Psychology dissertation optimal full matching builds variable-size subclasses containing both exposure groups, then uses subclass weights to estimate a clearly defined treatment effect.
It is a flexible design for observational research when fixed one-to-one pairs would discard useful participants or fail to balance important baseline characteristics. Each subclass can contain one exposed participant and several controls, or one control and several exposed participants.
Full matching can retain the complete eligible sample when overlap and restrictions permit, but retention alone does not guarantee precision or validity. This guide explains estimands, distances, subclass formation, weights, effective sample size, diagnostics, outcome analysis, and transparent reporting for psychology dissertations.
What optimal full matching does
Consider a dissertation comparing students who chose a therapist-supported anxiety programme with students who used standard university support. Before enrolment, the groups may differ in anxiety severity, prior therapy, medication, age, study demands, help-seeking attitudes, and financial pressure.
Full matching groups comparable participants into subclasses. Every subclass includes at least one participant from each exposure group. Some subclasses may contain one programme participant and several controls. Others may contain several programme participants and one control.
The official MatchIt optimal full matching documentation describes the method as assigning units to subclasses while minimising the sum of absolute distances between exposed and comparison participants within those subclasses. The resulting subclass membership is converted into matching weights for the selected estimand.
Why variable-size subclasses help
A fixed pair design forces every exposed participant to use the same number of controls. This can be inefficient when some regions of the covariate distribution contain many comparable controls and others contain few. Full matching adapts the local ratio to the data.
Suppose three students with high baseline anxiety have only one comparable control, while two students with moderate anxiety have six. Full matching can place the first group together and distribute the moderate-anxiety controls across other subclasses. It avoids demanding the same ratio where overlap differs.
The method is optimal only relative to the specified distance, restrictions, and objective. It does not prove that the covariates are sufficient, that subclasses are internally balanced, or that unmeasured confounding is absent.
| Feature | Pair matching | Optimal full matching |
|---|---|---|
| Subclass size | Usually fixed pairs or fixed ratio | Variable ratios across subclasses |
| Participant retention | Unmatched units are often removed | All units can be retained when feasible |
| Primary output | Pairs, subclasses, and matching weights | Variable-size subclasses and estimand weights |
| ATE support | Often limited for focal-group pair designs | Can target ATT, ATC, or ATE |
| Main precision concern | Discarded observations | Unequal weights and reduced effective sample size |
When psychology dissertation optimal full matching fits
The method suits binary observational exposures with adequate common support and a meaningful reason to preserve both groups. Psychology applications can include optional therapy uptake, school wellbeing programmes, digital mental health services, caregiver interventions, employee resilience courses, or naturally chosen support formats.
It is especially useful when the control-to-exposed ratio varies across covariate space. It can also be considered when one-to-one or fixed-ratio matching produces poor balance, discards many eligible participants, or does not align with an average treatment effect target.
Full matching is not appropriate when the exposure is poorly defined, matching covariates occur after exposure, or there is almost no overlap. Optimisation cannot recover counterfactual information that the data do not contain.
Full sample does not mean full identification
The word full refers to assigning eligible units to subclasses under the default unconstrained design. It does not mean that the design controls every source of bias. If a caliper, common-support rule, or exact restriction makes a unit impossible to match, it may still be omitted.
Even when every unit receives a subclass, extreme ratios can create highly variable weights. A dataset with 1,000 records can have far less than 1,000 records’ worth of effective information. Report nominal retention and effective sample size separately.
Define the estimand before choosing weights
The estimand identifies the target population for the causal contrast. The average treatment effect in treated participants asks what those who received the exposure would have experienced under the comparison. The average treatment effect in controls reverses the target. The average treatment effect asks about the full eligible population.
The estimand changes the subclass weights. It therefore belongs in the research question and analysis plan, not in a last-minute software setting. The MatchIt documentation supports ATT, ATC, and ATE for optimal full matching.
For the anxiety-programme example, an ATT may be appropriate if the question concerns students who chose the programme. An ATE may fit a policy question about offering the programme to all eligible students. The two effects can differ when programme benefits vary across baseline risk or willingness to participate.
Build a defensible time zero
Define eligibility, exposure assignment, follow-up start, outcome, and censoring on a common timeline. Baseline anxiety, previous therapy, and study load must be measured before programme enrolment. Do not match on adherence, session attendance, post-enrolment medication changes, mediators, or later attrition.
Matching on consequences of exposure can block part of the effect or introduce collider bias. A simple causal diagram and dated variable inventory help prevent this error.
Select baseline covariates carefully
Include credible pre-exposure causes of both exposure and outcome, as well as strong outcome predictors. Use theory, prior evidence, clinical or educational knowledge, and temporal ordering. Do not select covariates only because they predict exposure or have a small p-value.
Plausible variables in the example include baseline anxiety, depressive symptoms, previous counselling, medication status, age, disability accommodations, study level, workload, sleep quality, financial strain, and help-seeking attitudes. Site or referral pathway may require exact agreement if access rules differ.
Avoid redundant variables and deterministic combinations. Total scores entered alongside their component sums can destabilise covariance-based distances. With many weakly relevant covariates, participants can become distant in high-dimensional space without improving confounding control.
Plan missing-data handling before matching
Complete-case deletion can change the target population and damage overlap. If multiple imputation is appropriate, the full design and analysis should generally be repeated within each imputed dataset, followed by a justified pooling procedure. Matching once on a single completed dataset understates uncertainty and may misrepresent design variability.
Report missingness by exposure group, variables used in the imputation model, number of imputations, how subclass weights were handled, and how effect estimates were pooled.
Choose the distance and restrictions
Optimal full matching is an assignment method, not a distance. A propensity-score distance, Mahalanobis distance, prognostic score, or supplied distance matrix defines what it means for participants to be close.
A propensity score estimates exposure probability from baseline covariates. Its purpose here is design, not exposure prediction. A high area under a receiver operating curve does not show good matching. The relevant outcome is balance on the original covariates after weighting.
Mahalanobis distance can directly preserve closeness on a modest set of continuous and ordinal covariates. A hybrid can use Mahalanobis distance within a propensity-score caliper. Document which variables determine the score, which determine the Mahalanobis distance, and how restrictions are scaled.
Use calipers and exact matching to prevent implausible subclasses
A caliper prohibits links that exceed a chosen distance. Exact matching forces agreement on selected categories, such as site or diagnostic pathway. These rules can prevent unacceptable comparisons but may make full retention impossible.
Inspect common support before optimisation. If exposure groups occupy separated regions, redefine the target population, restrict eligibility, or acknowledge that the intended effect is not supported. Do not force distant matches merely to preserve a nominal full sample.
| Design decision | Question | What to report |
|---|---|---|
| Estimand | Whose effect is targeted? | ATT, ATC, or ATE and its population |
| Distance | What makes participants comparable? | Model, variables, link, transformations, or matrix |
| Exact restrictions | Which categories must agree? | Variables and substantive rationale |
| Calipers | Which links are too distant? | Width, scale, and resulting omissions |
| Ratio bounds | How unequal may subclasses become? | Minimum, maximum, and mean control constraints |
| Tolerance | How closely was the optimisation solved? | Setting, convergence, and software versions |
Control subclass ratios and computational demands
Unrestricted full matching may create subclasses with extreme exposed-to-control ratios. These can achieve the distance objective while producing unstable weights. Ratio constraints can limit the number of controls per exposed participant or the reverse.
The MatchIt matching-methods vignette describes controls for minimum, maximum, and mean subclass ratios. Treat these as design choices with consequences for balance, feasibility, and precision.
Large distance matrices can make optimal full matching slow or infeasible. Generalised full matching uses a faster clustering algorithm and may provide a near-optimal alternative for larger datasets, although its objective and available controls differ. Record any substitution and do not present it as the identical method.
Numerical convergence is not scientific validity
Report the solver tolerance and any convergence warnings. A tighter tolerance may matter for smaller optimisation problems, but numerical precision cannot repair weak overlap, measurement error, omitted confounders, or a poorly defined exposure.
Preserve code, package versions, random seeds used elsewhere, distance specifications, ratio bounds, and participant identifiers mapped to subclasses. Reproducibility requires more than reporting the function name.
Understand full-matching weights
Subclass membership determines how much each record contributes to the target estimand. Within each subclass, the smaller exposure group effectively represents comparable participants from the larger group. The exact formula depends on whether ATT, ATC, or ATE is targeted.
Do not overwrite matching weights with ordinary regression weights or treat them as frequencies. Inspect their range, distribution, and concentration by exposure group. A few large weights can dominate the estimate even when no participants were dropped.
If the original study uses sampling weights, incorporate them according to validated guidance for both the matching and effect-estimation stages. Explain whether the final analysis weight combines sampling and matching components.
Report effective sample size
Effective sample size summarises the information retained after unequal weighting. It typically falls as weights become more variable. The official matching-methods vignette cautions that full matching can use all units yet have a lower effective sample size than one-to-one matching.
Report the effective sample size separately for exposure groups before and after weighting. Compare it with nominal sample size, balance, and the precision of the effect estimate. Retention is valuable only when the retained observations contribute useful, stable information.
Assess balance in aggregate and within subclasses
The optimisation minimises distance, not covariate imbalance directly. The official MatchIt balance guide states that balance is not guaranteed and must be assessed.
Calculate weighted standardised mean differences before and after matching for every prespecified covariate. Add variance ratios, empirical cumulative distribution differences, and visual comparisons for important continuous variables. Report proportions for categorical levels.
Use the same standardisation denominator before and after matching. Do not rely on p-values because their power changes with sample size and weighting. A common heuristic is an absolute standardised mean difference below 0.10, with tighter expectations for strongly prognostic variables, but no threshold proves exchangeability.
Aggregate balance can hide poor subclasses
Weighted balance may look good overall even when particular subclasses compare dissimilar participants. Inspect subclass-specific balance, distance distributions, and extreme ratios. The balance guide notes that unadjusted effects within poorly balanced subclasses should not be treated as unbiased merely because aggregate balance is good.
Assess prespecified nonlinear functions and interactions. If age has a curved relationship with outcome, check a nonlinear age term. If baseline anxiety operates differently by previous therapy, inspect that interaction or subgroup.
| Diagnostic domain | Useful measure | Warning sign |
|---|---|---|
| Mean balance | Weighted absolute standardised mean differences | Important covariates remain meaningfully different |
| Distribution balance | Variance ratios, eCDF and density plots | Similar means conceal spread or tail differences |
| Subclass quality | Within-subclass distances and balance | Several internally implausible comparisons |
| Weight stability | Minimum, maximum, percentiles and plots | A few records dominate the analysis |
| Information | Effective sample size by group | Large nominal sample but little effective information |
| Target population | Participant flow and exclusions | Restrictions silently change the estimand |
Analyse outcomes after full matching
Freeze the design before examining exposure-effect results. Outcome analysis should preserve matching weights and account for subclass or cluster dependence. The appropriate model depends on the outcome type, estimand, sampling structure, and study design.
The MatchIt effect-estimation guide discusses weighted outcome regression, marginal contrasts, and uncertainty estimation after matching. Covariate adjustment in the matched sample can address small residual imbalance and improve precision, but it does not rescue a failed design.
For a continuous wellbeing score, a weighted regression can estimate an adjusted marginal mean difference. For a binary outcome, report an absolute risk difference when possible alongside relative measures. For clustered data, account for schools, clinics, therapists, families, or repeated observations as appropriate.
Estimate uncertainty consistently
Standard errors must reflect weighting and relevant clustering. Cluster-robust approaches and carefully implemented bootstrap procedures are common options, but suitability depends on the estimator. State exactly what was clustered, whether subclasses were used, and whether any resampling repeated the complete design.
Report the effect estimate, confidence interval, and interpretation on a meaningful scale. Avoid treating statistical significance as the only result. Discuss whether the interval includes effects that are clinically, educationally, or practically important.
Compare full matching with related methods
Optimal pair matching minimises total distance across fixed pairs or ratios. Full matching allows variable-size subclasses and can target the ATE. The flexibility changes weighting and precision, not just the visual arrangement of participants.
Propensity score matching usually refers to selecting pairs or fixed-ratio sets using the estimated score. Full matching can use a propensity score as its distance but creates variable subclasses and matching weights.
Cardinality matching maximises retained sample size under explicit balance constraints. Genetic matching searches for distance weights that improve balance. Coarsened exact matching groups participants using prespecified bins. Select among them by estimand, overlap, balance, effective sample size, transparency, and computation.
Sensitivity analyses for a credible dissertation
Compare a small set of prespecified, defensible specifications without consulting outcome effects. Options include ATT versus ATE when both questions are meaningful, propensity-score versus Mahalanobis distance, tighter calipers, exact matching on site, alternative ratio bounds, and optimal versus generalised full matching.
Summarise balance, omissions, weight distributions, effective sample size, and effect estimates for each. Do not keep changing settings until a preferred outcome appears.
Use a sensitivity analysis for unmeasured confounding that matches the effect measure and design. It cannot prove that no confounder exists. It quantifies how strongly an omitted factor would need to relate to exposure and outcome to alter the conclusion.
Maintain a design log
Record each attempted specification, why it was considered, its diagnostics, and whether outcomes remained hidden. This creates an audit trail that distinguishes legitimate design improvement from result-driven searching.
Common mistakes and repairs
Equating full retention with no bias: report balance, overlap, weights, effective sample size, and remaining assumptions.
Choosing the estimand after seeing results: define the target population before matching and calculate the corresponding weights.
Ignoring extreme subclass ratios: inspect ratios and weights, then add defensible bounds if needed.
Matching on post-exposure variables: rebuild the timeline and retain only genuine baseline covariates.
Reporting only aggregate balance: inspect within-subclass quality and important subgroups.
Using unweighted outcome analysis: preserve the matching weights and account for subclass or cluster dependence.
A reproducible workflow
- Define eligibility, exposure strategies, time zero, follow-up, outcome, and estimand.
- Select pre-exposure covariates through substantive knowledge and a causal model.
- Audit missingness, measurement, collinearity, outliers, and overlap.
- Choose the distance, exact restrictions, calipers, ratio bounds, and tolerance.
- Run full matching without viewing exposure-effect estimates.
- Inspect convergence, subclass composition, omissions, and distance distributions.
- Assess weighted and subclass-specific balance on original covariates and key functions.
- Review weight distributions and effective sample size, then freeze the design.
- Fit an outcome model aligned with the estimand, weights, subclasses, and clustering.
- Run prespecified sensitivity analyses and report all remaining limitations.
Frequently asked questions
Does optimal full matching always retain everyone?
No. It can assign all eligible units when a feasible match exists, but calipers, exact restrictions, missing data, or inadequate overlap may leave some units unmatched.
How is full matching different from optimal pair matching?
Pair matching uses fixed pairs or ratios. Full matching creates variable-size subclasses and uses subclass-derived weights, enabling more flexible retention and support for ATT, ATC, or ATE.
Can full matching use Mahalanobis distance?
Yes. It can use Mahalanobis distance directly or within a propensity-score support rule or caliper. State which variables contribute to each component.
Why can effective sample size fall if all records remain?
Unequal weights give some records more influence than others. When influence is concentrated, the weighted sample contains less independent information than its nominal record count suggests.
Should subclasses have fixed ratios?
Not necessarily. Variable ratios are central to full matching. However, bounds may be useful when extreme ratios create unstable weights or scientifically weak comparisons.
Is aggregate balance sufficient?
No. Check overall weighted balance and inspect subclass-specific distances, ratios, and important covariate patterns. Strong aggregate balance can coexist with weak local comparisons.
What should the methods chapter report?
Report the estimand, timeline, covariates, missing-data plan, distance, restrictions, calipers, ratio bounds, tolerance, software versions, subclass structure, weights, effective sample size, balance, outcome model, uncertainty method, and sensitivity analyses.
Conclusion
Psychology dissertation optimal full matching is most useful when variable-size subclasses improve balance or retention while supporting a clearly stated estimand. Its quality depends on the distance, overlap, subclass weights, effective sample size, diagnostics, and outcome analysis, not on retaining every row.
If you need support, seek ethical methodological guidance that helps you justify and reproduce each decision while leaving data ownership, interpretation, and authorship with you. A responsible review should test assumptions and strengthen your independent understanding rather than promise a particular result.
