data-raw/03_overarching_propensity_score.md

In bennysalo/placement-and-recidivism:

Calculate 'overarching propensity' for full sample

Benny Salo 2018-06-26

Here we calculate an overarching propensity score for the full sample. This is the propensity to be placed in open prison and to ,in addition, be granted and successfully conclude a conditional release.

This overarching propensity score will be used to make groups comparable in two ways. 1. We will create a jitter plot to compare the propensity in four groups - open prison with conditional release - open prison without conditional release - closed prison with conditional release - closed prison without conditional release

1. We will do a median split on propensity score to examine if effects are different for high and low propensity individuals. For this we want comparable propensity scores for each analysis.

For balancing we will instead use a custom propensity score for the specific sample to optimize balance. This will in effect be three different propensity scores: - The propensity to be placed in open prison. - The propensity to be granted conditional release. - The propensity to be placed in open prison and be granted conditional release.

That is, different propensity scores will be calculated for different groups for the purpose of balancing.

Load and setup

devtools::wd()
analyzed_data_plac <- readRDS("not_public/analyzed_data_plac.rds")

devtools::load_all(".")
library(tidyverse)

Analyses

Run logistic regression model. Predict placement in open prison plus successful conditional release from supervision of parole and all potential confounders.

overarch_propensity_fit <-
  glm(
    formula = write_formula(lhs = "open_and_cr01", 
                            rhs = c("supervisedParole", potential_confounders)),
    data    = analyzed_data_plac,
    family  = "binomial")

Calculate propensity scores as log odds and save in data frame.

analyzed_data_plac$overarch_propensity <- predict(overarch_propensity_fit)

Save the data including the new variable.

devtools::wd()
saveRDS(analyzed_data_plac, "not_public/analyzed_data_plac.rds")

Descriptive jitterplot

levels(analyzed_data_plac$openPrison) <- c("Closed", "Open")
levels(analyzed_data_plac$conditionalReleaseOutcome) <- c("No CR", "CR granted")
levels(analyzed_data_plac$supervisedParole) <- c("No supervision", "Parole supervised")

ggplot(analyzed_data_plac, aes(x = overarch_propensity,
                               y = supervisedParole)) + 
  geom_jitter() + 
  ylab("") +
  xlab("Propensity score (in log odds)") +
  facet_grid(openPrison + conditionalReleaseOutcome ~ .) +
  ggthemes::theme_tufte(base_family = "sans")

Median propensity

median_overaching_propensity <- median(analyzed_data_plac$overarch_propensity)

bennysalo/placement-and-recidivism documentation built on May 21, 2019, 9:47 a.m.