Why Model Plausibility Comes Before Calculation
A Poisson probability calculation can be performed whenever a rate and an exposure are supplied. That does not mean the model fits the situation. Before using a Poisson distribution, ask whether the count is recorded over a well-defined exposure, whether the event rate is reasonably stable across that exposure, and whether events occur independently rather than in clusters or dependent sequences.
These checks matter in actuarial settings because a single rate can hide important differences. A count of claims per policy-year, for example, is interpretable only if policy-years are defined consistently and the mix of risks does not make the claim rate vary substantially. Such settings are hypothetical illustrations; real applications require careful examination of the underlying data and assumptions.
The rate is not just a convenient number. It represents a stable average count per unit of exposure under the model. If the model applies, changing the exposure from \(E\) to \(cE\) changes the expected count from \(\lambda E\) to \(\lambda cE\). Counts over disjoint intervals are independent under the Poisson-process formulation. A constant average rate alone does not guarantee independence.
What to Check in a Count Scenario
A practical plausibility assessment starts with the data-generating setup, then looks for patterns in the observed counts. No single diagnostic proves that a Poisson model is correct. The goal is to identify whether its key assumptions are reasonable enough for the intended use.
State exactly what qualifies as an event and what one unit of exposure means. Check that exposure is measured in the same way across observations.
Ask whether the event rate could change by time, location, risk type, operating condition, or another relevant factor. A pooled rate may conceal substantial differences.
Consider whether one event changes the chance of another, or whether shared shocks and clustering can cause events to arrive together.
For equal exposures under one Poisson rate, the model implies that the count mean and variance are both \(\lambda E\). A large, persistent excess of variation is a warning, not a stand-alone proof of a particular alternative.
The mean-variance comparison is meaningful only when the counts being compared have equal exposures and are intended to share one rate. With unequal exposures, the Poisson model predicts different means, \(\lambda E_j\), for observations with exposures \(E_j\). Comparing their raw counts as if they had equal expected values can produce a misleading diagnosis.
Why Rate Heterogeneity Creates Extra Variation
Suppose each observation has a Poisson count conditional on its own rate, but the rate varies between observations. Let \(\Lambda\) denote that varying rate for a fixed exposure of one unit, and let \(N\mid\Lambda\sim\operatorname{Poisson}(\Lambda)\). Conditional on \(\Lambda\), the count has mean and variance \(\Lambda\). The law of total variance gives:
Thus, when rates vary, the variance exceeds the mean by \(\operatorname{Var}(\Lambda)\), unless the rate is constant. This is one reason pooled counts can be more variable than a single-rate Poisson model predicts. It also explains why overdispersion—variation greater than the Poisson mean-variance pattern—can indicate unmodelled rate differences. It does not, by itself, establish that heterogeneity is the cause; dependence or data problems may also contribute.
Worked Example: Similar Counts at Equal Exposure
A hypothetical maintenance team records equipment stoppages during eight shifts of equal length and similar operating conditions. The counts are \(1,2,0,1,3,1,0,2\). Assess whether a single-rate Poisson model is plausible as a working model.
Method. First confirm that the exposure units are comparable. Then calculate the empirical mean and sample variance and compare them with the Poisson mean-variance pattern.
Each observation is one shift, so all eight observations have equal exposure. The total count is \(1+2+0+1+3+1+0+2=10\). The empirical mean count per shift is therefore \(10/8=1.25\). A fitted rate for one-shift exposure is \(\hat{\lambda}=1.25\) stoppages per shift.
The sum of squared counts is \(1^2+2^2+0^2+1^2+3^2+1^2+0^2+2^2=20\). The sample variance is \((20-10^2/8)/(8-1)=(20-12.5)/7=7.5/7=1.0714\), to four decimal places.
The sample variance, \(1.0714\), is reasonably close to the mean, \(1.25\). The observed counts range from zero to three and do not show an obvious extreme cluster. These features are consistent with a Poisson model as a provisional description. They do not prove it: eight shifts provide limited evidence, and the team should still check whether operating conditions and stoppage dependence are genuinely comparable.
As a further rough check, a Poisson distribution with mean \(1.25\) gives \(P(N=0)=0.2865\), \(P(N=1)=0.3581\), \(P(N=2)=0.2238\), and \(P(N=3)=0.0933\), to four decimals. Across eight shifts, the corresponding expected frequencies are \(2.2920\), \(2.8648\), \(1.7904\), and \(0.7464\). The actual frequencies for zero through three are \(2,3,2,1\), respectively, which are broadly compatible with those small expected counts.
Check and common mistake. The sum is 10, the sum of squares is 20, and the variance calculation uses the sample-variance denominator \(7\). A common mistake is to treat a close mean and variance as confirmation of the model. Here they support plausibility only; they cannot establish stable rates or independent stoppages.
Worked Example: A Pooled Rate Hides Regional Differences
A hypothetical roadside service records vehicle breakdowns in two regions, each observed for 1,000 vehicle-hours. Region A records 80 breakdowns and Region B records 320. Assess whether one common Poisson rate is plausible for both regions.
Method. Calculate the regional rates and compare them with the rate pooled across both regions. If one rate applied, equal exposures would have equal expected counts.
Region A's rate is \(80/1{,}000=0.08\) breakdowns per vehicle-hour. Region B's rate is \(320/1{,}000=0.32\) breakdowns per vehicle-hour. The pooled count is \(80+320=400\), and the pooled exposure is \(1{,}000+1{,}000=2{,}000\) vehicle-hours. The pooled rate is \(400/2{,}000=0.20\) breakdowns per vehicle-hour.
Under a common rate of \(0.20\), each region's expected count would be \(0.20(1{,}000)=200\). The difference between each observed regional count and that expectation is 120 breakdowns. The Poisson standard deviation for a count with mean 200 is \(\sqrt{200}\approx14.14\), so each observed count is about \(120/14.14\approx8.49\) such standard deviations from the common-rate expectation. This is a substantial discrepancy, not a small fluctuation suggested by these data.
A single pooled rate is therefore not a plausible description of both regions without further explanation. The regions may have different operating conditions or risk mixes, or the observations may have another structural difference. Separate rates or a model that accounts for relevant factors may be more suitable.
Check and common mistake. The pooled rate calculation is correct as an overall average, but it does not show that both regions have that rate. A common mistake is to use \(0.20\) per vehicle-hour as if it described each region despite the observed rates of \(0.08\) and \(0.32\). Equal exposure makes the contrast especially clear here.
Worked Example: Repeated Clusters at Equal Exposure
A hypothetical monitoring program observes 100 sites for the same period and records the number of equipment failures at each site. Ninety sites have zero failures, while ten sites each have five. Assess a single-rate Poisson model and identify the main concern.
Method. Calculate the mean and sample variance, then compare the observed frequency of five failures with the frequency expected under a Poisson model fitted to the mean.
The total number of failures is \(90(0)+10(5)=50\), so the mean is \(50/100=0.5\) failures per site-period. The sum of squared counts is \(90(0^2)+10(5^2)=250\). The sample variance is \((250-50^2/100)/(100-1)=(250-25)/99=225/99\approx2.2727\), to four decimal places. This is much greater than the mean \(0.5\).
Under a Poisson model fitted with mean \(0.5\), the probability of at least five failures at one site is \(1-\sum_{k=0}^{4}e^{-0.5}(0.5)^k/k!\approx0.0001721\). Across 100 sites, the expected number of sites with at least five failures would be approximately \(100(0.0001721)=0.0172\). The observed number is ten. The large variance and concentration of failures at a small set of sites strongly challenge a single-rate Poisson description.
The pattern could reflect persistent site-to-site differences, dependence among failures at the same site, or a common shock affecting groups of sites. These explanations are not interchangeable, and the counts alone do not identify which one applies. Each would require checking the operating context.
Check and common mistake. The fitted mean is based on all 100 sites, including the ten high-count sites. A common mistake is to see that the total count is 50 and conclude that a Poisson model with mean 0.5 must fit. Matching the mean is not enough: the empirical variance and the concentration of high counts reveal serious lack of fit.
Interpreting the Evidence Carefully
A Poisson model is more credible when each observation has a clearly defined and comparable exposure, the rate is stable over the relevant time and population, and event occurrences do not materially trigger or prevent other events. The observed count variation should also be broadly consistent with the fitted model. These are contextual judgments as well as numerical checks.
The following warning signs deserve attention:
- Changing exposure definitions: one observation may represent a day while another represents a month, or recorded operating time may omit downtime.
- Rate differences: locations, time periods, or risk classes may have systematically different event rates.
- Clustering or inhibition: an event may make another event more likely, as with cascading failures, or less likely, as when an event removes an item from risk.
- Excess variability: equal-exposure counts may vary much more than their mean, suggesting heterogeneity, dependence, or a recording issue.
A common exam trap is to decide that a Poisson model applies merely because the quantity is a count and an average rate is available. The correct assessment is conditional: a count and a rate permit a Poisson calculation only if the model assumptions are reasonable for that setting. Conversely, a small discrepancy between sample mean and variance does not automatically invalidate the model; finite samples naturally fluctuate.
Check Your Understanding
For each setting, assess the Poisson assumptions rather than merely selecting a distribution from the fact that the data are counts.
- A hypothetical call center records incoming calls during six equal one-hour periods under the same staffing and routing rules. Counts are \(4,3,5,2,4,6\). Find the empirical mean and sample variance, then state whether these summaries alone rule out a Poisson model. (Answer: Mean \(=4\), sample variance \(=2\); the variance is lower than the mean, but these six observations alone do not rule out a Poisson model.)
- A hypothetical insurer observes 60 claims over 12,000 policy-years. The policies are equally exposed in time, but half are commercial and half personal, and the claim rates may differ by type. Find the pooled rate and identify what additional check is needed before using one rate for all policies. (Answer: Pooled rate \(=60/12{,}000=0.005\) claims per policy-year; compare the claim experience by policy type before assuming a common rate.)
- A hypothetical warehouse records 24 damaged parcels over 3,000 parcel-handling hours in one area and 96 over 3,000 hours in another. Calculate each rate and the pooled rate, then assess a common-rate model. (Answer: Rates \(=0.008\) and \(0.032\) damages per handling hour; pooled rate \(=120/6{,}000=0.020\); the fourfold regional rate contrast makes one common rate questionable.)
- In 80 equal-exposure hypothetical test batches, 72 batches have zero defects and eight batches have four defects. Calculate the empirical mean and sample variance, and state the main warning. (Answer: Mean \(=32/80=0.4\); sample variance \(=(128-32^2/80)/79=115.2/79\approx1.4582\); the substantial excess variation and concentration suggest heterogeneity or clustering.)
- A hypothetical fleet records one breakdown per vehicle on average over a year. A breakdown causes the vehicle to be removed from service for the rest of that year, so a second breakdown on that vehicle cannot occur during the same observation period. Explain which assumption needs particular scrutiny. (Answer: Independence and the event mechanism need scrutiny, because the first breakdown changes the vehicle's subsequent opportunity to experience another breakdown.)