We have presented a novel approach to calculate the future effect of exposures occurring now. This method can be used for a range of different exposures and diseases for which exposure data are available for an index year and there are appropriate data for predicted future rates of disease. We have written an R program which is available on request for other researchers wishing to use this method.
Comparisons with other methods
The FEF as calculated with this method is not directly comparable to the attributable risk percent as calculated with the attributable risk approach. As shown in Fig. 1, the approaches are estimating different concepts and are using different data. For example, occupational use of asbestos in Australia was much more widespread in the past than it was in 2012 and people who are still working in 2012 may have previously been occupationally exposed to asbestos. It might be more appropriate for the AF method to use the proportion of workers in 2012 who have ever been exposed to asbestos at work rather than a point prevalence of exposure in 2012. However such data are unavailable.
Only one other study has calculated the future burden of mesothelioma due to occupational exposure to asbestos - Hutchings et al. estimated that, by 2060, about 73 % of mesotheliomas will still be attributable to exposure to asbestos that is occurring at work beteween now and then [15]. The major difference between this method and the FEF method is that we do not include future exposure to asbestos.
Several previous studies have estimated the proportion of current mesotheliomas due to past occupational exposure to asbestos. A UK study used actual mesothelioma counts and concluded that 97 % of male mesothelioma deaths and 82 % of female mesothelioma deaths in 2004 were due to past occupational or environmental exposure to asbestos [16]. In the Australian Mesothelioma Registry in 2014, 74 % of mesotheliomas in males and 4 % in females had an occupational cause [17]. A French study examined work histories of mesothelioma cases and controls and calculated AFs of 83 % for men and 42 % for women [18]. Other investigators have assumed that all mesotheliomas are due to occupational exposure, that is, that the AF is 100 % [19]. In Australia, the number of people exposed to asbestos, particularly at high levels, has fallen markedly in the previous 50 years, and more exposure is occurring in non-occupational scenarios [20] so these proportions will decrease in the future.
Assumptions and implications of the assumptions
A major consideration in the use of the FEF method is the definition of the exposed population. In our example we used the prevalence of exposure in the index year. This is similar to the usual way of implementing the attributable fraction approach when a point prevalence of exposure in one historical year is used [3]. An alternative is to use the prevalence of “ever exposed” in the index year or historical year, if these data were available. Doing this would take account of the effect of duration of exposure which contributes to the estimates of risk (RR) used. The difference between point prevalence and prevalence of “ever exposed” would be likely to be greater for exposures which change over time. For example, occupational exposures change when a worker changes a job, and therefore the prevalence of “ever exposed” is likely to be higher than the prevalence of exposure at a single point in time and use of the point prevalence will underestimate the number of future cases. However, this may not be the case for relatively constant exposures such as obesity and alcohol intake which are less likely to change markedly over time. A method of adjusting point prevalences with turnover rates to more closely resemble the “ever exposed” prevalence has also been used [21].
When using the FEF method it is necessary that the chosen RR matches the definition of exposure used for the estimation of exposure prevalence. In our example, the exposure information was limited to whether there was or was not exposure to asbestos in the index year and we had no information on how long exposure had occurred nor the intensity of that exposure although a level of intensity may be available for other agents. We used a point prevalence estimate of exposure in the index year and implicitly assumed that there was a normal distribution curve with regard to the length of exposure and the level of exposure. There would be people exposed in the index year who had very low exposure for a short period as well as people who had high exposure for a long period. We used a RR from the general working population which would also have included people with a range of durations and levels of exposure. Although this RR is related to “ever exposed” it is appropriate to apply to the number of people exposed only in the index year, rather than also including those ‘ever exposed’ prior to but not in that year. Because the effect measure has a significant impact on the results it is important that the most appropriate and best justified measure is chosen. Meta-analyses and pooled analyses may be the best source for the effect measure if these are available.
We did not include a latency period in our estimates of lifetime risk. Similarly to the assumption of previous exposure, we assumed that there was no need for a latent period as some people exposed in the index year had been exposed for a long time and may have developed mesothelioma soon after the index year. An advantage of the LR approach is that it avoids the need to account for latency, as it estimates disease appearing over a lifetime rather than in a specific year, so that timing is not relevant.
Life tables were used for estimating the projected future person-years at risk. While mortality rates in the future may change, sensitivity analysis found that the FEN and FEF in the exposed were sensitive to changes in prevalence of exposure, population numbers, and relative risks, but much less sensitive to changes in mortality.
A limitation of the FEF method is the accuracy of projections of future disease. The validity of a projection will vary by the disease outcome being modelled. In our example, high quality national cancer registry data were available for the previous 30 years, but this is not always the case for other diseases or other countries. The future number of cases depends on the change in size and age structure of the population (which is relatively easy to predict), as well as changes in the rates of disease (which are more difficult to predict) [22]. We used CanProj which estimates likely future trends based on what has been observed in the past and does not take into account possible changes in risk factors (such as changes in smoking patterns). We projected rates for over 80 years into the future, so errors are likely to be large, others may limit the projection to a shorter interval. Even so, cancer is one of the diseases for which there is a relatively good range of options for projecting future rates. Most of the approaches are based on age-period-cohort models including a recently produced Bayesian method [23]. An alternative to using a forward projection model that takes account of competing risks causing the same disease outcomes (step 5 in the example) would be to apply a constant disease rate (for t = 0, e.g., 2012) to projected PYAR.
Policy applications
In the FEF method, exposure can easily be varied to predict the effect of various control measures. For example, in the case of cigarette smoking it is possible to model the effect of policies such as reducing tar content of cigarettes, mandating plain packaging, or increasing cost. Because the estimations are based on changes to current exposure, they are readily understood by policy makers. In addition, the exposures are current, so are of more salience to policy makers than exposures occurring in the past (e.g., it may be easy to disregard mesotheliomas arising from past exposure to asbestos in the belief that current exposure is negligible).
Optional extensions to the FEF model
Researchers with more detailed exposure data may be able to match their exposure categories with several RRs. For example, if there are population data on the prevalence of exposure to cigarette smoking for the index year, it may be possible to use different RRs for the different categories of pack-years. This would allow prediction of the effect of reducing exposures (e.g., smoking fewer cigarettes a day) rather than completely eliminating exposures.
Multiple exposures or outcomes can be accounted for using the product of complements method [2]: In this method the combined FEF for one outcome is equal to the complement of the product of the complement of each FEF for k exposures.
$$ FE{F}_{combined}=1 - {\displaystyle {\prod}_k\left(1-FE{F}_k\right)} $$
The combined FEF can be multiplied by the number of expected outcomes in the population to obtain the combined FEN for one outcome and multiple exposures. Combined FENs for different outcomes can be summed to obtain the overall number of attributable outcomes.
In conclusion, the future excess fraction method offers a different way of expressing the burden of disease which is readily understandable by the general public and by those with the ability to influence public health decisions.
Ethics
This manuscript describes a new method, and used publically available data sets for the example therefore no ethics approval was required.
Consent
No individual or identifiable data were used in this manuscript and therefore no consent was required.
Availability of data and materials
All data sources are referenced.
Trial registration
This manuscript does not report the results of a controlled health-care intervention and is not a trial.