UBC Faculty Research and Publications

How to use SAS® Proc Traj and SAS® Proc Glimmix in respiratory epidemiology Arrandale, Victoria; Koehoorn, Mieke; MacNab, Ying; Kennedy, Susan M. 2006-12-31

You don't seem to have a PDF reader installed, try download the pdf

Item Metadata

Download

Media
52383-Arrandale_guidance_doc.pdf [ 448.81kB ]
Metadata
JSON: 52383-1.0048205.json
JSON-LD: 52383-1.0048205-ld.json
RDF/XML (Pretty): 52383-1.0048205-rdf.xml
RDF/JSON: 52383-1.0048205-rdf.json
Turtle: 52383-1.0048205-turtle.txt
N-Triples: 52383-1.0048205-rdf-ntriples.txt
Original Record: 52383-1.0048205-source.json
Full Text
52383-1.0048205-fulltext.txt
Citation
52383-1.0048205.ris

Full Text

1How to use SAS? Proc Traj and SAS? Proc Glimmix in   Respiratory EpidemiologyVictoria Arrandale 1 , Mieke Koehoorn 2 , Ying MacNab 2 , Susan M. Kennedy 11 School of Occupational & Environmental Hygiene2 Department of Health Care & Epidemiology University of British Columbia, Vancouver, CanadaDecember, 2006 2Table of ContentsIntroduction   3Goal   3How to Use This Document   3SAS ?  Trajectory Procedure   4  Overview    4  Requirements    4  Data Organization    5  Dummy Variables   6  Missing Data   6  Types of Research Questions    6  Syntax   6  Selecting the Best Model   8  Output   1 0  User Information   1 1  Cautions   1 1  Reference Texts       1 1SAS? Glimmix Procedure   1 2  Overview    1 2  When to Use Mixed Effects   1 2  Requirements   1 2  Data Organization    1 3  Dummy Variables   1 3  Missing Data   1 4  Types of Research Questions    1 4  Syntax   1 4  Selecting the Best Model   1 5  Output   1 5  User Information   1 6  Cautions   1 6  Reference Texts   1 7Reference List   1 73How to use SAS? Proc Traj and SAS? Proc Glimmix in Respiratory EpidemiologyIntroductionThis document outlines the use of two procedures capable of modeling repeated respiratory symptom data in the software package SAS?: Proc Traj and Proc Glimmix.SAS? Proc Traj is a discrete mixture model which models the patterns of change over time in multiple subgroups within the population.  SAS? Proc Glimmix is a procedure that ?ts a generalized linear model to non-linear outcome data either with or without random effects.GoalThe goal of this document is to provide a concise user ?s guide for applying discrete mixture models (Proc Traj) and generalized linear mixed models (Proc Glimmix) in the analysis of longitudinal respiratory symptom data using SAS? software.   This document does not attempt to describe the statistical theory behind either of these techniques.How to Use This DocumentThis document presents an outline for setting up models in both Proc  Traj and Proc Glimmix for analyzing repeated respiratory symptom outcomes.  Data organization is explained, the modeling procedure is outlined, the basic syntax (appropriate for binary respiratory symptom outcomes, although Proc Traj will handle other outcomes) is described and the relevant modeling possibilities are discussed.  This document should be a starting point for modeling using Proc  Traj and Proc Glimmix.  Readers are advised to refer to the SAS? documentation as well as the noted reference texts for further explanations and for con?rmation that the models are appropriate for the data in use.Please Note:  This document was produced as part of an MSc thesis. 1   It is correct to the best of our knowledge; however, the authors take no responsibility for for the performance of the procedures or for the correctness of any research results.  This document is liscenced under the Creative Commons Attribution-Noncommercial-Share Alike 2.5 Canada License.  Please feel free to share this document under the following conditions:-  You must provide attribution to the authors.-   You may not use this work for commercial purposes.-  For any reuse or distribution, you must make clear to others the license terms of this work.-  Any of these conditions can be waived if you get permission from the authors.To request a paper copy of this document, please send a stamped, self-addressed envelope to: The Centre for Health and Environment Research, 2206 East Mall, Vancouver, BC, V6T 1Z3.  Correspondence of a scienti?c nature should be sent to: Dr Susan M. Kennedy, 2206 East Mall, Vancouver, BC V6T 1Z3; e-mail kennedy@interchange.ubc.ca1 Arrandale VH. An Evaluation of Two Existing Methods for Analyzing Longitudinal Respiratory Symptom Data [M.Sc. Thesis]. Vancouver: University of British Columbia; 2006.4SAS? Trajectory ProcedureOverview The SAS? Trajectory Procedure (Proc Traj) is a user-friendly ?nite mixture model procedure designed to run easily on the SAS? platform.  Proc Traj is a specialized mixture model that estimates multiple groups within the population, in contrast to a traditional regression or growth curve model that models only one mean within the population.  Designed by researchers, Proc Traj is not part of the base SAS? program and must be downloaded separately.Proc Traj is designed to address research questions focused on describing the trajectory, or pattern, of change over time in the dependent variable, speci?cally questions concerned with multiple distinct patterns of change over time and modeling unobserved heterogeneity in the data.  Proc Traj estimates a regression model for each discrete group within the population. The focus of the Proc Traj procedure is on group membership and identifying distinct subgroups within the population.  Proc Traj does not provide any individual level information on the pattern of change over time; subjects are grouped and it is assumed that every subject in the group follows the same trajectory.  There is no random effect capability within the Proc Traj model. The documentation for SAS? Proc Traj is a peer-reviewed publication by Jones, Nagin and Roeder (1).  A recent text authored by D. Nagin (2) is a valuable reference for users of Proc Traj and should be reviewed by those interested in the statistical theory behind Proc Traj.Requirements To apply Proc Traj to your data, you need (at a minimum) multiple measures of the outcome of interest and information on the timing of the repeated measures.  It would also be helpful to have multiple measures on a number of covariates you are also interested in.You must also download the Proc Traj application from B. Jones? website 2   and have copied the ?les to the folders as directed on the website.2  http://www.andrew.cmu.edu/user/bjones 5Data Organization In order to use Proc Traj you must organize your data in a multivariate, or ?wide? format, where there is only one row of data for each subject and multiple observations included in one line of data.  An example of data ready for use with Proc Traj is shown in Table 1, a description of each variable is provided in Table 2.  You can see in Table 1 that the outcome variable ?wheeze? is denoted by the variables W ez01, Wez02 and Wez03.  These three variables correspond to three repeated measurements taken at three different times.  The time at which each of these measurements was collected is represented by the variables Yr01, Yr02 and Yr03.  If a subject did not complete a visit, all variables corresponding to that visit are blank, in this case a ?.? is used to indicate missing data.Table 1 Mock data set up for analysis with Proc TrajID Sex Byr Csmk01 Csmk02 Csmk03 Wez01 Wez02 Wez03 Yr01 Yr02 Yr03001 0 1947 0 0 0 0 0 0 1992 1994 1999002 1 1953 1 . 0 0 . 1 1992 1994 1999003 0 1951 0 1 1 1 0 1 1992 1994 1999004 0 1946 0 0 . 1 1 . 1992 1994 1999005 1 1950 1 0 1 1 1 1 1992 1994 1999Table 2 Description of variables in mock data (Table 1)Variable Name Description ValuesID Subject ID as assignedSex Sex of subject 0= male 1= femaleByr Year of birth continuous, in yearsCsmk01 Current smoker at visit 10= never or former smoker   1= current smokerCsmk02 Current smoker at visit 2Csmk03 Current smoker at visit 3Wez01 Response to wheeze question at visit 1  0= no wheeze 1= wheeze Wez02 Response to wheeze question at visit 2Wez03 Response to wheeze question at visit 3Yr01 Date of visit 1  values corresponding to dataYr02 Date of visit 2Yr03 Date of visit 3 The variables that describe repeated measures of the same outcome must be numbered consecutively (i.e. Csmk01, Csmk02, Csmk03 etc.) before Proc  Traj will accept them; this will usually require some recoding.  SAS? will not accept the data if the variables are labeled alternatively (i.e. smk1992, smk1994, smk1999) even if this is logical given your data set.  By identifying the variables that contain information on the date of each repeat measure (i.e. Yr01, Yr02, Yr03) you are specifying the space between repeated measures.  Time varying covariates (i.e. Csmk01, Csmk02, Csmk03 for smoking information at each visit) must also be named with consecutive numbers corresponding to the visit.6Dummy VariablesIt is advisable to create dummy variables for each of your covariates that you plan to input into a Proc Traj model, as was done for Current Smoking in Table 1.  Covariates can be input in a binary (dummy) form or a continuous form but Proc  Traj does not handle categorical covariates.Missing DataProc Traj is able to handle data that is missing completely at random (MCAR), but is unable to handle data that is missing for more complex reasons (2).  Missing data can be entered in the dataset as shown in Table 1.Types of Research Questions In terms of respiratory symptom data, Proc Traj should be used when your research question is similar to one of the following:?  Are there multiple patterns of change in the outcome??  How many patterns of change are there in the outcome??  What is the shape of the change over time??  What predicts membership in each of these groups??  What are the characteristics that differ (or are similar) between the different groups?SyntaxThe entire Proc Traj syntax is outlined on B. Jones? website 1   and should be referenced for any further questions regarding syntax.A simple Proc Traj syntax for a two group model of the respiratory symptom wheeze is presented here:proc traj data=a.mockdata out=out outstat=os outplot=op;var wez01-wez03;indep year01-year03;model logit;ngroups 2;order 0 1;id ID;run;%trajplot (OP, OS, ?Title of graph?, ?Subtitle?, ?Y-axis label?, ?X-axis label?);  7As in all SAS? procedures, the Proc Traj statement outlines the data set to be used and in this case also de?nes the output from the procedure.The ?var ? statement de?nes the binary symptom outcome of interest.  ?Indep? de?nes the time variables that you are modeling the outcome over.  The ?model? statement identi?es the outcome as binary and the ?ngroups? states how many groups you want to model.  ?Order ? assigns the order of each equation that will describe the change over time in each group.  ?ID? identi?es the subjects in your population and also denotes which variable you want to use to uniquely assign subjects to a speci? c group in the output data set (in this case, ?out?).The ?%trajplot? is a macro statement that results in the graphical output from Proc Traj.  This macro includes references to the outplot and outstat statements in the ?Proc  Traj? statement.  If you make any changes in the ?Proc Traj? statement be sure to adjust the trajplot macro accordingly.When including time independent covariates into a Proc Traj model, the ?risk? or ?tcov? statements can be added to the syntax for time stable covariates and time varying covariates respectively.  For example, a time stable covariate for sex could be added:proc traj data=a.mockdata out=out outstat=os outplot=op;var wez01-wez03;indep year01-year03;model logit;ngroups 2;order 0 1;risk female;id ID;run;%trajplot (OP, OS, ?Title of graph?, ?Subtitle?, ?Y-axis label?, ?X-axis label?);Or, a time varying covariate for current smoking could be added:proc traj data=a.mockdata out=out outstat=os outplot=op;var wez01-wez03;indep year01-year03;model logit;ngroups 2;order 0 1;tcov csmk01-csmk03;id ID;run;%trajplot (OP, OS, ?Title of graph?, ?Subtitle?, ?Y-axis label?, ?X-axis label?);8Selecting the Best ModelThe model ?tting procedure with Proc Traj is iterative and requires a priori decisions based on substantive knowledge.  In the most basic process, the following  steps  should be followed:1.  Decide on the maximum number of groups using a priori knowledge2.  Fit number of groups to data (start by ?tting a one group model, and then ?t up to the maximum logical number of groups in a step wise manner)3.  Select the shape of the pattern of change for each group over time (e.g. linear)4.  Perform further modeling if required (addition of covariates, inclusion of second outcome etc.)Steps 1 & 2:  To decide on the optimum number of groups for your data you must begin by ?tting a basic one group model with all groups set to a second order (quadratic) equation.  Then ?t a two group, then three group model etc. until you have ?t the maximum number of groups based on your a priori decision.  Nagin suggests setting all group orders to second order during this process (2).  For each model you ?t in this ?rst step you will be given two Bayesian Information Criterion (BIC) values in the output: one relates to the overall sample size (total number of observations), and the other relates to the subject sample size (number of subjects).  The true BIC for the model lies between these values (2).  The BIC is the log-likelihood adjusted for the number of parameters and the sample size (1).  In the Proc  Traj procedure the BIC values given in the output are negative; the best ? t model is the one with the smallest negative number.Model selection in Proc Traj uses the BIC to select the best ?tting model via two different methods.  The ?rst, described by Jones, Roeder and Nagin (1) uses the change in the BIC between two models to measure the weight of evidence against the null model.  For each increasingly complex model that is tested, the BIC of the more complex (larger number of groups, or higher order equation) less the BIC of the less complex model is used to select the model that better ?ts the data.    ?BIC  =  BIC (complex)  -  BIC (null)The difference in BIC between the two models is a measure of the evidence against the null model.  Jones, Nagin and Roeder (1) suggest criteria for strength of evidence against the null model (Table 3).  Using the difference in the logged Bayes factor between successive models, the difference between the alternate and the null model can be quali?ed.  The null model is always the simpler model (i.e. less groups, or lower order equations).  The interpretation of the logged Bayes factor (2?BIC) in terms of model preference is shown in Table 3.  9Table 3 Interpretation of logged Bayes factor (2*?BIC) for model selection (Adapted from Table 2 in (1))2*?BIC Evidence against Ho0 to 2 Not worth mentioning2 to 6 Positive6 to 10 Strong> 10 Very StrongThe second method is called Jeffreys?s scale of the evidence and is described by Nagin (2). Jeffreys?s scale of the evidence uses the exponentiated difference between the BIC values of  models, i and j:  Bayes Factor ? e BICi-BICjIn this case is does not matter which model is the null model; only that the researcher remembers which model is which.  The interpretation of Jeffreys?s scale of the evidence is outlined in Table 4.  Further description and explanation can be found in Chapter 4 of Nagin (2005) (2).Table 4 Interpretation of Bayes Factor (eBICi-BICj) for model selection (Adapted from Table 4.2 in (2))Bayes Factor (Bij) InterpretationBij < 1/10 Strong evidence for model j1/10 < Bij  < 1/3 Moderate evidence for model j1/3 < Bij < 1 Weak evidence for model j1< Bij < 3 Weak evidence for model i3< Bij < 10 Moderate evidence for model iBij > 10 Strong evidence for model iWhen selecting the ?best? model it is important to base decisions on substantive knowledge about the research area, and remember the rule of parsimony to select the simplest model that best describes the data.Again, in reference to the example with respiratory symptoms, if we tested ? ve models (one group up to ?ve groups) we would have ?ve BIC values to review.  The comparisons are completed in a step-wise manner so that the two-group model is compared to the one-group model, and the three-group model to the two-group model and so on.  In each case, the model with the smaller number of groups is the null model. Step 3:  The next step in ?tting a model using Proc Traj is selecting the shape of each group?s trajectory over time.  Proc Traj can model up to a fourth order polynomial and can model both linear and non-linear trajectories within the same model.   This can be done using substantive knowledge (i.e. we expect one group to never report symptoms so this group?s trajectory will be a zero-order equation, or a straight line) or it can be done using the ?BIC.  It seems ideal to use a combination of substantive knowledge and statistical inference to make the decision regarding the shape of each group? s trajectory.10OutputThe output from Proc Traj includes the parameter estimates for each group (with standard errors), group membership probabilities (population level) and model ? t statistics.  The output data set (out= in ?Proc Traj? statement) includes all the variables included in the analysis (not all the variables in the original dataset), the variable identi?ed in the ?id? statement, posterior subject speci?c group membership probabilities and a group assignment for each individual.The parameter estimates can be used to construct regression equations for each group and a system of equations to describe the population.  The relative differences between the estimates for the same covariate between groups can be used to make inferences about differences between the groups.The posterior group membership probabilities and the group assignment variables in the output data set can be used to explore between group differences in covariates not included in the model and potentially as predictor variables in separate analyses.   The posterior group probabilities are calculated for each individual based on the estimated parameters, and the individual is assigned to a group based on their highest posterior group probability (2).The output from a basic model (no covariates) is shown in Figure 1.   The intercept parameters represent the estimated intercept for each group.  For Group 2 the linear parameter represents the estimated coef?cients for the linear time component of the regression equation.  The group membership probabilities indicate what proportion of the population is estimated to belong to each group.  And, the BIC values are the ?nal portion of the output.  Note the BIC values are shown for two sample sizes; ? rst for all the data points and second for the number of subjects.Figure 1 Text output from basic Proc Traj model with no covariates11Figure 2 Graphical output from Proc Traj showing a two-group model with group membership probability and shape of change over time User InformationBase SAS? is required to run Proc Traj.  You can download the procedure from the B. Jones? Proc Traj Home Page and copy the downloaded ?les into the appropriate folders on your hard drive (instructions provided on web page).  Proc Traj will then be installed and functional in the SAS? platform.  Because Proc Traj is an add-on to SAS?, there is no formal SAS? documentation in the traditional SAS? format.  Users are advised to thoroughly review the reference texts listed below.CautionsResearchers using Proc Traj are advised to remember that the multiple groups estimated are not rei?ed groups.  The identi?ed groups are estimations of multiple patterns of change within the population, and we must be careful not to think of group membership and trajectory shape as absolute certainties.Reference TextsNagin, Daniel S.  Group-based Modeling of Development.  Harvard University Press: Massachusetts (2005).Jones, B., Nagin, D., & Roeder, K. A SAS Procedure Based on Mixture Models for Estimating Developmental Trajectories.  Sociological Methods & Research (2001) 29: 374-393.SAS? Proc Traj Home.  http://www.andrew.cmu.edu/user/bjones.  12SAS? Glimmix ProcedureOverview Within the SAS? program there are several procedure for constructing mixed models.  The most common procedure is Proc Mixed, which models continuous outcomes.   A newer procedure, Proc Glimmix (general linear mixed models), models non-linear data.   Proc Nlmixed (non-linear mixed models) also models non-linear data but is primarily for use with advanced modeling and is programmatically complex. This guide is focused on modeling binary respiratory symptom data using Proc Glimmix.  Proc Glimmix is a fast, ?exible procedure capable of running linear models (?xed effects), generalized linear models (?xed effects), linear mixed models (?xed and random effects) as well as generalized linear mixed models (?xed and random effects).  The focus of Proc Glimmix for this guide is the generalized linear mixed model capability.  Proc Glimmix does not ship with the SAS? v.9, but the add-on and the documentation are both available for download on the SAS? support website. 2Mixed models are particularly useful in the modeling of longitudinal data because repeated measurements are collected over time on subjects and are inevitably correlated.  A ?xed effects model requires all of the measurements to be independent; in a longitudinal repeated measures data set this assumption is violated.  In a longitudinal mixed effects regression model the autocorrelation between repeated measures on individual subjects is accounted for, one overall group mean is modeled and subject speci?c deviations from the group mean are estimated.  When to Use Mixed EffectsIn reference to longitudinal study designs, random effects should be introduced into a regression model when there are correlated outcome measures (i.e. repeated measures on individuals) and when you want to allow the effect of a particular covariate on the outcome to vary randomly among your subjects.  RequirementsFor a mixed model you should have outcome measures that you expect are correlated; this occurs when you have collected repeated measures on individuals.   To use Proc Glimmix, you should also have a non-linear outcome variable, in this case binary symptom data.  If you are using a continuous measure of a respiratory symptom, you should consult SAS? Proc Mixed.2  http://support.sas.com13Data Organization For Proc Glimmix models, data must be organized in a univariate, or ?long?, format where there is one observation per line of data and multiple lines of data per subject.  An example of data organized this way is shown in Table 5.Table 5 Mock data set up for analysis with Proc GlimmixID Age Vyr First VisitSecond VisitSex Fsmk Csmk Wez001 53 1992 0 0 0 0 0 0001 55 1994 1 0 0 0 0 0001 60 1999 0 1 0 0 0 0002 49 1992 0 0 1 0 1 0002 56 1999 1 0 1 1 0 1003 40 1992 0 0 0 1 0 1003 42 1994 1 0 0 0 1 0003 47 1999 0 1 0 0 1 1004 60 1992 0 0 0 0 0 1004 62 1994 1 0 0 0 0 1005 36 1992 0 0 1 0 1 1005 38 1994 1 0 1 1 0 1005 43 1999 0 1 1 0 1 1Table 6 Description of variables in mock data (Table 5)Variable Name Description ValuesID Subject ID as assignedAge Subject?s Age  age in yearsVyr Year of Visit date in yearsVis2 Two complete visits yes/noVis3 Three complete visits yes/noSex Sex of subject 0= male 1= femaleFsmk Former Smoker yes/noCsmk Current Smoker yes/noWez Response to wheeze question  0= no wheeze 1= wheeze The data setup is quite straightforward, but note that the data includes dummy variables for otherwise categorical variables (smoking, visit number).  It is easier to deal with dummy variables, rather than categorical variables, in Proc Glimmix.Dummy VariablesProc Glimmix does have a ?class? statement in the syntax, and therefore theoretically you can input categorical variables without any recoding.  However , it is not easy to adjust the reference groups using the class statement.  Instead, researchers are advised to create dummy variables for each categorical variable. 14Missing DataProc Glimmix does handle missing data.  Observations are not excluded if variable values are missing within the observation.  However, if the amount of missing data is substantial the speci?ed models may not converge.  In this case, you can limit your dataset to subjects with less missing data in an attempt to run the models successfully , but this will result in a smaller sample size and a loss of power . Types of Research Questions In terms of longitudinal respiratory symptom data, Proc Glimmix should be used when your research question is similar to one of the following:?  Considering the repeated measures on individuals, what are the risk factors that predict the outcome??  How much variation exists between individuals for a given main effect??  How are the repeated measures on individuals correlated??  Does the probability of the outcome change over time?SyntaxThe syntax for a basic Proc Glimmix model is outlined here.  For further discussion of the Proc Glimmix syntax, including the speci?cation of a marginal model for estimating correlation structures, readers should refer to the of?cial SAS? documentation (3).The ?rst model presented is a mixed model estimating the risk factors for wheeze:proc glimmix data=a.mockdata ;model wez (event=?1?)= age sex vis2 vis3 fsmk csmk / s dist=binary link=logit or ;random intercept / subject=case ;ods output oddsratios=a.oddratio ;run;Again, the procedure statement speci?es the dataset to be used.  The model statement indicates that the outcome is ?wez? and that Proc Glimmix is modeling the probability of ?wez=1?.  Beyond that, the model statement lists the covariates to include in the model (in this case they are all dummy variables except age) and the model options.  The included model options in this example are ?s? (can also be written as ?solution?) to provide the ?xed effects parameter estimates, ?dist? to specify the distribution of the outcome, ?link? to specify the link function and ?or ? to provide the odds ratios for the ?xed effects.  An explanation of the ?dist? and ?link? options as well as a table of possible values is provided in the Proc Glimmix documentation (3).  When the outcome is a binary respiratory symptom, the ?dist? option will be binary and the ?link? will always be logit.The random statement speci?es the random variables.  In this case only a random intercept was speci?ed, but any other random variables would be listed before the forward slash.  The random statement options used here are ?subject?, which identi?es the variable for which there are repeated measurements.  The entry here will always be the subject or case identi?cation variable when the repeated measures are on individuals. 15The odds ratio option in the model statement gives a very long output table that is dif?cult to interpret from the SAS? window.  For this reason it is advisable to use the ?ods output? statement to specify that the odds ratio table be output as a dataset.  Once the odds ratio table is seen as a dataset ?le it is much easier to interpret.  For more information on ODS output and how to limit the output to speci?c portions (using the ?ods select? statement) or output speci?c tables to a new dataset, refer to the SAS/STAT? documentation (4).Selecting the Best ModelUnfortunately there is no easy way to select the best ?tting model using Proc Glimmix.  Proc Glimmix does not provide a likelihood value for the estimated models, instead pseudo-likelihood is calculated and this value cannot be used in a likelihood ratio test.Instead, users are advised to construct their model in a stepwise manner using substantive knowledge.  A priori hypotheses should drive decision making while constructing the model.  Once the model is assembled, the signi? cance of individual estimates and prior knowledge should guide what remains in the model.  Additional ?t statistics can be requested in the Proc Glimmix statement by including the following the command:  IC = PQWhen this command is included, pseudo-AIC and pseudo-BIC values will be included in the output ?t statistic table.  In the case of both pseudo-AIC and pseudo-BIC values, a smaller value indicates a better model ?t.More information on the complexities of ?tting models in Proc Glimmix can be found in the documentation (3).  There is also an on-going discussion of this and other pertinent SAS? issues on the SAS? user ?s list serve (5).OutputProc Glimmix provides extensive text output in SAS?.  A sample of Proc Glimmix output, limited to the key pieces of output, is shown in Figure 2.   The ?t statistics are shown, including the pseudo-likelihood mentioned previously.  Covariance parameter estimates are the estimates of variance in each of the speci?ed random effects, in this case only a random intercept was included in the random statement.   The covariance parameter estimates provide a measure of the between subject variability in the random variable.  The next table shown is the estimates of the ?xed effects included in the model statement.  These are the regression coef?cients describing the effect of each independent variable on the probability of reporting the symptom outcome.  If odds ratios had been requested in the output they would follow after the ? xed effects parameter estimates.The complete Proc Glimmix output is extensive, including information on the model optimization, iterative process of model ?tting and the convergence criteria.  Speci?c portions of the default output can be selected for viewing in the output window using the ?ods select? statement (4) as was done in Figure 2.16User InformationProc Glimmix does not ship with SAS?, instead the procedure and documentation can be downloaded from the SAS? Support website.  The ?les are self-extracting and will copy all necessary ?les to the correct location (unlike Proc Traj, where you have to manually move the downloaded ?les into the correct folders).The Glimmix procedure is supported by SAS? and has traditional SAS? documentation (3).  In addition, the book SAS? for Mixed Models contains an intensive chapter (with examples) on generalized linear mixed models that should be reviewed.Figure 3 Sample output from mixed model using Proc Glimmix.CautionsResearchers using Proc Glimmix should be aware that there are acknowledged issues with the estimation technique used in Proc Glimmix and that the procedure may result in biased coef?cient estimates.  A statistician can assess the magnitude of this problem using simulation techniques.  Researchers should consult a statistician to ensure that their results are not biased.17Reference TextsArrandale VH. An Evaluation of Two Existing Methods for Analyzing Longitudinal Respiratory Symptom Data [M.Sc. Thesis]. Vancouver: University of British Columbia; 2006.Diggle PJ HP, Liang K, Zeger SL. Analysis of Longitudinal Data. 2nd ed. New York: Oxford University Press; 2003.Fitzmaurice GM LN, Ware JH. Applied Longitudinal Analysis. New Jersey: John Wiley & Sons; 2004.Littell R, Milliken G, Stroup W, Wol?nger R, Schabenberger O. SAS for Mixed Models. 2nd ed. Cary NC: SAS Institute Inc.; 2006.SAS Institute. SAS/STAT 9.1 User ?s Guide: SAS Publishing; 2004.SAS Institute. The Glimmix Procedure: SAS Publishing; 2006. 18Reference List(1)  Jones BL, Nagin DS, Roeder K. A SAS procedure based on mixture models for estimating developmental trajectories. Sociological Methods & Research 2001;29(3):374-393.(2)  Nagin DS. Group-based Modeling of Development. Cambridge, Massachusetts: Harvard University Press; 2005.(3)  SAS Institute. The GLIMMIX Procedure: SAS Publishing; 2006.(4)  SAS Institute. SAS/STAT 9.1 User ?s Guide: SAS Publishing; 2004.(5)  Archives of SAS-L@LISTSERV.UGA.EDU.  Retrieved on multiple occasions, from http://listserv.uga.edu/archives/sas-l.html.

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.52383.1-0048205/manifest

Comment

Related Items