Review of causal pathways
As part of the Climate and Health Profile Report, several direct and indirect health impacts related to EH and EP were identified . Pathways by which EH affects health are direct and have been reviewed elsewhere [24, 36,37,38]. Briefly, mortality due to non-accidental causes was chosen as an indicator of disease burden for this analysis due to its known direct and immediate association with EH events [36, 37]. Non-accidental-cause ED visits and renal, respiratory and heat-related hospitalizations were chosen to best reflect the impact of EH on a wide variety of chronic diseases. Previous studies have shown associations between EH events and non-accidental-cause ED visits  and between EH and renal, respiratory and heat hospitalizations [24, 38]. Therefore, we included non-accidental cause mortality; renal, respiratory, and heat-related hospitalizations; and non-accidental ED visits in our EH burden of disease estimates, using studies that included Michigan and provided region-specific results, where possible, and from similar climates otherwise.
Multiple pathways between EP and health may exist, although we did not find sufficient quantitative estimates of EP and health for the majority of these pathways to sufficiently characterize the health effects in our quantitative burden of disease estimates. Therefore, we describe these pathways in detail below. EP leads to contamination of surface water by increasing turbidity, by increasing the chances of harmful algal blooms, which are fed by agricultural run-off, and by prompting combined sewer overflows (CSOs). EP may also contaminate surface water when EP leads to flooding and flood waters wash contaminants into the surface water body.
EP events have been found to be associated with GI illness in countries outside the U.S. with inadequate treatment of public drinking water [39, 40]. Even in the U.S. where public water supplies are treated, a small proportion of illness has been attributed to waterborne disease . These waterborne infections are caused by a variety of viruses, bacteria and protozoa. Though rare in Canada and the U.S., EP has been associated with waterborne disease outbreaks in recent decades: a Cryptosporidium outbreak in Milwaukee in 1993 and an Escherichia coli outbreak in Walkerton, Ontario in 2000 .
In addition to GI illnesses, a small number of studies have found an association between precipitation and the respiratory pathogen Legionella, or Legionnaires’ disease in its most severe form, which thrives in warm water [43, 44]. In a study of legionellosis incidence in five Mid-Atlantic states from 1990 to 2003, Hicks et al.  found both monthly temperature (another meteorological variable predicted to increase with climate change) and rainfall to be associated with legionellosis. Specifically, Hicks et al. found a 2.6% increased risk of legionellosis with each 1-cm increase in rainfall. A study in Switzerland did not find associations between precipitation and legionellosis, although the researchers did find associations of legionellosis with temperature and water vapor pressure . A case-crossover study of 240 Legionella cases in the Philadelphia area from 1995 to 2003 also did not provide evidence of an association between precipitation and legionellosis when controlling for other meteorologic factors, but the researchers did find associations between Legionella and relative humidity, with RRs of 3.93 (95% Confidence Interval: 2.18–7.09) and 3.59 (95% CI: 2.06–6.28) for the 4th and 5th quintiles of relative humidity, respectively, vs. the first quintile of relative humidity .
Turbidity, or water clarity, is often used as a proxy for microbial contamination, and the EPA has established regulations limiting levels of turbidity in public drinking water . EP events can increase the turbidity in surface water. However, all drinking water treatment plants in Michigan actively reduce levels of turbidity when processing raw water into finished drinking water. Studies of associations between gastrointestinal illness and turbidity, at levels as low as those measured in U.S. drinking water systems, have shown mixed results. Studies in Philadelphia, Milwaukee, Atlanta, New York City, Vancouver, and Quebec each found associations between turbidity, as measured at the treatment plant, and GI illness in subsets of age groups, subsets of seasons and/or subsets of days following the elevated turbidity, [49,50,51]. However, a study in Edmonton failed to find any association between turbidity and GI illness .
Evidence is strong for associations between GI health effects and toxins produced by cyanobacteria in harmful algal blooms (HABs) . The World Health Organization has set standards for microcystin levels in public drinking water systems , and these levels were exceeded in the 2014 Lake Erie algal bloom in Toledo, Ohio . Almost 100,000 Michigan residents use public water systems with intakes in Lake Erie. The potential health effects of HABs in these cities are of concern given that current water treatment methods do not remove all of the toxins produced by the cyanobacteria. Emergency response plans addressing a HAB, including bottled water distribution, have been put in place .
CSOs contaminate the receiving water with raw sewage and therefore with human pathogens. CSOs can lead to increased concentrations of pathogens in surface water [54,55,56,57]. Although treatment of the public water supply should remove these pathogens, a study in Massachusetts found associations between EP and GI illness in regions in which the public drinking water came from surface water and CSO discharges occurred . We explore the implications of this finding on the present and future burden of EP-associated GI illness in our quantitative burden of disease estimate.
Flooding, which can occur during EP, may also be associated with waterborne illness. Following severe flooding in the Midwest in 2001, the EPA  investigated the risk of GI illness among participants of a pre-existing drinking water intervention study. The incidence rate ratio for GI symptoms during the flood vs. prior to the flood was 1.29 (95% CI: 1.06–1.58), with an increased effect among individuals with increased sensitivity to GI illness. GI symptoms were also associated with floodwater contact, particularly in children. In a case-crossover study of 129 floods in Massachusetts from 2003 to 2007, Wade et al.  observed an odds ratio of 1.08 (95% CI: 1.03–1.12) for ED visits for GI illness in the 0–4 day period after flooding.
People may be exposed to pathogens in contaminated surface water by ingestion, inhalation or dermal contact. These exposures may be due to 1) inadequate treatment at the water treatment plant, 2) contamination of drinking water in the treated water delivery system or 3) direct contact with contaminated surface water. Flooding, in particular, may contribute to human exposure of contaminated surface water via pathways (2) and (3) . These indirect pathways between EP and illness are likely mediated by the quality of the drinking water treatment and delivery infrastructure as well as the sensitivity of the population to the microbes or their toxins.
Multi-GCM ensemble projections
By ZIP code, average annual EH days ranged widely, from 0 to 16 days historically and 0–46 days in the projected period (Table 2). By both ZIP code and county, the median projected average annual count of EH days was approximately 5 times higher than the historical count. The historical and projected counts of EH days were highly correlated (Spearman r = 0.99), and the six counties (St. Joseph, Wayne, Berrien, Cass, Kalamazoo, and Monroe Counties) with the highest number of EH days were the same in both the historical and projected periods (Additional file 1: Appendix 5).
In the historical period (1970–2000), there were, by definition, 7.3 EP days per year (2% of 365.25), so given an increase of 0.5–2.0 days per year statewide, EP days will increase to 7.8–9.3 days per year.
Historical disease burden of the health outcomes
The population age distribution changed from the historical to the projected period, with the percentage of older adults increasing from 12% to 18% statewide.
Across 83 counties, the daily baseline mortality rate for non-accidental deaths ranged from 1.8 to 5.3 deaths per 100,000 persons, with a median of 3.2 deaths per 100,000 persons in the historical period (Table 3). The median mortality rate in the projected period was slightly higher (3.3 deaths per 100,000 persons). We did not age-standardize these mortality rates, so the wide range of mortality rates partially reflects differing age distributions between counties. For renal hospitalizations, we estimated a statewide daily rate among non-whites in the warm season of 0.24 hospitalizations per 100,000 persons for both periods among individuals under 65 years of age. For older individuals, daily ED visits for non-accidental causes ranged from 34 to 56 visits per 100,000 persons in the historical period, and because of an increased percentage of older adults over time, the median in the projected period was again higher than in the historical period (53 vs. 44 visits per 100,000 persons).
For both periods, we estimated a daily ED visit rate for gastrointestinal illness in Michigan of 3 visits per 100,000 residents. Because our exposure-response association is only applicable to communities where drinking water may be exposed to CSOs, this result applies only to the 5.8 million Michigan residents served by a community public water system receiving drinking water from the Great Lakes or connecting channels or from inland rivers and lakes .
We did not see a significant association between either EH threshold (32.2–34.9, ≥35o C) and mortality in the 0–4 or 5–19 age groups (Additional file 1: Appendix 4, Table A2). Among men ages 20–54, we found a risk of mortality 1.12 (1.01–1.25) times that of women during ≥35o C EH days, and we found an added risk of mortality among individuals without a high school degree in this age group, as well. Among individuals 55–64, we found a 1.07 (1.00–1.14) risk of mortality during ≥35o C EH vs. non-EH (Additional file 1: Appendix 4, Table A2), but we did not find significant effect modification by any of the individual or area-level characteristics tested. Among individuals ages 65 and older, we found similar risks within 10-year age groups, so these were combined as in Gronlund et al. . In this age group, we found added risks among non-married individuals, and increased risk with increasing non-green space and increasing poverty at the ZIP-code level. We did not vary our AFs by marital status given that this characteristic varied little by ZIP code. Furthermore, we did not vary our AFs by non-green space given that similar land cover characteristics are used in deriving the historical and projected area-specific EH days. The increased vulnerability among residents of ZIP codes with high poverty rates was a different finding from Gronlund et al. , perhaps because our EH definition was more extreme than that used in Gronlund et al. .
For the 65 and older age group, we estimated an AF that was significantly greater than zero in 90% of the ZIP codes for EH ≥ 35 °C. Of these, the estimated AFs ranged from 0.12 to 0.69 (results not shown). Of the ZIP codes with AFs of at least 0.30, 36% were in Wayne County, MI. For EH ≥ 35 °C for men 20–49, the AF for all ZCTAs was estimated as 0.091. Likewise, for individuals ages 55–64, the statewide AF was estimated as 0.066.
Based on a statewide, all-ages RR of 1.31 for the risk of renal hospitalizations during EH days ≥32.2 °C vs. non-EH days among non-whites, we estimated an AF of 0.24 among non-whites less than 65 years of age. Among individuals 65 years and older, we estimated ZIP-code-specific AFs EH days ≥32.2 °C ranging as high as 0.63, among blacks 78 and older. Thirty-one percent of the 52 ZIP codes with AFs greater than 0.3 in any age-race group were in Wayne County, MI.
EH-ED visit association
Based on the RRs presented in Kingsley et al. for non-accidental ED visits, we estimated AFs for EH days 32.2–34.9 °C of 0.042 and 0.071 for the 0–18 and 65 and older age groups, respectively. For EH days ≥35 °C, we estimated slightly higher AFs of 0.052 and 0.088 for the two respective age groups. For heat-related visits in the 18–64 age group, the AFs were much higher: 0.61 and 0.69 for EH days 32.2–34.9 °C and EH days ≥35 °C, respectively.
EP-ED visit association
Based on an RR of 1.13 for the risk of an ED visit for a GI illness at the 99th percentile of EP, we estimated an AF of 0.12 for the residents receiving drinking water from a surface water source.
EH- and EP-attributable health burdens
We estimated the rate of all non-accidental mortality associated with EH days to increase sixfold from 0.46 per 100,000 adults aged 20 years and older in the historical period (33 deaths annually statewide, Table 4) to 2.9 per 100,000 adults (240 deaths annually statewide, Table 4) in the projected period. There was significant heterogeneity between counties, with a 19-fold variation in mortality rate between counties in the historical period and a nine-fold variation between counties in the projected period (Fig. 2a-b, Additional file 1: Appendix 5).
Mortality was highest among older adults, and the proportion of EH associated deaths that occurred in the 65 and older age group increased from the historical to the projected period (87% and 91%, respectively, Fig. 3), due to the increased percentage of the population of individuals 65 and older.
EH-associated hospitalization rates also ranged widely by county, ranging from 0.002 to 0.58 per 100,000 persons in the historical period, and 0.03 to 2.5 per 100,000 persons in the projected period. We estimated the annual statewide number of EH-attributable hospitalizations to increase from 28 in the baseline period to 185 in the projected period (Table 4).
EH-associated ED visit rates were substantially higher than EH-associated hospitalization rates, with rates of 12 per 100,000 persons (1218 visits statewide) in the historical period and 68 per 100,000 persons (7845 visits statewide) in the projected period (Table 4). Given that ED visit rates were much higher than hospitalization rates for EH, we focused the subsequent analyses and discussion of EH-associated morbidity on EH-associated ED visits. Significant heterogeneity between counties was seen, with a 16-fold variation in ED rate between counties at historical and a 6-fold variation in ED rate between counties in the projected period (Fig. 2c-d, Additional file 1: Appendix 5).
Multiplying our estimates of the AF, EP days, GI-illness ED visit incidence rate, and population, we estimated a historical burden-of-disease rate of ED visits for GI illness attributable to EP as 170 ED visits annually, or a rate of 1.7 visits per 100,000 Michigan residents. Assuming an increase of approximately 1 day of EP in the future period, we estimated a future burden of waterborne disease attributable to EP as 220 ED visits annually, or 1.9 visits per 100,000 Michigan residents (Table 4). Considering that the number of days by which EP events will increase is projected to range spatially between 0.5 days to 2.0 days, depending on the region in Michigan, the EP-associated ED visit rate increase may range spatially from 1.0 to 3.82 visits per 100,000 Michigan residents.
The cost of EH-associated mortality across the State of Michigan was $280 million in the projected period and $42 million in the historical period, based on a value per life-year of $129,000. The projected cost of EH-associated morbidity was dominated by EH-associated ED visits, estimated at $14 million, or $12 million higher than the historical cost (Table 4). EP costs were similar in the historical and projected periods: $390,000 and $480,000, respectively.
Table 5 summarizes the uncertainties in estimating the burden of disease due to climate.
Baseline health effect estimates
For the historical estimates, our uncertainty in the baseline mortality estimates is low given that over 99% of deaths in the U.S. are thought to be registered . Despite not having county-specific historical hospitalization data for our study, our estimate of uncertainty in hospitalization rates by race and age is also low, given that over 50% of the state’s black population lives in the counties for which we had detailed warm-season hospitalization rates by race and cause. Furthermore, the all-cause renal hospitalization rate in the 3 counties of 12.2% in 2014 is very close to the statewide rate of 12.5% . We did not have county-specific or even statewide ED visit rates available for this study. In comparing statewide rates among the Midwest states for which data were available, we found a maximum absolute percent difference around the Midwest estimate for the heat-related ED visits of 57%, giving us moderate uncertainty in our baseline ED-visit-rate estimates (Additional file 1: Appendix 3).
Uncertainties in the projected baseline health effect estimates were moderate, based on recent trends. Age-adjusted renal hospitalization rates increased 30% over 12 years, from 100 per 100,000 persons in 2001–2003 to 130 per 100,000 persons 2012–2014. Trends in ED visit rates from 2006 to 2011 varied depending on diagnosis, with the steepest increase of 74% for septicemia ; but overall, ED visit rates increased 4.5% in this 6-year time period.
Estimated mortality and morbidity impacts varied by population projection. The statewide EH-associated mortality rate estimated using the 2050 ICLUS A2 population estimates was 3% below that using the Woods & Poole central-case scenario while the EH-associated ED visit and hospitalization rates were 50% and 44% higher, respectively, than those using the Woods & Poole scenario. This suggests our estimates of future EH-attributable mortality are only mildly sensitive to assumptions about population growth and migration, but our morbidity estimates are moderately sensitive to population change assumptions.
A large source of uncertainty is that in the climate projections. The bias in number of days with maximum temperature > 32.2 °C for the A2 scenario from the six GCMs that were used to generate the Michigan-specific projections range ranged from − 50% to + 20% in Michigan. For EP, bias in wet days with > 3 in. of precipitation, the A2 scenario-specific biases range from − 90% to 100% in Michigan, depending on the GCM and the region . Therefore, projections of increases could reasonably range from almost no increase in EP days to twice as many days as projected by the ensemble climate projection. Global and downscaled climate projections may contain additional biases not quantified above, especially near the Great Lakes region. The spatial resolution of GCMs does not allow for accurate representation of the lakes’ influence on local and regional climate; at best, GCMs may capture the large-scale regional effects of the lakes . Furthermore, assigning EH and EP projections to ZIP codes, which are smaller than 1/8-degree, adds an additional source of uncertainty. With regards to historical EP exposure estimates, those used in deriving the precipitation exposure-response association were from single monitors rather than modeled data with high spatial resolution. Given the spatially heterogeneous nature of precipitation, this may result in bias of the effect estimates towards the null, as was demonstrated in a recent simulation study of precipitation health effects .
Another large source of uncertainty is in our estimate of the RRs. For the EH-hospitalization estimate among individuals under 65 years of age, the 95% confidence intervals ranged from ±80% around the point estimate. Among individuals 65 and older, the 95% confidence interval for the median RR estimate (1.13) of hospitalization ranged from ±40% around the point estimate. For the EH-mortality estimates, for which we had a large Michigan-specific data set, the 95% confidence intervals around the RRs corresponded to increases of more than 100% or decreases of close to 100% in the AFs. For the EH-ED associations, the 95% confidence intervals ranged as high as 52% above the point estimate. Additionally, the cumulative effects in the days following the exposure, i.e., the lagged effects, were not estimated in the EH-ED source study, further contributing to uncertainty in the net effect of EH on ED visits.
Several recent studies have found a substantial decrease in the association between EH and mortality over time , suggesting strong technological and/or behavioral adaptation to EH. Nordio et al. found a decrease in the RR for mortality at 27 °C vs. 16 °C for the climate region containing Michigan, from 1.25 in the 1962–1966 time period to 1.08 in the 2000–2006 time period . Likewise, Bobb et al. found that excess deaths per 1000 deaths attributable to each 5.6 °C increase in summer temperature declined from 50 per 1000 in 1987 to 11 per 1000 in 2005 in the Industrial Midwest . In New York City, in a climate similar to that of Michigan’s, Petkova et al. found a decline in RR for mortality at 29 °C vs. 22 °C over 11 decades, from 1.43 in the 1900s to 1.09 in the 2000s . Although all of these studies show a leveling off of this decline in recent years, changes in the RRs or attributable deaths over time of 60–80% for the mortality outcomes suggest that change over time for all of the exposure-response associations remains a source of moderate uncertainty.
External validity is another source of uncertainty in our EH-ED and EP-ED exposure-response estimates. For these estimates, we chose studies from similar climates, but population and infrastructure differences may affect the portability of RRs between two states. Michigan differs from Rhode Island in its demographic structure, with, for example, 14% vs. 8% of the population identifying as black . With regards to the EP exposure-response estimate, Michigan likely differs from Massachusetts in its water delivery infrastructure, its demographic structure, and the sensitivity of its population to waterborne pathogens.
For EP, several studies of the association between precipitation and gastrointestinal illness did not find evidence of an association . Jagai et al. may have been able to detect this association because they stratified their analysis by CSO exposure . Additionally, in estimating the burden of disease, we applied RRs derived from a study defining EP at the 99th percentile of daily precipitation to exposure estimates of EP defined at the 98th percentile of daily precipitation, which might slightly overestimate the burden.
Communities across the State of Michigan are in the process of eliminating CSOs. However, a recent study in Massachusetts  found associations between sanitary sewer overflows (SSOs) and GI ED visits, suggesting that separating the sanitary and storm sewers will not entirely eliminate the pathway by which sewage can affect GI ED visits during future EP events.
Our uncertainty in the costs associated with the historical and projected outcomes was moderate to high. For the costs associated with EH-attributable mortality, we were limited in our lack of estimates of years-of-life-lost. The time series and case-crossover study designs, on which our estimates were based, cannot estimate by how many months or years the deaths were advanced. Our approach of using the reduced life expectancies among individuals who die of cardiovascular or respiratory diseases attempts to account for this, but uncertainty in the degree of mortality displacement remains high. Uncertainty in the cost of a life-year is also high, with a reported range of $65,000 to $490,000 . For the morbidities, uncertainties in the estimated costs were moderate. The reported ranges were less than ±50% around the point estimates for hospitalization costs , but we are uncertain as to how well the estimates reflected the cost of an EH-associated visit that was not necessarily coded as heat-related. Specifically, we used heat-related hospitalization costs to estimate EH-associated renal and respiratory hospitalizations, and we used non-accidental ED visit costs to estimate ED visits that were presumably triggered by EH.
We have additional uncertainty in projecting the costs given trends in medical costs over time. For example, in 2011 dollars, the average cost of an ED visit for individuals age 65 and older increased 40% over just 10 years from $630 in 2001 to $880 in 2011 . Similarly, daily inpatient costs increased 30% from $2400 to $3200.