2027 Planning an Interrupted Time Series With Seasonal Patterns for MPH and MS Students
Co-Founder and Chief Data Scientist
A vaccination campaign, air-quality rule, or clinic policy may begin at a known date, but the outcome it targets may already rise and fall with the seasons. If students mistake that cycle for an intervention effect, their evaluation can mislead decision-makers.
This guide helps MPH and MS students decide whether a seasonal interrupted time series is feasible, assemble the right data, choose and check an analysis, and explain what the results can - and cannot - show.
Key Things You Should Know
- A seasonal interrupted time series compares outcomes before and after an intervention while accounting for recurring patterns; monthly data have a 12-observation annual cycle.
- Plan the intervention date, outcome denominator, expected delay, comparison series, and seasonal adjustment before fitting a model; several complete seasonal cycles on each side are preferable when feasible.
- Report both immediate and gradual changes with uncertainty, then test whether autocorrelation, competing events, or alternative seasonal specifications change the conclusion.
What is an interrupted time series analysis with seasonal patterns?
An interrupted time series (ITS) follows an outcome at regular intervals before and after a defined event. It estimates whether the outcome changed immediately, whether its subsequent trend changed, or both, relative to a projected continuation of the pre-intervention pattern. A seasonal pattern repeats at a calendar frequency - for example, a yearly cycle in monthly respiratory-illness visits.
Consider a hypothetical county that introduces a school-based vaccination program in September and tracks monthly visits for a specified illness. Visits might increase each winter regardless of the program. A comparison of winter with summer would therefore be misleading; the analysis must distinguish the intervention signal from the expected seasonal cycle. This is a population-level evaluation, not a comparison of the same individuals before and after treatment.
Students learning the distinction between descriptive public health trends and causal evaluation may first encounter it in undergraduate study; an online bachelor's public health education guide can help identify foundational pathways, but a credible ITS still requires supervised methods practice and suitable data.
When should MPH and MS students use an interrupted time series design?
Use ITS when a policy or program has a reasonably identifiable start, repeated outcome measurements exist on both sides, and random assignment is unavailable or inappropriate. It is especially useful for evaluating system-wide changes, such as a reporting rule or service-delivery policy. Define the target population, outcome, intervention, and expected effect before choosing the model: an immediate reporting change calls for a different interpretation than a program whose uptake builds over months.
ITS is less persuasive when the intervention date is vague, measurement changes at the same time, another major event coincides with implementation, or too little pre-intervention history exists to establish the usual seasonal pattern. An unaffected comparison series - or a comparable jurisdiction without the intervention - can strengthen an evaluation, provided its pre-intervention behavior and measurement practices are credible. It does not automatically remove confounding.
For an MPH applied project, negotiate a question the partner can use without promising a causal answer. An MS thesis may justify more extensive model comparisons. Neither degree label substitutes for data access or analytic supervision; students still exploring alternative entry routes to a public health master's should distinguish admission preparation from readiness to conduct this analysis.
How do you account for seasonality in interrupted time series models?
Specify a seasonal approach based on the measurement interval and the amount of history available. Calendar indicators allow each month or quarter to have its own typical level; harmonic terms, such as sine and cosine functions, represent a smoother repeating cycle. Seasonal time-series models can instead model recurring dependence directly. These approaches are alternatives, not adjustments to stack indiscriminately into a short series.
First check whether annual seasonality is plausible for the outcome and location. Monthly observations support a possible 12-period annual cycle; weekly data present additional complications because calendar years do not contain an identical number of complete weeks. Holidays, school terms, weather, and changes in care-seeking may also matter. Avoid choosing a seasonal specification solely because it makes the intervention estimate more favorable.
A practical analysis plan names a primary seasonal model and a plausible alternative before inspecting results. Students who need deeper preparation in regression and time-series methods can compare graduate biostatistics degree options for career advancement with a targeted methods course; an additional degree is not inherently necessary for one well-supervised project.
Use observations at consistent intervals, with the same outcome definition, geographic coverage, and reporting process throughout. Record the intervention announcement, implementation, and likely uptake dates separately if they differ. A long pre-period helps establish baseline trend and recurring seasonality; a long post-period helps distinguish a lasting change from a brief fluctuation. There is no universal minimum number of observations that makes a seasonal ITS valid: adequacy depends on cycle length, noise, effect timing, and model complexity.
For counts, establish whether population size or service volume changed; a rate or an appropriate model offset may be more informative than raw counts. Preserve a data dictionary covering denominators, missing periods, revisions, coding changes, and potential competing events. Ask a faculty advisor whether the available series can support the planned seasonal terms and effect estimates before securing a project deadline.
For restricted health-system or agency records, obtain the applicable data-use permission and determine whether institutional review, privacy safeguards, or community approval is needed. A public dashboard and a file containing identifiable patient records require different handling. If acquisition costs or limited staff time threaten feasibility, choose a smaller authorized dataset rather than assuming that enrollment in one of the budget-friendly online public health degrees will solve project-data constraints.
Plot the unmodeled outcome in time order, mark the intervention date, and identify any implementation or transition period. Display missing intervals rather than connecting across them without explanation. A separate plot by calendar month - or a year-over-year display - can reveal whether peaks recur and whether the pre- and post-intervention seasons differ. For rates, label the denominator and units clearly.
Then show fitted values and, where useful, a projected no-intervention counterfactual from the pre-intervention pattern. Label that projection as model-based, not observed fact. A second panel of residuals can expose recurring peaks that the fitted model missed. Keep the raw series visible: an attractive fitted line should not conceal an outbreak, reporting disruption, or sudden change in population coverage.
Visualization is a learnable skill rather than a credential-specific one. Students comparing accelerated courses in public health bachelor's study should look for enough time to practice data cleaning and graph interpretation; a compressed course schedule alone does not establish readiness for a graduate-level ITS.
Segmented regression is often an interpretable starting point. It includes time, an indicator for the post-intervention period, and time since intervention, alongside chosen seasonal terms. Depending on the outcome, students may use linear regression for an approximately continuous measure, or a count model with an exposure offset for event counts. The error structure needs separate attention: a count model does not, by itself, resolve serial correlation.
Other choices include regression with autocorrelated errors, generalized least squares, and seasonal ARIMA models with intervention terms. The best choice depends on the outcome distribution, series length, seasonal pattern, and whether the aim is an understandable policy contrast or a more detailed representation of time dependence.
The comparison below shows what each approach primarily contributes; it is not a ranking of model quality.
| Approach | Useful feature | Important limitation |
| Segmented regression with calendar or harmonic terms | Directly describes level and slope changes while representing seasonality | Requires checks for correlated errors and model fit |
| Regression with autocorrelated errors | Retains interpretable intervention terms while modeling serial dependence | Results depend on an appropriate error specification |
| Seasonal ARIMA with intervention terms | Can represent complex recurring and short-term dependence | Needs sufficient history and careful interpretation of intervention effects |
Methods-intensive research may justify advanced coursework or doctoral supervision, but financing that training is a separate decision. An epidemiology doctoral tuition assistance guide is relevant to a longer research pathway, not a prerequisite for fitting a student ITS model.
Autocorrelation means neighboring observations remain related after accounting for modeled trends and seasons. If ignored, it can make uncertainty estimates misleading and lead students to overstate precision. Inspect residual plots and autocorrelation across relevant lags, then compare defensible error specifications. Robust standard errors may help with some forms of dependence, but their performance in a short series is not guaranteed.
Nonstationarity means properties such as the underlying level, trend, or variance change over time. An ITS deliberately allows a change at the intervention; the problem is unmodeled drift or changing seasonal behavior elsewhere. Inspect the pre-period for trend changes, consider whether transformations or differencing are warranted, and check that those choices do not erase the effect the study was designed to estimate. A diagnostic test alone cannot settle the substantive question.
These checks require time and feedback, whether a student studies on campus or online. When comparing public health degrees by price and completion time, allow for the supervision and revision cycles that a credible time-series project needs; the shortest advertised timeline may not fit the available data.
A level change is an estimated immediate shift at the specified intervention point. A trend change is a change in the outcome's rate of increase or decrease per observation interval after that point. State the units: a change in monthly visits is not the same as a change in visits per population served. For a delayed program, an immediate-shift term may be less meaningful than a ramp-up or prespecified lag.
Interpret the intervention estimate against the modeled counterfactual and report uncertainty, not just whether a test crosses a significance threshold. A declining post-intervention line does not necessarily indicate success if the outcome was already declining faster before implementation. Conversely, an outcome can continue rising yet rise more slowly than the pre-intervention projection. Neither pattern alone proves causation, particularly when another policy or outbreak begins near the intervention date.
MPH students should translate the estimate into a decision the practice partner actually faces; MS students may also examine sensitivity to alternative effect shapes. Students considering public health master's options after a weak GPA should remember that admission history and the strength of an eventual causal interpretation are separate questions.
The most consequential mistakes often occur before modeling. Students may choose an intervention date after seeing the graph, treat a gradual rollout as an overnight change, compare unlike calendar seasons, or overlook a simultaneous change in coding or service access. Short series can also encourage overly complex seasonal models that fit historical noise rather than a stable pattern.
Watch for an outcome denominator that changes with testing volume, enrollment, or population coverage. Do not silently fill missing months with zero: "no event occurred" and "no report was received" mean different things. If implementation could affect reporting itself, a measured increase may reflect detection rather than a rise in underlying disease.
Plan these checks as part of the project, not as repairs after an appealing result appears. Coursework pace matters here: accelerated courses in public health bachelor's study can introduce the concepts, but a rushed analysis should not substitute for a documented intervention history and advisor review.
How can students report and validate interrupted time series findings?
Report the population, data source, observation interval, intervention and transition dates, outcome definition, denominator, number and range of pre- and post-intervention observations, seasonal specification, effect terms, and approach to serial correlation. Provide raw-data and fitted-series figures, effect estimates with uncertainty intervals, and a plain-language account of competing explanations. Document exclusions, missing periods, and any deviations from the original plan.
Validation should test whether the substantive conclusion survives plausible choices. Compare seasonal specifications, inspect residuals, vary a defensible implementation lag, and assess the influence of exceptional periods. If an appropriate comparison series exists, explain why it is comparable and whether it shows a similar change. A negative-control outcome or a prespecified placebo date may add perspective, but neither can rescue poor measurement or an implausible counterfactual.
Keep analysis code and a deidentified or synthetic demonstration dataset when sharing real records is prohibited; obtain permission before publishing partner figures. Ask the advisor and data steward what can appear in a thesis, practicum report, or portfolio. Students later considering doctoral research may consult a public health doctorate application accessibility resource, but the immediate standard for this project is reproducible reasoning and appropriately limited claims - not a future credential.
Other Things You Should Know About Public Health
Yes, if rollout timing is documented well enough to define the exposure. A single start-date indicator may misrepresent staged implementation; discuss a ramp-up or another prespecified exposure definition with your advisor.
Only if monthly aggregation preserves the decision-relevant outcome and enough observations remain. Aggregation may reduce noise but can hide short-lived effects and change how implementation timing is represented.
It can generate model-based projections, but uncertainty grows beyond observed data. Do not present a long-range projection as a measured or guaranteed program benefit.
No single package is required. Choose software that can fit the proposed model, check residuals, and reproduce figures and estimates; confirm that your advisor can review the workflow.
- The use of controls in interrupted time series studies of public health interventions - PubMed https://pubmed.ncbi.nlm.nih.gov/29982445/
- Time Series Analysis: The Basics | abs.gov.au https://www.abs.gov.au/websitedbs/d3310114.nsf/home/time+series+analysis:+the+basics
- Interrupted time series study | GOV. UK https://www.gov.uk/guidance/interrupted-time-series-study
- Three Statistical Approaches for Assessment of Intervention Effects: A Primer for Practitioners - PMC https://pmc.ncbi.nlm.nih.gov/articles/PMC7910529/
- 4 Seasonal Models - STAT 510 | Applied Time Series Analysis https://online.stat.psu.edu/stat510/Lesson04.html
- Confounders in Time-Series Regression | Columbia University Mailman School of Public Health https://www.publichealth.columbia.edu/research/population-health-methods/confounders-time-series-regression
- Interrupted Time Series: Segmented Regression | CASRAI https://casrai.org/guides/interrupted-time-series-analysis
- Use of Interrupted Time Series Analysis in Evaluating Health Care Quality Improvements - Academic Pediatrics https://www.academicpedsjnl.net/article/S1876-2859(13)00210-6/fulltext
