mirror of
https://github.com/agdamsbo/daDoctoR.git
synced 2024-11-25 05:01:54 +01:00
74 lines
2.0 KiB
R
74 lines
2.0 KiB
R
#' A repeated regression function for change-in-estimate analysis
|
|
#'
|
|
#' For bivariate analyses, binary logistic or linear regression. From "Modeling and variable selection in epidemiologic analysis." - S. Greenland, 1989.
|
|
#' @param meas Effect meassure. Input as c() of columnnames, use dput().
|
|
#' @param vars variables in model. Input as c() of columnnames, use dput().
|
|
#' @param string variables to test. Input as c() of columnnames, use dput().
|
|
#' @param data data frame to pull variables from.
|
|
#' @param logistic flag to set logistic (TRUE) or linear (FALSE,standard) analysis.
|
|
#' @param cut cut value for gating if including or dropping the tested variable. As suggested bu S. Greenland (1989).
|
|
#' @keywords estimate-in-estimate
|
|
#' @export
|
|
#' @examples
|
|
#' rep_reg_cie()
|
|
|
|
rep_reg_cie<-function(meas,vars,string,data,cut=0.1){
|
|
|
|
require(broom)
|
|
|
|
d<-data
|
|
x<-data.frame(d[,c(string)])
|
|
v<-data.frame(d[,c(vars)])
|
|
names(v)<-c(vars)
|
|
y<-d[,c(meas)]
|
|
dt<-cbind(y,v)
|
|
|
|
c<-as.numeric(cut)
|
|
|
|
if(!is.factor(y)){
|
|
|
|
meth<-"linear regression"
|
|
|
|
e<-as.numeric(round(coef(lm(y~.,data = dt)),3))[1]
|
|
df<-data.frame(pred="base",b=e)
|
|
|
|
for(i in 1:ncol(x)){
|
|
dat<-cbind(dt,x[,i])
|
|
m<-lm(y~.,data=dat)
|
|
|
|
b<-as.numeric(round(coef(m),3))[1]
|
|
|
|
pred<-paste(names(x)[i])
|
|
|
|
df<-rbind(df,cbind(pred,b)) }
|
|
|
|
di<-as.vector(round(abs(e-as.numeric(df[-1,2]))/e,3))
|
|
dif<-c(NA,di)
|
|
t<-c(NA,ifelse(di>=c,"include","drop"))
|
|
r<-cbind(df,dif,t) }
|
|
|
|
if(is.factor(y)){
|
|
meth="logistic regression"
|
|
|
|
e<-as.numeric(round(exp(coef(glm(y~.,family=binomial(),data=dt))),3))[1]
|
|
|
|
df<-data.frame(pred="base",b=e)
|
|
|
|
for(i in 1:ncol(x)){
|
|
dat<-cbind(dt,x[,i])
|
|
m<-glm(y~.,family=binomial(),data=dat)
|
|
|
|
b<-as.numeric(round(exp(coef(m)),3))[1]
|
|
|
|
pred<-paste(names(x)[i])
|
|
|
|
df<-rbind(df,cbind(pred,b)) }
|
|
|
|
di<-as.vector(round(abs(e-as.numeric(df[-1,2]))/e,3))
|
|
dif<-c(NA,di)
|
|
t<-c(NA,ifelse(di>=c,"include","drop"))
|
|
r<-cbind(df,dif,t)
|
|
}
|
|
return(list("method"=meth,"analyses"=r))
|
|
}
|