Zero Exposure And Offset Data Checks

Why Exposure Checks Matter

A frequency model uses earned exposure to describe how long, or how many insured units, were at risk. In a log-link model, exposure enters as an offset. That makes the validity of each exposure value a modeling requirement, not merely a formatting preference: the logarithm of zero is undefined, and the logarithm of a negative number is not real-valued.

A record with zero exposure can still be a useful data-quality signal. It may represent a legitimate record for a period before coverage began, or it may reveal a timing, coding, or aggregation problem. The correct treatment depends on what the record means. Automatically deleting it, changing it to a small positive value, or retaining a positive claim count without investigation can each distort the analysis.

Definition: Let $e_i$ be the earned exposure for record $i$, measured in a consistent, nonnegative unit, and let $N_i$ be its claim count. In a log-link frequency model, a typical specification is $$ \log E[N_i\mid \boldsymbol{x}_i] =\log(e_i)+\boldsymbol{x}_i^\mathsf{T}\boldsymbol{\beta}, \qquad e_i>0. $$ Equivalently, the fitted mean is $E[N_i\mid \boldsymbol{x}_i]=e_i\exp(\boldsymbol{x}_i^\mathsf{T}\boldsymbol{\beta})$. The term $\log(e_i)$ is a known offset, so its coefficient is fixed at one. This specification requires strictly positive exposure for every record included in the log-offset fit.

For example, if exposure is measured in policy-years, a value of $0.25$ represents three months of earned coverage under the exposure definition being used. A value of zero means no earned exposure in that measurement period. It does not mean a very small amount of exposure. A negative value is not a valid amount of earned exposure; it often signals a reversal, sign convention, or data-processing error that must be understood before modeling.

Classifying Exposure Records Before Fitting

First check the data as a set of records, not just the offset column in isolation. Verify that exposure uses the intended unit and period, that it is earned rather than written exposure if the model calls for earned exposure, and that counts and exposure refer to compatible populations and time windows. Then classify the records:

Positive exposure: eligible for a log-offset model, subject to the other model and data checks.
Zero exposure and zero claims: the record contains no observed at-risk time and no claims. It may have no likelihood information for the frequency rate, but it should still be reviewed and cannot be passed unchanged to software that computes $\log(e_i)$.
Zero exposure and positive claims: inconsistent with the ordinary exposure-based count model. Investigate attribution, timing, exposure calculation, and record linkage before fitting.
Negative exposure: invalid for the usual exposure interpretation and log offset. Correct the source or exclude the record from the fit only under a documented data-resolution rule.

A record with no earned exposure and no claims is not equivalent to a record with a small positive exposure and no claims. The latter contributes information about the rate; the former does not. Likewise, a positive count attached to zero exposure cannot be made coherent by choosing a tiny replacement exposure. That substitution invents at-risk time and can create an extreme implied frequency.

Confirm meaning and units.
Check the exposure definition, earning period, and unit. A unit conversion should preserve the underlying rate interpretation.

Flag nonpositive values.
Separate zero exposure from negative exposure; do not take logarithms or silently change either value.

Reconcile counts and exposure.
For zero exposure, check whether the claim count is also zero and whether both fields describe the same period and insured risk.

Fit only after resolving exceptions.
Use eligible positive-exposure records in the offset fit, document any exclusions, and retain an audit trail for records that were corrected or investigated.

Rate Estimation and Its Boundary Case

The edge case is easiest to see in a constant-rate Poisson model. Assume the counts on the eligible records are independent, with $N_i\sim\operatorname{Poisson}(e_i\lambda)$, where each $e_i>0$ and $\lambda$ is a common claim rate per exposure unit. The model gives $P(N_i=n)=\exp(-e_i\lambda)(e_i\lambda)^n/n!$. This simple model is useful for explaining the offset data check; it is not a substitute for a more detailed model when the rate varies across risks.

Result: Write $E=\sum_i e_i$ and $C=\sum_i N_i$, summing only over valid positive-exposure records. If $E>0$ and $C>0$, the constant-rate Poisson model has the interior maximum-likelihood estimate $\widehat{\lambda}=C/E$. If $E>0$ and $C=0$, that interior formula does not apply: the likelihood increases as $\lambda$ approaches zero. The maximum is at $\lambda=0$ if zero is allowed; if the parameter space requires $\lambda>0$, the likelihood has a supremum at zero but no attained maximum.

To justify the result, multiply the independent Poisson likelihood contributions. Apart from factors not involving $\lambda$, the likelihood is proportional to $\lambda^C\exp(-E\lambda)$. For $C>0$, its log-likelihood derivative is $C/\lambda-E$; setting it equal to zero gives $\widehat{\lambda}=C/E$. Its second derivative is $-C/\lambda^2<0$, confirming a maximum. If $C=0$, the log-likelihood is $-E\lambda$, whose derivative is $-E<0$ for every $\lambda>0$. There is no positive stationary point. The likelihood is $\exp(-E\lambda)$, which tends to one as $\lambda$ decreases to zero, so the boundary conclusion follows.

Worked Example: Separating Valid Exposure From Bad Records

A hypothetical portfolio extract has four policy-period records: A has exposure $0.50$ policy-years and $1$ claim; B has exposure $1.25$ policy-years and $2$ claims; C has exposure $0$ and $0$ claims; D has exposure $-0.25$ policy-years and $0$ claims. Identify the records eligible for a constant-rate log-offset fit, calculate the rate estimate from eligible records, and state what should happen to C and D.

Method. Validate the exposure values before summing. A and B have positive exposure, so they are eligible for this fit. C has zero exposure and no claims, while D has negative exposure; neither should be passed into the log-offset fit as recorded.

The valid claim total and exposure total are $$ C=1+2=3,\qquad E=0.50+1.25=1.75\text{ policy-years}. $$ Both totals are positive, so the interior estimate applies: $$ \widehat{\lambda}=\frac{C}{E}=\frac{3}{1.75}=1.714285\ldots $$ Thus the estimate is $1.714$ claims per policy-year, rounded to three decimal places.

Record C contributes no at-risk time or claims and therefore no information about this rate in the simple Poisson likelihood. It should be retained for audit and omitted from software input if the software requires a logarithm of exposure. Record D is not a negative amount of risk time; its source must be investigated and corrected or handled under a documented rule. Common mistake avoided: summing the negative exposure with the valid exposure would give $1.50$ policy-years and the incorrect calculation $3/1.50=2$. A negative value is not a valid denominator contribution.

As a check, multiplying the fitted rate by the valid total exposure gives $1.714285\ldots\times1.75=3$, the observed claim total, as required by the constant-rate estimate.

Worked Example: A Claim Count With Zero Earned Exposure

A hypothetical extract contains record P with $1.00$ policy-year of exposure and $1$ claim, and record Q with zero earned exposure and $1$ claim. An analyst proposes replacing Q’s exposure by $0.01$ policy-years so the offset can be computed. Explain why this is not a valid repair, calculate the estimate the substitution would produce, and state the appropriate next step.

Method. Check whether the count and exposure refer to the same risk and period before calculating a rate. Record Q’s combination of zero earned exposure and a positive claim count is inconsistent with the stated exposure-based model: when $e_i=0$, its Poisson mean $e_i\lambda$ is zero, and a positive count has probability zero under that model.

The proposed substitution would produce total exposure $1.00+0.01=1.01$ policy-years and total count $1+1=2$. Its apparent rate would be $$ \frac{2}{1.01}=1.980198\ldots $$ or $1.980$ claims per policy-year to three decimal places. That number is not a defensible estimate from the stated data because the extra $0.01$ policy-years were invented, not earned.

The analyst should investigate Q’s claim date and coverage period, the earning calculation, and whether the count was linked to the correct policy-period record. If the claim belongs to an exposed period, correct the record using verified data and then refit. If it does not, apply the appropriate documented attribution rule. Until the mismatch is resolved, do not use Q as a positive-exposure observation. Common mistake avoided: replacing zero with an arbitrary small value makes the offset computable but changes the data and can materially change the fitted rate.

For comparison only, record P by itself gives $1/1.00=1$ claim per policy-year. This is not automatically the final estimate for the two-record extract; Q must first be reconciled.

Worked Example: All Eligible Records Have Zero Claims

A hypothetical book has three records with exposures $0.40$, $0.60$, and $1.00$ policy-years. Each has zero claims. Under the constant-rate Poisson model, determine whether the usual ratio formula gives a positive interior estimate, and describe the maximum-likelihood result under two parameter-space conventions.

Method. Sum the positive exposures and the counts: $$ E=0.40+0.60+1.00=2.00\text{ policy-years},\qquad C=0+0+0=0. $$ Since the total count is zero, the formula $C/E$ gives zero, not a positive interior estimate. The likelihood, as a function of $\lambda$, is $$ L(\lambda)=\exp(-2\lambda),\qquad \lambda>0. $$ At $\lambda=0.50$, $L=\exp(-1)=0.3679$; at $\lambda=0.10$, $L=\exp(-0.2)=0.8187$. As $\lambda$ decreases toward zero, the likelihood increases toward $1$.

If the model allows $\lambda=0$, the maximum-likelihood estimate is the boundary value $\widehat{\lambda}=0$. If its parameter space requires $\lambda>0$, no maximum is attained; the likelihood has a supremum of $1$ as $\lambda$ tends to zero. The calculation $\widehat{\lambda}=C/E=0$ is therefore a boundary conclusion, not an interior stationary point. Common mistake avoided: claiming the positive-rate derivative has a zero when $C=0$; in this case $d\ell/d\lambda=-E=-2<0$ for every $\lambda>0$.

As a check, the zero-rate boundary assigns probability one to zero counts on these records, consistent with the limiting likelihood. The conclusion is specific to this intercept-only constant-rate model; with additional parameters, the behavior at a boundary can depend on the full model.

Common Data and Calculation Traps

The most consequential errors often occur before the model is fitted. A software error from $\log(0)$ is visible; a silent data repair may not be. Use the following distinctions when reviewing results:

Do not confuse zero with small positive exposure. Zero contributes no at-risk time; a small positive amount contributes a correspondingly small amount of information.
Do not retain a positive count at zero exposure without reconciliation. Under the stated model, its probability is zero, so the conflict indicates a data or model-definition issue.
Do not interpret a negative exposure as a rate adjustment. Find out whether it is a reversal, an error, or a different accounting quantity; it is not valid earned time at risk.
Do not apply the ratio estimate without checking its conditions. The interior estimate $C/E$ requires $E>0$ and $C>0$. With $C=0$, handle the boundary explicitly; with no positive exposure, there is no rate estimate from these records.
Do not remove flagged records without a rule. Preserve the original values, the reason for correction or exclusion, and the resulting fitted-data population so the analysis can be reproduced.

These checks matter because exposure defines the scale on which a frequency rate is measured. An offset cannot repair a mismatch between claims and exposure, and a valid-looking coefficient does not prove that every input record was coherent. Sound practice is to resolve the meaning of nonpositive exposure first, then fit and interpret the model using the exposure base that the data actually support.

Check Your Understanding

For each problem, decide what the record or model permits before calculating a rate.

A record has $0.75$ policy-years of earned exposure and zero claims. Is it eligible for a log-offset fit, and what is its exposure status? (Answer: It is eligible because exposure is positive; it contributes $0.75$ policy-years and zero claims.)
A record has zero earned exposure and $2$ claims assigned to the same period. What is the appropriate first action under an ordinary exposure-based Poisson frequency model? (Answer: Flag and reconcile the count-exposure mismatch; do not fit it as recorded or replace zero with an arbitrary positive exposure.)
Two valid records have exposures $0.25$ and $0.75$ policy-years, with $0$ and $3$ claims, respectively. Under a constant-rate Poisson model, what is the rate estimate? (Answer: $3/(0.25+0.75)=3$ claims per policy-year.)
Three positive-exposure records total $4.00$ policy-years and have zero claims. Does $C/E$ give an interior estimate? State the likelihood conclusion when $\lambda>0$, and when $\lambda\geq0$. (Answer: No; $C/E=0$ is not an interior estimate. For $\lambda>0$, the likelihood has no attained maximum and its supremum occurs as $\lambda\to0$; for $\lambda\geq0$, the maximum is at $\widehat{\lambda}=0$.)
A data extract contains one record with exposure $-0.10$ and zero claims, plus another with exposure $1.10$ and $2$ claims. Why is $2/1.00$ not justified as the rate from these two records? (Answer: The negative exposure is invalid and cannot be combined with the valid $1.10$ policy-years; investigate or handle the negative record under a documented rule, then use only resolved valid data.)

What You'll Learn