Here are the three potential proxy measures that we tested in the survey:
vax_chance_post - vax_chance_pre
best_treatment with “nothing would help” should be assigned a vax_chance_post equal to their vax_chance_pre (they are essentially saying that nothing would help them change their mind)vax_future
vax_conf_pre_yes vax_conf_pre_no
Objective: We need to decide whether or not we have a good proxy measure for treatment effect for our pilot. If we’re not confidence in the proxy measure, we’ll likely have to get more of Susan/Dean’s feedback on a better design.
## Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
## -4.0000 0.0000 0.0000 0.2367 1.0000 4.0000 2935
## Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
## -4.0000 0.0000 0.0000 0.2153 1.0000 4.0000 2920
Objective: We want to begin testing and hypothesizing which covariates are most important in predicting treatment effects. Impediments particularly are a key covariate we should test (to see if there are differences in proxy measures of treatment effect, given that a large part of the survey is designed to identify their impediment).
What covariates are likely to vary with our proxy treatment effects? One ideas is to order our participants by proxy measure of treatment effects (e.g., vax_chance_post - vax_chance_pre - but we should test for all three), and then separate them into quartile groups. For each quartile group, we can see what the average value of a covariate is, and how much that varies from the average value of a covariate across the entire survey population. Covariates that vary in the top/bot quartile are likely candidates for predicting our proxy treatment effects. Ultimately, we’re trying to answer: which covariates are likely to not have any impact on treatment effects, and what covariates look promising?
We mapped binary and ordinal variables to continuous variables.
Details on the mapping:
female: 1 if female, 0 if maleblack: 1 if black or african, 0 otherwiseincome: 0 if the participant is unemployed, 1 if income < R5,000, 2 if income in R5,000 – R9,999, …, 6 if income > R100,000education: 1 if the participant’s education < high school, 2 if education is high school, …, 6 if education is a graduate degreechristian: 1 if the participant identifies as christian, 0 otherwisereligiosity: 1 if the participant is not very religious, 2 if somewhat religious, 3 if very religiouspolitics: 1 if the participant is conservative, 2 if moderate, 3 if liberallocation: 1 if the participant lives in rural, 2 if suburban, 3 if urban,ability: 1 if the participant has the ability to get vax, 0 if notmotivation: 1 if the participant has the motivation to get vax, 0 if notDo any of our categorical variables have a significant difference (or approaching statistically significance) in proxy treatment measures? If so, what are the effect sizes?
What are the top covariates that predict variation in our three different measures for proxy treatment effects? If we haven’t done so, we should test to see if we get any results from a causal forest.