Description
An analyst-led procedure for choosing regression covariates that combines univariable screening, confounder checks, add-back, scale checks and interaction tests instead of mechanical stepwise selection.
Purposeful selection treats model building as a judgement the analyst owns, with statistics as support rather than substitute. Every candidate is first examined on its own; anything with a univariable p below 0.25, plus anything known to matter, enters one multivariable model. Variables are then removed only when they neither contribute (Wald and likelihood ratio tests) nor act as confounders, judged by a marked change in other coefficients when they are dropped. Excluded candidates are added back one at a time, continuous covariates are checked for linearity in the logit with quartile design variables or fractional polynomials, and plausible interactions are tested singly. Hosmer and Lemeshow contrast this with stepwise and best-subsets routines, which admit noise variables, report p-values that are not true significance levels and miss confounding that only appears collectively. The result is a staged sequence of models, each defensible on both statistical and subject-matter grounds.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Discovery, Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Analyse each candidate alone: likelihood ratio chi-square and odds ratios for categorical variables, a univariable fit plus a lowess smooth on the logit scale for continuous ones, and a check for zero cells.
- Enter every variable with univariable p below 0.25 and every variable of known substantive importance into one multivariable model.
- Remove variables that are non-significant by Wald test, refit, confirm with a likelihood ratio test, and restore any variable whose removal shifts other coefficients markedly (a common rule of thumb is a change above about 20 percent).
- Repeat delete-refit-verify, then add back each originally excluded variable one at a time to form the preliminary main effects model.
- Check the scale of each continuous covariate with quartile design variables or fractional polynomials and adopt a transformation only if it is significant and makes substantive sense.
- List plausible interactions, add each product term singly, and keep those significant by likelihood ratio test to form the preliminary final model.
- Pass the model to goodness-of-fit and diagnostic checks before interpreting it.
Sources
- Applied Logistic Regression (Hosmer & Lemeshow, 2000)
Also known as
- Hosmer-Lemeshow model-building strategy
- Purposeful variable selection
Categories
- Analytics & Behaviour
- Statistics & Modelling