Effect of an exposure on a binary outcome, with intervals that hold up with few clusters

Fits a logistic regression and reports the coefficient of the exposure with a confidence interval that accounts for clustering (persons in the same county or tract share unmeasured influences).

Usage

fitExposureEffect(
  formula,
  data,
  exposure,
  cluster,
  interval = c("jackknife", "cr1", "model", "glmm"),
  level = 0.95
)

Arguments

formula

A model formula with a 0/1 outcome, for example y ~ pm25 + ses + age + female.

data

A data frame.

exposure

Name of the exposure term whose coefficient is reported.

cluster

Name of the column of data that identifies the clusters, or a vector of cluster ids.

interval

"jackknife" (default): leave-one-cluster-out standard error, t quantile. "cr1": clustered sandwich standard error, t quantile. "model": model-based standard error, normal quantile (ignores clustering). "glmm": logistic mixed model with a random intercept for the cluster (lme4::glmer(), needs the lme4 package), Wald interval.

level

Confidence level.

Value

A one-row data frame with estimate, se, lower, upper, level, method, n_clusters and n. Coefficients are log-odds per unit of exposure.

Examples

set.seed(1)
d <- data.frame(county = rep(1:30, each = 40), x = rnorm(1200))
d$y <- rbinom(1200, 1, plogis(-1 + 0.3 * d$x + rep(rnorm(30, sd = 0.3), each = 40)))
fitExposureEffect(y ~ x, d, exposure = "x", cluster = "county")
#>    estimate         se     lower     upper level    method n_clusters    n
#> 1 0.2325906 0.05609663 0.1178601 0.3473211  0.95 jackknife         30 1200