-
-
Save krz/020ff09fc7c8e8eec48a4102cee89810 to your computer and use it in GitHub Desktop.
code for the paper "Tweedie distributions for fitting semicontinuous health care utilization cost data"
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| library(sampleSelection) # for RandHIE data | |
| library(dplyr) | |
| data(RandHIE) | |
| # combined costs | |
| RandHIE$cost <- RandHIE$outpdol + | |
| RandHIE$drugdol + | |
| RandHIE$suppdol + | |
| RandHIE$mentdol + | |
| RandHIE$inpdol | |
| # select relevant variables | |
| rh <- RandHIE %>% group_by(zper) %>% summarise(cost = first(cost), | |
| age=first(xage), | |
| disea=first(disea), | |
| physlm=first(physlm), | |
| logc=first(logc), #lc | |
| idp=first(idp),# | |
| lpi=first(lpi),# | |
| fmde=first(fmde), | |
| linc=first(linc), | |
| lfam=first(lfam), | |
| female=first(female), | |
| black=first(black), | |
| educdec=first(educdec), | |
| hlthg=first(hlthg) | |
| ) | |
| # select only people >= 18 and remove missing values | |
| rh <- rh %>% filter(age >= 18 & !is.na(educdec)) | |
| rh$zper <- NULL | |
| # remove implausible values | |
| rh <- rh %>% filter(physlm == 0 | physlm == 1 | black == 0 | black == 1) | |
| length(which(rh$cost ==0))/nrow(rh) # 18.1 % zeros | |
| ############## tobit | |
| library(VGAM) | |
| tob <- vglm(cost ~ ., tobit(Lower = 0, type.fitted = "censored"), data = rh, maxit=100) | |
| summary(tob) | |
| AICc(tob) | |
| ############## tweedie | |
| library(cplm) | |
| tw2 <- cpglm(cost ~ ., data=rh) | |
| summary(tw2) | |
| (AIC(tw2) - 2*length(coef(tw2)) ) / -2 | |
| ############## gamma twopart | |
| rh$zero <- ifelse(rh$cost == 0, 0, 1) # 0 if zero costs | |
| h1 <- glm(zero ~ . -cost, family=binomial(link=logit), data=rh) | |
| summary(h1) | |
| logLik(h1) | |
| h2 <- glm(cost ~ . -zero, family = Gamma(link = log), data=subset(rh, zero == 1)) | |
| summary(h2) | |
| logLik(h2) | |
| AIC(h1) + AIC(h2) | |
| ####################### | |
| # rmse estimation | |
| # split into train and test | |
| set.seed(41) | |
| idx <- sample(1:nrow(rh), 500, replace=F) | |
| train.x <- rh[-idx, ] | |
| test.x <- rh[idx,] | |
| ### tobit | |
| tob.rmse <- vglm(cost ~ . -zero, tobit(Lower = 0, type.fitted = c("censored")), data = train.x, maxit=100) | |
| summary(tob.rmse) | |
| preds <- predict(tob.rmse, test.x, type="response") | |
| #rmse | |
| sqrt(mean((preds - test.x$cost)^2)) | |
| ### gamma twopart | |
| h1.rmse <- glm(zero ~ . -cost, family=binomial(link=logit), data=train.x) | |
| pred2 <- predict(h1.rmse, test.x, type="response") | |
| #pred2 <- ifelse(pred2 < 0.5, 0, 1) | |
| h2.rmse <- glm(cost ~ . -zero, family = Gamma(link = log), data=subset(train.x, zero == 1)) | |
| # new | |
| preds2 <- pred2 * predict(h2.rmse, test.x, type="response") | |
| #rmse | |
| sqrt(mean((preds2 - test.x$cost)^2)) | |
| ### tweedie | |
| train.x$zero <- NULL | |
| tw.rmse <- cpglm(cost ~., data=train.x) | |
| preds3 <- predict(tw.rmse, test.x, type="response") | |
| # rmse | |
| sqrt(mean((preds3 - test.x$cost)^2)) | |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment