Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Do fathers care? Measuring mothers’ and fathers’ perceptions of fathers’ involvement in caring for young… Mercer, Gareth D. 2015

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

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata


24-ubc_2017_may_mercer_gareth.pdf [ 5.18MB ]
JSON: 24-1.0166376.json
JSON-LD: 24-1.0166376-ld.json
RDF/XML (Pretty): 24-1.0166376-rdf.xml
RDF/JSON: 24-1.0166376-rdf.json
Turtle: 24-1.0166376-turtle.txt
N-Triples: 24-1.0166376-rdf-ntriples.txt
Original Record: 24-1.0166376-source.json
Full Text

Full Text

Do fathers care? Measuring mothers’ andfathers’ perceptions of fathers’ involvementin caring for young children in South AfricabyGareth D. MercerB.Sc., The University of British Columbia, 2008A THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHE REQUIREMENTS FOR THE DEGREE OFDOCTOR OF PHILOSOPHYinThe Faculty of Graduate and Postdoctoral Studies(Population and Public Health)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)July 2015© Gareth D. Mercer 2015AbstractFathers can be an important source of support for children. However, in South Africa, many children donot reside with their biological father and little is known about fathers’ involvement in children’s care. Aquestionnaire that reliably measures fathers’ involvement and is adaptable to varied residential arrange-ments would facilitate future population-level research. We explored whether children who reside withtheir biological father have better health than children whose fathers live elsewhere. We also assessedwhether a questionnaire adapted from surveys in the United States would reliably measure South Africanfathers’ involvement in caring for infants. With data from the 1998 Demographic and Health Survey,we used multilevel logistic regression to estimate associations between father-child co-residence statusand four child health outcomes: breastfeeding for ≥6 months; immunization completeness; recent acuterespiratory infection; and recent diarrhea. We found that children with non-co-resident fathers were notat higher risk of these health outcomes. As part of a separate longitudinal cohort study in the WesternCape, we had a sample of mothers complete questionnaires about their infants’ fathers’ care involve-ment when infants were 2 weeks, 16 weeks and 6 months old. Using Item Response Theory models weestimated the distribution of the fathers’ levels of involvement in five hypothetically distinct modes ofcare. We used total information functions to assess the precision of father involvement estimates. Mostfathers were reportedly spending time with infants, doing routine care activities and providing finan-cially. Fewer fathers were involved in important care decisions or doing household chores. For mostfathers in the sample, the questionnaire gave precise estimates of involvement in three modes of care:Accessibility, Direct Caregiving, and Practical Support for Mother. In contrast, items measuring Mate-rial Provisioning and Responsibility gave imprecise estimates for the majority of fathers. Our findingsreinforce existing evidence that co-residence status is an inadequate proxy for care involvement. Futurepopulation-level research into fathers’ influences on children’s health should directly measure fathers’care practices. With further validation, the questionnaire assessed in this study could be used to measurethe more direct modes of infant care.iiPrefaceI had primary responsibility for conceiving of the project and research questions; designing, conducting,and interpreting the analyses; and writing the dissertation.Together with my research supervisor and the other investigators, I contributed to writing the proto-col for the Mother Infant Health Study. Data collection for the study (including the Fathering Sub-study)was done by local research assistants, with training and oversight given jointly by me and a professionalclinical research co-ordinator. Data collection was also directly overseen by the local principal inves-tigators (Mark Cotton and Monika Esser) and less directly by the UBC principal investigators (DavidSpeert, Tobias Kollmann and Julie Bettinger). I had primary responsibility for the following additionalaspects of the Fathering Sub-study: writing the protocol, adapting and revising the fathering question-naire, and assessing the quality of data collection. I also secured Canadian Institutes of Health Researchfunding for, and organized, a feedback meeting following the study.All steps of the research project were completed with input from and under the guidance of membersof my research supervisory committee.The Mother Infant Health Study (including the Fathering Sub-Study) protocol was approved bythe human research ethics committees at the Children’s and Women’s Health Centre of BC (Projecttitle: “The Mother Infant Health Study”, Certificate number: H12-01181) and Stellenbosch University(Certificate number: S12/01/009).iiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62.1 Theoretical basis for paternal influence on child health . . . . . . . . . . . . . . . . . 72.2 Empirical evidence for paternal influence on child health . . . . . . . . . . . . . . . . 143 Methods: Research Objective 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253.1 1998 South African Demographic and Health Survey . . . . . . . . . . . . . . . . . . 253.2 1996 South African census . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293.3 Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303.4 Inclusion and exclusion criteria and data linkage . . . . . . . . . . . . . . . . . . . . 403.5 Statistical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424 Results: Effects of Father’s Co-residence Status on Child Health Outcomes . . . . . . . 494.1 Description of the analytic sub-sample . . . . . . . . . . . . . . . . . . . . . . . . . . 494.2 Findings from multilevel regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . 574.3 Explaining variation in neighbourhood coefficients . . . . . . . . . . . . . . . . . . . 685 Discussion: Effects of Father’s Co-residence Status on Child Health Outcomes . . . . . 745.1 Do children who reside with their biological father tend to have better health outcomes? 755.2 Are the father’s co-residence status – child health outcome associations modified. . . . 75ivTable of Contents5.3 What is the magnitude of neighbourhood-level variation. . . . . . . . . . . . . . . . . . 775.4 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 795.5 Next steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 816 Methods: Research Objective 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 826.1 Study design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 826.2 Fathering questionnaire . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 846.3 Statistical methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 867 Results: Measuring Fathers’ Contributions to Infants’ Care . . . . . . . . . . . . . . . 987.1 Description of the MIHS fathering sub-study cohort . . . . . . . . . . . . . . . . . . 987.2 Description of fathers’ parenting practices . . . . . . . . . . . . . . . . . . . . . . . . 1037.3 Measurement properties of the fathering questionnaire . . . . . . . . . . . . . . . . . 1297.4 Agreement between father’s and mother’s reports of father’s parenting . . . . . . . . . 1548 Discussion: Measuring Fathers’ Contributions to Infants’ Care . . . . . . . . . . . . . 1698.1 Measurement properties of the fathering questionnaire . . . . . . . . . . . . . . . . . 1698.2 The potential to use mothers as proxy respondents in research on fathering . . . . . . 1798.3 Summary of contributions to South African fathering research . . . . . . . . . . . . . 1829 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1839.1 Overall summary of findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1839.2 Fathering practices may explain lack of association between co-residence status andchild health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1849.3 Implications and recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . 1869.4 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190AppendicesA Research Objective 1 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207A.1 Detailed description of analytic variables . . . . . . . . . . . . . . . . . . . . . . . . 207A.2 Kaplan-Meier curves for time-to-breastfeeding cessation . . . . . . . . . . . . . . . . 224A.3 Procedure for linking child and household member datasets from the 1998 SADHS . . 225A.4 Complete results for regressions of child health outcomes on father’s co-residence status 230A.5 Complete descriptive statistics and regression findings for extended household structureanalyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243vTable of ContentsB Research Objective 2 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254B.1 MIHS Fathering Sub-Study: Invitation letter for fathers . . . . . . . . . . . . . . . . . 254B.2 MIHS Fathering Sub-study: Informed consent script for fathers . . . . . . . . . . . . 258B.3 Items included in the fathering questionnaire . . . . . . . . . . . . . . . . . . . . . . 260B.4 MIHS Fathering Sub-study: Fathering questionnaire for father, 2-week visit . . . . . . 265B.5 Reasons for father’s non-participation in the MIHS fathering sub-study . . . . . . . . 275B.6 Description of cohort with complete follow-up to 6 months . . . . . . . . . . . . . . . 276B.7 Additional IRT findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280B.8 Results of sensitivity analysis comparing fathering practices of co-resident. . . . . . . . 287viList of Tables2.1 Terms used to search the Medline database for studies of the child-health effects ofresiding with a father . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.2 Terms used to search the ESBCO databases for studies of the child-health effects offather’s parenting practices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183.1 South African EPI schedule (1995-1999) and age groups used to derive immunizationcompleteness outcome variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324.1 Descriptive statistics for outcomes and potential confounding variables in the completesample of children and separately for strata of children with and without co-residentfathers, 1998 SADHS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504.2 Odds ratios (ORs) and 95% Credible Intervals (CIs) estimated from multilevel logisticregressions for four child health outcomes; children aged 0-4 years in the 1998 SADHS. 594.3 Adjusted Odds Ratios (ORs) and 95% Credible Intervals (CIs) for interactions betweenfather’s co-residence status and mother’s marital status estimated using multilevel lo-gistic regressions with intercepts varying by neighbourhood. Counts and population-weighted percentages in each group are also shown; children aged 0-4 years, 1998SADHS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644.4 Population-weighted descriptive statistics comparing children with co-resident fathersto those with non-co-resident fathers, the latter stratified by whether they reside withother adult relatives; children aged 0-4 years, 1998 SADHS. . . . . . . . . . . . . . . 664.5 Odds Ratios (ORs) and 95% Credible Intervals (CIs) comparing children with differentcombinations of co-resident relatives estimated by ordinary logistic regressions (unad-justed) or multilevel logistic regressions with neighbourhood-level varying intercepts(adjusted); children aged 0-4 years, 1998 SADHS. . . . . . . . . . . . . . . . . . . . . 674.6 Medians (M) and InterQuartile Ranges (IQR) for neighbourhood contextual variablesderived from 1996 SA Census data. Statistics are presented for the complete analyticsample of children from the 1998 SADHS and separately for strata of children with andwithout co-resident fathers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694.7 Odds Ratios (95% credible intervals) for neighbourhood covariates in models for neigh-bourhood varying intercepts and co-resident father slopes estimated using multilevellogistic regression. Standard Deviations (95% credible intervals) for the varying coeffi-cients are also shown. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71viiList of Tables6.1 Definition of terms commonly used in Item Response Theory . . . . . . . . . . . . . . 897.1 Descriptive statistics for complete cohort including a comparison by father’s participa-tion status, maternal report at enrolment and 2-week visit: MIHS Fathering Sub-study,2012-13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1017.2 Gamma statistics and results of Pearson’s chi-square tests for cross-tabulations of itemsmeasuring father’s Direct Caregiving with covariates, maternal report at 2-week visit;MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . 1137.3 Gamma statistics and results of Pearson’s chi-square tests for cross-tabulations of itemsmeasuring father’s Accessibility with covariates, maternal report at 2-week visit; MIHSFathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1157.4 Gamma statistics and results of Pearson’s chi-square tests for cross-tabulations of itemsmeasuring father’s Responsibility and Material Provisioning with covariates, maternalreport at 2-week visit; MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . 1177.5 Gamma statistics and results of Pearson’s chi-square tests for cross-tabulations of itemsmeasuring father’s Practical Support for Mother with covariates, maternal report at2-week visit; MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . 1197.6 Item slope and threshold estimates (standard errors) from an IRT model for father’s Di-rect Caregiving, maternal report data, 2-week visit: MIHS Fathering Sub-study, 2012-13.1307.7 Item slope and threshold estimates (standard errors) from an IRT model for father’sAccessibility, maternal report data, 2-week visit: MIHS Fathering Sub-study, 2012-13. 1347.8 Item slope and threshold estimates (standard errors) from an IRT model for father’sResponsibility, maternal report data, 2-week visit: MIHS Fathering Sub-study, 2012-13. 1377.9 Item slope and threshold estimates (standard errors) from an IRT model for father’sMaterial Provisioning, maternal report data, 2-week visit: MIHS Fathering Sub-study,2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1407.10 Item slope and threshold estimates (standard errors) from an IRT model for father’sPractical Support for Mother, maternal report data, 2-week visit: MIHS FatheringSub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1427.11 Correlations (standard errors) among latent modes of paternal influence, maternal reportdata, 2-week visit (n=178); MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . 1457.12 Regression coefficients (95% confidence intervals) from MIMIC models with father’sco-residence status as the independent variable, including tests of differential item func-tioning (DIF) for two items, maternal report at 2-week study visit: MIHS FatheringSub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1507.13 Estimated proportion of overall agreement (OA), proportions of specific agreement,Bhapkar and Bias test p-values, thresholds with equality test p-values <0.05 and meanscore (standard deviation) based on father (F) and mother (M) responses, Direct Care-giving items (plus four-category Accessibility and Responsibility items, where indi-cated): MIHS Fathering Sub-study, 2013-14 . . . . . . . . . . . . . . . . . . . . . . . 155viiiList of Tables7.14 Estimated proportion of overall agreement (OA), proportions of specific agreement, Mc-Nemar test p-value and mean (standard deviation) based on father and mother responses,Accessibility items: MIHS Fathering Sub-study, 2013-14 . . . . . . . . . . . . . . . . 1587.15 Estimated proportion of overall agreement (OA), proportions of specific agreement,McNemar test p-value, and mean (standard deviation) based on father and mother re-sponses, Responsibility items: MIHS Fathering Sub-study, 2013-14 . . . . . . . . . . 1597.16 Estimated proportion of overall agreement (OA), proportions of specific agreement,McNemar test p-value, and mean (standard deviation) based on father and mother re-sponses, Material Provisioning items: MIHS Fathering Sub-study, 2013-14 . . . . . . 1597.17 Estimated proportion of overall agreement (OA), proportions of specific agreement,Bhapkar and Bias test p-values, thresholds with equality test p-values <0.05 and meanscore (standard deviation) based on father (F) and mother (M) responses, PracticalSupport for Mother items: MIHS Fathering Sub-study, 2013-14 . . . . . . . . . . . . 1607.18 Pearson correlation and descriptive statistics for item total scores (mother and fatherreport) by mode of paternal influence . . . . . . . . . . . . . . . . . . . . . . . . . . . 1617.19 Estimated proportion of overall agreement (OA), lowest proportion of specific agree-ment (SA), Marginal Homogeneity (MH) and Bias test p-values, and thresholds withequality test p-values <0.05 based on father and mother responses, Father’s sociode-mographics and father-mother relationship characteristics: MIHS Fathering Sub-study,2013-14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1637.20 Estimated proportion of overall agreement (OA), lowest proportion of specific agree-ment (SA), Marginal Homogeneity (MH) and Bias test p-values, and thresholds withequality test p-values <0.05 based on father and mother responses, Father’s health:MIHS Fathering Sub-study, 2013-14 . . . . . . . . . . . . . . . . . . . . . . . . . . . 1647.21 Estimated proportion of overall agreement (OA), proportions of specific agreement(SA), Bhapkar and Bias test p-values, thresholds with equality test p-values <0.05 andmean score (standard deviation) based on father (F) and mother (M) responses, Father-hood beliefs: MIHS Fathering Sub-study, 2013-14 . . . . . . . . . . . . . . . . . . . . 166A.1 Description of variables used for statistical analyses including, where applicable, levelsof categorical variables and derivation from original SADHS variables . . . . . . . . . 208A.2 Estimated odds ratios (ORs) and 95% Credible Intervals (CIs) for having been breastfedfor six months or longer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231A.3 Estimated odds ratios (ORs) and 95% Credible Intervals (CIs) for being completelyimmunized. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234A.4 Estimated odds ratios (ORs) and 95% Credible Intervals (CIs) for having had a recentAcute Respiratory Infection (ARI). . . . . . . . . . . . . . . . . . . . . . . . . . . . 237A.5 Estimated odds ratios (ORs) and 95% Credible Intervals (CIs) for having had recentdiarrhoea. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240ixList of TablesA.6 Population-weighted descriptive statistics comparing children with co-resident fathersto those with non-co-resident fathers, the latter stratified by whether they reside withother adult relatives; children aged 0-4 years, 1998 SADHS.. . . . . . . . . . . . . . . 244A.7 Adjusted odds ratios (OR) and neighbourhood-level varying intercept standard devia-tions (95% credible intervals) estimated using multilevel logistic regressions comparingchildren living with different combinations of adult relatives; children aged 0-4 years,1998 SADHS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250B.1 MIHS Fathering Questionnaire items arranged by the mode of paternal influence theyare hypothesized to measure, along with associated response key on questionnaire andafter recoding for analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260B.2 MIHS questionnaire items assessing fatherhood beliefs, along with associated responsekey on questionnaire and after recoding for analysis . . . . . . . . . . . . . . . . . . . 263B.3 MIHS Fathering Questionnaire response keys and associated response options . . . . . 264B.4 Frequency distribution of reasons for father’s non-participation; MIHS Fathering sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275B.5 Descriptive statistics for cohort with complete follow-up to 6 months (N=109), maternalreport at enrolment and 2-week visit: MIHS Fathering Sub-study, 2012-13 . . . . . . . 276xList of Figures3.1 Map of South Africa showing the locations of rural (dark grey) and urban (white) 1996census Enumeration Areas in each province. . . . . . . . . . . . . . . . . . . . . . . . 274.1 Bar graphs representing the distribution of children in the complete analytic sub-sampleacross mothers (a), households (b) and neighbourhoods (c). . . . . . . . . . . . . . . 557.1 Flowchart of participants’ progress through MIHS Fathering Sub-study . . . . . . . . 997.2 Percentage distribution of father’s frequency of involvement in Direct Caregiving byitem, maternal report at 2-week visit (n=178); MIHS Fathering Sub-study, 2012-13. . . 1057.3 Percentage distribution of father’s Accessibility by item, binary items (a) and ordinalitems (b), maternal report at 2-week visit (n=178); MIHS Fathering Sub-study, 2012-13. 1067.4 Frequency distribution of number of nights per week father spent at the house where theinfant lives, maternal report at 2-week visit (n=178); MIHS Fathering Sub-study, 2012-13.1077.5 Percentage distribution of person with primary Responsibility for child’s care by item(a) and frequency father talked with mother about their child (b), maternal report at2-week visit (n=178); MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . 1087.6 Percentage distribution of person who mainly provided for child’s material needs (MaterialProvisioning) by item, maternal report at 2-week visit (n=178); MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1097.7 Percentage distribution of father’s frequency of giving Practical Support to Motherby item, maternal report at 2-week visit (n=178); MIHS Fathering Sub-study, 2012-13. 1107.8 Percentage distribution of father’s frequency of involvement in Direct Caregiving byitem at study visits at 2 weeks, 16 weeks and 6 months of age, maternal report (n=109);MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . 1217.9 Percentage distribution of father’s Accessibility measured by binary items at study visitsat 2 weeks, 16 weeks and 6 months of age, maternal report (n=109); MIHS FatheringSub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1227.10 Percentage distribution of father’s Accessibility measured by ordinal items at study vis-its at 2 weeks, 16 weeks and 6 months of age, maternal report (n=109); MIHS FatheringSub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1247.11 Percentage distribution of person who mainly provided for infant’s material needs (MaterialProvisioning) by item at study visits at 2 weeks, 16 weeks and 6 months of age, mater-nal report (n=109); MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . 125xiList of Figures7.12 Percentage distribution of father’s frequency of giving Practical Support to Mother byitem at study visits at 2 weeks, 16 weeks and 6 months of age, maternal report (n=109);MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . 1267.13 Percentage distribution of person with primary Responsibility for infant’s care by itemat study visits at 2 weeks, 16 weeks and 6 months of age, maternal report (n=109);MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . 1277.14 Percentage distribution of frequency father talked to mother about the infant (Responsibility)at study visits at 2 weeks, 16 weeks and 6 months of age, maternal report (n=109); MIHSFathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1287.15 Category characteristic curves for items measuring the frequency the father held (a) andsang (b) to the infant since birth in the IRT model for Direct Caregiving. These curvesshow the probability of being in each response category or higher as a function of thefather’s latent level of involvement in direct caregiving. . . . . . . . . . . . . . . . . . 1317.16 Category response curves for items measuring the frequency the father held (a) and sang(b) to the infant since birth in the IRT model for Direct Caregiving. These curves showthe probability of being exactly in each response category as a function of the father’slatent level of direct caregiving. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1317.17 Distribution of father’s location estimates on the latent Direct Caregiving mode of in-fluence, maternal report data, 2-week visit; MIHS Fathering Sub-study, 2012-13. . . . 1327.18 Total information curve for items measuring father’s Direct Caregiving, including par-tial information curves for Held, Sleep, Night, and Sang items, maternal report data,2-week visit; MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . 1337.19 Item characteristic curves for dichotomous items in the IRT model for father’s Accessi-bility. These curves show the probability of being in the “Yes” response category as afunction of the father’s latent level of accessibility. . . . . . . . . . . . . . . . . . . . 1357.20 Total and partial information curves for items measuring father’s Accessibility, maternalreport data, 2-week visit; MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . 1357.21 Distribution of father’s location estimates on the latent Accessibility mode of influence,maternal report data, 2-week visit; MIHS Fathering Sub-study, 2012-13. . . . . . . . . 1367.22 Item characteristic curves for dichotomous items in the IRT model for father’s Respon-sibility. These curves show the probability of being in the “Yes” response category as afunction of the father’s latent level of responsibility. . . . . . . . . . . . . . . . . . . . 1377.23 Category characteristic curves for item measuring the frequency the father talked to themother about the infant since birth in the IRT model for father’s Responsibility. Thesecurves show the probability of being in each response category or higher as a functionof the father’s latent level of responsibility. . . . . . . . . . . . . . . . . . . . . . . . 1387.24 Total and partial information curves for items measuring father’s Responsibility, ma-ternal report data, 2-week visit; MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . 138xiiList of Figures7.25 Distribution of father’s location estimates on the latent Responsibility mode of influ-ence, maternal report data, 2-week visit; MIHS Fathering Sub-study, 2012-13. . . . . . 1397.26 Item characteristic curves for items in the IRT model for father’s Material Provision-ing. These curves show the probability of being in the “Yes” response category as afunction of the father’s latent level of material provisioning. . . . . . . . . . . . . . . . 1407.27 Total and partial information curves for items measuring father’s Material Provision-ing, maternal report data, 2-week visit; MIHS Fathering Sub-study, 2012-13. . . . . . 1417.28 Distribution of fathers’ location estimates on the latent Material Provisioning mode ofinfluence, maternal report data, 2-week visit; MIHS Fathering Sub-study, 2012-13. . . 1417.29 Category characteristic curves for item measuring how often the father the washeddishes or cooking pots since the infant’s birth in the IRT model for father’s Practi-cal Support for Mother. These curves show the probability of being in each responsecategory or higher as a function of the father’s latent level of practical support for mother. 1437.30 Total and partial information curves for items measuring father’s Practical Support forMother, maternal report data, 2-week visit; MIHS Fathering Sub-study, 2012-13. . . . 1437.31 Distribution of father’s location estimates on the latent Practical Support for Mothermode of influence, maternal report data, 2-week visit; MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1447.32 Percentage distribution of father’s frequency of involvement in Direct Caregiving byitem and whether he was living with his infant, maternal report at 2-week visit: MIHSFathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1467.33 Percentage distribution of father’s Accessibility by item and whether he was living withhis infant for binary items (a) and ordinal items (b), maternal report at 2-week visit:MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . 1477.34 Percentage distribution of person with primary Responsibility for infant’s care by itemand whether father and infant were living together, maternal report at 2-week visit:MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . 1497.35 Percentage distribution of frequency father talked with mother about the infant strati-fied by whether he was living with his infant, maternal report at 2-week visit: MIHSFathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1507.36 Percentage distribution of person who provides for the infant’s material needs (MaterialProvisioning) by item and whether father and infant were living together, maternalreport at 2-week visit: MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . 1517.37 Percentage distribution of father’s frequency of giving Practical Support to Mother byitem and whether he was living with his infant, maternal report at 2-week visit: MIHSFathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1527.38 Comparison of marginal distribution of fathers’ and mothers’ responses to belief that itis less important for a father to spend time with his children than it is for him to providefinancially for them: MIHS Fathering Sub-study, 2013-14 . . . . . . . . . . . . . . . . 167xiiiList of Figures7.39 Comparison of marginal distribution of fathers’ and mothers’ responses to belief thatit is difficult for men to express affectionate feelings toward babies: MIHS FatheringSub-study, 2013-14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1677.40 Comparison of marginal distribution of fathers’ and mothers’ responses to belief thatthe father more than the mother should be the one to teach their children ‘right’ from‘wrong’: MIHS Fathering Sub-study, 2013-14 . . . . . . . . . . . . . . . . . . . . . . 1687.41 Comparison of marginal distribution of fathers’ and mothers’ responses to belief that afather should be as heavily involved as the mother in the care of their children: MIHSFathering Sub-study, 2013-14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168A.1 Plots of Kaplan-Meier product limit estimates (and 95% Hall-Wellner bands) of propor-tion of children still being breastfed stratified by father’s co-residence status . . . . . . 224A.2 Conservative, one-step procedure for matching records in the child and household mem-bers datasets of the 1998 SADHS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227A.3 Less conservative, multi-step procedure for matching records in the child and householdmembers datasets of the 1998 SADHS. . . . . . . . . . . . . . . . . . . . . . . . . . . 228B.1 Category characteristic curves (a) and Category Response Curves (b) for Soothed itemin the IRT model for the Direct Caregiving trait. . . . . . . . . . . . . . . . . . . . . . 280B.2 Category characteristic curves (a) and Category Response Curves (b) for Sleep item inthe IRT model for the Direct Caregiving trait. . . . . . . . . . . . . . . . . . . . . . . 281B.3 Category characteristic curves (a) and Category Response Curves (b) for Talked item inthe IRT model for the Direct Caregiving trait. . . . . . . . . . . . . . . . . . . . . . . 281B.4 Category characteristic curves (a) and Category Response Curves (b) for Night item inthe IRT model for the Direct Caregiving trait. . . . . . . . . . . . . . . . . . . . . . . 282B.5 Category characteristic curves (a) and Category Response Curves (b) for Walk item inthe IRT model for the Direct Caregiving trait. . . . . . . . . . . . . . . . . . . . . . . 282B.6 Category characteristic curves (a) and Category Response Curves (b) for Diaper item inthe IRT model for the Direct Caregiving trait. . . . . . . . . . . . . . . . . . . . . . . 283B.7 Category characteristic curves (a) and Category Response Curves (b) for item measuringthe number of days the father spends an hour or more with the infant in an average week(Spent hr.) in the IRT model for the Accessibility trait. . . . . . . . . . . . . . . . . . 283B.8 Category characteristic curves (a) and Category Response Curves (b) for Looked afteralone item in the IRT model for the Accessibility trait. . . . . . . . . . . . . . . . . . 284B.9 Percentage distribution of whether father had primary responsibility for child’s care byitem, maternal report at 2-week visit (n=178); MIHS Fathering Sub-study, 2012-13 . . 285B.10 Percentage distribution of whether father was the primary provider for child’s materialneeds by item, maternal report at 2-week visit (n=178); MIHS Fathering Sub-study,2012-13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286xivList of FiguresB.11 Category characteristic curves (a) and Category Response Curves (b) for Tidying itemin the IRT model for the Practical Support for Mother trait. . . . . . . . . . . . . . . . 287B.12 Category characteristic curves (a) and Category Response Curves (b) for Cooking itemin the IRT model for the Practical Support for Mother trait. . . . . . . . . . . . . . . . 288B.13 Category characteristic curves (a) and Category Response Curves (b) for Laundry itemin the IRT model for the Practical Support for Mother trait. . . . . . . . . . . . . . . . 288B.14 Category characteristic curves (a) and Category Response Curves (b) for Repairs itemin the IRT model for the Practical Support for Mother trait. . . . . . . . . . . . . . . . 289B.15 Percentage distribution of father’s frequency of involvement in Direct Caregiving byitem and whether he was living with his infant, maternal report at 2-week visit: MIHSFathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290B.16 Percentage distribution of father’s Accessibility by item and whether he was living withhis infant for binary items (a) and ordinal items (b), maternal report at 2-week visit:MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . 291B.17 Percentage distribution of person with primary Responsibility for infant’s care by itemand whether father and infant were living together, maternal report at 2-week visit:MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . 292B.18 Percentage distribution of frequency father talked with mother about the infant strati-fied by whether he was living with his infant, maternal report at 2-week visit: MIHSFathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293B.19 Percentage distribution of person who provides for the infant’s material needs (MaterialProvisioning) by item and whether father and infant were living together, maternalreport at 2-week visit: MIHS Fathering Sub-study, 2012-13. . . . . . . . . . . . . . . 294B.20 Percentage distribution of father’s frequency of giving Practical Support to Mother byitem and whether he was living with his infant, maternal report at 2-week visit: MIHSFathering Sub-study, 2012-13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295xvAcknowledgementsI am grateful to my supervisor – Julie Bettinger – and supervisory committee – Rachel Jewkes, YingMacNab, David Patrick, Jane Buxton, and Clyde Hertzman – for their expert guidance and warm en-couragement throughout the course of this project.I am also grateful to all who made the Mother Infant Health Study a reality: Monika Esser, MarkCotton, David Speert and Tobi Kollmann for their oversight, and especially to David for generouslyproviding funding for the MIHS Fathering Sub-study; Amy Slogrove and Moleen Zunza for their majorcontributions to the study design and for their emotional and intellectual support; and the staff membersof the Children’s Infections Diseases Clinical Research Unit (KID-CRU) and the Vaccine EvaluationCentre, with extra thanks to Carlos Moffatt, Marchalaine Hendricks, Kim Marty, and Wenli Zhang fortheir incredible dedication.Heartfelt thanks to Robert Morrell and Elena Moore for their generous mentorship and encourage-ment during my time in Cape Town.Thank you to the participants of the Diemersfontein, Rawdons and Pretoria Hegemonic Masculin-ity workshops, the “Families, Kin and the State in South Africa” workshop, and the “Promises andChallenges of Researching Fathering in South Africa” workshop for the feedback they gave while thisresearch was still in progress.Thank you, finally, to the Peter Wall Institute for Advanced Studies for funding the Mother InfantHealth Study, the Canadian Institutes of Health Research for funding the MIHS Fathering Sub-studydissemination events, and the UBC MD/PhD Program, Canadian Institutes of Health Research andVanier Canada Graduate Scholarships Program for my personal funding.xviThis work is dedicated to Family (in all its rich diversity) . . .. . . to my own family (Rachel, Leon, Mum, Dad, Terri, Reta, and Ian) . . .. . . and to the families who generously took part in the Mother Infant Health Study.xviiChapter 1IntroductionRates of child and infant mortality in South Africa are high relative to those of countries with similarper capita gross domestic product [1]. In addition, South Africa is one of the few countries wherethe under-five mortality rate has declined little over the last decade [2, p. 30]. Various data sourcessuggest that reductions in infant and child mortality observed over the latter half of the 20th centurywere halted and somewhat reversed during the 1990s. The Inter-agency Group for Child MortalityEstimates at the United Nations Children’s Fund estimate that the under-five mortality rate was 60.8 per1000 live births in 1990, rose to a high of 80.3 per 1000 in 2003/4 and had declined to 52.9 per 1000by 2010 [1]. The increase during the 1990s is thought to be the result of increasing HIV prevalenceand deteriorations in health care quality. The decline during the latter half of the 2000s coincides withimplementation of a national programme for preventing mother-to-child transmission of HIV, so is likelypartially attributable to reductions in vertical HIV transmission [2, p. 30][3, p. 62].Approximately one fifth of child deaths occur during the neonatal period (first four weeks of life),one half between one and twelve months of age, and the remainder between one and five years of age [3,p. 64]. The dominant causes of death vary somewhat by age-group but, outside of the neonatal period,HIV/AIDS and respiratory and gastrointestinal infections cause most deaths. National death notificationstatistics for 2007 suggest that intestinal and respiratory infections directly caused as many as one thirdof the deaths of children aged one month to four years old [3, p. 65]. However, these statistics are knownto underestimate the percentage of deaths having HIV/AIDS as an underlying cause. Statistical modelsand demographic surveillance sites, which are more sensitive to the contribution of HIV-infection, sug-gest that this may be the single largest cause of child deaths, accounting for around 35% [2, p. 31][3,p. 66]. Malnutrition is another important underlying cause, with around 60% of children who die inhospital reported to be underweight (greater than two standard deviations below mean weight-for-age)and half reported to be severely malnourished (greater than three standard deviations below the mean)[3, p. 66]. HIV infection and malnutrition compromise children’s immune system function, therebyincreasing their susceptibility to, and the severity of, common infections. In addition, illnesses, par-ticularly infections, increase children’s risk of malnutrition by reducing their dietary intake and/or bydisrupting their physiological nutrient utilization [4, p. 18]. In this way, the majority of child deaths inSouth Africa result from the synergistic effects of malnutrition and infectious illnesses.The relatively high average child mortality rate conceals large inequities in the distribution of childmortality risk across the South African population. The South African Demographic and Health Survey,a national household survey conducted in 1998, generated some of the most detailed data available todescribe these inequities. It documented that the under-five mortality rate was 65% higher in non-urban1Chapter 1. Introductionchildren than in urban children, 200-400% higher among Black children than among children in otherpopulation groups, and 250-300% higher among children whose mothers had completed no educationthan among those whose mothers had completed secondary school or some post-secondary education.There was also a two-fold difference in under-five mortality between provinces with the highest andlowest rates [5][6].There is a clear imperative to reduce child mortality and to increase equity in child health in SouthAfrica. It is likely that multifaceted strategies will be necessary to achieve improvements [7][2]. Somestrategies may focus on improving accessibility and quality of health services. The post-Apartheidgovernment has demonstrated a commitment to improving maternal and child health by investing innew primary-level clinics and by removing user fees for services at primary- and district-level facilities[7]. Nevertheless, there are large discrepancies between well-resourced and poorly-resourced areas ofthe country in the coverage of key maternal and child health interventions (e.g.: proportion of birthsoccurring in a medical facility and proportion of children fully immunized) [8, 2]. The government-funded public health sector is also grossly under-resourced relative to the for-profit private sector, whichdraws a disproportionate percentage of health professionals into providing care to a wealthy minorityof the population [9, p. 1027]. An example of one important health service intervention would be toensure that all HIV-infected pregnant women and their babies are offered vertical HIV transmissionprophylaxis [2, p. 37].However, the existence of long-established, even widening, socioeconomic inequities necessitatesadditional interventions outside of the health care system. Recognizing the central importance of fam-ilies in mediating and moderating the causes of child mortality, some authors have recommended thatinterventions ought to focus on strengthening and supporting families [10]. HIV, as an infection trans-mitted between family members during their most intimate interactions, has highlighted the potentialgains to be achieved by focusing on families, rather than individuals, in efforts to prevent HIV trans-mission [11]. Families are also of central importance for achieving adequate child nutrition and forpreventing and treating common childhood illnesses.Over the past couple of decades, there has been growing awareness internationally of the need to bet-ter integrate men into efforts to promote reproductive- and child-health. Prior to this, most interventionstended to target women [12, 13]. In one sense, this approach was justified because, in many cultures,child birth and caring for young children were traditionally the responsibility of women. However, ithas also been problematic because it ignores the reality that, in most societies, men have more powerthan women in decisions related to family members’ wellbeing [14]. Ignoring the importance of the roleplayed by men may have limited the reach and effectiveness of many of these interventions. In addition,the orientation of reproductive health services towards women tended to reinforce the perception amongmen that utilization of these services should be the responsibility of their female partners [12, 13].This awareness has been part of a wider ideological shift in the field of international developmentaround the necessity to actively involve men in efforts to achieve gender equity. This position wasfirst crystallized in a major way in the plan of action adopted at the 1994 UN-sponsored InternationalConference on Population and Development. The plan of action included explicit objectives to “...en-2Chapter 1. Introductioncourage and enable men to take responsibility for their sexual and reproductive behaviour...” and topromote “...equal participation of women and men in all areas of family and household responsibilities,including family planning, child-rearing and housework....” [14] The goals of this plan of action weretwofold: to improve the health of women, children and men as worthy outcomes in their own right, andto obtain more equal involvement of men in family planning and childcare as paths to achieving genderequitable societies.Achieving the above goals holds clear promise for improving infant and child health in South Africa.For example, as described above, HIV infection is a major contributor to child mortality. Maternal HIVdeath is a also significant risk factor for subsequent child mortality [15]. Including women’s partnersin antenatal HIV testing and counselling could prevent horizontal transmission among couples in whichone parter is infected and the other is not. Research in sub-saharan Africa has also demonstrated thatparticipation of male partners in prevention of mother-to-child transmission services is associated withlower risk of HIV transmission to children [16]. Furthermore, dominant gender norms in South Africahave been implicated as root causes in some of the most pressing public health concerns in the country.By legitimizing men’s physical and sexual dominance over women, these patriarchal norms promotehigh rates of HIV transmission [17] and contribute to the epidemic-levels of rape and intimate partnerviolence perpetrated by South African men [18, 19]. Finally, related ideals of masculine conduct canprevent men from accessing necessary health services, including HIV testing and treatment [17]. There-fore, transforming inequitable gender attitudes has immense potential to improve the health of women,men and children.In Africa, much of the research that has attempted to understand and promote men’s equal involve-ment in family responsibilities has focused on sexual and reproductive health, particularly HIV testing.In contrast, very little is known about men’s involvement in caring for children [20]. This is despite thefact that there is now considerable evidence from research in North America and Europe showing thatgreater involvement of fathers in caregiving improves children’s psychological, behavioural, and socialoutcomes [21]. Another limitation of existing research is that much of it focuses narrowly on men’sproblematic behaviours, such as their infidelity and sexual risk behaviours. This has contributed to astereotypical perception of men as “obstacles to health instead of partners in promoting family health.”[12, 13, p. 5] Examining the positive ways that men are involved in family life can help to challengenegative stereotypes and inform interventions that seek to engage men as agents of positive change [22].With these points in mind, the general aim of this study is to explore fathers’ contributions to thecare and wellbeing of young children in South Africa, and to identify factors which support and impedethese contributions. More specifically, we aim to evaluate the utility of two different approaches formeasuring biological fathers’ involvement in their children’s lives. Our research objectives and specificresearch questions are as follows:1. To determine whether children whose biological fathers reside with them tend to have better healthoutcomes than children whose biological fathers do not reside with them.(a) Among children under 5 years of age in South Africa, do those living with two biological3Chapter 1. Introductionparents compared to those living with a mother but not a father tend to be more likely to havebeen breastfed for at least six months or to have received all age-appropriate immunizations?(b) Among children under 5 years of age in South Africa, do those living with two biologicalparents compared to those living with a mother but not a father tend to be less likely to havehad a recent acute respiratory infection or diarrhoeal illness?(c) Does mother’s marital status or having additional adult family members living in the house-hold modify the association between living with two biological parents and experiencingbreastfeeding, immunizations, respiratory infection or diarrhoeal illness?(d) To what extend does the association between having a co-resident biological father and eachof the above child health outcomes vary across children living in different neighbourhoods?(e) Do neighbourhood levels of gender equality in educational levels, unemployment rate, andconcentration of low-income households modify the association between having a co-residentfather and each of the above child health outcomes?2. To evaluate a questionnaire designed to measure fathers’ involvement in caring for infants inSouth Africa.(a) In what ways and to what extent are a sample of South African fathers involved in caringfor their children at 2-weeks, 16-weeks, and 6-months after birth? What characteristics aremost strongly associated with father’s levels of involvement?(b) Does a questionnaire adapted from research on US fathers provide reliable and valid mea-sures of South African fathers’ involvement in five distinct modes of child care?(c) Do co-resident fathers have significantly higher levels of care involvement than non-co-resident fathers? Is there evidence that any questionnaire items function significantly differ-ently for non-co-resident compared to co-resident fathers?(d) What is the level of agreement between fathers’ and mothers’ reports of fathers’ involvementin different modes of child care?We intend for our findings to inform future population-level research into the involvement of men inchild-rearing. This future research could be used to design interventions to promote beneficial forms ofmen’s involvement, both for the purpose of improving child health and for transforming harmful gendernorms.We have chosen to restrict our focus to fathers of young children. We justify this focus becauseof our interest in contributing to reductions in child morbidity and mortality. However, research intothe nature and sources of men’s caregiving at other stages of the life course (for example, in parentingschool-aged children and adolescents, and in caring for sick individuals) is also warranted. We alsoacknowledge that better understanding men’s positive contributions is only one part of the problem.Further research is also needed on the ways in which men’s behaviour can be harmful for children.However, this will not be a focus of the present study.4Chapter 1. IntroductionIn the following chapter we review theoretical and empirical literature on the relationship betweenfathers’ contributions and children’s wellbeing. We highlight some gaps in the existing literature, whichthe above research questions are intended to address. In chapters 3-5 we present the analytic approach,results and discussion for research questions 1a-e. Chapters 6-8 address research questions 2a-c. InChapter 9 we present a unified discussion of our findings and recommendations for future research andinterventions.5Chapter 2Literature ReviewThis chapter reviews the theoretical and empirical evidence base for studying fathers’ influences ontheir children’s health. We have two aims in reviewing this literature: i) to describe various conceptualframeworks that this study draws from; and ii) to highlight the existing knowledge gaps that our researchquestions are intended to address. We begin by defining a few terms used throughout the review. Wethen review conceptual literature on the determinants of child health and on the sources and types ofcontributions fathers may make to their children. We synthesize ideas from this theoretical material toidentify pathways through which fathers may influence their children’s health. In the final section ofthis chapter, we review empirical evidence for the influences that fathers can have on children’s health.DefinitionsConventionally the word father is used to denote a child’s male genetic parent, or the man married tothe child’s mother. In this study we have chosen primarily to focus on fathers who meet this biologicaldefinition. Our reasons for this are that existing household surveys in South Africa use the biologicaldefinition to identify children’s fathers, most research on the influence of household membership onchild health focuses exclusively on biological fathers, and most theoretical research on the parentingpractices of fathers has been oriented toward biological fathers. For this reason, in this study we use theterm father to refer to a child’s biological father.However, because of demographic changes, including increasing rates of non-marital childbearing,divorce, step-parenting and adoption, the conventional biological definition does not adequately capturethe situation for many men who are doing fathering today [23]. Scholars have emphasized the needto understand the broader, social definition of being a father, that is, as “a role that is understood andexercised in different ways” [24, p. 14]. The term fatherhood is used to refer to the collective socialmeanings associated with being a father [25, p. 7]. In contrast to the biological concept of father, thesocial concept of fatherhood is fluid, both over time and between different cultures and social classes. Inmany traditional African cultures, and increasingly in western societies, the social meanings associatedwith being a father do not presuppose genetic relatedness [24]. Although we focus on biological fathersin this study, we are interested in the ways in which different men enact the role of father, given theresources and constraints they experience, rather than in the genetic endowments of fathers to theirchildren. Our intention is that the findings of this study should inform future research with biologicalfathers and with social fathers (i.e.: men who meet the social definition of father towards a child, eventhough they are not the child’s biological father).The meanings attached to being a father can also be distinguished from the actual parenting practices62.1. Theoretical basis for paternal influence on child healthof fathers, which are collectively referred to as fathering. Related to fathering is the concept of fatherinvolvement, which has been used to define the direct caregiving activities of fathers with children [26,23, p. 884]. Care, as used in this definition, refers to both the “set of activities and resources” given bythe carer as as well as “feelings of care....” [27, p. 17-18]2.1 Theoretical basis for paternal influence on child healthThe theoretical construct of primary interest in this study is fathers’ capacity to positively influencetheir children’s wellbeing. In the following sections we synthesize ideas from three separate conceptualframeworks to explain how this capacity becomes manifested in a variety of different fathering practices,which in turn may influence children’s health. The first framework distinguishes different pathways or“modes” through which fathers may influence the care their children receive. The second frameworkproposes a causal mechanism for how social and economic characteristics determine children’s risks ofmortality in low-to-middle-income settings. The third framework applies a systems ecological lens toorganize the various factors which influence the amount and nature of fathers’ involvement in children’scare. Although we draw ideas from multiple conceptual models, throughout we refer to this collectionof conceptual material as a singular “model”. Central to our thinking is the idea that the nature of aman’s fathering will depend on his motivation, but also on his skill at recognizing and managing hischild’s needs and, crucially, on the family and societal context [26]. Skill and context can be consideredas resources that fathers mobilize in order to realize their desires for child wellbeing. Considering theinfluences on fathering as a system is important because access to resources is not uniform across orwithin societies. Therefore, similar motivation for child wellbeing may manifest in different patterns offathering depending on the individual’s situation.2.1.1 Conceptualizing the modes of paternal influence on child healthTo explore fathers’ influences on children’s health, it is necessary to consider the various ways for fa-thers to be involved in their children’s lives. These different types of involvement have been referredto as “modes of paternal influence” [28]. Fathers may be directly involved in caregiving activities withchildren. In addition, fathers may indirectly influence children’s care through their economic provision-ing and by providing support to mothers and other caregivers [28, 29]. We describe these modes ofinfluence in more detail below.Direct caregiving activities, defined as ‘father’s involvement’ [26], have received considerable at-tention in research on children’s psychological development. To increase the conceptual clarity andcomparability of early research in this area, Lamb et al. proposed a model comprised of three distinctdomains of father involvement: engagement, accessibility and responsibility [26]. Engagement is de-fined as direct interaction with a child, for example during caregiving activities and play. Accessibilityincludes being available to and supervising the child but involves less direct interaction. Responsibil-ity is taking a primary role in recognizing and making arrangements for a child’s needs, for example,organizing to attend health clinics for immunizations. Early uses of the father involvement construct72.1. Theoretical basis for paternal influence on child healthfocused on the amount of time fathers directed to each domain of involvement and did not consider theactual content of caregiving activities. However, in more recent applications of this model, researchershave evaluated fathers’ amounts of involvement in specific care activities expected to be beneficial tochildren [30]. This revised construct has been labelled ‘positive father involvement’ [30]. There hasbeen relatively little research into the importance of fathers’ direct caregiving activities for child health.Possibly this is because caregiving is traditionally considered to be the responsibility of mothers, leadingmany to assume that fathers’ contributions in this domain will be of secondary importance [28, p. 7-8].Economic provisioning (or breadwinning) features prominently in popular perceptions of ‘good’fathering (in South Africa as well as in North America) [31, 32]. It includes ensuring that children’sneeds for nutritious food, clean drinking water and adequate clothing and shelter are met, and is thus animportant means for fathers to promote good child health. Nevertheless, there is frequently a trade offbetween fathers’ levels of economic provisioning and the amount of time they spend in direct child careactivities.Fathers may also contribute indirectly to their children’s wellbeing by supporting mothers or othercaregivers. Support may be emotional or practical, such as giving information or advice, caring for otherchildren, or doing household chores [28]. In terms of child health outcomes, support from fathers may beparticularly beneficial in helping to establish an environment conducive for mothers to practice health-promoting care behaviours like breastfeeding [33] and by enabling mothers to cope with emotionallyupsetting life experiences, which might otherwise increase the risk of adverse birth outcomes [34] orinterfere with maternal-infant bonding [35, 36].2.1.2 Distal and proximal influences on child health in South AfricaIn this section we describe a conceptual framework that integrates perspectives from biomedical andsocial science research to organize the major determinants of child morbidity and mortality in low-and middle-income countries [37]. We draw on this framework to identify child health determinantsover which fathers may have some control via the modes of paternal influence described above. Theframework also allows us to identify competing determinants that need to be accounted for in trying toisolate the effect of father’s influences.Consistent with findings from biomedical research, the framework identifies illness and malnutri-tion, operating in synergy, as the direct causes of most child deaths. Furthermore, the framework em-phasizes that children’s susceptibility to illness and their chances of recovering good health are relatedto other individual-level biological characteristics as well as to characteristics of the social, economic,and cultural contexts. The causal foundation of the framework, as described by Mosley and Chen, isthat “all social and economic determinants of child mortality necessarily operate through a common setof biological mechanisms, or proximate determinants, to exert an impact on mortality” [37, p. 25]. Theproximate biological determinants they identified were:• maternal factors, such as age, parity and birth interval;• environmental exposure to infectious organisms through ingestion, inhalation, and insect vectors;82.1. Theoretical basis for paternal influence on child health• nutrient deficiency;• injury; and• personal illness control, which includes measures to prevent illness and therapeutic measures torestore health [37].Millard extended the concept of proximal determinants to include child care practices and other be-haviours (for example, food preparation practices) that influence children’s risk of experiencing illnessand malnutrition [38].The more distal, socioeconomic determinants of child health identified by these authors includeindividual characteristics of children’s parents (such as knowledge and attitudes); household character-istics (including income and wealth, and division of labour and decision-making power); and featuresof the natural environment and the social, cultural, political and economic contexts in which familiesare located [37].Fathers may influence their children’s health by intervening directly on the proximal determinants.For example, through involvement in direct caregiving, as well as by supporting other caregivers, fatherscould promote care behaviours which reduce children’s exposure to infectious organisms and reducetheir risk of injury. Alternatively, fathers may make contributions at the level of the socioeconomicdeterminants to improve their children’s health. For example, by helping to provide adequate food,water, shelter and transport, fathers could reduce children’s exposure to infectious organisms, reducetheir risk of malnutrition and improve their access both preventative and therapeutic health care services.However, fathers require access to resources to effectively intervene on the determinants of childhealth. In their framework, Mosley and Chen identified parental knowledge and skill and householdincome as examples of important resources. Other theorists in economics and sociology have identifieddifferent types of family or household resources that are important for child wellbeing. For example,Becker’s time allocation (or household production) model [39] views families as units of production,capable of acting rationally to maximize their production of desired outcomes (reviewed by Ribar [40,p. 4-6], and Haveman and Wolfe [41, p. 1832-4]). Decision-makers in the family decide how muchof family members’ time will be allocated to generating economic resources and how these resourceswill be allocated to purchasing goods and services. To some extent, family members can substitutepurchased goods and services in place of contributions of their time in order to achieve desired outcomes.Importantly, these decisions are constrained by the total quantity of family members’ time, and by thewages and prices available to them, neither of which are constant across families. We apply this modelto suggest hypotheses about parents’ decision-making in relation to their children’s physical wellbeing,given the resources available to them. For example, parents choose between whether to allocate time tobreastfeeding or to substitute this with purchased infant formula. Holding prices and wages constant,mothers in households with fewer adult members (and, therefore, overall less time) may experiencegreater pressure to allocate time toward wage labour and away from breastfeeding. In contrast, givensimilar amounts of time, mothers in households with access to lower wages or higher formula prices,would be predicted to be more likely to breastfeed.92.1. Theoretical basis for paternal influence on child healthWe also acknowledge some important limitations of the time allocation model identified by feministscholars: first, there may be conflict between household members over the prioritization of differentoutcomes, and, second, power over the allocation of household resources is usually not shared equallyamong household members [42]. In addition, male and female parents may vary in how they make useof similar resources in securing child health. Research in South Africa suggests that the greater financialautonomy women enjoy in female-headed households may allow them to make food purchasing deci-sions that are more sensitive to children’s needs than would be made in similar male-headed households[43, p. 21]. Because access to resources is strongly influenced by gender, male and female parentsmay be able to mobilize different types of resources in differing amounts [28]. For these reasons, it isimportant to consider the nature of the relationships among household members as well as how accessto resources is shaped by wider gender relations. These influence each parent’s ‘bargaining power’ indecisions about how household resources will be allocated [42].While Becker’s model focuses on time and economic resources, Coleman describes the importanceof human and social capital in children’s development [44, p. S109-13]. Human capital includes par-ents’ knowledge and skill in caring for children, which, in turn, influence their health-related behavioursand care practices. In contrast, social capital derives from social structures that function as resourcesfor individuals. These include the relations between family members [44, p. S110]. For example, ac-cording to the patrilineal system of descent common among the African cultural groups of South Africa,acknowledged paternity confers on children recognition as a member of their father’s clan (or kin net-work) [45]. Structured relations among clan members include reciprocal obligations to provide care forchildren and assistance in times of need [45, 46].Considering the types of resources that parents require to influence their children’s wellbeing, aswell as how access to these resources is shaped by social structures such as race, gender and socialclass [38], is important for comparing father’s influence in different household and societal contexts.This understanding is critical for studying child health in the South African context, where the for-mer Apartheid political system enforced a grossly unequal distribution of resources and where largeinequities are still evident, despite radical legal reforms post-Apartheid. In addition, some biologicaldeterminants of child health will be out of parents’ control. Examining interactions between contextand individual behavioural and biological characteristics can help to elucidate the complex mechanismsproducing positive child health outcomes.2.1.3 A systemic ecological model of fathering and child healthTo conclude our review of theoretical literature on father’s influences on child health, we expand on theidea that both fathering and child health are shaped by a complex system of factors, spanning from thelevel of individual children and fathers up to the level of societies. We suggest that a number of thefactors identified as determinants of child health in Mosley and Chen’s model also influence fathering,meaning that it is important to consider these factors either as potential confounders or as modifiers offather’s influence on child health.Doherty et al. proposed a ‘systemic ecological’ model of the influences on fathering [47]. This102.1. Theoretical basis for paternal influence on child healthmodel is useful for our purposes because it is intended to be general enough to describe factors influenc-ing the practices of fathers who live with their children and those who live apart (hereafter referred to asco-resident and non-co-resident fathers, respectively). In addition, it depicts fathering as a dynamic pro-cess involving individual, inter-personal, and contextual factors. The central portion of the model is thechild-father-mother triad, in which individual characteristics of each member as well as characteristicsof the relationships between them are emphasized.• Father characteristics include knowledge and skill, commitment, psychological wellbeing, resi-dential status, employment characteristics, and identification with the father role.• Mother characteristics include attitudes toward and expectations of the father, support of thefather, and employment characteristics.• Child characteristics include sex, age, developmental status, temperament and behavioural diffi-culties.• Father-mother relationship characteristics include marital status, dual vs. single earner, relation-ship commitment, co-operation, mutual support, and conflict.The model also highlights the importance of the contextual environment in which the triad is embedded.Contextual characteristics influencing fathering include institutional practices, employment opportuni-ties and other economic factors, race-related resources and challenges, cultural expectations, and socialsupport.It is worth identifying factors in this general model that may be of particular importance for fatheringin South Africa due to the cultural, economic and political history of the country. The particular waythese factors intersect in South Africa may be somewhat unique. Nevertheless, there are clear similaritieswith other Southern African countries and even with the situations of African Americans in the UnitedStates [48].First, a large proportion of South African children and fathers live apart. Between 2002 and 2013the percentage of children under 5 years old who were residing with their biological father declinedfrom 43% to 39% (Calculated by the author from South African General Household Survey data [49]).Factors contributing to residential separation include labour migration, cultural traditions related tohousehold formation, and declining marriage rates.Temporary migration for work is a long-established livelihood strategy in South Africa. During theColonial and Apartheid eras, increasing numbers of Black South Africans were drawn into wage labourin cities and industrial centres [50]. A goal of the Apartheid government was to reserve well-resourcedareas of the country for Whites, while ensuring access to a steady supply of cheap Black labour. Tothis end, they forcibly removed Blacks to resource-poor “homelands” in rural areas of the country andestablished laws (known as “pass laws”) which prevented Blacks from living in “White areas” exceptwhere they were employed [50]. Together with short-term labour contracts, these government policiesresulted in a pattern of cyclical labour migration between permanent family homes in rural areas andtemporary accommodation near urban places of work [51]. Because movement was made so difficult,112.1. Theoretical basis for paternal influence on child healthmost migrants’ families remained at rural homes. As a result, Black households were often “stretched”between different physical dwelling places.The majority of migrant labourers were men [51]. Many retained strong ties with wives and chil-dren at their rural homes and visited them during periods between labour contracts [50]. It was alsonot uncommon for wives to defy pass laws to visit their husbands in cities, in some cases bringingchildren with them and in others leaving them with relatives [52]. This type of movement of householdmembers between different residences lead researchers to characterize Black South Africans’ residentialarrangements as “fluid” [53].Since the formal abolishment of pass laws in 1986, many Black men and women have moved per-manently to cities. However temporary labour migration from rural areas has remained common andonly recently appears to be declining. In 2008, it was estimated that 22% of rural Black households hadone or more member who was an absent migrant worker [54]. This contrasts with estimates around 36%in surveys conducted between 1999 and 2005 [54]. While an increasing proportion of migrant workersare female, still two thirds were male in 2008 [54].A second important feature of the context of fatherhood in South Africa are cultural norms and idealsrelated to parenting and the gendered-division of household labour (including care work). These normsinform fathers’ own expectations about their responsibilities to their children, as well as the expectationsof members of fathers’ social networks. These collective expectations may be difficult for individualmen to challenge [32].In the precolonial Southern African family system, power came with age, marriage, and being male[46]. The ideal of fatherhood was linked with becoming the patriarch of one’s family. Importantly, fam-ilies or households were not formed around a conjugal couple, as has been the long-standing traditionin northern western European cultures [45]. Instead, the principal of family formation was patrilinealdescent, i.e.: descent from the father’s lineage [45]. People sharing a common lineage were regarded asmembers of an extended family, or kin group. Marriage and payment of lobolo (bridewealth) resulteda woman’s subsequent children becoming members of her husband’s lineage [45]. Children born tounmarried mothers could either be accepted into the maternal grandfather’s kin group, or the father’sfamily could acknowledge them by paying ‘damages’ to the mother’s family [45, 55]. Caring andproviding for children was seen as a collective responsibility of the kin [56, 46]. Customarily, directcaregiving was the responsibility of mothers and female family members [46]. Fathers were expectedto be primary providers, but they were also expected to be involved in teaching children and in makingdecisions related to their wellbeing [46].It has been proposed that labour migration and degrading government policies undermined tradi-tional African ideals of fatherhood [56, 46]. During Apartheid, fathers who were migrant labourerswould rarely be able to spend time with their children. This may have lead their material provisioningrole to assume primacy [46]. Beginning during Apartheid, the increasing entry of women into paidlabour has been seen as a further threat to traditional patriarchal fatherhood norms [56, 57]. Finally, it isargued that many young men in cities began to rebel against older ideals of masculinity, which esteemedchildrearing and maturity. New ideals of masculine conduct placed greater emphasis on toughness, inde-122.1. Theoretical basis for paternal influence on child healthpendence and opposition [56]. Possibly these new ideals were mutually reinforcing with the increasinglyviolent political protests of the 1960s and ’70s [46]. Contemporary rigid gender norms that emphasizethe role of men as economic providers and equate caregiving with femininity, provide little impetus formen to take on non-traditional child care roles. These gender norms may even result in a lack of socialrecognition and support for the contributions made by men in caregiving roles [32].Also related to the undermining of traditional gender norms and cultural practices of householdformation is the dramatic decline in marriage rates among Black Africans, particularly those living inurban areas [58, 59]. For example, a 2006 study found that, among women 40-45 years old living inurban and peri-urban areas, 50% had never married, compared to 42% of their rural counterparts [58].Co-habitation of non-marital couples has become more common, particularly among people in urbanareas, but is still relatively rare [59]. For example, in the same study in 2006, less than half of all womenin regular, non-marital relationships were residing with their partners [58].Many Black South Africans still report positive attitudes towards marriage [59]. However, someresearchers have suggested that enduring effects of Apartheid policies on African family life, togetherwith rising unemployment overall and proportionally more women in the paid workforce, have resultedin many young people choosing to remain single [60, 43]. For women, in a society where men are nolonger necessary nor often sufficient for financial stability, remaining single may be seen as an attractiveway of avoiding patriarchal practices of men [60, 43]. Similarly, for men, unstable economic conditionsmay be a barrier to raising money for lobolo. Additionally, marriage itself may seem a less viablestrategy for attaining the masculine ideals of previous generations [56]. Other young people may bedelaying marriage until they have completed school and attained stable employment [61]. For example,among urban Black South Africans, those who are unemployed are less likely to be married [59]. Formany, these necessary conditions for marriage occur relatively late in life, if at all. For the above reasons,a large proportion of South African children are born to unmarried, non-cohabiting parents [62].A third important feature of the fathering context in South Africa is the high rate of unemployment.In 2013, the national unemployment rate was 25% [63]. One third of unemployed people in the coun-try had been without work for longer than a year. Importantly, unemployment is concentrated amongBlack and Coloured1 people with limited formal education who are in the age range where childrea-ring typically begins [63]. Limited employment opportunities and lack of income undermine fathers’ability to provide financially for their children [47, p. 286]. As mentioned above, they can also reduceparents’ chances of marrying and establishing an independent dwelling together, thereby further influ-encing their children’s material living conditions. There is also evidence that unemployment negativelyimpacts other aspects of fathers’ parenting and the quality of the co-parental relationship. Psychologicaldistress resulting from job loss [65] and perceived failure to live up to the socially ascribed ‘provider’role [32] may cause some men to ‘abandon’ their other care responsibilities (Steering by the stars: Beingyoung in South Africa (2002), as quoted in [66, p. 79]). Other men may respond by becoming abusive[67, p. 103-4].1Coloureds are people “of mixed European (‘white’) and African (‘black’) or Asian ancestry, as officially defined by theSouth African government from 1950 to 1991.” [64] During Apartheid, as with Black people, Coloureds were subjected tostrictly segregated occupational opportunities and were forcibly relocated to less desirable areas.132.2. Empirical evidence for paternal influence on child healthThe final feature we emphasize is that HIV-infection is widespread in South Africa and most preva-lent among women and men in their child-conceiving years. In 2012, the prevalence of HIV infectionnationally was 12%. However, among Black African females 20-34 years old, HIV prevalence was 32%and, among Black African males 25-49 years old, it was 26% [68]. Little is currently known aboutthe roles and responsibilities of fathers in HIV-affected families [20]. Nevertheless, HIV-infection isexpected to influence parenting in a variety of ways. HIV illness strains the economic, psychologicaland physical resources parents need to be able to provide good quality care to children [69, 70, 71].Since the initiation of a public antiretroviral-provision program in 2003, access to treatment hasincreased dramatically [68, 72]. However, it is estimated that approximately half of adults eligible forARVs are not receiving them [72]. Coverage is also lower among men than women [68]. As such, adultHIV morbidity is still an important problem.Caring for HIV-infected children also requires significant added emotional, financial, and time re-sources [73]. Current strategies for preventing mother-to-child transmission of HIV mean that evenuninfected infants of infected parents have additional care needs, including avoiding early mixing ofbreast- and formula-feeding and giving ARV prophylaxis throughout breastfeeding [74]. HIV-infectionmay also negatively impact the co-parental relationship and is associated with increased risk of relation-ship dissolution [75].2.2 Empirical evidence for paternal influence on child healthThe question of what influence fathers have on their children’s wellbeing has produced a long historyof empirical research. Our aim here is to synthesize findings for the effects of fathering on the physi-cal health outcomes of young children. We include evidence from two reasonably distinct domains ofresearch: one examining the child health consequences of residential separation of fathers and children(commonly referred to as father absence), and the other examining the effects of father involvement.For the purposes of this review, we define father involvement broadly as including direct involvement inchild nurturance as well as providing for children’s material needs and giving emotional and practicalsupport to mothers. Typically co-residence is not considered to be a component of father involvement,per se. This is because research evidence suggests that it is fathers’ positive contributions to children,rather than their mere physical presence, that is important for child wellbeing. However, residentialarrangements are one factor influencing fathering [47]. Because of the high rates of residential sepa-ration in South Africa and because of the very limited data on fathering, we propose that consideringresidential arrangements does provide insight into the ways that fathers can influence their children’shealth. However, we emphasize that there is ample research demonstrating that non-resident fathers canbe beneficially involved in their children’s lives. (In this sense, the term ‘father absence’ is misleading.)Indeed, while father-child co-residence may have an important influence on aspects of fathering that arecontingent on physical proximity, it may have relatively little influence on other aspects. For example,co-residence has been shown to be a poor predictor of whether fathers provide financial support fortheir children in South Africa [76]. Therefore, throughout this review we have attempted to complement142.2. Empirical evidence for paternal influence on child healthstudies of father absence with relevant studies of father involvement, wherever evidence is available. Inso doing, we aim to identify potential father involvement mechanisms by which residential separationmay manifest in differences in child health.The majority of studies of father absence and involvement have been conducted in the United States(US) and Western Europe where several decades of rising rates of divorce and extramarital childbearinghave resulted in increasing percentages of children living in ‘single-mother households’ [77, p. 5].Researchers’ interest in the issue of father absence has, in part, been driven by concern over whetherthese demographic changes are having a negative influence on child wellbeing. Over the same period,expectations for fathers to take a greater share of the responsibility for child nurturing have increasinglypermeated dominant fatherhood ideals [31]. Father involvement research has attempted to documentwhether fathers’ actual parenting practices have kept pace with these evolving ideals, and what theconsequences have been for children (as well as for mothers and fathers themselves) [78].While similar demographic changes have occurred in South Africa [79, 58], important contextualdifferences may limit the transferability of research findings from Western countries. For example,labour migration is a common and conceptually distinct reason for South African fathers to spend ex-tended periods away from their children’s households [48, p. 259]. In addition, only a minority of SouthAfrican children live in truly ‘single-parent’ households [80, p. 1022-3]. Many children with absentfathers reside together with their mothers and other maternal relatives [81, p. 17]. Furthermore, culturalexpectations which emphasize men’s responsibility for financial provisioning and cast child nurturingas ‘women’s work’ may deter many men from becoming highly involved in routine child care [32, 82].To summarize, research from high income countries suggest that father absence is associated withpoorer academic achievement, behavioural competence and psychological wellbeing, social relations,and economic security [83, 77, 84]. In contrast, greater father involvement predicts better outcomesin many of the same domains of child wellbeing [85]. There is more empirical support for the effectsof paternal engagement in direct caregiving than for other types of involvement [21]. However, theexisting evidence is not sufficiently rich to identify which specific aspects of engagement most effec-tively promote beneficial outcomes [21, p. 157]. Studies of non-resident fathers’ involvement, suggestthat payment of child support and practicing authoritative parenting (i.e.: parenting characterized by amixture of supportiveness and non-coercive control) are associated with improved academic outcomesand fewer behavioural and psychological problems [86]. However, high frequency of contact with non-resident fathers is not consistently associated with child wellbeing [86].Studies of the relationship between fathering and children’s physical health outcomes are rarer.Nevertheless, there is some evidence that young children with absent fathers tend to have higher risksof being diagnosed with, and of being hospitalized for, physical illnesses [87, 88], and of experiencingaccidental injuries [87]. In contrast, young children whose fathers are more engaged in routine careactivities tend to experience fewer injuries [89]. Experiencing parental divorce or living in a ‘single-mother household’ during childhood has also been found to predict poorer health in adulthood [90].A frequent criticism of father absence research is that the observed associations may be due to theeffects of factors that ‘select’ children into father-absent households rather than to any negative effect152.2. Empirical evidence for paternal influence on child healthof father’s absence itself. For example, parents’ pre-separation characteristics like low socioeconomicstatus, high levels of relationship conflict, and mental illness all make separation more likely and arealso, on average, associated with poorer outcomes for children [77, p. 16]. Studies of father involvementare also subject to potential selection bias and confounding effects. In addition, highly detailed studiesof involvement have typically been based on small, non-random (often culturally and socioeconomicallyhomogenous) samples, potentially limiting the generalizability of the findings [91, p. 24]. Nevertheless,studies rigorously designed to reduce potential selection bias and control for confounding still suggestfather absence can be detrimental for children [84], and that higher levels of involvement are beneficial[21]. Furthermore, recent, large-scale, population-based studies of father involvement in the UnitedStates are beginning to reveal patterns of fathering among low-income and minority groups [92], andconfirm the positive influence of father involvement on child development in these populations [93].While considerable interest has been expressed for better engaging men in child health programsand policies in lower-resource countries, evidence is still quite limited as to the influence that fathershave on child health in these contexts and the ways in which men can be most beneficially engaged [20].A particular challenge for researchers working in these countries is the almost complete lack of detaileddata on fathering [94, 48]. In addition, many of the studies that have been conducted suffer from themethodological limitations mentioned above.Nevertheless, there is a small body of literature documenting the child-health correlates of father-child residential arrangements and of fathers’ involvement in low- and middle-income countries. Wereview the findings below. We have chosen to restrict our review to studies of physical health out-comes because persistently high rates of child morbidity and mortality in South Africa (as well as theinequitable distribution of these risks by race and social class) mean that, in this and similar settings,there is an imperative to better understand the factors influencing children’s physical health2.We did not conduct a formal systematic review because we were interested in bringing together lit-erature from two quite distinct areas of study (1. Research on father-child co-residence, and 2. Researchon father’s involvement in child care). We were concerned that a narrowly focused review might preventus from identifying overlaps between the two sets of literature.To identify research on the child-health effects of residing with a father, we searched the Medlinedatabase using the terms listed in table 2.1 on the following page. In line with Mosley and Chen’s model,we sought to identify studies that included any of the following child health outcomes: breastfeeding orreceipt of immunizations (i.e.: proximal determinants); diarrhoeal or respiratory illnesses; malnutrition;or mortality. In addition, studies needed to include a measure of father-child residential arrangementsor a related characteristic of the household context (specifically: parents’ marital status, or presence ofadditional relatives) as a predictor of child health outcomes. We included the additional household con-textual characteristics because they may influence the nature and consequences of fathers’ parenting, asdescribed in section 2.1.3 above. We refer to these characteristics collectively as measures of householdstructure. In addition, we searched the Medline database for articles referencing Mosley and Chen’s2Morbidities, including malnutrition and infection, are important causes of impaired child development in low- and middle-income countries [95]. Therefore, the evidence addressed in this review is also relevant for understanding potential mechanismsby which fathers may contribute to developmental outcomes in later childhood and adolescence.162.2. Empirical evidence for paternal influence on child healthTable 2.1: Terms used to search the Medline database for studies of the child-health effects of residingwith a fatherSEARCH TERMS1. (household or family) and structure in: A2. father and (co-resident or co-residence or coresident or coresidence or non-resident or non-residenceor non-co-resident or non-co-residence or non-coresident or non-coresidence or absent or absence) in:A3. single and mother in: A4. Immunization or Vaccination in: B5. breastfeeding in: A or Breast Feeding in: B6. Diarrhea, Infantile or Diarrhea in: B7. Respiratory Tract Infections in: B8. Child, Preschool in: B9. Infant in: BA = title, abstract, original title, name of substance word, subject heading word, keyword heading word,protocol supplementary concept word, rare disease supplementary concept word, or unique identifierB = Medical Subject HeadingSEARCH STRATEGY:(1 or 2 or 3) and (4 or 5 or 6 or 7) and (8 or 9)model.We screened the complete search results for English-language articles that presented analyses ofdata from low- or middle-income countries and presented a measure of association (relative risk, riskdifference, odds ratio, or hazards ratio) between one or more characteristics of household structure andone or more of the child health outcomes of interest. We also scanned reference lists of included articlesfor additional articles meeting our criteria.We did not use a formal protocol for evaluating the quality of evidence for each association. How-ever, we describe whether studies were cross-sectional or longitudinal and whether they controlled forhousehold socioeconomic status, which was determined a priori to be an important potential confound-ing factor.Because there is limited research from low- and middle-income countries focusing specifically onfathers’ parenting practices, we used relatively general search terms to identify relevant studies. Wesearched the EBSCO databases for the terms listed in table 2.2 on the next page. We included pri-mary research articles which investigated linkages between fathers’ conduct and any of the child healthoutcomes described above.We have organized the review by type of child health outcome examined: proximal determinants,illnesses, malnutrition, and child mortality. Within each category of outcomes, we first synthesize ev-idence for the influence of household structure. Where available, we then present evidence for theinfluence of fathers’ involvement.172.2. Empirical evidence for paternal influence on child healthTable 2.2: Terms used to search the ESBCO databases for studies of the child-health effects of father’sparenting practicesDATABASES SEARCHED:MedlineBiomed CentralEconLitPsychArticlesHumanities databaseFamily studiesWomen studiesAnthropologyNursingSEARCH TERMS:1. (father and involvement) or (father’s and involvement) or (fathers’ and involvement) or fatherhoodor fathering or father or fathers or father’s or fathers’ in: A2. Africa in: A3. (child and wellbeing) or (child and health) in: AA = KeywordsSEARCH STRATEGY:1 or 2 or 3Proximal biological and behavioural determinantsDifferences in the following proximal determinants have been studied in relation to household structure:birth weight, breastfeeding, and immunization. We identified only a single study, in South Africa, whichspecifically explored the effect of fathers’ residential arrangements. This study found that co-residenceof fathers with mothers during pregnancy is associated with modestly, but statistically significant, higherbirth weights [96].A larger number of investigators focused on parents’ marital status. Mothers who are married orhave co-resident non-marital partners were found to give birth to significantly heavier babies than un-partnered women (South Africa) [96]. Similarly, mothers with co-resident spouses were significantlymore likely to be fully breastfeeding3 their babies between birth and six months, and also significantlymore likely to breast feed at all between birth and 12 months (Philippines) [97]. In addition, com-pared to children whose mothers were divorced or had never married, children with (monogamously)married mothers were found to have significantly higher probability of completing the polio immuniza-tion schedule (Kenya [98]), and higher probability of receiving all routine immunizations (Trinidad andTobago [99]; Jamaica [99]; Kenya [100]).Other studies examined the association between immunization completeness and living in extendedhouseholds (where children reside with their parent(s) and other adult relatives). These studies sug-gest that, after controlling for household socioeconomic status, there is no significant difference in the3Full breastfeeding is defined as feeding breast milk in combination with only non-nutritive quantities of other liquids [97].182.2. Empirical evidence for paternal influence on child healthprobability of being fully immunized for children living in two-parent households compared to thosein extended households (Niger, urban Nigeria, Trinidad and Tobago, and Jamaica) [99, 100]. Childrenin nuclear households in rural Nigeria are the exception, having been found to have lower probabilityof being completely immunized than their counterparts living in households which include their par-ents’ siblings [100]. However, the fact that the nuclear household category in this analysis includedboth single-parent and two-parent households makes the findings difficult to compare to those of otherstudies reviewed.Researchers have also documented that women who receive support from a male partner tend to havebetter antenatal care outcomes and breastfeeding practices. In sub-Saharan Africa, much of this researchhas focused on correlates of uptake and adherence to interventions to prevent vertical (i.e.: mother-to-child) transmission of HIV. While this research typically does not conceptualize of support from malepartners as a type of fathering, the findings clearly contribute to our understanding of potential modes ofpositive paternal influence. For example, participation of male partners in antenatal HIV counselling andtesting (HCT) is associated with a higher probability for HIV-infected women to receive antiretroviralprophylaxis during delivery [101, 102], better adherence to exclusive formula feeding or breastfeeding[101], and significantly reduced risks of vertical HIV transmission and infant mortality during the firstyear of life [16]. Receiving HIV test results and post-test counselling together as a couple, compared toindividually, is associated with an even greater increased probability of using antiretroviral prophylaxisduring delivery [102].A limitation of these studies is that they do not provide insight into the components of father involve-ment that produce the observed benefits. Antenatal HCT sessions include information about preventingHIV vertical transmission. Therefore some of the beneficial effect observed could have been achievedthrough better informed fathers being able to provide more encouragement and decision-making sup-port to mothers. However, participation of male partners in couples HCT interventions is voluntaryand uptake is generally very low. Therefore selection bias may explain some of the findings. Couplesagreeing to participate in HCT may be in more secure, communicative relationships. In these types ofrelationships, fathers may generally be more supportive of their partners and more positively involvedin caring for children [47]. It could be these (unmeasured) beneficial relationship characteristics, moreso than couples HCT, that explain the positive outcomes observed.Qualitative research on the infant-feeding experiences of HIV-infected mothers in South Africa pro-vides more detailed insight into beneficial fathering. These studies have consistently found that dissentby fathers is a significant barrier preventing some women from adhering to infant feeding recommen-dations [103, 104]. The corollary is that some fathers are an important source of emotional supportto mothers, particularly if they are aware of the mother’s HIV-status. In situations where economicconstraints influence infant feeding considerations, fathers may be able to support optimal infant feed-ing practices by providing material support to mothers, for example in the form of infant formula forHIV-infected women practicing exclusive formula-feeding [105, p. 106-7].192.2. Empirical evidence for paternal influence on child healthIllnessesWe identified three studies comparing risk of illness between children living in different householdstructures. Interestingly, using different approaches, all three studies examined the effect of father’sabsence due to migrant labour. It is also noteworthy that all three studies were restricted to householdsin rural areas.One study, using longitudinal data from Mexico, examined the effects of married fathers’ absencedue to migrant labour on their children’s risks of experiencing any illness and diarrheal illness [106].It found that father’s absence is associated with a significantly higher probability of both categoriesof illness, with the effect being stronger for diarrhoea. Notably, the detrimental effects were observedboth in models that adjusted for a small set of observed indicators of household SES, and in childfixed effects models, which account for unmeasured time-invariant differences between children andhouseholds. This suggests that the findings are not the result of confounding by pre-migration householdsocioeconomic conditions.The remaining two studies, from Kenya and Ghana, were cross-sectional and compared childrenliving in male-headed households to those living in de facto female-headed household (defined ashouseholds in which the self-declared male heads are absent for at least 6 months of the year) [107,108]. Migrant labour is identified as an important reason for the absence of male household heads inthese countries. In contrast to the study in Mexico, both African studies reported no significant dif-ference in illness prevalence between children living in de facto female-headed household and thosein male-headed households. The categories of illnesses examined were: any illness, diarrhoeal illnessand malaria. Neither set of analyses controlled for potential confounding (in one case because of smallsample size, and, in the other, because childhood illness was not the primary outcome of interest). Thedesign limitations of these two studies means that their results should be interpreted cautiously. How-ever, it is also important to consider that differences in social, cultural and economic context couldexplain the difference between these findings and those from Mexico.We did not identify any studies of the influence of fathers’ involvement on the incidence of child-hood illnesses. However studies of the dynamics of household treatment decision-making indicate thatfathers do have a role in influencing the outcomes of illness in Africa. Studies from Kenya and Senegalhave found that, although mothers are frequently the first to recognize illnesses, they rarely make treat-ment decisions alone [109, 110, 111]. Usually, care decisions involve the mother obtaining financialassistance or advice from another household member, with fathers being the most common sources ofboth types of support. Fathers appear to be particularly important in paying for treatments for childhoodillnesses [109, 110, 112, 111] Compared to mothers, fathers tend to be more likely to pay for more costlytreatments, including pharmaceutical therapies and visits to health facilities [110, 112]. An importantreason for this pattern may be that men are able to draw upon patriarchal cultural and societal norms toassert power over the allocation of household resources [109, 111]. The expectation is that mothers willconsult with husbands and elders around childhood treatment decisions. However, in reality, fathers’involvement in treatment decisions appears to depend on the type of illness, the cost of treatments beingconsidered, and the nature of the relationships among family members [109].202.2. Empirical evidence for paternal influence on child healthEven though parents of uninsured children are meant to be exempt from paying for child healthservices accessed through public primary- and district-level facilities in South Africa, a considerableminority are still inappropriately made to pay for these services [113]. In addition, transport costs canbe a significant barrier to accessing care, particularly for individuals in poor, rural households [113]. Forthese reasons, financial contributions from fathers may be important in determining whether and howquickly South African children receive medical care.MalnutritionMalnutrition appears to have received more research attention in relation to household structure thanthe other outcomes reviewed here. However, as with the other outcomes reviewed, it is uncommon forinvestigators to have specifically focused on fathers. Only one study, from South Africa, examined co-residential relationships of children with fathers. It found that neither residing with a father at the timeof the survey, nor the percentage of life spent residing with a father were associated with any changein children’s likelihood of malnutrition [114]. Living with both parents also showed no significantassociation, whereas residing with a mother (vs. not) was associated with a significant reduction inchildren’s likelihood of malnutrition. Similarly, a study from The Gambia, which examined parentalmortality, found that having a living mother was associated with children having significantly higherweight (adjusted for age and height). In contrast, having a living father was associated with no significantdifference in child weight [115]. A limitation of both studies is that they were not able to assess forconfounding by household socioeconomic characteristics.A larger number of studies compared children based on whether their mothers’ partners were house-hold residents4. Findings from these studies are mixed, with some suggesting that the absence of apartner is associated with a tendency for children to be more malnourished [116, 117], whereas othersfind no difference [118, 108] or that the opposite is true [107]. One factor that may contribute to thesecontradictory findings is whether investigators compared children of married to unmarried mothers, orrestricted their comparisons to children of different groups of married mothers. Analyses restricted tochildren of married mothers tend to find that whether the mother has a resident or a non-resident hus-band is not associated with a change in her child’s probability of malnutrition. In contrast, analysesthat include children of both married and unmarried women generally suggest that children of marriedmothers have the lowest probability of malnutrition [98], but children whose mothers have co-resident,non-marital partners may also have lower probability of malnutrition than those whose mothers have anon-resident partner or no partner [116].Studies in South Africa and Vietnam have examined associations between father’s involvement andchild malnutrition. In South Africa, receiving financial support from a father was associated with sig-nificantly reduced probability of malnutrition [114]. Due to small sample size, these analyses werenot stratified by father-child co-residence status. However, in itself, residing together was not found to4The relationship between mother’s partner and child is not described in these studies. For the purpose of this review wetreat the partner as the child’s father, but we recognize that the definition of father involves more than the man’s relationshipto the mother.212.2. Empirical evidence for paternal influence on child healthbe associated with probability of malnutrition. In Thailand, children whose fathers were involved intaking them for immunizations and in caring for them when sick had lower probability of malnutrition[119]. In contrast, neither the amount of time fathers spent in routine child care activities (e.g.: feeding,playing, bathing) nor the amount of time spent in household chores were consistently associated withmalnutrition. These analyses were restricted to children with co-resident fathers, potentially implyingthat the findings are not generalizable to children with non-resident fathers.Observations from two studies of household structure provide additional insight into involvement ofnon-resident fathers. A study using data from Kenya and Ghana, showed that households with absenthusbands were not inevitably economically disadvantaged compared to households with resident hus-bands [107, p. 683-4]. This could be because the former receive remittance income from husbands whoare migrant labourers. Because this study found no difference in malnutrition risk between householdswith and without resident fathers, this observation could provide support for the importance of fathers’economic provisioning in preventing child malnutrition.A separate study in Kenya, comparing nutritional outcomes of children of married mothers withresident versus non-resident husbands, included a detailed examination of household agricultural char-acteristics, and division of farming and financial decision-making between wife and husband [108]. Thefindings suggest quite different patterns in the division of decision-making and allocation of resourcesin the two types of households. For example, non-resident husbands tended to retain more control overfamily income but less control over farming decisions than did resident husbands. The authors proposethat differences in resource allocation and decision-making control reflect strategies of household mem-bers to cope with their particular residential arrangements. These strategies appear to result in childrenin both household types having similar caloric intake and similar probability of malnutrition, albeit inan area with relatively high rates of malnutrition overall.MortalityTwo studies (South Africa, The Gambia) estimated whether the death of a parent was associated with achange in children’s mortality risk. The study in South Africa also distinguished between father-childresidential connections (whether they live in the same dwelling) and father-child social connections(whether the father is recognized to be a member of the child’s “household”5). This study comparedmortality risks among children who physically resided with their fathers, children who were part ofthe same household as their fathers but did not reside together, and children whose fathers were notmembers of their households. Other studies compared child mortality risks by sex of household headand by mother’s marital status using data from a number of low- and middle-income countries.Father’s death was not significantly associated with children’s risk of mortality [81, 120]. However,children whose fathers were members of their household had a lower mortality risk than children whosefathers were not [81]. In addition, among children whose fathers were members of their household,5“Household” was defined differently in this study than in most other studies reviewed. In typical household surveys,households are composed of co-residing individuals. Whereas, in this study, households were defined “based on respondents’own perception that they belong to a social group which has a distinct identity and a recognized head of household.” [81, p. 8]222.2. Empirical evidence for paternal influence on child healthhaving always resided together was associated with a significantly lower probability of mortality thanhaving never resided together. Female, compared to male, household headship was not associated withchild mortality [118]. In about half of the countries examined, children of mothers who were continu-ously married to the same partner had significantly lower odds of mortality than children whose mothershad never been married, were formerly married, or had remarried during the child’s life [121].We did not identify any studies examining aspects of fathers’ involvement in relation to child mor-tality.SummaryOverall, it is relatively clear that fathers can influence the health of young children in low- and middle-income countries. However, the evidence is mixed as to whether residing with a father improves childhealth. Furthermore, studies of health-decision making suggest that, while fathers have the potentialto promote beneficial health outcomes for women and children, the fact that fathers tend to have morepower than mothers in household decision-making (particularly in allocating financial resources) can beharmful, leading to unnecessary delays in accessing health care [109]. This review highlights some gapsin the existing knowledge base and some important considerations for future research.First, relatively few studies of family structure have directly examined father-child residential ar-rangements. Given the present rarity of detailed data on fathering in low-resource countries, usingreadily available residential status data from population-level household surveys will likely continueas a preliminary approach to learning more about fathers’ influences on child health. However, usingvariables like mother’s marital status and household head’s sex as proxies for the residential status offathers does not seem suitable. Although these household characteristics are inter-related, they measuredistinct aspects of a complex set of domestic arrangements. For example, the definition of householdheadship may differ between studies, and is frequently subjective [122]. In addition, the wide range ofresidential arrangements in South Africa means that the sex of the household head may not be at allrelated to the residential status of a given child’s father. Declining marriage rates, modestly increasingrates of cohabitation, and the high frequency of marital partners residing apart all mean that mother’smarital status is also an inadequate proxy for father’s co-residence status [79, 58]. The findings reviewedhere suggest that these different dimensions of household structure may not show similar associationswith child health outcomes. However, few analyses have simultaneously examined different dimensionsof family structure or explored interactions between them.Second, the majority of studies of father-child residential arrangements have examined only oneor two child health outcomes. None used the same dataset to study associations between householdstructure and both illnesses and proximal behavioural outcomes intended to prevent illness. Doing sowould allow us to suggest potential mechanisms by which family structure influences child health.Third, considering historical and social context improves the interpretability of (sometimes contra-dictory) findings. The analyses of child malnutrition in rural Kenya and Ghana are good examples: byunderstanding that migrant labour is a common reason for husbands to spend extended periods awayfrom wives and children, a father’s absence can be acknowledged as a rational strategy for families to232.2. Empirical evidence for paternal influence on child healthdeal with poor economic conditions. While many studies highlight the importance of considering theinfluence of social context on family structure effects, none have attempted to estimate the magnitude ofcontextual variation in their effect estimates or to identify specific contextual factors which contributeto this variation.Fourth, the few existing studies of fathering in lower-resource settings have tended to focus on asingle type of involvement; some focus solely on financial provision or financial decision-making, otherson support of mothers, and a limited number on direct engagement in child care. We did not identify anystudies that included measures on all three of these types of involvement. Future research in this areawould benefit from applying broader conceptualizations of fathering and from using more comparablemeasurement instruments. Doing so would allow us to develop a more comprehensive understanding ofthe nature of men’s involvement in families. Researchers could then begin to identify the antecedentsand child health consequences of different patterns of involvement. Such an understanding would betterposition us to design interventions to engage men as partners for achieving positive health outcomes forchildren. A potential challenge is that most of the conceptual work on fathering has been informed byresearch involving North American and European fathers. Whether this work provides an appropriatemodel for studying the parenting practices of South African fathers has not been evaluated.24Chapter 3Methods: Research Objective 1The first objective of this study was to determine whether South African children who reside with theirbiological fathers tend to have better health outcomes than those whose fathers reside elsewhere. Al-though there are no large datasets with an explicit focus on fathering in South Africa, data are availablefrom a number of nationally representative household surveys. These surveys generally collect informa-tion about relationships among household occupants, making it possible to determine whether fatherswere residing in the household at the time of the survey. Some of these surveys also include measuresof children’s health status. We used data from one of the latter surveys, the 1998 South African De-mographic and Health Survey (SADHS), to address research objective 1. In addition, we aggregated1996 South African census data to construct measurements of different neighbourhood characteristicsfor the households surveyed in the SADHS. In the following sub-sections (3.1 and 3.2) we describe thedata collected during the SADHS and the 1996 Census, respectively. In sub-sections 3.3 and 3.4, wedescribe the variables used in our analyses and our steps in selecting the analytic sub-sample. We endthis chapter in sub-section 3.5 with a description of our analytic approach and the specific statisticalmethods used to address research questions 1a-e.3.1 1998 South African Demographic and Health SurveyThe SADHS was a collaborative project of the South African National Department of Health, the SouthAfrican Medical Research Council and the United States Agency for International Development. It wasconducted as part of an international program aimed at collecting comparable, nationally-representativedata on population and health indicators for a number of countries. In many countries these surveyshave been repeated periodically to allow for comparisons to be made over time. In South Africa, twoDemographic and Health Surveys have been completed, in 1998 and 2003. Because the 2003 data werenot publicly available, we limited our analyses to the 1998 data. We obtained these data through theMeasureDHS website [123]. Although they are over 15 years old, these data are valuable for investigat-ing factors that influence children’s health, particularly those that remain stable over time. They includedetailed information about a range of child health indicators and a variety of demographic and socioeco-nomic characteristics related to children’s health and health services utilization. In addition, the sampleis representative of almost the complete population of children under 5 at the time of the survey (withthe exception of children born to women younger than 15 or older than 45 at the time of the survey,who were not asked questions about their childbearing histories). Below we will briefly describe thesampling strategy and questionnaires used for the SADHS. More detailed descriptions are provided in253.1. 1998 South African Demographic and Health SurveyAppendices A and F of the full survey report [6].SamplingIn the SADHS, households were sampled using a stratified, two-stage cluster sampling approach. Anaim of the survey was to provide reliable statistics separately for rural and urban areas in eight of thenine Provinces. In the ninth Province, the Eastern Cape, it was desired that separate estimates could beproduced for each of five health regions. Consequently, the first-stage sample was stratified into urbanand rural areas of each Province, or into urban and rural areas of each health region in the Eastern Cape.Figure 3.1 (page 27) shows a map of South Africa. On the map, rural census Enumeration Areas (EAs)are coloured dark grey, and urban EAs are coloured white.In the first stage, a sample of EAs was systematically drawn from the list prepared for the 1996national census. The probability an EA was selected was proportional to the number of households itcontained. In the second stage, households were randomly sampled from lists of households in each EA,or from maps. Ten households were sampled from each urban EA and 20 from each rural EA. At eachsampled household, field workers completed a single “household questionnaire”. They also attemptedto complete a “women’s questionnaire” with all female household occupants aged 15-49 years.In total, 972 EAs and 12 860 households were sampled. Data collection was not completed in threeEAs, and data from another three EAs were lost, resulting in data being available for 966 EAs. Of thesampled households, 95.2% were successfully interviewed. Among those households that were found tobe occupied, the response rate was 97%. These households were comprised of a total of 12 327 femalemembers eligible to complete the women’s questionnaire, of whom 95% were successfully interviewed.Consequently the overall response rate for the women’s questionnaire (calculated as the product of thehousehold response rate and the women’s questionnaire response rate) was 92.3%.Sample weights, provided with the dataset, can be used to adjust statistics from the sample so thatthey are representative of the South African population at the time of the survey [124]. Except whereindicated, all of the descriptive statistics presented in the results section have been adjusted using thesample weights.263.1.1998SouthAfricanDemographicandHealthSurveyFigure 3.1: Map of South Africa showing the locations of rural (dark grey) and urban (white) 1996 census Enumeration Areas in each province.273.1. 1998 South African Demographic and Health SurveyQuestionnairesThe SADHS data used in these analyses were collected using two separate questionnaires. First, a sin-gle household questionnaire was completed for each household. This questionnaire included a roster ofusual members of and visitors to the household. For each person in the roster, their relationship to thehousehold head, and limited demographic, educational and employment information were ascertained.For each person younger than 15 years, questions were asked about whether her/his biological motherand father were still living, and whether each was a member of the household. In addition, the house-hold questionnaire included questions about household services (including access to electricity, sourceof drinking water, and type of toilet facility), whether the household members owned each item in alist of durable assets (including a television, automobile, and livestock), and an assessment by the inter-viewer of which materials were used in constructing the walls and floor of the household structure. Therespondent for the household questionnaire could be “any adult member of the household who [was]capable of providing information needed” to complete the questionnaire [125].Second, a woman’s questionnaire was completed with each female household member aged 15 to49 years identified in the household roster. This questionnaire included a fertility and marital history,detailed educational attainment and employment questions, and questions about health knowledge, be-liefs and attitudes. Using a birth roster, this questionnaire documented the date of all of the respondent’schildbirths. For each living child under five years of age, details about the child’s delivery, immunizationhistory, and recent health status were ascertained.As a consequence of the two-stage sampling strategy and data collection using these two question-naires, units of observation in the SADHS dataset are organized into four nested levels6: at the top areEAs within which are nested a number of sampled households. In each household there are potentiallymultiple women, each of whom could potentially have a number of children. Data were collected for atotal of 5066 children born in the five years preceding the survey. Of these, 4797 were alive at the timeof the survey.The complete data collected using these questionnaires are supplied in seven different datasets. Forthe analyses described here, we drew variables from the following two datasets:• the “Household member recode” dataset, which has a separate record for every member in thesampled households, and includes sociodemographic variables for each individual and character-istics of the household in which they live; and• the “Children’s recode” dataset, which has a separate record for each child born to the interviewedwomen. This large dataset includes variables specific to each child as well as variables specific tothe mother (i.e.: values are the same for all children born to the same mother) and some variablesspecific to the household (i.e.: values are the same for all children residing in the same household.)We describe how we linked records across these datasets in sub-section 3.4 on page 40.6 Nested means that each lower-level unit is found in just one higher-level unit283.2. 1996 South African census3.2 1996 South African censusThe 1996 South African census was conducted from October to November and aimed to collect in-formation about all persons living in the country on the night of the 9-10th October [126, 127]. Toenable enumeration, the country was divided into approximately 86000 EAs, each comprised of 100-250 households, and demarcated so as to align with municipal boundaries [128]. Enumerators visitedevery household in the country (as well as institutions and hostels) to collect information about theoccupants.For the present analyses, we were interested in combining information from the 1996 census withthat from the SADHS because, while the SADHS collected detailed information on households and thewomen and children living in them, the amount of information collected about EAs was very limited.Conversely, the information collected during the census on (nearly) every household and individual inthe country could be used to construct measures of a variety of characteristics of EAs by aggregating rel-evant information across all of the households or individuals comprising them. Because the list of EAsdesigned for the 1996 Census was used as the sampling frame for the first stage of the 1998 SADHS sam-ple, it is relatively simple to link information across these two data sources. This allowed for analysesof whether EA-level variables (such as unemployment rates) modify associations between child-levelvariables (such as the association between having a co-resident father and having been breastfed forsix months or longer). That the census data were collected less than two years before the SADHS datamakes it reasonable to assume that EA characteristics estimated from the census would correspond tothe time period covered by questions in the SADHS.A single census questionnaire was completed for each household7. The questionnaire consistedof 15 questions about the household (such as the type of dwelling, dwelling ownership, and types ofservices used by the household) and 50 questions about each household member (including their age,population group8, educational attainment, employment status, and income) [129]. Household memberscould choose to complete the questionnaire either by themselves or through a face-to-face interview withan enumerator, with the majority opting for the second option [127]. Separate questionnaires were usedfor people residing in institutions such as prisons and hospitals, and for homeless people. “The Countand How it was Done” report available through the 1996 Population Census website provides moredetailed information about the census methodology [127].We obtained the 1996 Census Community Profile databases through the University of Cape Town’sDataFirst service [130]. These databases come bundled within SuperTABLE version 3.6 software(Time-Space Research, Melbourne), which allows the user to produce cross-tabulations among a setof predefined census variables. One of the available variables is “Geographical Area”, which descendsto the level of EAs, thereby allowing simple aggregate statistics to be produced for every EA. The cross-tabulation results can be exported as comma-separated text files for manipulation in other statistical7For the 1996 Census, a household was defined as “a single person or a group of people who live together for at least fournights a week, who eat from the same pot and who share resources.” [126]8Population group refers to the official groups designated by the Apartheid government: “Asian/Indian”, “Black”,“Coloured”, and “White”. These do not correspond to specific ethnic groups. However, because of their historical politi-cal significance these categories are strongly associated with current socioeconomic status.293.3. Variablessoftware.It is important to recognize that EAs are demarcated to make census enumeration practical, ratherthan because they necessarily represent distinct social groupings in the ‘real world’. Researchers study-ing how place of residence influences individuals’ health and behaviour frequently do treat units ofcensus geography as being reflective of meaningful social groupings, largely because data are mostreadily available at these administrative levels of aggregation. However, the particular level of socialorganization selected for study is important and, therefore, should be justified by theories explaininghow characteristics of social environments influence individuals’ outcomes [131, 132]. In quantifyingand attempting to explain geographic variation in the associations between father’s residential statusand child health outcomes, we have focused on EAs as the geographic level of interest. This decisionwas based primarily on the fact of the shared geography of the SADHS and the census, which madeEAs the most convenient level at which to link individual-level data to group-level data. In present-ing our findings we refer to EAs as “neighbourhoods” because this makes the discussion less abstract.Because of our limited theoretical justification for selecting this geographic level, it is necessary to becautious when interpreting our effect estimates for neighbourhood-level variables and their interactionswith individual-level variables. In particular, associations should not be interpreted as causal. In addi-tion, lack of association at the neighbourhood level does not imply that the construct of interest wouldhave no association at a different geographic level.3.3 VariablesIn the following sections we describe the variables used for our statistical analyses. We first describethe father’s co-residence status exposure variable and the four child health outcomes. We then definevariables measuring other characteristics of child and mother which may act as potential confoundersin our analyses. Next, we describe neighbourhood variables which may modify associations betweenthe exposure and outcomes. In many cases we derived our analytic variables from variables in theSADHS datasets. Often our aim was to reduce the number of levels in complex categorical variables.Occasionally we created a new, more specific variable by combining information from a small set ofrelated variables. More detail on how each variable was derived is given in Table A.1 in Appendix A.1(page 208).ExposureThe father’s co-residence status exposure variable was binary, coded as “1” if the child’s biologicalfather was reported to be a de jure member of the child’s household, and “0” otherwise. We use the terms“co-resident” and “non-co-resident” rather than the more concise “present” versus “absent”, becausethe latter terms carry connotations about fathers’ levels of involvement in child care [24, 76], whichcannot be substantiated by the questions asked in the SADHS. Whereas, the terms “co-resident” and“non-co-resident” focus specifically on what has been ascertained in this survey. We only examinedbiological fathers because the SADHS household questionnaire specifically ascertained whether each303.3. Variableschild’s biological father was a member of her/his household. In contrast, the questionnaire did notascertain whether any other member(s) of the household act as social father(s) to the child. Given thatcare of children by members of the extended family is a feature of traditional kinship systems of manyAfrican cultural groups in South Africa, it is likely that many children in the SADHS dataset do havea social father. However, the effect on child health of residing with a social father cannot be addressedusing this dataset precisely because there is not way to identify who these social fathers are.OutcomesOur analyses treated four separate, binary child health-related outcome variables. Each variable wascoded “1” if the child had experienced the outcome, and “0” if not. These variables were defined asfollows:1. Was breastfed for six months or longerA count variable in the SADHS dataset records the child’s duration of any breastfeeding in months9.This information was ascertained by three questions in the SADHS women’s questionnaire: Did youever breastfeed [child’s name]?; Are you still breastfeeding [child’s name]?; and For how many monthsdid you breastfeed [child’s name]? Children still breastfeeding at the time of the survey were assigneda breastfeeding duration equal to their age in months.Initially we considered using survival methods to analyze children’s time to breastfeeding cessation.Comparing the Kaplan-Meier survival curves for children with and without co-resident fathers, we ob-served a large and reasonably consistent difference in the probability of still breastfeeding between thetwo groups of children between 2 months and 18 months of age (Appendix A.2 figure A.1 on page 224).Based on this observation, it seemed reasonable to dichotomize the breastfeeding duration data at a pointwithin this age range and model the resulting outcome using logistic regression, rather than to try to fit amore complex survival analysis model to the time-to-event data. We selected six months as the cut-pointfor the dichotomized variable because the greatest reduction in child mortality risk from breastfeedingappears to be during the first six months of life [133]. In addition, at the time of the SADHS, womenwere recommended to exclusively breastfeed for 3-4 months duration, and thereafter to continue breast-feeding with complementary foods for two years or more (A. Behr, personal communication. November25, 2013). Therefore it made sense to examine whether children had been breastfed for at least as longas the period recommended for exclusive breastfeeding.To derive our binary analytic variable we dichotomized the original count variable. First, we cal-culated child’s age in months as the difference between the interview date and child’s birth date (bothrecorded as century month codes [124, p.5]). We coded children younger than six months old as “notapplicable” and excluded them from analyses of this outcome. Among the remaining children, thosewith breastfeeding durations of six months or longer were coded having had the outcome, and children9Although it may have been informative to analyze data on the duration of exclusive breastfeeding, this information wasnot ascertained in the SADHS.313.3. VariablesTable 3.1: South African EPI schedule (1995-1999) and age groups used to derive immunization com-pleteness outcome variableImmunization doses Recommended age ofreceiptLimits of age group Implied range inlower limitPolio(0), BCG Birth 0 – <3 months 0 – 1 monthsPolio(1), DPT(1), HepB(1) 6 weeks 3 – <4 months 2 – 4 monthsPolio(2), DPT(2), HepB(2) 10 weeks 4 – <5 months 3 – 5 monthsPolio(3), DPT(3), HepB(3) 14 weeks 5 – <11 months 4 – 6 monthsMeasles 9 months ≥11 months 10 – 12 monthsNote: EPI=Expanded Programme on Immunization; BCG=Bacillus Calmette-Guérin;DPT=Diphtheria, Pertussis, Tetanus; HepB=Hepatitis B, (#)=Dose number for multi-dose vaccineswho were never breastfed or had breastfeeding durations of less than six months were coded not havinghad the outcome.2. Was completely immunized for her/his age groupA series of nine variables in the SADHS dataset indicate whether children received each dose of theroutine childhood immunizations recommended by the South African Expanded Programme on Immu-nization (EPI-SA): Polio, Diptheria-Pertussis-Tetanus, and Measles10. Based on the EPI-SA schedulein place from 1995 to 1999 [134], we identified five age groups in which children immunized accordingto schedule ought to have received distinct combinations of immunization doses (depicted in Table 3.1on page 32)11. Column three of Table 3.1 gives the limits used in calculating age groups. Columnfour gives the implied ranges in the lower age group cut-offs, taking into account our uncertainty aboutchildren’s ages. We selected age groups such that, for each lower limit, the lowest value in our rangeof uncertainty is just greater than the recommended age of receipt for the batch of immunization dosesthat define that age group. In other words, all children within a particular age group should have hadadequate opportunity to receive all of the immunization doses appropriate for their age group. Childrenwho had received all of the immunization doses appropriate for their age group were coded as havinghad the outcome, and all others were coded as not having had the outcome12.10Hepatitis B was added to the South African EPI in 1995 (whereas, all other dose recommendations in Table 3.1 were inplace prior to 1995), therefore four-year-old children in the 1998 SADHS dataset may not have had the opportunity to receivethis vaccine. In addition, according to the 1998 SADHS full report, “...hepatitis B vaccination had not been adopted as astandard for the whole country at the time of the survey....” [6, p. 121] For these reasons, we did not consider Hepatitis Bdoses when deriving our immunization completeness outcome variable.11Children’s birth dates and interview dates are recorded to the nearest month in the SADHS dataset. Therefore there isuncertainty of ±1 month in calculating children’s ages, making it necessary to use generous lower limits for the age windows.12We treated the following levels of the original variables as equivalent indicators that a particular vaccine dose had beenreceived: 1=yes, vaccination date on card; 2=yes, reported by mother, 3=yes; vaccination marked on card (but date missing).The purpose of doing this was to avoid excluding from our analyses the 4% of children whose health cards were reportedmissing, and the 22% whose health cards were not seen by interviewers.323.3. Variables3. Had symptoms of ARI in the two weeks before the surveyThis outcome was ascertained by asking the mother the following two questions: Has [child’s name]been ill or feverish with a cough at any time in the last 2 weeks?; and When [child’s name] was ill witha cough, did he/she breathe with difficulty or faster than usual with short, fast breaths?Only if a mother answered “Yes” to both of these questions was a child coded as having had theoutcome. This is consistent with the standard case definition for estimating period prevalence of acuterespiratory infection using DHS data [6].4. Had diarrhoea in the two weeks before the surveyThis variable comes directly from the SADHS dataset and was ascertained by asking the mother: Has[child’s name] had diarrhoea in the last 2 weeks? If a mother answered “Yes” to this question the childwas coded as having had the outcome, and not if the she answered “No”. Although mothers were askedadditional information about diarrhoeal episodes (i.e.: number of stools on the worst day of the episode,and whether the stools were bloody), this information was not considered in our outcome definition.This is consistent with the standard case definition for estimating period prevalence of diarrhoea usingDHS data [6].Additional household structure variables examined as effect modifiers of father’sco-residence statusWe examined whether two additional dimensions of household structure modified the association be-tween father’s co-residence status and child health outcomes: i) whether the child’s mother was in amarital union, and ii) whether there were additional adult relatives residing in the household.The first characteristic is measured by a dichotomous variable indicating whether a child’s motherwas married at the time of the survey. We derived this variable from the mother’s “current maritalstatus” variable by collapsing the following categories to form a “currently unmarried” reference group:never married, living like married, widowed, separated, and divorced. To test for effect modification offather’s co-residence status by whether the mother was currently married we included in our regressionmodels an interaction term involving these two variables13.The second characteristic was ascertained using information about the relationship of householdoccupants to the household head in each child’s household. We were particularly interested in comparingchildren with co-resident fathers to those with non-co-resident fathers, with the latter group stratified bywhether they lived with other adult relatives. We classified household members >17 years old as adults,based on evidence from the 2000 South African Time Use Survey showing that 20-39 year olds spendconsiderably more time engaged in economic work and in household maintenance (which includes care13The identity of the mother’s husband was not ascertained in the SADHS. Therefore we were not able to determine specifi-cally whether the child’s mother was married to her/his biological father. We assumed that, where the mother was married andthe biological father was co-resident, the parents were in a marital union. Where the mother was married and the biologicalfather was non-co-resident, the mother’s husband could be the biological father or someone else. This needs to be kept in mindwhen interpreting the regression coefficient for the interaction term.333.3. Variableswork) than 10-19 year olds[135, p. 38]. We were also interested in whether the gender of the co-residentrelatives was important. To make these comparisons, we constructed a 4-category extended householdstructure variable distinguishing children with:• Two co-resident biological parents (± other adult co-resident relatives)• Co-resident mother, non-co-resident father, >1 adult male co-resident relative• Co-resident mother, non-co-resident father, >1 adult female co-resident relative, no adult maleco-resident relatives• Co-resident mother, non-co-resident father, no other adult co-resident relativesFor analyses involving this variable we included children whose biological father was dead because thethe co-residence status of biological fathers was no longer our sole focus. To examine the effect of thisvariable we used it in place of the father’s co-residence status variable in regression models, and left allother covariates unchanged.Potential confoundersTo estimate unbiased associations between paternal co-residence and each child health outcome requiresidentifying and statistically controlling for the effects of potential confounding variables. Followingthe general criteria recommended in epidemiology, we identified potential confounding variables ascharacteristics that are: a) putative causes of the outcome, and b) expected to be associated with (but notcaused by) father’s co-residence status [136, p. 154-7].Mosley and Chen’s and Millard’s models of the socioeconomic determinants of child mortalityguided our consideration of putative causes of the four child health outcomes. To understand how otherinvestigators had operationalized the determinants identified in these theoretical models, we reviewedliterature concerning neighbourhood and household socioeconomic influences on child health in low-income countries [137, 138, 99, 139, 140, 141, 142, 106, 143]. Many of these studies used data fromhousehold surveys similar to the SADHS. To avoid missing determinants that are more specific to eachoutcome, we also made reference to studies and reviews of the main predictors of breastfeeding initiationand duration [144, 33], immunization uptake [145, 146, 147], acute respiratory infection [148], anddiarrhoeal illness [149, 150, 151, 152, 153, 154]. Most, but not all, of these latter articles were specificto low-income countries.To identify characteristics associated with fathers’ probability of residing with their children, wereferred to other South African research on this issue as well as Fein and colleagues’ categorizationof the determinants of marriage and co-habitation [155]. We reviewed this literature previously in thesection describing a systemic ecological model of the influences on fathering (section 2.1.3 on page 10).Based on our review of the above literature and results of exploratory regression models (describedin section 3.5), we included the following child and maternal covariates in regression models for all fouroutcomes. We describe neighbourhood covariates separately in the subsequent section (page 38).343.3. VariablesChild’s characteristics:• Birth order and preceding birth interval - A 7-level categorical variable derived by combininginformation about the number of children the mother gave birth to prior to the index child’s14birth and the time in months between the index child’s birth and the birth of her/his next oldestsibling. Categories were: 1 = First born; 2 = 2nd-4th born, birth interval <24 months; 3 = 2nd-4th born, birth interval 24-47 months; 4 = 2nd-4th born, birth interval >47 months; 5 = 5th+born, birth interval <24 months; 6 = 5th+ born, birth interval 24-47 months; 7 = 5th+ born, birthinterval >47 months. Birth order and birth interval are “maternal factors” identified in Mosleyand Chen’s model, with lower birth order and shorter preceding birth interval previously shown tobe associated with increased mortality risk for the index child [142, 156]. Higher parity and shortbirth intervals are also associated with more constrained household economic resources [157],which negatively predicts union formation and positively predicts union instability.• Place of delivery - A 3-level categorical variable identifying the location where the mother deliv-ered the index child. Categories were: 1 = Home; 2 = Public medical facility; 3 = Private medicalfacility. This variable was derived by collapsing the sub-categories in the original SADHS vari-able into the “major categories” defined in the dataset documentation [124]. Children reported tohave been born in an “Other” place of delivery (n=35) were treated as having missing data for thisvariable. Place of delivery is a proxy for the “environmental contamination” and “personal illnesscontrol” categories in Mosley and Chen’s model, in that pathogen exposure and availability ofmedical interventions differ between home and medical facility birth environments. Deliveringin a private medical facility would be more likely among women with greater income or wealthat the time of the child’s birth, which, in turn, my positively predict being in a marital union.Whereas, lower income women would be more likely to deliver at home or in a public medicalfacility, because user fees for maternity care at public facilities in South Africa were removed in1995 [7, p. 835].• Antenatal care provider - A 3-level categorical variable identifying the person who providedantenatal care to the mother during the pregnancy with the index child. Categories were: 1 =Nurse/midwife (with or without a doctor); 2 = Doctor only; 3 = Traditional birth attendant, othercare provider or no care. This variable is related to the mother’s access to health services, hersocioeconomic status during the pregnancy and her health care beliefs, all of which feature inmodels of the determinants of child mortality. Traditional health beliefs may co-occur with moretraditional beliefs favouring marriage. Whereas lower personal income as well as living in apoorer-resourced area may be negatively associated with marriage/cohabiting and positively as-sociated with cohabiting union disruption, for example from men’s entry into migrant labour. InSouth Africa antenatal care from government primary health clinics is usually provided by nursemidwives. We assumed that women who received antenatal care from a doctor only are likely14By “index child” we are referring to a child who is the unit of analysis in our models, i.e.: a child meeting the inclusioncriteria and for whom exposure, outcome, and covariate information were available.353.3. Variablesto have purchased this care privately. Also note, there is not a professional ‘Traditional BirthAttendant’ designation in South Africa. Traditionally, women would deliver alone or attended bya family member. We assumed that women who sought care from a traditional birth attendant orother care provider or who did not have any antenatal care were more likely to have experiencedbarriers to accessing medical facilities or were following specific beliefs that disfavoured medicalcare.• Whether the mother wanted the pregnancy with the index child at the time she becamepregnant - This variable was used directly as provided in the SADHS dataset. Response optionsfor the timing when the mother wanted to become pregnant were: 1 = Then; 2 = Later; 3 = Didnot want any more children. Women reporting unintended pregnancies are less likely to receiveadequate antenatal care, and children born of unintended pregnancies have lower probability ofreceiving routine childhood immunizations and increased probability of dying during the neonatalperiod [158]. Unintended pregnancy is also associated with decreased union stability. In addition,pregnancies occurring outside of a union (which may relate to whether the pregnancy is wanted)influence subsequent marriage and cohabiting union formation.Mother’s characteristics• Mother’s population group - A dichotomous variable (0 = Black; 1 = Non-black) derived bycombining the “Asian/Indian”, “Coloured” and “White” population groups to form the “Non-Black” category. During Apartheid, discriminatory laws controlled non-White population groups’rights to own property, job availability, wage levels, access and quality of services like educationand health care, as well as power and perceived status in society. As a consequence, populationgroup is persistently and strongly associated with socioeconomic status. Higher socioeconomicstatus is expected to improve children’s health outcomes and parent’s opportunities for marriageand/or cohabitation.• Mother’s highest completed level of education - A 3-level categorical variable (1 = No edu-cation or incomplete primary; 2 = Complete primary or incomplete secondary; 3 = Completesecondary or higher) derived by collapsing adjacent pairs of categories in the original six-levelvariable. Educational attainment is an important “socioeconomic determinant” in Mosley andChen’s model, with higher maternal educational attainment having been found to be associatedwith lower risk of various negative child health outcomes [142, 157]. Research shows that a ma-jority of Black South African women share the belief that marriage should come after completionof education [159, 61], suggesting that lower educational attainment may be associated with lowerprobability of being in a marital union.• Mother’s age at index child’s birth - A count variable derived by calculating the differencein years between the index child’s date of birth and mother’s date of birth. Mother’s age is animportant influence on children’s health outcomes, being related to knowledge and experience,363.3. Variablesand economic resources. Increasing age is also associated with increased likelihood of marriageand cohabitation [59].• Mother’s age at first child birth - A 3-level categorical variable (1 = <18 years old; 2 = 18-29 years old; 3 = >29 years old) derived from the continuous “age at first birth” variable inthe SADHS dataset. Early childbearing is associated with increased probability of low birthweight and infant mortality [160]. It is also associated with poorer economic outcomes for themother (school drop-out and household poverty) [161], which are, in turn, associated with poorerchild health outcomes. Early, non-marital pregnancy also reduces subsequent marriage prospects,although it may promote cohabitation.For models of breastfeeding duration and immunization completeness we also adjusted for a 4-levelcategorical variable measuring whether the mother spent her childhood in a rural or urban areaand whether she had migrated between childhood and the time of the survey. This variable wasderived by by creating a category for each combination of the levels of two dichotomous variables inthe SADHS dataset: mother’s childhood place of residence (with “City” and “Town” categories treatedtogether as “Urban” vs. “Countryside”), and type of place of residence at the time of the interview(“Urban” vs. “Rural”)15. This variable is intended to capture some of the influence of mother’s place ofupbringing on her later-life socioeconomic status and on her attitudes and beliefs. We are also interestedin capturing the effects of urbanization, which has been an important demographic change in SouthAfrica, especially since the end of Apartheid-era pass laws. Research suggests that moving to an urbanenvironment exposes women to better educational and employment opportunities and may reduce thesocial pressures to conform to traditional marriage norms [162, p. 93-4]. Urban women may be morelikely than rural women to live independently from, or in a cohabiting but non-marital union with, theirchildren’s fathers [52]. They may also have improved access to resources for keeping their childrenhealthy. Alternatively, women may migrate with their partners to urban centres, avoiding the commonpattern of migrant labour-induced familial/spousal separation.For models of immunization completeness, recent ARI and recent diarrhoea, we also adjusted for acount variable measuring the child’s age in months, which was derived as the difference between thedate of the interview and the child’s date of birth. Children’s risk of respiratory and diarrhoeal infectiondeclines with age [148, 153]. Children are also less likely to receive vaccine doses that are scheduledfor receipt at older ages [163]. Therefore, as children age, they are less likely to have received allimmunizations appropriate for their age-group. Finally, we anticipated that the probability of having aco-resident father might vary as children age.Finally, for models of recent ARI and recent diarrhoea we also adjusted for a dichotomous variablemeasuring child’s sex (0 = Female; 1 = Male), and a 3-level categorical variable measuring the seasonduring which the interview was conducted (1 = Summer [January-March]; 2 = Autumn [April-May];3 = Winter [June-September]). Male children have a higher incidence of diarrhoeal disease than females,and possibly a slightly higher incidence of respiratory illness [164, 148]. Female sex of first-born15For the purposes of this variable, “migration” means having moved from a rural childhood place of residence to an urbanplace of residence by the time of the survey, or vice versa373.3. Variableschildren has been found to be associated with reduced marriage quality [155, p. 13], while male childrenin the United States are more likely than females to have co-resident fathers [165]. We included seasonas a covariate in the models for the recent illness outcomes because incidence rates of respiratory anddiarrhoeal infections vary by season [148, 153] and because paternal non-co-residence associated withmigrant labour may also vary by season.Of note, we did not include sex as a covariate in the immunization completeness models. Althoughstudies of childhood immunization coverage in India have found that boys are more likely than girlsto be completely immunized [166], studies in African settings have not reported gender-differences inimmunization coverage [167].Neighbourhood-level variablesFathers’ parenting practices have been found to be sensitive to societal context [168, p. 50][47, p.285][26, p. 889]. As discussed in the literature review (Section 2.1.3 on page 10), employment avail-ability and gender norms are two important features of the context of fatherhood in South Africa, whichmay modify the effect of residing with a father on children’s health outcomes. Direct measures of em-ployment availability and gender norms are not available for the neighbourhoods in the SADHS dataset.However, by aggregating 1996 census data from individuals and households comprising the neighbour-hoods, we may be able to obtain indirect measures of the characteristics of interest. We hypothesize thatvariables measuring aspects of neighbourhood social context may explain neighbourhood-level varia-tion in co-resident father effects on children’s health outcomes. We explored the following variables forthis purpose:1. A dichotomous variable distinguishing urban from rural neighbourhoods2. Unemployment rate among adult male residents - A continuous variable calculated as thenumber of male EA residents 15 years old or older who were “unemployed, looking for work”expressed as a percentage of all male EA residents 15 years old or older who were either employedor unemployed, as per the “expanded definition” of unemployment used by Statistics South Africa[126, p. 2].3. Percentage of households having a female head - A continuous variable calculated as the per-centage of households in the EA who reported having a female household head.4. Percentage of households having an annual income <R600016 - A continuous variable calcu-lated as the percentage of households in the EA with derived annual household incomes in the“R2401-6000” range or lower. The chosen income cutoff roughly corresponds to the the upperlimit for the lowest quintile of household incomes nationally, as calculated in the 1995 SouthAfrican Income and Expenditure Survey [169, p. 28].16R6000 was equivalent to approximately $1900 CAD at the average exchange rate for the 1996 calendar year.383.3. Variables5. Percentage of female residents having completed high school education or higher - A contin-uous variable calculated as the percentage of female EA residents 20 years old or older who hadcompleted high school, including those with any level of post-high school education.6. Ratio of percentage female to percentage male residents having completed high school edu-cation or higher - A continuous variable calculated as the percentage of female EA residents 20years old or older who had completed high school divided by the percentage of male EA residents20 years old or older who had completed high school.Employment availability and cultural norms are expected to vary between urban and rural neighbour-hoods. Employment opportunities are expected to be more limited in rural neighbourhoods, neigh-bourhoods with higher rates of male unemployment, and neighbourhoods with greater percentages oflow-income households. Higher percentages of female headed households may indicate neighbourhoodswith higher numbers of absent male migrant labourers.Choosing to leave a neighbourhood offering limited opportunities to provide financially for one’schildren may reflect a responsible fathering decision and may benefit children’s health. In addition,if economic marginalization makes men less able to meet their fathering responsibilities, even whenthey physically reside with their children, having a co-resident father in these neighbourhoods may,on average, be of limited benefit to children. Furthermore, we expect that neighbourhoods with bet-ter employment opportunities and fewer low-income households would tend to have better access togood-quality public infrastructure and services17. Through their access to services, parents in theseneighbourhoods may be more able to ensure positive health outcomes for their children. We hypothe-size that in neighbourhoods with less unemployment and lower percentages of low-income households,having a co-resident father will be more strongly associated with positive child health outcomes.With respect to cultural norms, in neighbourhoods that empower women and emphasize greatergender equality, men and women may have more space to negotiate their obligations and responsibilitiesto one another and to their children. Neighbourhoods with more permissive gender norms are likely todisplay higher absolute and relative (compared to men) percentages of women having completed highschool. Urban neighbourhoods and those with greater percentages of female households heads may alsohave less traditional norms. In such neighbourhoods, fathers may become more involved in the intimatecare of their children [52, p. 77, 83]. In addition, women may be able to negotiate more control overdecision-making related to their children’s care [43, 60]. For these reasons, having a co-resident fathermay be more beneficial in these contexts.17Exploratory analyses support this assumption: using data from all neighbourhoods in the 1996 census dataset (results notshown), we found the percentage of households with incomes <R6000 to be modestly negatively correlated with the followingfive indicators of neighbourhood access to public services: percentage of households with water piped into dwelling or site,percentage of households having a flush or chemical toilet, percentage of households using electricity for lighting, percentageof households with a telephone in the dwelling, and percentage of households having refuse disposal services provided by thelocal authority (Spearman’s rank correlation coefficients between -0.32 and -0.40).393.4. Inclusion and exclusion criteria and data linkageAdditional variables used to describe the analytic sampleHousehold socioeconomic status is another important potential confounder in these analyses because itis intimately associated with household structure and affects child health through a number of pathways.Investigators using household survey data often operationalize household socioeconomic status using acombination of measures of household members’ educational attainments and employment statuses, andhousehold income, expenditure or wealth. The SADHS dataset includes household wealth variables, butthese cross-sectional measures only reflect the situation at the time of the survey. Because father’s co-residence status (also measured only at the time of the survey) could be a cause or a consequence ofhousehold socioeconomic status, adjusting for the latter could bias the co-resident father regressioncoefficient estimates (as a result of controlling for variables that mediate some of the effect of father’sco-residence on child health)18 [83, p. 358][77, p. 33][40, p. 12-13]. For this reason we do notadjust for these variables in our regression models but we do include them in the table of descriptivestatistics (Table 4.1). The two socioeconomic status variables we present are: (i) a dichotomous variablemeasuring whether the mother was working at the time of the survey (0 = No; 1 = Yes); (ii) and a5-level categorical variable measuring the quintile of household wealth, which was derived from aseries of questions about household dwelling structure and ownership of durable assets, as describedpreviously [170].In our descriptive statistics tables we also included an extra 9-level categorical variable measur-ing the Province in which the neighbourhood is located. This variable is of interest because it likelyinfluences both father’s co-residence and children’s health outcomes. Vastly differing employment op-portunities and lasting effects of the Apartheid government’s displacement of African people to rural“homelands” in some Provinces likely contribute to between-Province differences in the percentageof children with co-resident fathers. Differences in Provincial public infrastructure and health systemfunctioning likely contribute to variations in children’s health outcomes. However, we did not includethe Province variable in our regression models for the following reasons: a) we found it to be highlycorrelated with some of the individual-level potential confounders, and b) there are relatively few neigh-bourhoods per Province in our analytic sample. Both lead to large uncertainty in coefficient estimatesfrom models which include Province as a covariate.3.4 Inclusion and exclusion criteria and data linkageTo be eligible for inclusion in our analyses children had to be alive at the time of the survey and residingwith their mother. A total of 4,437 children in the child recode dataset met these criteria.To estimate associations between having a co-resident father (recorded for each child in the house-18Ideally we would adjust for household socioeconomic status at conception or during pregnancy. Not only does this precedefather’s co-residence status at the time of the survey, pregnancy is also likely to be a critical period in the processes leading towhether parents co-reside [155]. Unfortunately, household socioeconomic status during pregnancy was not directly assessedin the SADHS. We assume that mother’s population group and educational attainment, as well as antenatal care provider andplace of delivery will indirectly measure socioeconomic status during pregnancy. This is part of our rationale for adjusting forthese variables.403.4. Inclusion and exclusion criteria and data linkagehold member dataset) and children’s health outcomes (recorded for each child in the child dataset) itwas necessary to link each child’s record across the child and household member datasets. Optimally aunique child identifier that is common to both datasets would be used to link children’s records. Thiswould ensure accuracy of the linked data. Unfortunately the SADHS datasets do not include a uniqueidentifier for children. An alternative is to link pairs of records by matching their values on a set ofvariables present in both datasets. There are five variables common to the child and household mem-ber datasets: a unique number for each EA, a unique number for each household, a unique identifierfor each woman selected to complete the women’s questionnaire (referred to subsequently as mother’sidentifier), and the sex and age (in years) of each child. We tested two alternative matching strategies.One strategy was more conservative in that only those records matching on all five common variableswere linked. However, a concern with the conservative strategy is that it would fail to link records ofchildren whose ages and/or sexes were accidentally recorded differently in the household member andchild datasets.19 To overcome this limitation, the alternative strategy allowed records to be linked ifthey matched on only three or four of the five common variables, provided each record could only bematched to one record in the other dataset. At a minimum, linked records had to match on EA number,household number, and mother’s identifier. In other words, we assumed that if each dataset containedonly a single record for a child under 5 years old residing with the same uniquely identified mother thatthose records could be linked, even if the child’s age and/or sex variables did not match. We provide adetailed description of the two linking strategies and flow diagrams comparing them in Appendix A.3on page 225.We compared the datasets generated by the two linking strategies by assessing for differences in thedistributions of the exposure, outcomes and covariates of interest in each dataset. Because we foundno significant differences between the two datasets in the distributions of these variables, we used thedataset produced by the less conservative strategy for our analyses. The latter dataset was comprised oflinked records for 4010 children (90.4% of the 4,437 who met our inclusion criteria).To produce the final analytic sub-sample, we excluded 48 children because either child or motherwas recorded as a visitor to the household or because both the child’s and mother’s values of the house-hold resident status variable were missing. We excluded a further 152 children whose biological fatherwas dead at the time of the survey. Although it would be valuable to examine the health outcomes ofchildren whose fathers have died, the factors leading to, and consequences of, fathers’ deaths may bedistinct from those leading living fathers to reside apart from their children. As such, treating childrenwith dead fathers as equivalent to children with living, non-co-resident fathers in our analyses could biasour estimates of the effect of co-residing with a father. In addition, there were too few children in this19Children’s data in the household member dataset were collected using the household questionnaire, which could havebeen completed by any adult household member, not necessarily a close relative of the child. In contrast, data in the childdataset were reported by the child’s mother. In the former dataset children’s ages were collected simply by asking how oldchildren were at the time of the survey, possibly allowing ages to be rounded up or down at the discretion of the respondent.Whereas, in the child dataset, mothers reported the year and month of their children’s births, from which age in completedyears was calculated. Because of this difference in the way children’s ages were obtained, it is plausible for the same child tohave their age recorded differently in the two datasets. However, we would expect most discrepancies to be of only one year.In the dataset produced by our second linking strategy 91% of records having a mis-match on the child’s age were discrepantby only one year.413.5. Statistical analysisgroup to treat them as a separate group during analysis. For these reasons we excluded them from ourdataset. This leaves a total of 3,810 children in the analytic sub-sample, 85.9% of the total number ofchildren in the SADHS who met our inclusion criteria.Finally, we linked records of children in the analytic dataset to corresponding records for the neigh-bourhoods where they live. The latter records were included in a dataset of neighbourhood variablesthat we derived using 1996 SA census data (described in Table A.1 on page 208). The linkage involvedmatching records using a unique EA identifier. Each EA in the 1996 census is identified by a unique7-digit identifier. The same identifier for EAs in the SADHS dataset was obtained by multiplying the3-digit “district ID” variable by 10000 and added the 4-digit “EA number” (S.O. Manda, personal com-munication, April 17, 2012). We failed to link the records of 39 children in the analytic sample (1% ofthe total) because their EA identifier did not have a match in the census dataset. These children werefrom 10 EAs in the SADHS dataset. Exploratory analyses showed that two of the EAs were in theWestern Cape (one urban, one rural), one was in a rural area of the Eastern Cape, and the remainingseven were in urban areas of Gauteng (results not shown). We excluded these 39 children with missingneighbourhood-level data from all of our analyses.3.5 Statistical analysisThe following sections describe our approach to preparing initial descriptive statistics and to fitting aseries of multilevel regression models to address research questions 1a-e (Chapter 1).Descriptive data analysisFor all outcome, exposure, and potential confounding variables we prepared descriptive statistics withthe aims of understanding: a) the distribution of each variable in the analytic sample, and b) the pro-portion of missing responses. We examined categorical variables using single-variable frequency tables.For continuous variables, we calculated the percentages of records having missing data and prepared his-tograms of the valid response data. We calculated the following measures of central tendency and spreadfor each continuous variable: means and standard deviations for approximately normally-distributedvariables, and medians and interquartile ranges for non-normal variables. Assessment of normality wasbased on visual inspection of histograms.As a first step in determining whether the associations hypothesized from the conceptual model andliterature review were supported by the data, we examined associations among pairs of variables usingcontingency tables for categorical variables; Pearson and Spearman correlation statistics for continuousvariables; and, for pairs of continuous and categorical variables, means, medians, standard deviations,and interquartile ranges of the continuous variables within each stratum of the categorical variables.In the results chapter (Table 4.1 on page 50), we present population-weighted descriptive statistics forthe complete analytic sub-sample and for strata of children with and without co-resident fathers. Forpopulation weighting we used the person weights provided with the SADHS datasets.To describe the multilevel structure of the analytic sub-sample, we prepared frequency distributions423.5. Statistical analysisof the counts of children per mother, household and neighbourhood. As a gross assessment of whetherthe risk of each outcome is unevenly distributed across neighbourhoods, we prepared frequency dis-tributions for counts of each outcome per neighbourhood. Using these distributions, we can examinewhether outcomes appear to be ’concentrated’ in a subset of neighbourhoods.Multilevel regression modellingAnswering our research questions requires us to be able to isolate individual and neighbourhood sourcesof variation in the observed child health outcomes. Multilevel modelling is a powerful method for doingthis. The feature which distinguishes multilevel models from ordinary regression models is their abilityto model variation among lower-level units simultaneously with variation across higher-level units [171,p. 251]. Given the aims of this study, multilevel modelling afford the following advantages:• It allowed us to model neighbourhood-level variation in the regression intercept via treating theintercept as a random variable. This has two advantages. First, ordinary regression assumes thateach child has an independent probability of having the outcome. The complex sampling de-sign used in the SADHS produces data that likely violate this assumption because children whowere living in the same neighbourhood may have been more alike one another in their probabilityof having the outcome than children who were living in different neighbourhoods. Modellinggroup-level variation in the intercept allowed for outcomes of individuals in the same group to becorrelated, thereby relaxing the independence assumption. The second advantage of including in-tercepts that vary by group was that they “absorbed” unexplained variation in the individual-levelmodel that could be explained by group membership. In this way, varying intercepts accountedfor the effect of unmeasured group-level confounding variables.• We were able to model group-level variation in the regression co-efficient for the father’s co-residence status variable. This allowed us to estimate the magnitude of variation across neigh-bourhoods in the effect of having a co-resident father, and to explore the degree to which variousneighbourhood-level covariates explained this variation.The SADHS data are arranged in four nested levels: children > mothers > households > neighbourhoods.However, the sample sizes at the "mother" and "household" levels were too small in the hierarchicalstructure to facilitate useful inference at these levels of clustering. This can be seen in that fact that themajority of mothers and households have only a single eligible child (figure 4.1 on page 55). Therefore,for the analysis of this data we developed two-level (random intercept and coefficient) models withchildren nested under neighbourhood. We utilized maternal and household characteristics as individuallevel data.Our analyses proceeded in four stages. First, we estimated unadjusted odds ratios (ORs) for theassociation between covariates of interest and each child health outcome. These ORs were estimatedusing generalized linear models (GLMs), each with a single predictor variable. In the second stagewe fit a series of GLMs with father’s co-residence status and different sets of potential confounders.These models allowed us to assess the effect of confounder adjustment on the regression coefficient for433.5. Statistical analysisfather’s co-residence status. We also simultaneously modelled groups of potential confounders appear-ing to measure related characteristics (for example mother’s educational attainment and self-reportedliteracy). We reviewed the estimated correlation matrices of these models to identify covariates show-ing multicollinearity. Last, we tested interaction terms between father’s co-residence status and child’sage and mother’s population group. Using the Wald test, these interaction terms were not found to besignificant at the 5% level and so were not included in subsequent models.For the final individual-level models, we based our selection of potential confounders predominantlyon the findings of our literature search for variables having demonstrated associations with exposure andoutcome, and on our assessments for multicollinarity, rather than on the basis of statistical significance.Once we had decided on the final set of covariates for each adjusted model, we prepared analytic datasetshaving complete data for all included variables. The total number of records in these datasets varies byoutcome because of missing responses for the outcome variables themselves and because of differencesin the covariates included in the full models. We then refit the unadjusted models using the completecase datasets.Each GLM expresses the expected value, or mean, of the outcome variable as a linear combinationof explanatory variables and associated regression parameters (intercepts and slopes) [172, 173, 171]. Alink function is used to connect the mean of the outcome to the linear predictor [172]. All four outcomesin this study are binary (i.e.: for each child, the outcome, y, can only take on a value of 1=yes or 0=no).We assumed that the outcomes follow a Bernoulli distribution, a special case of the binomial distributionin which there is a single trial (i.e.: each child has only a single opportunity to have the outcome). Toaccommodate the binary outcomes, we used a log-odds or logit link function.Therefore, we modelled the log-odds of pii j, the probability that the outcome occurs for the ith childin the jth neighbourhood, as:logit(pii j) = log(pii j1−pii j)= β0 +β1x1i j +Xi jB (3.1)In the above equation, β0 is the intercept and β1 is the effect for the co-resident father variable,x1i j. The matrix Xi j is composed of children’s values for all additional covariates in the model. Columnvector B contains the regression coefficients for the additional covariates.In the third stage of our analysis, we constructed multilevel generalized linear models, in which weallowed some of the regression coefficients to vary by neighbourhood. The simplest multilevel modelinvolves an intercept that varies by neighbourhood and is otherwise identical to equation 3.1. This modelcan be written as:logit(pii j) = β0 j +β1x1i j +Xi jB, for i= 1,...,n and j= 1,...,Jβ0 j = β0 +u0 ju0 j ∼ N(0,σ2u0)(3.2)443.5. Statistical analysiswhere β0 j is the intercept, which varies by neighbourhood (as indicated by the subscript j). Inmultilevel models, the varying regression coefficients are modelled using a separate neighbourhood-level model which is fit simultaneously with the individual-level model20 [171, p. 251]. In the simple 2-level model shown above, the varying intercepts are modelled using a linear regression with an intercept,β0, and a residual term, u0 j. In this regression, β0 is the average of the neighbourhood intercepts, whilethe residual can be thought of as measuring each neighbourhood’s deviation from the average intercept.The neighbourhood-specific deviations were assumed to be normally distributed around a mean of zerowith a variance of σ2u0, which is estimated from the data.We used varying intercept models to explore the influence of unmeasured neighbourhood con-founders on the coefficient for father’s co-residence status. We also explored the effect of includingan interaction term involving father’s co-residence status and mother’s marital status. Finally, we re-placed the father’s co-residence status variable with the four-level household structure variable, to assesswhether living with additional relatives modified the effect of having a non-resident father.In model 3.2, the regression coefficient for the co-resident father predictor variable is fixed, i.e.: itsvalue is assumed not to vary by neighbourhood. To relax this assumption, we can extend model 3.2 byallowing the co-resident father coefficient to also vary by neighbourhood:logit(pii j) = β0 j +β1 jx1i j +Xi jBβ0 j = β0 +u0 jβ1 j = β1 +u1 j(u0 ju1 j)∼ MVN(0,Ω)Ω=(σ2u0 ρσu0σu1ρσu0σu1 σ2u1)(3.3)Similarly to the neighbourhood intercepts, the neighbourhood slopes are modelled using a linearregression without predictors. The intercept, β1, is equal to the mean slope across all neighbourhoods,and u1 j is the neighbourhood-specific deviation from the mean slope. We assumed that the deviationsfor the intercepts and slopes followed a joint multivariate normal distribution having mean zero and co-variance matrix, Ω, estimated from the data. The diagonal components of the covariance matrix are thevariance terms for the varying intercepts and slopes, σ2u0 and σ2u1, respectively. The off-diagonal compo-nents, ρσu0σu1, are equal to the covariance between the varying intercepts and slopes. The assumptionof a joint distribution allows for the varying intercepts and slopes to be correlated. We calculated thecorrelation, ρ , by dividing the covariance by the product of the intercept and slope standard deviations.In the final step of our analyses, we included neighbourhood-level covariates. To do this, we in-cluded the neighbourhood covariates as predictors in the linear regression models for the varying inter-20By individual-level we are referring to units at the lowest level of clustering, i.e.: children in this case.453.5. Statistical analysiscepts and slopes. Below we present an example model that includes a neighbourhood-level covariate,z1, which, for illustration, could be the percentage of households with an annual income <R6000:logit(pii j) = β0 j +β1 jx1i j +Xi jBβ0 j = β0 +u00 j +u01z1 jβ1 j = β1 +u10 j +u11z1 j(u00 ju01 j)∼ MVN(0,Ω)Ω=(σ2u0 ρσu0σu1ρσu0σu1 σ2u1)(3.4)In this model, z1 j is the percentage of low-income households in neighbourhood j. In the regressionfor the varying intercepts, β0 is the neighbourhood-average intercept and u01 is the coefficient for thelow-income household covariate, which does not vary by neighbourhood. Similarly, in the regressionfor the varying slopes, β1 is the neighbourhood-average slope and u11 is the fixed coefficient for thelow-income household covariate. The neighbourhood-specific deviation terms are given as u00 j andu10 j. If the covariates explain any of the deviation of the neighbourhood intercepts and slopes abouttheir respective means, we should see a corresponding decrease in the estimated variance parametersσ2u0 and σ2u1.In the model for the neighbourhood intercepts, the regression coefficient, u01, is the change in theintercept observed when comparing neighbourhoods that differ by one standard deviation in their valueof the neighbourhood covariate. In the model for the neighbourhood slopes, the coefficient, u11, can beinterpreted in two ways. Similarly to the term in the model for the intercepts, it can be interpreted as thechange in the slope observed in comparing neighbourhoods that differ by one standard deviation in theirvalue of the covariate. Alternatively, it can be interpreted as the coefficient for an interaction involvingthe value of the father’s co-residence status variable for child i and the value of the covariate for theneighbourhood where child i lives [171, p. 282-3]. The latter interpretation is more easily seen bysubstituting line 3 of model 3.4 into line 1. By either interpretation, the magnitude of this term indicateswhether the effect of having a co-resident father is modified by neighbourhood context.Model estimationThe varying coefficients in multilevel models make these models difficult to solve mathematically.However, model parameters can be estimated using procedures that maximize approximations to thelikelihood or using Bayesian Markov Chain Monte Carlo (MCMC) techniques [174, p. 128-31]. Resid-ual pseudo-likelihood is a simple and widely available method for estimating parameters of multilevelmodels [174, p. 130]. However, this method has been found to produce biased estimates of the varianceparameters for varying intercepts and slopes in models involving binary outcomes [175]. Two other463.5. Statistical analysisestimation procedures, Laplace Approximation and Gauss-Hermite Quadrature, are more accurate thanpseudo-likelihood but less able to fit complex models, such as ones including multiple random coeffi-cients. MCMC is both accurate and able to fit complex models, however it is time-consuming and bringsthe conceptual change of working within a Bayesian inferential framework [174, 176, p. 130].Prior to selecting a method for our regression analyses, we applied the four estimation proceduresdescribed above to fit a test model. The test model outcome was a binary indicator for whether childrenhad ever been breastfed and the model included a single predictor (the co-resident father variable) andvarying intercepts at household- and neighbourhood-levels. Using default settings, the Gauss-HermiteQuadrature method did not converge, so we were only able to compare the estimates from pseudo-likelihood, Laplace, and MCMC methods (results not shown). The co-resident father co-efficient es-timates were reasonably similar across the three methods. Whereas the estimated variance parametersfor the group-level intercepts were quite different. Given the ability to handle complex models and theaccuracy of the parameter estimates, we chose to use MCMC for our analyses.Using a Bayesian approach, parameters are estimated using the likelihood function for the observeddata, conditional on the unknown parameters, in combination with prior information about plausibleparameter values [177, p. 46]. The plausible values are specified using a prior distribution. Modelparameters are conceptualized as random variables having probability distributions. After the modelis fit, the parameter probability distributions are called posterior distributions. MCMC methods canbe used to estimate the posterior distributions. These methods work by iteratively sampling from thejoint posterior distribution of the model parameters. The serially correlated random sample values arecalled ‘chains’. After a sufficient number of samples have been drawn, and assuming the chains haveconverged (i.e.: further sampling will not dramatically change the estimates), it is possible to calculatesummary statistics for each of the model parameters (e.g.: means and 95% credible intervals). Bydefault, WinBUGS implements MCMC methods using a Gibbs sampling algorithm or a Metropolis-within-Gibbs algorithm for more complex conditional posterior distributions [178]. These algorithmsare described elsewhere (for example [177]).Statistical software, MCMC settings, and specification of prior distributionsWe used SAS version 9.3 to perform dataset manipulation (merging of records across datasets, deriva-tion of variables, selection of the analytic sub-sample) and to calculate descriptive statistics. We usedWinBUGS version 1.4.3 (Imperial College and MRC, United Kingdom) to implement Bayesian MarkovChain Monte Carlo methods for estimating our regression parameters. We used an open source statisticalsoftware package, R (version 2.14.1, The R foundation for Statistical Computing), and a complementaryuser interface, RStudio (version 0.97.551, RStudio, Inc.), to prepare the datasets from SAS for use inWinBUGS. Specifically, we used the writeDatafileR function to output datasets in array format [179].The MCMC implementation was carried out with weakly- or non-informative priors so that param-eter (posterior density) estimation was driven predominantly by the data. The estimates produced usingthis approach are typically similar to those from maximum likelihood-type methods [177, p. 46]. Forall non-varying regression coefficients we used normal prior distributions with mean of 0 and standard473.5. Statistical analysisdeviation of 100. To allow for correlation between the varying intercepts and slopes we modelled themusing a multivariate normal prior and an inverse-Wishart hyperprior with 2 degrees of freedom for thecovariance matrix. We tested models involving independent normal distribution priors for the neigh-bourhood intercepts and slopes, with uniform hyperpriors for the standard deviation parameters rangingfrom 0 to 100. We also tested a scaled inverse-Wishart hyperprior with 3 degrees of freedom for the co-variance matrix of the intercepts and slopes as recommended by Gelman and Hill [171, p. 376-7]. Thesetest models showed poor chain mixing for the neighbourhood slopes standard deviation parameter.In all MCMC runs, we used three chains. We determined whether chains had converged in two ways:a) by visually inspecting the chain sampling histories for good mixing, and b) by using ‘BGR diagrams’(prepared automatically by WinBUGS), which compare within-chain to between-chain variability acrossincreasing fractions of the total simulation run [178]. We identified convergence in the diagrams as anR value < 1.1 and reasonably stationary values for the estimates of between chain and within chainvariability.Our models appeared to converge within 2000 sampling iterations, so, by default, we discarded thefirst 2000 iterations when calculating chain statistics. We began monitoring the deviance informationcriterion following the first 2000 iterations. In the remaining iterations, we set thin to 10 (i.e.: we storedevery 10th sample). This level of thinning appeared to produce low auto-correlation between samplesfor most parameters. We found that running 7000 iterations per chain (i.e.: a total of 15000 samplesafter discarding the first 2000 from each chain) produced reasonably smooth posterior density plots, sowe used this as our default run length. In our results tables we report means and 95% credible intervalsfor the posterior densities.48Chapter 4Results: Effects of Father’s Co-residenceStatus on Child Health Outcomes4.1 Description of the analytic sub-sampleA total of 3810 children contributed data to these analyses. Of these children, 43% had co-residentfathers and 56% had non-co-resident fathers (the remaining 1% had missing values for father’s co-residence status). The top panel of Table 4.1 on the next page shows the numbers of mothers, householdsand neighbourhoods from which these children came, and the average number of children per higher-level unit. The average household and mother contributed a single child to the analysis, indicating thatmost children under 5 years old in South Africa did not reside with another child in this age-group.Interestingly, the average number of children per neighbourhood was higher for children with non-co-resident fathers (3.5) than for children with co-resident fathers (2.4). Another way to explore the multi-level structure of the dataset is to look at the frequency distribution of counts of children per mother, perhousehold, or per neighbourhood. Figure 4.1 on page 55 visually depicts these frequency distributionsfor the total analytic sub-sample. Approximately 15% of mothers and 20% of households in the samplehad more than a single child. Conversely, 81% of neighbourhoods in the sample had more than oneresident child and 48% had more than three resident children. Only about 5% of neighbourhoods hadmore than 10 resident children.494.1. Description of the analytic sub-sampleTable 4.1: Descriptive statistics for outcomes and potential confounding variables in the complete sam-ple of children and separately for strata of children with and without co-resident fathers, 1998 SADHS.All children,N=3810Children withnon-co-residentfathers, n=2141Children withco-residentfathers, n=1638CLUSTERING n (Ratio) n (Ratio) n (Ratio)Mothers 3271 (1.2) 1887 (1.1) 1355 (1.2)Households 3023 (1.3) 1701 (1.3) 1343 (1.2)Neighbourhoods 862 (4.4) 615 (3.5) 676 (2.4)OUTCOMES n (%) n (%) n (%)Breastfed 6 months or longerNo 958 (25.3) 450 (20.8) 498 (30.4)Yes 2329 (60.8) 1381 (64.5) 930 (56.2)Don’t know 22 (0.5) 11 (0.4) 10 (0.7)Missing 50 (1.5) 29 (1.5) 20 (1.6)N/A (<6 months old) 451 (11.9) 270 (12.8) 180 (11.2)Completely immunizedNo 1251 (31.6) 722 (32.2) 515 (30.4)Yes 2435 (64.9) 1359 (64.5) 1061 (65.7)Don’t know 118 (3.3) 57 (3.0) 59 (3.7)Missing 6 (0.2) 3 (0.2) 3 (0.2)Recent ARINo 3028 (77.8) 1708 (78.5) 1296 (77.0)Yes 743 (21.0) 412 (20.5) 326 (21.6)Don’t know 17 (0.5) 6 (0.2) 10 (0.8)Missing 22 (0.7) 15 (0.8) 6 (0.6)Recent DiarrhoeaNo 3255 (85.2) 1815 (84.5) 1416 (86.3)Yes 526 (13.9) 309 (14.6) 211 (12.7)Don’t know 5 (0.2) 3 (0.2) 2 (0.2)Missing 24 (0.7) 14 (0.6) 9 (0.8)Continued on next page504.1. Description of the analytic sub-sampleAll children,N=3810Children withnon-co-residentfathers, n=2141Children withco-residentfathers, n=1638POTENTIAL CONFOUNDERS n (%) n (%) n (%)Preceding birth interval & birthorder1st born 1232 (32.4) 897 (43.0) 315 (19.4)<24 months, 2-4th born 232 (6.3) 110 (5.1) 122 (7.8)24-47 months, 2-4th born 671 (17.9) 336 (15.4) 334 (21.2)>47 months, 2-4th born 1016 (26.9) 477 (22.2) 530 (32.2)<24 months, 5+ born 96 (2.4) 47 (2.1) 49 (2.8)24-47 months, 5+ born 304 (7.4) 141 (6.0) 162 (9.2)>47 months, 5+ born 254 (6.6) 131 (6.1) 123 (7.3)Missing 5 (0.1) 2 (0.1) 3 (0.2)Place of deliveryHome 610 (14.0) 380 (16.1) 228 (11.6)Public Medical 2862 (75.3) 1682 (79.5) 1155 (70.2)Private Medical 303 (9.7) 57 (3.2) 244 (17.3)Missing 35 (1.0) 22 (1.1) 11 (0.9)Antenatal care providerNurse/mid-wife (+/- doctor) 3005 (77.2) 1850 (85.9) 1133 (67.2)Doctor only 629 (18.3) 210 (10.7) 412 (27.0)TBA, Other, No antenatal care 140 (3.4) 66 (2.5) 73 (4.5)Missing 36 (1.1) 15 (0.8) 20 (1.4)When mother wanted child’sbirthAt that time 1810 (47.9) 771 (36.0) 1028 (62.3)Later 1307 (34.7) 928 (44.9) 366 (22.4)No more 683 (17.2) 435 (18.8) 242 (15.1)Missing 10 (0.3) 7 (0.3) 2 (0.2)Child’s sexFemale 1917 (50.9) 1071 (50.6) 806 (51.7)Male 1893 (49.1) 1070 (49.4) 832 (48.3)Child’s age in months, Median(IQR)27 (30) 24 (30) 29 (31)Continued on next page514.1. Description of the analytic sub-sampleAll children,N=3810Children withnon-co-residentfathers, n=2141Children withco-residentfathers, n=1638SeasonSummer 2169 (58.8) 1234 (61.9) 912 (54.8)Autumn 1315 (31.0) 769 (31.6) 541 (30.7)Winter 326 (10.2) 138 (6.5) 185 (14.6)Mother’s population groupBlack 3010 (81.2) 1862 (90.1) 1123 (70.6)Non-Black21 780 (18.3) 270 (9.4) 504 (28.8)Coloured 499 (10.0) 253 (8.8) 242 (11.3)White 190 (5.9) 11 (0.4) 179 (12.5)Asian/Indian 91 (2.5) 6 (0.3) 83 (5.0)Missing 20 (0.5) 9 (0.4) 11 (0.6)Mother’s highest completededucation levelLess than primary 1028 (26.4) 558 (26.0) 463 (26.8)Primary or incomplete secondary 1991 (51.1) 1203 (54.4) 772 (47.1)Secondary or higher 791 (22.5) 380 (19.5) 403 (26.0)Mother’s childhood place ofresidence and whether shemigratedRural area, did not migrate 1907 (46.3) 1252 (56.6) 649 (35.1)Urban area, did not migrate 1164 (33.7) 515 (25.8) 632 (42.2)Rural area, migrated to urban 520 (15.3) 250 (12.5) 265 (18.5)Urban area, migrated rural 170 (3.6) 95 (3.8) 73 (3.4)Missing 49 (1.1) 29 (1.3) 19 (0.8)Mother’s age at first child birth<18 years 959 (25.1) 592 (27.8) 361 (22.0)18-29 years 2739 (71.7) 1501 (69.7) 1214 (74.0)>29 years 112 (3.2) 48 (2.4) 63 (4.0)Mother’s age at index child’sbirth; Median (IQR)26 (11) 24 (10) 28 (9)Continued on next page21We use a binary population group variable in our regression analyses. The “non-Black” category was created by collapsingthe Coloured, White, and Asian/Indian categories. Statistics for the sub-categories are shown here for descriptive purposes.524.1. Description of the analytic sub-sampleAll children,N=3810Children withnon-co-residentfathers, n=2141Children withco-residentfathers, n=1638Type of neighbourhoodUrban 1702 (49.3) 773 (38.7) 907 (61.0)Rural 2108 (50.7) 1368 (61.3) 731 (39.0)ADDITIONAL VARIABLES (%) n (%) n (%)Mother’s current marital statusMarried 1855 (48.4) 534 (22.8) 1318 (79.5)Unmarried 1955 (51.6) 1607 (77.2) 320 (20.5)Never married 1314 (33.4) 1279 (60.2) 12 (1.0)Living like married 440 (12.9) 140 (8.3) 300 (18.8)Widowed 20 (0.4) 19 (0.8) 1 (0.0)Divorced 47 (1.4) 43 (2.3) 2 (0.2)Separated 134 (3.4) 126 (5.7) 5 (0.4)Mother worked in last 7 daysNo 2789 (72.5) 1677 (77.7) 1089 (66.4)Yes 1005 (27.0) 453 (21.7) 545 (33.3)Missing 16 (0.5) 11 (0.7) 4 (0.2)Household wealth score(quintile)1 (Lowest) 990 (22.7) 690 (28.7) 297 (16.0)2 887 (24.1) 555 (28.0) 327 (19.8)3 765 (19.8) 425 (20.4) 332 (19.3)4 658 (17.8) 352 (16.6) 295 (18.6)5 (Highest) 510 (15.5) 119 (6.4) 387 (26.2)Continued on next page534.1. Description of the analytic sub-sampleAll children,N=3810Children withnon-co-residentfathers, n=2141Children withco-residentfathers, n=1638ProvinceWestern Cape 266 (9.1) 108 (7.0) 154 (11.6)Eastern Cape 973 (14.3) 640 (17.2) 331 (11.1)Northern Cape 344 (2.2) 176 (2.1) 164 92.4)Free State 243 (5.3) 104 (4.3) 139 (6.7)KwaZulu Natal 532 (20.7) 282 (22.3) 246 (19.1)North West 282 (7.3) 159 (7.7) 120 (6.9)Gauteng 288 (18.5) 98 (11.7) 179 (25.3)Mpumalanga 400 (7.6) 243 (8.5) 156 (6.7)Northern Province (Limpopo) 482 (14.9) 331 (19.2) 149 (10.1)Note: Ratio=Ratio of children to each higher level unit; ARI=Acute Respiratory Infection;IQR=InterQuartile Range; TBA=Traditional Birth Attendant. Percentages, medians and IQRs areweighted to represent the 1998 South African population. Counts are unweighted.The second panel of Table 4.1 shows the prevalence of each outcome in the complete analytic sampleand separately for children with and without co-resident fathers. The third panel shows the frequencydistributions (or measures of central tendency and spread) of potential confounding variables. The con-founders are arranged from child- to neighbourhood-level. However, so that comparisons can be madeacross strata of the child-level exposure variable, all of the descriptive statistics have been calculated atthe child-level (for example, ‘type of neighbourhood’ reflects the percentage of children living in urbanneighbourhoods as opposed to the percentage of neighbourhoods that are urban). Medians and interquar-tile ranges for the five neighbourhood variables derived from 1996 census data are shown separately inTable 4.6 on page 69.The descriptive statistics show a clear tendency for women in the sample to have had relatively lowparity and long birth intervals. This is consistent with our observation that few children co-resided withother children in the under-5 age-group. This makes it difficult to use descriptive statistics to evaluatewhether children living in the same household were more similar to one another in their experience ofthe exposure and outcomes than they were to children in other households. Clustering can be moreeasily examined for children living in the same neighbourhood. We found that 50% of neighbourhoodshad a mixture of children with and without co-resident fathers. Whereas, in 22% of neighbourhoodsthere were only children with non-co-resident fathers, and in 29% there were only children with co-resident fathers. We also found that 65% of children who were breastfed for six months or longer livedin just 33% of neighbourhoods. Similarly, 65% of children who are completely immunized lived in35% of neighbourhoods. For the illness outcomes, 59% of all children who had a recent ARI wereconcentrated in the 20% of neighbourhoods where two or more children had this outcome. While 50%544.1. Description of the analytic sub-sample(a)(b)(c)Figure 4.1: Bar graphs representing the distribution of children in the complete analytic sub-sampleacross mothers (a), households (b) and neighbourhoods (c).554.1. Description of the analytic sub-sampleof children who had recent diarrhoea lived in the 15% of neighbourhoods where two or more childrenhad this outcome. These statistics suggests that the probability of living with a father and the probabilityof experiencing each health outcome was not evenly distributed across South African neighbourhoods.It is clear that, compared to children with non-co-resident fathers, fewer children with co-residentfathers were breastfed for six months or longer. In contrast, there were no differences between chil-dren with and without co-resident fathers in the probability of having experienced each of the otherthree outcomes. This is surprising given that the groups had clearly different distributions of potentialconfounders. For example, a slightly greater percentage of children with non-co-resident fathers weredelivered at home, while a dramatically greater percentage of children with co-resident fathers weredelivered in private medical facilities. These differences in birth environment might be expected tomanifest in different rates of immunization completeness because BCG and polio vaccines are given atbirth. Conversely, children with co-resident fathers tended to be older than those with non-co-residentfathers. Therefore, we may expect a greater percentage of the latter group to be completely immunizedbecause fewer would have had the opportunity to miss a vaccine dose. Possibly the reason there is nounadjusted association between father’s co-residence status and immunization completeness is becausethere is confounding in both positive and negative directions, leading to a net null effect.Another interesting observation is that children with co-resident fathers tended to be their mother’ssecond born child or higher, while a much larger percentage of children with non-co-resident fatherswere first borns. We might expect that having more siblings would increase children’s exposure toinfectious organisms, manifesting in the group of children with co-resident fathers having higher preva-lences of ARI and diarrhoea. On the other hand, a far greater percentage of children with non-co-residentfathers lived in rural neighbourhoods. Children in rural neighbourhoods tend to have greater risks for di-arrhoeal and respiratory illnesses because of poorer access to sanitation and running water and becauseof higher levels of indoor air pollution from cooking fuels [2, p. 35]. Again these opposing confoundingeffects may result in net null unadjusted associations between father’s co-residence status and recentARI or diarrhoeal illness.Examining indicators of household socioeconomic status (antenatal care provider and place of deliv-ery, mother’s population group and educational attainment, household wealth index and type of neigh-bourhood) it is clear that children with co-resident fathers tended to come from higher socioeconomicstatus households. However, it is interesting to note that this group does not seem to be a purely ‘highstatus’ group. For example, receipt of antenatal care from a non-medical provider (or receiving nocare), though rare in absolute terms, were twice as common for children with co-resident fathers thanfor those with non-co-resident fathers. Similarly, although having a mother who completed high schoolwas more common among children with co-resident fathers, the percentage whose mothers had lessthan primary-level education was equal in the two exposure groups. The potential confounding effectsof these variables on the co-resident father-child health outcome associations can be more readily ap-preciated by comparing adjusted to unadjusted odds ratios, presented in the following section.564.2. Findings from multilevel regressions4.2 Findings from multilevel regressionsPrincipal analyses of father’s co-residence status and child health outcomesIn this section we present results of regressions of each child health outcome on father’s co-residencestatus. In the top panel of Table 4.2 on page 59, odds ratios (ORs) and 95% Credible Intervals (CIs)22are from ordinary logistic regression models, without confounder adjustment. The remainder of theresults in Table 4.2 are from multilevel logistic regressions, which include neighbourhood-level varyingintercepts and co-resident father varying slopes and adjust for all potential confounders listed. Pointand interval estimates for the parameters of the neighbourhood varying coefficients are presented inthe bottom panel of the table. Tables A.2-A.5 (Appendix A.4) present findings of complementary re-gression models which do not model neighbourhood variation in the co-resident father slope or which,additionally, model household variation in the intercept. Note that, in models with varying slopes, theco-resident father regression co-efficient is conditional on neighbourhood membership. The interpre-tation thus becomes: the average expected change in the odds of having the outcome associated withhaving a co-resident father, comparing otherwise similar children living in the same neighbourhood.For models of breastfeeding duration, complete case data were available for 3126 children (93.1% ofchildren at least six months old in analytic sub-sample). 3499 children (91.8% of analytic sub-sample)contributed data to models of immunization completeness. For the models of recent ARI, completecase data were available for 3633 children (95.4% of analytic sub-sample) and, for recent diarrhoea,complete case data were available for 3642 children (95.6% of analytic sub-sample).Consistent with the descriptive statistics in the preceding section, without adjustment for potentialconfounding, having a co-resident father was associated with 39% lower odds of having been breastfedfor six months or longer. In contrast, the OR estimate from the full model is completely attenuated.This finding suggests that, for otherwise similar children living in the same neighbourhood, having aco-resident father was not associated with having been breastfed for six months or longer.Also consistent with the descriptive statistics, the unadjusted regression results suggest having a co-resident father was not significantly associated with having been completely immunized or with havinghad a recent illness. The multilevel models suggest similar results hold after controlling for potentialconfounding and neighbourhood membership.Referring to the bottom panel of the table, we observe relatively large standard deviations (SDs) forthe neighbourhood varying coefficients in all four models. The credible intervals give the range of mostlikely values for these parameters. These ranges are all reasonably large, but suggest that it is very un-likely that any of the varying coefficient parameters is zero (which would imply no neighbourhood-levelvariation). The varying slopes SDs for the immunization completeness and recent diarrhoea outcomesare of similar magnitude (SD: 0.59, 95% CI: 0.35 - 0.91; and SD: 0.57, 95% CI: 0.33 - 0.91, respec-tively). Whereas the varying slopes SD is noticeably larger in the model for recent ARI, though theassociated CI is very wide (SD: 1.10, 95% CI: 0.63 - 1.53).22A Credible Interval is interpreted similarly to a Confidence Interval in frequentist statistics. The 95% credible intervalgives the range of values which are 95% likely to include the true value of the parameter of interest.574.2. Findings from multilevel regressionsThe SD estimates are on the logit scale. To interpret the magnitude of variation on the odds ratioscale we can calculate the interval that would be expect to include odds ratios from 95% of neighbour-hoods in South Africa. In contrast to the 95% CIs (which reflect uncertainty in the parameter estimates)the following 95% OR intervals reflect the extent of variation observed in the co-resident father effectestimate across children living in different neighbourhoods. For the breastfeeding outcome the range ofodds ratios expected to be observed in 95% of South African neighbourhoods was 0.24 - 4.56. Similarly,for immunization completeness and recent diarrhoea the 95% OR ranges were 0.38 - 3.90 and 0.38 -3.58, respectively. For ARI the range was considerably wider than for the other outcomes because ofthe larger slopes SD (95% OR range: 0.14 - 10.40). The OR point estimates suggest that, on average,having a co-resident father was not associated with any of the child health outcomes. However, all ofthe above 95% intervals include very large and very small OR values. This indicates that in a percentageof neighbourhoods having a co-resident father was associated with increased odds of the child healthoutcomes and in other neighbourhoods having a co-resident father was associated with reduced odds ofthe outcomes.There appeared to be little to no correlation between the varying intercepts and slopes in models forbreastfeeding, immunization and diarrhoea. In contrast, in the ARI model the correlation was estimatedto be reasonably large and negative (ρ: -0.43, 95% CI: -0.74 – 0.03). This could indicate that there wereimportant covariates not accounted for in the model, which influence both neighbourhood intercepts andslopes, causing them to be correlated.584.2.FindingsfrommultilevelregressionsTable 4.2: Odds ratios (ORs) and 95% Credible Intervals (CIs) estimated from multilevel logistic regressions for four child health outcomes; childrenaged 0-4 years in the 1998 SADHS.Breastfed ≥6 months Completely immunized Recent ARI Recent diarrhoeaEst. (95% CI) Est. (95% CI) Est. (95% CI) Est. (95% CI)UNADJUSTEDCo-resident father 0.61 (0.52 - 0.71) 1.11 (0.96 - 1.27) 1.04 (0.88 - 1.22) 0.88 (0.73 - 1.07)VARYING INTERCEPT, VARYING SLOPECo-resident father 0.99 (0.78 - 1.25) 1.17 (0.95 - 1.43) 1.09 (0.81 - 1.42) 1.12 (0.83 - 1.45)Birth order & precedingbirth interval1st born 0.68 (0.49 - 0.91) 1.09 (0.82 - 1.41) 1.18 (0.85 - 1.60) 1.31 (0.89 - 1.84)2-4 born, <24 months 0.95 (0.59 - 1.46) 1.07 (0.72 - 1.54) 1.41 (0.89 - 2.11) 1.33 (0.79 - 2.09)2-4 born, 24-47 months REF REF REF REF2-4 born, >47 months 0.89 (0.65 - 1.19) 0.97 (0.74 - 1.25) 1.29 (0.94 - 1.73) 1.31 (0.91 - 1.83)5+ born, <24 months 0.64 (0.32 - 1.19) 0.58 (0.32 - 0.98) 1.67 (0.81 - 3.00) 2.16 (1.01 - 3.99)5+ born, 24-47 months 1.11 (0.65 - 1.77) 0.82 (0.54 - 1.20) 1.26 (0.77 - 1.94) 1.41 (0.84 - 2.25)5+ born, >47 months 1.18 (0.66 - 1.96) 1.02 (0.64 - 1.55) 1.62 (0.96 - 2.62) 2.06 (1.16 - 3.39)Continued on next page594.2.FindingsfrommultilevelregressionsBreastfed ≥6 months Completely immunized Recent ARI Recent diarrhoeaEst. (95% CI) Est. (95% CI) Est. (95% CI) Est. (95% CI)Place of deliveryPublic medical facility REF REF REF REFHome 1.31 (0.95 - 1.77) 0.60 (0.47 - 0.76) 1.03 (0.77 - 1.34) 1.41 (1.05 - 1.86)Private medical facility 0.68 (0.45 - 0.98) 0.89 (0.60 - 1.27) 0.80 (0.51 - 1.20) 0.60 (0.32 - 1.00)Antenatal care providerNurse/ midwife (+/- doctor) REF REF REF REFDoctor 0.64 (0.48 - 0.84) 0.81 (0.61 - 1.05) 1.03 (0.75 - 1.36) 1.10 (0.77 - 1.52)TBA, Other, no care 0.95 (0.54 - 1.56) 0.66 (0.42 - 0.99) 0.69 (0.38 - 1.11) 0.52 (0.25 - 0.90)Time mother wantedpregnancyThen REF REF REF REFLater 1.20 (0.95 - 1.50) 1.12 (0.92 - 1.36) 1.04 (0.83 - 1.29) 1.35 (1.04 - 1.71)No more 0.99 (0.74 - 1.30) 0.86 (0.68 - 1.09) 1.11 (0.84 - 1.44) 1.51 (1.11 - 2.00)Child’s sexFemale REF REFMale 0.99 (0.82 - 1.18) 1.35 (1.10 - 1.66)1 SD increase in child’s age(months)0.69 (0.63 - 0.75) 0.79 (0.72 - 0.87) 0.59 (0.52 - 0.65)Continued on next page604.2.FindingsfrommultilevelregressionsBreastfed ≥6 months Completely immunized Recent ARI Recent diarrhoeaEst. (95% CI) Est. (95% CI) Est. (95% CI) Est. (95% CI)SeasonSummer REF REFAutumn 1.03 (0.80 - 1.30) 0.99 (0.77 - 1.27)Winter 1.37 (0.91 - 1.99) 0.59 (0.34 - 0.92)Mother’s population groupBlack/African REF REF REF REFNon-Black/African 0.46 (0.34 - 0.60) 1.66 (1.25 - 2.17) 1.13 (0.84 - 1.50) 0.86 (0.60 - 1.18)Mother’s highest completededucation levelLess than primary 1.44 (1.10 - 1.86) 0.77 (0.62 - 0.94) 0.85 (0.66 - 1.07) 1.18 (0.90 - 1.51)Primary or incompletesecondaryREF REF REF REFSecondary or higher 0.87 (0.67 - 1.11) 1.08 (0.84 - 1.36) 0.84 (0.64 - 1.09) 0.87 (0.63 - 1.17)Mother’s childhoodresidence & migration statusRural area, did not migrate REF REFUrban area, did not migrate 0.53 (0.39 - 0.69) 1.21 (0.93 - 1.56)Rural area, migrated urban 0.74 (0.53 - 1.03) 1.08 (0.80 - 1.43)Urban area, migrated rural 0.77 (0.48 - 1.18) 1.75 (1.09 - 2.69)Continued on next page614.2.FindingsfrommultilevelregressionsBreastfed ≥6 months Completely immunized Recent ARI Recent diarrhoeaEst. (95% CI) Est. (95% CI) Est. (95% CI) Est. (95% CI)Mother’s age at first childbirth<18 1.15 (0.89 - 1.47) 0.91 (0.74 - 1.12) 1.16 (0.91 - 1.44) 1.09 (0.83 - 1.38)18-29 REF REF REF REF>29 0.67 (0.35 - 1.18) 0.90 (0.50 - 1.51) 0.72 (0.34 - 1.33) 0.85 (0.35 - 1.66)1 SD increase in mother’sage at index child’s birth0.91 (0.77 - 1.07) 1.10 (0.95 - 1.27) 0.92 (0.77 - 1.08) 0.90 (0.74 - 1.08)Varying coefficientsSD Neighbourhood intercepts 0.77 (0.57 - 0.99) 0.76 (0.58 - 0.95) 0.78 (0.56 - 1.02) 0.56 (0.38 - 0.77)SD Neighbourhood slopes 0.74 (0.43 - 1.13) 0.59 (0.35 - 0.91) 1.10 (0.63 - 1.53) 0.57 (0.33 - 0.91)Correlation between varyingcoefficients0.29 (-0.26 - 0.74) 0.01 (-0.50 - 0.56) -0.43 (-0.74 - 0.03) 0.00 (-0.58 - 0.56)Note: ARI=Acute Respiratory Infection; Est.=Estimate; REF=Reference category; TBA=Traditional Birth Attendant; SD=Standard Deviation.624.2. Findings from multilevel regressionsInteractions between father’s co-residence status and other dimensions of householdstructureIn this section we present results of analyses examining whether other dimensions of household structurewere interacting with father’s co-residence status in their association with each child health outcome.First, we considered mother’s marital status as a potential effect modifier, grossly distinguishingbetween married and currently unmarried mothers. The latter category includes never married andformerly married women. We treated children with currently unmarried mothers and non-co-residentfathers as the reference group. Note that the OR comparing children with married mothers and co-resident fathers to the reference group was calculated as the exponentiated linear combination of thecoefficients for father’s co-residence status, mother’s marital status, and the interaction term.There were modest differences in the ratio of odds ratios within each category of mother’s mari-tal status for the immunization, ARI and diarrhoea outcomes (Table 4.3 on the next page). However,the 95% CI estimate for the interaction term in all four models included the null (results not shown).Father’s co-residence status was not associated with immunization completeness among children of cur-rently unmarried mothers. Whereas, among children with married mothers, residing with a father wasassociated with a modest, but non-significant, increased odds of being completely immunized. Therewas also a non-significant trend for residing with a father to be associated with increased odds of bothtypes of illness, but only among children with currently unmarried mothers.634.2.FindingsfrommultilevelregressionsTable 4.3: Adjusted Odds Ratios (ORs) and 95% Credible Intervals (CIs) for interactions between father’s co-residence status and mother’s maritalstatus estimated using multilevel logistic regressions with intercepts varying by neighbourhood. Counts and population-weighted percentages in eachgroup are also shown; children aged 0-4 years, 1998 SADHS.Breastfed ≥6 monthsCompletelyimmunized Recent ARI Recent diarrhoean (%) OR (95% CI) OR (95% CI) OR (95% CI) OR (95% CI)Currentlyunmarried, notco-residing1607 (41.7) REF REF REF REFCurrentlyunmarried,co-residing320 (9.4) 1.07 (0.71 - 1.57) 0.96 (0.68 - 1.31) 1.44 (1.02 - 1.98) 1.45 (0.97 - 2.07)Married, notco-residing534 (12.3) 0.70 (0.49 - 0.97) 0.96 (0.71 - 1.25) 1.05 (0.76 - 1.40) 1.33 (0.92 - 1.84)Married,co-residing1318 (36.6) 0.73 (0.56 - 0.94) 1.15 (0.91 - 1.43) 1.12 (0.87 - 1.42) 1.32 (0.98 - 1.74)Note: ARI=Acute Respiratory Infection644.2. Findings from multilevel regressionsSecond, we examined whether living with other adult relatives affected the outcomes of childrenwhose fathers were non-co-resident as compared to those of children with co-resident fathers. (Completedescriptive statistics and regression results from these analyses are presented in Tables A.6 and A.7,Appendix A.5)Among children with non-co-resident fathers, approximately half had co-resident male relatives,one quarter had co-resident female relatives (but no co-resident male relatives) and the remaining onequarter had no additional co-resident relatives23 (Table 4.4 on the following page). Only a minorityof children with co-resident fathers had additional co-resident relatives of either sex. In contrast, themajority of children with non-resident fathers and co-resident male relatives also had co-resident femalerelatives. These types of households tended to have a greater number of adult members than the threeother household types.To relate these household types to the preceding analyses involving mother’s marital status, noticethat almost half of mothers who were the lone adult in their children’s households were married, whileabout one quarter had never been married. This could reflect that more lone mother households inSouth Africa arise because of married couples residing in separate dwellings than through non-maritalchildbearing. (Formal divorce appeared to be a very uncommon cause of lone mother households.) Itappears that the majority of children who were born outside of marriage and were living with theirmothers also had other adult relatives residing with them.Examining the unadjusted OR estimates in the upper panel of table 4.5 it appears that, in comparisonto living with a co-resident father, living in any other household structure was associated with greaterodds of having been breastfed for at least six months. However, similarly to what we observed inthe principal analyses, adjusting for potential confounders completely attenuated these associations.Children living with single mothers appeared to be less likely to be completely immunized than childrenwith two co-resident parents, even after adjusting for potential confounding (OR: 0.79, 95% CI: 0.61 -1.00). In contrast, children with additional co-resident relatives had similar odds of being completelyimmunized to children with co-resident fathers. All children with non-co-resident fathers showed atrend towards lower odds of recent ARI and diarrhoea compared to children with co-resident fathers,regardless of whether they had additional co-resident relatives. However, the 95% CIs included the nullin all cases.23Recall that all children included in these analyses live with their mothers.654.2. Findings from multilevel regressionsTable 4.4: Population-weighted descriptive statistics comparing children with co-resident fathers tothose with non-co-resident fathers, the latter stratified by whether they reside with other adult relatives;children aged 0-4 years, 1998 SADHS.Non-co-resident fatherCo-residentfather> 1 malerelativeFemalerelatives only Mother onlyn (%) 1638 (43.9) 1156 (27.8) 572 (13.7) 543 (13.1)Median number de jureadult household members(IQR)2 (1) 4 (2) 2 (1) 1 (0)Number co-resident adultfemale relatives0 76.9 16.3 0.0 100.01 16.4 39.9 63.6 0.02+ 6.4 43.7 36.4 0.0Missing 0.2 0.0 0.0 0.0Number co-resident adultmale relatives0 81.4 0.0 100.0 100.01 13.1 60.7 0.0 0.02+ 5.3 39.3 0.0 0.0Missing 0.2 0.0 0.0 0.0Mother’s current maritalstatusMarried 79.5 12.3 14.6 48.6Unmarried 20.5 87.7 85.4 51.4Never married 1.0 71.6 69.4 24.2Living like married 18.8 5.9 8.7 10.8Widowed 0.0 2.1 1.6 5.2Divorced 0.2 1.6 2.1 4.0Separated 0.4 6.5 3.5 7.2Note: IQR=InterQuartile Range.664.2.FindingsfrommultilevelregressionsTable 4.5: Odds Ratios (ORs) and 95% Credible Intervals (CIs) comparing children with different combinations of co-resident relatives estimated byordinary logistic regressions (unadjusted) or multilevel logistic regressions with neighbourhood-level varying intercepts (adjusted); children aged 0-4years, 1998 SADHS.Breastfed >= 6 months Completely immunized Recent ARI Recent diarrhoeaOR (95% CI) OR (95% CI) OR (95% CI) OR (95% CI)UNADJUSTEDCo-resident father REF REF REF REF> 1 male relative 1.52 (1.26 - 1.81) 1.01 (0.85 - 1.19) 0.99 (0.81 - 1.19) 1.17 (0.93 - 1.44)Female relatives only 1.63 (1.28 - 2.06) 0.88 (0.71 - 1.08) 0.98 (0.76 - 1.24) 1.24 (0.92 - 1.61)Mother only 1.84 (1.44 - 2.35) 0.67 (0.54 - 0.82) 0.84 (0.64 - 1.08) 1.02 (0.76 - 1.35)ADJUSTED, VARYING INTERCEPTSCo-resident father REF REF REF REF> 1 male relative 1.12 (0.86 - 1.42) 0.92 (0.73 - 1.14) 0.85 (0.67 - 1.07) 0.83 (0.63 - 1.07)Female relatives only 1.03 (0.74 - 1.40) 0.89 (0.68 - 1.14) 0.86 (0.64 - 1.14) 0.85 (0.61 - 1.15)Mother only 1.08 (0.79 - 1.45) 0.79 (0.61 - 1.00) 0.82 (0.60 - 1.09) 0.84 (0.60 - 1.14)SD neighbourhoodintercepts0.93 (0.75 - 1.11) 0.80 (0.65 - 0.94) 0.73 (0.56 - 0.90) 0.56 (0.37 - 0.74)Note: ARI=Acute Respiratory Infection; REF=Reference category; SD=Standard Deviation.674.3. Explaining variation in neighbourhood coefficients4.3 Explaining variation in neighbourhood coefficientsIn this section we present descriptive statistics for the neighbourhood-level covariates derived from 1996census data. We then present findings of analyses assessing the degree to which each neighbourhoodcovariate accounted for variation in the neighbourhood intercepts and co-resident father slopes.There were clear differences between the typical neighbourhood characteristics of children whohad co-resident fathers and those whose fathers were non-co-resident (Table 4.6). The former grouptended to live in neighbourhoods where a greater percentage of female residents had completed highschool. Interestingly, the IQR for the percentage of women having completed high school was higher forchildren whose fathers were co-resident than for children whose fathers were non-co-resident (IQR 22.7vs 10.7). Children with co-resident fathers also tended to live in neighbourhoods with lower percentagesof female-headed households, lower rates of male unemployment, and lower concentrations of low-income households.Table 4.7 on page 71 is a compilation of findings from seven multilevel models for each outcome.The first panel presents a subset of results from the ‘Varying intercept, varying slope models’ firstpresented Table 4.2 (reproduced here for ease of comparison). Each subsequent panel presents theresults of a model which includes a single neighbourhood covariate as a predictor in the models for thevarying intercepts and slopes (and is otherwise identical to the ‘Varying intercept, varying slope model’).The neighbourhood covariates are indicated in bold font in the first column of Table 4.7 on page 71. Theestimate for the neighbourhood covariate is the (exponentiated) regression coefficient from the modelof the neighbourhood intercepts. The ‘interaction’ estimate is the (exponentiated) regression coefficientfrom the model of the random slopes.Examining the OR estimates for each neighbourhood covariate, we observed few ORs with 95% CIsthat exclude the null. Two exceptions were the ORs for a) the rate of unemployment among male neigh-bourhood residents, and b) the percentage of female residents with completed high school education orhigher in models for the immunization outcome. These ORs are interpreted as the expected ratio of theodds of being completely immunized comparing otherwise similar children who live in neighbourhoodsthat differ by 1 SD on the characteristic of interest. For example, comparing otherwise similar children,each 17.9% increase in the neighbourhood percentage of female residents who completed high schoolwas associated with an 18% increase in the odds of being completely immunized. Conversely, a 26.5%increase in the unemployment rate for male neighbourhood residents was associated with a 14% declinein the odds of having been completely immunized, all other characteristics being equal. Including anyof these covariates in the model for the neighbourhood intercepts resulted in only a small reduction inthe estimated SD of the intercepts about their mean.Next we examine the estimated ORs for the interaction terms. In these models, the OR for the co-resident father variable is interpreted as the expected ratio of the odds of having the outcome comparinga child who had a co-resident father to an otherwise similar child in the same neighbourhood whohad a non-co-resident father, where the value of the neighbourhood-level predictor in question was at684.3. Explaining variation in neighbourhood coefficientsTable 4.6: Medians (M) and InterQuartile Ranges (IQR) for neighbourhood contextual variables derivedfrom 1996 SA Census data. Statistics are presented for the complete analytic sample of children fromthe 1998 SADHS and separately for strata of children with and without co-resident fathersAll childrenN=3810Children withnon-co-residentfathersn=2134Children withco-resident fathersn=1646M IQR M IQR M IQR% Female residentscompleted highschool10.7 (15.3) 8.7 (10.7) 13.8 (22.7)Ratiofemale-to-maleresidents completedhigh school0.9 (0.5) 0.9 (0.6) 0.9 (0.5)% Householdshaving a femalehead40.7 (28.6) 46.7 (27.5) 31.5 (26.7)% Male residentsunemployed37.6 (39.4) 44.7 (36.6) 25.9 (37.5)% Households withannual income <R600038.4 (44.6) 45.2 (42.8) 28.8 (41.2)Missing data; n (%) 39 (1.0) (1.2) (0.5)694.3. Explaining variation in neighbourhood coefficientsthe reference level24. On the other hand, the OR for the interaction term is the expected combinedeffect of having a co-resident father and a 1SD increase in the value of the neighbourhood covariate,comparing otherwise similar children, living in otherwise similar neighbourhoods (i.e.: neighbourhoodshaving the same value of the random slope). Few of the ORs were significant. The largest interactionswere for the urban neighbourhood indicator in the model of immunization completeness, and for thepercentage of female residents with completed high school education in the model for recent ARI. Theinteraction for the urban neighbourhood indicator suggested that, comparing otherwise similar childrenin a rural neighbourhood, father’s co-residence status was not associated with a change in odds of beingcompletely immunized. Whereas, comparing otherwise similar children in an urban neighbourhood,co-residing with a father was associated with about 50% higher odds of being completely immunized.For the ARI outcome, as the neighbourhood percentage of female high school graduates increased, theassociation between having a co-resident father and having had a recent ARI became more stronglypositive. Similarly to the case for the varying intercepts, none of the covariates appeared to significantlyexplain the variation in the neighbourhood-specific effects of having a co-resident father from theiroverall mean value.24The reference level for the urban neighbourhood indicator variable is rural. For all other neighbourhood covariates, thereference level is the mean across all neighbourhoods in the analytic sample.704.3.ExplainingvariationinneighbourhoodcoefficientsTable 4.7: Odds Ratios (95% credible intervals) for neighbourhood covariates in models for neighbourhood varying intercepts and co-resident fatherslopes estimated using multilevel logistic regression. Standard Deviations (95% credible intervals) for the varying coefficients are also shown.Breastfed ≥6 months Completely immunized Recent ARI Recent diarrhoeaEst. (95% CI) Est. (95% CI) Est. (95% CI) Est. (95% CI)Co-resident father 0.99 (0.78 - 1.25) 1.17 (0.95 - 1.43) 1.09 (0.81 - 1.42) 1.12 (0.83 - 1.45)SD neighbourhood intercepts 0.77 (0.57 - 0.99) 0.76 (0.58 - 0.95) 0.78 (0.56 - 1.02) 0.56 (0.38 - 0.77)SD neighbourhood slopes 0.74 (0.43 - 1.13) 0.59 (0.35 - 0.91) 1.10 (0.63 - 1.53) 0.57 (0.33 - 0.91)Co-resident father 0.90 (0.63 - 1.24) 0.95 (0.74 - 1.22) 0.99 (0.68 - 1.39) 1.24 (0.86 - 1.70)Urban neighbourhood25 N/A N/A 0.88 (0.64 - 1.19) 0.95 (0.69 - 1.28)Interaction 1.24 (0.79 - 1.86) 1.66 (1.13 - 2.37) 1.26 (0.79 - 1.91) 0.84 (0.52 - 1.28)SD neighbourhood intercepts 0.78 (0.57 - 1.00) 0.75 (0.58 - 0.94) 0.79 (0.56 - 1.02) 0.57 (0.38 - 0.78)SD neighbourhood slopes 0.73 (0.42 - 1.11) 0.58 (0.35 - 0.89) 1.10 (0.64 - 1.53) 0.58 (0.34 - 0.93)Co-resident father 1.01 (0.77 - 1.29) 1.12 (0.90 - 1.38) 1.13 (0.84 - 1.47) 1.06 (0.78 - 1.40)%male residentsunemployed1.03 (0.88 - 1.20) 0.86 (0.75 - 0.98) 1.08 (0.93 - 1.25) 0.89 (0.77 - 1.03)Interaction 1.07 (0.86 - 1.34) 0.99 (0.83 - 1.19) 0.89 (0.71 - 1.10) 1.10 (0.88 - 1.36)SD neighbourhood intercepts 0.76 (0.56 - 0.98) 0.74 (0.56 - 0.93) 0.78 (0.56 - 1.02) 0.57 (0.38 - 0.79)SD neighbourhood slopes 0.75 (0.45 - 1.13) 0.60 (0.36 - 0.90) 1.10 (0.66 - 1.53) 0.60 (0.34 - 0.95)Continued on next page25Because we used the ‘type of neighbourhood’ variable to derive the individual-level variable indicating whether the mother migrated since childhood, these variables are perfectlycorrelated and it is not possible to include them together in our models. In consequence, in models of the breastfeeding and immunization outcomes, we do not estimate the maineffect of type of neighbourhood on the outcome (as aspects of this effect are being reflected in the effects of having migrated versus continuously resided in the same type of areasince childhood). However, we do estimate the effect of the interaction between type of neighbourhood and father’s co-residence status.714.3.ExplainingvariationinneighbourhoodcoefficientsBreastfed ≥6 months Completely immunized Recent ARI Recent diarrhoeaEst. (95% CI) Est. (95% CI) Est. (95% CI) Est. (95% CI)Co-resident father 1.00 (0.77 - 1.27) 1.18 (0.96 - 1.46) 1.09 (0.81 - 1.43) 1.12 (0.83 - 1.49)% households having afemale head1.04 (0.89 - 1.22) 1.02 (0.88 - 1.16) 1.02 (0.87 - 1.18) 1.02 (0.87 - 1.18)Interaction 1.03 (0.83 - 1.27) 0.93 (0.77 - 1.10) 0.96 (0.77 - 1.18) 0.97 (0.78 - 1.20)SD neighbourhood intercepts 0.77 (0.56 - 0.98) 0.76 (0.59 - 0.95) 0.79 (0.57 - 1.02) 0.57 (0.38 - 0.79)SD neighbourhood slopes 0.74 (0.43 - 1.13) 0.59 (0.36 - 0.90) 1.11 (0.65 - 1.55) 0.59 (0.34 - 0.94)Co-resident father 1.00 (0.78 - 1.27) 1.16 (0.94 - 1.43) 1.13 (0.84 - 1.48) 1.06 (0.78 - 1.41)% households with annualincome < R60001.11 (0.94 - 1.30) 0.94 (0.81 - 1.09) 1.15 (0.99 - 1.33) 0.98 (0.84 - 1.14)Interaction 0.92 (0.74 - 1.14) 1.02 (0.85 - 1.22) 0.93 (0.74 - 1.15) 1.21 (0.96 - 1.50)SD neighbourhood intercepts 0.78 (0.57 - 1.00) 0.76 (0.58 - 0.94) 0.78 (0.55 - 1.02) 0.57 (0.38 - 0.79)SD neighbourhood slopes 0.71 (0.40 - 1.11) 0.59 (0.35 - 0.91) 1.11 (0.65 - 1.55) 0.59 (0.34 - 0.93)Co-resident father 1.03 (0.80 - 1.30) 1.15 (0.92 - 1.41) 1.14 (0.85 - 1.51) 1.08 (0.80 - 1.40)% female residents withhigh school education orhigher0.88 (0.75 - 1.03) 1.18 (1.03 - 1.35) 0.91 (0.78 - 1.06) 1.01 (0.86 - 1.19)Interaction 1.15 (0.92 - 1.43) 0.99 (0.83 - 1.19) 1.25 (0.99 - 1.55) 0.90 (0.71 - 1.14)SD neighbourhood intercepts 0.78 (0.57 - 1.00) 0.73 (0.56 - 0.92) 0.77 (0.54 - 1.01) 0.57 (0.39 - 0.78)SD neighbourhood slopes 0.72 (0.42 - 1.10) 0.59 (0.34 - 0.91) 1.07 (0.61 - 1.50) 0.59 (0.34 - 0.95)Continued on next page724.3.ExplainingvariationinneighbourhoodcoefficientsBreastfed ≥6 months Completely immunized Recent ARI Recent diarrhoeaEst. (95% CI) Est. (95% CI) Est. (95% CI) Est. (95% CI)Co-resident father 0.99 (0.77 - 1.26) 1.17 (0.94 - 1.43) 1.08 (0.80 - 1.40) 1.13 (0.84 - 1.49)Female-to-male ratioresidents with high schooleducation or higher0.91 (0.80 - 1.03) 0.98 (0.88 - 1.09) 0.93 (0.82 - 1.05) 0.95 (0.83 - 1.07)Interaction 1.00 (0.82 - 1.21) 1.05 (0.89 - 1.23) 1.15 (0.94 - 1.39) 1.08 (0.88 - 1.30)SD neighbourhood intercepts 0.76 (0.56 - 0.98) 0.76 (0.59 - 0.95) 0.78 (0.56 - 1.02) 0.57 (0.38 - 0.80)SD neighbourhood slopes 0.75 (0.43 - 1.15) 0.58 (0.35 - 0.91) 1.09 (0.62 - 1.52) 0.59 (0.35 - 0.95)Note: ARI=Acute Respiratory Infection; Est.=Estimate of Odds Ratio or Standard Deviation; CI=Credible Interval; SD=Standard Deviation.73Chapter 5Discussion: Effects of Father’sCo-residence Status on Child HealthOutcomesThe focus of this project was on the influences that biological fathers can have on their children’s health.Although father-child residential arrangements, the focus of the preceding set of analyses, only tells asmall part of the story, it does serve to highlight an important feature of the context for fathering in SouthAfrica: the nuclear family form - involving a married or cohabiting couple residing with their offspring- is far from the norm in South Africa (as previously emphasized, see for example [180, 122, p. 60-1]).Less than half of the children in our sample were residing with both parents. Nevertheless, father-childco-residence was considerably more common in our study than in two previous studies of households inrural areas of South Africa. Both of these found less than one third of children were residing with fathers[76, 81]. The difference between their findings and ours may be mainly because paternal co-residenceappears to be more common in urban than rural areas. That our findings are reflective of the situationnationally is supported by the fact that 46% of 0-4 year old children were estimated to be residing withboth biological parents in the 2002 South African General Household Survey, similar to the 43% in ouranalytic sample (calculated by the authors using Statistics South Africa data [49]).In our study, children with non-co-resident fathers were found to be living in a variety of householdstructures, only one quarter of which could be considered truly ‘single-mother households’. Manyprevious investigations of household structure effects on child health focus on a single dimension ofhousehold structure, such as mother’s marital status or female household headship. The complexity ofhousehold structures we observed in this dataset leads us to suggest that simultaneously consideringmultiple dimensions of household structure may produce more meaningful results, especially whenfathers are a focus of the investigation.With this set of analyses, we sought to address three questions. First, we asked whether childrenwho reside with both biological parents would tend to have better health-related outcomes than childrenwho reside with mothers but not fathers. Second, we assessed whether the (lack of) association betweenfather’s co-residence status and children’s health outcomes was modified by either a) mother’s maritalstatus, or b) co-residence status of other adult family members. Third, we estimated the magnitudeof neighbourhood-level variation in the association between residing with a father and children’s healthoutcomes, and explored whether neighbourhood contextual variables constructed from census data couldexplain variation in the co-resident father effect estimates. In the sections that follow, we discuss our745.1. Do children who reside with their biological father tend to have better health outcomes?findings to each question as well as some important limitations of our analyses.5.1 Do children who reside with their biological father tend to havebetter health outcomes?We did not find evidence to support this hypothesis. The lack of association with father’s co-residencestatus was consistent for all of the child health-related outcomes we examined. In general, controllingfor potential confounders did not change the conclusions of unadjusted analyses. Breastfeeding durationwas the exception: the unadjusted effect estimate suggested that children with co-resident fathers tendedto be less likely to be breastfed for six months or longer; whereas, the adjusted estimate suggestedno difference between the two groups of children. It is difficult, using the cross-sectional measuresavailable in the SADHS dataset, to control for socioeconomic conditions preceding the observed livingarrangements. Nevertheless, we were careful to control for indicators of socioeconomic status that wereunlikely to result from father’s co-residence status at the time of the survey. As such, our adjustedregression findings support the conclusion that the tendency for children in two-parent households to bebreastfed for shorter durations is more likely related to living in higher socioeconomic status householdsthan to co-resident fathers effecting earlier weaning.Previous research in low- and middle-income countries has found that more educated women,women who are employed and those with higher incomes tend to wean their children earlier [181,182, 97]. One of these studies also demonstrated that women with co-resident spouses tended to breast-feed for longer durations. The authors explained this association as possibly being because women withabsent spouses are “more likely to assume market economic roles, which often are incompatible withbreast-feeding.” [97, p. 67] Our analyses did not show the same association with father’s co-residencestatus. Possibly because, in our sample, a greater percentage of women who resided with their children’sfathers were working. This could relate to confounding by socioeconomic status: more highly educatedwomen tended to be more likely to co-reside with their child’s father. Additionally, more highly edu-cated women were more likely to be employed, reducing the chances that they breastfed for extendeddurations. Possibly, had we been able to adjust for more specific socioeconomic indicators, particularlywhether mothers worked prior to children’s births, we may have found a stronger positive associationbetween father’s co-residence and breastfeeding duration than we did.5.2 Are the father’s co-residence status – child health outcomeassociations modified by other characteristics of householdstructure?Our findings for this question are suggestive but not definitive. In all of our models, effect estimates forthe interaction terms were subject to large uncertainty, so need to be interpreted cautiously. However,it is interesting to note that children whose mothers were the only adult members of their households755.2. Are the father’s co-residence status – child health outcome associations modified. . .tended to be less likely to be completely immunized than children who resided with both parents. Thisassociation persisted even after adjusting for potential confounding variables. However, children whowere living with their mothers and one or more additional adult relatives were just as likely to be com-pletely immunized as children living with both biological parents. Whether the additional relativesincluded one or more males did not appear to be important for this association.These findings are consistent with predictions of the time allocation model: a greater number ofadults in a household means a greater pool of time resources to allocate to the common goal of produc-ing child wellbeing. According to this model, the sex of the adults and their relationship to the childshould be unimportant. Routine childhood immunizations had been available in South Africa withoutcost to parents from the beginning of the time period considered in this study. While this may havereduced the importance of financial cost as a barrier, the opportunity cost of time spent taking childrenfor immunizations may have been more burdensome for mothers who did not have ready access to sup-port from household members, contributing to a lower chance of immunization completeness for theirchildren.In conflict with the time allocation model, previous research in South Africa has suggested thatmothers are more supported in allocating household resources to their children’s health in householdsmanaged by co-operating female relatives than in households managed by men [43, 60]. Althoughthe information was not available in the SADHS dataset to properly identify households composedof co-operating female relatives, we used the presence of co-resident adult female relatives and theabsence of co-resident male relatives as an proxy indicator for these types of households. Our modelsdid not demonstrate higher rates of immunization completeness among children living with only femalerelatives compared to children living with both biological parents or with their mother and one or moreadult male relatives, however, future research into this question would be valuable.Our model findings for the immunization and recent illness outcomes also suggest that co-residenceof fathers tends to be more beneficial when parents are married. For example, among children whosemothers were married, a co-resident father was associated with a higher likelihood of being completelyimmunized, but not among children whose mothers were unmarried. In addition, children with cohabit-ing but unmarried parents tended to be more likely to have had a recent respiratory or diarrhoeal illnessthan children with unmarried mothers and non-resident fathers. But the same does not seem to be truefor children with married, co-residing parents. These findings do not fit with predictions of the timeallocation model: co-resident fathers, whether married or unmarried, should increase available house-hold resources. Nevertheless, they are consistent with findings of prior research, which often suggestsoutcomes are poorer among children whose parents are in cohabiting versus marital unions [96, 116].A possible explanation is that, at least among some segments of South African society, cultural normsfavour marriage over cohabitation, as described in the literature review (section 2.1.3). In families andcommunities with strong traditional values, couples living in non-marital unions may experience so-cial isolation, while their children may lack the benefits of recognized membership in either father’s ormother’s kin groups. These may result in more negative outcomes for children. However, it appearsthat attitudes towards cohabitation differ between Black and White South Africans and between urban765.3. What is the magnitude of neighbourhood-level variation. . .and rural Black South Africans [59]. Non-marital cohabitation is clearly a complex phenomenon. Thehealth consequences (if any) for children of living with cohabiting as compared to either married orsingle parents cannot be adequately addressed with the analyses presented here. Temporal trends to-ward fewer marriages and more cohabiting unions in South Africa indicate this question is deserving offurther research attention [79, 58].5.3 What is the magnitude of neighbourhood-level variation in theco-resident father effect estimates and which contextualcharacteristics contribute to this variation?For all outcomes we found reasonably consistently high levels of variation across neighbourhoods inthe co-resident father effect estimates. However, the standard deviation parameters for the distributionsof these neighbourhood effects were all estimated with large uncertainty. This uncertainty is due tothere being few 0-4 year old children per neighbourhood in the SADHS dataset, making it impossible toprecisely estimate the within-neighbourhood co-resident father effects. For this reason, our analyses aremore useful for describing the magnitude of variation than for making comparisons between neighbour-hoods. Our results support the conclusion that, for each outcome, in some neighbourhoods co-residentfathers are, on average, associated with significantly increased odds of the outcome, and in others theyare associated with a significant reduction in the odds of the outcome.In general, our analyses using neighbourhood-level variables did not provide much insight into thefactors responsible for variations in the co-resident father effect estimates. We hypothesized that father’sco-residence would be more weakly associated with beneficial child health outcomes in poorer neigh-bourhood and in neighbourhoods with lower levels of employment. We also hypothesized that father’sco-residence would be associated with stronger positive effects in more gender-equitable neighbour-hood contexts. The only findings to support our hypotheses were: a) the strongly positive associationof father’s co-residence with immunization completeness in urban neighbourhoods, compared to noassociation in rural neighbourhoods, and b) the increasingly negative association between father’s co-residence and diarrhoea as the percentage of low-income households decreased (although this estimatehad large uncertainty). In fact, for the ARI outcome, there is some evidence to contradict our hypotheses:although not statistically significant, the direction of the interaction between father’s co-residence statusand a) percentage of female residents having completed high school, and b) the ratio of female to malehigh school graduates suggest that, as these indicators of neighbourhood gender equity increased, hav-ing a co-resident father became more strongly positively associated with recent ARI. Even where thecross-level interaction terms were significant, the estimated standard deviation of the neighbourhoodco-resident father effect estimates did not noticeably decrease. This implies that the neighbourhoodcharacteristics did not explain much of the variation of the within-neighbourhood co-resident fathereffects away from the average effect across all neighbourhoods in the sample.We selected neighbourhood characteristics with reference to a theoretical model of the determinantsof positive fathering [47]. This model identifies characteristics of the economic and cultural context as775.3. What is the magnitude of neighbourhood-level variation. . .important. At least in some areas of South Africa, the dominant societal expectation of men is to beeconomic providers for their families [32]. These societal expectations appear to be mirrored in youngfathers’ expectations of their personal involvement with their children. Despite recognizing other posi-tive types of involvement, their accounts emphasize financial provision as the foremost requirement ofbeing a good father [82]. We had hypothesized that fathers should be able to contribute to improved childhealth through pathways not limited to economic provision. However, in reality, societal and personalexpectations may make financial provision a pre-requisite to other types of positive involvement.In economic contexts of limited employment for low- and semi-skilled labourers, realizing a stableposition as ‘breadwinner’ may be impossible for many men. Perceived as failures in light of traditionalgender norms, unemployed or unstably employed men may be discouraged from supporting their chil-dren through alternative (traditionally feminine) caregiving activities [32]. Evidence from the UnitedStates shows that in impoverished neighbourhoods, where few men are able to replicate traditional pa-triarchal models of masculinity, alternative models of masculinity - which celebrate violence and sexualpromiscuity and de-emphasize responsibility towards women and children - can become the norm [183,18].There are a few possible explanations for why our neighbourhood-level analyses did not producefindings consistent with this previous research. First, the limited number of children per neighbourhoodin the dataset resulted in very uncertain neighbourhood-level effect estimates, and prevented us frombeing able to consider multiple neighbourhood-level variables together in a single model. Second, ourarea-level variables may have been inadequate for two reasons: 1) EAs may not have been the mostrelevant geographic level at which to measure all of the theoretical constructs of interest; and 2) indicesderived from aggregated census data may not have adequately measured the constructs. Our conceptualframework hypothesizes that the nature and level of fathers’ involvement in caring for children are in-fluenced by economic conditions and gender norms (see section 2.1.3). Related to the first point, theinfluence of economic conditions on the co-resident father – child health associations may have beenmore apparent at the level of districts, provinces or even countries, because government economic poli-cies are applied at these higher geographic levels [132]. Related to the second point, “norms” are rules ofbehaviour that are enforced by social sanctions [184, p. 1494]. The immediate community or neighbour-hood may well be the most appropriate level at which to study how social norms influence behaviour.However, census data can only give indirect measures of norms. In addition, variables derived fromcensus data could simultaneously be measuring more than one theoretical construct [131]. For exam-ple, the percentage of female neighbourhood residents with high school education could be an indicatorof both gender norms and of the accessibility of secondary schools. Future studies using more directmeasures of the constructs of interest could improve upon our analyses. A final possible explanation forour findings is that our models may not have adequately accounted for individual-level variability in theexposure-outcome associations. This would have made making inference about neighbourhood-levelsources of variability challenging. This last point is related to inherent limitations of the exposure andoutcome measures, which we discuss in the following section.785.4. Limitations5.4 LimitationsThe child health outcomes we investigated were a pair of upstream child health-promoting behavioursand a pair of illnesses directly responsible for a large percentage of child deaths in South Africa. Our de-scriptive statistics showed that children with and without co-resident fathers clearly differed on a numberof variables known to influence these health outcomes. In light of differing background circumstances,it is surprising that even our unadjusted analyses did not show clear differences in the outcomes of thesetwo groups of children (save for breastfeeding duration). Previous empirical evidence is not consistentas to whether, and under what circumstances, co-residing with a father is beneficial for young children’shealth in South Africa (see literature review section 2.2). Our findings contribute to this evidence bysuggesting that, all else being equal, merely residing separately from one’s biological father does notnegatively impact children’s outcomes on two critical health-promoting interventions of early life andon a pair of important childhood illnesses. In the preceding sections we have described some nuancesto these findings and issues requiring clarification in future research. However, there are importantlimitations to these analyses which also need to be considered.The first limitation is related to an inherent challenge with cross-sectional studies: the fact thatexposure, outcome and potential confounder data are all ascertained at the same point in time. Thisusually prevents one from clearly identifying the temporal ordering of the variables and dramaticallylimits the ability to infer causation from associations in the data. With respect to our research questions,a particular limitation of the SADHS data is that they do not describe children’s residential arrangementssince birth, only at the time of the survey. Household transitions appear to be a common feature ofchildhood for many South Africans. Only a small minority of children will have spent all of theirlifetime residing with both biological parents (estimated at 8% for children under 5 in a rural area of theKwaZulu-Natal province [81]). The timing at which children experience major household transitionshas been found to be important for psychological and academic outcomes, and some negative effectsof household disruption have been found to weaken over time [84, 77]. The same may be true for thephysical health outcomes studied here. In particular, we might expect that current co-residence witha father would be less relevant for the breastfeeding and immunization outcomes of older children,because, for most, these outcomes would have been completed within the first 1-2 years of life. Wetested this possibility by including an interaction between father’s co-residence status and child’s agein our full regression models, but did not find evidence of interaction for any of the outcomes (resultsnot shown). Nevertheless, future analyses using either longitudinal data on father child co-residenceor more detailed retrospective data on the timing of transitions in children’s residential arrangementswould be beneficial and may produce more definitive conclusions than those of our analyses. Futurestudies involving more detailed residential data should also attempt to distinguish between the effects ofhousehold transitions (for example, migration to a new household, which might be expected to reducechildren’s health by disrupting their existing linkages with the health care system) and effects due toresiding in a given household structure.A second limitation is that, although maternal recall is a common method of estimating illness periodprevalence in household surveys, validation studies have identified some concerns with the accuracy795.4. Limitationsof these data. Maternal recall appears to decline for illnesses having occurred more than 2-3 daysbefore the survey [185]. In addition, recall errors may occur more frequently among less educatedmothers [185]. A similar bias has been observed in maternal recall of breastfeeding duration, withmisreporting becoming more common with increasing time since breastfeeding cessation, and withmore highly educated mothers tending to over-report their breastfeeding durations [186]. Although wecontrolled for child age and maternal education, differential recall errors for breastfeeding duration andillness episodes could be a source of bias in our effect estimates. Another limitation of examining periodprevalence is that diarrhoea and respiratory infection are relatively common childhood illness. As anindication of children’s burden of exposure to pathogens it may be more meaningful to examine thenumber of illnesses episodes experience over a period of time. Alternatively, examining illness durationand severity may give more insight into the quality of care children receive and their capacity to recoverfrom illness. Because of the synergistic relationship between illness and malnutrition, it would also bevaluable in future studies of fathering and child health to examine these outcomes together. This wasnot possible using the 1998 SADHS data because the survey did not include anthropometry for children.A third limitation is that the SADHS did not collect any additional data about non-co-resident fa-thers. This prevented us from being able to control for some potentially important confounding factors,such as father’s educational attainment. It also prevented us from examining whether the reason for fa-ther’s absence influenced the effect of co-residence status on children’s health outcomes. For example,it may be useful to distinguish between fathers who spent periods away from their children as migrantlabourers and those who were permanently residing elsewhere because of no longer having a relation-ship with the child’s mother. This limitation is common to most household survey datasets. We echoother authors in noting that high rates of residential separation makes it essential, in order to adequatelystudy fathering in South Africa, for us to develop study designs which include data collection on bothco-resident and non-co-resident fathers [48, p. 258-9].A final limitation is that defining households based solely on residential criteria limits the value ofthe SADHS (and similar household surveys) for studying fathers. Russell argues that this approach todelineating households is based on “assumptions about the coherence and stability and exclusiveness ofco-residential groups, which ... hold good only under limited cultural and economic conditions.” [45,p. 6] She points out that north-western Europeans (and their White descendants in South Africa) followa kinship tradition based on conjugal ties, whereas southern Africans’ kinship traditions are based on“patrilineal or agnatic descent, i.e. descent from father.” [45, p. 8] Therefore, in studying the allocationof resources within African domestic groups, including from fathers to children, it may be necessary toconsider kinship ties rather than simply co-residential and conjugal affiliations.Two studies based in rural demographic surveillance sites have improved on the co-residential def-inition of households in studying children’s health outcomes. Investigators on these studies have dis-tinguished between the social and residential connections of household members [81, p. 5]. Theirfindings support the idea that children who have social connections with their fathers, even when notresiding together, do have better health outcomes than children who have no connection to their fathers.(Unfortunately, thus far, these analyses have been limited in the child health outcomes explored.)805.5. Next stepsIn contrast, our reliance on co-residence status means we treated fathers who were socially butnot residentially connected to their children as equivalent to those who had no connection at all withtheir children. To some extent, parent’s marital status may reflect fathers’ social connections with theirchildren. In this respect, our finding that, among children of married parents, father’s co-residencestatus was associated with no difference in any outcome except immunization completeness would beconsistent with the hypothesis that social connections with fathers are more important for child healththan residential connections. However, as discussed, parents’ marital status is only one dimension ofkinship. This is clearly an important area for future investigation.5.5 Next stepsUltimately, not having data on father’s actual parenting practices prevents us from being able to explainwhy, for the sample as a whole, residing with a father was not associated with children’s health out-comes, while in some neighbourhoods the association was strongly positive and, in others, stronglynegative. It is likely, for example, that a non-co-resident father who contributes financially and isinvolved in making important care decisions would be more beneficial for a child’s health than a co-resident but uninvolved father. Similarly, a completely absent father may have less of a detrimentaleffect on child health than an abusive one. Confounding of the father’s co-residence status – child healthoutcome associations by unmeasured fathering practices could explain why we observed an overall nullassociation but with considerable neighbourhood-level variability.Fathering data is extremely limited in South Africa. To date, essentially all studies of the relation-ship between fathering and children’s health outcomes have relied on father’s household membership(whether residential or social) as a proxy for his contributions to children’s care. As we have discussedabove, this approach has clear limitations. In the rare cases where data have been collected on SouthAfrican fathers’ actual involvement in caring for children, the focus has been limited to financial con-tributions. A previous study using these data has demonstrated that co-residence is a poor proxy forfinancial support: co-resident fathers were found to be no more likely than some non-co-resident fathersto be financially supporting their children [76]. It is likely that co-residence status is also an inadequateproxy for other aspects of father’s involvement [48, 20, 94].In order to facilitate future population-level research, which moves beyond the limited reliance onfather’s co-residence status as a proxy for his involvement in child care, we conducted a study to assess acomprehensive questionnaire for measuring father involvement in South Africa. We describe the detailsof this study in the following chapter.81Chapter 6Methods: Research Objective 2As discussed in chapter 5, data on fathers’ involvement in caring for children is severely limited inSouth Africa. Richer data on fathers’ parenting practices, as well as the antecedents and child healthconsequences of different patterns of fathering would be valuable for a number of reasons. For example,household surveys that included data on fathering could be used to compare the practices of resident andnon-resident fathers. In turn, these comparisons might enable us to explain why in some cases residen-tial separation of fathers and children appears to be harmful, whereas in others it does not. Togetherwith existing research from other settings, such data would allow for cross-cultural comparisons of fa-thering to be made. Perhaps most important, population-level fathering data could be used to identifywhich fathering practices are most beneficial for children’s wellbeing and inform interventions aimed atpromoting positive fathering.Ideally, efforts to collect population-level fathering data should be informed by preliminary researchinto approaches for overcoming current theoretical and methodological challenges. The scarcity ofprevious research on the parenting practices of South African fathers is, itself, a major challenge becauseit means there are no validated data collection tools and limited data from which to develop and testconceptual models [48]. Research assessing the validity of conceptual models and data collection toolsdeveloped for use in other cultural settings could begin to address this challenge.Another challenge to collecting fathering data in South Africa, particularly using household surveys,is the high percentage of fathers and children who live apart. Methodological strategies to address thischallenge might include using secondary respondents, such as mothers, as sources of data on fathers’parenting practices. But the reliability of mothers’ reports would also need to be assessed. Alternatively,recruitment and data collection strategies that accommodate both co-resident and non-resident fatherscould be developed.Our second set of research questions (2a-c in Chapter 1) are directed at beginning to overcome theabove challenges. This chapter describes the study design and statistical methods used to address theseresearch questions.6.1 Study designSetting and sampleOur investigation of fathering was a sub-study of a longitudinal cohort study called the Mother InfantHealth Study (MIHS). It involved families living in Kraaifontein, a suburb of Cape Town in the WesternCape Province of South Africa. Kraaifontein is located in a relatively wealthy district of the City of826.1. Study designCape Town Municipality, but is comprised of census suburbs of which some are the least ‘well off’ inthe district (based on an index which combines the percentage of low-income households, percentageof adults with less than high school education, unemployment rate, and percentage of the labour forcein unskilled occupations) [187]. While the majority of dwellings are formal, stand-alone houses, thereare also some ‘informal settlements’, which consist of high concentrations of dwellings that were “noterected according to approved architectural plans or on planned sites...” [187][127]. In these areas,access to basic services (electricity, potable water, flush or chemical toilets, and refuse removal) islimited [187].The MIHS consisted of a convenience sample of infants delivered between July 16th, 2012, andJune 28th, 2013, at the Kraaifontein Midwife Obstetric Unit. This primary-level health centre managesall low-risk deliveries by women living in the area, excluding those of women who pay to deliver inprivate hospitals. To be eligible to participate in the MIHS, a mother had to be at least 18 years old, aresident of the clinic catchment area, and not planning to move within one year of enrolment. Infantshad to be live-born, singleton births, at least 35 weeks gestational age, weigh at least two kilograms atbirth, and could not have any severe congenital abnormalities.The sample for the Fathering Sub-study included all mothers and infants from the MIHS who com-pleted the first study visit at 2 weeks. The sub-study also included a convenience sample of fathersof infants in the MIHS. For a father to be eligible to participate in the sub-study, the infant’s motherhad to consent for us to enrol the father. In addition, the mother had to complete the first study visit.Participation of fathers in the Fathering Sub-study was not required for mothers and infants to continueon either the sub-study or the MIHS.Recruitment proceduresMIHSFor the MIHS, recruitment staff held regular group information sessions in the antenatal clinic waitingarea. These sessions explained the aims of the study and what would be involved in participating.Their purpose was to familiarize pregnant women with the study, should they later be approached forenrolment.Mothers and infants were enrolled at the time of delivery. Enrolment was offered to all eligible HIV-infected mothers, and to a roughly equal number of HIV-uninfected mothers. The uninfected womenwere selectively offered enrolment based on their self-identified racial group, with the aim of attaininga similar racial mix in both groups. There are significant racial and geographic socioeconomic inequal-ities in South Africa. By recruiting participants from a single community, and by balancing the racialmix across groups, we hoped to minimize socioeconomic differences between HIV-infected and HIV-uninfected mothers. In addition, social class has an important influence on father involvement [47]. Forthe present study, we wished to limit this source of variation.836.2. Fathering questionnaireFathering sub-studyBased on previous study findings [188] and past experiences of our research study staff, we antici-pated relatively few fathers would be present at the clinic during delivery. At enrolment into the MIHSwe asked each mother to take an introductory letter to her child’s father (Appendix B.1). The letterexplained the Fathering sub-study and invited the father to participate. We also collected father’s tele-phone number if mother consented to his participation. We enrolled and consented fathers by telephonewhen their infants were 2 weeks old. A copy of the informed consent script is shown in Appendix B.2.Sub-study visitsSub-study visits were completed when infants were 2-weeks, 16-weeks, and 6-months old. Mothers andinfants attended study visits at our research clinic located in a tertiary-level hospital approximately half-an-hour drive away from Kraaifontein. Including travel, they would typically spend the entire morningand early afternoon at study visits. A previous South African study reported many men were unable toparticipate because of employment obligations [188]. We also expected that many fathers would not beresiding with their children, with some of them potentially living in other cities. Therefore, to maximizeparticipation of fathers, we offered them the option to complete sub-study visits by telephone or in-person at the study clinic. At each visit, fathers and mothers completed an approximately 30-minutefathering questionnaire with an interviewer. Fathers and mothers completed different versions of thesame questionnaire.Participants were offered organized study transportation or reimbursement of their travel expenses toand from the study clinic. In addition, mothers were given a R100 (approximately $10 CAD) honorariumfor each visit attended. Because many fathers completed the entire study by telephone, we did not givehonoraria to any fathers even if they attended in-clinic visits. However, all telephone call charges werecovered by the study.Institutional review and consentThe study protocol was approved by the human research ethics committees at Stellenbosch Univer-sity (Reference number: S12/01/009) and the University of British Columbia Children’s and Women’sHealth Centre (Reference number: H12-01181). All participating mothers provided written informedconsent. Participating fathers gave verbal informed consent over the telephone.6.2 Fathering questionnaireWe constructed our fathering questionnaire by drawing from a pool of survey questionnaire items orig-inally developed for the Developing a Daddy Survey (DADS) project in the United States [189]. Agoal of the DADS project was to develop a common set of theoretically and methodologically rigorousquestionnaire items for measuring father’s involvement in three national studies. Guided by Lamb andcolleagues’ framework (described in literature review section 2.1.1), the DADS project developed items846.2. Fathering questionnaireintended to measure five modes of paternal influence: Accessibility, Direct Caregiving, Responsibility,levels of financial contribution toward children’s material needs (Material Provisioning), and involve-ment in household chores (Practical Support for Mother). For our questionnaire, we selected items thatwere designed for use with fathers of infants and were meant to be suitable for both co-resident andnon-co-resident fathers. The complete list of fathering items included in our adapted questionnaire ispresented in Table B.1, Appendix B.3.In addition to questions about father’s involvement, we collected information about potentially im-portant covariates. At each visit, to provide a measure of ongoing interparental relationship quality, wehad fathers and mothers complete the ‘negotiation’ scale of the Revised Conflict Tactics Scales, and asub-set of items from the ‘psychological aggression’ and ‘physical assault’ scales26 [190]. The ques-tionnaire asked respondents to separately rate the frequency they experienced and perpetrated each actof negotiation or violence in the past three months.Additionally, at the first visit we collected information (from both fathers and mothers) on father’sinvolvement during the pregnancy, father’s demographic and socioeconomic characteristics, number ofprevious children sired, and characteristics of the mother-father relationship. In order to assess father-hood beliefs, we also asked mothers and fathers to rate their level of agreement with each of a set ofopinions about fathers’ responsibilities towards children (questions shown in Table B.2, Appendix B.3).The belief questions were adapted from the DADS questionnaire items with reference to research onfatherhood ideals in South Africa.At the first visit, mothers reported on their own demographic and socioeconomic characteristics.They also completed a roster for the members of their household, and questions about the structure oftheir dwelling place and their access to basic services. These questions were adapted from the 1998South African Demographic and Health Survey household member questionnaire [6]. Finally, fatherswere asked questions about how frequently they saw each of their biological parents during childhood,and the quality of their relationship with each parent. These latter questions were drawn from the poolof DADS project items.The questionnaire was professionally translated from English into isiXhosa and Afrikaans, the twomost commonly spoken languages in Kraaifontein. Interviewers administered the questionnaire to par-ticipants in their choice of the three languages. Before beginning data collection, the translated ques-tionnaire was back-translated into English by members of the research study staff and reviewed by theprincipal investigator to ensure that questions retained their original meaning. In addition, we askedthe interviewers to give us feedback on the appropriateness of the content of the questionnaire. Theinterviewers had similar racial and cultural backgrounds to many of the study participants. Interview-ers were given another opportunity to suggest changes after completing the questionnaire with the firstcouple of participants. Before beginning data collection the interviewers agreed that the majority ofitems on the questionnaire were appropriate. However, some interviewers felt that two Direct Care-giving items (singing songs or nursery rhymes to the child, and taking the child for walks) may not be26Our questionnaire included the following two items from the CTS-II psychological aggression scale: “Frequency insultedor swore at [person]”, and “Frequency threatened to hit or throw something at [person]”; and the following item from thephysical assault scale: “Frequency punched or hit [person] with something that could hurt.”856.3. Statistical methodsculturally relevant. We decided to retain these items in the questionnaire, but we further consider theirappropriateness in a discussion of the questionnaire’s Face Validity in Chapter 8. After completing thequestionnaire with the first few participants, the interviewers requested that we change the wording offour Accessibility items that were difficult for participants to answer. The original questions asked thenumber of hours fathers spent with the infant on an average work day and on an average non-work day,and the number of hours fathers spent alone with the infant on an average work day and on an averagenon-work day. We revised these questions to ask whether fathers usually spent an hour or more with theinfant in the morning, afternoon and evening on work days and on non-work days, and whether fathersusually spent an hour or more alone with the infant on work days and on non-work days. The revisedquestions were translated from English into Afrikaans and isiXhosa by study staff members. The revisedquestionnaire was first used with the 10th mother and the 6th father to complete 2-week visits. Mothersand fathers who had completed the original version of the questionnaire were excluded from analysesinvolving the revised Accessibility items.For illustration, the final English-language version of the questionnaire administered to fathers atthe 2-week study visit is shown in Appendix B.4.6.3 Statistical methods6.3.1 Description of the cohortTo characterize families participating in the MIHS, we computed descriptive statistics for sociodemo-graphic data reported by mothers at study enrolment and the first study visit (together referred to as“baseline”). We calculated frequency distributions for categorical variables and medians and interquar-tile ranges for continuous variables. To compare the characteristics of families with a participating fatherto those without, we recalculated baseline descriptive statistics stratified by father’s participation status.Although the strata are not random samples, we present p-values from bivariate tests of independence asa way of indicating characteristics that contrast strongly between families with and without a participat-ing father. Pearson’s chi-squared tests were used to assess independence of father’s participation statusand categorical baseline variables. Wilcoxon rank-sum tests were used for continuous variables.6.3.2 Description of fathers’ parenting practicesTo describe the parenting practices of our cohort of fathers, we calculated the percentage frequencydistribution of responses for each fathering questionnaire item. We used response data from mothersat the 2-week study visit because our interest was in describing the reported practices of fathers of allinfants in the cohort (whereas, fathers’ reports were only available for a subset of infants). We preparedstacked bar charts to visually compare response frequencies across sets of items intended to measure thesame mode of paternal influence.Next, we examined whether the reported fathering practices of our sample were associated withvariables identified as determinants of positive fathering in the theoretical framework (Section 2.1.3).866.3. Statistical methodsUsing questionnaire data given by mothers at their 2-week visits, we prepared cross tabulations betweeneach fathering item and a small number of covariates. These covariates measured individual charac-teristics of father, mother and child; characteristics of the father-mother relationship; characteristics ofthe household context; and characteristics of father’s involvement during pregnancy. To present theseresults graphically, we prepared stacked bar graphs showing the frequency distribution of each fatheringitem grouped into strata specified by level of the covariate. We calculated gamma statistics to describethe monotonic trend association between each pair of variables in the cross tabulations. Gamma is cal-culated as the difference between the number of concordant and discordant pairs of observations in thecontingency table, divided by the sum of concordant and discordant pairs27 [191, p. 57-58]. It measuresthe difference between the probability of concordance and the probability of discordance for a pair ofvariables. In a cross-tabulation of variables X and Y , a pair of observations is concordant if the observa-tion having the higher value of X also has the higher value of Y . The pair is discordant if the observationhaving the higher value of X has the lower value of Y . Gamma has a range of -1 to 1, inclusive. Avalue of 1 indicates a monotonic increasing relationship (i.e.: zero probability of discordance), while-1 indicates a monotonic decreasing relationship (i.e.: zero probability of concordance). The gammastatistic is only suitable for measuring associations between ordinal (including binary) variables. Forthese analyses we dichotomized all nominal fathering variables, as described in section 6.3.3. In thecase of nominal covariates, we calculated associations between the fathering variables and each pair oflevels of the covariate. In addition, for all cross tabulations, we calculated Pearson’s chi-square tests toidentify associations greater than would be expected by chance.Finally, we assessed whether the frequency fathers were reported to perform each fathering itemchanged as their children aged. For this analysis we restricted our dataset to data reported by motherswho had completed a fathering questionnaire at all three study visits (i.e.: when infants were 2 weeks, 16weeks and 6 months old.) We calculated a percentage frequency distribution for each fathering item bystudy visit. We then prepared stacked bar graphs showing the frequency distribution at each visit side-by-side. As with our other analyses, we present these graphs grouped by the mode of paternal influenceeach item was intended to measure. Friedman tests [192] were used to evaluate whether, on average,father’s ratings on each item differed significantly across study visits. The Friedman test accounts forthe non-independence of measurements on the same individuals at different time points.We also evaluated how representative the mothers included in the latter analysis were of the entiregroup: using Pearson’s chi-square tests for categorical variables and Wilcoxon rank-sum tests for contin-uous variables, we compared the sociodemographic characteristics of mothers with complete follow-upto all who completed the 2-week visit.Results of the above descriptive analyses are presented in Sections 7.1 and 7.2 of the followingchapter.27For 2 x 2 tables, Gamma is equivalent to Yule’s Q statistic, which can also be calculated from the odds ratio (OR) as:Q = (OR−1)/(OR+1) [191, p. 67].876.3. Statistical methods6.3.3 Modelling the fathering questionnaire response dataReliability is the consistency of repeated measurements [193, p. 3]. It is inversely related to the amountof error in the measurements. In contrast, the validity of a measurement is the extent to which it evaluatesthe construct of interest [194, p. 183]. Reliability is a necessary, but not sufficient, component of validity[193]. Therefore, to establish the validity of measurements from a questionnaire, it is necessary to assesstheir reliability.We sought to assess the reliability of the measurements of fathering from our questionnaire in twoways. We assessed the consistency of responses from two different observers (fathers and mothers).These analyses are described in section 6.3.5. We also assessed the consistency of individual observers’responses to related questionnaire items. These items were related in the sense that they were hypothe-sized to provide measures of a common underlying construct. Although we repeated our measurementsof fathering at three study visits, we did not assess the consistency of responses across visits becausefathering behaviours have been shown to change as children age [30].In the following sub-section we describe the theoretical basis of the Item Response Theory methodused to evaluate the reliability of conceptually-related questionnaire items. We then present the statisti-cal models used, and our analytic approach.Theoretical basis of Item Response TheoryWe conducted these analyses within a latent variable framework. In other words, we assumed thatthe observed parenting practices measured by the questionnaire items were manifestations of fathers’levels on a smaller number of unmeasured (i.e.: latent) variables. We hypothesized that each mode ofpaternal influence described in the theoretical framework (section 2.1.1) would correspond to a distinctlatent variable. Thus the hypothesized latent variables would reflect father’s level of: Direct Caregiving,Accessibility, Responsibility, Material Provisioning and Practical Support for the Mother. We alsohypothesized that each set of conceptually-related questions would measure just one latent variable. Theitems hypothesized to measure each mode of paternal influence are identified in Table B.1, AppendixB.3.Specifically, we tested the fit between the questionnaire data and our hypothesized latent variablestructure using a special type of Confirmatory Factor Analysis called Item Response Theory (IRT) mod-elling. IRT is a theory for “establishing the correspondence between latent variables and their manifes-tations” [193, p. 4]. Some specific terms used in IRT are defined in table 6.1 on the next page.IRT models were suitable for our analyses because they are designed to model responses fromcategorical items, and they are concerned with how individuals having different levels of a latent traitperform differently on a set of items intended to measure the trait [195]. Similarly to Factor Analysis,the latent trait is treated as a continuous variable in IRT models. Parameter estimates from IRT modelsreflect properties of both the items and the individuals on whom the data were collected. Thereby, IRTmodelling allowed us to assess the measurement properties of the fathering questionnaire and to suggest886.3. Statistical methodsTable 6.1: Definition of terms commonly used in Item Response TheoryTerm DefinitionTrait The continuous latent construct of interest.§Item A measurement instrument (for example, a question on a questionnaire)intended to provide a measure of the latent trait.Discrimination A measure of the item’s ability to distinguish between individuals havingdifferent levels of the trait. Equal to the slope term in a regression of the itemresponses on the trait.Item/categoryboundary locationThe point on the latent trait scale at which a person is 50% likely to choosethe designated response option or higher. Inversely proportional to theintercept term in a regression of the item responses on the trait.§In the results chapter we refer to our latent constructs of interest as “modes of paternal influence”or just “modes” rather than “traits”. We chose this naming convention because our theoretical modeldescribes fathers’ parenting practices as dynamic, not as fixed, inherent traits.possible improvements. It also allowed us to describe the parenting practices of our sample of fathers interms of their levels of the latent constructs of interest.Three assumptions underlie the IRT models we used for our analyses. The first is that responses toeach set of related items are a function of just one latent trait. Although multidimensional IRT modelsare possible, this unidimensionality assumption applies to all of the models we used [193, p. 20]. Asecond assumption is that an item’s responses are conditionally independent of responses to other itemsgiven the respondent’s level of the latent trait [193]. This implies that the responses of individuals whoshare the same level of the trait are uncorrelated [195]. The third assumption is that the data fit thefunctional form specified by the model. We describe the functional form of our statistical models next.Statistical modelsThe general IRT model, which is applicable to dichotomous and ordinal items, can be written in theform of a proportional-odds model [191]. The model expresses the probability that, for a given item i,father j is in response category k or higher, given his score on the latent trait, θ j:P(yi j ≥ k∣∣θ j) =exp[αik +βiθ j]1+ exp[αik +βiθ j](6.1)Similarly to a regression equation, this model includes (item-specific) slope and intercept parame-ters, β j and α jk, respectively. The difference from ordinary regression is that the predictor in this model,θ , is a latent variable. The slope is an indication of how rapidly the response probability changes asthe level of the latent trait increases. It is proportional to the strength of association between the itemresponses and the latent trait. In IRT the slope is also called the discrimination. The higher the discrim-ination the better the item distinguishes fathers with high levels of the latent trait from those with lowlevels. For dichotomous items there is a single intercept, whereas for ordinal items there is an intercept896.3. Statistical methodsfor each response category but the lowest. The probability of selecting the lowest response category orhigher is 100% so that category does not have an associated intercept. For dichotomous items, the aboveequation corresponds to the 2-Parameter Logistic model and, for ordinal items, to Samejima’s GradedResponse model [193].The cumulative probability function in equation 6.1 specifies a logistic (sigmoid) function, with alower asymptote of 0 and an upper asymptote of 1 [193]. Furthermore, for ordinal items, categories arestrictly sequentially ordered and have a common slope. However, the slope is allowed to vary acrossitems. Whether these assumptions are justifiable can partially be judged by assessing the degree offit between the data and the statistical model. The approaches we used to assess model-data fit aredescribed in the next sub-section.Classically in IRT the following alternative formulation of the conditional probability expression isused:P(yi j ≥ k∣∣θ j) =exp[ai(θ j −bik)]1+ exp[ai(θ j −bik)](6.2)In this formulation, ai is the item discrimination (equal to βi in equation 6.1), and bik is the item (orcategory boundary) location28. The location gives the point on the latent trait scale at which a personis 50% likely to choose the designated response option or higher. A smaller, or more negative, valueindicates that a lower level of the latent trait is needed for that response (or higher) to be 50% likely.The majority of fathers would be expected to select this category or higher. Whereas, a larger locationvalue means that a higher level of the trait is needed before one is 50% likely to chose that response orhigher, and fewer fathers would be expected to do so.Note that equation 6.1 models the conditional probability boundary between response category k orhigher and response category k− 1 or lower. The plot of this cumulative probability function is calleda Category Characteristic Curve (or an Item Characteristic Curve in the case of a binary item) [193].For binary items there is only one category boundary, so the probability of being in category 1 or higherversus 0 is equivalent to the probability of being exactly in category 1. In contrast, for ordinal items, theprobability of being exactly in a specific category must be calculated by taking the difference betweenthe cumulative probability functions for adjacent response categories [193], i.e.:P∗(yi j = k) = P∗(yi j ≥ k)−P∗(yi j ≥ k−1)where P∗ is the conditional probability given the latent trait. A plot of the above function is called aCategory Response Curve.Together with item parameters, IRT models can be used to estimate each father’s level of the la-tent trait. These levels are called person locations and they are estimated on the same scale as theitem/category boundary locations. We used the Expected A Posteriori (EAP) method to estimate personlocations [196]. This method uses a Bayesian approach to combine information about a father’s position28An item or category boundary location can be calculated from the slope and intercept parameters in equation 6.1 as:bik =−βiαik .906.3. Statistical methodson the latent trait from the likelihood function of his observed item responses and from a prior prob-ability distribution reflecting our assumption about the distribution of the trait in the population [193,p. 76]. Typically, a standard normal prior distribution is used [193, p. 76]. By the EAP method, thefather’s location is calculated as the mean of the estimated posterior probability distribution. Includingprior information about what location values are most likely in the population allows location estimatesto be calculated for fathers who were in the lowest (or highest) possible response option on every item.Otherwise, the likelihood estimate for these fathers would be infinite [193, p. 75].From a fitted IRT model it is possible to estimate a Total Information Function for the set of itemsincluded in the model. Total Information is inversely proportional to the standard error of person locationestimates produced by the set of items. In contrast to a single measure of reliability, information variesas a function of the latent trait. In other words, total information is a measure of the precision of theperson location estimate at the specified point on the latent trait [193, p. 27]. A benefit of information isthat, in contrast to reliability, it does not depend on the characteristics of the questionnaire respondents,only on the characteristics of the items in the questionnaire [193, p. 29]. Total information for a set ofitems is calculated as the sum of the separate information functions for each item. For ordinal items,item information is the sum of the separate information functions for each response category.Analytic approachFor all of the analyses described in the current section we used questionnaire data reported by mothersat the 2-week visit. We used mothers’ reports because we were interested in modelling the fatheringpractices of the complete set of fathers, not only those who participated in the study. We used the2-week visit data because this was the single largest dataset available.For most Material Provisioning items, we observed few “Mainly mother” and “Mainly someoneelse” responses. Similarly, for the Responsibility items we observed few “Mainly father” and “Mainlysomeone else” responses. To make the analyses for these items more informative, we recoded them asbinary variables indicating whether the father had a primary-level of involvement in the item versus not.Fathers reported to be mainly involved or involved together with the mother were recoded as havingprimary involvement. If it was reported that mainly the mother or mainly someone else was responsiblefor the item, fathers were coded as not having a primary-level of involvement.For most ordinal items with six response categories, such as items measuring the frequency fathersperformed direct caregiving activities, we observed few fathers in the “Rarely”, “A few times a month”and “About once a day” response categories. In order to minimize small cell size problems in analysesinvolving these variables, while also retaining as much information as possible, we recoded these as4-level ordinal variables. To do so, we collapsed the “Rarely” and “A few times a month” categoriesand the “A few times a week” and “About once a day” categories, and left the “Not at all” and “Morethan once a day” categories unchanged.Our analytic steps were as follows. First, we calculated the correlation between all pairs of fatheringitems (Spearman rank-correlations for pairs of ordinal items, phi coefficients for pairs of binary items,and rank-biserial correlations for binary-ordinal item pairs). These pairwise associations allowed us to916.3. Statistical methodsassess whether items within sets were associated as hypothesized, and also whether there were any largeassociations between items in different sets.Next, we fit a unidimensional IRT model to each of the five sets of fathering items specified in tableB.1 (Appendix B.3). When fitting the Accessibility trait model we excluded items measuring time spentwith infants on work days because 10% of fathers were not working at the time of the 2-week studyvisit. We also excluded an item measuring whether the father spent time with his infant on non-work-day evenings, because it was highly correlated with a similar item measuring time spent together onnon-work-day afternoons. Similarly, when fitting the Responsibility trait model, we excluded an itemmeasuring whether the father took the baby to the clinic or doctor because it was highly correlated withthe Talk mother, Important Decisions, and Decided Name items. In addition, we excluded from our IRTanalyses items which had large percentages of “Not Applicable” responses. These were comprised of asingle item for the Responsibility trait model (decided when to introduce solid foods), and two items forthe Material Provisioning trait model (pays for medicines, and pays for toys).IRT model parameters were first estimated using Weighted Least Squares with mean- and variance-adjusted Chi-square test statistics (WLSMV), as implemented by default in Mplus software version7.3 [197]. Our initial models had a similar form to that shown in equation 6.1, except that we used acumulative standard normal distribution (with a probit link) instead of the logistic distribution29. Fromthese models we obtained the following commonly-used indices of model fit: chi-square fit statisticand p-value, Root Mean Square Error of Approximation (RMSEA), Comparative Fit Index (CFI) andTucker-Lewis Index (TLI) [198].We then re-estimated the model parameters using a Maximum Likelihood estimator with standarderrors and chi-square test statistic adjusted to be robust to the non-normality of our response data (MLR).Using the MLR estimator we were able to specify the model using a logistic functional form, exactlyas in equation 6.1. From these models we obtained chi-square fit statistics and p-values, item parameterestimates (and associated standard errors), and estimates of each father’s location on the latent trait. Wealso obtained estimates of total and item information at a series of points along the latent trait. In theresults, we present information graphically in the form of Information Curves. We prepared histogramsto graphically display the distribution of father’s location estimates along each latent trait.Note that, by default, Mplus estimates item (and category boundary) thresholds instead of intercepts.Thresholds are equal to the negative of item intercepts, i.e.: the βik’s in equation 6.1. Thresholds aretherefore proportional to, but not equal to, item locations (bik’s in equation 6.2). Because standard errorestimates calculated in Mplus are on the same scale as the thresholds, we present the thresholds estimatesin our results tables. But, because of their more intuitive interpretation, we also describe items in termsof their locations. We calculated item and category boundary locations by dividing each threshold bythe associated slope.As per the standard approach in IRT, we used the estimated distribution of the latent trait in thepopulation to fix the scale of the trait. We set the origin of the latent trait to be equal to the populationmean, and the unit of measurement equal to 1 standard deviation [199, p. 280]. Both the item and29The reason for this difference is that the WLSMV estimator in Mplus can only be used with a probit link.926.3. Statistical methodsperson location estimates were measured on this same scale. For person location estimation, we utilizedthe default standard normal prior distribution.Models were taken to have acceptable fit if both chi-square fit test p-values were >0.05 (implying thenull hypothesis of perfect fit was not rejected), RMSEA was <0.05, and CFI and TLI were both >0.95[198, 200]. In addition, for models fit by MLR, we examined tables of bivariate model fit informationto check that few standardized residuals had an absolute value >1.96.We included fathers with missing item response data in our IRT analyses, as per the default inMplus. With both the MLR and WLSMV estimators, this was accomplished using full informationmaximum-likelihood parameter estimation, which assumed the missing data were Missing at RandomLast, we estimated the correlation between each pair of latent traits. These correlations were es-timated in Mplus by fitting the five separate IRT models described above together in a single model.The result is a structural equation model comprised of five separate measurements models (one for eachlatent trait) and a structural part specifying that the correlation parameters among the continuous latenttraits are to be freely estimated. The parameters of this model were estimated using the WLSMV es-timator because it is less computationally intensive than the MLR estimator for models with multiplelatent variables. Model fit was assessed using the fit statistics described in the paragraph above. Whenfitting the model, we requested modification indices to be calculated. Mplus calculates these indices forall parameters in the model that are fixed. Modification indices give the expected drop in the chi-squarestatistic if the parameter were allowed to be freely estimated [197, p. 726]. In other words, they identifyparameter constraints that could be relaxed to improve model-data fit. In our model, the constraints ofinterest were those specifying zero slopes for items hypothesized not to measure a particular latent trait.We considered removing slope constraints associated with large modification indices in cases where itseemed theoretically reasonable for the item to be measuring the latent trait in question [201].Results of the IRT analyses are presented in Subsections 7.3.1 and 7.3.2 of the following chapter.6.3.4 Comparing co-resident and non-co-resident fathers and testing for differentialitem functioningTo compare the reported parenting practices of co-resident and non-co-resident fathers at 2 weeks, weprepared cross tabulations of the fathering items and a binary variable indicating whether father andinfant were living together at the 2-week visit. To present these results graphically, we prepared stackedbar graphs showing the frequency distribution of each fathering item stratified by the co-residence statusvariable. To assess for confounding caused by no longer being in a relationship with the mother, we con-ducted the following sensitivity analysis: we excluded fathers who were reportedly “no longer seeing”the mother at the 2-week visit and reran the fathering item by co-residence status cross tabulations. Wepresent the results of the sensitivity analysis graphically, as described above, in Appendix B.8 (FiguresB.15 to B.20).Next, to determine whether residing with the child was associated with having significantly differentlevels on each latent mode of paternal influence, we constructed a regression of each latent trait on thefather’s co-residence status variable. In each regression model, the latent trait is measured by its effect on936.3. Statistical methodsthe observed fathering items, exactly as in the IRT models described above. This type of model is calleda Multiple Indicators Multiple Causes (MIMIC) model [202]. The coefficient in the regression of thelatent trait on the covariate was interpreted similarly to the slope term in an ordinary linear regression.The difference from an ordinary regression is that, in this case, the dependent variable was latent.Differential Item Functioning (DIF) is when an item has a different relationship with the latent traitfor different sub-populations specified by a covariate [203, p. 324]. In this study, we were interested indetermining whether our questionnaire could be used to measure the contributions of both co-residentand non-co-resident fathers. Therefore, we assessed whether father’s co-residence status was associatedwith differential functioning of items on the questionnaire. To test for DIF, we extended the MIMICmodels described above by including regressions of selected fathering items on the father’s co-residencestatus variable. A statistically significant regression coefficient was interpreted as evidence for DIF.Because the fathering items were all categorical, these regression coefficients were interpreted as logodds ratios. What was being estimated, however, was the direct effect of father’s co-residence status onthe fathering item, after accounting for the indirect effect operating through the latent trait. We selecteditems for DIF testing based on the results of the cross-tabulations described in section 6.3.2, above. Weassessed for DIF when the cross-tabulation results suggested an item had a different association withfather’s co-residence status compared to other items measuring the same latent trait. Parameters in theabove models were estimated in Mplus using the MLR estimator.Results of the MIMIC models are presented in Subsection 7.3.3 of the following chapter.6.3.5 Agreement between fathers’ and mothers’ questionnaire responsesThe second means of assessing the reliability and validity of the fathering questionnaire data was tocompare agreement between the responses of paired fathers and mothers. We recognize that neitherparent’s responses can be treated as a ‘gold standard’ because each is likely to be influenced by theirpersonal perceptions about the adequacy of the father’s parenting, which are, in turn, likely to be in-formed by an array of individual, interpersonal, and societal factors. By assessing agreement we canidentify individual questions, or sets of related questions, that tend to elicit consistent responses fromboth parents. We would then have reason to be more confident about the validity of responses for theseitems. In addition, finding reasonable agreement between fathers’ and mothers’ questionnaire responseswould suggest that mothers could be used as reliable proxies for fathers in future research on fathering.In contrast, we would be less confident in the reliability and validity of items for which the level offather-mother agreement is low because at least one parent’s responses must differ from the father’s truepractices [204].For these analyses we used mothers’ and fathers’ responses from the 2-week study visit. Exceptwhere described below, mother-father pairs were excluded from the analyses for a given questionnaireitem if either parent had a non-valid response for that item (i.e.: a “don’t know” or missing response). Weanalyzed paired responses for all fathering questionnaire items listed in Appendix B.3, except for threeitems that had large percentages of “not applicable” responses. The excluded items measured whetherthe father paid for his child’s medicines or toys, and whether he had responsibility for deciding when his946.3. Statistical methodschild would begin eating solid foods. We also analyzed paired responses for a set of variables measuringthe father’s demographic, socioeconomic and health characteristics, and characteristics of the father’srelationship with the mother. We included “Don’t know” responses when estimating the proportionsof overall agreement for two variables measuring characteristics of father’s employment history and asingle variable measuring whether the father had ever had an HIV test. For these variables, 5-10% ofmothers had “Don’t know” responses. Therefore, excluding these mothers could have upwardly biasedthe estimates of overall agreement. A revision to the questionnaire caused the first 10% of mothersand fathers who completed the questionnaire to have missing data for six Accessibility items measuringwhether the father spent an hour or more with the infant at different times of the day. Because itwas caused by a questionnaire change, we considered these data to be missing completely at randomand, therefore, excluded mother-father pairs with missing data from the analyses involving these sixvariables.For each variable we calculated the observed proportion of overall agreement and the proportion ofagreement specific to each response category. The latter index gives the conditional probability, giventhat one randomly selected parent in a pair choses the response category, that the other parent will alsodo so [204]. We calculated 95% confidence intervals for these raw agreement indices using the non-parametric bootstrap method using 1000 simulated datasets. Only three variables had >5% (but <10%)“Don’t know” responses: “Father’s educational attainment”, “Number of months father worked in pastyear”, “Whether father has ever had a HIV test”. Excluding this large percentage of records because of“Don’t know” responses could upwardly bias estimates of overall agreement. Therefore, we retained“Don’t know” as a valid response option when calculating percentage overall agreement for these threevariables.Next we assessed for marginal homogeneity of father’s and mother’s responses on each item. Thepurpose of evaluating for marginal homogeneity was to compare fathers’ and mothers’ propensitiesto choose each response category. We used histograms to visually assess for differences in fathers’and mothers’ response profiles on each item. To test for departures from marginal homogeneity thatwere greater than would be expected by chance, we used the McNemar test for binary items and theBhapkar test for nominal and ordinal items. The Bhapkar test is for marginal homogeneity acrossall response categories [191, p. 424]. In addition, for ordinal and nominal items, we used the non-parametric bootstrap method to test for significant differences between fathers’ and mothers’ marginalproportions for each response category individually.All of the ordinal items in the questionnaire can reasonably be considered to be measuring un-derlying continuous variables, i.e.: the frequency the father performed a particular parenting practice.During the design of the questionnaire, these underlying variables were divided into discrete responsecategories. Both the definition of the underlying variable itself and the definition of the threshold be-tween each pair of adjacent response categories are somewhat subjective. Therefore, for each of ourordinal items, two components of disagreement can be distinguished: mothers and fathers may disagreeabout the definition of the fathering practice being assessed; alternatively, they may disagree about thethresholds between different frequencies of performing the fathering practice [204].956.3. Statistical methodsWhen a pair of ordinal variables are each assumed to measure an underlying continuous trait, thepolychoric correlation between them gives an estimate of the correlation between the underlying con-tinuous traits [205]. We estimated the polychoric correlation between mothers’ and fathers’ responseson each ordinal item. An important assumption of the polychoric correlation is that the underlying traitsare normally distributed [205]. We used the likelihood ratio chi-squared test to test whether our pairedquestionnaire response data fit this assumption. Both the polychoric correlation and likelihood ratio teststatistic were calculated using the “polycor” package in R software (version 2.14.1) [206]. In addition,we summed each parents’ responses on the sets of items included in the Item Response Theory models(described in the previous section), yielding an “item total score” for each latent mode of paternal in-fluence. These scores reflect mothers’ and fathers’ ratings of fathers’ positions on a continuous variablehypothesized to underly each related set of questionnaire items. Therefore, they offer an alternativeapproach for estimating the correlation between the continuous traits underlying mothers’ and fathers’questionnaire responses. We estimated the strength of association between mothers’ and fathers’ itemtotal scores using Pearson correlation coefficients.Next, to evaluate whether mothers and fathers agreed over the response category thresholds foreach item, we calculated tests of category threshold equality and tests of overall bias. We tested the nullhypothesis of equal response thresholds for mothers and fathers using the method described by Uebersax[204]. To describe this test briefly, consider a cross-tabulation of fathers’ and mothers’ responses to ahypothetical ordinal item with K response categories. Threshold equality for response category k canbe tested by collapsing rows and columns for response categories 1 to k-1 and collapsing rows andcolumns for response categories k to K, then performing a McNemar test on the resulting 2x2 table.This procedure is repeated for all K−1 thresholds.While tests of threshold equality focus on each category threshold separately, tests for bias evaluatewhether, overall, fathers tends to give responses that are higher or lower than mothers’ on a given item[204]. Again, the test is performed using a cross-tabulation of fathers’ and mothers’ responses for eachordinal item. Response frequencies above the main diagonal of the table are summed to give b andresponse frequencies below the main diagonal are summed to give c. As with the McNemar test, the teststatistic is then calculated as:X2 =(b− c)2(b+ c)and compared for statistical significance against a chi-squared distribution with one degree of free-dom (Discrete multivariate analysis: theory and practice (1975) as referenced in [204]).We did not adjust the type I error rate for multiple comparisons because we were more interested inidentifying patterns of agreement for sets of items thought to be measuring the same underlying modeof fathering than in the results of significance tests for individual items.Finally, because we assumed fathers’ and mothers’ responses to the fathering items would be, inpart, influenced by their fatherhood beliefs we used the same methods described above to calculateagreement on a set of six fatherhood belief questions (Table B.2, Appendix B.3). In contra