Psychology dissertation difference-in-differences analysis can estimate a programme or policy effect when one group is exposed and a credible comparison group is not. The design compares how outcomes change over time, rather than simply comparing post-intervention means. Its credibility depends on a defensible counterfactual, transparent timing, and assumptions that must be argued from evidence rather than accepted from software output.
This guide explains the classic two-group, two-period design and the complications created by multiple periods or staggered implementation. It is intended for psychology dissertations evaluating naturally occurring changes such as a school wellbeing programme, a service redesign, or a regional mental-health policy. It also shows when a broader quasi-experimental design or an interrupted time series would answer the question better.
What difference-in-differences estimates
Difference-in-differences, often shortened to DiD, compares two changes. First, calculate the treated group’s average outcome after the intervention minus its average before the intervention. Second, calculate the same change for the comparison group. Subtract the comparison-group change from the treated-group change. The resulting contrast removes stable differences between groups and common time changes under the design assumptions.
Imagine that a school introduces a peer-support programme while a similar school continues usual provision. Mean wellbeing rises by six points in the programme school and by two points in the comparison school. The simple DiD estimate is four points. It is not six points because two points of improvement also occurred where the programme was absent. That shared change could reflect season, examination timing, or a wider campaign.
The target is usually the average treatment effect on the treated, or ATT: the mean effect among units that received the intervention. The official R did package vignette defines the canonical design as two groups and two periods and states the counterfactual parallel-trends requirement. The treated group’s untreated post-intervention outcome is never observed, so identification rests on assumptions and design evidence.
| Group | Before mean | After mean | Observed change |
|---|---|---|---|
| Programme school | 52 | 58 | +6 |
| Comparison school | 51 | 53 | +2 |
| Difference-in-differences | Treated change minus comparison change | +4 | |
Decide whether psychology dissertation difference-in-differences fits
DiD is most useful when exposure begins at a known time for some units but not for a plausible comparison group. Treatment must have a reasonably clear definition. Examples include a university introducing a counselling triage system, selected clinics adopting a trauma-informed protocol, or one region changing school-screening guidance while another does not.
A before-and-after study with no comparison group is not DiD. Two groups observed only after treatment also cannot reveal differential change. If one population has many observations before and after a sharply timed intervention but no credible control, the site’s interrupted time-series guide describes a more suitable approach. If exposure selection can be explained by measured baseline characteristics without a clear time contrast, propensity-score methods may be considered, subject to stronger exchangeability requirements.
Define the treatment and estimand precisely
Write down who is treated, when treatment starts, whether it can stop, and what “usual care” or no exposure means. A policy announcement, staff training, and actual service delivery may occur on different dates. Choose the date that corresponds to the causal intervention and justify any transition period.
Specify the outcome scale, measurement window, eligible population, and follow-up horizon. “Effect of the programme on anxiety” is too vague. A clearer estimand is the average effect of access to the programme on a validated anxiety score six months after introduction among students enrolled before implementation. This wording distinguishes access from participation and avoids conditioning on a post-treatment choice.
Reconstruct why treatment happened
Investigate the assignment process before fitting a model. If the intervention went first to schools with worsening wellbeing, their outcome path may already differ from other schools. If high-performing clinics volunteered, organisational capacity may affect both adoption and outcomes. Draw a timeline, interview programme staff where appropriate, and examine implementation documents.
Do not adjust mechanically for variables affected by treatment. Post-treatment attendance, engagement, or service use may mediate the effect and controlling for them can introduce bias. A causal diagram can distinguish baseline confounders, mediators, and outcomes. The site’s confounding-variables guide explains this separation.
Understand the identifying assumptions
Parallel trends is about untreated potential outcomes
The central assumption is not that the two groups have equal outcome levels. It is that, without treatment, their average outcomes would have changed similarly over the study period. Different baseline means are compatible with DiD if the untreated trends would have remained parallel. This post-treatment counterfactual cannot be observed directly.
Multiple pre-treatment periods help assess whether observed trends were similar before exposure, but a non-significant pretrend test does not prove parallel trends. Such tests can have low power. A visibly different pretrend is a warning; visually similar histories strengthen but do not establish the assumption. Use theory, assignment knowledge, pre-period plots, and alternative controls together.
Roth and colleagues’ review of modern DiD methods organises recent work around multiple periods and treatment timing, possible parallel-trends violations, and inference. Their synthesis supports treating design assumptions and diagnostics as central rather than presenting the treatment-by-time coefficient as self-validating.
No anticipation and stable measurement
No anticipation means units do not change because they expect treatment before the recorded start. Students may seek help early after a public announcement, or staff may alter referrals during training. If anticipation is plausible, define an earlier exposure date, omit a transition window with justification, or estimate event-time effects that include leads.
Outcome measurement should remain comparable across groups and time. A new questionnaire version, altered response mode, changed eligibility criterion, or revised clinical recording system can mimic an effect. Verify scale invariance and data provenance. If samples are repeated cross-sections rather than the same individuals, changes in population composition also require close attention.
No differential co-intervention or spillover
No other event should affect the treated group at the same time in a way the comparison does not share. A programme introduced alongside extra staff, a funding increase, or new assessment software produces an inseparable package effect unless the estimand is explicitly that package. Document contemporaneous changes and test alternative explanations.
Spillovers weaken the contrast when comparison units indirectly receive treatment. Students may share digital materials across institutions, or clinicians may work at both treated and untreated sites. Define interference pathways before analysis. Consider wider geographic controls, exposure mapping, or a design that estimates spillover effects if contamination is substantial.
| Assumption | Threat in psychology research | Evidence or sensitivity check |
|---|---|---|
| Parallel untreated trends | Early adopters were already improving | Pre-period plot, assignment history, alternative controls |
| No anticipation | Behaviour changed after announcement | Event-time leads and alternative start dates |
| Stable measurement | Questionnaire or sampling changed | Protocol audit and harmonised outcome definition |
| No differential shock | Only treated sites received extra staff | Timeline, negative-control outcomes, adjusted estimand |
| Limited spillover | Control sites accessed programme materials | Contamination assessment and exposure mapping |
Build the dataset and comparison group
DiD can use panel data, where the same units are followed, or repeated cross-sections, where different individuals are sampled at each time. Panel data allow within-unit analysis but face attrition. Repeated cross-sections require the population and sampling process to stay comparable. Stata’s official DiD documentation distinguishes commands for repeated cross-sectional and panel designs and clusters uncertainty at the group level.
Create a data dictionary identifying the unit, group, period, treatment timing, outcome, baseline covariates, and cluster. Verify that treatment does not switch on before the recorded date. Inspect counts by group and period, missingness, scale distributions, duplicates, and impossible dates. Save code that constructs every variable.
Select controls using the assignment mechanism
The comparison group should represent how treated outcomes would have evolved without treatment. Geographic proximity alone is insufficient. Compare service structures, population composition, baseline outcome paths, exposure to other policies, and data collection. Choose controls before viewing post-treatment effects where possible.
Matching or weighting can improve baseline comparability, but it does not automatically create parallel trends. If combining DiD with weighting, estimate weights from pre-treatment information, check overlap and balance, and state the conditional parallel-trends assumption. Do not remove difficult treated units merely to produce an attractive estimate without redefining the target population.
Plan for missing data and attrition
Compare missingness across group and time. Treatment may change questionnaire response or service-record completion, making missingness an outcome-related process. Complete-case analysis can alter group composition and the estimand. Use the psychology dissertation missing-data workflow to justify imputation, weighting, or sensitivity analysis.
For panel data, report retention by treatment group and investigate whether attrition relates to baseline outcomes or predicted change. For repeated cross-sections, report sampling frames and response rates in every period. Do not impute outcomes after dropout without explaining the assumed missingness mechanism.
Estimate the classic two-period model
For one treated group, one comparison group, one pre-period, and one post-period, a linear model includes indicators for treated group, post period, and their interaction. The interaction coefficient is the DiD estimate. Include cluster-appropriate uncertainty and prespecified baseline covariates when they make conditional parallel trends more plausible or improve precision.
The group main effect represents the baseline level difference. The time main effect represents the comparison group’s change. The interaction represents the treated group’s additional change beyond that common time effect. Report the model formula, coding, reference groups, coefficient, confidence interval, and outcome units.
For binary or count outcomes, a nonlinear model’s interaction coefficient may not equal the DiD effect on the probability or mean scale. Define the estimand first and present marginal effects or standardised predictions that match it. A linear probability model can be transparent for an average probability difference, but predicted values and heteroskedasticity still require attention.
Use uncertainty that matches assignment
Errors are often correlated within schools, clinics, regions, or individuals. Cluster at the level at which treatment is assigned when appropriate, not simply at the row level. With few treated or total clusters, conventional cluster-robust standard errors can be unreliable. Consider randomisation inference when assignment supports it, wild-cluster bootstrap methods, or small-sample corrections, and interpret limited information honestly.
More people within one treated school do not create many independent treatment assignments. Sample-size planning should focus on groups, time points, plausible within-cluster correlation, outcome variation, and effect precision. Link this justification to the power-analysis guide rather than applying a generic participant rule.
Handle multiple periods and staggered treatment safely
Many real programmes start at different times across sites. In that setting, a conventional two-way fixed-effects regression can compare newly treated units with units treated earlier. When effects vary across cohorts or evolve after exposure, those comparisons can produce hard-to-interpret weighted averages and even misleading signs.
Callaway and Sant’Anna’s multiple-period DiD framework identifies group-time average treatment effects and permits never-treated or not-yet-treated comparison groups under corresponding parallel-trends assumptions. Its estimands can be aggregated by cohort, calendar time, or exposure duration. The maintained R did vignette explains these quantities and their implementation.
Sun and Abraham show in their event-study paper that conventional lead-and-lag coefficients can be contaminated under heterogeneous treatment effects. Use an estimator designed for staggered adoption, report which units serve as controls at each time, and explain the aggregation weights. Stata’s heterogeneous DiD tools similarly estimate cohort-specific and time-specific effects.
Use event studies as diagnostics and effect summaries
An event study indexes time relative to treatment. Pre-treatment estimates can reveal differential trends or anticipation, while post-treatment estimates show how effects evolve. Choose one pre-period as the reference and provide simultaneous or multiple-testing-aware uncertainty where the method supports it.
Do not interpret a failed pretrend significance test as proof of validity. Plot point estimates and intervals, state the available pre-period length, and assess whether substantively important violations remain compatible with the data. Avoid changing the event window merely because a different choice removes an inconvenient pretrend.
| Design | Main estimand or output | Key caution |
|---|---|---|
| Two groups, two periods | Single ATT contrast | Parallel trends remains unobserved after treatment |
| Multiple common periods | Dynamic or pooled effects | Serial correlation and outcome dynamics |
| Staggered adoption | Cohort-time ATT estimates | Avoid already-treated units as invalid controls |
| Event study | Effects by time from treatment | Use heterogeneity-robust estimation |
| Repeated cross-sections | Population mean changes | Population composition must remain comparable |
Run falsification and sensitivity analyses
Pre-specify analyses that could expose a weak design. Use placebo treatment dates before implementation, outcomes that should not respond, alternative credible comparison groups, and reasonable time windows. Test sensitivity to group-specific trends cautiously because adding them changes identifying assumptions and can absorb genuine dynamic effects.
Examine influential clusters and whether conclusions depend on a single treated site. Report estimates with and without major data-quality anomalies, but do not delete valid clusters merely because results change. If pretrends are concerning, quantify how large a post-treatment deviation would remain after allowing plausible violations rather than claiming the issue is solved.
Assess selection mechanisms, measurement changes, anticipation, spillovers, and co-interventions separately. A battery of similar regression specifications is not a substitute for design-specific falsification. Sensitivity results should modify the strength of the conclusion, not appear as decorative appendices.
Psychology dissertation difference-in-differences example
Suppose six universities introduce a same-day counselling triage service in September, while eight comparable universities retain usual booking. Monthly distress scores are available from routine anonymous surveys for 18 months before and 12 months after implementation. The target is the effect of service availability on mean distress among enrolled students during the first post-implementation year.
The researcher documents that treatment universities were selected because of waiting times, not because distress trends were rising. Pre-period plots show similar average paths, although one treated university experienced an unusual strike. Questionnaire wording and sampling remain stable, and no comparison university adopts a similar system during follow-up.
A common-adoption DiD model includes university and month effects, seasonal controls chosen in advance, and standard errors clustered by university. Because there are only 14 clusters, the researcher uses an appropriate small-cluster inference method and emphasises wide intervals. An event-study plot includes all pre-period estimates and the transition month.
The adjusted post-period estimate is a 1.8-point reduction in distress relative to the comparison trend, with a 95% confidence interval from a 3.2-point reduction to a 0.4-point reduction. The dissertation describes this as evidence consistent with a beneficial effect under parallel trends, no differential co-intervention, stable measurement, and limited spillover. It does not claim the service reduced every student’s distress or explain the mechanism without mediation evidence.
Report the method and results transparently
In the method chapter, name the unit of treatment and analysis, eligibility criteria, group-selection process, treatment date, outcome, estimand, comparison strategy, pre/post windows, covariates, missing-data approach, and inference method. Provide the model formula, software and package versions, and a diagram or table of timing. State every identifying assumption in substantive language.
In the results, begin with group-period counts and descriptive outcome trajectories. Show an unadjusted plot, the main estimate with its confidence interval, pre-period event-study estimates where available, and sensitivity analyses. If treatment timing varies, report cohort-time estimates or a clearly justified aggregation instead of only a conventional fixed-effects coefficient.
Archive analysis code, a data dictionary, and a decision log where ethically permissible. Protect participant confidentiality, especially when site-level plots could reveal small groups. Label exploratory model changes and avoid causal wording stronger than the design warrants.
Common mistakes to avoid
- Calling a one-group before-and-after comparison difference-in-differences.
- Choosing a control because baseline means match while ignoring pre-period trends.
- Treating a non-significant pretrend test as proof of parallel trends.
- Ignoring anticipation, spillovers, policy bundles, or changing measurement.
- Using participant-level standard errors when treatment is assigned by site.
- Applying conventional two-way fixed effects to staggered adoption without assessing effect heterogeneity.
- Adjusting for post-treatment mediators or colliders.
- Reporting only one coefficient without trajectories, diagnostics, or sensitivity analyses.
- Interpreting a nonlinear interaction as an effect on the probability scale.
- Claiming that DiD automatically removes all confounding.
Frequently asked questions
What data are needed for psychology dissertation difference-in-differences?
You need treated and comparison units observed before and after treatment, plus clear treatment timing and a consistently measured outcome. Multiple pre-periods materially improve design assessment, although they cannot prove the post-treatment parallel-trends assumption.
Does difference-in-differences require equal baseline means?
No. Stable level differences are allowed. The crucial requirement is a credible claim that untreated outcome changes would have been parallel, possibly conditional on prespecified baseline covariates.
Can I run a basic DiD in SPSS?
A two-group, two-period linear interaction can be estimated in general regression software. Complex clustering, staggered adoption, modern cohort-time estimators, and simultaneous event-study inference may require R, Stata, or specialist software.
How many pre-treatment periods should I include?
There is no universal number. Use all comparable periods relevant to the assignment and outcome process, while avoiding measurement regimes that are no longer comparable. More pre-periods help reveal dynamics, seasonality, and plausible violations.
Should I add group-specific time trends?
Only with a substantive and pre-specified rationale. Group trends change the estimand and may extrapolate strongly or absorb real treatment dynamics. Present them as sensitivity analyses when appropriate, not an automatic repair.
What if treatment starts at different times?
Use a staggered-adoption method that estimates cohort-time effects with valid never-treated or not-yet-treated controls. Do not assume a conventional two-way fixed-effects event study remains interpretable when effects vary.
Can difference-in-differences prove causality?
It can support a causal estimate when the counterfactual assumptions, timing, measurement, interference conditions, and inference are defensible. The design cannot verify all assumptions from observed data, so conclusions must remain conditional and evidence-based.
Conclusion
A strong psychology dissertation difference-in-differences analysis starts with treatment assignment, timing, and a credible comparison group. The calculation is simple, but the design is not. Parallel trends, no anticipation, stable measurement, limited spillover, suitable controls, clustered inference, and modern handling of staggered adoption determine whether the estimate is meaningful.
If you need methodological guidance, seek ethical support that helps you understand and defend your own design. An adviser can review the causal question, timeline, code, diagnostics, and reporting while you retain responsibility for the data, analysis decisions, and dissertation authorship.
