# Subgroup Analysis - Interactions and estimates The function can examine Cox regression, logistic regression and Poisson regression (Poisson regression for survival analysis) where the effect of one variable is of particular interest. This function systematically checks for effect modification with a list of other variables. In randomised studies the main regression analysis is often univariate and includes only the exposure of interest. In observational studies the main regression analysis can readily be adjusted for other variables including those which may modify the effect of the variable of interest. ```r subgroupAnalysis( object, data, treatment, subgroups, confint.method = "default", factor.reference = "extraline", ... ) ``` ## Arguments - `object`: - glm, coxph or cph object for which subgroups should be analyzed. - `data`: - Dataset including all relevant variables - `treatment`: - Must be numeric - 0/1 - `subgroups`: - A vector of variable names presenting the factor variables where subgroups should be formed. These variables should all be "factors" - `confint.method`: "default" creates Wald type confidence interval, "robust", creates creates robust standard errors - see regressionTable function. - `factor.reference`: "extraline" creates an extraline for the reference, "inline" avoids this line. - `...`: additional arguments such as case weights, which are passed on to `glm` and `coxph`. ## Returns A data.frame with subsgroup specifications, number in each subgroup, parameter estimates and p-value for interaction. A forest plot can be obtained with "plotConfidence". ## Details The function can only handle a bivariate treatment, which MUST coded as zero or one. The p-value for interaction is obtained with a likelihood ratio test comparing the main regression analysis with the interaction model. There are plot and print functions available for the function see helppages for plot.subgroupAnalysis and print.subgroupAnalysis ## Examples ```r #load libraries library(data.table) library(Publish) library(survival) data(traceR) #get dataframe traceR data.table::setDT(traceR) traceR[,':='(wmi2=factor(wallMotionIndex<0.9,levels=c(TRUE,FALSE), labels=c("bad","good")), abd2=factor(abdominalCircumference<95, levels=c(TRUE,FALSE), labels=c("slim","fat")))] traceR[,sex:=as.factor(sex)] # all subgroup variables needs to be factor traceR[observationTime==0,observationTime:=1] # remove missing covariate values traceR=na.omit(traceR) # univariate analysis of smoking in subgroups of age and sex # Main regression analysis is a simple/univariate Cox regression fit_cox <- coxph(Surv(observationTime,dead)~treatment,data=traceR) sub_cox <- subgroupAnalysis(fit_cox,traceR,treatment="treatment", subgroups=c("smoking","sex","wmi2","abd2")) sub_cox # to see how the results are obtained consider the variable: smoking fit_cox_smoke <- coxph(Surv(observationTime,dead)~treatment*smoking,data=traceR) # the last three rows of the following output: publish(fit_cox_smoke) # are included in the first 3 rows of the result of the sub group analysis: sub_cox[1:3,] # the p-value is obtained as: fit_cox_smoke_add <- coxph(Surv(observationTime,dead)~treatment+smoking,data=traceR) anova(fit_cox_smoke_add,fit_cox_smoke,test="Chisq") # Note that a real subgroup analysis would be to subset the data fit_cox1a <- coxph(Surv(observationTime,dead)~treatment,data=traceR[smoking=="never"]) fit_cox1b <- coxph(Surv(observationTime,dead)~treatment,data=traceR[smoking=="current"]) fit_cox1c <- coxph(Surv(observationTime,dead)~treatment,data=traceR[smoking=="prior"]) ## when the main analysis is already adjusted fit_cox_adj <- coxph(Surv(observationTime,dead)~treatment+smoking+sex+wmi2+abd2, data=traceR) sub_cox_adj <- subgroupAnalysis(fit_cox_adj,traceR,treatment="treatment", subgroups=c("smoking","sex","wmi2","abd2")) # subgroups as character string sub_cox_adj # When both start and end are in the Surv statement: traceR[,null:=0] fit_cox2 <- coxph(Surv(null,observationTime,dead)~treatment+smoking+sex+wmi2+abd2,data=traceR) summary(regressionTable(fit_cox)) sub_cox2 <- subgroupAnalysis(fit_cox2,traceR,treatment="treatment", subgroups=c("smoking","sex","wmi2","abd2")) # Analysis with Poisson - and the unrealistic assumption of constant hazard # and adjusted for age in all subgroups fit_p <- glm(dead~treatment+age+offset(log(observationTime)),family="poisson", data=traceR) sub_pois <- subgroupAnalysis(fit_p,traceR,treatment="treatment", subgroups=~smoking+sex+wmi2+abd2) # Analysis with logistic regression - and very wrongly ignoring censoring fit_log <- glm(dead~treatment+age,family="binomial",data=traceR) sub_log <- subgroupAnalysis(fit_log,traceR,treatment="treatment", subgroups=~smoking+sex+wmi2+abd2, factor.reference="inline") ``` ## See Also coxph, glm, plotConfidence ## Author(s) Christian Torp-Pedersen

subgroupAnalysis function