Skip to contents

Compute Prognostic Scores for Balance Assessment

Source: R/bal_prognostic_score.R
bal_prognostic_score.Rd

Calculates prognostic scores by fitting an outcome model on the control group and generating predictions for all observations. Prognostic scores represent the expected outcome under control conditions and are useful for assessing whether balance on treatment predictors ensures balance on outcome-relevant variables.

Usage

bal_prognostic_score(
  .data,
  outcome = NULL,
  .covariates = everything(),
  .exposure,
  formula = NULL,
  .reference_level = NULL,
  family = gaussian(),
  .weights = NULL,
  na.rm = FALSE,
  ...
)

Arguments

.data

A data frame containing the variables for analysis

outcome

The outcome variable. Can be specified as a bare name or character string. Ignored if formula is provided.

.covariates

Variables to include in the outcome model. Defaults to everything() which includes all variables except the outcome and treatment. Supports tidyselect syntax. Ignored if formula is provided.

.exposure

The treatment variable. Can be specified as a bare name or character string.

formula

Optional formula for the outcome model. If provided, outcome and .covariates arguments are ignored. The formula should not include the treatment variable.

.reference_level

The level of treatment that represents the control group. If NULL (default), the first level is used as control.

family

A family object for the GLM. Defaults to gaussian() for continuous outcomes. Use binomial() for binary outcomes.

.weights

Optional weights for the outcome model. Can be a numeric vector or bare name of a variable in .data.

na.rm

Logical. Should missing values be removed? Defaults to FALSE.

...

Additional arguments passed to glm().

Value

A numeric vector of prognostic scores with the same length as the number of rows in .data (after NA removal if na.rm = TRUE).

Details

The prognostic score method, introduced by Stuart et al. (2013), provides a way to assess balance that focuses on variables predictive of the outcome rather than just the treatment assignment. The procedure:

  1. Fits an outcome model using only control group observations

  2. Generates predictions (prognostic scores) for all observations

  3. Returns these scores for balance assessment

This approach is particularly useful when:

  • The outcome model includes non-linearities or interactions

  • You want to ensure balance on outcome-relevant variables

  • Traditional propensity score balance checks may miss important imbalances

The returned prognostic scores can be used with existing balance functions like bal_smd(), bal_vr(), or check_balance() to assess balance between treatment groups.

References

Stuart EA, Lee BK, Leacy FP. Prognostic score-based balance measures can be a useful diagnostic for propensity score methods in comparative effectiveness research. J Clin Epidemiol. 2013;66(8):S84-S90.

Examples

# Using tidyselect interface
prog_scores <- bal_prognostic_score(
  nhefs_weights,
  outcome = wt82_71,
  .exposure = qsmk,
  .covariates = c(age, sex, race, wt71)
)

# Using formula interface
prog_scores_formula <- bal_prognostic_score(
  nhefs_weights,
  .exposure = qsmk,
  formula = wt82_71 ~ age + sex + race + wt71 + I(age^2)
)

# Add to data and check balance
nhefs_with_prog <- nhefs_weights
nhefs_with_prog$prog_score <- prog_scores
check_balance(nhefs_with_prog, prog_score, qsmk, .weights = w_ate)
#> # A tibble: 8 × 5
#>   variable   group_level method   metric  estimate
#>   <chr>      <chr>       <chr>    <chr>      <dbl>
#> 1 prog_score 0           observed ks      0.124   
#> 2 prog_score 0           w_ate    ks      0.0316  
#> 3 prog_score 0           observed smd    -0.267   
#> 4 prog_score 0           w_ate    smd     0.00305 
#> 5 prog_score 0           observed vr      1.06    
#> 6 prog_score 0           w_ate    vr      1.01    
#> 7 NA         NA          observed energy  0.0442  
#> 8 NA         NA          w_ate    energy  0.000954

# For binary outcome
prog_scores_binary <- bal_prognostic_score(
  nhefs_weights,
  outcome = death,
  .exposure = qsmk,
  family = binomial()
)
#> Warning: prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases