Zero Exposure And Boundary Cases

Why the exposure boundary matters

A rate is a count divided by an amount of opportunity: events per person-year, failures per machine-hour, or cases per population at risk. When the exposure is zero, the observed count divided by exposure is undefined. But the record may still have a clear likelihood contribution. When exposure is positive but no events occur, the observed rate is zero, yet the record is not equivalent to a zero-exposure record. Keeping these cases separate is essential for fitting and interpreting a Poisson rate model.

Let \(Y_i\) denote a count and \(t_i\) its exposure, with \(t_i\geq 0\). In a Poisson rate model, the mean count is the exposure multiplied by the rate. Write the rate as \(\lambda_i=\exp(x_i^\mathsf{T}\beta)\), where \(x_i\) is the predictor vector and \(\beta\) is a finite parameter vector. Thus the mean is \(\mu_i=t_i\lambda_i\). The exponential link ensures a positive finite rate for finite \(x_i^\mathsf{T}\beta\); the exposure enters as an offset when \(t_i>0\).

The distinction is not merely arithmetic. Under the Poisson model with a finite rate, zero exposure forces the count to be zero with probability one. By contrast, zero events during positive exposure are possible under many positive rates, and their probability depends on the rate.

What the likelihood says at zero exposure

For a positive exposure \(t_i\), the probability mass function is

At \(t_i=0\), the mean is \(\mu_i=0\), and the Poisson distribution with mean zero is degenerate: \(P(Y_i=0)=1\), while \(P(Y_i=y_i)=0\) for every positive integer \(y_i\). This gives a precise interpretation of the boundary case.

For \(t_i=0,Y_i=0\), the contribution is the probability of the only possible count, namely one. It therefore adds zero to the log-likelihood and no score or curvature in \(\beta\). The log-offset expression \(\log(t_i)+x_i^\mathsf{T}\beta\) cannot be evaluated at zero exposure; this does not mean the likelihood is undefined. The probability model itself is well-defined at the boundary.

For \(t_i=0,Y_i>0\), it is not appropriate to insert a small positive exposure merely to make the offset computable. The recorded event and exposure must first be checked: perhaps exposure was measured over the wrong interval, was rounded to zero, or the event was assigned to the wrong record. If the data are correct, the assumed model needs reconsideration. Quietly changing zero to a small number changes the model and can produce an arbitrary rate contribution.

Zero events during positive exposure

Now suppose \(t_i>0\) but \(Y_i=0\). The rate \(Y_i/t_i\) is defined and equals zero, but the likelihood contribution is not one. Substituting \(y_i=0\) in the Poisson mass function gives

This is positive for every finite positive rate and depends on that rate. An event-free record can therefore provide evidence about the rate: it is less likely under larger rates, all else equal. Its contribution is not the same as an observation whose exposure is zero.

The parameter dependence can also be seen in a log-linear model. For \(t_i>0\), the log-likelihood contribution, score, and observed information matrix are

To verify the result, set \(Y_i=0\) in the log-likelihood. The factorial term is \(\log(0!)=0\), and the count multiplied by \(\log(t_i)+x_i^\mathsf{T}\beta\) vanishes, leaving \(\ell_i(\beta)=-t_i\exp(x_i^\mathsf{T}\beta)\). Differentiation gives the stated score. Differentiating that score once more gives \(-t_i\exp(x_i^\mathsf{T}\beta)x_ix_i^\mathsf{T}\), so its negative is the observed information. For finite \(\beta\) and positive \(t_i\), the scalar multiplier is positive. \(\square\)

Why very small exposures produce unstable crude rates

When \(t_i\) is positive but small, the offset \(\log(t_i)\) is finite and the likelihood model remains valid. The difficulty is that the crude rate \(Y_i/t_i\) can change sharply with a small change in the count. One event over \(0.01\) person-years gives a crude rate of \(100\) events per person-year; zero events over the same exposure gives a crude rate of zero. Neither number alone describes how uncertain the underlying rate is.

The variability of a crude rate follows directly from the Poisson mean and variance. For fixed \(t>0\) and rate \(\lambda>0\), if \(Y\sim\operatorname{Poisson}(t\lambda)\), then

The standard deviation of the crude rate is \(\sqrt{\lambda/t}\), and its standard deviation relative to its mean is \(1/\sqrt{t\lambda}\). As exposure decreases, both the variance and relative variability increase. This derivation assumes a fixed positive exposure and the stated Poisson model; it does not say that every small-exposure estimate must be wrong, but it explains why it is typically imprecise.

Practical handling and common pitfalls

A sound analysis begins by separating three situations rather than treating every zero alike:

A common mistake is to say, “The observed rate is zero, so this record contributes nothing.” That reasoning conflates a zero count with zero exposure. The correction is to check the denominator: with positive exposure, a zero count has a rate of zero and likelihood \(\exp(-t\lambda)\), which depends on the model rate. Only the zero-exposure, zero-count case has a parameter-free likelihood contribution.

Another mistake is to discard every zero count before fitting a rate model. That removes valid evidence and can bias the analysis by retaining only records in which events occurred. Conversely, retaining a positive count with zero exposure without investigation asks the model to explain an impossible observation. A third mistake is to report \(Y_i/t_i\) as a stable estimate for every record. For small \(t_i\), the arithmetic is valid but the estimate can be highly variable; interpret it in light of the exposure and the model, rather than as a precise rate.

In practice, verify that exposure is genuinely zero rather than missing, rounded, or measured on a different time scale. Keep the count and exposure definitions aligned to the same record and observation window. For a zero exposure and zero count, no log-offset row can be formed, but omitting that row from parameter estimation does not alter the likelihood because its contribution is constant. Retaining it in data-quality summaries can still be useful. For positive exposure, including a very small value, use the specified exposure in the likelihood rather than replacing it with an arbitrary threshold.