15.1 Detecting and Handling Overdispersion

Overdispersion in Regression Models

Overdispersion is one of the most common problems you'll encounter when fitting count or binary outcome models. It occurs when the observed variance in your data exceeds what the assumed distribution (Poisson or binomial) predicts. If you ignore it, your standard errors shrink artificially, your test statistics inflate, and you end up claiming predictors are significant when they aren't. This section covers how to detect overdispersion and what to do about it.

Definition and Consequences

In a standard Poisson model, the mean and variance are assumed to be equal. In a binomial model, the variance is a fixed function of the mean. Overdispersion is the situation where the observed variance exceeds these theoretical values.

The practical consequences are serious:

• Underestimated standard errors. Because the model thinks the data are less variable than they actually are, it produces standard errors that are too small.
• Inflated test statistics. Smaller standard errors mean larger z- or t-statistics, which leads to artificially small p-values.
• Incorrect inferences. You may conclude that predictors are statistically significant when they aren't, or build confidence intervals that are far too narrow.
• Poor model fit. A high deviance or Pearson chi-square statistic relative to the residual degrees of freedom signals that the model isn't capturing the true variability in the data.

Signs and Symptoms

Before running formal tests, watch for these red flags:

• A variance that increases faster than the mean predicts. For Poisson data, if you plot the sample variance against the sample mean across groups, the points should fall near the identity line. If they consistently fall above it, overdispersion is likely.
• An excess of extreme values. For example, in a Poisson regression predicting daily car accidents, you might see far more days with very high accident counts than the Poisson distribution can accommodate.
• A right-skewed distribution of residuals, or residuals with heavier tails than expected.

Failing to account for overdispersion doesn't just hurt model fit. In a binomial regression examining student dropout rates, for instance, ignoring overdispersion could cause you to flag predictors as significant that have no real association with dropout.

Diagnosing Overdispersion

Dispersion Parameter and Residual Plots

The most direct diagnostic is the dispersion parameter $\hat{\phi}$, estimated by dividing the Pearson chi-square statistic (or the deviance) by the residual degrees of freedom:

$\hat{\phi} = \frac{X^2}{n - p}$

where $X^2$ is the Pearson chi-square, $n$ is the number of observations, and $p$ is the number of estimated parameters.

• If $\hat{\phi} \approx 1$, the variance matches the assumed distribution. No overdispersion.
• If $\hat{\phi}$ is substantially greater than 1, overdispersion is present. A value of 2.5, for example, means the observed variance is about 2.5 times what the model expects.
• Values slightly above 1 (say, 1.1 or 1.2) are common and may not require correction, especially with large samples.

Residual plots offer a visual check. Plot Pearson residuals against fitted values:

• Under a well-fitting model, the residuals should scatter randomly with roughly constant spread.
• A fan-shaped pattern (variance increasing with fitted values) is a classic sign of overdispersion. In a Poisson model predicting hospital admissions, for instance, you'd see the residuals fanning out as predicted admission counts grow.

Formal Tests for Overdispersion

When you want a definitive answer, use a formal hypothesis test.

Score test (Lagrange multiplier test):

• Tests $H_0: \phi = 1$ against $H_a: \phi > 1$.
• A significant result (small p-value) provides evidence of overdispersion.
• This test is quick because it only requires fitting the simpler (e.g., Poisson) model.

Likelihood ratio test:

• Compares a model that allows overdispersion (e.g., negative binomial) to one that doesn't (e.g., Poisson).
• The test statistic is $\Lambda = 2(\ell_{\text{NB}} - \ell_{\text{Poisson}})$, where $\ell$ denotes the log-likelihood.
• A significant improvement in fit with the negative binomial model supports the presence of overdispersion.
• Note: because the Poisson model sits on the boundary of the negative binomial parameter space (the overdispersion parameter equals zero under $H_0$), the standard chi-square reference distribution is conservative. A mixture distribution $\frac{1}{2}\chi^2_0 + \frac{1}{2}\chi^2_1$ is more appropriate, though in practice many software packages handle this automatically.

Handling Overdispersion

Quasi-Likelihood Methods

Quasi-Poisson and quasi-binomial regression are the simplest fixes. They work by multiplying the model-based variance by the estimated dispersion parameter $\hat{\phi}$.

How quasi-likelihood adjusts the model:

1. Fit the standard GLM (Poisson or binomial) as usual. The coefficient estimates $\hat{\beta}$ stay the same.
2. Estimate $\hat{\phi}$ from the Pearson chi-square or deviance.
3. Multiply all variance estimates by $\hat{\phi}$, which inflates the standard errors by $\sqrt{\hat{\phi}}$.
4. Recompute test statistics and p-values using the adjusted standard errors.

The key trade-off: quasi-likelihood is simple and doesn't require specifying a full probability distribution, but because there's no true likelihood, you can't use AIC/BIC for model comparison, and you don't get a proper likelihood for prediction intervals.

Negative Binomial Regression

For count data where the Poisson assumption fails, negative binomial regression adds an explicit overdispersion parameter $\alpha$ (sometimes parameterized as $\theta = 1/\alpha$). The variance becomes:

$\text{Var}(Y) = \mu + \alpha \mu^2$

This is a full parametric model, so unlike quasi-likelihood, you get a genuine likelihood function. That means you can use AIC, BIC, and likelihood ratio tests for model selection.

Use negative binomial regression when:

• You have count data with variance clearly exceeding the mean.
• You need model comparison tools (AIC/BIC).
• You want proper prediction intervals.

For example, in a study of doctor visit counts, switching from Poisson to negative binomial regression often dramatically improves fit because visit counts tend to be highly variable across patients.

Zero-Inflated Models

Sometimes overdispersion is driven by an excess of zeros rather than general extra variability. Zero-inflated models handle this by combining two components:

1. A binary component (often logistic regression) that models whether an observation is a "structural zero" (an observation that could never have a nonzero count) versus a "potential count."
2. A count component (Poisson or negative binomial) that models the counts for observations that aren't structural zeros.

A zero-inflated Poisson (ZIP) model is appropriate when the non-zero counts follow a Poisson distribution but there are too many zeros. A zero-inflated negative binomial (ZINB) model handles both excess zeros and overdispersion in the non-zero counts.

Generalized Estimating Equations (GEE)

When your data have a longitudinal or clustered structure, overdispersion can arise from correlation among observations within the same cluster. GEE handles this by:

• Specifying a working correlation structure (exchangeable, autoregressive, unstructured, etc.) to model within-cluster dependence.
• Using robust (sandwich) standard errors that remain valid even if the working correlation is misspecified.

GEE is a population-averaged approach. It tells you about marginal effects across the population rather than subject-specific effects. In a longitudinal study of seizure counts in epilepsy patients, GEE accounts for both overdispersion and the fact that repeated measurements on the same patient are correlated.

Hierarchical and Multilevel Models

An alternative to GEE is to model the source of overdispersion directly using random effects. A Poisson model with a random intercept, for instance, introduces subject- or group-level variability that the fixed effects alone can't capture.

• The random effects induce a marginal distribution with variance greater than the mean, naturally accommodating overdispersion.
• These models estimate both fixed effects (regression coefficients) and random effects (variance components).
• In a study of crime rates across neighborhoods, a multilevel Poisson model with random intercepts for each neighborhood captures both the overdispersion and the spatial clustering in the data.

The choice between GEE and multilevel models depends on your research question: GEE for population-averaged inference, multilevel models for cluster-specific inference.

Overdispersion's Impact on Model Fit and Selection

Comparing Models

Use goodness-of-fit statistics to evaluate whether accounting for overdispersion improves your model:

• AIC and BIC: Lower values indicate better fit. Comparing the AIC of a Poisson model to a negative binomial model is a straightforward way to assess whether the extra parameter is justified. (Remember: quasi-likelihood models don't produce a true likelihood, so AIC/BIC aren't available for quasi-Poisson.)
• Coefficient changes: After accounting for overdispersion, examine how your coefficient estimates and their standard errors change. The point estimates in quasi-likelihood stay the same, but standard errors grow. In negative binomial regression, both coefficients and standard errors may shift. Predictors that appeared significant under Poisson may lose significance once overdispersion is properly handled.

In a study of hospital readmission risk factors, for instance, a Poisson model might flag five significant predictors while a negative binomial model retains only two. The difference reflects the Poisson model's artificially narrow confidence intervals.

Sensitivity Analysis and Predictive Performance

Running multiple approaches and comparing results strengthens your conclusions:

• Fit the data with quasi-Poisson, negative binomial, and (if appropriate) zero-inflated models. If the same predictors emerge as significant across methods, you can be more confident in those findings. Discrepancies signal that results are sensitive to modeling assumptions and deserve closer scrutiny.
• Evaluate predictive performance using cross-validation or posterior predictive checks. Models that account for overdispersion typically predict extreme values and rare events more accurately. In insurance claims modeling, for example, a negative binomial model often outperforms Poisson in cross-validated prediction error because it better captures the heavy right tail of the claims distribution.