Conditional logistic regression for matched case-control data
Conditional logistic regression compares cases with their matched controls within each set, estimating an exposure odds ratio while removing each set’s baseline intercept.
Fit
When to use it, and when not to
Use it when
- Cases were individually matched to controls and the set identifiers are available.
- You need within-set exposure comparisons while adjusting for covariates that vary within sets.
Not when
- Groups were only frequency matched in broad strata. An unconditional logistic model adjusting for those strata may be suitable. Read: Logistic regression
- Your cohort was propensity matched on treatment rather than sampled by case status. Use an analysis for that matched cohort outcome. Read: Propensity score matching
Before you start
What you need
- Case status, exposure and a matched-set identifierRetain every set’s sampling structure and control selection rule.
- Covariates that vary within setsExactly matched variables have no separate estimable coefficient.
- The source population and control sampling schemeThe sampling scheme determines how the exposure odds ratio relates to disease risk or incidence.
- Controls represent the relevant source population under the chosen sampling scheme.How to check: Document eligibility, selection and any risk-set sampling.
- The conditional log odds are described by the specified exposure and covariate terms.How to check: Inspect nonlinear terms and interactions; do not assume one coefficient captures every exposure effect.
- Enough informative within-set contrasts exist.How to check: For a lone binary exposure, concordant pairs add no exposure information; sparse discordant pairs can cause separation.
- Sets are independent, or dependence is addressed in inference.How to check: Repeated people, families or shared clinics may require a variance method suited to that dependence.
- Missing at random given observed information may justify imputation that preserves the matched structure.How to check: Explain how missingness affects whole sets. Compare sensitivity analyses when unseen values may drive missingness.
Plain language
How it works
Condition on the number of cases in each set. The likelihood compares who was the case among the sampled set members. This removes the set-specific intercept without fitting one coefficient per set. Exponentiating an exposure coefficient gives a conditional odds ratio. Breslow and colleagues develop this approach for matched case-control sampling. In the simple binary-exposure pair example, the estimate is exposed-case-only pairs divided by exposed-control-only pairs.
Concordant pairs contribute no exposure information in that simple model. Pearce explains why matching alone does not control confounding and why some matching schemes allow unconditional adjustment. Do not generalize the discordant-pair formula to multivariable or larger-set models. Use a conditional-likelihood profile interval or a log-scale Wald interval, with inference suited to any additional dependence.
Worked example
A worked example
- Setting
- A binary-exposure study has matched case-control pairs, with no additional covariates.
- Question
- What is the exposure odds ratio within matched pairs?
- Approach
- There are 30 pairs with an exposed case only and 10 with an exposed control only. Use the discordant-pair conditional likelihood.
- Result
- The odds ratio is 3.00. The log-scale standard error is 0.365; the approximate 95% Wald interval is 1.47 to 6.14.
- What it means
- This is an exposure odds ratio from matched pairs, not a threefold risk. The interval uses conditional-likelihood information, not a binomial proportion formula.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Matching is a sampling choice, not automatic control of confounding. Conditional logistic regression uses within-set information and cannot estimate a factor constant within every set.
- Using an unmatched exposure table and ignoring individual matching.Instead: Preserve matched sets or justify an alternative model that adequately adjusts for matching factors.
- Trying to estimate the main effect of an exactly matched factor.Instead: Its within-set contrast is absent. State which effects are identifiable.
- Using a Wilson interval for the odds ratio.Instead: Use a profile conditional-likelihood interval or exponentiate a Wald interval for the log coefficient; consider sparse-data methods when needed.
Inside Mallard
How Mallard uses this
The method registry links this method to its published analysis article and code.
conditional-logistic-matched-case-control).analysis.conditional_logistic, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Estimation of multiple relative risk functions in matched case-control studies (opens in a new tab) · Am J Epidemiol · doi 10.1093/oxfordjournals.aje.a112623
- Analysis of matched case-control studies (opens in a new tab) · BMJ · doi 10.1136/bmj.i969