R Model Diagnostics

How it works

Residuals compare observed and fitted responses. Leverage describes how predictor geometry influences a fitted value. Neither quantity alone certifies a model. Check row alignment, assumptions, unusual cases and held-out behaviour. This fictional six-row example has noise so the diagnostic calculations are not based on a perfect zero-residual line.

Run this R example

These examples run locally with R, not PHP, Python or JavaScript. The reviewed baseline is R 4.5.2 on Windows x64/UCRT, using only packages shipped with R. This is a tested version, not a newest-release claim. Install your own R from the official R project; an IDE is optional. Save the complete source as model-diagnostics.R in a separate learning folder and run Rscript --vanilla model-diagnostics.R. Keep code and fictional fixtures separate from a real project.

Bhylu has no R browser runner. There is no Try in playground link because the existing Python and JavaScript editors do not execute R. No package installation, account, network dataset, Apache, MySQL or PHP setting is needed to run these base-R sources.

The complete download uses a local scope, restores its temporary options and treats unexpected warnings as errors. Examples that handle an expected warning do so narrowly. Random examples change only their R process’s RNG state. File examples own their temporary output; figure examples leave their named PNG in the current learning folder and refuse to overwrite it. Use a new empty folder to rerun a figure.

Download the complete model-diagnostics.R · Download all source files with hash checks

For the optional collection downloader, save download-examples.mjs in an empty folder and run node download-examples.mjs. Local preview before publishing uses node download-examples.mjs http://127.0.0.1:8000/examples/r/. It refuses an existing bhylu-r-examples folder and does not install or execute R.

Example

local({
  previous_options <- options(warn = 2, digits = 8, stringsAsFactors = FALSE)
  on.exit(options(previous_options), add = TRUE)
  sessions <- data.frame(visits = 1:6, minutes = c(7, 10, 10, 14, 14, 18))
  fit <- lm(minutes ~ visits, data = sessions)
  errors <- residuals(fit)
  leverage <- hatvalues(fit)
  cat("residuals:", length(errors), "all finite:", all(is.finite(errors)), "\n")
  cat("leverage sum:", sprintf("%.3f", sum(leverage)), "\n")
  cat("estimated noise positive:", summary(fit)$sigma > 0, "\n")
  invisible(NULL)
})

sessions <- data.frame(visits = 1:6, minutes = c(7, 10, 10, 14, 14, 18))
fit <- lm(minutes ~ visits, data = sessions)
errors <- residuals(fit)
leverage <- hatvalues(fit)
cat("residuals:", length(errors), "all finite:", all(is.finite(errors)), "\n")
cat("leverage sum:", sprintf("%.3f", sum(leverage)), "\n")
cat("estimated noise positive:", summary(fit)$sigma > 0, "\n")

Expected result

residuals: 6 all finite: TRUE
leverage sum: 2.000
estimated noise positive: TRUE

The leverage sum is two for this full-rank model with an intercept and one predictor. The result is a property of its design matrix, not a claim that every observation is trustworthy. Fitted residuals are not independent held-out prediction errors.