Psychology dissertation negative binomial regression helps you analyse overdispersed count outcomes without forcing them into a continuous or binary model. This guide explains when the method fits a research question, how it differs from Poisson and zero-inflated models, what to diagnose, and how to report incidence-rate ratios responsibly.
What negative binomial regression answers
Negative binomial regression models a non-negative integer outcome, such as the number of panic episodes in a month, therapy sessions attended, aggressive incidents recorded during observation, or errors made on a sustained-attention task. It estimates how the expected count changes with predictors while respecting the discrete and often right-skewed nature of the outcome.
The method is most useful when counts vary more than a Poisson model expects. A sample may contain many participants with low counts and a smaller group with much higher counts because people differ in unmeasured vulnerability, opportunity, exposure, or context. Negative binomial regression introduces a dispersion parameter that can represent this extra heterogeneity.
It does not automatically solve every problem involving counts. Your design, measurement window, dependence structure, missingness, excess zeros, functional form, and influential cases still matter. Model choice should follow a defensible account of how observations were generated rather than a single rule based on the sample variance.
Poisson, negative binomial and other count models
Poisson regression is a useful starting point. Conditional on the predictors, it assumes the outcome variance equals its mean. Negative binomial regression relaxes this equality by allowing additional variance. In the common NB2 parameterisation, variance rises approximately as the mean plus a dispersion term multiplied by the squared mean.
| Model | Best considered when | Main caution |
|---|---|---|
| Poisson regression | Counts are conditionally equidispersed after relevant predictors are included | Unmodelled overdispersion can make standard errors too small |
| Negative binomial regression | Counts remain overdispersed and a one-process count model is plausible | It may absorb zero inflation without explaining a separate zero-generating process |
| Zero-inflated model | Theory supports a structural-zero process plus a count process | Extra complexity needs substantive justification and adequate data |
| Hurdle model | Any occurrence and the number of occurrences may have distinct mechanisms | Results have two linked but separate interpretations |
| Count mixed model or GEE | Counts are clustered or repeatedly measured | Independence assumptions, random effects or working correlation must be addressed |
Overdispersion is conditional
A raw variance larger than a raw mean is a warning, not a complete diagnosis. Dispersion should be considered after modelling relevant predictors and exposure. Omitting an important nonlinear term, interaction, cluster, or offset can create apparent overdispersion. Conversely, a negative binomial model fitted when conditional dispersion is negligible may add an unstable parameter without improving inference.
Zeros need a theory
Many zeros are not automatically excessive. If the mean count is low, an ordinary Poisson or negative binomial distribution may predict many zeros. A zero-inflated model claims that some observations belong to a separate always-zero state. For example, zero help-seeking contacts could reflect no perceived need, no access, or a low expected count. These explanations imply different mechanisms and should not be selected only because a more complex model has a lower information criterion.
Define the psychology outcome precisely
A valid count has a clear event, observation window, unit, and recording process. “Number of symptoms” differs from “number of episodes,” and a sum of Likert items is not automatically a count outcome suitable for count regression. State what qualifies as one event, whether repeated events can occur, and how ambiguous records were handled.
Counts should normally be integers at or above zero. Check impossible values, duplicated records, heaping, top-coding, and whether measurement accuracy changes across participants. Self-recorded daily events may be affected by recall, fatigue, device access, or declining adherence. Administrative counts may depend on service availability as much as psychological need.
Use an offset for unequal exposure
Participants may be observed for different lengths of time or have different opportunities for an event. A participant monitored for 28 days has more opportunity to record an episode than one monitored for 7 days. Include the logarithm of valid observation time as an offset when the goal is to model a rate. Its coefficient is fixed at one, so the remaining coefficients compare expected rates at equal exposure.
Exposure must be measured rather than guessed. For behavioural observation, it could be minutes coded; for service use, months enrolled; for online activity, active days. If exposure itself is affected by the outcome or treatment, explain the causal complication and consider sensitivity analyses.
Plan a psychology dissertation negative binomial regression
Begin with a question that names the count outcome, predictors, population, period, and estimand. A suitable question might ask whether baseline social anxiety predicts the rate of avoided social events during six weeks, adjusting for valid diary days, age, and treatment condition. This is clearer than asking whether anxiety “affects behaviour.”
Specify predictors before looking for significance
Distinguish focal predictors from adjustment variables, moderators, and exploratory variables. Code reference categories deliberately. For continuous predictors, use meaningful units or standardisation when it aids interpretation. Inspect whether a linear effect on the log-mean scale is plausible; splines or carefully justified polynomial terms may represent curvature better than categorising a continuous measure.
Avoid selecting covariates solely because their bivariate p-values are small. Theory, design, temporal order, and a causal diagram can help identify confounders without adjusting for mediators or colliders. The site’s guide to confounding variables explains this distinction in more depth.
Plan sample size for the actual design
There is no universal observations-per-predictor rule for negative binomial regression. Precision depends on the expected mean, dispersion, exposure distribution, predictor prevalence, correlation among predictors, effect size, missingness, and clustering. Sparse binary groups and rare events can produce unstable estimates even when the total sample appears large.
Use simulation or software designed for count outcomes whenever possible. Simulate counts under plausible means and dispersion, reproduce the planned predictor structure and missingness, fit the intended model, and estimate power, bias, interval coverage, convergence, and the frequency of extreme estimates. Document assumptions and test less favourable scenarios. See the psychology dissertation power analysis guide for the broader planning logic.
Prespecify a model sequence
A transparent sequence might include a Poisson model, a negative binomial model, and theory-driven alternatives for zeros or clustering. Prespecify which comparisons are diagnostic and which model will answer the primary question. This reduces the temptation to select whichever specification produces the most attractive p-value.
Fit and compare the models
Most negative binomial regressions use a log link. The linear predictor therefore describes the logarithm of the expected count, while exponentiated coefficients give multiplicative effects. Maximum-likelihood software can estimate the regression coefficients and dispersion parameter jointly. The statsmodels negative binomial documentation describes formula-based fitting, prediction and likelihood methods. R users can consult the official MASS glm.nb documentation.
Check parameterisation
Software packages do not always define dispersion in the same direction or use the same variance function. NB1 permits variance to increase linearly beyond the mean, whereas NB2 commonly uses a quadratic mean-variance relationship. A parameter called alpha in one package may correspond to the inverse of theta in another. Name the package, function, family, link, parameterisation, and version so readers can reproduce the analysis.
Compare nested and non-nested specifications carefully
Likelihood-ratio tests can compare appropriately nested models, but testing a dispersion parameter at the Poisson boundary requires care. Information criteria such as AIC compare relative expected out-of-sample fit among models fitted to the same outcome observations. Lower AIC is useful evidence, not proof that a model represents the data-generating process.
For zero-inflated or hurdle alternatives, combine fit statistics with predicted-versus-observed frequencies, residual checks, convergence, parameter stability, and substantive plausibility. An open-access simulation comparison of count models shows why performance depends on dispersion, zero generation, and sample size rather than one model being universally superior.
Diagnose the final model
Diagnostics ask whether the fitted model reproduces important features of the data and whether conclusions depend on a few observations or choices. Save code and decisions as you proceed rather than reconstructing them after results are known.
| Diagnostic question | Practical check | Possible response |
|---|---|---|
| Does the model reproduce the count distribution? | Compare observed and simulated frequencies, zeros, upper tail and fitted means | Revisit dispersion, exposure, nonlinear terms or zero process |
| Are residual patterns systematic? | Plot appropriate residuals against fitted values, predictors and time | Adjust functional form or dependence structure |
| Are cases influential? | Inspect leverage, case-deletion change and unusually large counts | Verify records and report sensitivity analyses |
| Did estimation succeed? | Check warnings, gradient, Hessian, iterations and starting values | Simplify unjustified complexity or investigate separation and sparsity |
| Is dependence represented? | Review repeated measures, sites, therapists, schools or families | Use a count mixed model, GEE or cluster-aware inference |
Use residuals suitable for discrete outcomes
Ordinary raw residuals are difficult to interpret for discrete, heteroscedastic outcomes. Deviance, Pearson, or simulation-based quantile residuals can be more informative. Plot them against fitted values, each important predictor, observation time, and exposure. A smooth pattern may indicate a missing nonlinear effect; changing spread may signal a poor mean-variance relationship.
Investigate influential observations without deleting inconvenience
A very high count may be a data error, a valid rare case, or evidence that the model misses a subgroup. Verify the source record and eligibility first. Then compare results with and without the case only as a labelled sensitivity analysis. Do not remove an observation merely because it changes significance. The outlier analysis guide offers a defensible workflow.
Address clustering and repeated counts
Counts collected repeatedly from the same participant are not independent. Neither are observations nested within classrooms, clinics, therapists, or families. A standard negative binomial model can estimate the mean structure but give misleading uncertainty if dependence is ignored. Consider a negative binomial mixed model for subject-specific effects or a GEE approach for population-average effects, aligned with the research question. The GEE guide explains working correlations and robust standard errors.
Interpret coefficients and incidence-rate ratios
With a log link, exponentiating a coefficient produces an incidence-rate ratio, often abbreviated IRR. An IRR of 1.20 for a one-unit predictor increase means the expected count or rate is multiplied by 1.20, holding other variables constant. It is commonly described as 20% higher. An IRR of 0.80 corresponds to an expected rate 20% lower, not 80% lower.
Always state the predictor unit, reference category, covariates, exposure basis, confidence interval, and outcome period. For an interaction, a main-effect IRR is conditional on the other interacting variable. Present predicted counts or rates at meaningful values because they are often easier to understand than a table of conditional multiplicative coefficients.
A psychology example
Suppose 240 students record the number of intrusive thoughts each day for 30 days. The outcome is overdispersed and observation days vary because some diaries are incomplete. A negative binomial model uses valid diary days as an offset and includes baseline stress, sleep quality, study stage, and a stress-by-sleep interaction.
If the adjusted IRR for a five-point stress increase is 1.18 with a 95% confidence interval from 1.07 to 1.30, the expected daily rate is estimated to be 18% higher at that stress difference, conditional on the covariates and reference values. This is an association unless the design and assumptions support a causal claim. Predicted rates across realistic stress and sleep values can communicate the interaction more clearly.
Handle missing data and measurement limitations
Separate a genuine zero from a missing observation. Recording no panic attacks is a count of zero; failing to complete the diary is missing. Replacing missing days with zeros depresses the rate and can create artificial group differences. Define valid exposure, completion rules, and whether partially observed periods contribute.
Complete-case analysis assumes a missingness process that may be implausible. Multiple imputation can be appropriate for missing predictors, but the imputation model must respect distributions, interactions, clustering, and the outcome. Missing outcome counts usually require design-specific reasoning. Report missingness by variable and group, reasons when known, and sensitivity analyses. Consult the missing data guide for planning principles.
Report the analysis transparently
A reader should be able to reconstruct why the model was chosen, how variables were coded, what was tested, and whether alternatives changed the conclusion. Avoid reporting only a coefficient table.
| Report element | Include |
|---|---|
| Design and outcome | Population, event definition, period, unit, inclusion rules and exposure |
| Model | Negative binomial parameterisation, log link, predictors, interactions, offset and dependence structure |
| Estimation | Software, package, version, convergence checks and uncertainty method |
| Diagnostics | Dispersion evidence, residual checks, predicted frequencies, influence and model comparisons |
| Results | Coefficients or IRRs, confidence intervals, exact p-values where relevant, and predicted rates |
| Sensitivity | Reasonable Poisson, zero-process, influence, missing-data or clustering alternatives |
| Limits | Measurement error, unmeasured heterogeneity, generalisability and causal restrictions |
Keep descriptive statistics on the count scale, including the mean, standard deviation, median, range, proportion of zeros, exposure distribution, and a frequency plot. Describe missingness and exclusions. Report the dispersion estimate using the software’s definition. Present confidence intervals and practical quantities, not only significance labels.
Common mistakes to avoid
- Applying linear regression to a strongly skewed count and accepting possible negative predictions.
- Choosing negative binomial regression only because the raw variance exceeds the raw mean.
- Treating all unrecorded events or incomplete diary days as zeros.
- Ignoring unequal observation time instead of using a justified offset.
- Assuming many zeros prove that a zero-inflated model is required.
- Interpreting an IRR as an absolute count difference or a causal effect.
- Comparing AIC values from models fitted to different observations.
- Ignoring clustering, convergence warnings, influential cases, or nonlinear predictors.
- Trying several models and reporting only the one with the smallest p-value.
Frequently asked questions
When should I use negative binomial rather than Poisson regression?
Consider it when the outcome is a count and meaningful overdispersion remains after exposure, predictors, functional form, and dependence have been addressed. Confirm the choice with fit, residuals, predicted frequencies, convergence, and substantive reasoning.
Is a Likert-scale total a count outcome?
Usually not in the event-count sense. A summed rating scale has bounded support and reflects item responses rather than recurring events. Its measurement properties and distribution should guide analysis; do not select negative binomial regression merely because values are integers.
What is the difference between overdispersion and zero inflation?
Overdispersion means conditional variability exceeds that expected by a Poisson model. Zero inflation proposes an additional structural-zero mechanism. Zero inflation can contribute to overdispersion, but overdispersion can also arise from unobserved heterogeneity, dependence, omitted predictors, or other distributional features.
Do I need an offset?
Use an offset when observation time or opportunity differs and the estimand is a rate. The offset is usually the logarithm of a positive exposure measure. If everyone has equal exposure, an offset may be unnecessary, although the common period should still be reported.
How do I report an incidence-rate ratio?
Give the IRR, confidence interval, predictor unit or reference group, covariate conditions, exposure basis, and outcome period. Add predicted counts or rates at meaningful values when that improves interpretation.
Can negative binomial regression handle repeated measures?
The distribution can be used within mixed-effects or GEE frameworks, but an ordinary independent-observations model does not automatically handle repeated data. Choose a subject-specific or population-average framework that matches the research question.
What if the negative binomial model does not converge?
Check coding, sparse categories, extreme counts, scaling, starting values, redundant predictors, dispersion near the Poisson boundary, and unjustified zero-inflation or random-effects complexity. Do not conceal warnings. Simplify only when the resulting model remains theoretically defensible.
Conclusion
Psychology dissertation negative binomial regression is a defensible option for overdispersed event counts when the outcome, exposure, mean structure, and dependence are specified carefully. A strong dissertation explains why extra-Poisson variation is plausible, compares reasonable alternatives, checks residual and predictive behaviour, and interprets IRRs with clear units and limits.
If you need methodological tutoring, seek support that helps you understand, run, and explain your own analysis. Keep your data secure, follow institutional ethics, academic integrity, and authorship rules, and retain responsibility for every decision and claim in the submitted dissertation.
