Addressing Complications in Exposed-to-Risk Calculations

The Challenge: From Ideal Data to Census Approximations

In an ideal world, for every year of age $x$, we would know the exact population size at the start of the year and the end of the year. This would allow a straightforward calculation of the central exposed to risk. However, population-level data is rarely so convenient. Typically, an actuary or demographer has access to two main sources:

1. Periodic Census Data: A full population count by age and sex, but only conducted on a specific day every 5 or 10 years.
2. Vital Statistics: A register of births and deaths that occur during each calendar year.

The core problem is this: how do we calculate the exposure to risk for a specific calendar year, say 2023, if our last census was in 2020 and the next one won't be until 2030? We need to approximate the population size for the years between the censuses. This is the essence of handling "census approximations."

Method 1: Inter-Census Survival Rates

A direct, if somewhat crude, way to analyze mortality between two censuses is to calculate a cohort survival rate. The logic is simple: if we assume the population is "closed" (meaning no migration in or out), then the change in the size of a cohort from one census to the next must be due to deaths.

For example, the population aged 40-44 in a 2010 census is the same cohort as the population aged 50-54 in a 2020 census. By comparing the size of the cohort at these two points in time, we can calculate a 10-year survival rate.

The following resource details this straightforward but powerful technique.

This method provides a survival rate, $S_x$, which is an estimate of the probability $p_x$. The approximation for the 5-year rate from the 10-year rate, ${}_5{S}_x\approx \sqrt{{}_{10}{S}_x}$, stems from the relationship for survival probabilities. The probability of surviving 10 years from age $x$ is the product of surviving the first 5 years and then the next 5 years: ${}_{10}{p}_x{=}_5{p}_x\cdot {}_5{p}_{x+5}$. If we assume mortality doesn't change much over that 10-year span (${}_5{p}_x\approx {}_5{p}_{x+5}$), then we get ${}_{10}{p}_x\approx {(}_5{p}_x{)}^2$, which leads to the square root approximation.

While useful, this method is sensitive to its assumptions. The "no migration" assumption is the most significant weakness. As the reading notes, this is why the method is best applied to national-level data (where international migration is a small fraction of the total population) and, if possible, restricted to the native-born population to minimize this distortion.

Method 2: The Intercensal Survival Method

The previous method gives us an overall survival rate for the period, but it doesn't give us what we really need for exposure calculations: an estimated population count for each year between the censuses. For that, we turn to a more sophisticated technique known as the intercensal survival method.

The concept is elegant. We start with the population of a cohort from the first census. We then "survive" this population forward, year by year, by subtracting the registered deaths for that cohort. When we reach the year of the second census, our calculated population will inevitably differ from the actual census count. This difference, or "closing error," represents the net effect of everything we didn't account for, primarily migration and data inaccuracies. The final step is to distribute this error back over the intercensal period, usually uniformly, to create a consistent set of annual population estimates.

The Human Mortality Database (HMD) provides the definitive protocol for this method.

Let's illustrate with a simplified numerical example.

• Census 1 (1 Jan 2010): Population aged 50 is 100,000.
• Census 2 (1 Jan 2020): Population aged 60 is 92,000.
• Vital Statistics: Total registered deaths for this cohort between 2010 and 2020 is 7,500.

1. Estimate survivors without migration:
   $\text{Estimated Survivors}=100,000-7,500=92,500$
2. Calculate the closing error ($\Delta$):
   $\Delta =\text{Actual Survivors}-\text{Estimated Survivors}=92,000-92,500=-500$
3. Interpret the error: A negative error of 500 implies a net outflow of 500 people over 10 years, either through emigration or data discrepancies.
4. Distribute the error: We assume this outflow happens evenly, at a rate of 50 people per year.
5. Estimate annual population: To estimate the population aged 51 on 1 Jan 2011, we would start with the 100,000, subtract the deaths at age 50, and then subtract the first year's share of the error (50).

By applying this method, we can generate a complete set of population estimates by age for January 1st of every year. With these estimates, we can then calculate the central exposed to risk for any calendar year, often using the approximation $E_x^c\approx \frac{1}{2}(P_x+P_{x+1})$, where $P_x$ is the population at the start of the year and $P_{x+1}$ is the population at the end of the year (assuming the time between $P_x$ and $P_{x+1}$ is one year).

Other Practical Complications

Handling census approximations is the main challenge, but it's worth noting another common issue: the format of the data. Death data often comes in aggregate form, such as "total deaths in calendar year 2023 for the 60-64 age group." For precise cohort analysis, we need to split these deaths into their respective cohorts.

A death in 2023 at age 60 could belong to the cohort born in 1963 or 1962. A simple Lexis diagram illustrates this. The deaths in the 1x1 square defined by age 60 and year 2023 are split between an upper triangle (cohort born in 1962) and a lower triangle (cohort born in 1963). The simplest approach is to split these deaths 50/50. More advanced methods, like those in the HMD protocol, use regression models to perform this split more accurately. For the CS2 exam, the key is to be aware that this data-splitting step is a necessary part of the process.

Conclusion

This lesson has moved us from the clean world of theory into the messy reality of demographic data. You've seen how actuaries and demographers bridge the gaps in their data to produce the population estimates necessary for robust mortality analysis.

Key Takeaways:

• Real-world mortality studies often rely on periodic census data and annual death counts, not complete individual records.
• The inter-census survival rate is a straightforward method for estimating cohort survival by comparing population sizes at two censuses, but it assumes a closed population.
• The intercensal survival method is a more robust technique for creating annual population estimates. It works by surviving a cohort forward with death data and then systematically distributing the closing error (due to migration and other factors).
• Once annual population estimates are derived, they form the basis for calculating the central exposed to risk for each year.

Preview of the Next Lesson

We have now explored how to define exposure, relate different exposure measures, and handle the practical challenge of estimating population counts from census data. In the next lesson, we will tie these concepts together by looking at the actuarial estimate. This is a specific formula used to calculate age-specific mortality rates ($q_x$) directly from life table data, which are themselves constructed from the population and death data we've been discussing. This will formalize the link between survival analysis and rate estimation.