Matched Case Comparison¶
Quasi-experimental method — instantiates Counterfactual Comparison
Pairs cases or periods that are similar on key attributes so differences in outcomes can be interpreted relative to a more credible counterfactual baseline.
Matched Case Comparison constructs the counterfactual out of real, existing cases: it pairs the focal unit with one or a few actual others deliberately chosen to resemble it on the attributes that drive the outcome, so the matched twin's result stands in for what the focal case would have done untreated. Its whole discipline is balance — making the treated and comparison cases indistinguishable on the measured covariates — and its whole vulnerability is the covariate it failed to measure. Unlike a method that projects a trend or blends many units into a composite, it never manufactures a comparison unit; it goes and finds one that already exists and defends the match.
Example¶
A school district rolls out a new structured-literacy curriculum in twelve elementary schools and wants to know whether reading scores rose because of the curriculum or for reasons that had nothing to do with it. There is no randomized rollout to lean on, so the evaluators build matched pairs: for each adopting school they find a non-adopting school in the same district that is close on the factors known to move reading scores — share of students on free lunch, English-learner percentage, prior-year reading level, and class size. A common way to collapse those factors into a single closeness score is propensity-score matching.[1] Once the pairs are balanced, they read the difference in this year's reading gains between each adopting school and its matched partner. The matched schools drifted up too — demographics and district-wide coaching lifted everyone — but the adopting schools rose noticeably more, and that pair-by-pair gap is the estimated curriculum effect, far more credible than comparing the adopters against the district average.
How it works¶
The method turns on choosing the right things to match on and then proving the match is good. First, name the covariates that actually drive the outcome — matching on irrelevant traits buys nothing, and missing a real driver lets its effect masquerade as the treatment's. Second, find comparison cases balanced on those covariates, whether by exact matching, nearest-neighbor within a tolerance, or a propensity score that summarizes many covariates at once. Third, check balance: confirm the matched groups really are similar on every covariate before reading any outcome. Only then is the treated-minus-matched difference taken. The comparison unit is real throughout; the work is in selection and balance, not construction.
Tuning parameters¶
- Matching covariates — which traits the pair must balance on. Omit a driver of the outcome and the estimate silently absorbs its effect.
- Matching method — exact, caliper/nearest-neighbor, or propensity-score. Tighter matching improves comparability but shrinks how many cases can be paired at all.
- Match ratio — one comparison case per treated case, or several. More comparators steady the estimate but usually at the cost of match quality.
- Balance tolerance — how close is "close enough" before a pair is accepted. Strict tolerance improves fairness; loose tolerance keeps sample size but weakens the counterfactual.
When it helps, and when it misleads¶
Its strength is fairness with transparency: it uses observed, real cases anyone can inspect, and it neutralizes the confounders it matches on, which makes it far more defensible than comparing a treated unit to a raw average. When randomization is impossible but comparable cases exist, it is often the strongest counterfactual available.
Its defining failure mode is unobserved confounding — matching can only balance the covariates you measured, so any hidden difference between treated and comparison cases (motivated principals, a school that self-selected into the program because it was already improving) survives the match and biases the estimate in an unknown direction. Thin overlap (no genuinely comparable case exists) and accidentally matching on a consequence of the treatment are the other traps. The classic misuse is presenting a well-balanced table of observed covariates as if it proved the groups were comparable on everything. The guarding discipline is to match only on pre-treatment variables, report balance honestly, acknowledge the residual unobserved gap plainly, and prefer randomization wherever it is feasible.
How it implements the components¶
actual_path— records the focal case's real course and the context needed to identify what a fair comparator must resemble.comparison_unit_matching— the core act: selecting real comparison cases balanced on the outcome-driving covariates, and proving the balance.outcome_comparison— the pair-by-pair difference between the focal case and its matched twin is the estimated effect.
It does not build a weighted composite reference_baseline from many donor units, nor validate one with placebo sensitivity_probe tests — that construction-and-falsification is Synthetic Control Method, its nearest twin; Matched Case Comparison rests on a single real, hand-matched partner instead of a synthesized one.
Related¶
- Instantiates: Counterfactual Comparison — Matched Case Comparison supplies a real comparable case as the counterfactual when no experiment is possible.
- Sibling mechanisms: Synthetic Control Method · Baseline Comparison · Scenario Contrast · Counterfactual History Review · What-If Analysis · Control Group Comparison · A/B Test
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Matched Case Comparison operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it pairs cases or periods that are similar on key attributes so differences in outcomes can be interpreted relative to a more credible counterfactual baseline.
Independent corroboration: The frozen evidence defines Matched Case Comparison as 'Pairs cases or periods that are similar on key attributes so differences in outcomes can be interpreted relative to a more credible counterfactual baseline', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: Matching as counterfactual construction is a central observational-study method in statistics and causal inference.
Related originating lineages:
- Ethnography & Qualitative Methods — Comparative case-study traditions independently developed matched-case logic for controlled comparison.
Review resolution: Both independent reviews place the primary provenance in statistics_experimental_design. The queued differences (reported_ambiguity, alternate_origin_disagreement, origin_mode_disagreement) concern secondary metadata, not primary lineage. The final retains ethnography_qualitative_methods only where a reviewer supplied a formative-lineage rationale; downstream use or broad applicability by itself is not treated as origin. origin_mode=convergent because the supplied rationales identify independently established or materially co-developing traditions. domain_reach=multi_domain records established application breadth separately from provenance. confidence=medium preserves the more cautious evidence assessment. encyclopedia_synthesis=false records whether either reviewer identified deliberate corpus-level composition.
Attribution caveat: The label spans quantitative matching and qualitative comparative-case design.
Review outcome: Reconciled after independent review; medium confidence.
References¶
[1] Rosenbaum, P. R., & Rubin, D. B. "The Central Role of the Propensity Score in Observational Studies for Causal Effects". Biometrika 70(1), 41–55 (1983). Defines the propensity score as a scalar summary of observed covariates and identifies matched sampling on it as an application. registry ↩