###############################################
# HOW TO USE THIS FILE
# This is an HTML version of the R script file ex4.R
# It is intended to show you the code and to allow links.
# NOT to use interactively to run the code
# For that download the script file ex4.R
# to view the ouptut from these commands in html go to R results page
red stuff (R commands)
#

#
Links in this page

Equal opportunities policies weighted and unweighted - Table 4.3 E O policy by workplace size Table 4.4 Other factors by equal opportunities policy Table 4.5 Logistic regression Table 4.6 Logistic regression Table 4.7 Finite population corrections Table 4.8 # to get names of variables names(wers) # # the ethnic group percentages are in a column labelled ETHNIC # to get a histogram of this variable do # hist(wers$ETHNIC) # # now it is time to try using the survey package # to give R access to the survey functions do - see mini guide if this fails # library(survey) Back to topuparrow ## ################ table 4.3 ############################### # # # The first step is to create a SURVEY DESIGN OBJECT that # will contain both the data and the information about the design # # This design has WEIGHTS (EST_WT),but to get the correct # design effects without using the option deff="replace" # we need to use the weights that add to the population (GROSSWT) # and it is stratified by size of workplace # and sector. There are 71 strata indexed by STRATA # Since it is not clustered we set the sampling unit ids # to be the serial number SERNO, different for each workplace wers.des <- svydesign(id=~SERNO, strata=~STRATA,weights=~GROSSWT,data=wers) # # # now we can look at the survey weighted mean of the proportions # of workplaces with an EO policy using the svymean command # giving the design as the second argument to the function # svymean(~EO,wers.des) # # sadly this just gives a missing value since EO has 7 missing values # but the answer is to remove the missing data from the survey design # see the help for is.na and for subset.surveydesign # svymean(~EO,subset(wers.des,!is.na(EO)),deff=T) # # we can compare this with the unweighted mean had the design been different # wers.des <- svydesign(id=~SERNO, strata=~STRATA,data=wers) svymean(~EO,subset(wers.des,!is.na(EO)),deff=T) # Back to topuparrow ###################Table 4.4################################## # # to get the weighted means for subgroups by employment size # we can use the svyby command to apply svymean to subgroups # first resetting the design # wers.des <- svydesign(id=~SERNO, strata=~STRATA,weight=GROSSWT,data=wers) svyby(~EO,~NEMPSIZE,subset(wers.des,!is.na(EO)),svymean,keep.var=TRUE) # # the warning messages refer to lonely PSUs (see theory section on checking data) # This means that the standard errors masy be slightly underestimated # Back to topuparrow #######################Table 4.5######################################## # # Now we get means of percent female by EO # svyby(~FEMALE,~EO,subset(wers.des,!is.na(FEMALE)),svymean) # # again the easisest way to get the standard error of # the difference is with a with svyglm (see help) # summary(svyglm(FEMALE~EO,wers.des) ) # # obviously workplaces with an EO policy have a higher % women # we can now do similar analyses for % ethnic minority employees # svyby(~ETHNIC,~EO,subset(wers.des,!is.na(ETHNIC)),svymean) summary(svyglm(ETHNIC~EO,wers.des)) # # like females a bigger effect in the weighted analysis # but clear evidence of more ethnic minority employees when policy in place # # now find weighted percent in each categories of disabled worker %s # since DISABGRP is a factor R gives proportions # svymean(~DISABGRP,subset(wers.des,!is.na(DISABGRP))) # # now means of EO by disabled groups # svyby(~EO,~DISABGRP,subset(wers.des,!is.na(DISABGRP) & !is.na(EO)),svymean) # # svyglm compares percent in each disabled group category # between eo and non-eo workplaces # summary(svyglm((DISABGRP=='none')~EO,subset(wers.des,!is.na(wers$DISABGRP) & !is.na(EO)))) summary(svyglm((DISABGRP=='under 3%')~EO,wers.des[!is.na(wers$DISABGRP),])) summary(svyglm((DISABGRP=='3%+')~EO,wers.des[!is.na(wers$DISABGRP),])) ############################################################## # now the same things for unweighted data wers.des <- svydesign(id=~SERNO, strata=~STRATA,data=wers) # svyby(~FEMALE,~EO,subset(wers.des,!is.na(FEMALE)),svymean) summary(svyglm(FEMALE~EO,wers.des) ) svyby(~ETHNIC,~EO,subset(wers.des,!is.na(ETHNIC)),svymean) summary(svyglm(ETHNIC~EO,wers.des)) svymean(~DISABGRP,subset(wers.des,!is.na(DISABGRP))) svyby(~EO,~DISABGRP,subset(wers.des,!is.na(DISABGRP) & !is.na(EO)),svymean) summary(svyglm((DISABGRP=='none')~EO,subset(wers.des,!is.na(wers$DISABGRP) & !is.na(EO)))) summary(svyglm((DISABGRP=='under 3%')~EO,wers.des[!is.na(wers$DISABGRP),])) summary(svyglm((DISABGRP=='3%+')~EO,wers.des[!is.na(wers$DISABGRP),])) Back to topuparrow ########################################################## # Table 4.6 # now the svyglm command and the glm commands to # fit logistic regressions # first restore design wers.des <- svydesign(id=~SERNO, strata=~STRATA,weights=~GROSSWT,data=wers) summary(svyglm(EO~ETHNIC+FEMALE+DISABGRP,wers.des,family='binomial')) # # and unweighted # wers.des <- svydesign(id=~SERNO, strata=~STRATA,data=wers) summary(svyglm(EO~ETHNIC+FEMALE+DISABGRP,wers.des,family='binomial')) Back to topuparrow ################################################################## # table 4.7 # now explore models that include workplace size wers.des <- svydesign(id=~SERNO, strata=~STRATA,weights=~GROSSWT,data=wers) summary(svyglm(EO~ETHNIC+FEMALE+DISABGRP+NEMPSIZE,wers.des,family='binomial')) # # and without weighting # wers.des <- svydesign(id=~SERNO, strata=~STRATA,data=wers) summary(svyglm(EO~ETHNIC+FEMALE+DISABGRP+NEMPSIZE,wers.des,family='binomial')) Back to topuparrow ############################################################### # # table 4.8 with finite population corrections # wers.des <- svydesign(id=~SERNO,fpc=~SAMPFRAC, strata=~STRATA,weight=~GROSSWT,data=wers) svymean(~EO,subset(wers.des,!is.na(EO))) svyby(~EO,~NEMPSIZE,subset(wers.des,!is.na(EO)),svymean,keep.var=TRUE) # # and without wers.des <- svydesign(id=~SERNO, strata=~STRATA,weight=~GROSSWT,data=wers) svymean(~EO,subset(wers.des,!is.na(EO))) svyby(~EO,~NEMPSIZE,subset(wers.des,!is.na(EO)),svymean,keep.var=TRUE) # ############################################################# # now some more modelling stuff to make plots # wers.des <- svydesign(id=~SERNO, strata=~STRATA,weights=~GROSSWT,data=wers) # disabled now dropped out # summary(svyglm(EO~ETHNIC+FEMALE+NEMPSIZE+ETHNIC,wers.des,family='binomial')) # now make a factor to test linearity of percent female # wers.des<-update( wers.des,GPFEMALE=factor(ceiling(FEMALE/20)*(FEMALE>0)+(1*FEMALE==0), labels=c('0-<20','20-<40','40-<60','60-<80','80-100') ) ) summary(svyglm(EO~ETHNIC+GPFEMALE+NEMPSIZE,wers.des,family='binomial')) # # Linearity for grouped size looks not too bad # But there seems to be steep increase for workplaces with >50% # female. For robustness, stick with this model where we fit both of these as categories # # fit as simple (not survey) glm # # # fit as simple (not survey) glm # summary(glm(EO~ETHNIC+GPFEMALE+NEMPSIZE,wers,family='binomial')) # # we can see that the effect of % of female employees is less marked in the bigger workplaces # note also that the efect of ethnic group still stands out clearly in these # analyses with the odds of EO increasing by around 1.8 (exp(0.3*20)) when the # percentage ethnic goes from 0 to 20%. # # This final bit of code makes a plot of %EO by FEMALE and SIZE to understand findings. # wers$GPFEMALE=factor(ceiling(wers$FEMALE/20)*(wers$FEMALE>0)+(1*wers$FEMALE==0), labels=c('0-<20','20-<40','40-<60','60-<80','80-100') ) forpred<-data.frame(NEMPSIZE=rep(0:5,5),ETHNIC=rep(2,30), FEMALE=rep(c(10,30,50,70,90),rep(6,5))) forpred$GPFEMALE=factor(ceiling(forpred$FEMALE/20)*(forpred$FEMALE>0)+(1*forpred$FEMALE==0), labels=c('0-<20','20-<40','40-<60','60-<80','80-100') ) forpred$NEMPSIZE<-factor(forpred$NEMPSIZE,labels=c('10-24','25-49','50-99','100-199', '200-499','500+')) wt.fit<-predict(svyglm(EO~ETHNIC+GPFEMALE+NEMPSIZE,wers.des,family='binomial'),forpred) unwt.fit<-predict(glm(EO~ETHNIC+GPFEMALE+NEMPSIZE,wers,family='binomial'),forpred) invlogit<-function(x){1/(1+exp(-x))} par(mar=c(5,5,1,1),cex.axis=.9,cex.lab=.9,oma=c(0,0,0,0)) plot(c(0,5),c(40,100),type='n',ylab='% with eo policy',xlab='workplace size group') for (i in 1:5) points(0:5,100*invlogit(wt.fit)[1:6+(i-1)*6],pch=1,lty=1,type='b',col=i,lwd=2) for (i in 1:5) points(0:5,100*invlogit(unwt.fit)[1:6+(i-1)*6],pch=2,lty=2,type='b',col=i,lwd=2) legend(3.3,67,c('10% female' ,'30% female',' 50% female', '70% female',' 90% female'),col=1:5,pch=rep(1,5),lty=rep(1,5),bty='n',cex=.8) legend(4.5,67,c('' ,'',' ', '',''),col=1:5,pch=rep(2,5),lty=rep(2,5),bty='n',cex=.8) text(3.5,69,'weighted',cex=.8) text(4.7,69,'unweighted',cex=.8) #################################################################