Create an estimates_vcov from estimates and a vcov you already have

Constructor for estimates_vcov objects from an estimates data frame and a variance-covariance matrix supplied directly, rather than read off fitted models by as_estimates_vcov().

Use it whenever the covariances do not come out of a single regression. The main case is estimates that are correlated because they share subjects but cannot be stacked into one model: several experiments run on overlapping samples, where the covariance between their estimates is obtained by bootstrapping the whole design and taking cov() of the replicate estimates. It also serves the plumbing case of recombining the output of get_estimates_df() and get_vcov().

The vcov is matched to the estimates by position: row i of estimates_df is row and column i of vcov_matrix. Any dimnames on vcov_matrix are discarded and replaced with the object's id, so build both from the same ordered set of terms.

Usage

make_estimates_vcov(estimates_df, vcov_matrix)

Arguments

estimates_df: A data frame or tibble of coefficient estimates. Must contain an estimate column to be usable downstream; a std.error column (typically sqrt(diag(vcov_matrix))) is recommended so that rescale_estimates_vcov() has standard errors to rescale.
vcov_matrix: A variance-covariance matrix, in the same row order as estimates_df. Must be square, symmetric, and of the same dimension as nrow(estimates_df). A base matrix or a Matrix object are both accepted and neither is converted, so the storage of the result follows what you supply.

Value

An object of class estimates_vcov

Examples

# Two experiments on overlapping subjects: every subject takes the survey
# experiment, a random third also takes the lab experiment. The two effect
# estimates are correlated, but there is no single regression to read the
# covariance off, so bootstrap the design and use cov() of the replicates.
set.seed(123)
n <- 400
dat <- data.frame(
  Z_survey = rbinom(n, 1, 0.5),
  in_lab = rbinom(n, 1, 1 / 3)
)
dat$Z_lab <- ifelse(dat$in_lab == 1, rbinom(n, 1, 0.5), NA)
dat$Y_survey <- 0.2 * dat$Z_survey + rnorm(n)
dat$Y_lab <- 0.5 * dat$Z_lab + 0.6 * dat$Y_survey + rnorm(n)

estimate_both <- function(d) {
  c(
    survey = coef(lm(Y_survey ~ Z_survey, data = d))[["Z_survey"]],
    lab = coef(lm(Y_lab ~ Z_lab, data = d[d$in_lab == 1, ]))[["Z_lab"]]
  )
}

# Resample subjects, not rows within experiment, so the shared-sample
# correlation is what the replicates reproduce
boots <- t(replicate(200, estimate_both(dat[sample(n, n, replace = TRUE), ])))
V <- cov(boots)
point <- estimate_both(dat)

estimates_df <- data.frame(
  study = "study_1",
  term = names(point),
  estimate = point,
  std.error = sqrt(diag(V))
)

ev <- make_estimates_vcov(estimates_df, V)
ev
#> <estimates_vcov>
#> # 2 estimates with 2x2 vcov matrix
#> 
#> # A tibble: 2 × 5
#>   id    study   term   estimate std.error
#>   <chr> <chr>   <chr>     <dbl>     <dbl>
#> 1 1     study_1 survey    0.385     0.106
#> 2 2     study_1 lab       0.720     0.228
get_vcov(ev)
#>             1           2
#> 1 0.011197075 0.003696527
#> 2 0.003696527 0.052076563

Create an estimates_vcov from estimates and a vcov you already have

Usage

Arguments

Value

See also

Examples