Backend Guide

Overview

ReproStat supports multiple model-fitting backends through the same high-level API. That means you can often keep the same reproducibility workflow while changing only the modeling engine.

Supported backends are:

• "lm" for ordinary least squares
• "glm" for generalized linear models
• "rlm" for robust regression via MASS
• "glmnet" for penalized regression via glmnet

This article explains when to use each one and what changes in the returned diagnostics.

Common interface

The same entry point is used across backends:

run_diagnostics(
  formula,
  data,
  B = 200,
  method = "bootstrap",
  backend = "lm"
)

The key differences are in:

• how the model is fit
• which quantities are available
• how to interpret selection-related outputs

Backend: lm

"lm" is the default backend and is the best place to start for standard linear regression.

diag_lm <- run_diagnostics(
  mpg ~ wt + hp + disp,
  data = mtcars,
  B = 100,
  backend = "lm"
)

reproducibility_index(diag_lm)
#> $index
#> [1] 90.13777
#> 
#> $components
#>       coef     pvalue  selection prediction 
#>  0.9218519  0.9000000  0.8100000  0.9736588

Use "lm" when:

• the response is continuous
• ordinary least squares is the intended analysis
• you want the simplest interpretation of all components

Backend: glm

Use "glm" when you need a generalized linear model, such as logistic or Poisson regression.

diag_glm <- run_diagnostics(
  am ~ wt + hp + qsec,
  data = mtcars,
  B = 100,
  backend = "glm",
  family = stats::binomial()
)

reproducibility_index(diag_glm)
#> $index
#> [1] 74.76995
#> 
#> $components
#>       coef     pvalue  selection prediction 
#>  0.2498896  1.0000000  0.8333333  0.9075752

Notes:

• if you provide family = ... while leaving backend = "lm", the function promotes the fit to "glm"
• prediction stability for GLMs uses response-scale predictions
• p-value and selection summaries remain available

Backend: rlm

Use "rlm" when you want robustness against outliers or heavy-tailed error behavior.

if (requireNamespace("MASS", quietly = TRUE)) {
  diag_rlm <- run_diagnostics(
    mpg ~ wt + hp + disp,
    data = mtcars,
    B = 100,
    backend = "rlm"
  )

  reproducibility_index(diag_rlm)
}
#> Warning in rlm.default(x, y, weights, method = method, wt.method = wt.method, :
#> 'rlm' failed to converge in 20 steps
#> $index
#> [1] 88.09575
#> 
#> $components
#>       coef     pvalue  selection prediction 
#>  0.9103266  0.7800000  0.8600000  0.9735035

Use "rlm" when:

• a few influential observations may distort OLS results
• you want a more robust regression baseline
• you still want coefficient, selection, prediction, and RI summaries in a familiar regression framework

Backend: glmnet

Use "glmnet" when you want penalized regression such as LASSO, ridge, or elastic net.

if (requireNamespace("glmnet", quietly = TRUE)) {
  diag_glmnet <- run_diagnostics(
    mpg ~ wt + hp + disp + qsec,
    data = mtcars,
    B = 100,
    backend = "glmnet",
    en_alpha = 1
  )

  reproducibility_index(diag_glmnet)
}
#> $index
#> [1] 84.08971
#> 
#> $components
#>       coef     pvalue  selection prediction 
#>  0.7308482         NA  0.8250000  0.9668430

The en_alpha argument controls the penalty mix:

• 1 gives LASSO
• 0 gives ridge
• values in between give elastic net

Important differences for "glmnet":

• p-values are not defined, so the pvalue component is NA
• selection stability measures non-zero selection frequency
• RI values are therefore based on a different component set than the non-penalized backends

Backend comparison summary

Choosing a backend in practice

A simple decision pattern is:

1. Start with "lm" if a standard linear model is appropriate.
2. Move to "glm" when the response distribution requires it.
3. Use "rlm" when outlier resistance matters.
4. Use "glmnet" when shrinkage, regularization, or sparse selection is the main modeling goal.

Comparing RI values across backends

Be careful when comparing RI values between penalized and non-penalized backends.

For "glmnet", the p-value component is unavailable, so the composite score is formed from a different set of ingredients. That makes cross-backend RI comparisons descriptive at best, not strictly apples-to-apples.

Model comparison with repeated CV

All backends can also be used in cv_ranking_stability():

models <- list(
  compact = mpg ~ wt + hp,
  fuller  = mpg ~ wt + hp + disp
)

cv_obj <- cv_ranking_stability(
  models,
  mtcars,
  v = 5,
  R = 20,
  backend = "lm"
)

cv_obj$summary
#>     model mean_rmse   sd_rmse mean_rank top1_frequency
#> 1 compact  2.709462 0.1712591         1              1
#> 2  fuller  2.824818 0.1573067         2              0

This is especially valuable when you are choosing between competing formulas and want to know not just which model is best on average, but which one is consistently best.

Next steps

For a broader conceptual explanation, read the interpretation article. For a complete first analysis, start with vignette("ReproStat-intro").