Predictive marginal means for a generalised linear model, using the method of Korn and Graubard (1999) and matching the results of SUDAAN. The predictive marginal mean for one level of a factor is the probability-weighted average of the fitted values for the model on new data where all the observations are set to that level of the factor but have whatever values of adjustment variables they really have.
svypredmeans(adjustmodel, groupfactor, predictat=NULL)
A generalised linear model fit by
svyglm with the adjustment variable but without the factor for which predictive means are wanted
A one-sided formula specifying the factor for which predictive means are wanted. Can use, eg,
~interaction(race,sex) for combining variables.
This does not have to be a factor, but it will be modelled linearly if it isn't
A vector of the values of
groupfactor where you want predictions. If
groupfactor is a factor, these must be values in the data, but if it is numeric you can interpolate/extrapolate
An object of class
svystat with the predictive marginal means and their covariance matrix.
Graubard B, Korn E (1999) "Predictive Margins with Survey Data" Biometrics 55:652-659
Bieler, Brown, Williams, & Brogan (2010) "Estimating Model-Adjusted Risks, Risk Differences, and Risk Ratios From Complex Survey Data" Am J Epi DOI: 10.1093/aje/kwp440
It is possible to supply an adjustment model with only an intercept, but the results are then the same as
It makes no sense to have a variable in the adjustment model that is part of the grouping factor, and will give an error message or
Worked example using National Health Interview Survey data: https://gist.github.com/tslumley/2e74cd0ac12a671d2724
data(nhanes) nhanes_design <- svydesign(id=~SDMVPSU, strata=~SDMVSTRA, weights=~WTMEC2YR, nest=TRUE,data=nhanes) agesexmodel<-svyglm(HI_CHOL~agecat+RIAGENDR, design=nhanes_design,family=quasibinomial) ## high cholesterol by race/ethnicity, adjusted for demographic differences means<-svypredmeans(agesexmodel, ~factor(race)) means #> mean SE #> 2 0.114596 0.0065 #> 3 0.084718 0.0106 #> 1 0.123081 0.0058 #> 4 0.108770 0.0304 ## relative risks compared to non-Hispanic white svycontrast(means,quote(`1`/`2`)) #> nlcon SE #> contrast 1.074 0.0722 svycontrast(means,quote(`3`/`2`)) #> nlcon SE #> contrast 0.73928 0.0923 data(api) dstrat<-svydesign(id=~1,strata=~stype, weights=~pw, data=apistrat, fpc=~fpc) demog_model <- svyglm(api00~mobility+ell+hsg+meals, design=dstrat) svypredmeans(demog_model,~enroll, predictat=c(100,300,1000,3000)) #> mean SE #> 100 699.19 10.3203 #> 300 684.29 9.5242 #> 1000 632.13 9.3377 #> 3000 483.11 22.8229