```{r setup, include=FALSE} knitr::opts_chunk$set(echo = TRUE, warning = FALSE, message = FALSE, error = FALSE) ``` # Prerequisite ```{r} library(questionr) library(tidyverse) # load data unions <- read_csv(file = "data/union_sweden.csv") ``` We will be using data from the European Social Survey, specifically from wave 5 in 2010. The dataset, unions_sweden.csv, includes information on socioeconomic characteristics for over 1,289 respondents from Sweden. Our focus is on labor unions, and we will be utilizing the following variables: -------------------------------- ---------------------------------------------------------- Name Description -------------------------------- ---------------------------------------------------------- `Country` Name of the respondent's country of residence. `union` Whether the respondent is an active member of a labor union (`0`= No, `1`= Yes) `gndr` Respondent's gender (`0` = male, `1`= female) `income` Respondent's disposable income. `sector` Sector in which the respondent is employed. It follows the Industry, NACE classification. `lrscale` Placement on left right scale (where `0` = left, `1`= right) `gincdif` Respondent's answer to question "Government should reduce differences in income levels" (where `5` = Agree strongly, `4` = Agree, `3` = Neither agree nor disagree, `2` = Disagree, `1` = Disagree strongly) ------------------------------------------------------------------------------------------- # Union's treatment effects Estimate the differences between the control and treatment groups based on the "union" variable for the outcomes `income` and `gincdif.` Report these estimates. Do these estimates **identify** the causal effects of the "union" variable? Check the balance between the control and treatment groups. ```{r} ## check balance between treatment and control table(unions$union_f) ## estimate union effect on income mean(unions$income[unions$union_f=="treatment"]) - mean(unions$income[unions$union_f=="control"]) ## estimate union effect on gincdif. mean(unions$gincdif[unions$union_f=="treatment"]) - mean(unions$gincdif[unions$union_f=="control"]) ``` # Confounding Since this sample has been adjusted to achieve a nearly equal number of non-union (**control**) and union membership (**treatment**) respondents, we would anticipate an approximate 50% proportion of union members across various observable factors if, and only if, the **assumption of covariate balance** holds. This balance serves as a tentative assumption of the causal effects of union membership, conditioned on other factors. To assess this assumption, let's examine the proportion of union members at each level of the left-right ideological scale (`lrscale`). Does the data corroborate this assumption? ```{r} ## Look at the proportion of union memberhsip with tapply tapply(unions$union, unions$lrscale, mean) ## new function: aggregate aggregate(union ~ lrscale, data = unions, FUN= mean) ``` Additionally, consider identifying another variable that may confound union membership. ```{r} ## check balance of union membersip based on gender mean(unions$gndr[unions$union_f=="treatment"]) - mean(unions$gndr[unions$union_f=="control"]) ## check proportion of union membership conditional on sector aggregate(union ~ sector, data = unions, FUN= mean) ``` # Adjusting confounding Use the `lm()` function to regress the outcomes `income` and `gincdif` on the treatment variable union using a linear model. Save the output in an object and use the `summary()` function to examine the estimated coefficients. Compare these coefficients with the previously calculated differences for the union variable. ```{r, results='asis'} m1 <- lm(income ~ union, data = unions) m2 <- lm(gincdif ~ union, data = unions) library(stargazer) stargazer(m1,m2, type="latex") # change to type="text" to display the results in the console ``` Repeat the same exercise, but this time include the variables `lrscale` and `sector`. How much have change your estimations after **adjusting for confounding**? Can we say that we have identified a causal estimate? ```{r, results='asis'} m3 <- lm(income ~ sector + lrscale + union, data = unions) m4 <- lm(gincdif ~ sector + lrscale + union, data = unions) stargazer(m3,m4, omit = "sector", # for space concern, omit reporting sector coefficients. type="latex") # change to type="text" to display the results in the console ``` Are there differences in redistributive attitudes between male and female union members? Compare the `gincdif` variable between the two groups. ```{r} # group_by with dplyr unions %>% group_by(union_f,gndr_f) %>% summarize(gincdif=mean(gincdif,na.rm=T)) %>% pivot_wider(names_from = c("union_f","gndr_f"), values_from = gincdif) %>% mutate(gndr_diff=treatment_Female-treatment_Male) # with base R women_union <- mean(unions$gincdif[unions$union_gndr=="trt_female"]) men_union <- mean(unions$gincdif[unions$union_gndr=="trt_male"]) women_union - men_union ``` Also, examine union-gender differences by sector and save your visualizations to your local folder. Are these differences considered **causal effects**? ```{r} unions %>% # group by categories group_by(sector,union_f,gndr_f) %>% # estimate the summaries of interest summarize(ginunions=mean(gincdif,na.rm=T)) %>% # pivot from long to wide pivot_wider(id_cols = "sector", names_from = c("union_f","gndr_f"), # note the two groups values_from = "ginunions") %>% mutate(diff=treatment_Female-treatment_Male, diff = round(diff,2)) # save the data in an object for visualization vis <- unions %>% group_by(sector,union_f,gndr_f) %>% summarize(ginunions=mean(gincdif,na.rm=T)) %>% pivot_wider(id_cols = "sector", names_from = c("union_f","gndr_f"), values_from = "ginunions") %>% mutate(diff=treatment_Female-treatment_Male, diff = round(diff,2)) vis %>% ggplot(aes(x=diff, y=(sector=reorder(sector, diff)))) + geom_point() + geom_vline(xintercept = 0,linetype="dashed") + theme_minimal() + labs(y="sector", x="Gender-union treatment differences", title="Unionized gender effects on solidarity by sector") # width = 6 # ggsave("output/gndr_diff.punions", width = width, height = width/1.618) ``` 