Why the Number of Opportunities Matters
A count of events does not, by itself, determine its response distribution. The key question is whether the count is a tally of successes across a fixed number of eligible opportunities. If each opportunity can produce at most one success, then the count has a natural upper bound and a binomial model may be appropriate. If events can recur without a fixed trial-by-trial maximum, a Poisson model may be appropriate. A known observation period or a known number of objects does not automatically settle the choice: what matters is how an event is generated within that unit.
This distinction applies to GLMMs as well as to models without random effects. A GLMM describes the response distribution conditional on its random effects. The count family must therefore make sense for the response under those conditions; random effects do not turn an unbounded event count into a bounded one or remove a genuine maximum.
A Poisson count \(Y_{ij}\) has conditional mean \(\lambda_{ij}(b_j)>0\) and support \(0,1,2,\ldots\). Its conditional probability mass function is \(P(Y_{ij}=y\mid b_j)=e^{-\lambda_{ij}(b_j)}\lambda_{ij}(b_j)^y/y!\). In the Poisson response model, the conditional variance equals the conditional mean.
Model Assumptions and Their Consequences
For a binomial GLMM, a common conditional model is
Here \(i\) indexes observations and \(j\) indexes clusters; \(m_{ij}\) is the known number of trials, \(x_{ij}\) and \(z_{ij}\) are covariate vectors, \(\beta\) is a fixed-effects vector, and \(b_j\) is a cluster-specific random-effects vector. Conditional on \(b_j\), the mean is \(m_{ij}p_{ij}(b_j)\), the variance is \(m_{ij}p_{ij}(b_j)\{1-p_{ij}(b_j)\}\), and no count above \(m_{ij}\) is possible. The binomial model assumes the same conditional success probability for each trial in that observation. If trial probabilities differ, simply adding the trials does not generally give a binomial count.
A common Poisson GLMM instead specifies
Conditional on the random effects, the Poisson model has mean and variance \(\lambda_{ij}(b_j)\). It allows every nonnegative integer, including values larger than a number of objects that might initially seem like a plausible maximum. A Poisson model is appropriate only if that unbounded support and its conditional mean–variance relationship are credible for the response. In settings described by event arrivals, a constant rate and independent event occurrence are common motivations; the Poisson distribution is also used as a count-response model when those assumptions are expressed directly as a distributional approximation.
In either model, the random effects represent dependence among observations in the same cluster. The distributional assumptions above are conditional on those effects. Marginally, after averaging over random effects, observations can be dependent, and the response variance need not equal the marginal mean. Thus, the Poisson conditional variance rule should not be mistaken for a claim about every marginal variance in a Poisson GLMM.
What a Fixed Opportunity Count Does—and Does Not—Tell You
If the count records successes among \(m_{ij}\) known, eligible, one-success-per-trial opportunities, its support is bounded at \(m_{ij}\). That is strong structural evidence for a binomial response, provided the conditional independence and common-probability assumptions are also reasonable. An observed count below the maximum does not make that maximum irrelevant: a sample may simply not have reached it.
Conversely, knowing how many people, machines, or sites were observed does not guarantee a binomial outcome. A machine might produce several incidents during a shift; a site might experience several events during a defined period. If the recorded count can exceed the number of units, it is not a tally of one binary success per unit. A Poisson model could be suitable if its conditional assumptions are reasonable.
For a binomial count, the Poisson distribution can be a useful approximation when \(m\) is large, \(p\) is small, and the product \(mp\) is moderate. The approximating Poisson mean is \(mp\). This approximation does not change the data-generating structure: when there are exactly \(m\) Bernoulli trials, the exact family remains binomial. It is important to distinguish an approximation used for calculation from a decision about which response family represents the process.
Worked Example: A Count From Eight Binary Opportunities
In a hypothetical quality check, each of eight components is classified as either defective or not defective. Suppose each component has an independent 0.10 probability of being defective, and let \(Y\) be the number classified as defective. The trial count and conditional probability are fixed, so \(Y\sim\operatorname{Binomial}(8,0.10)\). The maximum possible count is eight.
The probability of exactly two defects is
A Poisson approximation would use \(\lambda=mp=8(0.10)=0.8\), giving
The approximation is close here, but not exact. The exact model respects the maximum of eight; the Poisson model assigns a small positive probability to counts above eight. Since the number of independent binary opportunities is explicit, the binomial model is the direct choice.
Why the Poisson Approximation Can Work
The approximation follows by taking the binomial limit with \(m\) increasing and \(p=\lambda/m\), so that the mean \(mp=\lambda\) stays fixed. For any fixed nonnegative integer \(y\), the binomial probability is
Rewrite the combinatorial factor as \(m(m-1)\cdots(m-y+1)/y!\). As \(m\) grows with \(y\) fixed, the product of the first \(y\) factors divided by \(m^y\) tends to one. Also, \((1-\lambda/m)^m\) tends to \(e^{-\lambda}\), while \((1-\lambda/m)^{-y}\) tends to one. Therefore
The limit justifies using a Poisson distribution to approximate binomial probabilities in the large-\(m\), small-\(p\), fixed-\(mp\) regime. It does not imply that a Poisson model is always preferable for rare binary outcomes: an individual binary outcome remains binary, and a known finite set of trials still has binomial support.
Worked Example: Checking the Approximation With 100 Opportunities
Suppose a hypothetical screening procedure has 100 independent eligible opportunities, each with a 0.02 probability of success. Let \(Y\) be the number of successes. The exact distribution is \(\operatorname{Binomial}(100,0.02)\). Its mean is \(100(0.02)=2\), so compare the exact probability of one success with the Poisson approximation having \(\lambda=2\).
For the exact binomial probability,
For the approximation,
The probabilities agree to four significant figures in this calculation. The known 100 opportunities still make the exact binomial model the structural model; the Poisson result is a convenient, accurate approximation for this probability. For other values of \(m\) and \(p\), the approximation may be less accurate, so it should not be treated as an identity.
A Count That Can Repeat Within the Observation Unit
A fixed observation unit is not necessarily a fixed set of binary trials. For example, in a one-hour period, a service desk may receive several calls. If calls can occur repeatedly in the same hour, the count is not bounded at one per hour. A Poisson response may be a reasonable model if counts in comparable periods are conditionally independent and share a common rate, with no meaningful event maximum imposed by the process.
Worked Example: Repeated Events Across Equal Time Periods
In a hypothetical log, a service desk records 0, 2, and 3 calls in three separate one-hour periods. Assume for this simple calculation that the periods have the same underlying rate and are conditionally independent. Each count can exceed one, so a binomial model with one opportunity per hour cannot represent these data. Under a common Poisson model, the likelihood for rate \(\lambda\) is proportional to \(e^{-3\lambda}\lambda^{0+2+3}\), apart from factors that do not depend on \(\lambda\). Maximizing it gives the rate estimate \(\hat{\lambda}=5/3\) calls per hour.
Using that estimated mean, the fitted probability of exactly two calls in a one-hour period is
This calculation illustrates the support-based choice, not proof that the Poisson assumptions hold. The observed count of three is already incompatible with a binomial model that allows only one success per hour. A binomial model with a larger maximum would only be defensible if that maximum represented actual one-success-per-trial opportunities and the trial assumptions were plausible.
Comparing the Families and Checking a Proposed Model
| Feature | Binomial response | Poisson response |
|---|---|---|
| Natural data structure | Successes among a known number of eligible binary trials | Events counted on an unbounded nonnegative scale |
| Support | 0 through \(m_{ij}\) | 0, 1, 2, and all larger integers |
| Conditional mean | \(m_{ij}p_{ij}(b_j)\) | \(\lambda_{ij}(b_j)\) |
| Conditional variance | \(m_{ij}p_{ij}(b_j)\{1-p_{ij}(b_j)\}\) | \(\lambda_{ij}(b_j)\) |
| Key response assumption | Fixed trials, at most one success per trial, common conditional success probability, and conditional independence | Unbounded count support and a credible conditional Poisson mean–variance specification |
A practical check begins with the mechanism and support, not a goodness-of-fit statistic. Write down what one opportunity would be, if opportunities exist, and ask whether each can generate at most one recorded event. Verify that the proposed denominator includes only eligible trials and that the trials can reasonably share a probability. For a Poisson response, check that the count can genuinely exceed any proposed finite trial maximum and that the conditional Poisson variation is plausible.
When examining fitted models, investigate residual patterns and whether observed counts are systematically inconsistent with the fitted conditional distribution. A binomial model can fail if its trial probabilities differ substantially or if trials remain dependent after conditioning on the random effects. A Poisson model can fail when its conditional variance or event dependence is inconsistent with the model. In a GLMM, marginal variation alone is not enough to reject the conditional Poisson variance assumption, because random effects can add variation after averaging over clusters.
The main alternative to a binomial model in the large-\(m\), small-\(p\) case is often a Poisson approximation, not a different account of the underlying opportunities. Conversely, choosing a binomial model just because the analyst knows the number of observed units is not a valid substitute for checking whether those units are actual binary trials. Neither family is universally more flexible: the binomial encodes a real upper bound, while the Poisson encodes an unbounded count and a particular conditional variance. If neither set of assumptions is credible, the model choice requires a different response distribution or a revised description of the data-generating process.
Common Mistake: Treating Every Known Number as a Denominator
Incorrect reasoning: “There were 50 patients in the clinic, so the number of events must be binomial with 50 trials.” This is wrong if a patient can experience multiple events during the observation period, or if the recorded count is not one binary outcome per patient. In that case, 50 is a sample size, not necessarily a binomial denominator.
Correction: A binomial denominator counts eligible opportunities that each yield at most one success. If a patient can have repeated events, then patient count alone does not create that trial structure. Define the response and its generating opportunities first, then choose a family whose support and assumptions match them.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A hypothetical sensor reports whether each of 12 separate tests detected a signal. What response family fits the count of positive tests if the tests are conditionally independent and have the same success probability?
- For the situation in question 1, calculate the probability of exactly two positive tests when the success probability is 0.10.
- A hypothetical counter records 4, 1, and 2 events across three equal periods. Could a binomial model with one opportunity per period describe these data? Explain.
- For \(m=50\) opportunities and \(p=0.02\), state the Poisson approximation and its mean. What is the probability of zero events under that approximation?
Solutions. (1) The count is binomial with \(m=12\), because each test is one binary opportunity and the stated conditional independence and common-probability assumptions hold. (2) \(P(Y=2)=\binom{12}{2}(0.10)^2(0.90)^{10}=66(0.01)(0.348678)\approx 0.2301\). (3) No. A binomial model with one opportunity per period has support only at 0 and 1, whereas the recorded counts include 4 and 2. A Poisson model could be considered if its assumptions were plausible. (4) The approximation is \(\operatorname{Poisson}(\lambda)\) with \(\lambda=mp=50(0.02)=1\). Thus \(P(Y=0)\approx e^{-1}(1^0/0!)=e^{-1}\approx 0.3679\). The exact binomial model remains the structural model if those 50 opportunities are genuine binary trials.