Description
Logistic (or Poisson) regression for matched pairs, matched sets or small clusters that conditions out each set's own baseline, so effects are estimated only from within-set comparisons.
In 1-1 or 1-M matched designs, or when each person contributes only a few observations, fitting a separate intercept per set makes the number of parameters grow with the sample and badly biases the coefficients (by 100 percent for one binary covariate in 1-1 data). Conditioning on the number of events in each set removes these nuisance intercepts; the remaining coefficients are ordinary log odds ratios. Only sets that differ on a covariate inform it: for a binary covariate in matched pairs the estimate is the ratio of the two kinds of discordant pairs, consistent with McNemar's test and the Mantel-Haenszel estimator. Matching variables cannot enter as main effects but can interact with others. The same idea gives conditional Poisson regression for counts. It needs no distribution for the set effects, is nearly as efficient as full likelihood, but cannot estimate set-level covariates and becomes heavy beyond about 20 observations per cluster.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Discovery, Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Confirm the design: many matched sets or clusters, few observations in each, and within-set variation in the predictors of interest.
- Cross-tabulate case versus control values of each binary covariate to count discordant pairs; if one discordant type is absent the estimate is undefined.
- Select variables purposefully with univariable conditional fits, check continuous scales with quartile design variables, and test plausible interactions, including with matching variables.
- Fit by conditional maximum likelihood (for example Stata clogit or SAS PHREG); for 1-1 data, a standard logistic program on case-minus-control differences with outcome 1 and no intercept gives the same fit.
- Report exponentiated coefficients as within-set odds ratios with confidence limits; for counts, fit the conditional Poisson model and report rate ratios.
- Compute leverage, delta chi-square and delta beta per matched set, refit without influential sets, and judge them with subject-matter experts.
- If clusters are large use fixed intercepts instead; if both clusters and their sizes are moderate or large, use a random-intercept model.
Sources
- Applied Logistic Regression (Hosmer & Lemeshow, 2000)
- An Introduction to Categorical Data Analysis (Agresti, 1996)
- Mixed Models: Theory and Applications (Demidenko, 2004)
- Statistical Methods for Survival Data Analysis (Lee & Wang, 2003)
Also known as
- Conditional logistic regression
- Matched case-control logistic regression
- clogit
- Matched-pairs logit model
- Conditional Poisson regression
- Conditional likelihood for clustered binary and count data
Categories
- Analytics & Behaviour
- Statistics & Modelling