Psychology dissertation cardinality matching helps researchers select the largest feasible subset of exposed and comparison participants that meets prespecified covariate-balance constraints. It treats observational study design as a constrained optimisation problem before outcomes are analysed.
This guide explains the method, its assumptions, the distinction between cardinality and profile matching, and the diagnostics a psychology student should report. A running example compares students who used therapist-guided digital support with students who used self-guided support. Matching can improve comparability on measured baseline variables, but it cannot recreate random assignment or balance variables that were never measured.
What cardinality matching does
Many matching procedures choose observations because they are close according to a propensity score or multivariate distance. Cardinality matching begins with a different question: what is the largest subset that can satisfy explicit balance requirements? The optimisation retains as many eligible observations as possible while constraining differences in selected baseline covariate moments.
The foundational Zubizarreta, Paredes and Rosenbaum study separated two design tasks. First, cardinality matching selected a maximum balanced sample. Second, observations could be paired within that sample to reduce heterogeneity. This separation means selecting a balanced subset does not automatically create matched pairs.
The current MatchIt cardinality matching documentation describes a mixed integer programming approach. Users specify covariates, balance tolerances, a group ratio and an optimisation solver. The program seeks the largest subset meeting those conditions.
| Method | Primary design criterion | Typical output | Main caution |
|---|---|---|---|
| Nearest-neighbour matching | Choose close observations using a distance | Pairs or sets | Close pairs may still leave aggregate imbalance |
| Optimal pair matching | Minimise total within-pair distance | Optimised pairs | Distance choice drives the solution |
| Cardinality matching | Maximise retained sample under balance constraints | Balanced subset, optionally paired later | Selected subset may change the target population |
| Profile matching | Maximise subsets balanced to a fixed target profile | Target-aligned subsets | Requires a defensible target distribution |
| Weighting | Reweight rather than remove observations | Continuous analysis weights | Extreme weights can reduce precision |
The site’s propensity score matching guide explains the broader matching workflow. This article focuses specifically on maximum balanced subset selection, constraint specification, optimisation and the population represented after selection.
When psychology dissertation cardinality matching fits
The method can suit a dissertation using observational data with a clearly defined binary exposure, credible pre-exposure confounders and enough overlap to form a meaningful comparison. Secondary datasets may work well when sample size is adequate and variable timing is documented.
Suppose the exposure is therapist-guided rather than self-guided online anxiety support. Baseline anxiety, previous therapy, age, financial strain, academic workload and service access may influence both support choice and follow-up symptoms. Cardinality matching can retain the largest subset satisfying chosen balance tolerances for these variables.
A second example compares schools that introduced a wellbeing programme with schools that did not. Baseline wellbeing, attainment, school size, staffing, prior initiatives and region might define the balance constraints. If all programme schools are urban and all comparison schools are rural, no algorithm can recover the missing counterfactual comparison for rural programme schools.
Situations that require another approach
Do not use cardinality matching simply because a solver returns an answer. Repeated exposures with time-varying confounding need longitudinal methods. Continuous treatments need methods designed for dose-response questions. Clustered assignment may require school- or clinic-level design constraints and cluster-aware inference.
The method is also unsuitable when the exposure precedes important eligibility measurements, exposure versions are ambiguous, post-exposure variables are used as confounders, or the retained subset is too small to support the planned outcome model. Optimisation cannot repair a flawed causal timeline.
Define the causal question before matching
Write the target contrast in words before setting any tolerance. For example: among eligible university students represented by the matched sample, what is the average difference in twelve-week anxiety outcomes under guided rather than self-guided support? Define eligibility, treatment versions, time zero, follow-up and outcome measurement.
Cardinality matching between groups does not inherently preserve the average treatment effect among all treated participants. If treated observations are excluded, the selected sample represents the people who could be balanced under the constraints, not necessarily the original treated population. The MatchIt documentation explicitly warns that excluding treated units can make an ATT label inappropriate.
Profile matching addresses a related but different problem. It selects observations that resemble a prespecified target distribution, such as the original treated group or full eligible sample. In current MatchIt terminology, setting a non-fixed ratio differently changes whether the procedure acts as cardinality or profile matching. Describe what the software actually did, not just the function name.
Describe the matched-sample target
Compare included and excluded observations using baseline information only. Report how many participants remain in each exposure group and which eligibility profiles were lost. This makes the trade-off between internal comparability and generalisability visible.
Do not claim a population-wide effect if the matched subset excludes participants with severe symptoms, rare demographic combinations or limited service access. The estimate may be useful for a narrower overlap population, but that population must be described honestly.
Select covariates using causal reasoning
Use theory, prior evidence and a causal diagram to identify common causes of exposure and outcome. Include variables measured before exposure. A baseline outcome measure is often important because it can be strongly prognostic and may capture prior severity.
Avoid selecting covariates solely because they predict exposure or have small p-values. Do not balance mediators, exposure descendants or colliders. For instance, engagement recorded after support assignment may mediate part of the effect. Conditioning on it can remove a pathway of interest or introduce bias.
The site’s confounding variables guide provides a fuller selection framework. Document the timing, source, reliability and missingness of every proposed balance variable.
Choose balance functions, not only variable names
A mean constraint on age balances average age, not its variance, tails or interactions with baseline severity. Preserve continuous variables and add scientifically justified functions such as squared terms, categories at clinically meaningful thresholds, or interactions. Each added constraint consumes information and can reduce the achievable sample.
For a binary covariate, a mean difference is a proportion difference. For multi-category variables, represent all necessary indicator levels while avoiding redundant columns. Rare categories may require exact matching, aggregation justified before outcome inspection, or a narrower target population.
Set defensible balance tolerances
A tolerance states the largest imbalance the optimisation may accept for a covariate function. MatchIt currently uses standardised mean-difference tolerances and documents a default of 0.05, but a software default is not a universal scientific threshold. Predefine tolerances with attention to measurement quality, prognostic importance and feasible sample size.
Tighter constraints generally improve measured balance but may exclude more observations or make the problem infeasible. Looser constraints retain more observations but tolerate greater residual imbalance. Treat this as a design trade-off, not a search for whichever setting produces a significant outcome.
| Design decision | Question to answer | Evidence to retain | Warning sign |
|---|---|---|---|
| Target population | Who should the effect describe? | Eligibility and target definition | Target changes after results |
| Covariate set | Which pre-exposure causes require control? | Causal diagram and timing table | Mediators or colliders included |
| Balance functions | Are means sufficient? | Terms, transformations and interactions | Important distributional features omitted |
| Tolerances | How much residual difference is acceptable? | Prespecified thresholds and rationale | Repeated tuning against outcomes |
| Ratio | Is 1:1 or another ratio justified? | Group counts and precision rationale | Ratio chosen only to maximise significance |
| Solver | Was a feasible optimum found? | Status, runtime, gap and warnings | Time limit mistaken for proven optimum |
Use a design-stage tolerance sequence
A dissertation can prespecify several candidate tolerance sets, such as a strict primary design and a slightly relaxed sensitivity design. Evaluate them using sample retention, balance and population coverage without examining outcomes. Record every attempted specification so that unsuccessful designs do not disappear from the audit trail.
If no useful subset meets defensible constraints, report the lack of common support. Abandoning an unsupported causal comparison is a valid research finding. Quietly widening every tolerance until the solver succeeds is not.
Run the optimisation reproducibly
In R, MatchIt can perform cardinality matching with method = "cardinality". Record the package version, formula, estimand argument, ratio, tolerance vector, standardisation choice, exact-matching variables, solver, time limit and any pairing variables.
Current MatchIt documentation lists HiGHS, GLPK and Gurobi as solver options. HiGHS is the documented free default, while solver performance can differ. Solver choice is a computational decision that belongs in the methods section because different time limits or optimisation gaps can affect the selected subset.
Cardinality matching uses mixed integer programming. A result may be feasible without being proven globally optimal if the solver stops early. Retain the termination status, objective value, bound or optimisation gap when available, runtime and warnings. Re-run the saved script to confirm that the design is reproducible.
Distinguish subset selection from pairing
The selected subset may contain equal group counts without assigning each exposed participant a specific comparison partner. If the dissertation needs paired analysis or reduced within-pair heterogeneity, perform and document a second pairing stage using baseline variables only.
When additional pairing creates matched clusters, uncertainty estimates should respect that structure. The current MatchIt documentation recommends robust standard errors after cardinality or profile matching and cluster-robust standard errors when additional pairing is used.
Assess balance after selection
Do not assume that constraints guarantee every relevant form of balance. Produce before-and-after summaries for all prespecified variables and functions. Verify that the realised differences satisfy the intended constraints, then inspect distributional features not explicitly constrained.
The official cobalt documentation supports covariate balance tables and plots for matched and weighted designs. Report standardised mean differences, raw proportion differences for binary variables, variance ratios and empirical cumulative distribution differences when useful.
A Love plot offers a compact overview, but it should accompany a numerical table. Significance tests are poor balance diagnostics because their values depend on sample size. The goal is descriptive comparability, not failure to reject a null hypothesis.
Audit exclusions and positivity
Plot propensity or prognostic summaries for included and excluded observations, but remember that cardinality matching is not defined solely by a propensity distance. Tabulate retention by exposure group and substantively important baseline categories. Large losses in one subgroup can reveal practical positivity problems.
The JAMA guide indexed by PubMed describes cardinality matching as a design tool for balanced and representative observational samples. Representativeness is conditional on how the target and constraints are defined. It should be demonstrated, not assumed from the algorithm’s name.
Estimate outcomes and uncertainty
Only examine outcomes after the matched design passes predefined checks. Use an effect scale aligned with the question: a mean difference for a continuous symptom score, a risk difference for a binary recovery outcome, or another prespecified contrast that readers can interpret.
Outcome regression can adjust small residual imbalances and improve precision, but its variables and functional form should be specified before seeing treatment effects. Include the matching weights or subset indicators correctly. Do not analyse the retained sample as though it were an unstructured random sample.
Use robust or cluster-robust uncertainty as appropriate. A bootstrap must repeat all stages that contribute meaningful uncertainty, with care because optimisation can select different subsets across resamples. Explain the chosen variance estimator and its assumptions rather than reporting software defaults without justification.
| Report section | Minimum information | Interpretive safeguard |
|---|---|---|
| Design | Eligibility, exposure, time zero, outcome and target | Defines the causal contrast |
| Optimisation | Constraints, tolerances, ratio, solver and status | Shows how the subset was selected |
| Retention | Included and excluded counts by group | Reveals the represented population |
| Balance | Before-and-after tables and plots | Separates measured balance from assumptions |
| Effect | Estimate, confidence interval and variance method | Discourages p-value-only reporting |
| Limitations | Unmeasured confounding, overlap and generalisability | Prevents claims of randomisation |
Plan sensitivity analyses
Compare a prespecified strict design with a justified alternative tolerance set. Report how retention, balance, target composition and the estimated effect change. Do not present only the specification that produces the narrowest confidence interval.
Compare cardinality matching with a conventional propensity-score design using the same eligibility rules and confounder set. The site’s overlap weighting guide provides a useful contrast when retaining observations through bounded weights is scientifically preferable to subset selection.
Test plausible unmeasured confounding with a method suited to the effect scale and design. Examine whether missing-data assumptions affect membership in the matched subset. If multiple imputation is justified, explain whether matching occurs within each imputed dataset and how effects are combined.
Common cardinality matching mistakes
Calling every selected observation a matched pair
Cardinality matching selects a balanced subset. Pairing is optional and separate. Report pair identifiers only if a pairing stage actually occurred.
Claiming the original ATT after excluding treated participants
Exclusions can change the target population. Describe the retained treated group and avoid an ATT label that the final design does not support.
Using outcomes to tune tolerances
Balance choices belong to the design stage. Outcome-guided tuning creates researcher degrees of freedom and can bias inference.
Constraining means but ignoring distributions
Mean balance does not guarantee comparable variance, tails or interactions. Inspect the full distribution of important covariates.
Hiding infeasibility or solver limits
A time-limited feasible solution is not always a proven optimum. Report solver status, warnings and the gap when available.
Reproducible workflow for psychology dissertations
- Define eligibility, exposure versions, time zero, follow-up and outcome.
- State the intended target population and causal contrast.
- Select pre-exposure confounders using theory and a causal diagram.
- Specify balance functions, tolerances, exact constraints and group ratio.
- Inspect raw overlap without examining outcomes.
- Run cardinality or profile matching with a documented solver.
- Record optimisation status, retained counts and excluded profiles.
- Verify mean and distributional balance with tables and plots.
- Estimate the outcome using design-aware uncertainty.
- Run prespecified sensitivity analyses and preserve all code.
Keep a variable dictionary, analysis log, software versions and dated outputs. Protect confidentiality in any participant-level diagnostic display. Reproducible research does not require publishing restricted records.
Frequently asked questions
What is psychology dissertation cardinality matching?
It is a design method that selects the largest subset of observational units satisfying prespecified covariate-balance constraints.
Is cardinality matching the same as propensity-score matching?
No. Propensity-score methods commonly use a distance or caliper. Cardinality matching directly maximises retained sample size under balance constraints, although propensity information may still support diagnostics.
Does cardinality matching create pairs?
Not necessarily. It first selects a balanced subset. An optional second stage can pair observations within that subset.
What balance tolerance should I use?
No threshold is universally correct. Prespecify tolerances using scientific importance, measurement quality, overlap and the sample-size trade-off, then report the rationale.
Can cardinality matching estimate the ATT?
Only when the final design represents the intended treated population. If treated observations are excluded, the matched-sample target may no longer be the original ATT.
What if the optimisation finds no feasible match?
Check coding and redundant constraints, then reconsider overlap, target population and scientifically defensible tolerances. Do not relax constraints merely to obtain a preferred effect.
Does good balance remove unmeasured confounding?
No. Balance applies to observed, specified covariate functions. Causal interpretation still requires consistency, exchangeability, positivity and sound measurement.
Conclusion
Psychology dissertation cardinality matching is valuable when the goal is to retain the largest defensibly balanced observational sample under transparent constraints. Its strength is design-stage control over measured balance, not automatic representativeness or protection from hidden bias.
If you need support, choose ethical dissertation coaching that helps you define the target, audit code and interpret diagnostics while you remain the author. Responsible assistance should never invent data, conceal exclusions, guarantee significance or replace work required by your institution.
