What each component does to a profile, on the scale the data are reported in

A loading is a number on the standardized modelling scale, so a table of them says which visits a component is concentrated on and not what moving along it looks like. This runs the model's own inversion instead: from the cohort centre, move one component by a given number of score standard deviations, invert the standardization and the endpoint's log transform, and report the feature values that come back.

Usage

pca_component_effect(x, sds = 1)

Arguments

x

A dataset from synpmx_pca(), or its trial summary.

sds

How far to move along each component, in score standard deviations. Defaults to 1. The spread used is the pooled residual standard deviation across arms, which is the scale pca_scores() reports per arm.

Value

A data frame with one row per component, feature and displacement: component, feature, kind, endpoint, time, covariate, level, score_sd (-sds, 0 or +sds), and value on the reported scale. Marked "restricted_not_releasable": the centre is a mean of real data.

Details

The result is the component in the units the study measured. A component that raises the whole concentration profile shows as a curve shifted up at every time; one that separates early visits from late shows as a curve that crosses the centre. Read it as a description of the fitted basis, not as pharmacology: no clearance was estimated and a score standard deviation is a variance decomposition, so naming a mechanism for a curve is a claim the fit does not support.

The reference is the cohort centre — score zero on every component — which is the whole cohort's mean feature vector rather than any arm's. Arms differ by their mean score vectors, which pca_scores() reports.

See also

pca_components() for the loadings themselves, pca_features() for the grid, pca_scores() for the per-arm score model.

Examples

data <- pmx_simulated_fixture(60)
roles <- pmx_roles(
  id = "ID", time = "TIME", nominal_time = "NTIME", dv = "DV", amt = "AMT",
  evid = "EVID", cmt = "CMT", dvid = "DVID", mdv = "MDV"
)
effect <- pca_component_effect(synpmx_pca_summarize(data, roles))
head(effect)
#>    feature          kind endpoint  time covariate level component score_sd
#> 1 dv_cp__1 endpoint_cell       cp  0.25      <NA>  <NA>       PC1       -1
#> 2 dv_cp__2 endpoint_cell       cp  1.00      <NA>  <NA>       PC1       -1
#> 3 dv_cp__3 endpoint_cell       cp  2.00      <NA>  <NA>       PC1       -1
#> 4 dv_cp__4 endpoint_cell       cp  6.00      <NA>  <NA>       PC1       -1
#> 5 dv_cp__5 endpoint_cell       cp 12.25      <NA>  <NA>       PC1       -1
#> 6 dv_cp__6 endpoint_cell       cp 13.00      <NA>  <NA>       PC1       -1
#>      value
#> 1 1.085682
#> 2 8.686082
#> 3 4.342978
#> 4 1.628557
#> 5 1.302831
#> 6 9.228971