I am defining a function to get the predicted values of a regression model with survey data for different subgroups(subpopulations). I use the svyglm function from the survey package.
My problem concerns handling the subset option in the svyglm function. As it uses non-standard evaluation, which I understand means it doesn't take the column names as a string. I tried just using the columns name without strings and enquoting (enquo) and unquoting it (!!). However both options do not work. I also played around with ensym() and expr() but did not get any results.
library(dplyr)
library(survey)
library(srvyr)
library(purrr)
library(rlang)
mtcars <- read.table("https://forge.scilab.org/index.php/p/rdataset/source/file/master/csv/datasets/mtcars.csv",
sep=",", header=TRUE)
mtcars_cplx <- mtcars %>% as_survey_design(id = cyl, weights = qsec)
carb <- c(1:8)
cyl <- c(4:8)
new_data <- expand.grid(carb, cyl)
colnames(new_data) <- c("carb", "cyl")
subpop_pred <- function(formula, data, subpop, new_data) {
subpop_quo <- enquo(subpop)
subpop_txt <- data$variables %>% select(!!subpop_quo) %>% colnames()
for(i in min(data$variables[subpop_txt]):max(data$variables[subpop_txt])){
reg <- svyglm(formula, data, subset=!!subpop_quo==i)
pred <- predict(reg, newdata=new_data)
if(exists("reg_end")==TRUE){
pred <- cbind(new_data, pred, confint(pred))
pred[subpop_txt] <- i
reg_end <- rbind(reg_end, pred)
} else {
reg_end <- cbind(new_data, pred, confint(pred))
reg_end[subpop_txt] <- i
}
}
}
subpop_pred(mpg ~ carb + cyl + carb*cyl,
data=mtcars_cplx,
new_data=new_data,
subpop=gear)
Error: Base operators are not defined for quosures.
Do you need to unquote the quosure?
# Bad:
myquosure == rhs
# Good:
!!myquosure == rhs
Call `rlang::last_error()` to see a backtrace
8. stop(cnd)
7. abort(paste_line("Base operators are not defined for quosures.",
"Do you need to unquote the quosure?", "", " # Bad:", bad,
"", " # Good:", good, ))
6. Ops.quosure(subpop_quo, i)
5. eval(subset, model.frame(design), parent.frame())
4. eval(subset, model.frame(design), parent.frame())
3. svyglm.survey.design(formula, data, subset = !!subpop_quo ==
i)
2. svyglm(formula, data, subset = !!subpop_quo == i)
1. subpop_pred(mpg ~ carb + cyl + carb * cyl, data = mtcars_cplx,
new_data = new_data, subpop = gear)
subpop_pred <- function(formula, data, subpop, new_data) {
subpop_quo <- enquo(subpop)
subpop_txt <- data$variables %>% select(!!subpop_quo) %>% colnames()
for(i in min(data$variables[subpop_txt]):max(data$variables[subpop_txt])){
reg <- svyglm(formula, data, subset=subpop==i)
pred <- predict(reg, newdata=new_data)
if(exists("reg_end")==TRUE){
pred <- cbind(new_data, pred, confint(pred))
pred[subpop_txt] <- i
reg_end <- rbind(reg_end, pred)
} else {
reg_end <- cbind(new_data, pred, confint(pred))
reg_end[subpop_txt] <- i
}
}
}
subpop_pred(mpg ~ carb + cyl + carb*cyl, data=mtcars_cplx, new_data=new_data, subpop=gear)
Error in eval(subset, model.frame(design), parent.frame()) :
object 'gear' not found
5. eval(subset, model.frame(design), parent.frame())
4. eval(subset, model.frame(design), parent.frame())
3. svyglm.survey.design(formula, data, subset = subpop == i)
2. svyglm(formula, data, subset = subpop == i)
1. subpop_pred(mpg ~ carb + cyl + carb * cyl, data = mtcars_cplx,
new_data = new_data, subpop = gear)
Do you have an idea how to get the function to work?
I could get things working with the subset argument by mixing expr() and rlang::tidy_eval().
The model line in your function could then read:
reg <- svyglm(formula, data = data,
subset = rlang::eval_tidy( expr( !!subpop_quo == i), data = data) )
I don't know robust this is, though, or if there is some more straightforward approach to tidyeval. Working on this made me realize that the subset() function/argument are difficult to work with in functions. :-P