Psychology dissertation Poisson regression is designed for non-negative event counts, such as the number of panic episodes, help-seeking contacts, aggressive acts, missed appointments, or correct responses observed during a defined opportunity. It respects the discrete outcome and models an expected count or rate, but only when the outcome, exposure period, variance structure, and independence assumptions are handled carefully.
This guide uses a hypothetical study of daily reassurance-seeking episodes among adults completing a 14-day diary. The main question is whether baseline intolerance of uncertainty is associated with the expected episode rate after accounting for observation time. The example shows why choosing Poisson regression is more than noticing that the outcome contains whole numbers.
What psychology dissertation Poisson regression estimates
Poisson regression belongs to the generalised linear model family. It models the logarithm of an expected count as a linear combination of predictors. Exponentiating a coefficient produces an incidence-rate ratio or count ratio, depending on whether exposure time is included. A ratio of 1.20 means the expected rate is 20% higher for a one-unit predictor increase, holding other model variables constant.
The model’s predictions are non-negative and its mean-variance relationship reflects count data. Ordinary linear regression can predict impossible negative counts and generally gives unsuitable uncertainty when residual variance changes with the mean. Logistic regression is for binary outcomes, not counts. Do not discard information by converting “number of episodes” into “any versus none” solely to use a familiar test.
Count outcomes need a defined opportunity
A count is meaningful only relative to an observation window or opportunity. Ten interruptions during a 30-minute task differ from ten interruptions during a five-hour session. If all participants have equal exposure, modelling counts may answer the question. If exposure varies, an offset commonly converts the model into a rate comparison.
When Poisson regression fits a psychology dissertation
The outcome should be an integer count beginning at zero, with events defined consistently. Examples include therapy attendances, self-harm incidents, intrusive thoughts recorded per day, peer interactions during observation, or errors on a task. The events should occur within a specified interval and be sufficiently comparable across people or units.
| Outcome | Possible model | Key question |
|---|---|---|
| Number of help-seeking contacts | Poisson or negative binomial | Are follow-up periods comparable? |
| Whether any contact occurred | Binary logistic regression | Is losing frequency information justified? |
| Time until first contact | Survival analysis | Are censoring and event timing central? |
| Continuous symptom severity score | Linear or another continuous-outcome model | Is the score genuinely continuous enough? |
| Repeated daily counts | Multilevel count model or GEE | Are observations nested within participants? |
The model is unsuitable merely because values are skewed. A bounded scale, ordinal score, proportion, or number-correct outcome may require a different distribution. Use the site’s operational definitions guide to specify what qualifies as an event, who records it, the window, and the denominator.
Define the count-generating process
Before modelling, describe how events arise. Ask whether one event changes the chance of another, whether opportunities vary, whether some participants cannot experience the event, and whether repeated recording is reliable. Reassurance-seeking episodes may cluster during acute stress, so conditional independence can be questionable even when daily totals are counts.
Specify time zero and the unit of analysis. One row per participant with a 14-day total differs from one row per participant-day. The latter creates repeated observations and requires a model that recognises within-person dependence. Avoid collapsing repeated data simply to satisfy an independence assumption when timing is scientifically useful.
Understand the Poisson assumptions
Conditional mean and variance
The standard Poisson model assumes the conditional variance equals the conditional mean after accounting for predictors and exposure. Raw variance exceeding the raw mean is not, by itself, proof that the fitted model is wrong because predictors may explain heterogeneity. Diagnose dispersion from the model and residuals.
Conditional independence
Counts should be independent after conditioning on model variables. Violations arise with repeated measures, participants nested in clinics, contagion between peers, or event dependence over time. Cluster-robust uncertainty, generalised estimating equations, or multilevel Poisson models may be needed. The multilevel modelling guide explains nested structures.
Correct mean structure
The log of the expected count should relate appropriately to continuous predictors. A straight line on the log scale is not guaranteed. Inspect functional form using theory, plots, splines, or planned transformations. Interactions should be included only when the question concerns conditional associations and the sample contains relevant information.
Consistent event definition
Events must be comparable across participants. If one person logs every urge while another logs only actions, the outcome mixes measurement processes. Training, clear codebooks, reminders, and inter-rater procedures can reduce measurement error in observational counts.
Use exposure and offsets correctly
An exposure variable represents the amount of time, opportunity, population, or space in which events could occur. In rate regression, the natural logarithm of exposure enters as an offset with its coefficient fixed at one. The model then compares event rates while retaining the count likelihood.
For the diary example, a participant observed for 14 valid days has more opportunity to record episodes than one observed for seven days. Using log(valid days) as an offset treats twice the exposure as implying twice the expected count, all else equal. This proportionality assumption should be plausible rather than automatic.
| Scenario | Exposure | Interpretation |
|---|---|---|
| Daily diary episodes | Valid observation days | Expected episodes per day |
| Behavioural observations | Minutes observed | Expected acts per unit time |
| Errors across tasks | Number of eligible trials | Expected errors per trial opportunity |
| Clinic incidents | Person-days at risk | Expected incidents per person-day |
| Equal follow-up for everyone | Constant exposure | Count ratios equal rate ratios up to the constant |
Do not enter exposure as an ordinary predictor unless estimating its coefficient is substantively intended. Do not offset by zero or negative exposure. Investigate participants with no valid observation opportunity separately because their event count is structurally undefined rather than simply zero.
Plan predictors and sample size
Choose predictors from the research question, design, and causal reasoning. Include confounders because they are plausible common causes, not because they pass univariate screening. Avoid controlling for mediators or colliders when estimating a total effect. The site’s confounding guide provides a structured selection workflow.
Count-model information depends on the number of events, exposure distribution, predictor variation, zero frequency, overdispersion, and model complexity. A universal participants-per-predictor rule is inadequate. Simulation based on plausible rates and dispersion is preferable. If simulation is unavailable, justify the sample using anticipated events, smallest meaningful rate ratio, predictor distributions, attrition, and desired precision.
Prepare and explore count data
Report the mean, variance, median, range, zero proportion, and exposure distribution. Plot counts overall and across important predictors. Check impossible negative values, fractional counts, duplicated records, changes in definitions, and observation periods. A histogram alone cannot choose the model, but it can reveal coding mistakes and extreme cases.
Distinguish true zeros from missing records. “No episode occurred” differs from “the diary was not completed.” Coding missing days as zero lowers the apparent rate and may introduce differential measurement bias. Follow the site’s missing-data guide and preserve an auditable distinction.
Fit the Poisson model in a reproducible sequence
- Define the outcome count, exposure, unit, time window, and target estimand.
- Plot counts and exposure, audit zeros, and confirm event coding.
- Specify predictors, reference categories, interactions, and functional forms in advance.
- Fit the Poisson model with the correct log exposure offset if needed.
- Inspect fitted values, residuals, influential observations, and dispersion.
- Compare defensible alternatives when assumptions fail.
- Present incidence-rate ratios, predicted counts or rates, and uncertainty.
- Run sensitivity analyses and report every material decision.
Preserve syntax rather than relying only on menus. R can fit a standard model with glm(..., family = poisson); SPSS can use a generalised linear model with a Poisson distribution and log link. Software defaults differ, so record link, offset, covariance estimator, reference categories, convergence, and diagnostic choices.
Diagnose overdispersion
Overdispersion means the conditional variance exceeds the Poisson mean. It can arise from omitted predictors, unobserved heterogeneity, dependence, excess zeros, measurement differences, or an unsuitable functional form. If ignored, conventional Poisson standard errors can be too small, creating overconfident intervals and tests.
Inspect Pearson and deviance dispersion measures, residual plots, observed versus predicted frequencies, and subject-matter causes. A single threshold is not a mechanical model switch. UCLA’s count-model guidance recommends checking specification before treating overdispersion as only a distribution problem.
Quasi-Poisson and robust uncertainty
Quasi-Poisson scales the variance while retaining the mean structure. Robust sandwich standard errors can protect inference against some variance misspecification. Neither automatically repairs a wrong mean model, dependence, omitted nonlinearities, or zero-generating mechanisms. They also differ from full likelihood models in fit comparison and prediction.
Negative binomial regression
Negative binomial regression adds a dispersion parameter so conditional variance can exceed the mean. It often suits heterogeneous psychology counts, but its parameterisation must be reported. Compare predicted distributions and diagnostics, not just significance. A similar coefficient across models with wider negative-binomial intervals may reveal that the Poisson model understated uncertainty.
Investigate underdispersion and excess zeros
Underdispersion occurs when conditional variance is below the mean and may result from bounded opportunities, regular behaviour, or dependence. Standard Poisson uncertainty may then be conservative or otherwise unsuitable. Generalised Poisson or other flexible models can be considered with specialist guidance.
Many zeros do not automatically require a zero-inflated model. A low expected rate naturally produces zeros. Zero inflation assumes an additional process that creates structural zeros, such as participants who cannot experience an event, alongside a count process for those at risk. This distinction must be theoretically plausible and identifiable from the design.
Hurdle models separate any occurrence from the positive count process. They can fit questions where starting an action differs from repeating it. Avoid choosing zero-inflated or hurdle models solely because they reduce an information criterion. Explain the behavioural mechanism and examine classification uncertainty.
Check influential observations and model fit
Large counts can have strong influence, especially with low exposure or unusual predictor combinations. Inspect Pearson or deviance residuals, leverage, influence measures, and case-level context. Do not delete a legitimate high-frequency participant simply because the model fits better without them. Correct errors, justify exclusions, and report sensitivity analyses.
Compare observed and predicted count distributions, zero proportions, tail behaviour, and predicted rates across meaningful predictor values. Residual patterns against fitted values, exposure, and continuous predictors can reveal nonlinearity or heterogeneity. Absolute fit and scientific plausibility matter alongside AIC or likelihood comparisons.
Interpret incidence-rate ratios accurately
Exponentiated coefficients are multiplicative. An incidence-rate ratio of 0.80 corresponds to a 20% lower expected rate, not a decrease of 0.80 events. For a continuous predictor, state the unit: a one-point change may be trivial, while a five-point or one-standard-deviation contrast may be interpretable.
Present predicted counts or rates for concrete profiles with confidence intervals. For nonlinear models, an interaction means the rate ratio for one predictor changes across the other predictor, but the coefficient alone may be hard to understand. Plot predicted rates across the observed range and avoid extrapolation.
| Result | Useful wording | Avoid |
|---|---|---|
| IRR 1.25 | Expected rate was 25% higher per stated unit | The outcome increased by 1.25 events |
| IRR 0.70 | Expected rate was 30% lower | Risk fell by 70% |
| Wide confidence interval | Estimate was imprecise and compatible with varied effects | No effect was proven |
| Observational association | Predictor was associated with the expected rate | Predictor caused the events |
Repeated and clustered count outcomes
Daily events nested within participants violate ordinary independence. A generalised estimating equation estimates population-average associations under a working correlation structure, while a generalised linear mixed model includes participant-level random effects and estimates conditional associations. Their coefficients answer different questions.
Also consider serial dependence, irregular observation days, time trends, and time-varying predictors. A random intercept does not automatically handle every temporal pattern. Align the model with the longitudinal design and avoid treating thousands of daily rows as thousands of independent participants.
Psychology dissertation Poisson regression example
Suppose 180 adults complete up to 14 diary days. The outcome is the total number of reassurance-seeking episodes; valid diary days form the exposure. Predictors are baseline intolerance of uncertainty, age, and prior therapy. The primary model uses a log valid-days offset and reports adjusted incidence-rate ratios.
Exploration shows many low counts, several high counts, and 18% zeros. The fitted Poisson model has residual overdispersion and underpredicts the upper tail. After verifying coding and functional form, the researcher fits a pre-specified negative binomial sensitivity model. Coefficients are similar, but intervals widen. The negative binomial model is retained because it represents heterogeneity better, while the Poisson result remains a transparent sensitivity analysis.
The conclusion states an association, not causation: higher intolerance of uncertainty was associated with a higher expected daily reassurance-seeking rate, conditional on measured covariates. Unmeasured distress, self-monitoring differences, missing diary days, and within-person variation remain limitations.
Report Poisson regression transparently
Methods should name the outcome, event definition, observation window, exposure and offset, predictors, coding, transformations, link, distribution, independence structure, missing-data approach, dispersion assessment, influence diagnostics, alternatives, software, and planned sensitivity analyses.
Results should provide sample flow, total events, exposure summaries, mean, variance, zero frequency, coefficient or incidence-rate ratios with confidence intervals, model diagnostics, dispersion evidence, fit comparisons, predicted rates, influential-case analyses, and deviations from the plan. Follow the site’s effect-size guide for proportionate interpretation.
Common Poisson-regression mistakes
- Using Poisson because data are skewed: confirm the outcome is a count with defined opportunity.
- Ignoring unequal exposure: counts from different observation windows are not directly comparable.
- Testing only raw mean versus variance: dispersion is conditional on the fitted model.
- Switching models only for significance: choose alternatives from assumptions and diagnostics.
- Treating all zeros as structural: low rates naturally generate zeros.
- Ignoring repeated measures: within-person counts are dependent.
- Interpreting coefficients additively: exponentiated effects are multiplicative ratios.
- Claiming causes from observational predictors: model adjustment does not remove every bias.
Frequently asked questions
Does a count outcome always require Poisson regression?
No. The process may be binomial, negative binomial, zero-inflated, hurdle-based, ordinal, time-to-event, or multilevel. Choose from the outcome definition, opportunity, dispersion, dependence, and research question.
How do I know whether data are overdispersed?
Inspect model-based Pearson or deviance dispersion, residuals, predicted frequencies, and design causes. Do not rely only on raw variance exceeding the raw mean or one automatic test.
When should I use negative binomial regression?
It is often defensible when conditional variance exceeds the Poisson mean because of unobserved heterogeneity and diagnostics support its variance structure. First check coding, dependence, omitted structure, and exposure.
What is an offset?
An offset is typically the log of exposure entered with a coefficient fixed at one. It lets the model compare rates while using count outcomes, such as events per valid observation day.
Can Poisson regression include categorical predictors?
Yes. Code reference categories clearly and interpret their exponentiated coefficients as expected rate ratios relative to the reference, conditional on other model variables.
Can SPSS run Poisson regression?
Yes. Generalised linear model procedures can fit Poisson models with a log link and exposure or offset. Availability varies by version, so preserve syntax and verify every default.
Should I report coefficients or incidence-rate ratios?
Report enough detail for reproducibility, but incidence-rate ratios and predicted rates are usually easier to interpret. Always include confidence intervals, units, reference groups, and the exposure basis.
Conclusion
Psychology dissertation Poisson regression is valuable when the outcome is a genuine event count observed over a defined opportunity. A strong analysis specifies the count process, uses exposure correctly, diagnoses dispersion and dependence, compares defensible alternatives, checks influence and fit, and communicates multiplicative effects as rates or predicted counts.
If you need ethical dissertation support, use feedback to refine the question, review model assumptions, check syntax and diagnostics, or improve reporting. Retain ownership of the analysis and final writing, protect participant data, and follow your university’s rules on permitted assistance.
