Statistical Analysis of Case-Control Studies: Conditional Logistic Regression, Matching and Reporting

Most published errors in case-control analysis arise from treating logistic regression as a single technique applicable to any dataset. A matched design analysed unconditionally yields an attenuated estimate, and a matched design adjusted for the matching variable in the wrong functional form yields intervals that are substantially incorrect rather than merely wide. The simulation evidence on this point is sufficiently specific to guide practice, and the correct analysis is no more difficult to conduct than the incorrect one.

Derivation of the analysis from the study design

Three features of the design determine most of the analytic plan, and they should be specified in the analysis plan before any data are examined.

The control sampling scheme determines what the odds ratio estimates. Controls sampled from those at risk at the time of each case’s event — density or risk-set sampling — give an odds ratio that estimates a rate ratio 25,29. Controls sampled as a random subset of the baseline cohort — case-cohort sampling — give a risk ratio. Controls sampled from those still free of the outcome at the end of the risk period — cumulative sampling — give a disease odds ratio that approximates a risk ratio only when the outcome is uncommon 25. See our page on case-control designs for the sampling detail.

Whether the design was matched determines whether the analysis must condition on the matched sets. If cases and controls were selected in matched sets, the matched structure is part of the data and discarding it changes the estimand.

What was matched on determines which variables can be studied and which cannot. A matched variable has had its variation between cases and controls removed by design, so its effect is not estimable from the matched comparison 2,29.

A practical consequence is that the statistical methods paragraph of a case-control paper should be derivable from the design paragraph. Where it is not — for example, where the design describes matching and the analysis describes unconditional regression — a reviewer will identify a problem before reading any results.

Conditional logistic regression and the matched set

Conditional logistic regression is the standard analysis for matched case-control data. Rather than estimating a separate parameter for every matched set — which would require a large number of nuisance parameters, each estimated from very little data — it conditions on the matched set, using only the within-set comparison of exposure between the case and its controls 2,28.

Two properties follow. The matching variables drop out of the model, because they do not vary within a set, which is why they should not be entered as covariates: doing so is either redundant or, with imperfect matching, actively misleading 2. Variables that were not matched on, but which confound the exposure-outcome relationship, must still be included as covariates, because conditioning on the matched set does not address them.

Conditional logistic regression performs well when the design is respected. In the simulation scenario where the functional form of the matching variable was misspecified, conditional logistic regression returned an estimate of -1.209 against a true value of -1.2, with confidence interval coverage of 0.953 — essentially nominal 2. This is the benchmark against which the alternatives should be judged.

Its cost is a loss of efficiency in a specific sense: matched sets in which all members share the same exposure status contribute nothing to the conditional likelihood, so a study with little within-set exposure variation will produce wide intervals 2,28. This is a design problem manifest in the analysis rather than an analytic defect, and it is a reason to avoid over-matching. Our biostatistics team and statistical review service both check this before submission.

Confidence interval coverage failure and its consequences relative to bias

The principal simulation finding is as follows. In the scenario where the functional form of the matching variable was misspecified in the model, adjusted unconditional logistic regression gave an estimate of -0.793 against a true value of -1.2, with confidence interval coverage of 0.479. Conditional logistic regression on the same data gave -1.209 with coverage of 0.953 2.

Both results are unfavourable to the unconditional analysis, and the second is more serious than the first. A point estimate of -0.793 for a true value of -1.2 represents substantial attenuation, with roughly a third of the effect lost. However, coverage of 0.479 means that a nominal 95% confidence interval contained the true value less than half the time. The interval is neither conservative nor merely imprecise; it is incorrect, while appearing entirely ordinary in a results table, and a reader of the paper could not detect the error.

This illustrates a more general principle of model misspecification. Bias in a point estimate is, at least in principle, detectable through sensitivity analysis and comparison with other studies. Coverage failure cannot be detected from within a single study, because the interval appears the same whether its coverage is 0.95 or 0.48 2. The protection is therefore structural, namely use of the analysis that the design implies, rather than diagnostic.

A separate and milder finding from the same work: under exact matching, unadjusted unconditional logistic regression underestimated the exposure effect in all six scenarios tested, with an estimate of 2.203 against a true value of 2.27 in one 2. That attenuation is small, but it is systematic and readily avoided.

Analysis of unmatched designs

Where controls were sampled without matching, unconditional logistic regression is the appropriate analysis, with confounders entered as covariates 2,28. The analytic questions are then those of any regression model, with two points specific to case-control designs.

First, the intercept is not interpretable. Because the ratio of cases to controls was chosen by the investigator rather than observed, the model cannot estimate absolute risk, and nothing in the output should be presented as one 17,28. The exposure coefficients are interpretable; the baseline is not.

Second, the covariate set should be chosen on substantive grounds and pre-specified, not by stepwise selection against the data. Stepwise procedures in a case-control analysis produce a model fitted to the sample’s idiosyncrasies and standard errors that do not account for the selection, which is a second route to intervals whose coverage is not what it claims.

Stratified analysis retains a role even where regression is used. Computing the exposure association within levels of a strong confounder, and presenting the stratum-specific estimates alongside the adjusted one, allows a reader to assess whether the adjustment is averaging over heterogeneity. Where the stratum-specific estimates differ materially, the single adjusted estimate conceals heterogeneity that a reader needs to know about.

Continuous exposures, functional form and dose-response relationships

The coverage finding above reflects a failure of functional form, and functional form also warrants attention for the exposure.

Entering a continuous exposure as a single linear term imposes the assumption that the log odds change by a constant amount per unit, across the whole range. That assumption is usually unexamined and frequently wrong, particularly where the exposure has a threshold or a plateau. The alternatives — categorisation into quantiles, fractional polynomials, splines — each have costs. Categorisation discards information and makes the estimate depend on arbitrary cut points. Flexible functions preserve information but are harder to report and can overfit with few cases.

A defensible default is to pre-specify the primary analysis with the exposure in the form supported on substantive grounds, to report a dose-response presentation by quantile as a secondary analysis for readability, and to check the shape with a flexible function as a sensitivity analysis. All three should be stated in the analysis plan, so that the form reported as primary is not the one that gave the smallest p-value.

Dose-response, where present, merits explicit presentation because it is one of the few features of an observational association that is difficult to produce by confounding alone. It is not proof of causation, and a monotone gradient can be produced by a confounder that is itself graded, but its absence where a mechanism predicts it is informative in the opposite direction. Our page on effect size covers how these estimates behave when pooled.

Interaction, subgroup analyses and multiplicity

Three cautions specific to this design.

Interaction is tested on a scale. An interaction present on the multiplicative scale may be absent on the additive scale and vice versa, and because a case-control study estimates ratios, the model naturally yields multiplicative interaction. Where the question concerns absolute joint effects — as it often does for public health or clinical decisions — the design does not supply the necessary baseline risk without external information.

Subgroup analyses are badly powered. A case-control study powered for a main effect is substantially underpowered for interaction, typically requiring several times the sample size 18. A subgroup finding in a case-control study should be presented as a hypothesis, and the power limitation stated rather than left to the reader.

Multiplicity is structural here. The design’s ability to examine many exposures at once is an advantage for hypothesis generation and a hazard for inference 15,18. The primary exposure and the primary analysis should be pre-specified, and the remaining analyses reported and labelled as exploratory. The alternative — presenting the strongest of twenty associations as the finding — is the mechanism behind a large share of non-replicable observational results 10.

Classification trees, random forests and model variability

Tree-based methods have been adapted for matched case-control data, with conditional logistic regression trees recursively partitioning the data while retaining the conditional likelihood so that the matched structure is respected 1. Their appeal is that interactions and non-linearities are identified rather than specified, which directly addresses the functional form problem that produced the coverage failure described above.

The principal limitation is instability. Single conditional logistic regression trees were found to be highly variable, meaning the structure recovered depends substantially on the particular sample — a well-known property of trees in general and a serious one when the number of matched sets is modest. A random forest built on conditional logistic regression trees was proposed in 2024 specifically to reduce that variability, by averaging over many trees grown on resampled data 1.

Two caveats apply to this approach. First, the ensemble gains stability at the cost of interpretability: there is no longer a single reportable structure, and the output is a prediction or a variable-importance ranking rather than an adjusted odds ratio for a named exposure. Second, the approach is recent, so the behaviour of these methods across the range of matched designs is not yet as well characterised as that of conditional logistic regression. The reasonable position is to use them as an exploratory complement — to identify interactions and non-linearities that are then modelled explicitly — rather than as the primary analysis of a confirmatory question. Our page on R for trial analysis covers implementation and reproducibility, and predictive analytics is where prediction-focused work sits.

Reporting requirements

Eight items are required, and a statistical reviewer typically checks most of them within the first two minutes.

The control sampling scheme, described specifically enough to be named, and the measure the odds ratio therefore estimates 7,25. The matching variables with their tolerances, and the number of controls per case. The model fitted, stated as conditional or unconditional with the reason. Confirmation that matching variables were not entered as covariates in a conditional model. The functional form used for every continuous variable, including the matching variables where any adjustment was made. The covariate set and how it was chosen, with pre-specification stated. The number of matched sets contributing to the conditional likelihood, which is lower than the number of sets whenever exposure does not vary within a set. And which analyses were pre-specified and which were exploratory.

Two sentences carry most of the weight: one naming the sampling scheme and the measure estimated, and one stating the model and why it follows from the design. A paper containing both can be assessed; a paper containing neither cannot, whatever its results show. See the STROBE case-control checklist and our analysis plan work.

Key points

Application to your study

Frequently asked questions

@misc{statistical_analyses_of_case_control_s_2026,
  author       = {{Pubrica Academy}},
  title        = {Statistical Analysis of Case-Control Studies: Conditional Logistic Regression, Matching and Reporting},
  howpublished = {Pubrica},
  year         = {2026},
  month        = {oct},
  url          = {https://pubrica.com/academy/research/statistical-analyses-of-case-control-studies/},
  note         = {Accessed: <span data-cite-acc="iso">2026-10-15</span>}
}

CASE-CONTROL ANALYSIS PLAN STUB

1. DESIGN FACTS THAT DETERMINE THE ANALYSIS
   Control sampling scheme:
     [ ] density / risk-set  -> OR estimates a RATE RATIO
     [ ] case-cohort         -> OR estimates a RISK RATIO
     [ ] cumulative          -> disease ODDS RATIO only
   Matched?  [ ] yes  [ ] no
     Matched on: ______________________________________________
     Tolerances: ______________________________________________
     Controls per case: ______
   Variables matched on CANNOT be studied as exposures: ________

2. PRIMARY MODEL
   [ ] Conditional logistic regression (matched design)
       [ ] matching variables NOT entered as covariates
   [ ] Unconditional logistic regression (unmatched design)
       [ ] intercept will not be interpreted as risk
   Software and function: ______________________________________
   Reason the model follows from the design, in one sentence:
     __________________________________________________________

3. EXPOSURE SPECIFICATION
   Primary exposure: __________________________________________
   Form: [ ] binary [ ] categorical, cut points: ______________
         [ ] continuous linear  [ ] spline, knots: ____________
         [ ] fractional polynomial
   Pre-specified PRIMARY form: ________________________________
   Secondary presentation (e.g. quartiles): ___________________
   Sensitivity check on shape: ________________________________

4. COVARIATES
   Confounders to include (substantive grounds, list them):
     __________________________________________________________
   [ ] NO stepwise or data-driven selection will be used
   Functional form for each continuous covariate:
     Variable ................ Form ...........................
     Variable ................ Form ...........................

5. INFORMATIVE DATA
   Expected number of matched sets: ______
   [ ] Number contributing to the conditional likelihood will be
       reported
   Over-matching risk assessed?  [ ] yes - note: ______________

6. SECONDARY AND EXPLORATORY ANALYSES  (label honestly)
   Pre-specified secondary: ___________________________________
   Exploratory, including subgroups: __________________________
   [ ] Paper will state the study is powered for the main effect
       only
   Interaction scale to be reported: [ ] multiplicative
                                     [ ] additive - source of
                                         baseline risk: ______

7. SENSITIVITY ANALYSES
   [ ] Unconditional analysis reported alongside conditional,
       with the expected direction of discrepancy explained
   [ ] Alternative functional form for matching variable
   [ ] Alternative covariate set
   [ ] Exclusion of sets with incomplete matching

8. REPORTING CHECK BEFORE SUBMISSION
   [ ] Sampling scheme named and measure estimated stated
   [ ] Model stated as conditional or unconditional, with reason
   [ ] Matching variables confirmed excluded from a conditional model
   [ ] Functional form given for every continuous variable
   [ ] Informative matched set count reported
   [ ] Pre-specified vs exploratory clearly separated
   [ ] No absolute risk reported from these data alone
   [ ] STROBE case-control checklist mapped