Tabulate Truncated Weights Over a Grid of Bounds
Source:R/wt_trunc_sensitivity.R
wt_trunc_sensitivity.Rdwt_trunc_sensitivity() truncates a set of weights with wt_trunc() at
each bound in a grid and tabulates what each truncation does to them: the
bound as given and as applied, the number of weights moved, and the range
and mean of the weights that result. The first row describes the weights
before any truncation, so every other row can be read against it. This is
the weight side of the sensitivity table Cole and Hernán (2008) recommend
for choosing a bound.
Usage
wt_trunc_sensitivity(
.weights,
method = c("pctl", "wt", "count"),
lower = NULL,
upper = NULL
)Arguments
- .weights
A psw vector, such as one returned by
wt_ate(), or a numeric vector of weights. Propensity scores, including those returned byps_trim(),ps_trunc(), andps_calibrate(), are refused with an error of classpropensity_type_error, and weights that have already been truncated with an error of classpropensity_already_modified_error.- method
The rule that reads each bound, one of
"pctl"(the default),"wt", or"count", as inwt_trunc(). The"adaptive"method sets a single bound from the number of weights, so it has no grid to vary and is not accepted.- lower
An optional grid of lower bounds, read according to
methodaswt_trunc()readslower. A single value is used with every upper bound; otherwiselowermust have the same length asupper, and the two are paired element by element.- upper
The grid of upper bounds, read according to
methodaswt_trunc()readsupper. For"pctl"it defaults toc(0.99, 0.975, 0.95, 0.90); for"wt"and"count"it is required.
Value
A tibble with one row for the untruncated weights followed by one row per bound, and the columns:
lower,upper: the bounds as given,NAin the first row andlowerNAwhen no lower bound was given.lower_value,upper_value: the bounds as applied, on the scale of the weights, recorded bywt_trunc(). Both areNAin the first row, andlower_valueisNAwhen no lower bound was given.n_truncated: the number of weights moved, 0 in the first row.min,max,mean: the smallest, largest, and mean weight present.range_ratio:max / min.
Details
Each row after the first is computed by calling wt_trunc() with that
row's bounds, so it describes exactly the weights that call returns. Every
bound in the grid must be one wt_trunc() accepts: a single invalid bound
refuses the whole grid, with the error wt_trunc() would raise for it,
rather than leaving a gap in the table. The rows follow the grid in the
order given.
The summaries are taken over the weights that are present, so missing
weights, such as those of units set aside by ps_trim(), take no part in
them and are never counted as truncated. A grid over weights of which none
are present is refused with an error of class propensity_range_error.
Weights can be zero, and a zero weight makes range_ratio infinite.
Weights that have already been truncated are refused, because the first
row would then describe weights that were already bounded rather than the
weights before truncation. Pass the weights as they were before
wt_trunc() instead.
What is left to the caller
The table does not include effect estimates or effective sample sizes.
propensity fits no outcome models, so to add estimates to the table, map
over the grid, truncate the weights at each bound with wt_trunc(), and fit
the outcome model with each set of weights. For the effective sample size,
add the weights before and after truncation to a data frame as separate
columns and pass them to halfmoon::check_ess().
References
Cole, S. R., & Hernán, M. A. (2008). Constructing inverse probability weights for marginal structural models. American Journal of Epidemiology, 168(6), 656–664.
See also
wt_trunc() to truncate the weights at a chosen bound.
Examples
set.seed(1)
n <- 200
x <- rnorm(n)
z <- rbinom(n, 1, plogis(2 * x))
fit <- glm(z ~ x, family = binomial)
w <- wt_ate(fit, .exposure = z)
#> ℹ Treating `.exposure` as binary
# The default percentile grid. The first row is the untruncated weights.
wt_trunc_sensitivity(w)
#> # A tibble: 5 × 9
#> lower upper lower_value upper_value n_truncated min max range_ratio mean
#> <dbl> <dbl> <dbl> <dbl> <int> <dbl> <dbl> <dbl> <dbl>
#> 1 NA NA NA NA 0 1.02 31.6 31.0 2.05
#> 2 NA 0.99 NA 7.94 2 1.02 7.94 7.79 1.91
#> 3 NA 0.975 NA 6.44 5 1.02 6.44 6.32 1.87
#> 4 NA 0.95 NA 4.31 10 1.02 4.31 4.23 1.79
#> 5 NA 0.9 NA 3.33 20 1.02 3.33 3.27 1.72
# A grid of absolute bounds, and a lower tail bounded at every one of them
wt_trunc_sensitivity(w, method = "wt", upper = c(20, 10, 5))
#> # A tibble: 4 × 9
#> lower upper lower_value upper_value n_truncated min max range_ratio mean
#> <dbl> <dbl> <dbl> <dbl> <int> <dbl> <dbl> <dbl> <dbl>
#> 1 NA NA NA NA 0 1.02 31.6 31.0 2.05
#> 2 NA 20 NA 20 1 1.02 20 19.6 1.99
#> 3 NA 10 NA 10 2 1.02 10 9.81 1.93
#> 4 NA 5 NA 5 8 1.02 5 4.91 1.82
wt_trunc_sensitivity(w, method = "wt", lower = 1.05, upper = c(20, 10, 5))
#> # A tibble: 4 × 9
#> lower upper lower_value upper_value n_truncated min max range_ratio mean
#> <dbl> <dbl> <dbl> <dbl> <int> <dbl> <dbl> <dbl> <dbl>
#> 1 NA NA NA NA 0 1.02 31.6 31.0 2.05
#> 2 1.05 20 1.05 20 12 1.05 20 19.0 1.99
#> 3 1.05 10 1.05 10 13 1.05 10 9.52 1.93
#> 4 1.05 5 1.05 5 19 1.05 5 4.76 1.83
# The largest one, three, and ten weights
wt_trunc_sensitivity(w, method = "count", upper = c(1, 3, 10))
#> # A tibble: 4 × 9
#> lower upper lower_value upper_value n_truncated min max range_ratio mean
#> <dbl> <dbl> <dbl> <dbl> <int> <dbl> <dbl> <dbl> <dbl>
#> 1 NA NA NA NA 0 1.02 31.6 31.0 2.05
#> 2 NA 1 NA 12.5 1 1.02 12.5 12.3 1.95
#> 3 NA 3 NA 7.58 3 1.02 7.58 7.43 1.90
#> 4 NA 10 NA 4.29 10 1.02 4.29 4.21 1.79