A psychology dissertation quasi-experimental design evaluates a plausible cause without randomly assigning every participant to a condition. It can be the right choice when a university, school, clinic, employer, or public authority has already introduced a programme, policy, threshold, or service change. The design is not simply a weaker experiment. Its credibility depends on explaining how exposure occurred, constructing a defensible counterfactual, testing rival explanations, and matching the analysis to the assignment process.
This guide shows how to turn a naturally occurring comparison into a feasible psychology dissertation. It covers research questions, comparison groups, interrupted time series, difference-in-differences, regression discontinuity, matching, ethics, assumptions, analysis, sensitivity checks, and transparent reporting. The aim is not to make causal language sound impressive. It is to make every claim proportionate to what the design and data can support.
What is a quasi-experimental design in psychology?
A quasi-experiment studies the effect of an exposure or intervention when assignment is not fully controlled through randomisation. Participants may receive a programme because of location, eligibility, timing, need, institutional choice, or a numerical threshold. The researcher then uses design features and statistical evidence to estimate what would probably have happened without that exposure.
That missing outcome is the counterfactual. A student who attended a stress-management programme cannot simultaneously be observed at the same time as the same person without the programme. A comparison group, repeated pre-intervention measurements, or an assignment cutoff must therefore stand in for the unobserved alternative. The quality of that substitute matters more than the label “quasi-experimental.”
A conventional psychology dissertation experimental design uses random assignment to make groups comparable on average before the intervention. A quasi-experiment does not receive that protection automatically. It must show why the groups or trends are comparable, identify threats such as selection and history, and check whether the assumptions needed for the chosen estimator are plausible.
| Feature | Randomised experiment | Quasi-experiment | Simple observational comparison |
|---|---|---|---|
| Exposure assignment | Random process controlled by the study | Rule, event, timing, threshold, or institutional process | Observed without a design-based assignment strategy |
| Main causal protection | Expected baseline balance | Credible counterfactual plus assumptions | Measured adjustment, often with greater residual uncertainty |
| Key task | Protect randomisation and limit attrition | Reconstruct assignment and test rival explanations | Describe associations and avoid causal overstatement |
| Typical conclusion | Effect under the trial conditions | Estimated effect for a defined population, time, or threshold | Association after stated adjustments |
When is the design suitable for a dissertation?
Use a quasi-experimental design when the research question concerns the effect of a clearly defined exposure and random assignment is impossible, unethical, or outside the researcher’s authority. Suitable opportunities often arise when implementation differs across places or dates, an eligibility score determines access, or a service retains repeated outcome records before and after a change.
For example, a university may introduce a peer-support programme in two departments before extending it elsewhere. A dissertation could examine whether distress scores changed more in the early departments than in comparable departments not yet receiving the programme. The intervention, timing, units, outcomes, and comparison are identifiable. By contrast, comparing students who voluntarily downloaded a wellbeing app with those who did not may be heavily confounded by motivation, initial distress, digital confidence, and help-seeking. Statistical adjustment cannot guarantee recovery of information that was never measured.
Begin with an effect question
Write the question so that it names the population, exposure, comparator, outcome, and timing. A useful form is: “Among first-year students, what was the change in weekly anxiety scores after the introduction of an academic-transition programme, compared with students at a similar campus where implementation occurred later?” This is more precise than asking whether the programme “worked.”
Specify the estimand, meaning the effect you intend to estimate. It might be the average effect among eligible participants, the effect near an eligibility cutoff, or the change attributable to implementation during a defined period. This decision determines whom the conclusion describes. The research-question guide can help align the wording with available evidence.
Reconstruct the assignment mechanism
Ask who decided exposure, when the decision occurred, which information was used, whether participants could influence the decision, and whether implementation followed the rule. Obtain policy documents, enrolment criteria, dates, referral forms, or administrative notes where possible. If the assignment mechanism remains unclear, the analysis is built on an assumption that cannot be examined.
Choose the strongest feasible quasi-experimental design
The best method follows the real assignment process. Do not choose matching because software makes it convenient or call two measurements an interrupted time series. The design should reflect why some observations were exposed and others were not.
| Design | Useful setting | Core identifying idea | Major check |
|---|---|---|---|
| Non-equivalent comparison group | Programme implemented in one group but not another | Groups are sufficiently comparable after design and adjustment | Baseline balance and differential attrition |
| Difference-in-differences | Exposed and comparison groups observed before and after change | Outcome trends would have remained parallel without exposure | Multiple pre-period trends and concurrent events |
| Interrupted time series | Many repeated observations around a known intervention date | A level or slope departure exceeds the expected pre-intervention pattern | Seasonality, autocorrelation, and co-interventions |
| Regression discontinuity | Exposure assigned by a cutoff on a continuous score | Cases just above and below the cutoff are locally comparable | Manipulation at the threshold and model sensitivity |
| Matching or weighting | Rich baseline data for exposed and unexposed cases | Measured covariates can create a balanced comparison | Overlap, balance, and unmeasured confounding |
Non-equivalent comparison groups
This design compares an exposed group with a group that did not receive the intervention through random assignment. Pre-intervention measurement is crucial because similar post-intervention means do not show that the groups started in the same place. Compare baseline outcome levels and important prognostic variables using plots, standardised differences, and substantive judgement rather than relying only on significance tests.
The U.S. Institute of Education Sciences provides detailed guidance on designing quasi-experiments. Its standards emphasise baseline equivalence, appropriate statistical adjustment, concurrent outcome measurement, and the danger of a single unit being perfectly confounded with condition. These principles transfer well to psychology. One intervention classroom compared with one control classroom cannot separate programme effects from teacher, classroom, or site effects.
Difference-in-differences
Difference-in-differences compares the change over time in an exposed group with the change over the same period in a comparison group. Suppose sleep quality improves by four points after a timetable reform at one campus and improves by one point at a comparison campus. The basic difference-in-differences estimate is three points. It subtracts common change but does not automatically remove group-specific shocks.
The central assumption is that, without the reform, the groups would have followed parallel outcome trends. One pre-intervention measurement cannot reveal a trend. Several pre-period observations are much more informative. Plot them, examine whether slopes differ, identify anticipation of the intervention, and document other events occurring at either site. A recent methodological comparison describes both single-group and multiple-group quasi-experimental approaches and their assumptions in an accessible peer-reviewed overview.
Interrupted time series
An interrupted time series estimates whether an outcome’s level, slope, or both changed at a defined intervention point. It requires many observations before and after the interruption, not merely one pretest and one post-test. Outcomes might be weekly counselling referrals, monthly recorded distress, or daily mental-health service use.
Plot the raw series before fitting a model. Check seasonality, secular trend, autocorrelation, delayed effects, changes in data collection, and unusual events. Specify whether the expected effect is immediate, gradual, temporary, or delayed. Penfold and Zhang explain why interrupted time series can be a strong quasi-experimental approach and discuss segmented regression and difference-in-differences in their methodological tutorial. A controlled interrupted time series is generally stronger when a credible unexposed series experiences the same background conditions.
Regression discontinuity
Regression discontinuity is appropriate when a continuous assignment score and known cutoff determine exposure. For example, students scoring at or above a screening threshold may be offered an intensive support programme. The estimated effect applies locally to cases near that threshold, not automatically to all students.
Record the assignment variable before treatment, preserve its continuous form, and investigate whether people could manipulate scores around the cutoff. Plot the outcome against the assignment score on both sides. Test reasonable bandwidths and functional forms, but do not search repeatedly until a preferred result appears. The What Works Clearinghouse maintains a dedicated review guide for regression discontinuity designs, reflecting the design-specific checks required for credible interpretation.
Matching and weighting
Matching pairs or groups observations with similar measured baseline characteristics. Propensity scores estimate the probability of exposure from observed covariates and can be used for matching, stratification, weighting, or covariate adjustment. They are design tools, not proof of randomisation.
Choose covariates using a causal rationale and information available before exposure. Do not match mechanically on post-intervention variables or consequences of treatment. Inspect common support and report balance after matching or weighting. Austin’s introduction to propensity score methods explains the major approaches and their estimands. Even excellent balance on measured variables cannot remove bias from unmeasured confounding.
Build a defensible comparison and causal model
Draw the assumed causal structure
Before analysing data, list variables that may cause both exposure and outcome. A simple directed acyclic graph can clarify whether a variable is a confounder, mediator, collider, or outcome predictor. For a mentoring programme, prior attainment and baseline wellbeing may affect both enrolment and later stress. Programme attendance may affect social support, which then affects stress; controlling for that mediator would change the question from a total effect to a direct effect.
Record which variables are measured, when they were measured, how they will enter the model, and which important variables are unavailable. This makes the adjustment strategy auditable. It also prevents the common practice of selecting covariates only because they were significant in preliminary tests.
Protect time order and measurement consistency
Confirm that confounders and baseline outcomes precede exposure. Apply the same validated measure, scoring rule, language version, administration mode, and observation window to each condition. A change from paper questionnaires to online collection at the same time as an intervention can imitate an intervention effect. If measurement changed, analyse and discuss that threat rather than hiding it in a limitations paragraph.
Use the reliability and validity guide to justify instruments. Measurement equivalence matters when groups differ by language, culture, age, or administration mode. A scale can be internally consistent while measuring the construct differently across groups.
Plan the sample, data, and ethics
Power depends on the design, number of units, number and spacing of time points, exposure prevalence, clustering, autocorrelation, effect size, and covariate distribution. A standard two-group calculator may be inappropriate. Simulate data under plausible assumptions or use software designed for the chosen estimator. Report the values and code used. The power-analysis guide explains how to distinguish prospective planning from sensitivity analysis.
Account for the level of assignment. If a whole school or clinic receives the programme, individuals within it are not independent assignments. One exposed school and one comparison school provide two organisational units, regardless of the number of participant questionnaires. Use multiple units where possible and model clustering appropriately.
Quasi-experimental research still needs ethical review. Using an existing policy does not remove responsibilities for consent, privacy, proportional data access, secure linkage, distress procedures, and fair interpretation. Clearly distinguish research decisions from service decisions. Researchers should not withhold a beneficial service merely to manufacture a comparison unless this has independent ethical and institutional justification. Consult the site’s psychology dissertation ethics guide when preparing the application.
Create an analysis plan before inspecting effects
A concise preregistration or time-stamped protocol should define the assignment rule, intervention date, inclusion criteria, outcome, comparison, estimand, model, functional form, bandwidth or time window, covariates, clustering, missing-data approach, exclusions, and sensitivity analyses. Separate confirmatory analyses from exploratory work. The preregistration guide provides a practical structure.
| Stage | Primary task | Evidence to retain |
|---|---|---|
| Design | Document exposure assignment and the counterfactual | Policy, dates, threshold rule, site information |
| Preparation | Define variables and sample without viewing preferred effects | Codebook, flow diagram, missingness summary |
| Diagnosis | Check balance, trends, overlap, manipulation, and model assumptions | Plots, balance table, diagnostic statistics |
| Estimation | Fit the prespecified model with appropriate uncertainty | Effect estimate, confidence interval, model output |
| Robustness | Test reasonable alternative specifications and rival explanations | Sensitivity table and negative-control results |
| Reporting | Link each conclusion to its population and assumptions | Transparent methods, limitations, reproducible code |
Diagnose before interpreting
For comparison groups, report pre-intervention balance and overlap. For difference-in-differences, plot pre-trends and consider placebo intervention dates. For interrupted time series, inspect residual autocorrelation and seasonal patterns. For regression discontinuity, show the assignment-score distribution around the threshold and test alternative bandwidths. For matching, show balance before and after adjustment and the number of discarded observations.
Missing data may differ by exposure and outcome. Describe missingness by group and time, then justify complete-case analysis, multiple imputation, weighting, or another approach. Include a sensitivity analysis when conclusions could depend on unverifiable assumptions. The missing-data guide covers prevention, diagnosis, imputation, and transparent reporting.
Estimate effects with uncertainty
Report the effect in interpretable units with a confidence interval, not only a p value. For a wellbeing scale, state the adjusted point difference and uncertainty. For an interrupted series, distinguish immediate level change from change in slope. For regression discontinuity, state that the estimate is local to the cutoff. When observations are clustered or serially correlated, use an uncertainty estimator that reflects that dependence.
Do not automatically interpret statistical adjustment as causal identification. The model estimates an effect only under the design’s assumptions. A regression with many covariates cannot repair a comparison group measured in a different historical period or a threshold that participants manipulated.
Run sensitivity and falsification checks
Robustness analysis should target plausible weaknesses rather than generate a wall of alternative models. Change one defensible decision at a time and explain why it matters.
- Use alternative but reasonable time windows, bandwidths, or trend specifications.
- Test placebo dates or cutoffs where no intervention should occur.
- Examine outcomes that should not respond to the intervention.
- Compare complete-case and justified missing-data analyses.
- Assess whether a modest unmeasured confounder could change the conclusion.
- Repeat the analysis after excluding an unusual site, period, or implementation disruption.
Interpret stability carefully. Similar estimates across models increase confidence only if the models address meaningful alternatives. A result that disappears under a reasonable specification is not a failed dissertation. It is evidence that the conclusion is design-sensitive and should be stated cautiously.
Evaluate bias and report transparently
Review bias from confounding, participant selection, exposure classification, deviations from intended intervention, missing data, outcome measurement, and selective reporting. The peer-reviewed ROBINS-I framework organises these domains for non-randomised intervention studies. It is intended for systematic evaluation, not for converting a complex study into a reassuring score.
Use the TREND reporting statement where relevant. Report eligibility, recruitment, assignment, intervention delivery, participant flow, baseline data, analysis, adverse events, and limitations. Include enough detail for another researcher to understand exactly how the counterfactual was constructed.
Use proportionate causal language
Reserve strong causal wording for designs with well-supported identifying assumptions and successful diagnostic checks. “The programme was associated with a larger reduction in distress than the concurrent comparison condition” may be more accurate than “the programme caused lower distress.” If the design supports a causal estimate near a threshold, state that local scope explicitly.
Discuss external validity separately. A rigorous local effect among students near an eligibility cutoff may not generalise to students far from that cutoff, other universities, or different implementation conditions. Precision about scope strengthens the dissertation.
Common mistakes to avoid
- Calling every pre-post study quasi-experimental: two observations without a credible comparison rarely distinguish intervention effects from history, maturation, regression to the mean, or measurement change.
- Choosing analysis before design: the assignment process should determine the estimator.
- Using one cluster per condition: the intervention becomes inseparable from the school, clinic, teacher, or department.
- Controlling for post-treatment variables: this can block part of the effect or introduce collider bias.
- Claiming matching created randomisation: balance applies only to measured variables included in the design.
- Ignoring implementation: a policy date does not prove that participants received the intended programme.
- Hiding specification sensitivity: readers need to see when conclusions depend on a bandwidth, time window, or trend choice.
Frequently asked questions
Can a quasi-experimental dissertation establish causation?
It can support a causal estimate when the assignment process creates a credible counterfactual and the design-specific assumptions are plausible. Causal strength varies. A well-executed regression discontinuity or controlled interrupted time series may be more persuasive than a weak randomised study with severe attrition, but the assumptions must be explicit.
Is a pretest-post-test design automatically quasi-experimental?
No. A single group measured once before and once after an intervention provides limited protection against history, maturation, testing effects, and regression to the mean. Add a concurrent comparison, more time points, or a defensible assignment rule where feasible.
Which software can analyse quasi-experimental data?
R, Stata, SAS, Python, and some SPSS procedures can estimate relevant regression, mixed, time-series, weighting, and matching models. Software choice is secondary to correct design specification, diagnostics, reproducible code, and an estimator aligned with the assignment mechanism.
How many time points does an interrupted time series need?
There is no universal number. Requirements depend on noise, seasonality, autocorrelation, expected effect shape, and available comparison series. More well-spaced pre-intervention and post-intervention observations generally improve trend estimation. Justify the chosen series through design-specific power or simulation.
Should I use propensity score matching?
Use it only when there is adequate overlap, rich pre-exposure covariate information, and a clear estimand. Other weighting or regression strategies may retain more information. Whichever method is chosen, report balance, discarded cases, assumptions, and sensitivity to unmeasured confounding.
Conclusion
A strong psychology dissertation quasi-experimental design starts with a real assignment process, not a preferred statistical technique. Define the counterfactual, select the design that mirrors exposure, protect time order, measure plausible confounders, diagnose assumptions, estimate uncertainty, and test meaningful alternatives. Report what the evidence supports, including where it remains uncertain.
If you need help turning an existing programme, policy, cutoff, or institutional change into a defensible dissertation plan, Psychology Dissertation Help can support ethical question refinement, design alignment, analysis planning, and reporting. Use support to strengthen your own research decisions and comply with your university’s academic-integrity and authorship requirements.
