vignettes/alternative_power.Rmd
alternative_power.Rmd
The alternative power framework summarizes how often simulated coefficient estimates meet a chosen effect threshold. Its logo represents the distribution of estimates across replications, with a highlighted tail beyond a threshold.
For a nonnegative threshold, simglm computes the
proportion of estimates greater than or equal to that threshold. For a
negative threshold, it computes the proportion less than or equal to the
threshold. Thresholds are on the fitted coefficient scale; for a
logistic model, for example, this is the log-odds scale.
This quantity describes threshold attainment by point estimates. It does not require a significance test or a confidence interval to exclude zero, and should be interpreted separately from conventional hypothesis-test power.
Specify thresholds as a named list under
power$thresholds. Names must match the fitted model terms
returned by broom::tidy(). Each term can have more than one
threshold.
set.seed(321)
sim_arguments <- list(
formula = y ~ 1 + x,
fixed = list(
x = list(var_type = "continuous", mean = 0, sd = 1)
),
error = list(variance = 1),
sample_size = 100,
reg_weights = c(0, 0.3),
model_fit = list(model_function = "lm"),
extract_coefficients = TRUE,
replications = 500,
power = list(thresholds = list(x = c(0.2, 0.3, 0.4)))
)
simulation_results <- replicate_simulation(sim_arguments)
threshold_results <- compute_statistics(
simulation_results,
sim_arguments,
power = FALSE,
type_1_error = FALSE,
precision = FALSE,
alternative_power = TRUE
)
threshold_results[
threshold_results$term == "x",
c("term", "threshold", "alt_power", "replications")
]
#> # A tibble: 3 × 4
#> term threshold alt_power replications
#> <chr> <dbl> <dbl> <dbl>
#> 1 x 0.2 0.826 500
#> 2 x 0.3 0.526 500
#> 3 x 0.4 0.162 500Here the generating coefficient for x is 0.3. The
alt_power column reports the fraction of its simulated
estimates reaching each requested threshold. As the positive threshold
increases, the fraction meeting it cannot increase. The 500 replications
keep this example quick; use more replications when greater Monte Carlo
precision is needed.
For an effect in the negative direction, specify negative thresholds for that term. A threshold of -0.2 counts estimates at or below -0.2. A threshold of zero uses the nonnegative rule and counts estimates at or above zero.
The same threshold summary can be used with the package’s other simulation designs. See Tidy simulation for the broader workflow and Spline simulation for thresholds applied to individual spline basis coefficients.