Regression & Modeling Guard
Regression models come with assumptions that are routinely violated in practice: linearity, homoscedasticity, non-influential outliers, absence of multicollinearity. Violations do not produce errors — they produce biased estimates and wrong standard errors, silently. The Regression & Modeling Guard runs diagnostic checks automatically and surfaces violations before they reach the page.
What is Regression & Modeling Guard?
The guard covers three specific failure modes in regression analyses: assumption violations detectable through standard diagnostics (residual plots, Cook's distance, VIF); inappropriate application of OLS to compositional outcomes (addressed in 4.4.4 for test selection, here extended to regression); and sample size adequacy for the complexity of the model (number of predictors relative to n).
Diagnostics are not optional. OLS regression diagnostics (residual vs. fitted, Q-Q plot, Cook's distance) are standard practice in any statistics textbook. Most published regression results in biology, psychology, and medicine are reported without them. This is the gap the Regression & Modeling Guard addresses.