Method-of-Moments Estimation for Poisson Counts

Why the Sample Mean Estimates a Poisson Parameter

Poisson models are commonly used for counts over a fixed period or within a fixed unit of exposure. For example, a hypothetical insurer might model the number of claims reported by each policy during one year. The parameter $\lambda$ is both the expected count and the count variance under the model. This equality is useful when estimating the parameter, and it also gives a way to check whether the data look compatible with a Poisson model.

Definition: Suppose $X_1,\ldots,X_n$ are independent and identically distributed Poisson random variables with common parameter $\lambda>0$. For each $i$, $X_i$ is a nonnegative integer count with $$ P(X_i=x)=\frac{e^{-\lambda}\lambda^x}{x!},\qquad x=0,1,2,\ldots, $$ and $E[X_i]=\operatorname{Var}(X_i)=\lambda$. The method-of-moments estimate based on the first population moment is $\hat{\lambda}_{\mathrm{MM}}=\bar{X}$, where $\bar{X}=\frac{1}{n}\sum_{i=1}^n X_i$.

The assumptions matter. The observations must be counts measured on comparable units, such as the same one-year period for every policy, and the common-parameter model assumes each unit has the same Poisson mean. Independence and a common mean are needed for the usual sampling properties and for interpreting the sample variance as a diagnostic of Poisson equidispersion. If exposure differs between observations, an unadjusted comparison of the count variance with the count mean may be misleading.

Deriving the Method-of-Moments Estimate

The method of moments equates a population moment implied by the model to the corresponding empirical moment. Here the model has one unknown parameter, and its first raw population moment is $E[X_i]=\lambda$. The empirical first raw moment is the sample mean. Equating them gives the estimate directly.

\bar{X}=\frac{1}{n}\sum_{i=1}^n X_i \quad\text{and}\quad E[X_i]=\lambda \quad\Longrightarrow\quad \hat{\lambda}_{\mathrm{MM}}=\bar{X}.

This is not just a convenient rule: under the assumed model, the sample mean targets the parameter because its expectation is the population mean. In fact, independence also gives its variance. Since each observation has variance $\lambda$,

E[\bar{X}]=\frac{1}{n}\sum_{i=1}^n E[X_i]=\lambda, \qquad \operatorname{Var}(\bar{X}) =\frac{1}{n^2}\sum_{i=1}^n\operatorname{Var}(X_i) =\frac{\lambda}{n}.

Thus $\bar{X}$ is unbiased under the model, and its variance decreases as the sample size increases. Substituting $\hat{\lambda}_{\mathrm{MM}}$ for the unknown $\lambda$ gives the estimated standard error $\sqrt{\bar{X}/n}$. This substitution summarizes sampling variability under the Poisson assumptions; it does not establish that those assumptions fit the data.

Using the Sample Variance as a Diagnostic

The Poisson model requires the population variance to equal the population mean. A sample will not generally have exactly equal sample mean and sample variance: both fluctuate from one sample to another. Nevertheless, a large discrepancy can alert the analyst to investigate the model, data definitions, or exposure structure.

Diagnostic: For $n>1$, define the usual sample variance by $$ s^2=\frac{1}{n-1}\sum_{i=1}^n(X_i-\bar{X})^2. $$ Under independent, identically distributed Poisson observations, both $\bar{X}$ and $s^2$ estimate the same population quantity $\lambda$. Comparing them is a check of the Poisson mean-variance relationship, not a way to estimate two separate model parameters.

The sample variance is calculated from the same observations as the sample mean, so the two summaries are not independent. A variance noticeably above the mean is sometimes called overdispersion relative to the Poisson model; a variance noticeably below it is underdispersion. Either pattern is a signal to examine, not proof by itself that a particular alternative model is correct. Small differences are expected in finite samples.

There is a subtle distinction between using the variance as a diagnostic and constructing another moment equation. Since the Poisson variance is also $\lambda$, one could equate it to a sample variance and obtain a variance-based estimate. But this does not introduce a second unknown parameter, and it is not an independent estimate: it uses the same data and may differ from the sample-mean estimate because of sampling variation. For the one-parameter Poisson method-of-moments calculation in this tutorial, use the first moment to estimate $\lambda$, then use the variance comparison to assess the model.

Worked Example: Estimating a Mean Claim Count

A hypothetical portfolio study records the number of claims in one year for six comparable policies. The observed counts are $0,1,1,2,2,3$. Assume the counts are independent and identically distributed as $\operatorname{Poisson}(\lambda)$. Estimate $\lambda$, calculate the sample variance, and interpret the comparison.

Method. Use the first population moment for the method-of-moments estimate, and use the usual $n-1$ sample variance as a model diagnostic.

Step 1: Calculate the mean. The total count is $0+1+1+2+2+3=9$, and $n=6$. Therefore,

\hat{\lambda}_{\mathrm{MM}}=\bar{X}=\frac{9}{6}=1.5.

Step 2: Calculate the sample variance. The sum of squared observations is $0^2+1^2+1^2+2^2+2^2+3^2=19$. The sum of squared deviations can be calculated as $\sum X_i^2-n\bar{X}^2=19-6(1.5)^2=19-13.5=5.5$. Hence

s^2=\frac{5.5}{6-1}=1.1.

Step 3: Interpret and check. The fitted Poisson mean is $1.5$, while the observed sample variance is $1.1$. The sample variance is lower, but the difference alone does not show that the Poisson model is wrong. Under the stated model, the same population quantity, $\lambda$, governs both; finite samples need not produce identical summaries. The estimate is a count per policy-year.

Common mistake avoided: Replacing the mean-based estimate $1.5$ with $1.1$ because the sample variance is another number related to $\lambda$. The requested first-moment method-of-moments estimate is $1.5$; the variance comparison is diagnostic.

Calculations From Grouped Counts

Count observations are often summarized by the number of units having each possible count. If $f_x$ units have count $x$, then the sample size is $n=\sum_x f_x$, the total count is $\sum_x xf_x$, and the sum of squared observations is $\sum_x x^2f_x$. These quantities are enough to calculate the sample mean and variance without expanding every observation in a list.

A useful identity follows by expanding each squared deviation:

\sum_{i=1}^n(X_i-\bar{X})^2 =\sum_{i=1}^nX_i^2-n\bar{X}^2, \qquad s^2=\frac{\sum_{i=1}^nX_i^2-n\bar{X}^2}{n-1}.

Worked Example: Estimating From a Frequency Table

A hypothetical survey records annual service calls for ten identical devices. Four devices had zero calls, three had one call, two had two calls, and one had three calls. Assume independent $\operatorname{Poisson}(\lambda)$ counts. Find the method-of-moments estimate and the sample-variance diagnostic.

Method. Use the frequency table to compute the total count and the sum of squared counts, then apply the first-moment estimate and the sample-variance identity.

Step 1: Count the observations and total calls. The number of devices is $4+3+2+1=10$. The total calls are $0(4)+1(3)+2(2)+3(1)=0+3+4+3=10$. Thus

\hat{\lambda}_{\mathrm{MM}}=\bar{X}=\frac{10}{10}=1.

Step 2: Find the sum of squared observations. It is $0^2(4)+1^2(3)+2^2(2)+3^2(1)=0+3+8+9=20$. The squared-deviation sum is therefore $20-10(1)^2=10$, giving

s^2=\frac{10}{10-1}=\frac{10}{9}=1.1111\ldots.

Step 3: Interpret and check. The fitted mean is $1$ call per device-year and the sample variance is about $1.1111$ calls-squared. They are close, so this particular comparison does not show a large discrepancy from the Poisson mean-variance relationship. It is still only a diagnostic, not a proof of model adequacy.

Common mistake avoided: Dividing the squared-deviation sum by $n$ when reporting the usual sample variance. The denominator here is $n-1=9$. If a problem explicitly defines a variance with divisor $n$, follow that definition and identify it; do not silently switch conventions.

When the Comparison Does Not Apply Directly

The equality of mean and variance concerns observations with one common Poisson parameter. If units have different exposure periods, a reasonable model may instead be $X_i\sim\operatorname{Poisson}(\lambda e_i)$, where $e_i$ is the exposure and $\lambda$ is a rate per exposure unit. Then the expected count differs across units. The raw sample mean estimates an average count for the particular set of exposures, not the common rate $\lambda$, and comparing the raw count variance with that mean is not the same simple diagnostic.

For independent counts with known exposures under that model, the total count has expectation $\lambda\sum_i e_i$. Equating the observed total count to that expectation gives the rate estimate $\sum_i X_i/\sum_i e_i$. This exposure-adjusted calculation is appropriate only when the stated model and exposure information support it; it is not a substitute for the common-parameter, equal-unit analysis above.

Worked Example: Avoiding an Unequal-Exposure Error

A hypothetical claims study observes four policies with exposure amounts $0.5,1,1.5,$ and $2$ policy-years, and counts $0,1,2,$ and $5$, respectively. Assume independent counts with $X_i\sim\operatorname{Poisson}(\lambda e_i)$, where $\lambda$ is the annual claim rate. Estimate $\lambda$. Explain why the unweighted sample mean is not the rate estimate.

Method. Match the total observed count to its expectation under the exposure-specific Poisson means.

Step 1: Add the counts and exposures. The total count is $0+1+2+5=8$. Total exposure is $0.5+1+1.5+2=5$ policy-years. Since $E[\sum_iX_i]=\sum_i\lambda e_i=\lambda\sum_i e_i$, the moment equation is $8=5\lambda$.

\hat{\lambda}_{\mathrm{MM}}=\frac{8}{5}=1.6 \quad\text{claims per policy-year}.

Step 2: Check the tempting alternative. The unweighted sample mean count is $8/4=2$ claims per observed policy. That describes the average count across these four observations, which have unequal exposure; it is not the exposure-adjusted annual rate. The rate estimate $1.6$ uses all $5$ policy-years in the denominator.

Step 3: State the diagnostic limitation. The individual means in this example are $\lambda e_i$, not a common $\lambda$. Therefore, comparing the sample variance of these four raw counts directly with their raw sample mean is not the common-parameter Poisson diagnostic. The exposure structure must be accounted for before assessing fit.

Common mistake avoided: Dividing the total count by the number of policies and reporting $2$ as an annual rate. The correct rate denominator is total exposure, so the estimate is $1.6$ claims per policy-year.

Exam Traps and Practical Checks

A reliable solution states the data unit and model before calculating. For independent, identically distributed Poisson counts with a common parameter, the first-moment estimate is the arithmetic sample mean. The sample variance is then compared with that estimate to assess the model’s equidispersion implication. Keep these roles separate.

Using the variance as the requested estimate: If the question asks for the first-moment method-of-moments estimate, calculate $\bar{X}$. Do not substitute $s^2$ just because the Poisson variance is $\lambda$.
Calling the two summaries independent estimates: They are calculated from the same observations. A discrepancy is useful information, but neither summary is an independent replication of the data.
Mixing variance conventions: The usual sample variance uses $n-1$; a defined empirical central moment may use $n$. Read the stated convention and show the denominator.
Ignoring exposure differences: If expected counts vary by exposure, an unweighted sample mean is not generally the rate per exposure unit, and the common-mean variance diagnostic does not apply directly.
Overinterpreting a small sample: Sample mean and variance need not match exactly even when the Poisson model is correct. A discrepancy prompts investigation; it is not by itself a formal test conclusion.

Key takeaway: For independent, identically distributed $\operatorname{Poisson}(\lambda)$ counts, the first-moment method-of-moments estimate is $\bar{X}$. Compare $s^2$ with $\bar{X}$ to examine the Poisson mean-variance assumption, while remembering that the comparison is based on the same sample and is not a second independent parameter estimate.

Check Your Understanding

For each problem, state the model assumption used and keep the parameter estimate distinct from the variance diagnostic.

A hypothetical sample of five comparable one-year policy counts is $1,0,2,3,1$. Under an independent common-parameter Poisson model, find the first-moment estimate. (Answer: $\hat{\lambda}_{\mathrm{MM}}=1.4$ claims per policy-year.)
For counts $0,2,2,1,3$, calculate the sample mean and the usual sample variance using denominator $n-1$. (Answer: $\bar{X}=1.6$, $s^2=1.3$.)
A frequency table for eight comparable units has two units with count zero, three with count one, two with count two, and one with count three. Find the method-of-moments estimate and the usual sample variance. (Answer: $\hat{\lambda}_{\mathrm{MM}}=1.25$, $s^2=1.0714$.)
Three observations have exposures $1,2,$ and $3$ years, with counts $1,3,$ and $4$. Under independent $X_i\sim\operatorname{Poisson}(\lambda e_i)$ counts, estimate the annual rate. (Answer: $\hat{\lambda}=8/6=1.3333$ counts per exposure-year.)
Eight independent, comparable hypothetical units have counts $0,0,1,1,4,4,4,2$. Find the first-moment estimate and usual sample variance, then describe the direction of the diagnostic comparison. (Answer: $\hat{\lambda}_{\mathrm{MM}}=2$, $s^2=22/7=3.1429$; the sample variance is above the fitted mean, indicating a possible overdispersion signal.)

What You'll Learn