Sweep instrument strength

Runs a custom DGP that varies the G->X coefficient (pi_GX) and records IV2SLS performance binned by first-stage partial F. Uses fit_iv2sls() with min_f = 0 (no weak-IV guard) so bias is visible across all instrument strengths.

Usage

sweep_instrument_strength(
  pi_GX_grid = c(0.02, 0.05, 0.1, 0.15, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8),
  n_iter = 50,
  n = 200,
  k = 1,
  conf_str = 0.8,
  tau = 0.25,
  coverage = 0.7,
  n_cores = 1
)

Arguments

pi_GX_grid

Numeric vector of G->X coefficients to sweep.

n_iter

Number of replications per grid point.

n

Sample size.

k

Number of confounders.

conf_str

Confounding strength.

tau

True total effect.

coverage

Negative-control coverage.

n_cores

Number of parallel workers for the replicate loops. Default 1 (sequential). Uses parallel::mclapply on Unix and a PSOCK cluster on Windows.

Value

A data frame with columns pi_GX, arm, iter, partial_F, beta, se, pvalue, rejected.

Examples

res <- sweep_instrument_strength(pi_GX_grid = c(0.05, 0.2), n_iter = 2)
#> Running instrument-strength sweep...
#>  pi_GX=0.05: 4 tasks (sequential)
#>   pi_GX=0.05: 25% (1/4) [0s]
#>   pi_GX=0.05: 50% (2/4) [0s]
#>   pi_GX=0.05: 75% (3/4) [0s]
#>   pi_GX=0.05: 100% (4/4) [0s]
#>  pi_GX = 0.05 done (mean F = 1.9)
#>  pi_GX=0.2: 4 tasks (sequential)
#>   pi_GX=0.2: 25% (1/4) [0s]
#>   pi_GX=0.2: 50% (2/4) [0s]
#>   pi_GX=0.2: 75% (3/4) [0s]
#>   pi_GX=0.2: 100% (4/4) [0s]
#>  pi_GX = 0.20 done (mean F = 37.4)
head(res)
#>   pi_GX  arm iter  partial_F       beta        se    pvalue rejected
#> 1  0.05  alt    1  4.4765442 0.35939447 0.3463920 0.3007575    FALSE
#> 2  0.05 null    1  2.3100322 0.60615488 0.4923835 0.2197672    FALSE
#> 3  0.05  alt    2  0.2283631 1.46302545 2.5320641 0.5640593    FALSE
#> 4  0.05 null    2  0.4112861 1.71448760 2.6166955 0.5130963    FALSE
#> 5  0.20  alt    1 16.5591634 0.32253791 0.1957043 0.1009286    FALSE
#> 6  0.20 null    1 48.5586431 0.01714582 0.1021277 0.8668451    FALSE