Psychology dissertation interrupted time series research tests whether an outcome changed after a clearly dated intervention while accounting for its prior trajectory. It is useful when random assignment is impossible but repeated observations exist before and after a policy, programme, service change, or naturally occurring event. This guide explains how to define the counterfactual, plan segmented regression, diagnose time-series problems, and report conclusions without overstating causality.
An interrupted time series is more than a pretest and post-test. Its strength comes from showing what the outcome was doing before the interruption, projecting that pattern forward, and comparing the observed post-intervention series with that projection. For psychology dissertations, suitable outcomes include weekly counselling referrals, monthly distress scores from repeated service samples, daily crisis contacts, or aggregated wellbeing indicators collected under a stable protocol.
What is a psychology dissertation interrupted time series?
An interrupted time series, often abbreviated ITS, is a longitudinal quasi-experimental design. It uses multiple observations before and after a defined intervention date to estimate whether the outcome’s level, slope, or both departed from the pattern expected without the intervention. The expected post-intervention path is the counterfactual trend.
The design is especially relevant when an institution introduces a change for everyone at once. A university may launch an out-of-hours mental health service, a school system may change its behaviour policy, or a clinic may introduce routine anxiety screening. Randomisation may be unavailable because implementation has already occurred. A well-designed ITS can still offer stronger evidence than a simple before-after comparison.
The Lopez Bernal, Cummins, and Gasparrini tutorial describes segmented regression as a practical ITS approach and highlights impact models, seasonality, autocorrelation, and time-varying confounding. The design belongs within the wider family explained in our quasi-experimental design guide, but it requires its own time-specific planning and diagnostics.
| Design | Measurements | Main comparison | Principal limitation |
|---|---|---|---|
| One-group pretest-post-test | Usually one before and one after | Mean before versus mean after | Cannot separate intervention from prior trend or history |
| Interrupted time series | Multiple points before and after | Observed post-period versus projected pre-period trend | Concurrent events can mimic the intervention effect |
| Controlled interrupted time series | Repeated treated and comparison series | Relative change across time and series | Control must share relevant background influences |
| Longitudinal cohort | Repeated observations on people | Change and predictors within or between people | Does not automatically identify a dated intervention effect |
When is interrupted time series suitable?
Use ITS when the intervention has a defensible start date, the outcome is recorded consistently across time, and enough observations surround the interruption to estimate a baseline pattern and a post-period departure. The question should concern a population-level or service-level change rather than only individual development.
A feasible psychology example is a university that introduced same-day counselling appointments on 1 September. Weekly waiting time, missed appointments, and symptom severity at intake were recorded under the same definitions for eighteen months before and twelve months after implementation. The series provides a visible pre-intervention pattern, a clear interruption, and enough post-period information to examine immediate and gradual change.
By contrast, anxiety measured once in April and once in October is not a time series. Nor is a dataset suitable simply because it contains dates. If eligibility rules, instruments, recruitment, or recording systems changed at the intervention point, observed discontinuity may reflect measurement rather than psychological change.
Define the intervention precisely
Record what changed, who was exposed, the implementation date, rollout duration, adherence, and any transition period. A formal announcement date may differ from the date participants first experienced the programme. If implementation was gradual, a single immediate step may be unrealistic. You may need a phase-in term, a delayed onset, or a sensitivity analysis that shifts the interruption date.
Do not select the interruption date by searching for the largest visible change. That turns a confirmatory evaluation into an undisclosed change-point search and makes conventional uncertainty misleading. If the date is genuinely unknown, treat estimation of a change point as a different, explicitly exploratory problem.
Write an estimable question
A strong question names the unit, outcome, frequency, intervention, and expected response. For example: “Did the introduction of 24-hour crisis text support change the immediate level or monthly trend in recorded self-harm presentations among students, relative to the pre-implementation trajectory?” This distinguishes level change from slope change and defines the time scale.
Clarify whether the estimand is the immediate change at implementation, the change in trend, the difference between observed and expected outcome at a meaningful follow-up point, or an accumulated effect. The methodological framework by Lopez Bernal and colleagues emphasises defining both the counterfactual and the anticipated impact model.
Plan the time series before analysing it
Choose the unit and outcome
The analytical row might represent a day, week, month, school term, or survey wave. Choose a frequency that matches how quickly the intervention could plausibly affect the outcome and how reliably data are collected. Daily data may be noisy and strongly seasonal. Annual data may provide too few points. Aggregation can improve stability but can also hide short-lived effects.
Use the same outcome definition throughout. If symptom scores come from different participants each month, describe the design as repeated cross-sectional rather than individual longitudinal follow-up. Adjusting for changes in sample composition may be necessary. If the same people contribute repeatedly, dependence exists both across time and within participants, which may require a richer model than a simple aggregate series.
Decide how many time points are defensible
There is no universal minimum that guarantees a credible ITS. Required observations depend on variability, seasonality, autocorrelation, effect shape, outcome family, and the number of parameters. Three points before and after may meet a minimal label in some review rules, yet usually provide weak trend estimation and little diagnostic power.
Plan with simulation where possible. Generate series under plausible baseline slopes, residual variation, autocorrelation, seasonal cycles, intervention effects, and missing points. Fit the intended model repeatedly and examine bias, interval coverage, and power. This is more informative than applying an ordinary two-group calculator. Our psychology dissertation power analysis guide explains simulation-based planning.
Protect measurement consistency
Create a timeline of instrument versions, scoring rules, service access, staff roles, software changes, recruitment channels, and reporting requirements. A new electronic record introduced in the intervention month can create an apparent step even if behaviour did not change. Document such changes and, where feasible, calibrate overlapping measures or restrict analysis to consistently observed outcomes.
Reliability and construct validity remain essential. A monthly average score is not automatically comparable if language versions, administration modes, or respondent populations changed. Consult the reliability and validity guide before treating a stable variable name as stable measurement.
Specify the impact model
A basic segmented regression commonly includes elapsed time, an indicator for the post-intervention period, and time since intervention. The intercept describes the estimated baseline level. The time coefficient describes the pre-intervention slope. The post indicator estimates an immediate level change, while time since intervention estimates a change in slope.
These coefficients answer different questions. A student wellbeing programme might reduce distress immediately but leave the subsequent slope unchanged. Another intervention might show no immediate jump yet produce gradual monthly improvement. Reporting only “the intervention coefficient” obscures the assumed response.
| Impact shape | Model representation | Psychology example | Interpretive caution |
|---|---|---|---|
| Immediate, sustained | Level change | New referral rule immediately increases recorded assessments | Could reflect recording rather than behaviour |
| Gradual | Slope change | Peer-support participation slowly changes wellbeing trend | Long follow-up is needed to observe the trajectory |
| Immediate and gradual | Level and slope changes | Crisis service reduces waiting immediately and keeps improving | Parameters may be correlated in short series |
| Delayed | Lagged interruption | Staff training affects outcomes after implementation settles | Delay should be justified before inspecting results |
| Temporary | Pulse or decaying effect | Awareness campaign briefly increases help-seeking | A permanent step model would misrepresent the effect |
Centre time for useful interpretation
Code time so that the intervention boundary or another meaningful point has a clear value. Coding post-intervention time as zero before the interruption and then 0, 1, 2, and onward after it makes the slope-change term interpretable. Check the exact convention used by your software, especially whether the intervention observation is treated as pre or post.
For a continuous outcome, linear regression may be adequate if residual assumptions are reasonable. Counts may require Poisson or negative binomial models, rates may require an offset, and proportions may need binomial methods. A bounded mean score can sometimes use Gaussian errors, but that decision should follow the distribution and inferential target rather than habit.
Handle autocorrelation, seasonality, and nonlinearity
Diagnose autocorrelation
Adjacent observations often resemble one another. Weekly counselling demand this week may be close to last week’s demand even after accounting for trend and season. If residual autocorrelation is ignored, standard errors and tests may be inaccurate. Plot residuals, inspect the autocorrelation and partial autocorrelation functions, and use diagnostics suited to the fitted model.
Possible responses include generalized least squares with an autoregressive error structure, Newey-West style standard errors, autoregressive integrated models, or other time-series methods. The choice depends on the outcome, spacing, number of points, and residual pattern. Do not add many lags until significance appears. Pre-specify a plausible range and report sensitivity.
Model seasonality and secular change
Psychological outcomes can vary by academic term, examination period, holidays, weather, or service schedules. If the intervention occurs near a recurring seasonal change, an unadjusted step can be misleading. Use calendar indicators, harmonic sine and cosine terms, splines, or another parsimonious seasonal representation supported by the measurement frequency and series length.
A nonlinear baseline trend also matters. Extrapolating a straight line beyond the observed range may create an implausible counterfactual. Plot the pre-period carefully and compare a limited set of scientifically justified forms. The intended model should be selected from subject knowledge and pre-intervention fit, not from which post-intervention effect looks most favourable.
Strengthen inference with a control series
A single ITS controls for stable population characteristics and the established outcome trajectory, but it remains vulnerable to another event occurring at the same time. A controlled ITS adds an unexposed series that shares relevant seasonal patterns, measurement processes, and historical shocks. The key contrast becomes whether the treated series changed more than the comparison series.
The International Journal of Epidemiology article on ITS controls explains how a control can reduce confounding by simultaneous events while introducing its own selection requirements. A nearby university may be a poor control if it changed counselling capacity at the same time or serves a very different population.
Choose controls by mechanism
Select a control because it would have experienced the important background influences but not the intervention. Compare pre-intervention levels and trends, data quality, outcome definition, seasonality, and exposure to other policies. Similar averages alone are insufficient. A control with a different pre-trend may still be usable in a correctly specified comparative model, but the required extrapolation becomes less convincing.
Negative-control outcomes can also help. If a counselling access reform is expected to affect waiting time but not unrelated library incidents, a discontinuity in both series may signal a broader recording change. Such checks do not prove validity, but they can expose rival explanations.
Address bias, missing data, and ethics
The Cochrane Handbook risk-of-bias chapter treats ITS as a non-randomized intervention design with repeated aggregate measurements before and after implementation. Key threats include confounding, intervention classification, deviations from intended intervention, missing data, outcome measurement, and selective reporting.
Create a dated catalogue of co-interventions and shocks. Examples include examination-policy changes, industrial action, an epidemic, a publicity campaign, new staff, or revised eligibility rules. Explain whether each event could affect the outcome and whether a control, covariate, exclusion window, or sensitivity analysis addresses it. Some threats cannot be adjusted away and must constrain the conclusion.
Do not hide missing time points
Missingness may reflect system outages, incomplete surveys, clinic closure, or selective nonresponse. Plot denominators and missingness across time. Imputation should respect temporal dependence, intervention status, and uncertainty. Interpolating a missing outcome from neighbouring points and treating it as observed will understate uncertainty.
For individual-level repeated data, dropout may depend on prior symptoms or exposure. Aggregating available cases can alter the population composition over time. Use the missing-data guide to distinguish prevention, description, modelling, and sensitivity analysis.
Protect people in administrative data
Existing records are not ethically unrestricted. Seek the required institutional review, data-owner permission, and lawful basis. Minimise identifiers, use secure linkage, define retention, and avoid reporting small cells that could expose participants. Where intervention rollout affected access to care, discuss fairness and unintended harms rather than measuring only average improvement.
Run the analysis in a transparent sequence
| Stage | Action | Evidence to retain |
|---|---|---|
| Design | Fix the intervention date, outcome, unit, estimand, and impact shape | Protocol, implementation record, causal diagram |
| Preparation | Build a complete ordered series and audit definitions | Codebook, missingness plot, denominator checks |
| Description | Plot raw and seasonally annotated outcomes | Graph with intervention marker and key events |
| Primary model | Fit the pre-specified segmented model | Formula, software, coefficients, intervals |
| Diagnostics | Check residuals, autocorrelation, fit, and influential points | Residual plots and diagnostic statistics |
| Sensitivity | Test plausible dates, trends, error structures, and controls | Comparison table with unchanged and changed conclusions |
| Reporting | Present observed and counterfactual trajectories | Effect at meaningful times and reproducible code |
Fit diagnostics before interpreting effects
Inspect residuals for remaining trend, changing variance, seasonality, outliers, and serial dependence. Check whether a few extreme dates drive the result. Investigate those observations using records, but do not delete them merely because the coefficient weakens. A documented leave-one-out or robust sensitivity analysis is more transparent.
Compare fitted values with the raw series and display the counterfactual post-intervention trajectory. Confidence intervals should reflect the model and extend around relevant effect estimates. The Penfold and Zhang tutorial explains the role of segmented regression in evaluating policies and programmes when randomisation is infeasible.
Use software without surrendering judgement
R, Stata, Python, SAS, and other platforms can fit segmented regression and time-series error structures. Report the software and version, data frequency, outcome family, time coding, intervention coding, seasonal terms, lag structure, variance estimator, and estimation method. Save code and outputs so every reported estimate can be reproduced.
Software defaults are not design decisions. Ordinary least squares may produce a coefficient even when errors are serially correlated. Automated model selection may prefer a complex pattern that does not represent the intervention mechanism. Use the general regression analysis guide alongside time-series-specific checks.
Interpret level and slope changes responsibly
Suppose weekly mean distress was rising by 0.10 points before a support programme. The fitted immediate change is -1.20 points and the slope change is -0.08 points per week. The post-intervention slope is therefore 0.02, not -0.08. Interpret the immediate contrast, the new slope, and the estimated difference from the counterfactual at a pre-specified follow-up week.
Report effect sizes in outcome units with confidence intervals. If a standardized scale is used, explain what its points mean and whether the estimated difference is practically important. Statistical evidence of a slope change does not establish clinical benefit, and absence of significance does not demonstrate equivalence.
Causal language must follow the design. A controlled series with aligned measurement, stable pre-trends, no plausible concurrent shocks, and robust sensitivity analyses may support a cautious intervention-effect claim. A single noisy series with a coincident policy change should be described as evidence of a temporal association.
Psychology dissertation interrupted time series example
Imagine a university introduced a stepped-care triage pathway at the beginning of an academic term. The dissertation examines weekly median days from referral to first appointment over 72 weeks before and 48 weeks after implementation. A comparison university continued its usual process and used the same waiting-time definition.
The primary model estimates immediate level and slope changes in the treated series relative to the control. Term indicators account for recurring academic cycles, and an autoregressive error structure addresses residual dependence. The protocol defines a four-week implementation phase, excludes no outcomes, and specifies sensitivity analyses using alternative phase lengths and a count of completed appointments.
The results graph shows observed data, fitted trajectories, the interruption, and the counterfactual. The student reports the immediate relative change, the difference at 26 weeks, confidence intervals, missing weeks, control-series fit, and concurrent events. The conclusion concerns access under the observed university conditions, not all psychological services worldwide.
Common interrupted time series mistakes
- Calling one pretest and one post-test an interrupted time series.
- Choosing the interruption date after seeing the outcome pattern.
- Assuming every intervention has an immediate permanent effect.
- Ignoring seasonality, autocorrelation, nonlinear trend, or changing variance.
- Using an outcome whose definition changed at implementation.
- Treating a comparison series as valid because its average is similar.
- Overfitting many lags, breakpoints, or polynomial terms.
- Reporting only a p-value without observed and counterfactual trajectories.
- Claiming causality without examining concurrent events and co-interventions.
Frequently asked questions
How many observations does an interrupted time series need?
No fixed number guarantees adequacy. More points improve trend, seasonality, and autocorrelation estimation, but required length depends on noise, effect shape, frequency, and model complexity. Justify the series using design knowledge and simulation rather than a universal rule.
Is interrupted time series the same as pretest-post-test?
No. A pretest-post-test comparison uses one or a few measurements and cannot estimate the established trajectory. ITS uses repeated observations before and after a defined interruption to compare observed outcomes with a projected counterfactual trend.
What is the difference between level change and slope change?
A level change is an immediate step at the intervention point. A slope change is a change in the rate of increase or decrease after implementation. A model can include either or both when justified by the expected impact.
Does interrupted time series prove causation?
Not automatically. Credibility depends on a fixed intervention date, stable measurement, a plausible counterfactual, appropriate time-series modelling, and the absence or control of competing events. A suitable control series can strengthen inference.
Can individual symptom scores be analysed with ITS?
Yes, but the data structure matters. A single person’s intensive series differs from aggregate repeated samples or many participants measured repeatedly. Dependence, missingness, generalisability, and the analytical level must match the design.
Should the intervention month be excluded?
Only with a pre-specified implementation rationale. You may model a transition period, delayed effect, or phase-in. Excluding points because they weaken the result is inappropriate. Report and test plausible coding choices transparently.
What should an ITS results figure show?
Show the raw ordered outcomes, intervention marker, fitted pre- and post-period trajectories, the projected counterfactual, and uncertainty where useful. Annotate major pre-specified events without crowding the graph.
Conclusion
A strong psychology dissertation interrupted time series begins with a real intervention date and consistent repeated outcome data. It defines the counterfactual and impact shape before effect inspection, distinguishes level from slope change, models temporal dependence, and treats concurrent events as central validity questions.
Transparent work shows the raw series, assumptions, diagnostics, sensitivity analyses, and effects at meaningful times. If you want ethical support checking an ITS design, model specification, code, or reporting plan, Psychology Dissertation Help can provide structured feedback while leaving all decisions, analysis, and academic authorship under your control.
