and the village bonding social capital indicator ( SC j ): the villagelevel of trust in local neighbours. The informational term in equation 5 is then given by:  (6) I j(TPj ,SC j ) = λ1TPj + λ2TPj2 + λ3 SC j  Before moving any further in this section, I would like to briefly discuss the inclusion of a quadratic term in equation (6). In first place, I must emphasize that improved stove monitoring visits took place approximately 8 to 12 months after the distribution of these devices; that is, at very initial adoption stages. Then, as it has been previously discussed by Bandiera and Rasul (2006), we have to take into account that at these stages of adoption individuals are very likely to behave strategically. In the empirical section of their paper, Bandiera and Rasul (2006) find that the effect of network adoption on the individual adoption likelihood is nonlinear and decreasing in the proportion of network adopters; moreover, after a given threshold the network effect becomes negative. In other words, their study strongly suggest that the higher the levels of adoption within the reference network at early adoption stages, the stronger the incentives the household has to delay its own adoption and free ride on its neighbours’ learning and experimentation. Taking into account this previous finding, and given that this chapter also focuses on effective usage decisions at very initial adoption stages (within the first year of improved stove distribution), I allow the effect the village proportion of improved stove users has on the individual usage likelihood to be nonlinear (a linear and a quadratic term for this proportion are included in the main regressions). Now, if strategic behaviour is a relevant feature of beneficiaries’ usage decisions at early adoption stages, our model specification is the appropriate one; and under correct model specification our nonlinear in means model is not likely to be affected by the reflection problem (Manski (1993)). As originally defined, the reflection problem is an issue of collinearity. In Manski’s (1993) seminal paper endogenous effects are not identified 45  This is the total proportion of beneficiaries using the improved stove (includes users with and without problems), and it is equal to the total number of users in the village divided by the number of beneficiaries.  37  because they are a linear combination of exogenous and correlated effects. In other words, non identification in social interaction models due to the presence of the reflection problems is intrinsically linked to linearity; for the case of nonlinear in means social effects and under correct model specification, social effects are normally identified 46 (Brock and Durlauf (2001b) and (2005)). In Appendix B at the end of this thesis, I provide additional technical detail for the argument by Brock and Durlauf (2001b) on the possibility of local identification in the context of nonlinear in means models.  Table 2.7: Village level determinants of the household’s decision to use the improved stove as the main cooking device I II III IV 0.0243*** Village total proportion of users (0.0062) -0.0003*** Village total proportion of users^2 (0.0001) 0.0184*** 0.0199*** Village proportion of users without problems (0.0047) (0.0051) Village proportion of users without -0.0002*** -0.0003*** problems^2 (0.0001) (0.0001) 0.0054 -0.0011 Village proportion of users with problems (0.0105) (0.0101) -0.0002 -0.0002 Village proportion of users with problems^2 (0.0003) (0.0003) Village-level trust in local neighbours -0.0299 -0.0578 -0.0324 -0.0794 (bonding social capital indicator) (0.0809) (0.0824) (0.0842) (0.0913) 0.0032 0.0054** 0.0047*** 0.0044** Village proportion of beneficiaries (0.0016) (0.0019) (0.0020) (0.0020) 283 N 283 283 283 24 Villages 24 24 24 R2  0.21  0.21  0.19  0.23  The dependent variable is the decision to use the improved stove as the main cooking device. All regressions in this table control for watershed dummies and include as household level controls the household head´s sex and age, household’s head level of education, household’s number of adults, presence of a female adult member in the household, household’s wealth (measured by the value of farm assets), farm size, household’s participation in women and environmental organizations, household’s elaboration of processed products and usage of fertilizer and household’s participation in local activities in the previous 12 months. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  Moving back to our estimation results; in column I in Table 2.7 we can observe that the marginal effect of the total proportion of improved stove users on the household’s usage 46  As Brock and Durlauf also discuss, non identification due to the presence of the reflection problem is not likely to fail in the binary choice model. For the binary choice model, formal statements of conditions for identification appear in Brock and Durlauf (2001a, b) for the case when the random terms εi are logistically distributed, and in Brock and Durlauf (2004) for general distribution functions. I have also estimated the main regressions in this section using the probit and logit models with almost identical results.  38  likelihood is clearly nonlinear. Moreover, this marginal effect will be positive only if the total proportion of users within the village is below 44%. Note also that the coefficient for the bonding social capital indicator in this specification appears as not statistically significant and relatively small in absolute size. In column II in Table 2.7 I replace the total proportion of improved stove users “ PT j ” for the proportion of users without problems47 “ P1 j ” in the village informational term. In my opinion, only this group of users played a positive role influencing the individual decision to effectively use the new stove as the main cooking device. In other words, information on how to use the new technology or positive information on the improved stove real performance and associated benefits is more likely to have been disseminated by this group of users. The village informational effect is then defined as: (7) I j(P1j ,SC j ) = λ1 P1j + λ2 P12j + λ3 SC j We can observe in column II that the proportion of users without problems has a statistically significant effect on the household´s usage decision, and that (as it was the case for the total proportion of users) this effect is clearly nonlinear. Compared to the estimation results in column I, we can see that the coefficients’ point estimates for the village-level usage terms are relatively lower in absolute value. The results indicate that the marginal effect of the village proportion of users without problems on the household’s decision to effectively use the new stove will be positive in column II only if this proportion is below 42%. Note also that, as in column I, the coefficient for the bonding social capital indicator (the village-level trust in local neighbours) appears as not statistically significant and relatively small in absolute size. To confirm that only users without problems have a significant nonlinear effect on the individual household’s usage decision, column III only includes the proportion of users that reported problems48 with the improved stove “ P 2 j ” in the village informational term. 47  As mentioned before, this is just equal to the village number of beneficiaries using the improved stove “without” problems divided by the total number of beneficiaries. 48 This is equal to the village number of beneficiaries using the new stove “with” some problem divided by the total number of beneficiaries.  39  As expected, the results show that the effect of this proportion of users on the household´s likelihood to use the stove is not statistically significant. Finally, the empirical specification that corresponds to column IV in Table 2.7 includes at the same time both user proportions: stove users with problems " P 2 j " and stove users without problems " P1 j " (the linear and the quadratic term for each proportion are included in the regressions). The village informational term is then given by: (8) I j(P1 j ,P 2 j , SC j ) = λ1 P1 j + λ2 P12j + λ3 P 2 j + λ4 P 2 2j + λ5 SC j  The results in column IV confirm that the village proportion of users without problems has a significant nonlinear effect on the household’s usage decision; while the effect of the proportion of users with problems is not statistically significant 49 . As monitoring visits were performed at very initial adoption stages, it is very likely that the observed nonlinear effect the proportion of stove users without problems has on the usage decision reflects the presence of strategic behaviour: the higher the proportion of beneficiary users without problems in the village, the more likely is the household to delay its own usage decision and free ride on others’ learning and experimentation. Interestingly, in all the model specifications in Table 2.7, the coefficient for the bonding social capital indicator (the village-level trust in local neighbours) always appears as not statistically significant and relatively small in absolute size50. As mentioned before, this research’s main hypothesis states that the informational effect of social capital on the household´s usage decision should be closely linked to the initial usage patterns within the village. Then, to capture the informational effect of village bonding links on the usage likelihood, we must allow the marginal impact of the bonding social capital indicator to 49  I have also estimated a linear regression in which the effect of the total proportion of users with problems is considered to be only linear (that is no quadratic term for this proportion is included); the coefficient for the linear term is in this case also not statistically significant. 50 As discussed before, in Table 2.7 I include watershed location dummies in order to control for geographic and other characteristics that tend to be relatively similar among villages within the same watershed. As a robustness check, I have also estimated the same regressions in Table 2.7 adding some specific village characteristics such as altitude, road access and trust in strangers. As it can be observed in Table C.1 in the Appendix C, when these additional village controls are included, the results are very similar to the ones obtained in Table 2.7.  40  depend on the initial village-level usage patterns. Moreover, if information diffusion is the process behind the strong correlations observed in Table 2.7; it makes sense to expect that the individual usage decision will be more sensitive to the within village usage patterns in those communities with stronger bonding links, where information should be more intensively diffused. All the specifications in Table 2.8.A take this issue into account, and introduce in the village informational term an interaction term between the within village usage patterns and the village bonding links indicator (the village-level trust in local neighbours).  Table 2.8.A: Village level determinants of the household’s decision to use the improved stove as the main cooking device (allowing for an interaction term between village usage patterns and the village bonding social capital indicator) Village total proportion of users Village total proportion of users^2 Village total proportion of users * Village-level trust in local neighbours  I 0.0203*** (0.0069) -0.0003*** (0.0001) 0.0045 (0.0044)  II  IV  0.0215** (0.0086) 0.0002 (0.0004) -0.0220** (0.0078)  0.0191*** (0.0065) -0.0004*** (0.0001) 0.0025 (0.0035) 0.0098 (0.0113) 0.0004 (0.0003) -0.0238** (0.0094)  0.0121* (0.0062) -0.0002*** (0.0001) 0.0056** (0.0026)  Village proportion of users without problems Village proportion of users without problems^2 Village proportion of users without problems * Villagelevel trust in local neighbours Village proportion of users with problems Village proportion of users with problems^2 Village proportion of users with problems *Village-level trust in local neighbours Village-level trust in local neighbours (bonding social capital indicator)  III  -0.2561 (0.2134) 0.0052***  -0.2867** (0.1285) 0.0048**  0.1989** (0.0950) 0.0019  0.0438 (0.2136) 0.0056**  N  (0.0016) 283  (0.0019) 283  (0.0022) 283  (0.0024) 283  Villages  24  24  24  24  R2  0.21  0.21  0.20  0.23  Village proportion of beneficiaries  The dependent variable is the decision to use the improved stove as the main cooking device. All regressions in this table control for watershed dummies and include as household level controls the household’s head sex and age, household head’s level of education, household’s number of adults, presence of a female adult member in the household, household’s wealth (measured by the value of farm assets), farm size, household’s participation in women and environmental organizations, household elaboration of processed products and usage of fertilizer and household participation in local activities in the previous 12 months. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  41  As in Table 2.7, all the model specifications in Table 2.8.A control for household´s characteristics, the village usage patterns, the bonding social capital indicator, the total proportion of beneficiaries and watershed location. The specification that corresponds to column I in Table 2.8.A defines the informational component in equation (5) as a nonlinear function of the total proportion of users “ TPj ”, the bonding social capital indicator and include an interaction term between these two village level variables. The village informational term is then given by: (9) I j(APj ,SC j ) = λ1TPj + λ2TPj2 + λ 3TPj .SC j + λ4 SC j As it can be observed in column I in Table 2.8.A, the coefficient for the interaction term between the total proportion of users and the social capital indicator is positive but not statistically significant, and the coefficient for the village-level trust in local neighbours also appears as not statistically significant. As previously discussed, not all stove users influence the usage decision in the same manner, and the social network is expected to disseminate the information provided by different user types in dissimilar ways. As the specification in column I does not distinguish between the proportions of stove users with and without problems, it is not a surprise that the interaction term and the linear social capital indicator term appear as not statistically significant in this regression. The model specification that corresponds to column II in Table 2.8.A includes in the village informational term the proportion of stove users without problems instead of the total proportion of users, and introduces an interaction term between this proportion and the village bonding social capital indicator. The village informational term is then given by:  (10) I j(APj ,SC j ) = λ1 P1j + λ2 P12j + λ 3 P1 j .SC j + λ4 SC j The results in column II show that the interaction term between the proportion of users without problems and the bonding social capital indicator is positive and statistically significant at the 5% significance level. This results confirm the key role the proportion of successful users has played at influencing the household´s usage decision, and clearly indicate that the impact of this group of users will be higher in villages in which the  42  village-level trust in local neighbours (our indicator for bonding social capital) is higher; probably because in these villages the information provided by them will be more intensively diffused. For example, when evaluated at the sample means, the results in column II indicate that if the proportion of users without problems marginally increases by one percentage point, the usage likelihood increases by 0.2 percentage points; while if we evaluate this effect at the observed maximum level of village bonding social capital, a one percentage point increment in the proportion of users without problems will increase the usage likelihood by 0.6 percentage points (that is the marginal effect will be three times higher). Note that the coefficient for the linear bonding social capital term is statistically significant in column II, but has a negative sign. This result implies that although the marginal effect of the village bonding links indicator on the individual usage decision is increasing in the proportion of users without problems, it will be positive only if this proportion is relatively high (above 51%). This finding supports the hypothesis that the effect of village social capital on the individual usage decision is critically linked to the initial performance of the technology within the village: if the village level of success in adopting the technology (measured by the village proportion of users without problems) is relatively low, the village network is more likely to diffuse “negative” information about the new technology. To confirm that only village users without problems positively influence the household’s usage decision through its interaction with the bonding social capital indicator, the specification in column III only includes the village proportion of users with problems in the village informational term, as well as its interaction with the social capital indicator. Interestingly, the results indicate that the coefficient for this interaction term is negative and statistically significant, suggesting that this proportion of users is more likely to negatively influence the usage decision in those communities with higher levels of trust in local neighbours. For example, when evaluated at the sample means, if the proportion of users with problems increases by one percentage point, the usage likelihood decreases by 0.4 percentage points; while if we evaluate this effect at the maximum level of village bonding social capital, the usage likelihood will decrease by 2.1 percentage points (this effect will be five times higher!). Note also in column III that the coefficient for the linear  43  bonding social capital term is statistically significant but has a positive sign. Taken together, these results indicate that if the proportion of beneficiaries that experience problems while using their new stoves is relatively high (more than 10%), the marginal impact of bonding social capital on the household’s usage decision is more likely to be negative. In other words, if a relatively high proportion of beneficiaries encounter problems with the new stove, the communal social networks are more likely to diffuse “negative” information about the new technology. So far the results in Table 2.8.A provide strong support for the main hypothesis51 in this chapter: information diffusion seems to be increasing in the strength of village bonding links and, interestingly, social capital appears to diffuse the village-level usage information in the “right” direction: village usage levels without problems encourage individual usage decisions through its social capital effect (measured by the interaction term between the proportion of users without problems and the bonding social capital indicator); while village usage levels with problems appears to do exactly the opposite. In column IV both user proportions are included in the regression, as well their interaction terms with the bonding social capital indicator. The informational term is then defined as: (11) I j (APj ,SC j ) = λ1 P1 j + λ2 P12j + λ 3 P1j *SC j + λ4 P 2 j + λ5 P 2 2j + λ 6 P 2 j *SC + λ7 SC j The result in column IV in Table 2.8.A corroborates that only the proportion of users without problems positively influences the household’s usage likelihood through its interaction with the bonding social capital indicator; however the coefficient for this interaction term appears as not statistically significant. As before, the proportion of users with problems appears to negatively affect the individual usage decisions through its interaction with the bonding social capital measure, and the coefficient for this interaction term is statistically significant. Interestingly, when both interaction terms are added at the same time, the coefficient for the linear bonding social capital indicator term appears as not statistically significant and very small in absolute size. This specific result suggests 51  I have also estimated the regressions in Table 2.8 using two alternative measures for bonding social capital: village-level trust in local organizations and a within village communication index. Very similar results (not shown here) were obtained in both cases.  44  that the effect bonding social capital has in the context of technology adoption in rural communities, is closely linked to the initial performance of the new technology at the village level. Then, when village usage levels with and without problems are included at the same time, their respective interaction terms with the within village social capital indicator fully capture the informational effect of bonding links52. Finally, it is important to highlight that the results in Table 2.8.A indicate that the absolute size of the interaction term coefficient is significantly higher for the village proportion of users with problems, suggesting that the “bad” news about a new technology that are diffused through the communal network may have a higher impact on the usage likelihood than “good” ones, particularly in villages with high levels of bonding social capital. The finding that peer effects are heterogeneous in the bonding social capital indicator, as well as the finding that the marginal effect of the bonding social capital indicator is closely linked to initial village usage patterns (the effect of social capital may be negative if initial village-level technology usage success is low), provide strong empirical support towards the significant role bonding social capital has played in the diffusion of information related to improved stove usage and performance. Furthermore, as it was measured before the intervention, it is unlikely that the village-level trust in local neighbours has been critically influenced by the improved stove adoption process; in other words, our estimates should be free from endogeneity problems due to reverse causality between social capital and individual adoption decisions. The baseline results in Table 2.8.A are not significantly affected when we include additional village level controls which may not being properly captured by the inclusion of the watershed dummies. As we can see in Table 2.8.B, when we also include village altitude, road access and the village level of trust in strangers, the estimation results are very similar to those in Table 2.8.A. Village altitude and road accessibility have been identified by the improved stove program technicians as two of the most important factors influencing stove performance at initial adoption stages. In first place, the stove 52  Interestingly, if I exclude the social capital linear term from the regression in columns IV (keeping the interaction terms), both interaction terms have the expected signs and are statistically significant at the 5% significance level.  45  design did not take into account the heating needs of households at high altitude villages; in second place, road accessibility was very likely to influence the effort levels of the NGO’s members during the distribution and training stages. Then, the fact that our results are robust to the inclusion of these controls alleviates some concerns in the direction that village specific characteristics may be the ones driving the estimation results.  Table 2.8.B: Village level determinants of the household’s decision to use the improved stove as the main cooking device (allowing for an interaction term between village usage patterns and the village bonding social capital indicator and including additional village controls) Village total proportion of users Village total proportion of users^2 Village total proportion of users * Village-level trust in local neighbours  I 0.0225 *** (0.0078) -0.0005*** (0.0002) 0.0092 (0.0057)  II  IV  0.0284*** (0.0090) 0.0005* (0.0003) -0.0277***  0.0145* (0.0078) -0.0005*** (0.0002) 0.0053 (0.0038) 0.0177 (0.0145) 0.0006* (0.0003) -0.0330***  (0.0069)  (0.0126)  0.0062 (0.0081) -0.0004*** (0.0001) 0.0102*** (0.0034)  Village proportion of users without problems Village proportion of users without problems^2 Village proportion of users without problems * Villagelevel trust in local neighbours Village proportion of users with problems Village proportion of users with problems^2 Village proportion of users with problems *Village-level trust in local neighbours Village-level trust in local neighbours (bonding social capital indicator)  III  -0.4867 (0.3094) 0.0064***  -0.4882** (0.1997) 0.0082***  0.2550** (0.1024) 0.0002  0.0188 (0.2566) 0.0087**  (0.0021) 0.0669  (0.0019) 0.0556  (0.0021) -0.1695  (0.0033) 0.0823  (0.2148)  (0.1518)  (0.1304)  (0.1637)  -0.0378**  -0.0613***  -0.0396**  -0.0645**  (0.0142)  (0.0219)  (0.0164)  (0.0305)  0.1775**  0.1801**  0.1613***  0.2231**  (0.0811)  (0.0701)  (0.0418)  (0.0878)  N  283  283  283  283  Villages  24  24  24  24  0.24  0.24  0.23  0.26  Village proportion of beneficiaries Village-level trust in strangers Village Altitude Village Road Access  R2  The dependent variable is the decision to use the improved stove as the main cooking device. All regressions in this table control for watershed dummies and include as household level controls the household’s head sex and age, household’s head level of education, household’s number of adults, presence of a female adult member in the household, household’s wealth (measured by the value of farm assets), farm size, household’s participation in women and environmental organizations, household elaboration of processed products and usage of fertilizer and household participation in local activities in the previous 12 months. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  46  Despite the strong evidence provided in Tables 2.8.A and 2.8.B; it can be argued that households in villages with higher levels of intra-village trust (stronger bonding links) are more likely to share the same type of unobservable factors and shocks; and that the higher impact the village proportion of users without problems appears to have in village with stronger bonding social capital is just capturing this situation. In order to provide additional support for the social learning hypothesis, in Section 2.7 it will be shown that the social capital information diffusion hypothesis also applies to the decision to dismantle the improved stove among beneficiaries that reported not using the new stove during the monitoring visits, which makes it unlikely that correlated unobservables or other social process are driving the main results in this chapter. Before addressing this issue, I will first briefly comment on the main household-level factors that influence the household’s decision to use the improved stove as the main cooking device.  2.6.1. Household Level Determinants of Improved Stove Usage This section focuses on the household-level determinants of stove usage that have been included as controls in the main estimations in Section 2.6: household head’s sex, age and level of education, number of adults, presence of a female adult member, wealth (farm assets value), farm size, participation in women and environmental organizations, elaboration of processed products, fertilizer usage as well as participation in communal activities in the 12 previous months. Table 2.9 presents the household-level coefficients that correspond to columns I to IV in Table 2.8.B, where in addition to village usage patterns and social capital, also the village proportion of beneficiaries, watershed location, altitude, road access and village trust in strangers were controlled for53. As we can observe in Table 2.9, the coefficients for the household-level controls are relatively stable in terms of size, sign and statistical significance across the four columns. The results in this table indicate that the household’s composition has a significant impact on the improved stove usage decision. In first place, the household’s number of adults has a nonlinear statistically significant effect on the usage likelihood: the linear term’s coefficient is negative and the quadratic term coefficient is positive (the effect is however 53  The household´s coefficients for the estimations in Table 2.8.A (not shown here) are very similar.  47  always negative for the observed sample values), indicating that this effect is decreasing in absolute value in the total number of adults. A higher number of adults in the household implies that labor for firewood collection is relatively more abundant, which decreases the cost of collecting firewood and then has a negative impact on the usage decision; however, this negative effect decreases in absolute size as the number of adults increases, probably due to the presence of decreasing returns in firewood collection as well as due to a higher firewood demand that results from increasing cooking and heating needs. The number of adults is not the only household composition factor that significantly influences the usage likelihood; Table 2.9 also shows that having at least one adult woman member increases the usage likelihood by 15 to 18 percentage points. This result is in line with the fact that women are the main expected beneficiaries of the new stove, as in most cases they are in charge of cooking tasks, firewood collection, and usually spend a higher amount of time inside the dwelling unit. Also, women are likely to assign a higher weight on infants’ health outcomes; then, they are more likely to support the usage of the new stove as a way to reduce the incidence of respiratory illnesses among these family members. The results in Table 2.9 also indicate that wealthier households are significantly more likely to effectively use the improved stove; a wealth increase of 10.00 “nuevos soles” (approx. CAN$ 3.50) increases the usage likelihood by approximately 0.6 percentage points. This result suggests that richer households are more likely to bear the initial costs of technology adoption; for example, they are more likely to afford the higher amounts of firewood consumption that may be incurred during initial trials with the new technology. The results also show that having been involved in communal activities in the 12 months previous to the survey increases the usage likelihood by 15 percentage points. Probably this group of households has been more likely to participate in development projects in the past and by so are more proactive and open to new technologies. It can also be the case that as these households are more likely to participate in the communal social life, they tend to be more exposed to information related to the use of the new device54.  54  When I average at the village level the responses to the past participation in local activities question, and use it as a potential measure of social capital, the interaction between this variable and the village usage  48  Table 2.9: Household-level factors affecting the improved stove usage decision II IV VI Household head´s sex Household head´s age Household’s number of adults Household’s number of adults^2 Adult female member present in the household (yes=1) Household’s head has formal education at the maximum level of primary school (yes=1) Household’s head has formal education at the level of secondary school or higher (yes=1) Household’s members participated as organizers or supporters in local activities in the past 12 months (yes=1) Household’s value of farm assets in Peruvian Soles Household belongs to a local environmental group (yes=1) Household belongs to a local women based organization (yes=1) Household uses fertilizer (yes=1) Household elaborates processed products (yes=1) Household’s farm size  VIII  -0.062  -0.021  -0.039  -0.027  (0.074)  (0.071)  (0.077)  (0.074)  0.002  0.002  0.002  0.002  (0.002)  (0.002)  (0.002)  (0.002)  -0.132**  -0.119**  -0.111*  -0.127**  (0.056)  (0.056)  (0.057)  (0.056)  0.013**  0.013**  0.012*  0.012**  (0.006)  (0.006)  (0.006)  (0.005)  0.127  0.153*  0.182*  0.169*  (0.078)  (0.085)  (0.092)  (0.086)  0.003  0.039  0.070  -0.013  (0.122)  (0.122)  (0.115)  (0.128)  0.069  0.094  0.120  0.074  (0.139)  (0.141)  (0.129)  (0.159)  0.138**  0.131**  0.147**  0.135**  (0.055)  (0.049)  (0.052)  (0.051)  0.006***  0.006***  0.006***  0.005***  (0.002)  (0.002)  (0.002)  (0.002)  -0.011  -0.033  -0.045  -0.034  (0.081)  (0.076)  (0.069)  (0.077)  -0.018  -0.022  -0.005  0.033  (0.066) -0.038  (0.065) -0.057  (0.064) -0.058  (0.071) -0.053  (0.062)  (0.062)  (0.061)  (0.061)  0.051  0.025  0.042  0.058  (0.062)  (0.058)  (0.057)  (0.059)  0.010  0.006  0.008  0.010  (0.011)  (0.011) 283  (0.012) 283  0.23  0.26  N  283  (0.011) 283  R2  0.24  0.24  The dependent variable is the decision to us the improved stove as the main cooking device. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  As it can be observed in Table 2.9, the household’s previous experience in fertilizer usage or in the elaboration of processed products does not have a significant impact on the likelihood of stove usage. Also note that the coefficient sign for the dummy variable that patterns appears as non statistically significant. One problem in using average village-level past participation in communal activities is that we do not know the quality of this participation or the exact type of activities in which people tend to participate. Due to these factors, this variable may not be able to properly capture how strong the social links within the communities of study are.  49  takes the value of one if the household has secondary education or higher is positive in all columns but it is not statistically significant. Finally whether the household belongs or not to an environmental or women based organization does not seem to significantly influence the individual usage decision.  2.7. Stove Dismantling The results in Section 2.6 indicate that the village proportion of beneficiaries that use the improved stove without problems has a positive effect on the household’s usage decision through its interaction with the village-level trust index in local neighbours; which suggests that information diffusion is stronger in villages which are likely to have strong levels of bonding social capital. However, it can be argued that households in these villages are more likely to share or be affected by the same type of unobservable factors, and that the positive and significant effect the interaction term between successful adoption and the bonding social capital indicator has in the individual usage regression is just reflecting this situation. Up to some extent, the fact that the village proportion of users with problems also has a significant (negative) effect on the individual usage decision mainly through its interaction with the village bonding links indicator alleviates the previous concern; as users with problems were mainly affected by materials problems, which the evidence suggests were not systematically caused by usage, installation or maintenance, but by random and unobservable differences in improved stoves’ material quality (see Section 3 in Chapter 3 for more on this) In order to provide stronger empirical support for the informational role of social capital in the context of the adoption decision; this section analyzes only among beneficiary households that by the time of the 2004 monitoring visits reported not using the improved stove, how the village-level usage patterns and the bonding social capital indicator affect their decision to dismantle the new cooking device (which can be interpreted as a decision to abandon any attempt to use the new technology). If the village proportion of users without problems is relatively high and bonding social links are strong, non users may reasonably expect to be able to effectively use the new stove at some point in the future (as they are more likely to learn from others), which decreases the likelihood of  50  dismantling the new device. On the other hand, if the proportion of users with problems is high and this information expands through a strong social network; then, it is very likely that a non user will decide to dismantle the new technology. In the light of the findings in the previous sections, it is expected that a higher proportion of users without problems will have a negative impact on the individual decision to dismantle the improved stove through its interaction with the bonding social links indicator; while the opposite will be true for a higher proportion of users with problems.  Table 2.10: Village factors affecting the household’s likelihood of dismantling the improved stove I II III 0.0063 Village total proportion of users (0.0127) Village total proportion of users * Village-level trust in local -0.0096 neighbours (0.0079) 0.0158 Village proportion of users without problems (0.0101) Village proportion of users without problems * Village-level -0.0140** trust in local neighbours (0.0061) -0.0442** Village proportion of users with problems (0.0204) 0.0329** Village proportion of users with problems *Village-level trust in local neighbours Village-level trust in local neighbours (bonding social capital indicator) N Villages R2  (0.0130) 0.3854 (0.4575) 102 23 0.29  0.4461 (0.3259) 102 23 0.32  -0.4084* (0.1992) 102 23 0.32  The dependent variable is the decision to dismantle the improved stove. Only beneficiary non users that installed their improved stove are considered in the estimations. All columns in Table 2.10 control for the same household level controls as Tables 2.7 and 2.8, watershed location as well as for the proportion of village beneficiaries. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  As it was the case in Section 2.6, all the specifications in Table 2.10 have been estimated using a linear probability model with clustered standard errors at the village level55. The regressions in Table 2.10 control for the same household’s characteristics included in tables 2.7 and 2.856, the village proportion of beneficiaries and include watershed location dummies57. The specification corresponding to column I in Table 2.10 defines the village  55  The probit estimates are relatively similar to the OLS ones. All regressions in Table 2.10 also control for whether the non user reported or not a material problem. 57 As it was the case in Table 2.7, I initially estimated the regressions in Table 2.10 without considering an interaction term between social capital and the village usage proportions. Regardless of whether or not I 56  51  informational term as a function of the total proportion of users, the bonding social capital indicator (village-level trust in local neighbours) and the interaction term between these variables. As we can see, the effect of the total proportion of stove users on the uninstalling decision is not statistically significant; neither the total proportion of users’ linear term nor its interaction with the social capital indicator appears to influence the dismantling decision. These results should not surprise us; as discussed earlier, not all improved stove users influence the individual household’s decision in the same manner. Column II in the same table includes the village proportion of users without problems instead of the total proportion of stove users. As we can observe, the coefficient for the interaction term between the proportion of users without problems and the village bonding links indicator is negative and statistically significant at the 5% significance level. The result in column II evaluated at the sample means indicate that if the village proportion of users without problems increases by one percentage point, then the dismantling likelihood decreases by 0.4 percentage points; while if we evaluate this effect at the maximum level of village social capital, the dismantling likelihood will decrease by 1.5 percentage points. In order to confirm that only users without problems have a negative significant impact on the decision to dismantle the improved stove through its interaction with the village bonding links indicator, in column III I only include in the informational term the proportion of users with problems. As expected, the interaction term is in this case positive and statistically significant; that is, the proportion of users with problems is more likely to encourage the dismantling decision in villages with strong levels of bonding social capital. Evaluated at the sample means, the results in column III indicate that if the village proportion of users with problems increases by one percentage point, the dismantling likelihood will increase by 0.1 percentage points; while if we evaluate this effect at the maximum observed level of social capital, the dismantling likelihood will increase by 2.8 percentage points. Note that the coefficient for the social capital linear  considered a quadratic term for the village usage proportions, the coefficients for the village usage proportions and the village social capital term appeared as not statistically significant.  52  term in column III is negative and statistically significant at the 10% significance level. This result indicates that the informational impact of social capital on the decision to dismantle the new stove will be positive (encourage the dismantling decision) only if the proportion of users with problems is relatively high58 (above 12%). The results in this section confirm that the bonding social capital indicator has a significant multiplier effect on the impact village-level usage patterns have on households’ decisions: the village proportion of users without (with) problems is more likely to negatively (positively) influence the dismantling decision in villages with higher levels of bonding social capital. These findings strongly suggest that it is unlikely that the interaction term between the village proportion of successful users and the bonding social capital indicator in the usage regressions in Section 2.6 is just capturing the effect of some unobservable village factor, which happens to influence the usage decision with more intensity in those villages with stronger levels of trust in local neighbours (as households in strong bonding links communities are more likely to be affected by the same unobservable components).  2.8. Additional Identification Issues 2.8.1. Social Acceptability As noted earlier in this chapter, it can be argued that the higher impact the village proportion of successful users has on individual usage decisions in communities with higher level of trust in local neighbours (our measure for bonding social capital) just reflects the fact that in such circumstances the new cooking device appears as more socially acceptable or fashionable. However, fashion or social acceptability effects are likely to be more relevant in a context in which the decision of study is related to “getting the new stove” or “having it installed” during the distribution stages. In the specific case of this research, all households in the sample already have the new device, and this 58  The results in Table 2.10 also hold when we separately introduce as controls village altitude, road accessibility and trust in strangers. When these controls are introduced at the same time, statistically significance is lost for the interaction term between the social capital indicator and the village proportion of users without problems (probably due to the small sample size); however the coefficient´s size is very similar.  53  research focuses on the decision to “effectively” use it as the principal cooking stove. Although it may be impossible to completely discard the presence of some form of social fashion or social acceptability; for the case of the effective usage decision, such effects are less likely to play a major role in comparison to learning effects, which allow access to crucial information that may facilitate stove usage and desired performance. This section provides “indirect” evidence in order to alleviate the reader’s concerns that social effects other than social learning may be driving the observed results for the interaction term between the bonding links indicator and the village proportion of successful users.  Table 2.11: The effect of social fashion on individual usage decisions (indirect evidence) Village proportion of users without problems Village proportion of users without problems ^2 Village proportion of users without problems * Village-level trust in local neighbours Village proportion of beneficiaries  I 0.0181*** (0.0052) -0.0003*** (0.0001)  II 0.0141 (0.0156) -0.0003*** (0.0001)  III 0.0035 (0.0189) -0.0004*** (0.0001) 0.0105*** (0.0035)  0.0066*** (0.0018)  Village proportion of beneficiaries* Village proportion of users without problems  0.0050 (0.0080) 0.0001 (0.0002)  0.0071 (0.0059) 0.0000 (0.0002)  Village proportion of beneficiaries*Villagelevel trust in local neighbours Village-level trust in local neighbours (bonding social capital indicator) N Villages R2  IV 0.0176*** (0.0049) -0.0003*** (0.0001)  V 0.0002 (0.0061) -0.0003*** (0.0001) 0.0127*** (0.0019)  0.0084 (0.0083)  0.0209*** (0.0067)  -0.0014  -0.0097*  (0.0061)  (0.0054)  -0.0732  -0.0644  -0.4924**  0.0623  0.3208  (0.1017) 283 24 0.22  (0.1130) 283 24 0.23  (0.2032) 283 24 0.24  (0.5621) 283 24 0.22  (0.5616) 283 24 0.25  The dependent variable is the decision to use the improved stove as the main cooking device. All columns in Table 2.11 control for the same household level controls as Table 2.8, watershed location, altitude, road access and trust in strangers. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  In first place, and given that the stove was freely provided to anyone who requested it; in villages where fashion or social acceptability effects tend to be stronger, one should expect to observe a higher proportion of beneficiary households during the distribution stages. Then, we can use the village proportion of improved stove beneficiaries as an indicator for the degree of village fashion or acceptability tendencies in the context of this intervention. If these effects are indeed driving the results in the usage regressions, the marginal effect of the village proportion of successful users (users without problems)  54  should be higher in villages in which our proposed indicator of fashion and social acceptability tendencies is higher. In order to test for this I introduce an interaction term between the proportion users without problems and the village proportion of beneficiaries. As we can observe in column II in Table 2.11, the coefficient for this interaction term is very close to zero and not statistically significant. In column III I include at the same time the village bonding links and the village “social fashion” indicators, as well as their interaction terms with the village proportion of users without problems; as it was the case in the previous sections, only the coefficient for the bonding social capital interaction term appears positive and significant, which suggests that pure fashion or social acceptance effects are not likely to be behind the observed data patterns. It can also be the case that fashion or social acceptability effects are “per se” increasing in village bonding social capital, and that the interaction term between the village proportion of users without problems and the bonding links indicator is just capturing this pattern. If the degree of fashion or social acceptability in the context of the improved stove intervention can be approximated by the village proportion of beneficiaries; then the effect of this proportion in the usage regression should be higher in villages with stronger bonding social capital. Column IV in Table 2.11 includes an interaction term between the village proportion of beneficiaries and the village bonding links indicator; as we can observe, the coefficient for this term appears as not statistically significant. Column V includes at the same time the interaction term between the village proportion of beneficiaries and the bonding links indicator, as well as the interaction term between the proportion of users without problems and the bonding links indicator; note that only the last interaction term appears to positively and significantly impact usage decisions. As we can also observe, the coefficient for the interaction term between the village proportion of beneficiaries and the village bonding links indicator is negative and statistically significant at the 10% confidence level. This result is just telling us that for a given proportion of beneficiaries, social capital is more likely to positively influence usage decisions in villages with higher success levels in technology usage, which is just the same conclusion obtained in the previous sections of this chapter.  55  2.8.2. Unobservable Correlates Serious identification problems will arise in this study if villages with a higher proportion of successful stove users were more likely to be visited by NGO members during the monitoring process, and also if within villages, households using the improved stove were more likely to be interviewed (or probably those not using the improved stove may have systematically refused to be interviewed). As it was mentioned in Section 2.3, visits to all the beneficiary villages were planned and the order of visits was not done as a function of the expected number of working stoves. In fact, special emphasis was set in visiting villages in high altitude less accessible areas; where, if anything, low usage rates were expected. Note also that on average 82% of the beneficiaries per village were visited, and that from information provided by MIRHASPERU and Universidad de Piura it is known that the improved stove monitoring team did not report any situation in which beneficiary households refused to be interviewed 59 . In addition, it is important to emphasize that visits were not planned in terms of the expected quality of the village bonding links, as the results of the social capital survey were not available by the time of these visits. Another important identification issue that is common to social capital studies is the problem of reverse causality. For example, in papers that try to relate village economic performance to social capital, it is not only the case that social capital affects economic performance; but also that good economic performance may allow building better social capital (Narayan et al (1998)). In the context of technology adoption in rural communities, households that decide to adopt a new technology may also decide to invest more in their social relationships, affecting in this way the equilibrium level of bonding links. However, the estimated regressions in this chapter are not likely to suffer from this problem, as the household survey measuring social capital was implemented before the improved stove adoption process and almost one year before the monitoring survey. In other words, the measures of bonding social capital used in this chapter are not likely to have been influenced by the program intervention or by the nature of the adoption process. 59  During the summer of 2008 I carried a survey in the area in order to evaluate the current situation of improved stove usage, from the all the households visited, only one household refused to be interviewed.  56  In studies on social interactions, it is very important to take into account the potential presence of the reflection problem (Manski (1993)). As it was originally defined, the reflection problem is an issue of collinearity; in Manski’s seminal paper, endogenous effects are not identified because they are a linear combination of exogenous and correlated effects. However, as Brock and Durlauf (2000) clearly explain, non identification in social interaction models is intrinsically linked to linearity; for the case of nonlinear in means social effects -as it is the case of the present research essay- and assuming correct model specification, social effects are normally identified60. The presence of village-level correlated unobservables is probably the most important issue in terms of identifying the role social capital has played in facilitating social learning during the adoption process of improved stoves. Households’ stove usage decisions within the village may be correlated not because a social learning process is present but just because they share common unobservable characteristics or are subject to the same shocks and environments, especially in villages with strong bonding links. Although it is not possible to address all the possible alternative hypothesis, in my opinion this chapter provided solid evidence in order to support the information diffusion hypothesis as the generating process behind the strong observed correlation between household decisions, village usage patterns and village social capital. In first place it was shown that not all types of technology usage patterns influence individual decisions in the same manner through its interaction with the bonding social capital indicator. While the interaction term between the proportion of improved stove users without problems and the bonding social capital indicator has a positive and statistically significant impact on the household’s usage decision, the interaction term between the village proportion of users with problems and the bonding social capital measure has a negative one. Interestingly, when the total proportion of users was included in the village informational term, no multiplier role for the social capital indicator was found; the multiplier effect appeared to be statistically significant only when we distinguished between user types. 60  For more detail see appendix B or refer to: Durlauf, Steven and Brock, William (2001b). “Interaction Based Models”. Handbook of Econometrics 5, James H. Heckman and Edward Leamer, eds. pp. 44 to 45.  57  Another result that supports the social learning hypothesis as the driving process in the improved stove usage data, is that the marginal effect of bonding social capital on the individual usage decision was shown to be closely linked to village-level stove performance: bonding social capital only appears to have a positive impact on the usage likelihood if the initial village-level success (failure) rates are relatively high (low). That is, if initial usage success (failure) levels in the village are relatively low (high), the information that the new technology is not a good one will be intensively diffused through the communal network, and this will negatively affect the household’s decision to use the new stove as the main cooking device. This finding strongly suggests that information diffusion has been the main role played by social capital in the context of this study. The results obtained in the previous sections are robust to different indicators of bonding social capital and remain significant after key geographical factors influencing stove performance and NGO’s effort during the diffusion process, such as altitude or road access, were controlled for. As a final step to convince the reader that the results in this chapter are not driven by unobservable village factors which are more likely to be shared by households in villages with stronger bonding links; in Section 2.7 it was shown that the social capital indicator has also a multiplier effect on the decision to dismantle the improved stove among beneficiary non users. The results in this section indicate that the interaction term between the bonding social capital indicator and the proportion of users without problems has a negative effect on the decision to dismantle the new stove; while the reverse is true for the proportion of users with problems, even after controlling for household own failure. In my opinion, no other village process or alternative story is likely to generate the patterns that have been consistently identified in the improved stove usage data61.  61  I have also estimated the baseline regressions including the proportion of village households participating in environmental organizations and women organizations, no change in the main results was observed. The main regressions were also estimated including a dummy variable taking the value of one if the household “self reported” most influential village member was using the stove without any problem, the coefficient for this variable was not significant in the regressions and the main results remained totally unaffected.  58  2.9. Bonding versus Bridging Social Capital In his seminal work on social capital, Woolcock (1998) suggests that this social variable has multiple and dynamic dimensions (e.g. bonding vs. bridging links), and that these dimensions have different and very specific roles in the communal social life. In the previous sections, it was shown how bonding social capital influences within village information diffusion by expanding the effect the within village usage patterns have on the household’s likelihood of using the improved stove as the main cooking device. As an indicator of village bonding social capital this research used the village-level of trust in local neighbours; however, very similar results are observed when other indicators potentially related to the strength of communal bonding links are used in the main regressions 62 . This section presents empirical evidence suggesting that only bonding (within village) social capital significantly influences the effect within village usage patterns have on individual decisions; while bridging (across villages) social capital does not play any significant role in this situation. Bridging social capital only seems to influence the effect usage patterns outside the household’s village of residence have on the household’s usage decision. In this section, the village-level trust in people from other communities is used as an indicator of bridging social capital, and the proportion of improved stove users without problems in the closest neighbour village is employed as an indicator of successful usage levels outside the household’s village of residence. The closest neighbour village is defined as the one which geographical center is situated at the minimum linear distance from the geographical center of the household’s village of residence63. The estimation results in column I in Table 2.12 (which estimates the same regression as in column II in Table 2.8.B) indicate that the impact of the within-village proportion of users without problems on the household’s usage likelihood is increasing in the village bonding social capital indicator, and that the bonding social capital indicator may have a “negative” effect on the individual usage decisions if the village proportion of users without 62  For example, the village-level trust in local organizations. Given the area’s geography, it may be argued that using linear distances may not accurately identify the “closest neighbour village”; however, in most of the cases my defined “closest village” coincides with the opinion of people that have worked in the area of Chalaco District during the Chalaco Program and have then a better knowledge of the terrain and approximate walking distances across villages. 63  59  problems is relatively low. In column II a very similar regression is estimated, the only difference with respect to column I is that trust in people from other villages is used as a measure of social capital instead of trust in local neighbours. As we can see, the coefficient for the interaction term between the level of trust in people from other villages and the proportion of users without problems within the village is very close to zero and not statistically significant. Column III in Table 2.12 explores how the proportion of improved stove users without problems in the closest neighbour village and the village bridging social capital indicator mutually influence the individual usage decision. Column III does not allow for an interaction term between these variables. As we can see, neither the proportion of successful users in the closest neighbour village, nor the village bridging links indicator appear to significantly influence the household’s usage likelihood. When an interaction term between these variables is included in column IV, we can observe that the results are significantly affected. In first place, the effect that the proportion of users without problems in the closest neighbour village has on the individual usage decision appears to be increasing in the bridging social capital indicator64; in second place we can note that if successful usage levels in the closest neighbour village are relatively low, the marginal effect of the bridging social capital indicator may be negative. In column V I estimate a very similar regression to the one estimated in column IV, but instead of using an indicator of bridging social capital, the standard indicator for bonding social capital is used in the regression. Note that only the interaction term appears to be statistically significant in this case, but its coefficient is much lower than the one estimated in column IV, in which trust in people from other villages was used as an indicator of bridging social capital65. This result suggests that only the bridging social capital indicator (trust in people from other villages) influences the effect usage levels outside the village have on individual usage decisions. 64  The linear term coefficient for successful usage in the closest village in column IV appears negative; but when we add to this the interaction term, the total effect is positive at the mean value of bridging links. 65 Given the relatively high correlation between usage levels across villages in the same watershed, the specification in column V seems to be just mimicking the specification in column I. Note the very similar sizes of the coefficients for the interaction and the linear social capital terms in column V as compared with the ones obtained in column I.  60  Table 2.12: The effect of bonding vs. bridging social capital on the household’s likelihood of using the improved stove as the main cooking device Village proportion of users without problems Village proportion of users without problems^2 Village proportion of users without problems* Village-level trust in local neighbours Village proportion of users without problems* Village-level trust in people from other villages  I 0.0062 (0.0081) -0.0004*** (0.0001) 0.0102*** (0.0034)  II 0.0221* (0.0112) -0.0004*** (0.0001)  III  IV  -0.0005 0.0116 -0.0135 (0.0112) 0.0001 (0.0001)  Closest village proportion of users without problems^2  -0.0429*** (0.0097) -0.0000 (0.0001)  Closest village proportion of users without problems*Village-level trust in local neighbours  -0.0082 (0.0111) -0.0001 (0.0000) 0.0092**  0.0295 (0.0301) 0.0002 (0.0004)  (0.0044)  Closest village proportion of users without problems*Village-level trust in people from other villages  Village-level trust in people from other villages (bridging social capital indicator) N Villages R2  VI 0.0221 (0.0186) -0.0009** (0.0003) 0.0148* (0.0779)  Closest village proportion of users without problems  Village-level trust in local neighbours (bonding social capital indicator)  V  -0.0200  0.0287*** -0.4079 (0.2788)  (0.0276) -0.6632 (0.4197)  276 22 0.21  0.0552 (0.7681) 276 22 0.28  (0.0058) -0.4882** (0.1997)  283 24 0.24  -0.5896 (0.5653) 283 24 0.24  -0.2532 (0.1671) 276 22 0.20  -1.0763*** (0.1994) 276 22 0.20  The dependent variable is the decision to use the improved stove as the main cooking device. All columns in Table 2.12 were estimated by OLS with clustered standard errors at the village level, and control for the same household level controls as Table 2.8, watershed location, altitude, road access and trust in strangers. When the proportion of successful users within the village is included in the regression, the proportion of beneficiaries within the village is also included. When the proportion of successful users in the closest village is included in the regression, the proportion of beneficiaries in the closest village is also included. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  61  Finally, in column VI in Table 2.12 I include both proportions of users without problems: within and outside the village. Given the results in the previous columns, the within village proportion is interacted with the bonding social capital indicator; while the outside village proportion is interacted with the bridging social capital measure. The results show that only the interaction term between the within village proportion of users without problems and the bonding social capital indicator has a significant effect on the individual usage decision (its coefficient size is relatively close to the one estimated in column I). This evidence confirms that the within village usage patterns as well as the village bonding social capital had a prevailing role during the diffusion of information related to the new cooking technology in the area of study. Up to my current knowledge, this research is one of the first ones presenting clear empirical evidence on the effective presence of different dimensions of social capital at the village level, as well as on the very specific roles bonding and bridging social capital play in the context of technology usage decisions.  2.10. Conclusion This chapter empirically explored how the household’s decision to use a new firewood stove as the main cooking device is mutually influenced by the strength of the villagelevel trust in local neighbours (an indicator for bonding social capital) and by the villagelevel technology usage patterns at initial adoption stages. The main results suggests that bonding social capital played a crucial role facilitating social learning during the adoption process of improved stoves in the rural communities of the Chalaco District. More precisely, it was empirically demonstrated that the impact of village-level usage patterns on the usage likelihood was higher in villages in which the bonding social capital indicator was also higher. In addition, it was also shown that only the proportion of village users without problems has a positive effect on the household’s usage likelihood through its interaction with the village bonding social capital indicator; while the reverse is true for the proportion of users with problems. When we considered the total proportion of users as an indicator of the within village usage patterns, its social capital effect (captured by the interaction between this proportion and the village social capital indicator) appeared as not statistically significant. Our results also confirm that the  62  marginal impact of the village bonding social capital indicator on the individual usage decision is closely linked to the village success rates in technology usage: if the proportion of users without problems is relatively low or the proportion of users with problems is relatively high, the marginal impact of the bonding social capital indicator is more likely to be negative. In other words, if a new technology does not perform well on average (or does not perform as expected), the village network is more likely to intensively diffuse “negative” information about the new device. In order to argue that correlated unobservables are not likely to drive the main findings, this research analyzed how village usage patterns and the bonding social capital indicator influence the decision to dismantle the improved stove among beneficiary non users. The results pointed in the right direction: village usage levels without problems have a negative impact on the likelihood to dismantle the new stove mainly through its interaction with the bonding social capital indicator, while village usage levels with problems appear to encourage this decision. These results constitute strong evidence towards the social capital information diffusion hypothesis in the context of the improved stove usage program in the Chalaco District. Moreover, as the social capital indicators were obtained before the intervention, the reverse causality problem between usage decisions and village social capital is not likely to be present in our estimations. The empirical results in this study also show that only the bonding links indicator influences the effect within village-level usage patterns have on individual decisions; while the bridging social capital indicator only influences the effect usage patterns outside the village have on individual usage decisions. This chapter also provides indirect empirical evidence to rule out the possibility that fashion or social acceptability effects are driving the results for the higher impact usage levels without problems have on villages with stronger bonding links. Using the proportion of beneficiaries as an indicator of fashion or social acceptability in the context of this intervention, it was shown that these social effects do not play any significant role at influencing the individual stove usage decisions; and the baseline results were not affected when I allowed for different specifications in which fashion or social acceptability effects were controlled for.  63  The main findings in this study suggest that the nature and extent of communal social relationships play an important role at influencing the size of peer effects during technology adoption processes in rural communities. This result has important policy implications for development programs related to the dissemination of new technologies in rural areas of developing countries. The first implication that can be derived from this study is that technology diffusion programs that intend to rely on informational spillovers during early adoption stages must first obtain a clear understanding of the nature of communal social relationships. We should not expect to observe a strong social learning process if the bonding links are extremely weak. Having said this, it is important to note that “creating” social capital is not an easy task; neither is to decide the optimal levels of bonding and bridging links that are appropriate to promote economic development. The second relevant implication in this chapter is that in the context of development programs introducing new technologies in rural communities, it is crucial to properly and constantly monitor the adoption process, especially at initial adoption stages. A low initial level of success in technology usage, or a poor initial performance of the technology, may lead to its complete rejection and abandonment, especially if the bonding social links (or the within village level of trust) are relatively strong. In such scenarios the social network is likely to intensively disseminate “negative” information about the new technology. Even a small failure rate can have disastrous effects in terms of the adoption process; our results indicate that rural households tend to react to bad news more drastically than they react to good news about a new technology. In our estimations the coefficient for the interaction term between the proportion of users with problems and the bonding links indicator is approximately 2.5 times higher than the coefficient for the interaction term between the village proportion of users without problems and the bonding social capital measure. A final implication that can be derived from this chapter is that when empirically studying the impact of social capital on individual decisions in rural communities, it is of critical importance to clearly define which social capital dimensions are expected to play a crucial role in the context of study, and the specific function they are likely to perform.  64  3. Improved Stove Adoption and Firewood Consumption 3.1. Introduction “Forest resources have the potential to make a major contribution to development by meeting basic needs in energy as well as other forest products, by contributing to food security, by sustaining industries which provide employment and income, and by maintaining environmental stability. But if this potential is to be fully realised, uncontrolled exploitation must be replaced by appropriate management of the entire forest production chain, from the establishment through to the maintenance and harvesting of forest crops to processing, marketing and fuel use in the home and elsewhere. The domestic stove, as a key element of the end-use of forest products, plays an important part in this process.” Boy, Erick et al. 2000  In the last three decades, the distribution of improved “more efficient” firewood cookstoves has been one of the most popular strategies implemented in rural areas of developing countries (such as China, India and Mexico) to alleviate forest degradation 66. At first glance, the logic behind the massive distribution of these cooking artefacts appears quite appealing. Since improved stoves are more efficient at generating cooking energy -as usually supported by laboratory and controlled “in situ” cooking tests- they will unequivocally and significantly reduce the firewood extraction needs of rural households in developing countries, who are daily faced with increasing forest resources scarcity. However, this is not totally clear. In first place, laboratory settings and “in situ” standardized cooking tests supervised by program technicians are not likely to fully capture the complexity of daily cooking tasks (Johnson et al. 2010); then, efficiency gains under real conditions are very likely to differ from those observed in such controlled trials67. In second place, it is commonly assumed that cooking behaviour will remain unaffected after the introduction of the new stove; however, now that you have a more efficient cooking device, you may optimally decide to switch your consumption towards goods that require more cooking energy; or you can increase the consumptions levels of 66  It must be understood however that deforestation and forest degradation are complex processes, and that firewood extraction for cooking and heating purposes is just one of the many factors affecting them. 67 In addition to this, different qualitative studies (e.g. Gill, 1985) suggest that rural households not only care about efficiency or firewood savings -speeding cooking for example can be more important to themand that they tend to modify and adapt the new technology to their specific needs, affecting in this way the expected performance of the device.  65  the cooked (heated) goods you are currently consuming 68 . In this sense, whether a particular improved stove design reduces firewood consumption in a given context is an empirical question which must be addressed using household level data corresponding, as closely as possible, to real patterns of firewood consumption and stove usage conditions. The formal empirical evidence related to the impact of improved stove usage on firewood consumption is relatively scarce and surprisingly inconclusive69. Most of the evaluation studies in the literature rely on standardized cooking tests (such as the Water Boiling Test – WBT or the Controlled Cooking Test - CCT) generally performed inside the beneficiaries households’ units and closely supervised by program technicians (McCracken et al. 1998, Boy et al. 2002, Barrueta et al. 2008). However, as mentioned before, there is consistent evidence indicating that standardized cooking tests are not representative of stove performance during normal cooking activities, and that no clear conclusions can be obtained from such type of evaluation designs (Johnson et al 2010). In order to estimate the effect of improved firewood stoves under more real usage conditions, a few studies in the literature (Wallmo et al 1998, Boy et al 2002, Barrueta et al 2008) have also performed Kitchen Performance Tests (KPTs), in which daily firewood consumption during normal cooking tasks is closely monitored and measured by the evaluation team. Nevertheless, these studies presents several problems; for example, Wallmo et al (1998) and Barrueta et al (2008) fail to take into account the endogenous nature of the stove usage decision; while Boy et al (2002) do not take into consideration that the current cooking habits of improved stove users are probably quite different than the ones they had before adopting the new technology. Improved stove possession has also been included as a control variable in a few empirical studies in which the main focus is some other factor related to rural household’s firewood extraction and consumption patterns (Amacher et al 1996, Heltberg et al 2000 and Chen et al 2006). However, these studies control for stove ownership instead of effective usage; furthermore, the stove variable is generally included as an exogenous regressor, and little 68  If the stove is also used as a heating device, then even if the new technology reduces firewood consumption for cooking purposes, you may decide to keep the stove fired for a longer time period to warm your house in the winter months, increasing in this way your firewood consumption. It can also be the case that the stove design is more efficient at cooking tasks but less efficient at heating the household unit. 69 Section 2 in this chapter provides a more detailed discussion on the related literature.  66  attention has been given to self-selection issues related to improved stove adoption and usage decisions. Given the limited empirical evidence on the effect of improved stoves usage on firewood extraction and consumption, as well as the identification flaws of the few existing studies, this research contributes to the literature by empirically exploring how “effective” usage of an improved stove design, freely distributed in the year 2003 in the villages within the Chalaco District (Northern Peruvian Andes), influences household’s firewood consumption. More precisely, in this chapter I estimate the improved stove treatment effect on the treated; that is, the effect of stove usage on those households that selfselected as adopters of the new cooking device during the 2003 improved stove distribution program. This study is based on self-reported firewood consumption data, collected in 21 villages in the summer of 2008, and exploits evidence on random and exante unobservable differences in improved stoves’ material quality to identify the effect of improved stove usage. While it is true that this research’s results cannot be generalized to all contexts, the villages here analyzed are relatively similar to the typical village in this area of the Northern Peruvian Andes, where deforestation and forest degradation have been linked to the increasing negative impacts of intense precipitations during the winter season, such as land erosion, declining soil fertility and flooding. Probably the most challenging problem in terms of identifying the causal effect of improved stove usage is related to the fact that the stove usage decision is likely to be correlated with unobservable household´s and village factors which are also correlated with firewood consumption (such as household’s ability, forest preferences, women´s empowerment, etc). In the context of this research, the special circumstances of the 2003 Improved Stove Distribution Program in the Chalaco District provide a plausible identification strategy which may allow us to overcome this econometric issue. Reports from the monitoring visits carried out between April and August 2004 (approximately eight to twelve months after improved stove distribution in the Chalaco District70) clearly indicate that a proportion of households that decided to adopt the new technology as their 70  Internal Report on Stove Performance, MIRHASPERU and Universidad de Piura, August 2004  67  main cooking device presented chimney and iron frame material problems; and that in the particular case of iron frame material failures (deformations and cracks), these were not likely to have been systematically caused by deficiencies in stove installation, improper usage or maintenance, but by poor materials quality. By the time of the 2004 monitoring visits close to 50% of the beneficiaries with material problems stopped using their new stoves. Members of the monitoring team at Universidad de Piura, who I interviewed in 2008, confirmed that deficiencies in materials quality were associated with iron frame failures and the abandonment decision of the new cooking technology. The evidence also suggests that deficiencies in the stove iron frame were random and exante unobservable to the beneficiaries. In fact, households which experienced iron frame material problems appear to be ex-ante very similar (in terms of their observable characteristics) to households that did not present iron frame failures. Also, there is no evidence on households or village leaders trying to influence the responsible NGO in order to get stoves of better iron frame quality; and complains about materials issues did not appear during the distribution stages, but after the beneficiaries made effective use of their new stoves (Ureta (2007)). Furthermore, the stoves were produced in small local workshops in the main coastal city, and there is no evidence on materials quality inspection prior to its distribution. As I mentioned before, this evidence provides a suitable identification strategy, as experiencing a material problem appears to be exogenous to households’ and villages characteristics; that is, you were just randomly allocated an iron frame of lower material quality. Then, in this study I use self-reporting an iron frame failure as an indicator for having received an iron frame of lower material quality, and employ this indicator as an instrumental variable to identify the effect of improved stove usage. While it is true that the self-reported nature of my instrument can raise some identification concerns, I asked approximately half of the households that reported an iron frame failure if I could take a look at their iron frames and in almost all the cases and I was able to visually confirm the presence of material failures. Nevertheless, identification issues do not stop here; the reader may point out that iron frame’s deformations could have been influenced by geographic factors (e.g. a stove of bad quality was more likely to deform at higher altitudes than at lower ones, even if  68  operated by similar individuals) or by household’s stove usage patterns (particularly in the long term). Although this is not likely to be the case -if properly built an iron frame was expected to work without problems for at least 10 years- I deal with these issues by controlling for several household factors and by estimating a village fixed effects regression; as we will see in the estimation section, the main results in this chapter are robust to different model specifications. Estimating the effect of improved stove usage on firewood extraction and consumption, and linking this effect to forest resources degradation has several practical difficulties. In first place, while in some seasons rural households collect the small pieces of wood available from fallen tree branches, which are relatively abundant, in others more “serious” degrees of firewood extraction take place. In the context of the Chalaco District it is known that the most severe levels of firewood extraction are observed in the last two months of the year (November and December), when households collect the firewood they will use in the winter (rainy) season. In this case, firewood extraction is more connected to cutting down trees than to collecting the small pieces of firewood available from fallen tree branches. Households must collect all the firewood they will consume during the winter in the immediate previous months, since the weather conditions make it extremely difficult to collect firewood on a frequent basis during this season. Due to the particular nature of this collection process and to the fact that they have a specific storage space for the firewood collected, households in the area have also a more accurate idea of the total amount collected in the rainy season than in the dry season, when collection takes place on a more frequent basis (from weekly to daily). In addition, the measurement units used to report firewood collection for the winter season are relatively uniform, and most households are able to report their collection in “cargas” (approximately 30 Kg of firewood). Finally, while in the dry season other biomass fuels such as crop residues are also employed; in the winter season firewood is generally the only source of energy available71. Given this context, this research only focuses on firewood consumption in the winter season; the instrumental variable results in this paper suggests that usage of an  71  Less than 1.5% of the households report selling firewood and only 3% of the households in the district report using a gas or kerosene stove for cooking and/or heating.  69  improved stove reduces monthly firewood consumption by approximately 40% during this time of the year. This chapter develops as follows: Section 3.2 discusses in more detail the existing studies in the literature, Section 3.3 briefly comments on the 2003 Improved Stove Program and discusses the identification strategy; Section 3.4 describes the data collected by myself in the summer of 2008 in the Chalaco District; Section 3.5 presents the empirical results and finally Section 3.6 concludes.  3.2. Related Literature During the initial years of improved stoves distribution (late 70’s and early 80’s), the empirical evidence on the efficiency gains related to the effective usage of these artefacts was mainly anecdotal or restricted to laboratory tests, where it was shown that some stove designs were more efficient than the traditional “open fire” stove (Lou Ma, 1981; Gill, 1985). During the 80’s and early 90’s a variety of standardized tests were developed to evaluate the different improved cookstoves programs around the world, being the Water Boiling Test72 (WBT) the one which most attention received during these years. One of the first formal studies using such standardized methods “in situ” was performed by McCracken et al (1998) in rural villages in Guatemala. In order to evaluate the performance of an improved stove design know as the “plancha” (a design very similar to the one distributed in the Chalaco District), the authors implemented the high power version of the WBT as well as the Controlled Cooking Test (CCT, which basically involves the task of cooking a given quantity of beans or some other local food); their results indicated that the “plancha” was not more efficient than the traditional open fire stove used in the area. Subsequent studies employing similar testing procedures have 72  The standard WBT developed by Baldwin (1986) has 3 components: two high power tests, one conducted at cold starting conditions and the other at warm starting conditions, and a low power test designed to simulate slow cooking tasks (tasks that require low heat). The high power cold start test begins with the stove at room temperature and uses a pre-weighted bundle of wood to boil 3 L. of water in a standard pot. In the high power warm start test, a fire is reset immediately after the WBT cold start phase and the test repeated, with the main intention of identifying differences in performance between a stove when it is cold and when it is warm. In the low-power simmering phase test, a fire is reset using a preweighted bundle of wood after the high power tests and used to simmer water 3CO degrees below boiling temperature for 45 min. These tests assess a) stove thermal efficiency: defined as the ratio of work done by heating and evaporating water to the energy consumed by burning wood; b) stove firepower: the ratio of the wood energy consumed by the stove per unit of time during each phase of the tests; and c) the stove specific fuel consumption: defined as the ratio of the amount of fuelwood consumed to the amount of water remaining at the end of the trial (this last one should be considered as the fuelwood required to produce a unit of output) (Barrueta et al. 2008)  70  been performed by Boy et al (2002) also in rural Guatemala (to tests the same “plancha” design) and more recently by Barrueta et al (2008) in rural Mexico (to tests the “Patsari” stove, an enhanced design of the “plancha” concept) with more positive results73 in terms of efficiency gains. However, even if significant firewood savings are observed in these controlled settings, its conclusions cannot be extended to predict the performance of the stove under real usage conditions. As it is clearly documented by Johnson et al (2010), there is consistent evidence indicating that standardized cooking tests are not representative of stove performance during daily cooking activities74. In addition to the standardized tests above mentioned, Kitchen Performance Tests (KPTs), in which firewood consumption during normal daily cooking activities is closely monitored and measured by program technicians for a relatively short period of time, have been implemented to evaluate the performance of improved stoves under more real usage circumstances. Wallmo et al (1998) for example monitored during three days the firewood consumption of improved stove users (of the “Lorena” stove design, which is relatively similar to the “plancha”) and users of a traditional open fire stove in villages in Uganda. In this study, the authors could not find significant differences between the two groups in terms of firewood consumption. However, this study suffers from obvious identification issues, as the authors fail to take into account that the usage decision of an improved stove is clearly endogenous in the firewood consumption equation. In a similar evaluation study, Boy et al (2000) requested a group of improved stove users in rural Guatemala to cook with their improved stoves (after some modifications were made by program technicians) for a week, and then to cook with a traditional open fire stove for another week. These households were closely monitored by program technicians, firewood was freely provided to them and they were explicitly asked not to modify their cooking habits. Their findings indicate that improved stove usage significantly reduces firewood consumption by approximately 60%. Nevertheless, it is important to take into 73  In laboratory tests, the Patsari stove performed better at low power tests (e.g. cooking beans) than the traditional one, but poorer at high power tests (e.g. boiling water from cold starting temperatures). 74 “The nonrepresentative carbon emissions and efficiency estimates should not be surprising given that controlled burn cycles for specific tasks cannot encompass the variety of daily stove use activities, with up to 90% of stove tasks in some regions not involving boiling water […] In addition, since efficiency varies significantly as a function of power output during the different phases of the burn cycle, a single efficiency is not a good performance indicator.” (Johnson et al 2010)  71  account that household’s cooking habits are likely to be influenced by the availability of the more efficient technology, especially in the medium and long term, which makes it complicate to obtain clear implications from this evaluation setting (also household’s “normal” behaviour and incentives are likely to differ from what is observed during supervised trials). In addition to this, once a household adapts to a new technology and uses it for a long period of time, switching to another one (even the traditional stove you used before) may have some adjustment costs. It follows that for the conclusions of this study to be correct –in terms of efficiency gains-, households that have been using the improved stove for a long time period should be able to use the traditional one as efficiently as they did before adopting the improved cooking device. In a very recent paper, Barrueta et al (2008) implemented a KPT to estimate the changes in firewood consumption for a group of households in rural communities in Mexico which were randomly assigned an improved stove. Their results show that the improved stove reduces firewood consumption by approximately 67% after one year of usage. Unfortunately, in this study no control group was included to account for time variant factors potentially affecting consumption levels, and then no clear conclusions can be obtained. There are a few empirical studies which focus on some other factors related to firewood extraction and consumption patterns by rural households, which also include improved stove possession as a covariate in their main specifications (Amacher et al 1996, Heltberg et al 2000 and Chen et al 2006). In these papers, which are mainly based on self-reported firewood consumption data, the improved stove variable appears to decrease firewood consumption (the sign for the dummy variable indicating its possession appears negative), but its impact is generally not statistically significant. One of the main problems with the stove variable in these studies is that stove ownership does not necessarily implies “effective” usage for cooking or heating; furthermore, the stove variable is included as an exogenous regressor, and then it is not possible to separate the effect of stove usage from the effect of unobservable factors potentially correlated with firewood consumption and with the stove usage decision. All things considered, in my opinion there is not conclusive evidence on the causal effect of improved stove usage, particularly in the long term, and there is plenty of room for empirical research based on firewood consumption  72  data corresponding to improved stove “real” usage conditions. However, unless a suitable instrumental variable or a natural experiment is available, studies based on observational data may not be able to fully identify the causal impact of effective improved stove usage.  3.3. Identification Strategy In the fall of 2003, improved firewood cookstoves were freely distributed and installed in 37 of the 39 villages within the 5 watersheds in the Chalaco District, in the Northern Peruvian Andes75. Stove distribution was on charge of the local NGO MIRHASPERU and was funded by the Spanish International Cooperation Agency. During the distribution process the responsible NGO contacted the most representative watershed and village organizations, and with their support called to open meetings where the expected benefits of the new technology were explained. Approximately 85% of all households76 in the district received an improved stove (Ureta (2007), Urday (2006)), and close to 96% of those households who requested an improved stove were allocated one77. It is important to mention that beneficiary households were not required to immediately abandon their traditional cooking technology in order to receive the new stove and have it installed78. A second stage of stove distribution was originally planned, but had to be cancelled due to administrative and budgetary reasons. The same stove design was provided in all the villages, as well as the same installation, usage and maintenance instructions. As it can be observed in figure 1 in Appendix A, the stove design has an iron frame placed on a combustion box made from locally provided mud bricks and a metallic chimney designed to reduce exposure to IAP. The responsible NGO trained two local craftsmen in every village, selected by the local beneficiaries in a public meeting, to install the new stove. From April to August 2004, the responsible NGO under the supervision of Universidad  75  These villages are located in five watersheds at altitudes between 1000 and 3500 m. The minimum temperature at high altitude areas ranges around 2 to 5 Co during the winter season (Mid December to Mid May approximately). 76 The number of households in Chalaco District as estimated during the year 2003 was around 2000 units. 77 Information obtained from the 2008 survey carried out in 21 villages in the Chalaco District. 78 Although I do not exactly know the number of households that installed the new stove without uninstalling their open fire stove; the 2003 stove program guidelines suggest that these probably were the majority of beneficiaries, as the program intended to allow for a gradual transition from the old technology to the more efficient improved one.  73  de Piura monitored the performance of the new cooking technology in 26 villages79. After monitoring visits additional training in all the beneficiary villages was provided on stove usage and maintenance. Interestingly, during the monitoring visits a proportion of improved stoves beneficiaries reported material problems with the stove iron frame and/or metallic chimney80. During these visits program technicians had full access to the household’s kitchen area and were able to visually confirm the presence of such problems. By the time of the visits, close to 50% of these beneficiary households stopped making use of their improved stoves, and those affected by this problem claimed that stoves of poor quality were provided to them (Ureta (2007)). The evidence from the monitoring visits suggests that in the particular case of iron frame’s material problems, these were not likely to have been systematically caused by deficiencies in stove installation, improper usage or inadequate maintenance81, and it has been confirmed by members of Universidad de Piura involved in the monitoring visits, who I interviewed in 2008, that the deficiencies (mainly cracks and deformations) observed in the iron frames during the monitoring visits were likely to have been caused by poor material quality82. Furthermore, the iron frames were produced in small workshops in the main coastal city, and there is no evidence on materials inspection prior to its distribution. All this evidence suggests that households experiencing iron frame problems were just given an iron frame of lower quality and that this deficiency affected the adoption decision of the new stoves. 79  Initially visits to all beneficiary villages were planned, but due to budgetary, administrative and security reasons 11 villages were finally not visited. However, it is important to emphasize that villages in which high rates of stove usage were expected were not particularly targeted. Moreover, at the beginning of the monitoring process special emphasis was set in visiting villages in which low rates of improved stove usage were expected. 80 In this chapter I focus on iron frame failures as the evidence suggests that it is very unlikely that deformations or cracks in this big piece of iron could have been caused by improper usage, maintenance or installation. In any case, when I also include those households that only presented chimney problems in the “materials failure” group, the results for the effect of stove usage are relatively similar. 81 The monitoring report indicates that in only 17% of the cases with materials problems also installation deficiencies were detected; and that in only 25% of the cases with material problems excessive use of firewood was also reported. Furthermore, from all the households reporting excessive firewood usage, only 20% reported materials problems, and from all the cases in which an installation deficiency was detected, only 28% reported materials problems (MIRHAS PERU – Universidad de Piura internal report 2004). 82 In the case of the metallic chimney, circumstantial evidence also suggest that different types of materials quality were distributed in some villages; however, some of the problems with the chimney were also related to defects in installation and water filtration issues during the winter, mainly at high altitude areas.  74  During recent visits to the intervention area in 2008, I was able to visually confirm the presence of such deficiencies, and that they were among the main reasons why households stopped using the new cooking device. It is very important to emphasize that there is no evidence on beneficiary households complaining about iron frame’s materials quality prior to stove installation and usage; these problems were reported after households made effective use of their new devices (Ureta (2007)); in other words, these problems appeared among households that decided to adopt the new technology as their main cooking device. Moreover, as stove distribution was on charge of NGO members, it is unlikely that households could have selected their stoves based on observable characteristics (i.e. thickness of the iron frame), and it is also unlikely that villages leaders may have successfully influenced the NGO in order to get stoves of better quality. Improved stove distribution progressed as these artefacts were made available by the production workshops, and the path of distribution basically followed the main route of access into the Chalaco District. Iron frame failures were reported in all the five watersheds within the Chalaco District and there is no evidence which may lead us to think that the allocation of the improved stoves with the lowest material quality was done in a systematic way; that is, as a function of certain village characteristics (e.g. poor quality stoves were allocated to the poorest villages83 or to the remotest ones). In fact, the monitoring reports indicate that the higher proportion of beneficiary households which reported iron frame problems relative to the number of beneficiaries using the improved stove without problems, has been observed within villages in the Ñoma Watershed, which is the most accessible watershed in the district as well as the first watershed in which improved stoves were distributed (which makes it very unlikely that worse stoves were “intentionally” distributed in this area by the NGO84). In any case, in the estimation section in this chapter I control for key household characteristics and I also allow for village fixed effects in order to control for any systematic household or village factor that may have influenced the allocation of improved stoves with poor quality iron frames. 83  The standards of living are relatively similar across the villages in the area of the Chalaco District. (Estudio Socioeconómico del Distrito de Chalaco, Universidad de Piura, 2004). 84 In this sense, the appearance of iron frame failures does not seem to have been ex-ante anticipated by NGO members; they started dealing with iron frame problems after improved stoves were effectively adopted by the beneficiary households.  75  To provide additional supporting evidence on the fact that improved stove iron frame materials failures were exogenous to households’ and village characteristics; Table 3.1 (below) reports the main observable ex-ante features for a sample of beneficiary households divided in two groups: beneficiaries making effective use of their improved stoves without any type of material problem, and beneficiaries that reported an iron frame failure. These households were visited during the 2004 monitoring visits and were also interviewed during a socioeconomic survey in the months prior to the improved stove intervention in 2003. As we can observe, these two groups of households were ex-ante very similar in terms of their observable characteristics. For example, it can be argued that younger households are better at adopting and using new technologies and less likely to experience problems of any sort; however, we can see in Table 3.1 that the age of the household head is relatively similar between the two groups of households. It can also be pointed out that ability is a key factor in terms of being able to adopt and use a new technology without complications, and that probably low ability households were more likely to damage their stove iron frame. Nevertheless, note that there are not statistically significant differences in the years of education of the household head (a variable closely linked to ability) between the two groups analyzed in Table 3.1. Moreover, note that households using the stove without problems and households presenting iron frame problems are equally likely to use fertilizer, which alleviates concerns related to systematic differences in the household’s ability to adopt and use new technologies. Another valid point that can be raised is that poorer households were the ones that ended with the worst stoves (and that wealthier ones got the best ones); however, note that there are not statistically significant differences between the two groups in terms of farm assets value, farm size and per capita number of rooms (all these variable potentially related to household’s wealth). Additionally, it could have been the case that iron frames of lower quality were more likely to break down or deform under certain usage conditions; for example, probably this piece of metal was more likely to deform when used by a 6 member family than by a 2 member family; note however that there are not statistically significant differences between the two groups in terms of household’s size. In the same sense, probably iron frame deformations were more likely to appear at higher altitudes, maybe due to the specific nature of the combustion process or because the stove is kept  76  fired for a longer time period in high altitude villages, as it is also used as a heating device during the winter months. However village altitude is only 50 m. higher for those households presenting iron frame problems; moreover, the difference in means is not statistically significant.  Table 3.1 Ex-ante characteristics for beneficiary households using the improved stove without problems and beneficiary households who experienced iron frame problems during the 2004 monitoring visits Beneficiaries Beneficiaries with P-value for the using the stove iron frame difference in without problems problems means test N=129 N=34 50.9 51.5 Household head’s age 0.86 (13.8) (15.4) 5.67 4.84 δ Household head’s years of education 0.20 (3.64) (3.32) Household head’s sex 0.86 0.82 0.60 4.76 4.94 Household’s size 0.68 (2.18) (1.74) Household’s wealth (farm assets value 82.7 73.1 0.71 † in 2003 Peruvian “soles”) (14.5) (5.9) 2.53 3.19 † Household’s farm size (in has) 0.20 (2.40) (3.76) 1.00 0.97 Per capita number of rooms 0.79 (0.78) (0.59) Household uses fertilizer (yes=1) 0.64 0.62 0.78 Household’s village altitude (in 16.54 17.15 0.38 hundred meters) (3.68) (3.15) Standard deviations shown in parenthesis. δ Due to missing observations, in these case there are only 120 users without problems and 31 households with material problems for which the variable “years of education” is available. † I have also compared the per capita values for these variable, with very similar conclusions in terms of the difference in means test among groups  In my opinion, the evidence strongly suggests that, conditional on adopting the new stove as the main cooking device during the 2003 distribution stages, being affected by an iron frame failure constituted an exogenous process which can be exploited to identify the effect of stove usage on firewood consumption. In other words, conditional on having decided to use the new stove as the main cooking device, experiencing an iron frame failure seems to be exogenous to households’ and villages characteristics: you were just randomly given a stove with a poor quality iron frame. In line with this evidence, to estimate the improved stove average treatment effect on the treated (the effect on households that self-selected as adopters of the new cooking device during the distribution stages), in this chapter I use self-reporting an iron frame failure during the  77  2008 survey (which is an indirect indicator for having received a stove of poor material quality) as an instrument for improved stove usage. While it is true that the self-reported nature of my instrumental variable can raise some valid concerns, I asked approximately half of the households that reported an iron frame failure if I could take a look at their iron frames and in almost all the cases and I was able to visually confirm the presence of material failures.  3.4. Data In the summer of 2008, I implemented a household survey in 21 villages85 within the Chalaco District in which improved stoves were freely distributed in 2003. The survey collected information on two important aspects: a) households’ usage patterns of the improved cooking device; and b) households’ levels of firewood collection and consumption. The survey also gathered information on household’s characteristics, economic activities, social capital and access to social programs. It is important to highlight that almost all households in the area collect firewood for self-consumption, with just 1.5% of them reporting firewood sales86; also, firewood stoves are the main devices used for cooking and heating purposes among households in the area, with only 3% of the households reporting usage of gas or kerosene stoves87. In order to design the survey, initial visits to some villages were performed in the month of May 2008. The objective of these visits was to obtain a first insight on stove usage patterns and a better understanding of the process of firewood collection and consumption in the area of study. As a result of the preliminary visits, I decided to divide the firewood section into two subsections. The first centers on firewood collection and consumption during the winter (rainy) season (Mid December to Mid May). Almost all the firewood that households in the Chalaco District consume during this season is collected in the immediate previous 85  Originally, I intended to survey all the 26 villages that were visited during the 2004 monitoring. For the non visited villages, in one case I arrived to the village when all households were participating in a communal party, and then no one was interviewed; in other case I underestimated the travel time, so we had to go back to the main town without making any interview; the other three villages were finally not visited due to time and budgetary constraints. 86 I have excluded these households from the estimation sample. 87 This pattern of behaviour suggests a theoretical model in which firewood production (extraction) and consumption decisions are non-separable.  78  months (November and December) and stored in the household unit; the rainy season is relatively intense in this area, which makes it extremely difficult to collect firewood on a frequent basis during this period. In this subsection, interviewed households were asked for the total amount of firewood from all sources they collected for the winter season88. The second subsection centers on collection patterns during the summer (dry) season. As during these months firewood collection takes place in a very frequent basis, households were asked for the time frequency of firewood collection (e.g. daily, weekly, etc) and the amount collected at each time. As the most serious levels of forest extraction in the area are observed when households collect the firewood they will consume in the winter season, this chapter focuses on households’ firewood consumption during a typical winter month89. In this time of the year firewood collection is closely connected to cutting down trees; while in the dry season collection of small fallen tree branches and the use of other biomass fuels (such as crop residues and animal dung) is more common. In second place, given the specific nature of the collection process (all firewood must be collected at once and stored at the household unit); households have a precise estimate of the total amount they collected for the winter season, probably because they accurately know the dimensions of their storage space. Also importantly, the measurement units used to report winter firewood collection are relatively uniform90, and most households were able to report their consumption in “cargas” (which contains approximately 30 kg. of firewood). In the survey section on improved stove usage patterns, households were asked if they requested the new stove during the distribution stages in the year 2003, if they effectively received the stove, if they installed the stove and made effective use of it, and if they  88  Household mainly reported their answer in two local measurement units: “Cargas” and “Pircas”. A “carga” contains approximately 30 Kilos of firewood. When the household reported an answer in “Pircas”, we asked them for the approximate number of “cargas” a “Pirca” contains. 89 This is calculated by simply dividing the total firewood amount collected by five, which is the approximately duration of the winter season in the area of study. 90 On the other hand, during the dry season collection takes place on a more frequent basis and the large variety of measurement units reported by households (“palos”, ‘tercios”, “chamizas”, ‘brazadas”, etc.) makes it very difficult to come up with a uniform estimate. For example, during the dry season in many cases households reported as a measurement unit “as much as I can carry on my shoulders”.  79  were currently using it as the main cooking device91. Households that in the 2008 survey reported using the improved stove as their main cooking device were specifically asked if in the past they had to repair or change their improved stove iron frame, which is clearly connected to having experienced a material problem with this piece of metal. On the other hand, households that received the improved stove during the distribution stages in 2003, installed it and made effective use of the new device, but in the 2008 survey reported not using it; were asked for the main reason why they stopped using the enhanced cooking device, so I am able to observe which households stopped using their improved stoves due to a material problem with the stove iron frame92. As I mentioned before, initial evidence from the 2004 monitoring reports indicates that conditional on deciding to adopt the improved stove as the main cooking device, iron frame materials problems were random and non ex-ante observable to beneficiary households: individuals experiencing iron frame problems were just given an iron frame of lower material quality. Using this information, I create an indicator variable for having experienced a problem with the improved stove iron frame, which the evidence discussed in Section 3.3 suggests is directly connected to having received a poor quality stove. This indicator takes the value of one in the case of non users that stopped using their improved stove due to an iron frame material problem, and also in the case of current users that experienced iron frame material problems at some point in the past. In the empirical section, I use this selfreported indicator as an instrument for improved stove usage. This identification strategy allows me to identify the improved stove average treatment effect on the treated; that is the effect of stove usage on those households that received the new stove during the 2003 distribution stages and self-selected as improved stove users (that is, in the empirical 91  This section was directly asked to the female spouse or to an adult female member. While current users were directly asked if they experienced a problem with the iron frame, unfortunately non users were asked in a general way for the main reason why they stopped using the stove; then, for those non users that did not report iron frame problems I cannot tell for sure if they effectively received an iron frame of poor quality or not. However, non users that stopped making use of their stove due to a problem other than a material failure seem to have made use of the stove for a very short period of time. For example, more than 50% of these non users reported using the stove for no more than 6 months; which may not be enough usage time for the material problem to reveal (assuming that they indeed made any use of the stove). In the case of non users that reported material problems, the proportion that made use of the improved stove for less than 6 months very small, lower than 15%, which suggest that they were very likely to have initially decided to adopt the stove as the main cooking device (and were then “forced” to stop using it due to an iron frame failure). 92  80  section our estimation sample only include households that decided to effectively adopt the new stove as the main cooking device during the distribution stages).  Table 3.2 Main characteristics for households reporting and not reporting iron frame material problems: 2008 survey I II III Did not reported Reported material p-value material problems problems 51.91 52.96 Household’s head age 0.65 (14.94) (13.81) Household’s head sex 0.87 0.84 0.58 5.78 4.91 Household’s head years of education 0.10 (3.41) (3.24) Years of education of the adult female 5.89 5.64 0.69 member with the highest level of education (4.09) (3.44) 4.13 4.71 Household’s size 0.06 (1.93) (1.99) 0.88 0.88 Per capita number of rooms 0.99 (0.62) (0.58) Household’s wealth (value of farm assets in 194.20 192.69 0.97 2008 Peruvian soles) (270.15) (203.61) 2.40 1.74 Household’s farm size in has. 0.14 (3.17) (1.63) Household uses fertilizer 0.57 0.68 0.15 Household belongs to the local “tree 0.45 0.43 0.79 nursery” Household’s village altitude (in hundred 16.47 17.41 0.08 meters) (3.93) (3.02) Observations 138 56 Standard deviations shown in parenthesis. Column III presents the p-value for the difference in means significance test.  Table 3.2 reports the main observable characteristics for the beneficiary households that reported iron frame failures as well as for those that did not report this problem during the 2008 interviews. As we can observe, these two groups of households are relatively similar in most of their main observable ex-post characteristics. In the few cases in which it appears to be some sort of difference among the characteristics of the two groups, this is only statistically significant at the 10% significance level93 and relatively small in size. The information contained in Table 3.2 is relatively similar to the one presented in Table 3.1, and provides additional supporting evidence towards the exogenous nature of self-  93  I have also estimated an OLS regression (results not shown) for reporting an iron frame problem in the 2008 survey in which I control for all the variables included in Table 3.2; the F-statistic for the joint significance of the model coefficients is in this case equal to 1.51 (with a p-value equal to 0.11).  81  reported iron frame material problems. It is also important to note that it is not unlikely that certain villages received a higher proportion of improved stoves with low quality iron frames; however if this was the case, it will not represent a problem in terms of the identification strategy as long as this allocation was not done in a systematic way (i.e. the NGO did not purposively distributed bad stoves in poorer villages or leaders in some villages were more successful at getting better stoves), and if within villages stove allocation was not influenced by observable improved stove material characteristics. In terms of the identification strategy that will be followed in this research, it is important to take into account that some of the beneficiaries that presented iron frame problems were able to obtain a new improved stove (generally of better quality) while others were able to repair their iron frames (these iron frames were mainly reinforced with longitudinal metallic bars to prevent deformations). Then, if a repaired stove is “substantially different” than a stove that did not present any material problem, operating an improved stove of lower iron frame quality is likely to have its own direct effect on firewood consumption; however, the main responsible NGO members involved in the improved stove program reported that “repaired” stoves performed in a very similar way as improved stoves without iron frame problems.  Table 3.3 Firewood consumption during the 2008 winter season: improved stove users and non users All Stove Users Non Users p-value Households 39.24 Total firewood collected for the last winter 37.71 46.96 0.09 (28.23) season (in cargas) (27.95) (28.79) 7.84 Monthly firewood consumption during the last 7.54 9.39 0.09 (5.64) winter season (cargas per month) (5.59) (5.75) Observations 194 162 32 Standard deviations shown in parenthesis. The fourth column presents the p-value for the difference in means test.  The next section presents the OLS and instrumental variables results for the effect of improved stove usage on firewood consumption, which is measured in “cargas” per month (a “carga” contains approximately 30 kilos of firewood). As the winter season in the area last for approximately five months (mid December to mid May) and households collect all the firewood they will consume in this season in the immediate preceding  82  months (mainly November and December), I simply divide the total amount of firewood the household collected for the winter season by five in order to obtain the household´s monthly firewood consumption during the winter season94. Table 3.3 shows the total and monthly firewood consumption levels for the households in our sample. As we can see, users of the improved stove appear to consume approximately 1.85 less “cargas” (56 Kg.) of firewood per month during the winter season than users of the traditional stove, and the difference in raw means is statistically significant at the 10% significance level.  3.5. Empirical Estimations 3.5.1. Baseline OLS Estimations In order to estimate the partial correlation between improved stove usage and household’s monthly firewood consumption during the winter season in the Chalaco District, in this section I estimate the following fixed effects linear regression: (1) Fij = α 0 + α 1 Stoveij + θ . X ij + c j + eij  In equation (1) the dependent variable Fij represents the natural log of the monthly firewood consumption for household “i” in village “j” during the winter season (in “cargas”)95. The term Stoveij is a dummy variable which takes the value of one if the household uses an improved stove as the main cooking device and zero if it does not. X ij is a vector of household’s characteristics, and includes the household head’s age, sex and years of education; the years of education of the adult female member with the maximum education level in the household; the household’s size, per capita number of rooms, farm size (in hectares), wealth (measured by the per capita value of the farms assets in Peruvian soles), use of fertilizer, processing of farm and animal processed products (such as alcohol, cheese, clothing, etc) and participation in the communal “tree nursery” (as a 94  To confirm this I asked households for the number of months the firewood they collected lasted, the average number reported by users and non users is very similar: 5.27 and 5.17 respectively, and the difference in raw means is not statistically significant (p-value=0.70), which suggest that it is unlikely that the adoption of the improved stove also affected the time horizon for firewood collection. 95 I take the natural log of monthly firewood consumption because the distribution is relatively skewed otherwise.  83  measure of the household´s pro-forest preferences). The term c j represents a village fixed effect and eij is a random disturbance which is assumed to be correlated among households in the same village. The regression that corresponds to column I in Table 3.4 only controls for improved stove usage; the results in this column indicate that on average improved stove usage reduces log firewood consumption by 0.27 log points during a typical winter month (that is approximately a 30% reduction in firewood consumption), and the improved stove dummy coefficient is statistically significant at the 1% significance level. The regression corresponding to column II in the same table also controls for household’s characteristics; as we can observe, the estimated coefficient for the improved stove usage dummy is very similar to the one estimated in column I and it is statistically significant at the 5% significance level. In column III I also allow for village fixed effects; as we can see, the improved stove usage coefficient is very similar in sign and absolute size to the coefficients estimated in columns I and II, and it is statistically significant at the 1% significance level96.  Table 3.4 The effect of improved stove usage on monthly firewood consumption I II Households effectively uses the improved -0.27*** -0.27** stove as the main cooking device (yes=1) (0.10) (0.10) Household Controls Included NO YES Village fixed effects included NO NO R2 0.03 0.25 N 194 194 Villages 19 19  III - 0.26*** (0.08) YES YES 0.37 194 19  The dependent variable is the natural log of the amount of firewood consumed in a typical winter month (in cargas). All regressions have been estimated clustering the standard errors at the village level. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels. Standard errors shown in parenthesis.  In Table 3.4, the user category included all households that reported using the firewood improved stove as the main cooking device; however, a few of them also own a traditional open fire firewood stove, which they report to use in a very occasional basis. In Table 3.5 I just include in the “users group” those beneficiary households that “only” 96  I have estimated alternative specifications which control for other household variables available in the data; the effect of improve stove usage in these cases is very close to the obtained in Tables 3.4 and 3.5.  84  use the improved firewood stove as the household’s cooking device. As we can observe, the results in all the columns in Table 3.5 are very similar in terms of sign, absolute size and statistical significance to the results obtained in Table 3.4.  Table 3.5 The effect of improved stove usage on monthly firewood consumption (user group just includes households that use the improved stove as the only cooking device) I II III Households effectively uses the improved -0.30*** -0.31*** -0.28** stove as the only cooking device (yes=1) (0.09) (0.10) (0.10) Household Controls Included NO YES YES Village controls included NO NO YES R2 0.03 0.26 0.38 N 173 173 173 Number of Villages 19 19 19 The dependent variable is the natural log of the amount of firewood consumed in a typical winter month (in cargas). All regressions have been estimated clustering the standard errors at the village level. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels. Standard errors shown in parenthesis.  3.5.2. Instrumental Variables The estimated coefficient for the improved stove usage dummy in Tables 3.4 and 3.5 is very stable across all model specifications. As we could see, adding household controls and villages fixed effects does not significantly affect the point estimates for the effect of improved stove usage on the natural log of monthly firewood consumption. However, it is practically impossible to control for every household factor (many of them unobservable to the econometrician) simultaneously correlated with firewood consumption and the household decision to use the improved stove as the main cooking device (such as women’s empowerment, forest preferences or unobserved ability to adopt new technologies and manage forest resources); then, the baseline OLS results in tables 3.4 and 3.5 may not be able to capture the causal effect of improved stove usage on firewood consumption. In this section, an instrumental variable approach is proposed in order to deal with this identification problem. As previously discussed in Section 3.3, there is strong circumstantial evidence indicating that: a) some of the iron frames distributed in the year 2003 were of inferior material quality than others, b) receiving a low quality iron frame was exogenous to household’s characteristics, and c) this condition affected the improved stove long term usage decision.  85  In the 2008 survey I asked non user households (only those that initially received and made effective usage of the improved stove) for the main reason why they stopped using the new cooking device; and I asked current user households if in the past they experienced a problem with their improve stove iron frame. Then, among beneficiary households which decided to effectively use the improved stove, I am able to observe which ones were likely to be affected by iron frame material problems, and I use this information to construct an indirect indicator for having being allocated an improved stove of poor iron frame material quality. In this section I employ this indicator as an instrument for improved stove usage. The second and first stage regressions are correspondingly given by equation 1 (in Section 3.5.1 above) and equation 2 below: (2) Stoveij = β 0 + β1 IF ij +δ . X ij + c j + vij  In equation (2), IF ij takes the value of one if the household reported problems with the stove iron frame, which is an indirect indicator for whether the household was allocated an iron frame of poor material quality or not. The interpretation for the other controls included in equation (2) is the same as in equation (1).  Table 3.6.A The effect of improved stove usage on monthly firewood consumption Instrumental variables: second stage regressions I II Households effectively uses the improved -0.38** -0.36* stove as the main cooking device (yes=1) (0.16) (0.17) Household Controls Included NO YES Village fixed effects included NO NO R2 0.02 0.25 N 194 194 Number of Villages 19 19 Table 3.6.B Instrumental variables: first stage regressions I II Household Reported an iron frame problem -0.57*** -0.58*** (instrument) (0.04) (0.04) R2 0.48 0.55  III -0.36* (0.19) YES YES 0.37 194 19  III -0.54*** (0.05) 0.60  The dependent variable is the natural log of the amount of firewood consumed in a typical winter month. All regressions have been estimated clustering the standard errors at the village level. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels. Standard errors shown in parenthesis. ^ In this case the p-value is equal to 0.12.  86  Table 3.6.A shows the instrumental variables regression results. The first column in this table only controls for improved stove usage; as we can see, the results indicate that using an improved stove as the main cooking device reduces monthly log firewood consumption by 0.38 log points on average (that is a 46% reduction in firewood consumption), and this effect is statistically significant at the 5% significance level. The instrumental variables regression that corresponds to column II in Table 3.6.A also controls for household’s characteristics; note that in this case the effect of improved stove usage appears very similar in sign and absolute size to the effect estimated in column I and it is significant at the 10% significance level. Adding village fixed effects in column III does not have any significant impact on the estimation results; the estimated coefficient for the effect of improved stove usage on the log of firewood consumption in this column is very similar to the coefficients obtained in columns I and II, and it is statistically significant at the 10% significance level97. As was the case in Table 3.5, in Table 3.7.A (below) I just include in the “user category” those households that use the improved stove as the “only” cooking device. The coefficients for the effect of improved stove usage are in all columns in Table 3.7.A slightly higher (in absolute size) than the ones obtained in Table 3.6.A; as well as statistically significant at the standard significance levels. Note that in Tables 3.6.A and 3.7.A, the coefficient for the effect of stove usage is relatively stable in terms of sign,  97  The IV regression results in Table 3.6.A confirm the negative and significant effect of improved stove usage; however, just by taking a look at the standard deviations for the coefficient of improved stove usage, we can intuitively predict that if we perform the Hausman’s test, we will be not able to reject the null hypothesis that the IV coefficient is not systematically different than the OLS one. In general this type of test has being implemented as an endogeneity test, in which failure to reject the null is interpreted as failure to reject the null hypothesis that the regressor in the OLS regression is exogenous (so if you do not reject the null hypothesis, the OLS estimation is the appropriate one). The Hausman’s test can be estimated in STATA, however from the STATA help file we have that: “The assumption (in the Hausman’s test) that one of the estimators is efficient (i.e., has minimal asymptotic variance) is a demanding one. It is violated, for instance, if your observations are clustered or pweighted, or if your model is somehow misspecified. Moreover, even if the assumption is satisfied, there may be a "small sample" problem with the Hausman’s test. Hausman's test is based on estimating the variance var(b-B) of the difference of the estimators by the difference var(b)-var(B) of the variances. Under the assumptions (1) and (3) (see STATA Hausman´s test help for details) var(b)-var(B) is a consistent estimator of var(b-B), but it is not necessarily positive definite "in finite samples", i.e., in your application. If this is the case, the Hausman test is undefined. Unfortunately, this is not a rare event. STATA supports a generalized Hausman test that overcomes both of these problems. See suest for details”. Unfortunately, the suest command does not support the ivreg command.  87  absolute size and statistical significance among all the instrumental variables specifications. All in all, the second stage results in the instrumental variables regressions confirm the sizeable and statistically significant effect improved stove usage has in reducing monthly firewood consumption in the area of study during the winter season.  Table 3.7.A The effect of improved stove usage on monthly firewood consumption Instrumental variables: second stage regressions (user group includes just households that use the improved stove as the only cooking device) I II III Households effectively uses the improved -0.46** -0.42** -0.40* stove as the main cooking device (yes=1) (0.17) (0.16) (0.19) Household Controls Included NO YES YES Village fixed effects included NO NO YES R2 0.03 0.25 0.38 N 173 173 173 Number of Villages 19 19 19 Table 3.7.B Instrumental variables: first stage regressions I II III Household Reported an iron frame problem -0.58*** -0.59*** -0.55*** (instrument) (0.05) (0.05) (0.05) R2 0.48 0.55 0.63 The dependent variable is the natural log of the amount of firewood consumed in a typical winter month. All regressions have been estimated clustering the standard errors at the village level. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels. Standard errors shown in parenthesis. ^ In this case the p-value is equal to 0.12.  At this point, it is important to highlight that for the instrumental variable approach to be valid, not only the random assignment condition must be satisfied, but the exclusion restriction must also hold. Moreover, as Imbens and Wooldridge (2009) clearly state, the assumption that the instrument does not directly affect the outcome is not implied by the random assignment, and the argument must be made in a case by case basis. In the context of this study, it must be the case that having received an iron frame of lower quality does not have a direct effect on firewood consumption once current improved stove usage is controlled for; in other words, having received a stove with a lower quality iron frame during the 2003 distribution stages should only influence the variable of interest trough the improved stove current usage decision. In the case of this study, it is known that some of the current improved stove users reported initial problems with their iron frame; while many of them were able to obtain a new iron frame, others had their iron frame repaired. Then, if a repaired stove is “substantially different” than a stove that  88  did not present any material problem, having an operative improved stove of poor iron frame quality is likely to have its own direct effect on firewood consumption. As I only have one instrumental variable, it is not possible for me to test if the instrument has been correctly excluded from the second stage regression; however, the main responsible NGO members involved in the program informed me that “repaired” stoves were able to perform in a very similar way as improved stoves without iron frame deficiencies98. As I mentioned before, this chapter attempts to estimate the average treatment effect for those households that received the improved stove, installed it and decided to effectively use the new technology as the main cooking device during the 2003 distribution stages. However, in the presence of heterogeneity, endogeneity creates serious problems for identification of the population of interest averages in our instrumental variables approach. As discussed by Imbens and Wooldridge (2009), “population average causal effects are only estimable under very strong assumptions on the effect of the instrument on the endogenous regressors (sometimes referred to as “identification at infinity”, Chamberlain, (1986)) or under the constant treatment effect assumptions”. In the presence of heterogeneity, our instrumental variable approach is more likely to identify a Local Average Treatment Effect. More precisely, in the presence of heterogeneous treatment effects, our instrumental variable approach will be only informative about the average treatment effect on compliers, that is, on households who would keep making use of the new stove if they were allocated a good materials one, and would stop making use of the new device if they were allocated a poor iron frame quality stove99. For this to be true, the monotonicity assumption should be satisfied. In simple terms, the monotonicity condition means that we should be able to rule out what is known as the “defier” behaviour. In other words, in our population of interest, there should not be individuals that would keep making use of the improved stove in a situation in which they were given 98  To provide evidence on the fact that a repaired stove performs in a very similar way than a stove of good quality, I estimated a regression only including current users of the improved stove, in this regression firewood consumption is the dependent variable and I control for whether the household experienced an iron frame problem (which should be exogenous); the results (not shown here) indicate that having experiencing an iron frame failure does not have a significant effect on current firewood consumption. 99 I thank Kevin Milligan for pointing out the identification problem generated by the presence of heterogeneous treatment effects, and David Green for pointing out that if the monotonicity assumption is satisfied, at least the estimates should be able to identify a LATE.  89  a stove with a poor quality iron frame and would stop making use of the new stove in the hypothetical case in which they were allocated a stove of good iron frame quality. For the specific context of the stove intervention in the Chalaco District, I think that this is a reasonable assumption to make100.  3.5.3. Household Factors Table 3.8 Natural log of monthly firewood consumption: household level controls I II 0.01 0.01* Household head’s age (0.00) (0.00) -0.03 -0.01 Household head’s sex (0.11) (0.13) 0.05** 0.05** Household head’s years of education (0.02) (0.02) Years of education for the female member with the highest 0.01 0.01 education (0.02) (0.02) 0.37*** 0.35*** Household’s size (0.08) (0.09) -0.03*** -0.02*** Household’s size^2 (0.01) (0.01) 0.06 0.06 Farm size (0.04) (0.04) -0.00 -0.00 Farm size^2 (0.00) (0.00) 0.00*** 0.00*** Per capita Farm Assets Value (in soles) (0.00) (0.00) 0.19** 0.18* Per capita number of rooms (0.08) (0.09) Household elaborates farm or animal derivate products (liquor, -0.02 -0.02 cheese, floor, clothing, etc) (0.09) (0.09) -0.12 -0.13 Household belongs to the local “tree nursery” (0.11) (0.11) -0.03 -0.03 Household uses fertilizer (0.12) (0.12) Village fixed effects included YES YES R2 0.37 0.37 N 194 194 Number of Villages 19 19 The dependent variable is the natural log of the amount of firewood consumed in a typical winter month. Column I shows the household level factors that correspond to the estimation in column III in Table 3.4. Column II shows the household level factors that correspond to the estimation in column III in Table 3.6.A. All regressions have been estimated clustering the standard errors at the village level. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels. Standard errors shown in parenthesis.  100  Heckman et al (2006) provide an interesting discussion on how to interpret instrumental variable regressions results in the presence of essential heterogeneity (heterogeneity in treatment effects or outcomes).  90  Columns I and II in Table 3.8 present the estimation results for the household level controls included in the fixed effects regressions corresponding to column III in Tables 3.4 and 3.6.A respectively. As we can observe, the size and significance levels for the household’s control coefficients included are relatively stable across both columns. Note in first place that the household head’s years of education appears to be positively and significantly correlated with firewood consumption. Although this result seems to be contradictory, as one may expect than more educated (and probably high ability) households are more efficient at processing firewood for energy purposes; we must also take into account that households with higher education may be better at collecting (producing) firewood (Chen et al (2006)), and then have more firewood available for consumption during the winter season. As both effects are likely to be present, the results in Table 3.8 suggest that the later effect seems to be dominated the first one. Household’s size appears to have a non linear effect on firewood consumption; this result can be explained by the potential presence of increasing returns to scale in firewood usage for food preparation and heating. Also, as household’s size increases more labor is available for firewood collection, and the negative size for the quadratic term could also indicate the presence of decreasing returns to labor in firewood collection tasks. The results in Table 3.8 also indicate that households with a higher number of rooms per capita appear to significantly consume more firewood during the winter season, probably due to higher heating needs. Finally, the coefficient for the household’s per capita value of farm assets is positive and statistically significant in both columns; this result just indicates that, as expected, richer households have more access to firewood resources.  3.6. Conclusion The massive distribution of firewood improved stoves has been one of the main strategies implemented in the developing world to alleviate forest resources degradation due to firewood extraction for energy purposes. Quite surprisingly, the current empirical evidence on the effect of these devices on rural households’ firewood consumption is relatively scarce and mostly inconclusive. Moreover, most of the studies in the literature fail to take into account the endogeneity of the stove usage decision; and then, it is not  91  possible to separate the effect of stove usage from the effect of unobservable factors simultaneously correlated with firewood consumption and the household´s stove usage decisions. This chapter tries to help filling this gap by estimating the effect of using an improved firewood stove as the main cooking device on households’ firewood consumption during the winter season in the rural communities within the Chalaco District, in the Northern Peruvian Andes. In order to identify the causal effect of stove usage, this chapter exploits a quasi-experiment related to the presence of iron frame materials problems in some of the stove distributed. Initial evidence from the 2004 monitoring reports suggests that material deficiencies were random and ex-ante unobservable; moreover, these deficiencies influenced the long term usage decision. In line with this evidence, in the 2008 survey implemented in the intervention area I asked non user households for the main reason why they stopped using the new cooking device; and I asked current users if they experienced a problem with their stove iron frame in the past. This information allows me to construct an indicator which identifies those households that were allocated an improved stove of poor material quality during the distribution stages, which I then use as an instrument for current improved stove usage. The instrumental variable results confirm the baseline OLS findings, and indicate that using an improved stove reduces monthly firewood consumption by approximately 40% during the winter season. In the presence of heterogeneous treatment effects and if the monotonicity or “no defiers” assumption is satisfied (which is very likely to be the case in the context of this study), this estimated effect can be interpreted as a LATE: the effect of improved stove usage on compliers´ firewood consumption levels. Although the results in this chapter cannot be extended to all contexts, the villages analyzed in this study are very similar in their main characteristics to the typical village in the northern Peruvian Andes, where deforestation and forest degradation have been identified among the main causes behind land erosion, declining soil fertility and flooding. Up to my current knowledge, this is one of the first formal evaluation studies on the environmental effects of improved firewood stoves which carefully intends to address the endogenous nature of the improve stove usage decision.  92  4. Improved Stove Adoption and Health Outcomes 4.1. Introduction Approximately three billion people around the globe rely on biomass (fuelwood, charcoal, dung and crop residues) and coal as their main source of domestic energy (Reddy et al (1996)); and it is estimated that biomass accounts for 50% to 95% of the primary energy consumption in low income countries (Werecko-Brobby et al (1996)). Different studies in the epidemiological literature indicate that there is a clear connection between biomass and coal usage and the incidence of acute respiratory illnesses and chronic pulmonary diseases due to increased exposure to indoor air pollution (IAP), especially among adult women and infants (Ezzati and Kammen (2002)); and the World Health Organization has ranked IAP from solid fuels as the 8th most important risk factor for attributable preventable loss of disability-adjusted life years (Diaz et al (2006)). Adding to this evidence, in a recent report based on a household survey in rural Orissa (India), Duflo et al (2008b) show that there is a strong correlation between using a clean fuel101 stove and having better respiratory health, particularly among infants and women; which suggests that the use of biomass traditional stoves is a critical factor behind the incidence of respiratory problems102. In the recent years a few studies, mainly in the epidemiological literature, have started to pay attention to the effect of improved firewood stoves with a metallic chimney (specially designed to reduce exposition to IAP) on rural households’ health (Masera et al (2007), Chengapa et al (2007), Diaz et al (2006) and Smith-Sivertsen et al (2009)). From all these research works, the papers by Diaz et al (2006) and Smith-Sivertsen et al (2009) are in terms of their identification strategy the most important studies on the health benefits 101  LPG or electricity stoves. However, as the authors clearly state, the decision to use a clean stove is likely to be correlated with other factors (many of them unobservable to the econometrician) which also affect respiratory health, making it complicate to identify the causal effect of using a clean fuel cooking device. The authors are currently involved in a program that randomly distributed improved cooking stoves with a chimney in rural Orissa, and are conducting follow up studies to determine whether the improved stove improves respiratory health and households’ welfare. As the distribution of stoves was done in a random way, the authors indicate that any difference in outcomes can be attributed solely to the improved cooking stoves. The results from this study are expected to be available in the near future. 102  93  of improved firewood stove usage. These two papers exploit a randomized trial (the RESPIRE program) which distributed improved firewood stoves (“plancha”) with a metallic chimney mechanism among 500 households in Mayan villages in Guatemala, and periodically evaluated adult women’s respiratory illnesses and other discomfort symptoms at initial stages of improved stove adoption (within the first 6, 12 and 18 months). The paper by Diaz et al (2006) shows that effective usage of an improved stove significantly reduces the self-reported incidence of eye discomfort among adult women; however, no significant effect was found on the self-reported incidence of headache or back pain. The findings in Smith-Sivertsen et al (2009) indicate that effective usage of an improved firewood stove significantly reduces IAP as well as the self-reported incidence of respiratory symptoms; nevertheless, no significant effect was found when an objective measure of respiratory health (lung function measured by a spirometry) was used as the dependent variable in the main estimations. As previously noted, these studies were carried at relatively early adoption stages; however, for policy makers it is of critical importance not only to know the short term benefits of this type of intervention, but also how the continuous usage of the device will affect the health of beneficiary households in the long term. Moreover, recent research by Chapman et al (2005) suggests that to identify a clear effect on self-reported symptoms related to chronic obstructive pulmonary diseases, a follow up of approximately 10 years would be required. This chapter contributes to the existing literature by exploring how “long term” improved firewood stove usage with an operative chimney mechanism affects rural households’ self-reported health. This research uses household level data I collected during the summer of 2008 in the Chalaco District, in the Northern Peruvian Andes, where improved stoves with a metallic chimney were freely distributed in the fall of 2003. In the 2008 survey, households were asked about the incidence of respiratory illnesses and eye discomfort symptoms in the previous 12 months, and reported which members were the affected ones. This chapter specially focuses on the health effects on adult women that reported being on charge of housekeeping activities (housewives). These women are the main responsible persons for food preparation and spend a relatively higher amount of time inside the household dwelling, which makes them more exposed to episodes of IAP  94  than other individuals in the household. To identify the casual effect of firewood improved stove usage with an operative chimney on self reported respiratory health and eye discomfort symptoms, I follow the same identification strategy as in Chapter 3. That is, I exploit a “quasi-experiment” related to differences in materials quality among the stoves distributed. Monitoring visits performed in 2004 indicate that a proportion of households that decided to use the new stove as the main cooking device presented chimney and iron frame material problems; and that in the case of iron frame problems (mainly cracks and deformations), these were not likely to have been systematically caused by deficiencies in stove installation, maintenance or improper usage, but by poor materials quality. Members of the monitoring team at Universidad de Piura interviewed in the year 2008 confirmed that iron frame deformations were likely associated with poor materials quality, and that this problem prevented the permanent adoption of the new stove. The evidence also suggests that iron frame material deficiencies were random and ex-ante non observable to the beneficiaries103. Given the above mentioned evidence, in this chapter (as in Chapter 3) I use self-reporting an iron frame failure during the 2008 survey as an indicator for having received a stove of lower material quality, and I employ this indicator as an instrument to identify the effect of improved stove usage. While it is true that the self-reported nature of my instrument can raise some concerns, I asked approximately half of the households reporting this problem if I could take a look at their iron frames and for almost all of them I was I able to visually confirm the presence of such material problems. The unconvinced reader may suggest that material deformations could have been also influenced by geographic characteristics (a stove of bad quality was more likely to deform at higher altitudes, even if operated by similar individuals) or by households’ improve stove usage patterns (particularly in the long term). Although this was not likely to be the case, as if properly built an iron frame was expected to work without problems for at least 10 years, I deal with these issues by additionally controlling for several household factors as well as by  103  However, it is not unlikely that certain villages may have received (although in a no systematic way) a higher proportion of stoves with material problems than others.  95  estimating a village fixed effects regression. As we will see in the empirical section, the baseline results are robust to different model specifications. The results in this chapter indicate that using an improved stove with an operative chimney significantly decreases the housewives’ likelihood of suffering from respiratory illnesses and eye discomfort symptoms by at least 18 and 17 percentage points respectively. No significant effect on housewives’ respiratory health or eye discomfort symptoms was found when a regression including only improved stove users without an operative chimney was estimated, which indicates that reductions in IAP is indeed what drives the observed results in this essay. No significant effect of stove usage was found on eye discomfort symptoms among adult male individuals, which suggests that the improved stove is more likely to reduce exposition to IAP in those situations in which adult women are more likely to be exposed to pollutant concentrations, such as cooking activities. This chapter develops as follows; Section 4.2 describes the data, Section 4.3 presents the empirical methodology and the main results for the impact of stove usage on health outcomes; finally, Section 4.4 concludes104.  4.2. Data During the summer of 2008, I implemented a survey in the Chalaco District, in the Northern Peruvian Andes. As already discussed in Chapter 3, in this survey I asked those households making effective use of the improved stove if in the past they had to change or repair their stove iron frame, which is clearly connected to having experienced a material problem with this piece of metal. On the other hand, households that received, installed and initially made effective use of the improved stove, but that were not making use of it during the 2008 survey, were asked for the main reason why they stopped using the device; so I am able to observe which program beneficiaries stopped using their new stoves due to a material problem with the stove iron frame105. Using this information, I 104  For a more detailed discussion on the identification strategy please refer to Section 3 in Chapter 3. While current users were directly asked if they experienced a problem with the iron frame; unfortunately, non users were asked in a general way for the main reason why they stopped using the stove; then, for those non users that did not report iron frame problems I cannot tell for sure if they effectively received an iron frame of poor quality or not. However, non users that stopped making use of their stove due to a problem other than a material failure seem to have made use of the stove for a very short period of time. For 105  96  create an indicator variable for having experienced a problem with the stove iron frame; this indicator takes the value of one in the case of non users that stopped using their stove due to an iron frame failure and also in the case of current users that experienced iron frame problems in the past. In the empirical section, I use this indicator as an instrument for improved stove usage. As mentioned in the previous sections, initial evidence from the 2004 monitoring reports indicates that conditional on deciding to adopt the stove as the main cooking device, iron frame materials problems were random and ex-ante non observable to beneficiary households: individuals experiencing this problem were likely given an iron frame of lower material quality. This identification strategy allows me to estimate the improved stove average treatment effect on the treated; that is the effect of stove usage on households that received the improved stove during the 2003 distribution stages and initially self-selected as effective users (adopters) of the new technology. Although the survey main focus was on the effect of improved stove usage on firewood consumption, it also included a couple of simple questions related to households’ respiratory health and eye discomfort symptoms. More precisely, the household head’s spouse (or the most informed female member present at the time of the interview) was asked the following health related question: “En los últimos doce meses algún miembro de su familia se ha enfermado de las vías respiratorias o de los bronquios (ha sufrido de irritacion en los ojos); si fuera ese el caso indicar quien” (Did any of the household’s members suffered from respiratory related diseases or bronchitis (eye irritation) in the last 12 months, if so please indicate who?). Unfortunately the survey health questions do not distinguish among different types of respiratory illnesses (or symptoms); so although it is related to the household’s respiratory health status, it is relatively broad in nature. Using the responses to these questions, I try to identify the effect of “long term” improved stove usage on “general” self-reported respiratory health and eye discomfort symptoms106. example, more than 50% of these non users reported using the stove for no more than 6 months; which may not be enough usage time for the material problem to reveal (assuming that they indeed made any use of the stove). In the case of non users that reported material problems, the proportion that made use of the improved stove for less than 6 months very small, lower than 15%, which suggest that they were very likely to have initially decided to adopt the stove as the main cooking device (and were then “forced” to stop using it due to an iron frame failure). 106 A few households that stopped using the new stove within a year before the interview are excluded from the main estimations.  97  It is important to take into account that there are some current users of the improved stove which chimney was not it place or was in non operative conditions (many of them broken or burned) by the time of the 2008 survey. As the main purpose of this research is to evaluate the effect of an improved stove with an operative chimney as compared to the traditional open fire one, I exclude this group of users from the estimation sample107. In total, the estimation sample contains 384 individuals which belong to 90 households within 19 villages108.  Table 4.1 Incidence of respiratory illnesses and eye discomfort symptoms among different groups of household members: 2008 household survey Incidence of self-reported Incidence of self-reported eye respiratory illnesses discomfort All Individuals (N=384) 0.32 0.15 Housewives (N=96) 0.30 0.32 Adult males (N=139) 0.22 0.13 Infants (≤5 years) (N=38) 0.46 0.02  Table 4.1 reports the incidence of respiratory illnesses and eye discomfort symptoms among different groups of household´s members in the sample. As we can see, approximately 31% of all the individuals in the sample reported being affected by respiratory illnesses in the last 12 months; while this percentage is only equal to 15% in 107  When I include all stove users in the sample and I estimate the average effect of “using the improved stove” (whether with a properly working chimney or not), the (OLS and IV) results are relatively close to the ones that will be discussed in the next sections, in which users without a chimney have been excluded from the estimation sample. Also, as we will see in Section 4.3.3, when I estimate a regression in which among users I only included those using the stove without a chimney, the effect of improved stove usage is very small and not statistically significant, which suggest that indeed improved stove usage with a chimney is what drives improvements in health through a reduction in indoor air pollution. Including everyone in the sample and adding an interaction term between improved stove usage and having an operative chimney presents several difficulties. In first place for users of the improved stove without a chimney I do not know the exact time at which the chimney stopped working. In second place, improved stove usage may have a positive effect on respiratory health by channels other than reductions in IAP; for example, as less firewood is required for cooking and heating, adult women need to make fewer trips to collect firewood, which in the end may improve women’s general health, and by consequence will also affect respiratory health. In third place, and more importantly, I only have one instrument, and including an interaction term in this case will require an additional one. 108 The main ex-post characteristics for households reporting and not reporting material problems in this sample are shown in Table C.3 in Appendix C. As we can see in Table C.3, in general both groups look relatively similar in terms of their observable ex-post characteristics; in the cases in which there seems to be a significant difference in simple means, this appears to be relatively small in size. Moreover, when we estimate a multiple OLS regression for reporting an iron frame problem including all the variables in Table C.3, none of them appears as statistically significant and the model F-statistic is very low (this regression is shown in Table C.4 in Appendix C).  98  the case of eye discomfort symptoms. We can also observe that approximately 1 in 3 housewives reported suffering from respiratory illnesses and eye discomfort symptoms, while only 1 in 5 adult males reported being affected by respiratory illnesses and only 1 in 7 adult males reported being affected by eye discomfort symptoms. Note also that approximately 1 in 2 infants (children 5 years old or younger) reported suffering from respiratory illnesses (which makes them the most affected group of household members); on the other hand, only 1 in 50 infants were affected by eye discomfort symptoms. Interestingly, for the case of respiratory illnesses, the affected proportions of housewives and infants in my small sample are relatively close to the ones reported by Duflo et al (2008b) for a sample of 2220 households in rural Orissa, India.  4.3. Empirical Methodology and Results As it has been discussed in the epidemiological literature, rural household members’ exposition to IAP is a complex issue which depends on a variety of individual´s, household’s, social, cultural and geographical factors. Nevertheless, it is clear that some household’s members are more likely to be exposed to higher levels of IAP than others. In particular, housewives tend to be more exposed to episodes of high concentrations of pollutant particles than other household’s members, as the they are on charge of food preparation, being in closer contact with the stove during tasks such as adding or moving the fuel, lifting the stove, placing the pots, etc (Ezzati and Kammen (2002)). Then, the main focus in this chapter will be on the health benefits of stove usage on this group of adult women. In order to estimate the effect of stove usage on housewives’ health, the following empirical equation will be estimated:  (1 ) R ihv = θ 0 + θ 1 SCH  hv  + θ 2 NH  ihv  + θ 3 SCH  hv  *NH  ihv  + α . X ihv + β .Y hv  + V v + e ihv In equation 1, Rihv is a dummy variable which takes the value of one if the individual “i” in a given household ‘h” in a particular village “v” reported suffering from respiratory illnesses (eye discomfort) in the previous 12 months, and zero otherwise. SCH hv is a dummy variable taking the value of one if the individual belongs to a household that uses 99  an improved stove with an operative chimney as the main cooking device (and zero if not); NH ihv is a dummy variable that takes the value of one if the individual is not a housewife (and zero if it is) and θ 3 SCH hv * NH ihv is the interaction term between the non housewife status and the stove usage dummies. Note that given the way the dummy variable for the housewife status has been defined, the coefficient θ 1 directly captures the effect of stove usage on housewives. In the same equation, X ihv and Yhv are vectors of individual and household characteristics respectively, and include variables such the individual’s age and sex, the household head’s years of education, household’s overcrowding (number of rooms per person) and household’s wealth. The term Vv represents a village fixed effect and eihv is a random disturbance which is assumed to be correlated among individuals in the same household (so the error terms are clustered in all regressions at the household level).  4.3.1. Baseline OLS Estimations In this section, I use a linear probability model instead of a probit (or logit) model to estimate the effect of improved stove usage on the binary variables of interest. The main reason for this is that the linear probability model allows us to control for village fixed effects without biasing the other coefficients in the model (Bandiera and Rasul (2006)).  Table 4.2 The effect of improved stove usage with an operative chimney on housewives’ eye discomfort symptoms I II III IV Household uses an improved stove with an 0.007 0.071** 0.010 operative chimney: effect on all other household (0.031) (0.032) (0.036) members Household uses an improved stove with an -0.177** -0.235** -0.166 operative chimney: effect on housewives (0.089) (0.094) (0.164) Village Fixed Effects Individual and Household controls R2 Observations  NO YES 0.20 384  NO YES 0.22 384  YES YES 0.25 384  YES YES 0.27 96  Observations in columns I to IV correspond to a total of 90 households in 19 villages. The dependent variable is the incidence of eye irritation in the last 12 months. All columns control for age, sex, years of education of the adult household member with the highest level of education, per capita number of rooms, per capita value of farm assets, and a dummy variable which takes the value of one if there is a household member 11 years old or younger. In columns I to III the standard errors are clustered at the household level; while in column IV standard errors are clustered at the village level. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  100  Table 4.2 reports the results for the eye discomfort symptoms regression. The first column in this table does not allow for an interaction term between the improved stove usage and the housewife condition dummies; as we can see, the coefficient for the stove usage dummy (which in this case captures the average effect of stove usage on any household´s member) is very small in absolute size and not statistically significant. In column II, I allow for an interaction term between the stove usage and the non housewife status dummies. In this case, the effect of stove usage on housewives appears negative and significant; housewives in households with an improved stove (with a working chimney) are 18 percentage points less likely to suffer from eye irritation problems than those in households using the traditional stove technology. On the other hand, the effect of stove usage on any other household member109 also appears as statistically significant, but it has the “wrong” sign and it is relatively small in absolute size. Column III in the same table estimates a village fixed effects regression; note that the effect of improved stove usage on housewives remains statistically significant and is slightly higher in absolute size than the effect obtained in column II; while the effect of stove usage on other household members is non significant and close to zero in absolute size. It is important to mention that in the regressions corresponding to columns II and III, the coefficient for the interaction term between the stove usage dummy and the non housewife condition dummy (that is θ3 in equation (1), which is given by the difference between the two group effects) is negative and statistically significant, which indicates that the effect of stove usage on eye discomfort symptoms is significantly higher (in absolute size) on housewives than on any other household member. Column IV estimates a separate regression for housewives; the stove usage effect is very similar to the estimated in columns II and III; although it appears as not statistically significant (probably due to the small sample size). Table 4.3 shows the results for the regressions in which the incidence of respiratory illnesses in the previous 12 months is the dependent variable. As it was the case in Table 4.2, column I in Table 4.3 does not allow for an interaction term between the improved stove usage and the non housewife condition 109  This is equal to the sum of the stove usage dummy coefficient and the coefficient for the interaction term between stove usage and the non housewife status dummies.  101  dummies; as we can see, the effect of stove usage (in this case the average effect on any household member) appears to be negative; although it is not statistically significant. Column II includes an interaction term between the non housewife status dummy and the improved stove usage dummy; as we can observe, the stove usage effect appears only statistically significant for housewives (at the 10% significance level). The third column in Table 4.3 estimates a village fixed effects regression110; note that improved stove effect on housewives is statistically significant at the 1% significance level and relatively higher in absolute size than the effect estimated in column II. The effect of improved stove usage on all other household members also appears to be negative, and it is statistically significant at the 5% significance level. Note that the average effect of stove usage on other household´s members is relatively lower in absolute size than the housewives’ effect; however, the difference between these effects (which is captured by θ3 in equation (1)) is not statistically significant, so I cannot rule out the null hypothesis that the improved stove usage effect on housewives is not significantly different than the effect this device has on any other individual in the household. The small sample size may be one of the reasons why I am not able to find a statistically significant difference between both group specific estimates.  Table 4.3 The effect of improved stove usage with an operative chimney on housewives’ respiratory health I II III IV Household uses an improved stove with an -0.104 -0.074 -0.211** operative chimney: effect on all other household (0.091) (0.096) (0.097) members Household uses an improved stove with an operative chimney: effect on housewives  -0.187* (0.109)  -0.321*** (0.109)  -0.237 (0.167)  Village Fixed Effects  NO  NO  YES  YES  Individual and Household controls R2 Observations  YES 0.07 384  YES 0.08 384  YES 0.23 384  YES 0.37 96  Observations in columns I to IV correspond to a total of 90 households in 19 villages. The dependent variable is the incidence of respiratory illnesses in the last 12 months. All columns control for age, sex, years of education of the adult household member with the highest level of education, per capita number of rooms, per capita value of farm assets, and a dummy variable which takes the value of one if there is a household member 11 years old or younger. In columns I to III the standard errors are clustered at the household level; while in column IV standard errors are clustered at the village level. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  110  Bronchitis and respiratory illnesses have a high degree of incidence at high altitude areas due to the extremely cold weather in the winter season. Adding fixed effects is then important in order to control for this other factor behind the incidence of respiratory illnesses.  102  Finally, column IV in Table 4.3 estimates a separate regression for housewives; note that the estimated coefficient for the stove usage dummy is very similar to the one obtained in columns II and III; however, it is not statistically significant (probably due to the small sample size).  So far the results in Tables 4.2 and 4.3 suggest that housewives in  households that use the improved stove with an operative chimney are less likely to be affected by eye discomfort symptoms and respiratory illnesses than housewives in households that use the open fire traditional stove. In the case of eye irritation problems, improved stove usage appears to have a statistically significant effect only among housewives; while there does not seem to be any statistically significant effect on other household members. The effect of improved stove usage in the respiratory health regression also appears to be significant on household’s members other than housewives; and I am not able reject the null hypothesis that the housewives’ effect is not significantly different than the effect of improved stove usage on any other individual in the household (nevertheless, the point estimate for the effect of stove usage is higher in the case of housewives).  4.3.2. Instrumental Variables The baseline OLS regressions provide some evidence on the negative and significant effect improved stove usage (with an operative chimney) appears to have on the incidence of eye discomfort symptoms and respiratory health among housewives, even after controlling for village fixed effects. However, it is practically impossible to control for every individual and household factor which is correlated at same time with the stove usage decision and housewives’ health outcomes (such as women’s empowerment or unobserved ability). Also, remember that I am excluding from the estimation sample those households which use the improved stove as the main cooking device but do not have an operative chimney (as I want to evaluate the effect of an improved stove with an operative chimney as compared with an “open fire” stove), which introduces sample selection issues in our estimations. Therefore, the results in Tables 4.2 and 4.3 may not be able to fully identify the causal effect of stove usage on adult women’s eye irritation symptoms and respiratory health. In order to address this issue, I follow the same empirical strategy as in Chapter 3. That is, I use self-reported iron frame material  103  problems as an instrument for improved stove usage with an operative chimney. As discussed in Chapter 3, the allocation of poor quality iron frame failure was completely exogenous to household’s and village level characteristics; and experiencing an iron frame problem only has an effect on respiratory health and eye irritation symptoms through the improved stove usage (with an operative chimney) decision..  Table 4.4 The effect of improved stove usage with an operative chimney on housewives’ eye discomfort symptoms: instrumental variables approach I II III IV Household uses an improved stove with an -0.014 0.108* 0.002 operative chimney: effect on all other household (0.042) (0.055) (0.065) members Household uses an improved stove with an -0.287*** -0.379*** -0.491** operative chimney: effect on housewives (0.109) (0.119) (0.201) Village Fixed Effects Individual and Household controls R2 Observations  NO YES 0.20 384  NO YES 0.21 384  YES YES 0.24 384  YES YES 0.21 96  Observations in columns I to IV correspond to a total of 90 households in 19 villages. The dependent variable is the incidence of eye irritation symptoms in the last 12 months. All columns control for age, sex, years of education of the adult household member with the highest level of education, per capita number of rooms, per capita value of farm assets, and a dummy variable which takes the value of one if there is a household member 11 years old or younger. In columns I to III the standard errors are clustered at the household level; while in column IV standard errors are clustered at the village level. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  Table 4.5 The effect of improved stove usage with an operative chimney on housewives’ respiratory health: instrumental variables approach I II III IV Household uses an improved stove with an -0.153 0.132 -0.323** operative chimney: effect on all other household (0.121) (0.136) (0.146) members Household uses an improved stove with an -0.236* -0.429*** -0.387** operative chimney: effect on housewives (0.129) (0.132) (0.174) Village Fixed Effects Individual and Household controls R2 Observations  NO YES 0.07 384  NO YES 0.07 384  YES YES 0.22 384  YES YES 0.36 96  Observations in columns I to IV correspond to a total of 90 households in 19 villages. The dependent variable is the incidence of respiratory illnesses in the last 12 months. All columns control for age, sex, years of education of the adult household member with the highest level of education, per capita number of rooms, per capita value of farm assets, and a dummy variable which takes the value of one if there is a household member 11 years old or younger. In columns I to III the standard errors are clustered at the household level; while in column IV standard errors are clustered at the village level. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  104  Tables 4.4 and 4.5 show the instrumental variable second stage results for the effect of stove usage on self reported eye discomfort and respiratory health respectively111. As we can see, the effect of stove usage on housewives health (respiratory health and eye discomfort) is statistically significant in both tables, and the point estimates for the effect of stove usage are relatively higher than the ones obtained in Tables 4.2 and 4.3. As was the case in the OLS estimations, the effect of stove usage on eye discomfort symptoms is only statistically significant among housewives; while the effect of stove usage on selfreported respiratory health seems to be also statistically significant among other household’s members. In the later case, although the coefficient for the effect of stove usage appears to be higher (in absolute size) for housewives, I cannot reject the null hypothesis that the difference between the two effects is not significantly different from zero112.  4.3.3. Additional Estimations To provide additional support for our hypothesis that the observed effect of stove usage with an operative chimney on housewives’ self-reported respiratory health and eye discomfort is associated to reduced exposure to IAP, in Table 4.6 I only include in the user category those housewives living in households that reported using the improved stove without an operative chimney. If a reduction in IAP is what drives the main results in Sections 4.3.1 and 4.3.2, then there should not be any significant effect of improved stove usage without an operative chimney on housewives’ eye discomfort symptoms or self reported respiratory health. Column I in Table 4.6 (below) estimates an OLS regression for the effect of stove usage (without an operative chimney) on eye discomfort only among housewives; while column II estimates the instrumental variable regression in which I use reporting an iron frame problem as an instrumental variable. Columns III and IV are equivalent to columns I and  111  The 2nd stage regressions for the estimations in Tables 4.5 and 4.6 are shown in Table C.2 in Appendix  C.  112  As discussed in Section 5.2 in Chapter 3, in the presence of heterogeneous treatment effects and if the monotonicity assumption is satisfied, the effects estimated in Tables 4.4 and 4.5 should be interpreted as LATEs, more precisely as the effect of improved stove usage on compliers’ health outcomes.  105  II but focus on housewives’ self-reported respiratory health. As we can observe, the effect of improved stove usage appears as not statistically significant in all columns in Table 4.6. These results support the hypothesis that decreased exposure to IAP is likely to be the key factor behind the main results in the estimations in the previous sections.  Table 4.6 The effect of improved stove usage without an operative chimney on housewives’ eye discomfort and respiratory health I II III IV Household uses an improved stove without an -0.085 -0.078 -0.065 -0.127 operative chimney: Effect on Housewives (0.144) (0.155) (0.101) (0.108) Village Fixed Effects Individual and Household controls R2 Observations  YES YES 0.24 125  YES YES 0.24 125  YES YES 0.27 125  YES YES 0.26 125  The dependent variable in columns I and II is the incidence of eye irritation symptoms in the last 12 months. The dependent variable in columns III and IV is the incidence of respiratory illnesses in the last 12 months. All columns control for age, years of education of the adult household member with the highest level of education, per capita number of rooms, per capita value of farm assets, and a dummy which takes the value of one if there is a household member 11 years old or younger. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels. Standard errors have been clustered at the village level.  The fact that in all columns in Table 4.6 we always observe a negative -although not statistically significant- effect of stove usage without an operative chimney, particularly for the case of respiratory illnesses in column IV, deserves some brief discussion. In first place, I must emphasize that for users without an operative chimney, I do not know the exact time at which the chimney stopped working; if in most cases the chimney got damaged in the very immediate months to the 2008 survey, then observing a negative sign for the stove usage dummy (without an operative chimney) does not necessarily appear as a contradiction to the previous findings. In second place, improved stove usage may have a positive effect on respiratory health by channels other than reductions in IAP; for example, as less firewood is required for cooking and heating, adult women may need to make fewer trips to collect firewood, which in the end may improve women’s general health, and by consequence will also affect respiratory health. In the previous sections, it was shown that improved stove usage only seems to have a significant effect on eye discomfort symptoms among housewives; while in the case of respiratory health, there also seems to be a significant effect on other household´s members. Table 4.7 presents the results for the effect of stove usage on eye discomfort 106  symptoms among other household´s members. Column I shows the OLS estimates for the effect of stove usage only among adult males, while column II shows the instrumental variables regression estimates for the same group of individuals. Columns III and IV are equivalent to columns I and II but focus only on household’s members 14 years old or younger. As we can observe, in all columns in Table 4.7 the effect of stove usage is not statistically significant and very small in absolute size.  Table 4.7 The effect of improved stove usage with an operative chimney on eye discomfort symptoms: adult males and children specific group regressions I II III IV Household uses an improved stove with an operative -0.009 0.065 chimney: Effect on adult males (0.092) (0.121) Household uses an improved stove with an operative chimney: Effect on household members age≤14 Village Fixed Effects Individual and Household controls R2 Observations  YES YES 0.26 139  YES YES 0.25 139  0.009 (0.055)  0.130 (0.124)  YES YES 0.17 133  YES YES 0.13 133  The dependent variable is the incidence of eye irritation symptoms in the last 12 months. All columns control for age, sex, years of education of the adult household member with the highest level of education, per capita number of rooms, per capita value of farm assets, and a dummy which takes the value of one if there is a household member 11 years old or younger. The odd columns present the OLS results and the even columns present the instrumental variable regression results. Columns I and II focus on adult males, while columns III and IV focus on children 14 years old or younger. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels. Standard errors have been clustered at the village level.  Table 4.8 The effect of improved stove usage with an operative chimney on respiratory health: adult males and children specific group regressions I II III IV Household uses an improved stove with an operative -0.146 -0.289 chimney: Effect on adult males (0.108) (0.201) Household uses an improved stove with an operative chimney: Effect on household members age≤14 Village Fixed Effects Individual and Household controls R2 Observations  YES YES 0.31 139  YES YES 0.31 139  -0.268 (0.192)  -0.319 (0.213)  YES YES 0.25 133  YES YES 0.24 133  The dependent variable is the incidence of respiratory illnesses in the last 12 months. All columns control for age, sex, years of education of the adult household member with the highest level of education, per capita number of rooms, per capita value of farm assets, and a dummy which takes the value of one if there is a household member 11 years old or younger. The odd columns present the OLS results and the even columns present the instrumental variable regression results. Columns I and II focus on adult males, while columns III and IV focus on children 14 years old or younger. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels. Standard errors have been clustered at the village level.  107  Table 4.8 is equivalent to Table 4.7, but focus on self-reported respiratory health. As we can see, the effect of improved stove usage with an operative chimney always appears negative and relatively big in size, but it is not statistically significant in all cases. Note that the effect of stove usage on adult males’ respiratory health is relatively lower in size than the effect of stove usage on housewives obtained in column IV in Tables 4.3 and 4.5; however, as it was also the case in the previous sections, I cannot reject the null hypothesis that the difference between the two effects is not significantly different from zero, probably due to the small sample size used in this study. In addition to the effect of improved stove usage with an operative chimney, it is also of relevant interest to analyze how other individual and household level factors included as controls in the regressions are related to respiratory health and eye discomfort symptoms for different groups of household´s members. This analysis is performed in Tables 4.9 and 4.10. Table 4.9 focuses only on housewives; while Table 4.10 focuses only on adult males. Columns I and II in both tables present the estimation results for the controls included in the eye discomfort specific group OLS and instrumental variables regressions respectively (that is column IV in Tables 4.2 and 4.4 in the case of housewives, and columns I and II in Table 4.7 in the case of adult males); while columns III and IV in both tables present the estimation results for the individual and household controls included in the respiratory health specific group OLS and instrumental variable regressions respectively (that is column IV in Tables 4.3 and 4.5 in the case of housewives, and columns I and II in Table 4.8 in the case of adult males). As we can see in columns I and II in Tables 4.9 and 4.10, in the case of housewives as well as in the case of adult males, the age coefficient in the eye irritation regressions appears statistically significant and positive in sign. That is, older individuals are more likely to suffer from eye irritation problems, probably because they tend to spend a relatively higher amount of time inside the household unit. In the case of the respiratory health estimations, the age coefficient always appears as not statistically significant; and it only has a positive sign in the case of housewives (the p-value for the age coefficient in column III in Table 4.9 is equal to 0.105). With respect to household’s education  108  (measured by the years of education of the adult member who attained the highest education level), in general its coefficient appears negative in sign (mainly in the respiratory regressions), however it is not statistically significant113.  Table 4.9 Housewives’ eye discomfort and respiratory health: individual and household level controls I II III IV 0.004 0.010*** 0.008*** 0.005 Age (0.003) (0.003) (0.003) (0.003) -0.029 Years of Education of the Household 0.002 0.004 -0.030 (0.018) Member with the Highest Education (0.019) (0.018) (0.018) 0.001 Household’s per capita wealth (value of -0.000 0.000 0.001 (0.001) farm assets in 2008 Peruvian soles) (0.001) (0.001) (0.001) -0.008 -0.082 -0.099 0.016 Household’s per capita Number of Rooms (0.124) (0.146) (0.149) (0.139) 0.255** Household has a member 11 years old or -0.053 0.019 0.226** (0.107) younger (yes=1) (0.149) (0.155) (0.092) Observations 96 96 96 96 The dependent variable in columns I and II is the incidence of eye irritation symptoms in the last 12 months. The dependent variable in columns III and IV is the incidence of respiratory illnesses in the last 12 months. Columns I and II show the coefficients for the individual and household’s controls corresponding to the OLS and instrumental variables eye discomfort regressions in column IV in Tables 4.2 and 4.4 respectively, while columns III and IV show the coefficients for the individual and household’s controls corresponding to the OLS and instrumental variables respiratory health regressions in column IV in Tables 4.3 and 4.5 respectively. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels. Standard errors have been clustered at the village level.  Table 4.10 Adult males’ eye discomfort and respiratory health: individual and household level controls I II III IV 0.006*** 0.006*** -0.003 -0.003 Age (0.002) (0.002) (0.002) (0.002) Years of Education of the Household -0.007 -0.006 -0.017 -0.013 Member with the Highest Education (0.014) (0.014) (0.010) (0.014) Household’s per capita wealth (value of 0.001** 0.001** 0.001 0.001 farm assets in 2008 Peruvian soles) (0.000) (0.000) 0.001 0.001 Household’s per capita Number of -0.062 -0.062 -0.198* -0.184* Rooms (0.081) (0.079) (0.097) (0.098) Household has a member 11 years old 0.005 -0.015 0.148 0.189 or younger (yes=1) (0.074) (0.080) (0.109) (0.111) Observations 139 139 139 139 The dependent variable in columns I and II is the incidence of eye irritation symptoms in the last 12 months. The dependent variable in columns III and IV is the incidence of respiratory illnesses in the last 12 months. Columns I and II show the coefficients for the individual and household’s controls corresponding to the OLS and instrumental variables eye discomfort regressions in columns I and II in Table 4.7 respectively, while columns III and IV show the coefficients for the individual and household’s controls corresponding to the OLS and instrumental variables respiratory health regressions in columns I and II in Table 4.8 respectively. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels. Standard errors have been clustered at the village level.  113  Similar results are obtained if we include the years of education of the adult female member with the maximum level of education.  109  In the case of the per capita number of rooms, interestingly this variable only has a statistically significant effect in the adult males’ respiratory health regression (columns III and IV in Table 4.10); while the coefficient is not statistically significant and very small in absolute size in the housewives’ respiratory health regression (columns III and IV in Table 4.9). In other words; in the case of housewives, having a more spacious house does not make a significant difference in terms of respiratory health outcomes, probably because they are directly and heavily exposed to IAP during food preparation. On the other hand, adult males in a more spacious dwelling may considerably benefit from reduced exposure to IAP concentrations. Exactly the opposite is observed when we focus on the presence of a child 11 years old or younger114. The coefficient for this variable in the adult males’ respiratory health regression appears as not statistically significant (columns III an IV in Table 4.10); while in the housewives’ respiratory health regression this variable seems to play a significant role (columns III and IV in Table 4.9), probably because when young children are present, adult women have to spend a higher amount of time inside the house and are then more exposed to IAP as well as to contagion episodes.  4.4. Conclusion In the last years, special attention has been given to the diffusion of firewood efficient stoves with a chimney mechanism as a strategy to reduce exposure to indoor IAP in rural areas of developing countries, where biomass such as firewood and coal constitute the main source of cooking and heating energy. Although new evidence suggests that improvements in respiratory health and eye discomfort symptoms are observed during the early stages of improves stove adoption (within the first 6, 12 and 18 months) due to reduced exposure to IAP, to the best of my knowledge there is little empirical evidence on the long term benefits of continuous improved stove usage. This research contributes in such direction, and its results suggest that in the long term, improved stove usage (after 5 years of its distribution) with an operative chimney may have sizeable and statistically significant benefits on respiratory health and eye discomfort symptoms, especially among 114  I consider children 11 years old or younger because these ones are more likely to stay at home than children older than this age. Children older than 11 years old are more likely to be enrolled in secondary education, which in the majority of the cases means that they have to attend a school outside the village; also, when they come back from school, they are more likely to help their parents in farm related activities.  110  housewives. The presence of unobservable factors simultaneously correlated with stove usage decisions and health is the main identification concern in this type of evaluation exercises. In this chapter, I exploit exogenous and ex-ante not observable differences in the stove iron frame material quality to identify the casual impact of stove usage. The evidence presented in this chapter strongly suggests that having the stove chimney in operative conditions is crucial in order to achieve improvements in respiratory health; it is shown that an improved stove without a working chimney does not have any significant effect on the incidence of respiratory related illnesses or eye discomfort symptoms, which confirms that reductions in IAP is indeed what drives our main results. Our results also confirm that individual and household factors affect different groups of household members in very specific and dissimilar ways, for example in households where a child 11 years old or younger is present, housewives are more likely to suffer from respiratory illnesses but not adult males, as in such circumstances housewives probably have to spend a higher amount of time inside the household unit. On the other hand, in households with a higher number of rooms per capita, adult males are less likely to suffer from respiratory illnesses but not adult women. As the last ones are already in direct contact with IAP during cooking tasks, having a bigger or more ventilated house doesn’t seem to make a significant difference among them. One of the main disadvantages in this study is that the health indicators used in the estimations are self reported; however I intend to obtain objective health measures in a future research (such as lung function measured by a spirometry) as well as to consider a relatively larger sample size.  111  5. Conclusion In recent years, the economic literature115 has paid special and increasing attention to the concept of social capital and to its potential benefits for economic development (Knack and Keefer (1997), Narayan et al (1999), Guiso et al (2004), Dasgupta (2005), Francois and Zabojnik (2005), among others). Probably one of the most important “benefits” (social capital may also have negative effects) associated with social capital at the community level is the facilitation of information diffusion (Dasgupta (2005)). In the context of new technology adoption processes in rural communities (e.g. adoption of fertilizers, HYV seeds, telecommunication technologies, etc.), this implies that social capital may have an important role accelerating social learning. However, although a significant variety of issues related to social learning has been explored in great detail in the economic development literature, not enough attention has been given at understanding how village social capital impacts social learning; neither at how the initial performance of a new technology influences the type of effects village social links will have on individual household’s decisions. The second chapter in this thesis aims to help filling this gap and contributes to the literature by providing strong empirical evidence that supports the information diffusion role social capital has played during the adoption process of a new firewood cooking technology in the Northern Peruvian Andes. The main strength of my research in relation to others in the literature is that the social capital indicators used in this thesis were obtained before the introduction and adoption stages of the new stove technology. As a result, reverse causality should not be a major concern. The results in the second chapter clearly suggest that information about a new technology is indeed more effectively diffused in villages with higher levels of trust in local neighbours (which is a potential indicator for the level of bonding social capital which is present within a community), and that the effect of this bonding social capital indicator is closely linked to the initial performance of the new technology. That is, the  115  The sociology and political science literatures have also made important contributions towards a better understating of the social capital concept as well as on its potential benefits on economic development (see for example Coleman 1987, Putman 1993 and 1995, Woolcock 1998)  112  village social capital (informational) effect is more likely to be negative if village initial adoption success rates are relatively low. In order to rule out the possibility that unobservable correlates are the ones driving the results, different household´s decisions associated to the adoption process of the new technology were analyzed: a) the decision to use the stove as the main cooking device; and b) the decision to uninstall the new technology among non users. In all cases, the hypothesis that the village bonding social capital plays a crucial role facilitating information diffusion and social learning was strongly supported. In addition to testing the information diffusion role of social capital in the context of technology adoption, this research also contributes to the social capital literature by providing empirical evidence towards the effective presence of different dimensions of social capital at the village level (bridging and bonding social capital), as well as on the specific roles these different dimensions have in the communal social life. For example, it was shown that only the bonding links indicator (village-level trust in local neighbours) have a multiplier effect on how adoption patterns within the village influence individual decisions, while the bridging social capital indicator (village-level trust in people from other villages) does not play any significant role in this situation. In the other hand, it was also shown that only the bridging social capital indicator influences the way information available in other villages affects individual decisions; while the bonding social capital indicator does not play any role in this case. The research results in the second chapter provide important lessons applicable to improved stove dissemination programs (or any program diffusing new rural technologies). Probably the most critical lesson that can be learned is that bad news about a new technology can have “disastrous” effects during the adoption process; especially in villages with strong levels of within village trust (as in this case bad news will be strongly disseminated). Our estimates suggest that the marginal network effect of village usage patterns with problems is 2.5 higher that the network effect of village usage patterns without problems. These results not only highlight the importance of continuous monitoring and high quality extension services, but also the importance of involving the  113  beneficiary population in the design and implementation stages of the intervention, as well as to keep the expectations about the benefits of the new technology at a realistic level. Considering the adoption of a cooking technology, one might expect that women’s social capital within the community plays a more important role in terms of facilitating social learning relative to men’s social capital. In other words, as this specific technology is mainly operated by the female members in the households; one would expect that in communities in which women have strong bonding links, they will share more intensively with each other their experiences and learning with the new cooking device. Unfortunately, the 2003 survey does not allow us to distinguish between all the male and female respondents of the social capital survey, and then we cannot test this hypothesis in the data. However, one possible extension of this research will be to explore if female social capital is indeed the main social factor facilitating the diffusion of information in the context of other improved stoves dissemination programs, or in the context of any other rural technology. Another possible, and quite interesting, line of research will be to empirically explore how different technology diffusion interventions affect the equilibrium levels of bonding and bridging links in the village. For example, as mentioned before, individuals may try to invest more in their social relationships during technology adoption processes, and this may affect the equilibrium levels of bonding social links. It can also be the case that non adopters (or adopters) of a new technology may suffer different degrees of exclusion from the village social network, which in the end will affect village social capital. Also, a successful (unsuccessful) technology may increase (decrease) the level of trust people in the village have in outsiders, which has a positive (negative) effect on the village levels of bridging/linking social capital. Furthermore, the new technology may have a direct effect over different spheres of the communal social organization. For example, a more efficient technology which significantly decreases rural household’s firewood needs may decrease people’s incentives to participate in local pro-forest organizations and this may have a negative impact on the village levels of social capital. In an alternative way, a  114  more efficient firewood technology may increase the amount of forest environmental benefits (due to reduced forest degradation), and then may encourage the participation in communal pro-forest groups, which may positively affect the communal levels of bonding links. Dissemination programs of improved firewood cooking stoves (with a metallic chimney) have lately started to capture increasing attention from development economists (e.g. Duflo et al 2008b), especially due to their potential benefits on rural household’s welfare (as a result of improved respiratory health and reduced firewood needs); however, the empirical evidence in the literature on the effects of these devices is still limited to a few number of (mostly inconclusive) works (as clearly discussed by Duflo et al (2008a) and Johnson et al (2010)). The main difficulty faced by most of these studies is related to the endogenous nature of the improved stove usage decision, which makes it complicate to obtain clear implications about the environmental and health effects of these devices. Chapters 3 and 4 in this dissertation empirically explored the firewood consumption savings and health benefits of the improved stove design distributed in the Chalaco District. These chapters contribute to the literature firstly by providing a methodological identification example which relies on unique evidence suggesting that some stove beneficiaries received iron frames of lower quality that others, and that this allocation was exogenous to households and village characteristics. Secondly, the evidence in these chapters confirmed the statistically significant impact improved stove usage has in reducing firewood consumption (a clear environmental benefit of this strategy) as well as in improving housewives’ respiratory and eye related health indicators  116  . By  consequence, we expect that improved stove usage will also have a positive impact on household’s welfare due to fewer resources allocated to medical expenses and/or firewood collection, as well as to increased productivity (as a result of improved health). Accordingly, the natural extension of this research will be to test the welfare effects of 116  It was shown that stove usage with an operative chimney has a significant effect on eye discomfort problems only among housewife’s; whereas in the case of respiratory health, although the improved stove effect on housewives seems to be higher than the effect on other household members, we could not rule out the null hypothesis that these effects were significantly different (probably due to the small sample sizes).  115  improved stove usage; for example, to test if women generated income in households with improved stoves is higher or if children school attendance also improves due to better health outcomes. One of the main limitations in Chapters 3 and 4 is that the firewood collection and health indicators were self reported by the individuals. However, in a future research I intend to test the effect of stove usage using more objective measures (indicators), especially for the case of respiratory health. For example we can use lung function responses as measured by a spirometry to test the objective health respiratory benefits of stove usage. Another problem in Chapters 3 and 4 is related to the small sample size in which our estimations are based; however note that we were still able to find statistically significance for most of our coefficients of interest. In the introductory section of this dissertation, it was said that each of the chapters in this thesis constituted an independent research essay; in fact, each chapter exploited different features of the improved stove data and tested its own research hypothesis. However, it is not hard at all to see the connection line between the three essays; taken together they constitute an integral analysis of the improved stove dissemination program in the Chalaco district. Any technology diffusion intervention sets special emphasis on two key aspects: in first place on the adoption process of the technology and the factors that influence this process, and in second place on the evaluation of the real program’s impacts. Chapter 2 in this dissertation focused on the first aspect, and highlighted the importance of social capital and the hazards of poor initial successful adoption rates; while Chapters 3 and 4 evaluated the promised environmental and health related benefits of the new cooking technology. In this sense, this thesis constitutes the first comprehensive study related to the dissemination of improved stoves in the Peruvian Andes, where this strategy has been intensively promoted in the last years117.  117  See for example: http://www.myproworld.org/group_programs/environmental_conservation_peru.htm  116  6. Bibliography AGURTO, MARCOS (2010): “The Role of Social Capital on Improved Stoves Usage Decisions in the Peruvian Andes”, working paper version, Department of Economics, University of British Columbia. __________________ (2004): “Social Capital: A Measurement Exercise”, in Spanish, MIMEO, Programa Académico de Economía, Universidad de Piura. AMACHER, GREGORY et al (1996): “Household Fuelwood Demand and Supply in Nepal’s Tarai and Mid-Hills: Choice between Cash Outlays and Labor Opportunity”, World Development, 24(11), 1725-1736. BANDIERA, O. and RASUL, I. (2006): “Social Networks and Technology Usage in Northern Mozambique”, Economic Journal, 116(514), 869-902. BARRUETA, VICTOR M. et al (2009): “Energy Performance of Word-burning cookstoves in Michoacan, Mexico”, Renewable Energy, 33(5), 859-870. BOY, ERICK et al (2000): “Fuel Efficiency of an Improved Wood-burning Stove in Rural Guatemala: implications for Health, Environment and Development”, Energy for Sustainable Development, 4(2), 23-31. BROCK, W and DURLAUF, S. (2007): “Identification of binary choice models with social interactions”, Journal of Econometrics, 140(1), 52-75. ____________________________ (2003): “Discrete Choice with Social Interactions”. The Review of Economic Studies, 68(2), 235-260.  ____________________________  (2001):  “Interaction  Based  Models”,  Handbook  of  Econometrics 5, James H. Heckman and Edward Leamer, editors. BRUCE, NIGEL et al (2000): “Indoor Air pollution in developing countries: a major environmental and public health challenge”. Bull World Health Organ, 78(9), 1078-1092.  117  CHAPMAN, Robert et al (2005): “Improvement in household stoves and risk of chronic obstructive pulmonary disease in Xuanwei, China: retrospective cohort study”, British Medical Journal, 331(7524):1050. CHEN, LE et al (2006): “Energy Consumption in Rural China: A Household Model for Three villages in Jiangxi Province”, Ecological Economics, 57(2), 407-420. CHENGAPPA, CHAYA et al (2007): “Impact of Improved cookstoves on indoor air pollution on indoor air quality in the Bundelkhand region in India”, Energy for Sustainable Development, 11(2), 33-44. COLEMAN, JAMES (1987): “Families and Schools”, Educational Researcher, 16(6), 32-38. CONLEY, T. and UDRY, C. (2010): “Learning About a New Technology: Pineapple in Ghana”, American Economic Review. 100(1), 35-69. DASGUPTA, PARTHA (2005): “The Economics of Social Capital”, Economic Record, 81(1), 2-21. DIAZ, ESPERANZA et al (2001): “Eye discomfort, headache and back pain among Mayan Guatemalan women taking part in a randomised stove intervention”, Journal of Epidemiology and Community Health, 61(1), 74-79. DUFLO, ESTHER et al (2008b): “Cooking Stoves, Indoor Air Pollution and Respiratory Health in Rural Orissa”. Economic and Political Weekly, 43(32), 71-76. ________________________ (2008a): Indoor Air Pollution, Health and Economic Well-being, MIMEO, MIT. DURLAUF, STEVEN. (2002): “On the Empirics of Social Capital”, Economic Journal, 112(483), 459-479. DURLAUF, S. and FACHAMPS, M. (2004): “Social Capital”, MIMEO, University of Wisconsin at Madison.  118  EDWARDS, RUFUS et al (2007): “Design Considerations for field studies of changes in indoor pollution due to improved stoves”, Energy for Sustainable Development, 11(2), 71-81. EZZATI, M. and KAMMEN, D. (2002): “The Health Impacts of Exposure to Indoor Air Pollution from Solid Fuels in Developing Countries: Knowledge, Gaps and Data Needs”, Environmental Health Perspectives, 110(11), 1057-1068. FEDER, G. and SLADE, R, (1984): “The Acquisition of Information and the Adoption of a New technology”, American Journal of Agricultural Economics, 66(3), 312-320. FOSTER, A and ROSENZWEIG, M. (1995): “Learning by Doing and the Choice of Technology”, Econometrica, 64(6), 1299-1310. FRANCOIS, P. and ZABOJNIK, J. (2005): “Trust, Social Capital and Economic Development”, Journal of the European Economic Association, 3(1), 51-94. GILL, JAS (1985): “Stoves and Deforestation in Developing Countries”, Paper presented to the UK-ISES conference: “Energy for Development – what are the solutions?” held at Reading University, December 13th 1985. GLAESER, EDWARD et al (2000): “Measuring Trust”, Quarterly Journal of Economics, 115(3), 811-846. GUISO, LUIGI et al (2004): “The Role of Social Capital in Financial Development”, American Economic Review, 94(3), 526-556 GRADOS, NORA et al (2004): “Estudio Socioeconómico del Distrito de Chalaco”, MIMEO, Universidad de Piura. HECKMAN, JAMES et al (2006): “Understanding Instrumental Variables in Models with Essential Heterogeneity”, The Review of Economics and Statistics, 58(3), 389-432.  119  HELTBERG, RASMUS et al (2000): “Fuelwood Consumption and Forest Degradation: A Household Model for Energy Domestic Substitution in Rural India”, Land Economics, 76(2), 213-232. HILDELMANN, LYNN et al (2009): “Indoor Air Pollution and Household Cookstove Choice in Bangladesh”, working paper presented in the 2009 NEUDC conference celebrated at Tufts University. IMBENS, G. And WOOLDRIDGE, J. (2009): “A Course in Panel Data Econometrics”. Lectures notes prepared for a seminar organized by the Canadian Economics Association and the Canadian Labour Market and Skill Researcher Network at Ryerson University Conference Center, Toronto. ISHAM, JONATHAN (2002): “The Effect of Social Capital on Fertiliser Usage: Evidence from Rural Tanzania”, Journal of African Economics, 11, 39-60. JOHNSON, MICHAEL et al (2010): “New Approaches to Performance Testing of Improved Cookstoves”, Environmental Science and Technology, 44(1), 368-374. JOVANOVIC, B. and NYARKO, Y. (1996): “Learning by Doing and the Choice of Technology”, Econometrica, 64(6), 1299-1310. KARLAN, DEAN (2005): “Using Experimental Economics to Measure Social capital and Predict Financial Decisions”, American Economic Review, 95(5), 1688-1699. KNACK, S. and KEEFER, P (1997): “Does Social Capital have an Economic Payoff? A Cross Country Investigation”, Quarterly Journal of Economics, 112(4), 1251-1288. LA PORTA, RAFAEL et al (1997): “Trust in Local Organizations”, American Economic Review, 87(2), 333-338. MASERA, OMAR et al (2007): “Impact of improved cookstoves on indoor quality in Michoacan Mexico”. Energy for Sustainable Development, 11(2), 45-56.  120  MANSKI, CHARLES (1993): “Identification of Social Effects: The Reflection Problem”, Review of Economic Studies, 60(3), 531-542. MCCRACKEN, JOHN et al (1998): “Emissions and Efficiency of Improved Wood-burning Cookstoves in Highland Guatemala”, Environment International, 24(7), 739-747. NARAYAN, D. et al (1999): “Cents and Sociability: Household Income and Social Capital in Rural Tanzania”, Economic Development and Cultural Change, 47(4):871–897. PENNISE, DAVID et al (2009): “Indoor Air quality impacts of an improved stove in Ghana and an ethanol stove in Ethiopia”, Energy for Sustainable Development, 13, 71-76. PUTNAM, ROBERT (1995): “Bowling Alone, America’s Declining Social Capital”, Journal of Democracy, 6(1), 65-78 SMITH-SIVERTSEN, TONE et al (2009): “Effect of Reducing Indoor Air Pollution on Women’s Respiratory Symptoms and Lung Function: The RESPIRE Randomized Trial, Guatemala”, American Journal of Epidemiology, 170, 211-220. URDAY, GONZALO (2006): “Estado del Sistema de Cocinas de Bajo Uso de Lena”, reporte interno MIRHASPERU. URETA, IVAN (2007): “Experiencias y Lecciones de Desarrollo Rural: Sistematización del Programa de Desarrollo Sostenible de Ecosistemas de Montana del Perú – Programa Chalaco”, Free Online Resource, Full text available at www.eumed.net/libros/2007a/228/ WANG, SHUN (2009): “Social capital and Rotating Labor Associations: An Instrumental Variable Approach”, working paper, Department of Economics, UBC. WALLMO, KRISTY et al (1998): “A social and environmental evaluation of fuel-efficient cook-stoves and conservation in Uganda”, Environmental Conservation, 25(2): 99-108. WOOLCOCK, MICHAEL (1998): “Social capital and economic development: Toward a Theoretical Synthesis and Policy Framework”, Theory and Society, 27(2), 151-208.  121  Appendix A: Figures and Maps Figure 1: the improved firewood original design  source: MIRHASPERU  Figure 2: traditional firewood cooking technology  Source MIRHAPSERU  122  Figure 3: An improved firewood stove as observed in the summer 2008  Own source  Figure 4: The Chalaco District in the Piura Region  Source: SIG Universidad de Piura  123  Figure 5: The Chalaco District*,**  Source: SIG Universidad de Piura *The black dotes in the graph represent all the villages within the Chalaco District **The red lines represent the main access routes  124  Appendix B: Local Identification in the Non-Linear in Means Models As suggested by Brock and Durlauf (2001b), the differences between the binary choice and linear-in-means models suggest that nonlinearity has a fundamental effect on the identification problem due to the presence of the reflection problem (the issue of identification in the presence of nonlinearity has also been discussed by McManus, 1992). In this section I follow, almost literally, Section “v” in the chapter “Interaction Based Models” by Brock and Durlauf (2001b) published in the Handbook of Econometrics 5, edited by James H. Heckman and Edward Leamer. In this chapter the authors show that it is possible to demonstrate a basic role for nonlinearity in identifying the parameters of interaction based models, by examining deviations from the linear-in-means model. Following Brock and Durlauf (2001b) suppose that the household´s behavioral equation (again consider only beneficiaries) for the improved stove usage decision (represented by wij in equation (1)) is given by:  wi = k + c´ X i+ d ´ X n(i ) + Jwne(i ) + ei  (1)  As the objective here is to focus on the role played by non-linearity and to make the analysis as simple as possible, I assume that all households belong to the same social e  group (i.e. to the same community). The term wn (i ) gives us the expected proportion of improved stove effective users (adopters); the vector X n (i ) represents contextual effects which are averages of the individual household controls given by X i . In the spirit of McManus (1992) the objective of this section is to make precise the idea that for the class e  e  of models, when the model is not linear in wn (i ) but rather is linear in a function of wn (i ) , lack of identification is pathological.  125  2  To do this, first let g ( w) be a C function such that g is non linear in w and let  G ( wne (i ) ) = wne(i ) + ξg ( wne (i ) )  (2)  e  represent a class of functions which are perturbations around the linear function wn (i ) . Brock and Durlauf consider the following nonlinear-in-means model:  wi = k + c´ X i+ d ´ X n(i ) + JG ( wne(i ) ) + ei  (3)  Associated with this equation is a conditional mean function H:  H ( X i, X n(i ) , wne(i ) ) = k + c´ X i+ d ´ X n(i ) + JG ( wne(i ) ) (4) e  For this model, self-consistency of wn (ij ) requires:  wne (i ) = wn(i ) = k + (c´+ d ´) X n(i ) + JG ( wn(i ) ) (5) The goal in this section is to determine whether the model with  ξ = 0 is special in terms  of nonidentifiability of the parameters in (1). In doing so, Brock and Durlauf assume that when there are multiple solutions to this equation, there is a selection rule which selects a particular solution wn (i ) so that the observed wn (i ) = w( X n (i ) ).  In analyzing this equation the authors work with a notion of local identification. The model equations (2)-(5) define a “structure” for each particular parameter vector A = (k, c, d, J). Brock and Durlauf focus here on identification at the level of the conditional mean function (4). Following Rothernberg (1977) or McManus (1992), the authors state that a parameter point A0 is locally identified if it fulfills the following definition. In their 126  context, this condition is equivalent to requiring that the gradient vector of (4) with respect to A has full rank.  Definition: Local identification in the nonlinear-in-means model with interactions and self-consistent beliefs For the model described by equations (2) to (5), the parameter vector A0 is locally identified if there exists an open neighbourhood N A of A0 such that no other parameter 0 vector in N A gives the same conditional mean in equation (4) and such that the self0 consistency condition equation (5) holds as well. The authors also state that the concept of local identifiability has value as argued in Rothenberg (1971) pg. 578: “It is natural to consider the concept of local identification. This occurs when there may be a number of observationally equivalent structures but they are isolated from each other”  Rothernberg (1971) demonstrates that there is a close connection between local identification and the full rank assumption of particular derivative matrices of a likelihood function. In our context, this means that one must show the gradient of the conditional mean function with respect to A is of full rank. In addition, we need to account for the self-consistency condition in the sense that the full rank condition must hold when the gradient is evaluated at a solution w(i ) to the self-consistency condition. The following theorem has been verified by Brock and Durlauf (2001b):  Theorem: Local Identifiability for models in a neighbourhood of the linear-in-means model: Assume:  127  i) supp( X n (i ) ) is not contained in a proper linear subspace of Rr ii) There exist at least one open neighbourhood n0 such that conditional on X n0 , X i is not contained in a proper linear subspace of Rr. iii) J ≠ 1  {  }  iv)The population data X i, X n (i ) , wn (i ) is such that there is an open set O such that  m( X n(i ) ) is differentiable on O and non constant on O. Further there are two distinct values in O, call them X 1 and X 2 , such that m1 = m1 ( X 1 ) ≠ m2 = m2 ( X 2 ) and that  dg (m1 ) dg (m2 ) ≠ dm dm Then, there exist an open neighbourhood N of  ξ = 0 such that ∀ξ ∈ N − {0}, the model  defined by equations (2)-(5) is locally identified. In the authors’ opinion, what it is important about this theorem is that it highlights the importance of linearity in generating nonidentification. For a permutation of the linear in means model in the direction of any non linear function g, identification will hold. As nonlinearity seems to be a very standard feature of models with interactions, in the authors opinion this result provides a relatively optimistic perspective on the identification problem, at least for correctly specified models.  128  Appendix C: Additional Tables  Table C.1 Village level determinants of the household’s decision to use the improved stove as the main cooking device (including additional village controls) I II III IV 0.0298 *** Village total proportion of users (0.0071) -0.0004*** Village total proportion of users^2 (0.0001) Village proportion of users without 0.0179 *** 0.0191*** problems (0.0051) (0.0059) Village proportion of users without -0.0003*** -0.0003*** problems^2 (0.0001) (0.0001) Village proportion of users with 0.0113 0.0054 problems (0.0124) (0.0134) Village proportion of users with -0.0002 -0.0002 problems^2 (0.0003) (0.0004) Village level of trust in local -0.0255 -0.0670 -0.0386 -0.0802 neighbours (bonding social capital) (0.0999) (0.0970) (0.0886) (0.1035) 0.0023 0.0077*** 0.0059*** 0.0067*** Village proportion of beneficiaries (0.0018) (0.0018) (0.0019) (0.0022) -0.1362 0.0660 0.0426 0.0375 Village level of trust in strangers (0.1673) (0.2059) (0.2093) (0.2079) Village Altitude Village Road Access N  -0.0473**  -0.0418**  -0.0502**  (0.0132)  (0.0203)  (0.0164)  (0.0224)  0.1459*  0.1358**  0.1116*  0.1197  (0.0729)  (0.0637)  (0.0577)  (0.0732)  283  283  283  283 24 0.23  -0.0346**  Villages  24  24  24  R2  0.23  0.23  0.21  All regressions in this table control for watershed dummies and include as household level controls the household’s head sex and age, household’s head level of education, household’s number of adults, presence of a female adult member in the household, household’s wealth (measured by the value of farm assets), farm size, household’s participation in women and environmental organizations, household’s elaboration of processed products and usage of fertilizer and household’s participation in local activities in the previous 12 months. As it is standard ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  129  Table C.2 Instrumental variables: first stage regressions for the effect of improved stove usage with an operative chimney on housewives’ eye discomfort symptoms and respiratory health. I II III IV First stage regression A. Dependent variable: household uses an improved stove Instrument: Household reported an iron frame failure  -0.679*** (0.032)  -0.760*** (0.064)  -0.594*** (0.038)  -0.677*** (0.072)  First stage regression B. Dependent variable: (household uses an improved stove)*(household member is not a housewife) Instrument: (Household reported an iron frame failure)*(household member is not a housewife)  -0.693*** (0.067)  -0.665*** (0.038)  The results in column I correspond to the instrumental variable regression in column I in Tables 4.4 and 4.5. The results in column II correspond to the instrumental variable regression in column II in Tables 4.4 and 4.5. The results in column III correspond to the instrumental variable regression in column III in Tables 4.4 and 4.5. The results in column IV correspond to the instrumental variable regression in column IV in Tables 4.4 and 4.5. Observations in columns I to IV correspond to a total of 382 individuals in 90 households in 19 villages. All columns control for age, sex, years of education of the adult household member with the highest level of education, per capita number of rooms, per capita value of farm assets, and a dummy variable which takes the value of one if there is a household member 11 years old or younger. In columns I to III the standard errors are clustered at the household level; while in column IV standard errors are clustered at the village level. ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  130  Table C.3 Main characteristics for households reporting and not reporting iron frame material problems: 2008 survey I II III Did not reported Reported material p-value material problems problems 49.3 54.4 Household head’s age 0.09 (14.35) (13.40) Household head’s sex 0.93 0.79 0.04 6.47 5.17 Household head’s years of education 0.06 (2.85) (3.56) Years of education of the adult female member 6.00 5.40 0.49 with the highest level of education (4.26) (3.8) 4.08 4.50 Household’s size 0.32 (1.82) (2.12) 0.93 0.94 Per capita number of rooms 0.94 (0.51) (0.59) Household’s wealth (value of farm assets in 228.4 208.4 0.72 2008 Peruvian soles) (306.4) (214.8) 2.56 1.98 Household’s farm size in has. 0.33 (3.53) (1.77) Household uses fertilizer 0.58 0.69 0.30 Household belongs to the local “tree nursery” 0.54 0.45 0.40 15.61 17.35 Household’s village altitude (in hundred meters) 0.01 (3.07) (3.10) Observations 42 48 As I intend to compare the effect of improved stove usage with an operative chimney as compare to the traditional open fire stove, only households with an operative chimney are included in the user group. Standard deviations shown in parenthesis. Column III presents the p-value for the difference in raw means test.  131  Table C.4 Dependent variable: reporting an iron frame material problems. OLS multiple regression results  Household head’s age Household head’s sex Household head’s years of education Years of education of the adult female member with the highest level of education Household’s size Per capita number of rooms Household’s wealth (value of farm assets in 2008 Peruvian soles) Household’s farm size in has. Household uses fertilizer Household belongs to the local “tree nursery” Household’s village altitude (in hundred meters) Observations  0.003 (0.005) -0.197 (0.177) -0.017 (0.021) 0.014 (0.017) 0.004 (0.004) 0.097 (0134) 0.000 (0.000) -0.014 (0.012) 0.068 (0.117) -0.069 (0.114) 0.031 (0.02) 90  The F-statistic for the model is equal to 1.38 ***, ** and * indicate statistical significance at the 1%, 5% and 10% significance levels.  132  

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0071618/manifest

Comment

Related Items