Cold Spells and Cause-Specific Mortality in 47 Japanese Prefectures: A Systematic Evaluation

Abstract

Background: Many studies have investigated the devastating health effects of heat waves, but less is known about health risks related to cold spells, despite evidence that extreme cold may contribute to a larger proportion of deaths.

Objectives: We aimed to systematically investigate the association between cold spells and mortality in Japan.

Methods: Daily data for weather conditions and 12 common causes of death during the 1972-2015 cold seasons (November-March) were obtained from 47 Japanese prefectures. Cold spells were defined as $\ge 2$ consecutive days with daily mean temperatures $\le 5\text{th}$ percentile for the cold season in each prefecture. Quasi-Poisson regression was combined with a distributed lag model to estimate prefecture-specific associations, and pooled associations at the national level were obtained through random-effects meta-analysis. The potential influence of cold spell characteristics (intensity, duration, and timing in season) on associations between cold spells and mortality was examined using a similar two-stage approach. Temporal trends were investigated using a meta-regression model.

Results: A total of 18,139,498 deaths were recorded during study period. Mortality was significantly higher during cold spell days vs. other days for all selected causes of death. Mortality due to age-related physical debilitation was more strongly associated with cold spells than with other causes of death. Associations between cold spells and mortality from all causes and several more specific outcomes were stronger for longer and more intense cold spells and for cold spells earlier in the cold season. However, although all outcomes were positively associated with cold spell duration, findings for cold spell intensity and seasonal timing were heterogeneous across the outcomes. Associations between cold spells and mortality due to cerebrovascular disease, cerebral infarction, and age-related physical debility decreased in magnitude over time, whereas temporal trends were relatively flat for all-cause mortality and other outcomes.

Discussion: Our findings may have implications for establishing tailored public health strategies to prevent avoidable cold spell-related health consequences. https://doi.org/10.1289/EHP7109.

Figure 1.

Overall lag structure (0–21 d) in effects of cold spell on cause-specific mortality in 47 Japanese prefectures during the period 1972–2015. Cold spells were defined as $\ge 2$ consecutive days with daily mean temperatures lower than the prefecture-specific 5th percentile cold season daily mean temperature (for 1972–2015). Prefecture-specific cold spell–mortality associations were estimated using quasi-Poisson regression with a distributed lag model, which were then pooled using random-effect meta-analyses with restricted likelihood estimation. Gray bands are 95% confidence intervals.

Figure 2.

Estimated percentage increase in all-cause mortality on cold spell days compared with non–cold spell days during the 1972–2015 cold seasons (November–March) in 47 Japanese prefectures over lag 0–21 d. Cold spells were defined as $\ge 2$ consecutive days with daily mean temperatures lower than the prefecture-specific 5th percentile cold season daily mean temperature (for 1972–2015). Prefecture-specific cold spell–mortality associations were estimated using quasi-Poisson regression with a distributed lag model, which were then pooled using random-effect meta-analyses with restricted likelihood estimation. Note: CI, confidence interval.

Figure 3.

The annual change in coefficient of the association between cold spell and cause-specific mortality in 47 Japanese prefectures during the period 1972–2015. Cold spells were defined as $\ge 2$ consecutive days with daily mean temperatures lower than the prefecture-specific 5th percentile cold season daily mean temperature (for 1972–2015). Prefecture-season-specific cold spell–mortality associations were estimated using quasi-Poisson regression with a distributed lag model, which were then pooled at the national level through random-effect meta-regression models, with “season” as the independent continuous variable (linear) and prefecture-season specific coefficients as the dependent variable. Gray bands are 95% confidence intervals. $p$-values are for the slopes of the annual trends. Note: CI, confidence interval.