Mortality and Macroeconomic Conditions: What Can We Learn From France?

Brenner (1971) and Brenner and Mooney (1983) conducted similar research, but their empirical methods have subsequently been heavily criticized.

An exception to the rule that higher unemployment is associated with fewer deaths due to cardiovascular diseases might be Sweden (Svensson 2010), even though Tapia Granados and Ionides (2011) cast doubt on this conclusion.

Personal communication with Buchmueller (February 28, 2017) confirmed that the project was abandoned because this research group discovered that the results were not robust.

See Stevens et al. (2015) for a similar analysis in the United States.

In contrast to these more recent theories, earlier research tended to explain procyclical mortality chiefly with people’s own employment status. In this light, the direction of the effect of unemployment on health is not evident a priori. On the one hand, unemployment can be a psycho-social stressor that reduces quality of life and subjective well-being, possibly resulting in higher morbidity and mortality. Downward mobility could also be associated with increased morbidity and mortality. On the other hand, economic downturns could reduce stress and overtime hours, which should particularly reduce deaths due to cardiovascular diseases. Rising opportunity cost of time that accompanies better labor market opportunities might lead to higher mortality (Miller et al. 2009) because it makes it more costly for individuals to undertake time-intensive, health-producing activities. Finally, income growth may increase risky activities, such as drinking and driving (Ruhm and Black 2002). Stress-induced increases in alcohol consumption were found to be more than offset by reductions in incomes during economic crises.

These latter controls are obtained by interpolating census data from 1982, 1990, 1999, 2009, and 2014.

Coefficients in this specification continue to be interpreted as semi-elasticities. For a discussion of the advantages of PPML over a log-linear model estimated by ordinary least squares (OLS), see Manning and Mullahy (2001) and Santos Silva and Tenreyro (2006, 2011).

When we split our sample in two parts, the value was .03 from 1982 to 1998 and was stable from 1998 to 2014 (i.e., .03). Ruhm (2015) found this value to be .09 for the period 1999–2010 in the United States.

https://​www.​cepidc.​inserm.​fr/​causes-medicales-de-deces/​centre-collaborateur-oms.

As an additional robustness check, we present results with the age-adjusted mortality rate as dependent variable in the online appendix (section D).

Anderson et al. (2001) provided comparability ratios for the United States. For the specific sources of mortality we are considering, most of the estimated comparability ratios are close to 1, suggesting that a similar number of deaths are reported using either ICD system.

http://​www.​insee.​fr/​fr/​insee-statistique-publique/​default.​asp?​page=​connaitre/​ddar.​htm.

For details on how unemployment statistics are calculated in France, see Fougère et al. (2009) or http://​www.​insee.​fr/​en/​methodes/​default.​asp?​page=​sources/​ope-taux-chomage-localises.​htm.

In comparison, the average state in the United States had a population of 6,280,000; the average county, 100,000. The size of geographic unit is expected to have a moderate influence on the estimated coefficients. The channels through which the unemployment rate influences mortality might vary with the level of aggregation. Smaller geographical units will have higher spillover effects (Lindo 2015).

Section B of the online appendix shows that the inclusion of département-specific time trends similarly changes most of the results of Buchmueller et al. (2007).

See also Ruhm (2015) for a discussion of the influence of weights on the unemployment coefficient for the case of the United States.

Another reason for weighting that is sometimes given is to achieve more efficient estimation given heteroskedasticity. However, as shown in Table 1, the standard errors between WLS and OLS estimates hardly differ, indicating that weighting by area population does not do a better job of dealing with heteroskedasticity. See Solon et al. (2015) for a discussion of the relative advantages of WLS and OLS.

As an additional check, we ran the département-level regression with the full set of controls, weighting, year fixed effects, and département fixed effects, but we replaced the département-specific time trends by region-specific time trends. This yielded a coefficient estimate of 0.0014 (standard error = 0.0011). This should make our analysis more robust to the claim that the département-specific time trends take too much variation out of the analysis. Unlike in panel B of Table 1, the old regions were used to create the region-specific time trends because these were the administrative levels in place during the analysis period. Using the new regions for the time trends in an otherwise similar specification yields an estimate of –0.0001 (standard error = 0.0015).

In unreported regressions, we also used the nontransformed mortality rate as a dependent variable and estimated by OLS. However, because mortality rates can be viewed as count data, we prefer to use a count model and thus keep the preferred functional form and the interpretation of coefficients as semi-elasticities.

In unreported results, we found that the coefficients steadily decline with age. Younger people who are inexperienced drivers are involved in more accidents. It also makes sense that the effect would be weakest among people above the retirement age given that this demographic group is least likely to participate in auto travel, partly because they no longer commute to work.

As a robustness check, we created the same figures excluding the weighting by population. Results were qualitatively the same but coefficients were closer to 0, as can be expected from Table 1.

Interestingly, in his reanalysis of his 2000 paper, Ruhm (2015) actually added such a state-specific time trend.

We can compare unemployment with GDP by region only from 1990 to 2014.

Importantly, we did not have access to the full time series for all variables, but this is the most complete data set we can find for France.