Why Exposure Matters for Frequency Estimates
A raw claim count is difficult to compare across groups if those groups were observed for different lengths of time or contained different numbers of insured risks. Forty claims during 200 policy-years and forty claims during 400 policy-years are not evidence of the same claim frequency. Dividing the count by exposure puts the estimate on a common basis; quantifying its sampling uncertainty indicates how much that estimate might vary across repeated observations.
This tutorial considers uncertainty in an exposure-adjusted claim rate. The principal model is Poisson: claims occur independently at a constant rate per unit of exposure. The resulting confidence interval is approximate, and its reliability depends in part on the observed claim count.
Poisson Model and Standard Error
Suppose group \(j\) contributes a known exposure \(e_j\) and an observed count \(N_j\). Assume that, conditional on these fixed exposures, the counts are independent and \(N_j\sim\operatorname{Poisson}(\lambda e_j)\), with the same rate \(\lambda\) applying to every group. This assumes a stable claim rate over the observation periods, consistent claim counting, and no dependence among the counts beyond the model. The exposure units must match the rate: a rate per policy-year requires exposure measured in policy-years.
The total exposure is \(E=\sum_j e_j\), and the total claim count is \(N=\sum_j N_j\). Independent Poisson counts add, so \(N\sim\operatorname{Poisson}(\lambda E)\). The estimate \(N/E\) is unbiased, and its variance follows directly from the Poisson mean and variance.
The variance depends on the unknown rate, so estimate it by substituting \(\hat{\lambda}=N/E\). Taking the square root gives either equivalent expression for the estimated standard error:
The numerator is the square root of the claim count, not the square root of the exposure-adjusted rate alone. Equivalently, the expression \(\sqrt{\hat{\lambda}/E}\) includes both the estimated rate and the exposure.
Approximate Confidence Interval
When the count is large enough for a normal approximation to the Poisson distribution to be reasonable, an approximate two-sided confidence interval for \(\lambda\) is the estimate plus or minus a normal critical value times its estimated standard error. For an approximate 95% interval, the standard normal critical value is \(1.96\).
The interval estimates uncertainty in the underlying rate, not the range of claim counts that might occur next year. Its interpretation is in terms of repeated samples: if the same observation process were repeated many times and this method used each time, approximately 95% of the resulting intervals would contain the fixed underlying rate, provided the model and approximation are suitable.
This is a normal-approximation, or Wald, interval. If its lower endpoint is negative, that is a warning that the approximation may be poor; a claim rate cannot be negative. In particular, the formula gives a zero standard error when \(N=0\), which is not a credible statement that the rate is known exactly. With very small counts, a Poisson-based interval designed for low counts is preferable to relying on this approximation.
Worked Calculations
Worked Example: A Claim Rate from Policy-Year Exposure
A hypothetical insurer observes 100 claims over 250 policy-years. Assume independent Poisson claim counts at a constant rate per policy-year. Calculate the estimated rate and an approximate 95% confidence interval, and check the units.
Method. Divide the observed count by exposure, calculate the plug-in standard error, and apply the 1.96 normal critical value. The assumption of a common rate and the use of total policy-year exposure are essential.
Rate estimate. \(\hat{\lambda}=100/250=0.40\) claims per policy-year.
Standard error. \(\widehat{\operatorname{SE}}=\sqrt{100}/250=10/250=0.04\) claims per policy-year.
Margin of error. \(1.96(0.04)=0.0784\) claims per policy-year.
Interval. The lower endpoint is \(0.40-0.0784=0.3216\), and the upper endpoint is \(0.40+0.0784=0.4784\). The approximate 95% interval is therefore \((0.3216,\ 0.4784)\) claims per policy-year.
Check. The interval is centered at \(0.40\), and each endpoint is \(0.0784\) from the estimate. The rate and standard error both have the required units, claims per policy-year. The calculation avoids using the claim count as the denominator: exposure, not count, sets the rate basis.
Worked Example: Comparing Precision at Different Exposure Levels
Two hypothetical portfolios have the same estimated rate of 0.25 claims per policy-year. Portfolio A records 25 claims over 100 policy-years; Portfolio B records 100 claims over 400 policy-years. Under the same Poisson assumptions, compare their standard errors and approximate 95% confidence intervals.
Method. Calculate each estimate from its own count and exposure, then use \(\sqrt{N}/E\). A common mistake is to conclude that equal estimated rates imply equal precision; the total exposure and resulting claim volume also matter.
Portfolio A. Its estimate is \(25/100=0.25\). Its standard error is \(\sqrt{25}/100=5/100=0.05\). Its margin of error is \(1.96(0.05)=0.098\), so its interval is \((0.25-0.098,\ 0.25+0.098)=(0.152,\ 0.348)\).
Portfolio B. Its estimate is \(100/400=0.25\). Its standard error is \(\sqrt{100}/400=10/400=0.025\). Its margin of error is \(1.96(0.025)=0.049\), so its interval is \((0.25-0.049,\ 0.25+0.049)=(0.201,\ 0.299)\).
Check. Both intervals are centered at 0.25. Portfolio B has four times the exposure and four times the observed claims, but half the standard error: the standard error varies with the square root of the count, while the rate divides by exposure. Its interval is narrower, as expected from the larger volume of experience.
Worked Example: Combining Counts from Different Observation Periods
A hypothetical block of policies produces 12 claims during 60 policy-years in one observation period and 28 claims during 140 policy-years in another. Assume the periods are independent and share one constant Poisson claim rate. Find the combined rate and an approximate 95% confidence interval, reporting the result per 100 policy-years.
Method. Add counts and exposures before calculating the rate. This is valid under the stated common-rate model; simply averaging the two separately calculated rates would not in general give the exposure-weighted result.
Combine the data. The total count is \(N=12+28=40\), and the total exposure is \(E=60+140=200\) policy-years.
Estimate per policy-year. \(\hat{\lambda}=40/200=0.20\) claims per policy-year.
Standard error per policy-year. \(\widehat{\operatorname{SE}}=\sqrt{40}/200\approx6.3249/200\approx0.03162\).
Interval per policy-year. The margin of error is \(1.96(0.03162)\approx0.06198\). Thus the endpoints are approximately \(0.20-0.06198=0.1380\) and \(0.20+0.06198=0.2620\).
Convert the rate basis. Multiplying each endpoint and the estimate by 100 gives an estimate of 20.00 claims per 100 policy-years and an approximate 95% interval of \((13.80,\ 26.20)\) claims per 100 policy-years.
Check. The estimate and interval use the same per-100-policy-year basis. The combined estimate is also \(100(12+28)/(60+140)=20\). The calculation avoids treating the two periods as equally weighted when their exposures differ.
What Drives the Uncertainty?
The estimated standard error can be written as \(\sqrt{\hat{\lambda}/E}\). This form separates two influences: the estimated rate and total exposure. For a fixed underlying rate, increasing exposure produces more expected claims, but the rate estimate becomes more precise. Since its standard error is proportional to \(1/\sqrt{E}\) when the rate is fixed, quadrupling exposure approximately halves the standard error and the interval's margin of error.
At a fixed exposure, a larger observed count increases the absolute standard error \(\sqrt{N}/E\). However, the estimate itself also increases with the count. A useful measure of relative uncertainty is the standard error divided by the estimate, when \(N>0\):
Thus the estimated relative standard error is smaller when the observed claim volume is larger. This relationship is not a substitute for checking the assumptions: dependence, changing rates, or materially different underlying risks can make the simple Poisson standard error misleading. If groups have different claim rates, combining them estimates an exposure-weighted average rate, and the common-rate model used for the interval may not describe the data.
When to Use the Approximation
Before reporting an interval, check the data basis and the model. Confirm that exposure is measured in the units of the stated rate, that counts and exposures cover comparable risks and periods, and that a common-rate Poisson model is defensible for the intended estimate. Then check whether the observed count is large enough for a normal approximation. There is no universal count threshold that guarantees accuracy in every setting, but a very small count, a negative Wald lower endpoint, or zero claims should prompt caution.
A confidence interval describes uncertainty in the rate estimate due to observing a finite amount of experience. It does not incorporate every source of uncertainty. For example, it does not automatically allow for changes in coverage, an evolving risk mix, or extra variation caused by clustering. The interval is therefore useful only in the context of the assumptions that produced its standard error.
Check Your Understanding
For each problem, assume the stated Poisson model and use \(1.96\) for an approximate 95% confidence interval unless the question asks you to assess the approximation. Give rates per policy-year unless another basis is requested.
- A hypothetical portfolio has 36 claims over 180 policy-years. Calculate the rate estimate, plug-in standard error, and approximate 95% confidence interval. (Answer: Estimate \(0.2000\), standard error \(0.0333\), interval \((0.1347,\ 0.2653)\) claims per policy-year.)
- A portfolio records 64 claims over 320 policy-years. Find its estimated rate and approximate 95% interval, then compare its standard error with the portfolio in Question 1. (Answer: Estimate \(0.2000\), standard error \(0.0250\), interval \((0.1510,\ 0.2490)\); the standard error is smaller.)
- A hypothetical coverage segment records 49 claims over 140 policy-years. Calculate the approximate 95% interval. (Answer: Estimate \(0.3500\), standard error \(0.0500\), interval \((0.2520,\ 0.4480)\) claims per policy-year.)
- Two independent segments have a common rate: the first has 18 claims over 90 policy-years, and the second has 30 claims over 150 policy-years. Combine the experience and report the approximate 95% interval per 100 policy-years. (Answer: \(48\) claims over \(240\) policy-years; estimate \(20.00\), standard error \(2.8868\), interval \((14.34,\ 25.66)\) claims per 100 policy-years.)
- A group has zero observed claims over 25 policy-years. What does the plug-in Wald formula report, and why should that result not be interpreted as evidence that the underlying rate is exactly zero? (Answer: The plug-in estimate and standard error are both zero, producing the degenerate interval \((0,\ 0)\); this is not a reliable uncertainty statement, so a low-count Poisson interval is needed.)