Checking Poisson Regression Assumptions for Count Outcomes

Why Check the Poisson Assumptions?

A count outcome takes nonnegative integer values: for example, the number of service calls received by a fictional assistance program in a month. Poisson regression is a natural model for such data, but the fact that an outcome is a count does not by itself make the Poisson model appropriate. Its key restriction is that, conditional on the modeled information, the count’s variance equals its mean. Counts may be more variable than this restriction allows, may contain more zeros than predicted, or may be dependent because observations share unmodeled conditions.

The assessment is conditional: compare variation among observations with similar modeled characteristics, rather than automatically pooling unlike groups. A pooled collection of counts with different means can appear overdispersed even when each group follows a Poisson distribution. The question is whether the specified Poisson model adequately describes counts after accounting for the information it includes.

Definition: In a Poisson regression, an observation $Y_i$, conditional on its modeled predictors $x_i$, has a Poisson distribution with mean $\mu_i>0$: $$ Y_i\mid x_i\sim\operatorname{Poisson}(\mu_i). $$ The model implies $E[Y_i\mid x_i]=\mu_i$ and $\operatorname{Var}(Y_i\mid x_i)=\mu_i$. The usual independent-observations assumption is that the counts are conditionally independent given their modeled predictors. A Poisson model is questionable when these conditional distributional or independence assumptions are not plausible.

The word conditional matters. The mean $\mu_i$ can differ across observations because their predictors differ. Thus, the model does not require every count in the data set to have the same mean. It requires the variance for an observation with given modeled characteristics to equal that observation’s conditional mean.

Why the Mean and Variance Are Equal

For a Poisson random variable $Y$ with mean $\mu$, the probability of observing $y$ is $P(Y=y)=e^{-\mu}\mu^y/y!$, for $y=0,1,2,\ldots$. Its mean and variance follow from its factorial moments.

Find the mean.
Using the Poisson probability mass function, $E[Y]=\sum_{y=1}^{\infty}y e^{-\mu}\mu^y/y!=\mu\sum_{y=1}^{\infty}e^{-\mu}\mu^{y-1}/(y-1)!=\mu$.

Find the second factorial moment.
$E[Y(Y-1)]=\sum_{y=2}^{\infty}y(y-1)e^{-\mu}\mu^y/y!=\mu^2\sum_{y=2}^{\infty}e^{-\mu}\mu^{y-2}/(y-2)!=\mu^2$.

Convert to the variance.
Since $Y^2=Y(Y-1)+Y$, $E[Y^2]=\mu^2+\mu$. Therefore, $\operatorname{Var}(Y)=E[Y^2]-(E[Y])^2=\mu^2+\mu-\mu^2=\mu$.

The same probability mass function gives the zero probability $P(Y=0)=e^{-\mu}$. This is a useful check: given a fitted mean, the model predicts not only a variance but also a specific proportion of zero counts. Neither check is conclusive by itself, but a substantial, systematic discrepancy can signal that the assumed count distribution needs further investigation.

Checking Variation in Comparable Counts

For a simple initial check, examine counts from observations that have similar modeled characteristics. Compare their observed spread with the spread expected from the fitted means. In an intercept-only Poisson model, the fitted mean is the sample average. A sample variance much larger than that average is a warning sign of possible extra-Poisson variation; a sample variance close to or below the average is not evidence of overdispersion in that sample. With varying fitted means, the comparison should use the observation-specific means rather than treating all counts as if they shared one mean.

A Pearson-style summary can help organize this comparison. If the model supplies fitted means $\hat{\mu}_i$, calculate the squared standardized differences and divide their sum by an appropriate residual degrees of freedom:

X^2=\sum_{i=1}^{n}\frac{(y_i-\hat{\mu}_i)^2}{\hat{\mu}_i}, \qquad \hat{\phi}=\frac{X^2}{\text{residual degrees of freedom}}.

For a well-fitting Poisson model, $\hat{\phi}$ is often expected to be around 1 as a rough diagnostic. A noticeably larger value can indicate extra-Poisson variation; it does not identify its cause or prove that the model is wrong. A small sample, an incorrect mean structure, or influential observations can affect the summary. Detailed residual and deviance diagnostics are separate tools; this calculation is an initial assessment of dispersion, not a replacement for them.

Worked Example: Comparing Sample Spread With a Poisson Mean

A fictional support desk records the number of overnight requests on eight comparable weekdays. The counts are $0,1,1,2,2,3,3,4$. Use the intercept-only Poisson model to assess whether the observed spread is an obvious sign of extra-Poisson variation.

Method. For an intercept-only Poisson model, estimate the common mean by the sample average. Then compare the sample variance with that mean. This small-sample comparison is descriptive, not a formal declaration of model fit.

Step 1: Estimate the mean. The sum of the counts is $0+1+1+2+2+3+3+4=16$, so $\hat{\mu}=16/8=2$.

Step 2: Calculate the squared deviations. From the mean 2, the squared deviations are $4,1,1,0,0,1,1,4$, whose sum is $12$.

Step 3: Calculate the sample variance. Using the sample-variance denominator $n-1=7$, $s^2=12/7=1.7143$ to four decimal places.

Step 4: Compare. The fitted Poisson variance is $\hat{\mu}=2$, while the observed sample variance is about $1.7143$. The sample variance is not larger than the fitted mean, so these eight counts do not show an obvious excess-variation signal.

Check and trap avoided. The sum of squared deviations is 12, but that is not the sample variance: it must be divided by 7. Dividing by 8 instead would give the empirical average squared deviation, not the usual sample variance. In either calculation, eight observations are too few to establish that the Poisson assumptions hold.

Extra Zeros and Unobserved Heterogeneity

If observations have different underlying means but the model ignores the reason, the pooled counts can have greater variation than their pooled mean. This is one way a Poisson model can appear overdispersed. More zeros than the fitted model predicts can also arise from such unobserved heterogeneity: some observations have very small means and are therefore especially likely to be zero, while others have larger means.

The following result explains why heterogeneity can increase variance even if each count is Poisson conditional on its own mean. Let $\Lambda$ be a random, unmodeled mean and suppose $Y\mid\Lambda\sim\operatorname{Poisson}(\Lambda)$. The law of total variance gives:

\begin{aligned} \operatorname{Var}(Y) &=E[\operatorname{Var}(Y\mid\Lambda)] +\operatorname{Var}(E[Y\mid\Lambda])\\ &=E[\Lambda]+\operatorname{Var}(\Lambda). \end{aligned}

The overall mean is $E[Y]=E[\Lambda]$. If $\Lambda$ varies across observations, then $\operatorname{Var}(\Lambda)>0$, making the overall variance greater than the overall mean. Thus, extra-Poisson variation may reflect omitted differences in means rather than a failure of every conditional Poisson distribution. The probability of zero in this mixture is $E[e^{-\Lambda}]$, not generally $e^{-E[\Lambda]}$.

Worked Example: How Ignored Groups Create Overdispersion

In a hypothetical study, half the observation periods belong to a low-activity setting with a Poisson mean of 1, and half belong to a high-activity setting with a Poisson mean of 5. Suppose a proposed model ignores which setting applies. Find the pooled mean, pooled variance, and probability of zero, then compare them with a single Poisson model using the pooled mean.

Method. Treat the conditional mean $\Lambda$ as 1 or 5 with probability $1/2$ each. Apply the law of total expectation and variance, and average the conditional zero probabilities.

Step 1: Calculate the pooled mean. $E[\Lambda]=(1+5)/2=3$. Hence the pooled mean count is 3.

Step 2: Calculate variation among the means. $E[\Lambda^2]=(1^2+5^2)/2=13$, so $\operatorname{Var}(\Lambda)=13-3^2=4$.

Step 3: Calculate the pooled variance. The conditional Poisson variance averages to $E[\Lambda]=3$. Adding the variance among the means gives $\operatorname{Var}(Y)=3+4=7$. The pooled variance 7 exceeds the pooled mean 3.

Step 4: Compare zero probabilities. The mixture probability of zero is $(e^{-1}+e^{-5})/2=(0.3679+0.0067)/2=0.1873$ to four decimal places. A single Poisson model with mean 3 predicts $e^{-3}=0.0498$ zeros to four decimal places.

Check and trap avoided. The pooled mean is indeed 3, but using $3$ as if it described every observation misses the variation between settings. That single-mean model predicts both a variance of 3 and far fewer zeros than the mixture. If the setting information were included and the conditional Poisson assumptions were reasonable within each setting, the apparent pooled overdispersion would not by itself rule out conditional Poisson modeling.

Clustering and Dependence

Counts may also be associated because multiple observations share a latent condition, location, period, or source of risk. Conditional independence may then be plausible given that shared condition, while independence fails when the condition is omitted. This dependence can raise the variance of an aggregate count and make a model that treats observations as unrelated too optimistic about how much information the data provide.

For two counts $Y_1$ and $Y_2$, the variance of their sum includes a covariance term:

\operatorname{Var}(Y_1+Y_2) =\operatorname{Var}(Y_1)+\operatorname{Var}(Y_2) +2\operatorname{Cov}(Y_1,Y_2).

If the counts are positively associated, the covariance is positive and the total varies more than it would under independence with the same marginal variances. Looking only at the overall average can therefore miss an important failure of the model’s independence assumption.

Worked Example: A Shared Condition Links Two Counts

Consider two fictional service areas observed over the same period. A shared activity level $\Lambda$ equals 1 or 3 with probability $1/2$ each. Conditional on $\Lambda$, the areas’ counts $Y_1$ and $Y_2$ are independent and each is Poisson with mean $\Lambda$. Find the marginal mean and variance of each count, their covariance, and the mean and variance of the combined count.

Method. Apply conditional expectation and variance to each count. Use conditional independence to calculate the covariance induced by the shared $\Lambda$.

Step 1: Find the mean of each count. $E[\Lambda]=(1+3)/2=2$, so $E[Y_1]=E[Y_2]=2$.

Step 2: Find the variance of each count. $E[\Lambda^2]=(1+9)/2=5$, so $\operatorname{Var}(\Lambda)=5-2^2=1$. Therefore, each marginal variance is $E[\Lambda]+\operatorname{Var}(\Lambda)=2+1=3$.

Step 3: Find the covariance. Conditional independence gives $\operatorname{Cov}(Y_1,Y_2\mid\Lambda)=0$. Their conditional means both equal $\Lambda$, so the covariance induced by the shared mean is $\operatorname{Var}(\Lambda)=1$. Thus, $\operatorname{Cov}(Y_1,Y_2)=1$.

Step 4: Find the combined count’s moments. Its mean is $2+2=4$. Its variance is $3+3+2(1)=8$, not 4.

Check and trap avoided. Treating the two counts as independent would omit the covariance and understate the combined variance by 2. The calculation illustrates why clustering or shared conditions can undermine an independence assumption even when the counts are conditionally independent once the shared condition is known.

Reading Scenario Clues Carefully

A large observed zero proportion relative to the fitted Poisson expectation is a clue, not a diagnosis. Before concluding that a special zero-generating mechanism is needed, check whether the model omitted relevant predictors, combined unlike groups, or treated related observations as independent. Similarly, a variance greater than a mean can arise from unmodeled variation in the conditional means, dependence, an incorrect mean specification, or other features of the data.

A practical assessment keeps three questions separate:

Mean structure: Do modeled characteristics plausibly explain how the conditional mean changes across observations?
Dispersion: Is the observed conditional spread broadly compatible with variance equal to the conditional mean?
Independence and zeros: Are observations plausibly conditionally independent, and is the observed number of zeros compatible with fitted means?

These checks provide evidence, not a mechanical pass-or-fail rule. In particular, comparing a raw pooled variance with a raw pooled mean is not a valid conditional diagnostic when the fitted means differ. For unequal observation windows, the mean may also need to reflect the amount of observation time; the count’s raw average alone does not address that issue.

Common Mistakes

Assuming every count must be Poisson. Nonnegative integer support makes the family a candidate; it does not establish the mean–variance or independence assumptions.
Equating excess zeros with a unique cause. More zeros than predicted can result from omitted groups or predictors, heterogeneity, dependence, or other model limitations.
Comparing pooled variance and mean without checking the means. A mixture of Poisson counts with different means can have variance greater than its overall mean.
Calling a rough dispersion summary a formal verdict. A value noticeably above 1 flags possible extra-Poisson variation; it does not reveal the cause or prove the model unusable.

Key takeaway: Poisson regression assumes that, conditional on modeled predictors, each count has variance equal to its mean and observations are conditionally independent. Excess variation, excess zeros, or clustering are reasons to investigate the model’s conditional structure—not automatic proof that a Poisson model cannot be used.

Check Your Understanding

For each problem, use the stated assumptions and keep conditional and pooled quantities distinct.

A count is modeled as Poisson with mean $1.5$. State its model-implied variance and calculate the probability of zero to four decimal places. (Answer: variance $1.5$; $P(Y=0)=0.2231$.)
Six comparable observations have counts $0,0,1,1,2,2$. Calculate their sample mean and sample variance, using denominator $n-1$, and say whether the sample variance exceeds the mean. (Answer: mean $1$; sample variance $4/5=0.8$; it does not exceed the mean.)
A fitted Poisson model gives a common mean of $0.5$ for 120 comparable observations, of which 80 are zero. Calculate the expected number of zeros and the observed zero proportion, rounding both to four decimal places. Does the observed zero count exceed expectation? (Answer: expected zeros $120e^{-0.5}=72.7837$; observed proportion $80/120=0.6667$; yes, the observed zero count exceeds expectation.)
Half of a population has conditional Poisson mean $0.5$, and half has conditional Poisson mean $2.5$. Find the pooled mean, pooled variance, and pooled probability of zero; compare the pooled variance with the mean. (Answer: pooled mean $1.5$; pooled variance $2.5$; pooled zero probability $(e^{-0.5}+e^{-2.5})/2=0.3443$; variance exceeds mean.)
Two counts each have marginal variance 3 and covariance 1. Their marginal means are both 2. Find the mean and variance of their sum, and identify the error in treating the counts as independent. (Answer: sum mean $4$, sum variance $3+3+2(1)=8$; assuming independence would omit the covariance term and give variance $6$.)

What You'll Learn