Truncation and missing family links in population-based registers

Abstract

Studies of familial aggregation of disease routinely use linked population registers to construct retrospective cohorts. Although such resources have provided numerous estimates of familial risk, little is known regarding the sensitivity of the estimates to assumed disease models and incompleteness of the data, such as truncation and/or missing family links. Furthermore, there are no standard tools for testing the validity of estimates from standard epidemiologic designs using register data.

We first introduce a software package, Poplab for simulating realistic populations of related individuals, with complete family information, using easily available vital statistics (fertility and mortality rates) and disease incidence rates. Incidence events are influenced by the familial model of disease, and other disease-related features such as value of familial association, case mortality ratio or biological relationship of aggregation.

We illustrate by mimicking the Swedish population evolving dynamically over the calendar period 1955 - 2002, with female breast cancer aggregating in families. The simulated population agrees well on important demographic features with the real population, and the simulated parameters of familial aggregation are faithfully recovered.

Next, virtual populations are used to investigate the impact of left-truncation of family history and missing family links due to death on familial risk estimates. The missing familial links had no effect, except when there was differential mortality for familial and non-familial cases. Bias due to left-truncation is most pronounced for high familial risks and for registers with a short life-span. The age distribution of disease and the magnitude of background incidence rates also affected the magnitude of bias.

In the third study, we develop a method for correcting the bias in familial risk estimates due to left-truncation. The required sensitivity of exposure is estimated from virtual populations, and was found non-differential for cases and healthy individuals. In all the situations studied, the bias-corrected estimates are in excellent agreement with the true values.

In our last study, we use the bias-correction methodology to evaluate the bias in the apparent familial risks due to left-truncation for the common cancer sites in Sweden. The study cohorts are based on the Swedish MultiGeneration Register linked to the Swedish Cancer Register for the period 1961-2002. We found that corrected age-group specific and overall estimates of the familial risks for colorectum, lung, breast and prostate cancer were close to the apparent relative risks, with overall values of 1.99 95%CI (1.85, 2.14), 2.05 (1.86, 2.26), 1.84 (1.76, 1.92) and 2.33 (2.19, 2.48), respectively. For melanoma, the apparent estimate, 2.68 (2.35, 3.07) was somewhat smaller than the corrected estimate, 3.18 (2.73, 3.64), and was dramatically different with the exposure defined as a parent affected at a younger age, thus changing from 4.07 (3.21, 5.16) to 5.67 (4.51, 6.83) after correction.

In conclusion, we found that left-truncation induces bias in familial risk estimates especially for high values of risk. For common cancers, analyses of the Swedish MultiGeneration cohorts produced generally unbiased familial risk estimates, which was in agreement with our expectations for diseases with older ages of onset and familial risks of relatively low magnitude. However, where the exposure of interest is early age of onset in a parent, commonly considered to be an indication of genetically determined cancers, estimates may be biased, especially where familial risk is high. Our simulation method can correct for such biases and offers a feasible alternative to the use of validation samples.