具有第一差异的系数图

Question

I am running a logistic model in r. I am trying to present the differences in probabilities for my independent variables with a coefficient plot. Specifically, I would like to create the differences in probabilities by shifting the variables of interest from their minimum to their maximum value (while holding the other variables at their means or modes).

In the picture attached, I would like for my graph to look similar to the top half.

I have ran this code:

 mydata <- read.csv("http://www.ats.ucla.edu/stat/data/binary.csv")
 mylogit <- glm(admit ~ gre + gpa + rank, data = mydata, family =
 "binomial")

Afterwards I calculated the predicted probability for each variable for the minimum value and the maximum value and subtracted the two. I repeated this process for the upper and lower bounds of the interval. Attached is my code

plotdat <- data.frame(gre=c(.220, 800), gpa=mean(mydata$gpa, na.rm=TRUE), rank=c(2) ) 
preddat <- predict(mylogit, newdata=plotdat, se.fit=TRUE)

Grebeta<-(exp(preddat$fit[2])/(1+exp(preddat$fit[2])))-(exp(preddat$fit[1])/(1+exp(preddat$fit[1])))
Gremin<-(exp(preddat$fit[2]+1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]+1.96*preddat$se.fit[2])))-exp(preddat$fit[1]+1.96*preddat$se.fit[1])/(1+exp(preddat$fit[1]+1.96*preddat$se.fit[1]))
Gremax<-exp(preddat$fit[2]-1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]-1.96*preddat$se.fit[2]))-exp(preddat$fit[1]-1.96*preddat$se.fit[1])/(1+exp(preddat$fit[1]-1.96*preddat$se.fit[1]))


plotdat <- data.frame(gpa=c(2.26, 4), gre=mean(mydata$gre, na.rm=TRUE), rank=c(2) )
preddat <- predict(mylogit, newdata=plotdat, se.fit=TRUE)

GPAbeta<-(exp(preddat$fit[2])/(1+exp(preddat$fit[2])))-(exp(preddat$fit[1])/(1+exp(preddat$fit[1])))
GPAmin<-(exp(preddat$fit[2]+1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]+1.96*preddat$se.fit[2])))-exp(preddat$fit[1]+1.96*preddat$se.fit[1])/(1+exp(preddat$fit[1]+1.96*preddat$se.fit[1]))
GPAmax<-exp(preddat$fit[2]-1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]-1.96*preddat$se.fit[2]))-exp(preddat$fit[1]-1.96*preddat$se.fit[1])/(1+exp(preddat$fit[1]-1.96*preddat$se.fit[1]))


plotdat <- data.frame(rank=c(4, 1), gre=mean(mydata$gre, na.rm=TRUE), gpa=mean(mydata$gpa, na.rm=TRUE ))
preddat <- predict(mylogit, newdata=plotdat, se.fit=TRUE)

Rankbeta<-(exp(preddat$fit[2])/(1+exp(preddat$fit[2])))-(exp(preddat$fit[1])/(1+exp(preddat$fit[1])))
Rankmin<-(exp(preddat$fit[2]+1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]+1.96*preddat$se.fit[2])))-exp(preddat$fit[1]+1.96*preddat$se.fit[1])/(1+exp(preddat$fit[1]+1.96*preddat$se.fit[1]))
Rankmax<-exp(preddat$fit[2]-1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]-1.96*preddat$se.fit[2]))-exp(preddat$fit[1]-1.96*preddat$se.fit[1])/(1+exp(preddat$fit[1]-1.96*preddat$se.fit[1]))

Afterwards, I created three vectors containing the difference in probabilities and bands. Attached is my code:

se.max<- c(Gremax   , GPAmax  , Rankmax  )
coef.vec<- c( Grebeta  ,GPAbeta  , Rankbeta ) 
se.min<-c(Gremin , GPAmin, Rankmin)



var.names <- c("gre", "gpa", "rank")

Finally, I plotted my graph.

y.axis <- c(length(coef.vec):1)

par(mar=c(2, 13, 0, 0))

plot(coef.vec, y.axis, type = "p", axes = F, xlab = "", ylab = "", pch = 19, cex = 1.2,  xlim = c(-2,2.5), xaxs = "r", main = "")

segments(se.max, y.axis,se.min, y.axis, lwd =  1.5)

axis(1, at = seq(-1,1,by=.25), labels = NA, tick = T,cex.axis = 1.2, mgp = c(2,.7,0))
axis(1, at = seq(-1,1,by=.5), labels =  c(-1,  -.5,  0, .5,1), tick = T,cex.axis = 1.2, mgp = c(2,.7,0))

axis(2, at = y.axis, label = var.names, las = 1, tick = T, ,mgp = c(2,.6,0), cex.axis = 1.2)
segments(0,0,0,17,lty=2)

However, I can't get my confidence intervals to plot. Attached below is my final output.

It appears my confidence bands won't plot. If anyone could provide assistance and point to errors in my calculations or code, I would greatly appreciate it.

Answer 1

plotdat <- data.frame(gre=c(.220, 800), gpa=mean(mydata$gpa, na.rm=TRUE), rank=c(2) ) 
preddat <- predict(mylogit, newdata=plotdat, se.fit=TRUE)

#GRE High
GREbetahigh<-(exp(preddat$fit[2])/(1+exp(preddat$fit[2])))
GREminhigh<-(exp(preddat$fit[2]+1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]+1.96*preddat$se.fit[2])))
GREmaxhigh<-exp(preddat$fit[2]-1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]-1.96*preddat$se.fit[2]))


#GRE low
GREbetalow<-(exp(preddat$fit[1])/(1+exp(preddat$fit[1])))
GREminlow<-(exp(preddat$fit[1]+1.96*preddat$se.fit[2])/(1+exp(preddat$fit[1]+1.96*preddat$se.fit[1])))
GREmaxlow<-exp(preddat$fit[1]-1.96*preddat$se.fit[2])/(1+exp(preddat$fit[1]-1.96*preddat$se.fit[1]))

#GRE Diff
GREbeta.diff<-GREbetahigh-GREbetalow
GREmax.diff<-GREmaxhigh-GREmaxlow
GREmin.diff<-GREminhigh-GREminlow

#GPA
plotdat <- data.frame(gpa=c(2.26, 4), gre=mean(mydata$gre, na.rm=TRUE), rank=c(2) )
preddat <- predict(mylogit, newdata=plotdat, se.fit=TRUE)

#GPA high
GPAbetahigh<-(exp(preddat$fit[2])/(1+exp(preddat$fit[2])))
GPAminhigh<-(exp(preddat$fit[2]+1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]+1.96*preddat$se.fit[2])))
GPAmaxhigh<-exp(preddat$fit[2]-1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]-1.96*preddat$se.fit[2]))

#GPA low
GPAbetalow<-(exp(preddat$fit[1])/(1+exp(preddat$fit[1])))
GPAminlow<-(exp(preddat$fit[1]+1.96*preddat$se.fit[2])/(1+exp(preddat$fit[1]+1.96*preddat$se.fit[1])))
GPAmaxlow<-exp(preddat$fit[1]-1.96*preddat$se.fit[2])/(1+exp(preddat$fit[1]-1.96*preddat$se.fit[1]))

#GPA Diff
GPAbeta.diff<-GPAbetahigh-GPAbetalow
GPAmax.diff<-GPAmaxhigh-GPAmaxlow
GPAmin.diff<-GPAminhigh-GPAminlow

#Rank

plotdat <- data.frame(rank=c(4, 1), gre=mean(mydata$gre, na.rm=TRUE), gpa=mean(mydata$gpa, na.rm=TRUE ))
preddat <- predict(mylogit, newdata=plotdat, se.fit=TRUE)

#Rank high
Rankbetahigh<-(exp(preddat$fit[2])/(1+exp(preddat$fit[2])))
Rankminhigh<-(exp(preddat$fit[2]+1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]+1.96*preddat$se.fit[2])))
Rankmaxhigh<-exp(preddat$fit[2]-1.96*preddat$se.fit[2])/(1+exp(preddat$fit[2]-1.96*preddat$se.fit[2]))

#Rank Low
Rankbetalow<-(exp(preddat$fit[1])/(1+exp(preddat$fit[1])))
Rankminlow<-(exp(preddat$fit[1]+1.96*preddat$se.fit[1])/(1+exp(preddat$fit[1]+1.96*preddat$se.fit[1])))
Rankmaxlow<-exp(preddat$fit[1]-1.96*preddat$se.fit[1])/(1+exp(preddat$fit[1]-1.96*preddat$se.fit[1]))


#Rank Diff
Rankbeta.diff<-Rankbetahigh-Rankbetalow
Rankmax.diff<-Rankmaxhigh-Rankmaxlow
Rankmin.diff<-Rankminhigh-Rankminlow

#Graph
se.max<- c(GREmax.diff   , GPAmax.diff, Rankmax.diff)
coef.vec<- c( GREbeta.diff , GPAbeta.diff, Rankbeta.diff)
se.min<-c(GREmin.diff , GPAmin.diff, Rankmin.diff)

var.names <- c("gre", "gpa", "rank")

y.axis <- c(length(coef.vec):1)

par(mar=c(2, 13, 0, 0))


plot(y.axis, coef.vec, type = "p", axes = F, xlab = "", ylab = "", pch = 19, cex = 1.2,  ylim = c(-1,1), xlim=c(1,3.3), xaxs = "r", main = "")
segments(y.axis, se.max,y.axis, se.min, lwd =  1.5)

axis(2, at = seq(-1,1,by=.25), labels = NA, tick = T,cex.axis = 1.2, mgp = c(2,.7,0))
axis(2, at = seq(-1,1,by=.5), labels =  c(-1,  -.5,  0, .5, 1), tick = T,cex.axis = 1.2, mgp = c(2,.7,0))

axis(1, at = y.axis, label = var.names, las = 1, tick = T, ,mgp = c(2,.6,0), cex.axis = 1.2)
segments(1,0,3.3,0,lty=2)