rstanarm

Misc

Linear Regression

• Interaction: continuous x categorical

  stan_mod <- stan_glm(log_gdp_std ~ rugged_std_c * cid, 
                      data = dd, 
                      family = gaussian(link = "identity"), 
                      seed = 12345)

  • Linear models still use stan_glm but with a gaussian distribution family

• Interaction: categorical x categorical

  rstan_mod1 <- stan_glm(
    admitted ~ gender*dept,
    data = ucb_01,
    family = binomial
  )

Discrete Distribution Models

• Binomial (SR, Ch. 11)

  tot_mod <- stan_glm( 
    cbind(admit, reject) ~ 0 + applicant.gender, 
    prior = normal(0,1), 
    data = ucb, 
    family = binomial 
  )

  • Can also use a logistic model, but need case-level data (e.g. 0/1)

    • Deaggregate count data into 0/1 case-level data

    data(UCBadmit, package = "rethinking")
    ucb <- UCBadmit %>%
      mutate(applicant.gender = relevel(applicant.gender, ref = "male"))
    
    # deaggregate to 1/0
    deagg_ucb <- function(x, y) {
      UCBadmit %>%
        select(-applications) %>%
        group_by(dept, applicant.gender) %>%
        tidyr::uncount(weights = !!sym(x)) %>%
        mutate(admitted = y) %>%
        select(dept, gender = applicant.gender, admitted)
    }
    
    ucb_01 <- purrr::map2_dfr(c("admit", "reject"),
                              c(1, 0),
                              ~ disagg_ucb(.x, .y)
    )

  • Logistic

    rstan_mod1 <- stan_glm(
      admitted ~ gender*dept,
      data = ucb_01,
      family = binomial
    )