Abstract

This study examines the health impacts of recent heat waves statewide and for six subregions of California: the north and south coasts, the Central Valley, the Mojave Desert, southern deserts, and northern forests. By using canonical correlation analysis applied to daily maximum temperatures and morbidity data in the form of unscheduled hospitalizations from 1999 to 2009, 19 heat waves spanning 3–15 days in duration that had a significant impact on health were identified. On average, hospital admissions were found to increase by 7% on the peak heat-wave day, with a significant impact seen for several disease categories, including cardiovascular disease, respiratory disease, dehydration, acute renal failure, heat illness, and mental health. Statewide, there were 11 000 excess hospitalizations that were due to extreme heat over the period, yet the majority of impactful events were not accompanied by a heat advisory or warning from the National Weather Service. On a regional basis, the strongest health impacts are seen in the Central Valley and the north and south coasts. The north coast contributes disproportionately to the statewide health impact during heat waves, with a 10.5% increase in daily morbidity at heat-wave peak as compared with 8.1% for the Central Valley and 5.6% for the south coast. The temperature threshold at which an impact is seen varies by subregion and timing within the season. These results suggest that heat-warning criteria should consider local percentile thresholds to account for acclimation to local climatological conditions as well as the seasonal timing of a forecast heat wave.

1. Introduction

The devastating effects of extreme-heat events have been seen in recent years. The 2003 European heat wave and the 2010 Russian heat wave each resulted in tens of thousands of deaths (Agence France-Presse 2013; Robine et al. 2007), and the 2006 California heat wave killed more than 600 (Trent 2007; Ostro et al. 2009) and resulted in over 16 000 excess hospital emergency-department visits (Knowlton et al. 2009). Adding to the tragedy of these losses is the fact that most heat-related deaths are preventable with adequate warning tools and effective emergency planning. Since climate change has the potential to increase the frequency of these types of events (Meehl and Tebaldi 2004; Solomon et al. 2007; Allen et al. 2012), improved heat-warning systems are urgently needed. This would require a better knowledge of the full impact of extreme heat on morbidity and mortality.

California has unique challenges for heat-wave preparedness owing to its diversity of population and climate zones. Some residents live in desert conditions just inland of coastal populations who are used to relatively mild temperatures. In addition, many residents lack air conditioning, especially along the coast, making them particularly vulnerable during extreme-heat events (Sailor and Pavlova 2003; Reid et al. 2009). This vulnerability was apparent during the 2006 California heat wave that affected most of the state. Health-impact studies of that heat wave showed that although temperatures were hotter inland the health impacts were stronger along the coast (Knowlton et al. 2009; Gershunov et al. 2011). The 2006 heat wave was unusually humid and nighttime temperatures were unprecedented; therefore, the nighttime recovery that is typical for California was stifled. Recent work shows a clear trend in humid heat waves in the western United States, with a disproportionate increase in nighttime temperatures (Gershunov et al. 2009; Bumbaco et al. 2013). In fact, California coastal communities are becoming increasingly susceptible to midsummer, humid heat waves to which they are not accustomed (Gershunov and Guirguis 2012, hereinafter GG12). Thus, coastal populations may be at a higher risk for heat-related illness in the short- and long-term future since they are neither physiologically nor technologically acclimatized to this type of heat.

California heat-alert criteria rely on the heat index, which is based on empirical relationships between temperature/humidity thresholds and mortality in a few major U.S. cities. The national heat index threshold of 105 used by the National Weather Service (NWS) to issue a heat warning does not work well in California. Desert communities regularly exceed this threshold, but residents are well adapted to extreme heat. Coastal communities rarely exceed this threshold but are much more vulnerable to heat illness because they are accustomed to much milder conditions. For effective weather warnings, events posing a danger to health should be identified locally with higher or lower thresholds directed at populations living in hotter or cooler climates, respectively (Robinson 2001). In addition, mortality only accounts for a small portion of acute health effects; therefore, for effective preparedness nonfatal illness should also be considered. Local NWS offices typically modify the alert criteria to better suit California conditions, but these decisions are put in place with only limited information about local heat–health relationships. Such information would be highly beneficial for making informed decisions about when to issue a warning, which could prevent heat-related illnesses and save lives.

There have been many studies investigating the impacts of extreme heat on human health (e.g., Basu and Samet 2002; Martiello and Giacchi 2010). Most of this work has focused on health impacts that are related to daily ambient apparent temperatures observed throughout the summer (e.g., Basu 2009; Basu et al. 2008, 2012), however, with much less attention given to health outcomes from one heat wave to another. Health-impact studies of heat waves (multiple days of hot weather) have primarily focused on a few notorious heat waves such as those in Chicago in 1995 (e.g., Kaiser et al. 2007), Europe in 2003 (e.g., Le Tertre et al. 2006), or California in 2006 (e.g., Knowlton et al. 2009). Long-term heat-wave studies require some definition of a heat wave, and there is no universal definition. Heat waves are usually defined by magnitude as days exceeding a set temperature or percentile threshold and may also include a duration requirement (e.g., Hajat et al. 2006; Mastrangelo et al. 2007; Son et al. 2012; Vaneckova and Bambrick 2013) or by synoptic weather type (e.g., Sheridan et al. 2012; Sheridan and Kalkstein 2010; Vaneckova et al. 2008), and the impact of heat on health is subsequently quantified. Our method takes a new approach by using health and meteorological data simultaneously to identify dangerous historical heat waves that occurred in California between 1999 and 2009. The advantage is that we make no a priori assumption about the kinds of conditions that affect human health. This approach leaves open the possibility of detecting a health impact during events that might not typically be considered as extreme. Coastal communities that are not well acclimated to heat, for example, may be adversely impacted during a heat wave at lower temperatures than would inland communities.

In addition to statewide impacts, this study also investigates regional impacts using six regions defined empirically on the basis of heat-wave expression over the state’s complex geography. These are the north and south coasts, the Central Valley, the Mojave Desert, southern deserts, and northern forests. Heat-risk warnings issued by the NWS during specific heat waves are considered in the context of actual health risks as measured by excess hospitalizations. Beyond improved understanding of the meteorological impacts on heat illness, we aim for the results presented below to be practical and useful for optimizing the effectiveness of regional heat warnings.

2. Data

a. Climate and weather data

The daily maximum temperatures (Tmax) are from Maurer et al. (2002) and are composed of daily station data interpolated onto a regular 12 km × 12 km grid with temperature lapsed to gridcell center elevations. The source station data are from the National Climatic Data Center (NCDC) first-order Automated Surface Observing System and cooperative observer (coop) summary of the day (NCDC 2009). Grid-to-station differences would depend on location, with larger differences seen in areas of complex elevation or strong spatial temperature gradients over relatively short distances. For example, the 1999–2009 summertime temperature record from the downtown San Francisco coop station is approximately 0.13°C warmer than that of its nearest grid cell. This is because the grid cell represents temperatures over a larger area, including parts immediately on the coast. For this study, grid cells that were collocated with the zip-code-level health data (described below) were averaged over six predefined California subregions to give regional daily maximum temperatures from 1999 to 2009. These subregions were defined empirically using principal components analysis applied to temperature data, as described in GG12. As a result of the regionalization, localities grouped within a given region exhibit a similar temporal variability in heat-wave activity. These six regions are shown in Fig. 2b, which is described in more detail below.

Daily specific humidity (SH) data are from the North American Regional Reanalysis (Mesinger et al. 2006). The SH data were processed just as for Tmax to give daily, regional averages of specific humidity.

b. Health-outcome data

Data were compiled from the Office of Statewide Health Planning and Development patient discharge (PD) data for the warm season (May–September) spanning 1999–2009. These data were limited to include only hospitalizations at acute-care facilities that were designated as unscheduled so that they would represent a subset of emergency-department visits in which conditions were serious enough to require hospitalizations. The data were aggregated to zip-code level to protect patient privacy and include zip code, date of hospital admission, day of week, counts for each health outcome category, and stratification by age and race/ethnicity. The outcome categories (Table 1) included all cardiovascular diseases and cardiovascular subcategories (ischemic heart disease, acute myocardial infarction, cardiac dysrhythmias, and essential hypertension), all respiratory diseases, acute renal failure, mental health, dehydration, and heat illness. We also considered “all causes” as an outcome category, which was taken as the sum of all outcome categories listed in Table 1.

Table 1.

Average daily increase in hospital admissions with confidence intervals for HHE span and peak from a two-sample t test. An asterisk indicates a nonsignificant health impact (at the 95% level). Also shown are the daily average number of hospitalizations in California over the 1999–2009 record and the excess admissions seen on peak heat-wave days, expressed as percent above normal. ICD is an acronym for International Classification of Diseases.

Average daily increase in hospital admissions with confidence intervals for HHE span and peak from a two-sample t test. An asterisk indicates a nonsignificant health impact (at the 95% level). Also shown are the daily average number of hospitalizations in California over the 1999–2009 record and the excess admissions seen on peak heat-wave days, expressed as percent above normal. ICD is an acronym for International Classification of Diseases.
Average daily increase in hospital admissions with confidence intervals for HHE span and peak from a two-sample t test. An asterisk indicates a nonsignificant health impact (at the 95% level). Also shown are the daily average number of hospitalizations in California over the 1999–2009 record and the excess admissions seen on peak heat-wave days, expressed as percent above normal. ICD is an acronym for International Classification of Diseases.

For the purposes of this study, the PD data were aggregated regionally in the same regions as for Tmax. The regional PD data were then filtered to remove periodic signals and long-term trends. Because of annual variation in viral activity and other causes, morbidity is higher in spring and autumn than in midsummer. This annual and semiannual harmonic seasonal cycle was removed using least squares regression analysis. Hospitalizations are notably lower on weekends and holidays. The weekly cycle was removed by subtracting the long-term day-of-the-week average from each daily admissions count, and the holiday effect was removed by subtracting the average holiday admission count from each Memorial Day, Labor Day, or Independence Day holiday. Very low admissions also occurred on 5 July when Independence Day occurred on a Sunday, since most Americans would have received a work holiday on that Monday. During those years, 5 July was treated as a holiday. Last, any long-term trend in the data was removed using locally weighted polynomial regression and scatterplot smoothing (LOWESS; Cleveland 1979), which fits a quadratic curve to the nearest 30% of data points. Figure 1 provides an illustration of the filtering process for the north-coast region.

Fig. 1.

(a) PD data, (b) PD data and weekly plus seasonal cycle for 2006, (c) filtered PD data after removing trend, seasonal and weekly cycles, and holiday effects for 2006.

Fig. 1.

(a) PD data, (b) PD data and weekly plus seasonal cycle for 2006, (c) filtered PD data after removing trend, seasonal and weekly cycles, and holiday effects for 2006.

c. Historical heat-alert/heat-advisory information

Historical information about heat advisories or alerts issued by the NWS was obtained from the National Oceanic and Atmospheric Administration (NOAA) Hierarchical Data Storage System (available online at http://hurricane.ncdc.noaa.gov/pls/plhas/HAS.FileAppSelect?datasetname=9957ANX) for the nonprecipitation warnings, watches, and advisories category as well as from other internal NWS communications.

d. Historical electrical-alert information

Information on historical power alerts issued by the California Independent System Operator (ISO) was obtained from the ISO alert, warning, and emergency records from 1998 (available online at http://www.caiso.com/Documents/Alert_WarningandEmergenciesRecord.pdf). Local utilities generally follow the ISO recommendations and issue their own alerts to promote conservation. On occasion, however, a local utility may issue an alert unaccompanied by the ISO. For this research, we only had access to those alerts issued by the California ISO.

3. Methods

a. Canonical correlation analysis

This study uses canonical correlation analyses (CCA) to identify space–time patterns of heat-wave expressions optimally related to morbidity. CCA is a multivariate statistical approach used to linearly summarize information contained in the cross-correlation matrix between two sets of variables, in this case, daily maximum temperatures and morbidity as represented by hospitalizations. CCA transforms the original data pairs (x and y) into new variables called canonical variates, defined as

 
formula
 
formula

where x (Tmax) and y (PD) are the centered data vectors (standardized for this study), I is the number of elements in x, J is the number of elements in y, and m is the number of pairs of canonical variates that can be obtained from the two datasets and is equal to the lesser of I and J. Each canonical variable υ and w is a linear combination of elements of the respective data vectors, or in other words, a weighted average with weights given by a and b in the above equations (Wilks 2006). Pairs of canonical variates are ordered sequentially by the degree of correlation between υ and w, such that the first pair (CC1) exhibits the maximum canonical correlation. CCA was originally developed by Hotelling (1935, 1936) to identify and quantify associations between two sets of variables and was initially used in the social sciences. In climate prediction, CCA has been used to match patterns in two fields of variables, typically with the intention to forecast one with the other, that is, the predicted with the predictor (Barnett and Preisendorfer 1987; Gershunov and Cayan 2003; Alfaro et al. 2006). Here, we use CCA as a purely diagnostic tool to identify periods in the recent historical record in which daily maximum temperatures and morbidity were strongly correlated. We have previously used CCA to identify heat and humidity effects on county-level emergency-department visits over California in a limited emergency department dataset spanning only 1 yr (2006) and resolving the impacts of only one heat wave (Gershunov et al. 2011). Here, the daily input Tmax (x′) and PD (y′) data arrays span 11 years, are regionally averaged, and are filtered and standardized to filter out local noise and to remove population-density bias while focusing on meteorologically relevant regions. For the purposes of discussion, we refer to the first canonical variable υ1 as CC1Tmax and the first canonical variable w1 as CC1Health (Fig. 2a). These variables represent the simultaneous pattern of strongest linear coevolution of temperature and hospitalizations throughout California (Fig. 2b). Higher-order canonical modes were poorly related to temperatures and health, and therefore, although they do explain some heat–health covariability, we focus our analysis on the primary mode, which best represents the spatial–temporal pattern of heat-related health outcomes in California.

Fig. 2.

(a) Canonical variables CC1Tmax (black curve) and CC1Health (green curve) and (b) homogeneous correlation maps showing the correlation coefficients between the input data vectors and their associated canonical variables [e.g., correlation (left) between x′ and CC1Tmax and (right) between y′ and CC1Health]. The input data were regionalized using empirically defined California subregions: Central Valley (CV), southern deserts (SD), north coast (CN), south coast (CS), northern forests (NF), and Mojave Desert (MJ).

Fig. 2.

(a) Canonical variables CC1Tmax (black curve) and CC1Health (green curve) and (b) homogeneous correlation maps showing the correlation coefficients between the input data vectors and their associated canonical variables [e.g., correlation (left) between x′ and CC1Tmax and (right) between y′ and CC1Health]. The input data were regionalized using empirically defined California subregions: Central Valley (CV), southern deserts (SD), north coast (CN), south coast (CS), northern forests (NF), and Mojave Desert (MJ).

In preliminary analyses, we tested our method with daily minimum temperatures (Tmin). The Tmax results were found to be more robust, however, in terms of both the correlation between the canonical variable and the source data (i.e., the correlation between CC1 and x′) and the correlation between the two canonical variables (i.e., υ and w). Therefore, the results using Tmax were superior in describing the heat–health relationship in California. This is likely because California experiences both dry and humid heat waves and because Tmax is elevated during both varieties whereas Tmin may not be strongly elevated during dry events.

b. Identifying heat–health events

We identified those heat waves in the 11-yr record that had an impact on human health by looking for cases in which three criteria were met: 1) canonical variables CC1Tmax and CC1Health (Fig. 2a) were significantly correlated (at the 95% level, correlation coefficient r > 0.51) using a running 15-day window, 2) a strong temperature anomaly was observed, as represented by canonical variable CC1Tmax crossing a threshold of 1 standard deviation, and 3) a strong health anomaly was observed as represented by canonical variable CC1Health crossing a threshold of 1 standard deviation. To allow for some flexibility in timing, a heat–health event (HHE) is defined to span the full duration of the heat anomaly from when it first becomes warm (CC1Tmax is positive), to when it peaks (at least once), and then to when it drops back to normal again. The health impact can occur at any point within this heat event, which allows for lags in response and additionally allows us to quantify the full health impact of an individual heat wave.

4. Results

a. Heat–health events

Figure 2a shows the first pair of canonical variates CC1Tmax and CC1Health. These time series are only moderately, yet significantly, correlated (r = 0.3) over the 11-yr record, signifying that, as expected, disease processes associated with heat are not the main cause of morbidity. Over shorter intervals, however, the relationship between heat and illness can become much stronger than the long-term average (Fig. 3). For example, using a 15-day running window, the correlation reaches 0.79 and 0.82 during the July 2006 and July 2003 heat waves, respectively. Figure 2b gives the homogeneous correlation maps showing how well each of the input data vectors are represented by their canonical variates. CC1Tmax best represents Tmax on the north coast (r = 0.94) and also does reasonably well in capturing Tmax variability on the south coast (r = 0.62), the Central Valley (r = 0.55), and the southern deserts (r = 0.51), while the Mojave Desert and the northern forests are weighted less strongly (r = 0.41 and r = 0.40, respectively). A similar regional pattern is observed for the health results, although with generally weaker correlations. CC1Health best represents hospitalizations in the north coast (r = 0.82) followed by the south coast (r = 0.62) and the Central Valley (r = 0.54), while hospitalizations in the southern deserts, the Mojave Desert, and the northern forests are not well represented by CC1Health (r < 0.23). The heavy weighting of the north coast in both CC1Tmax and CC1Health highlights the sensitivity/vulnerability of this region’s population to extreme heat. This is a region for which summers are typically cool because of the proximity of cool coastal Pacific Ocean waters, which is further enhanced by marine layer clouds. Coastal heat waves therefore do not need to be as hot as those over the hotter and air-conditioned inland to come as a stark contrast to typical conditions and to catch residents unprepared. This result is consistent with recent studies that considered the health impacts of the 2006 heat wave and also found an increased sensitivity to heat in this region (Knowlton et al. 2009; Gershunov et al. 2011).

Fig. 3.

Panels with green and black time series represent the canonical variables shown for each year of the analysis period. The 15-day running correlation between CC1Tmax and CC1Health is shown in blue if statistically significant. HHEs are shown in red.

Fig. 3.

Panels with green and black time series represent the canonical variables shown for each year of the analysis period. The 15-day running correlation between CC1Tmax and CC1Health is shown in blue if statistically significant. HHEs are shown in red.

Using the criteria described in section 3 to identify HHEs—namely, strong positive anomalies observed in both CC1Tmax and CC1Health as well as a significant correlation between them over a 2-week period, we identified 19 heat waves with a significant impact on human health. These HHEs are outlined in red in Fig. 3, and additional details such as peak date, duration, and whether a power alert or NWS heat advisory/warning was issued are provided in Table 2. These results show that at least one HHE occurred each year except in 1999 and 2005 and that five years (2000, 2001, 2003, 2006, and 2009) had more than one event. Records show an NWS heat warning was issued for only 6 of the 19 events. The strongest health signal is seen for 12–16 June 2000, when CC1Tmax and CC1Health both exceeded 3.9 standard deviations above normal (Fig. 3). This event occurred during the California energy crisis (e.g., Sweeney 2002), when market deregulation and high energy prices caused power shortages. In fact, on the peak date of 14 June, rolling blackouts affected 97 000 customers in northern California (Bergman 2001) while temperatures in San Francisco reached 105°F. High energy prices were passed on to consumers during this time, which could have influenced personal decisions about air-conditioning (AC) use, even where AC was available. Of the five HHEs that occurred during the 2000–01 crisis, four were accompanied by a power alert. Other summers that stand out as remarkable are the summer of 2003, which had six HHEs of varying strengths and duration, and the summer of 2006 for the duration and intensity of the midsummer heat wave.

Table 2.

Heat–health events and associated statewide health impact. Boldface font indicates statistical significance at the 90% level. The “electrical alert” column indicates whether a stage-1 (S1), stage-2 (S2), or stage-3 (S3) electrical alert was issued by the California ISO.

Heat–health events and associated statewide health impact. Boldface font indicates statistical significance at the 90% level. The “electrical alert” column indicates whether a stage-1 (S1), stage-2 (S2), or stage-3 (S3) electrical alert was issued by the California ISO.
Heat–health events and associated statewide health impact. Boldface font indicates statistical significance at the 90% level. The “electrical alert” column indicates whether a stage-1 (S1), stage-2 (S2), or stage-3 (S3) electrical alert was issued by the California ISO.

Figure 4a (Fig. 4b) shows the peak daily maximum and minimum temperatures (standardized anomalies) for each HHE. Also in Fig. 4a are the number of days exceeding the warm-season (May–September) 95th percentile for each HHE and region and an indicator of whether the monthly 95th percentile was reached (marked with asterisks). From Fig. 4a, the Central Valley and southern deserts were hottest during all events, with daytime temperatures usually exceeding 37°C (~98°F). From Fig. 4b, the north coast tends to reach hotter temperatures (especially daytime) relative to its climatological values as compared with other regions. This means that, while the temperatures may be lower, these coastal residents are experiencing heat conditions that are very extreme relative to what they are used to, and they may experience health impacts at lower temperatures than inland populations that are more acclimatized to heat. This figure also highlights the notorious 2006 heat wave affecting most of California, which was unprecedented in magnitude and spatial extent since at least the late 1940s (Gershunov et al. 2009), and nighttime temperatures are shown to be even more extreme than those experienced during the day. Figure 4b also shows that, during nighttime-accentuated events, when Tmin is extremely elevated, multiple regions are affected (e.g., July 2002, July 2003, and July 2006), indicative of a particularly expansive heat wave with the potential for large-scale, statewide impacts.

Fig. 4.

(a) Regionally averaged peak temperature for (top) Tmax and (bottom) Tmin during each HHE, with the number of days exceeding the summertime 95th percentile shown in blue text and an asterisk indicating whether the monthly 95th percentile was reached. (b) Standardized Tmax and Tmin anomaly on peak days. Note here that the peak day is calculated regionally for each variable (Tmin and Tmax do not necessarily peak on the same day) and may vary slightly from the peak day given in Table 2.

Fig. 4.

(a) Regionally averaged peak temperature for (top) Tmax and (bottom) Tmin during each HHE, with the number of days exceeding the summertime 95th percentile shown in blue text and an asterisk indicating whether the monthly 95th percentile was reached. (b) Standardized Tmax and Tmin anomaly on peak days. Note here that the peak day is calculated regionally for each variable (Tmin and Tmax do not necessarily peak on the same day) and may vary slightly from the peak day given in Table 2.

It is interesting that in terms of summertime temperatures these events often do not fall in the top 5% of daytime or nighttime highs, except on the north coast. If we look at monthly percentiles, however, the majority of events do fall in the top 5% (above the 95th percentile) for all regions except the south coast. This finding means, for example, that, although it might not meet the summertime threshold for extreme temperatures, a May event would be extremely hot for that time of year. This result suggests populations may be more heat sensitive during cooler parts of the season. An increased vulnerability early in the season has also been found in other studies (e.g., Basu and Samet 2002; Ebi et al. 2004). This result is generally attributed to the loss of acclimatization that occurs during the winter as well as to mortality displacement whereby the most vulnerable populations succumb to the first dangerous event of the season (Basu and Malig 2011).

b. Statewide health impact

The statewide heat–health impacts are shown in Fig. 5 and Tables 1 and 2. Figure 5a shows the distribution of daily hospitalization anomalies during non-HHE days, during the span of a HHE, and during the peak HHE day for all causes. There is a dramatic increase in hospitalizations during HHEs, especially at the peak day. This difference is statistically significant at the 95% level using a two-sample t test to compare sample means. For all causes, there was an average daily increase of 102 hospitalizations during the HHE span, which increased to 173 excess hospitalizations on the peak day. In California during the 1999–2009 record, there were, on average, 2519 hospitalizations per day. Therefore, 173 excess hospitalizations at the peak represent nearly a 7% increase above what would occur on an average day.

Fig. 5.

(a) Boxplot showing statewide hospitalization anomalies for non-HHE days (n = 1544), HHE span (n = 139), and HHE peak (n = 19), and (b) morbidity associated with each event using the resampling method (see text). In (b), the boxplots show the distribution of historical non-HHE days spanning the same duration as the HHE, and green markers give the cumulative health impact for each HHE (filled markers indicate statistical significance at the 90% level).

Fig. 5.

(a) Boxplot showing statewide hospitalization anomalies for non-HHE days (n = 1544), HHE span (n = 139), and HHE peak (n = 19), and (b) morbidity associated with each event using the resampling method (see text). In (b), the boxplots show the distribution of historical non-HHE days spanning the same duration as the HHE, and green markers give the cumulative health impact for each HHE (filled markers indicate statistical significance at the 90% level).

A similar comparison was done for each of the disease categories, accounting for unequal variances as necessary for some categories. A statistically significant increase in hospitalizations was seen for all outcomes except essential hypertension, a cardiovascular subcategory. This nonsignificant result for essential hypertension could be physiological, because blood pressure goes down with increased heat exposure (Basu et al. 2012), or it could be due to the small sample size (Table 1). Hospitalizations due to all cardiovascular diseases increased by an average of 36 (49) per day for the HHE span (peak), or 3.5% (4.7%) above the typical cardiovascular-disease admission rate (Table 1). Admissions for respiratory diseases, mental health, acute renal failure, dehydration, and heat illness increased by 20 (42), 2 (6), 5 (10), 8 (16), and 4 (10) per day, respectively, on average for HHE span (peak). Expressed as a percentage of daily mean hospitalizations during the record for these disease categories, these numbers translate to an increase of 7.7%, 9.8%, 17.7%, 22.5%, and 505% at the peak of the heat wave.

Figure 5b gives the cumulative statewide health impact for each of the 19 events. To quantify the cumulative impact, hospitalization anomalies y′ were summed over the span of each event, and this value was compared with all non-HHE days in the historical record spanning the same duration. For example, the impact of the 13–26 July 2006 event was determined by comparing that 14-day sum of hospitalization anomalies with those obtained by resampling the record for all consecutive, non-HHE days spanning 14 days. In a similar way, the 12–16 June 2000 event was compared with non-HHE days spanning 5 days. The health impact is said to be significant (using the 90th-percentile level for a one-sided test) if it falls above the 95th percentile of the resampled distribution. From Fig. 5b, a significant statewide health impact is observed for 15 of the 19 HHEs identified. Taken together, these 15 events are associated with more than 11 000 excess hospitalizations statewide. The number of excess hospitalizations associated with each of these 15 events (Table 2) ranges from 367 for the short, 3-day 2006 HHE to 1657 for one 17-day heat wave in September of 2004. During the notorious 14-day July 2006 heat wave that affected most of the state, there were 1254 excess hospitalizations in California, a result that is similar to that of Knowlton et al. (2009), who found 1182 excess hospitalizations and 16 166 excess emergency-department visits. The magnitude of the impact from one event to another is strongly associated with its duration. By looking at the health impact in terms of quantiles of the resampled data, which accounts for duration, we see that the health impacts of these 15 events are all in the top 3%, with several in the top 1%, as compared with non-HHE days spanning the same duration (Table 2); that is, they are highly significant.

c. Regional health impact

Table 3 shows the regional impacts for HHE span and peak for each of the six California subregions using a two-sample t test comparing non-HHE days with HHE span and HHE peak (same method as for disease categories in Table 1). A significant health impact is observed for four of the six regions. Daily admissions for the south coast, north coast, Central Valley, and southern deserts increased by 50 (71), 23 (47), 24 (43), and 4 (9), respectively, for HHE span (peak). Expressed as a percentage of average daily hospitalizations for these regions, these numbers translate to an increase of 5.6%, 10.5%, 8.1%, and 6.3% at the peak of the heat wave. While the largest impact in terms of total admissions is greatest for the south coast owing to its larger population, the north coast contributes disproportionately to the health impact during heat waves. This region represents 18% of all California hospitalizations during 1999–2009, and this percentage increases to 27% during heat waves. There are a few possible explanations for this situation. First, there may be a poor acclimation to extreme heat both physiologically and through air-conditioning use [air-conditioning coverage is low in the north coast; for example, San Francisco has only 21% air-conditioning saturation (Sailor and Pavlova 2003)]. Second, residents there typically have easier access to hospitals and better insurance coverage than do residents of other parts of the state, making it more likely that they would seek medical treatment. Third, heat waves on the north coast are hotter relative to the mean climate (cf. Fig. 4b) simply because of the regional temperature distribution.

Table 3.

Similar to Table 1, but for the six California subregions and using the all-causes outcome category.

Similar to Table 1, but for the six California subregions and using the all-causes outcome category.
Similar to Table 1, but for the six California subregions and using the all-causes outcome category.

Figure 6 gives the cumulative health impact of the 19 HHEs for each California subregion using the same resampling method described above. Eighteen HHEs are associated with a significant health impact in at least one subregion of California (7–17 August 2009 is the only exception). There is very little impact seen in the deserts with only two events (one event) associated with a significant health impact in the southern deserts (Mojave Desert). This result is likely due to the fact that CC1 does not represent desert heat waves well, especially for the Mojave (cf. Fig. 2b). CC1 also does poorly in representing heat waves in the northern forests, and here we see only a modest impact (four significant events). Our method is designed to explain the strongest co-relationships between heat and health in California. Not all heat waves are represented, however. There are likely some additional, more regionally focused heat waves that may have a regional health impact, but these would need to be studied on a smaller spatial scale.

Fig. 6.

As in Fig. 5b, but for the six California subregions and with a color scale that shows the impact in terms of the percentile of the resampled distribution.

Fig. 6.

As in Fig. 5b, but for the six California subregions and with a color scale that shows the impact in terms of the percentile of the resampled distribution.

For the remaining regions, there is a strong health signal. Morbidity in the Central Valley was significantly affected during nine events, and the north and south coasts each experienced 11 impactful heat waves. The south coast, the most populous region, shows the strongest overall impact in terms of patient numbers, with excess hospitalizations in the range from 300 to more than 800, depending on event. In general, excess hospitalizations on the south coast are 1.5–3 times those in the north coast or Central Valley, where the most intense heat waves cause typically 300–400 excess hospitalizations. In terms of quantiles of historical observations (color scale in Fig. 6), which equalizes the regions in terms of population, the health impact is similar across the three regions, with a mean percentile rank of 86%–89% for the 19 events, although the coastal regions see a larger number of impacts in the top 5% (i.e., more are significant at the 90% level).

Figure 7a shows the peak temperatures for those HHEs identified as having a significant regional health impact (significant HHEs; hereinafter SHHEs) in the context of the full Tmax distribution by region and timing within the season, and Fig. 7b gives the results in terms of degrees above normal. A health impact is seen in the Central Valley for Tmax in the range of 33°–42°C (92°–108°F) depending on month. For the north and south coasts, where mean summertime temperatures are much lower, a health impact is seen for temperatures reaching 27°–36°C (81°–97°F). Most impacts occur at temperatures above the 90th percentile in the Central Valley and north coast, with many falling above the 95th or even the 99th percentile. An exception is one event (5–9 September 2000) that affected the Central Valley for which temperatures peaked at 33°C (92°F), which is approximately the 70th percentile for that region in September. The south coast is more vulnerable, with health impacts seen for four SHHEs having peak temperatures at or below the 85th percentile. Relative to monthly normal conditions, Tmax is elevated by an average of 6.1°C in the Central Valley, 9.1°C on the north coast, and 4.5°C on the south coast. From Fig. 7a there are several observations above the 99th percentile, including 11-yr highs in on the south coast that were not associated with a significant health impact. We analyzed those 13 days, which spanned five separate events. Four of the five events were not identified as an HHE by the CCA method and therefore were not examined in terms of health impacts. The reason they did not meet the HHE criteria is because, on the large scale, there was no strong correlation between temperatures and health in California. It is possible there were more-localized health impacts, however. The fifth event (13–26 July 2006) was identified as an HHE, but no significant health impact was found for the south coast. This was a particularly impactful heat wave statewide (cf. Table 2), and on the south coast there were 3 days above the 99th percentile and an 11-yr high that occurred on 22 July. While the health impact was not statistically significant by our criteria, on the south coast there were 515 excess hospitalizations during that 14-day period, which is above the 93rd percentile as compared with all other 14-day periods in the record.

Fig. 7.

Distribution of Tmax by region and month: (a) Tmax on the peak day of those HHEs identified as having a significant health impact and (b) showing results as degrees Celsius above normal.

Fig. 7.

Distribution of Tmax by region and month: (a) Tmax on the peak day of those HHEs identified as having a significant health impact and (b) showing results as degrees Celsius above normal.

d. Effect of humidity

Current heat-warning systems attempt to account for the effect of humidity on heat-wave morbidity and mortality. The heat index uses relative humidity or dewpoint temperature to estimate the human health impact on the basis of empirical relationships with mortality. More-sophisticated systems use empirical relationships between morbidity and forecast synoptic air masses (e.g., Ebi et al. 2004). Quantifying the heat–humidity–health relationship is beyond the scope of this study, but we attempt to describe the relative impact of dry versus humid heat waves using regional anomalies in specific humidity. An event is categorized as dry or humid depending on whether the daily, regional specific humidity was below or above normal, respectively. The anomalies are calculated using the 1999–2009 May–September climatological averages.

Figure 8 shows the proportion of hospitalizations by month and heat-wave type for those HHEs found to have a significant regional impact on health. For the Central Valley and north coast, humid heat waves account for 65% of the impactful heat waves in each region (6 of 9 in the Central Valley and 7 of 11 on the north coast). In terms of health impact, humid heat waves account for 66% of the hospitalizations in each of these regions, with no appreciable difference seen within our sample of impactful events in terms of health outcome during dry versus humid heat waves. Because health data were used directly in the identification of these events, however, the fact that the majority of impactful heat waves are humid suggests that humid heat waves are the more dangerous variety in the Central Valley and north coast. For the south coast, dry and humid events are approximately equal both in occurrence rate and health impact. In terms of seasonal timing, midsummer events have the strongest impact on health in the Central Valley, but early season heat waves have the strongest impact in the coastal regions. In general, early and midseason heat waves tend to be getting stronger and more frequent in California because of climate change and are specifically trending toward the humid variety (GG12). Since the coastal populations show more vulnerability early in the season because of loss of acclimation over the winter and mortality displacement as discussed above and are more prone to heat illness during humid events (at least in the northern part of coast), this means heat waves are changing toward the most dangerous variety in terms of human health.

Fig. 8.

Proportion of hospitalizations by month and heat-wave type. Individual events are separated by white horizontal lines.

Fig. 8.

Proportion of hospitalizations by month and heat-wave type. Individual events are separated by white horizontal lines.

5. Discussion and conclusions

This study investigated the health impacts of recent heat waves from 1999 to 2009. Using canonical correlation analysis applied to daily maximum temperatures and hospitalization data, we identified 19 heat events spanning 3–15 days in duration that had a significant impact on human health. Taken collectively, these events resulted in more than 11 000 excess hospitalizations statewide. A heat advisory or warning from the National Weather Service was only issued during six of them, however. In terms of individual heat waves, the 17-day September 2004 heat wave showed the greatest impact, with 1657 excess hospitalizations. The 14-day July 2006 heat wave and the 15-day July 2003 heat wave were also very harmful to health with 1254 and 1063 excess hospitalizations, respectively. These events were not only long in duration but were particularly extreme, with temperatures exceeding the 95th percentile for several days. The 2003 and 2006 events were additionally humid, with very high nighttime temperatures that hindered physiological recovery at night. Previous research has shown that heat waves in the Southwest are becoming more durable and spatially expansive, especially the humid variety (Gershunov et al. 2009). Therefore, local and statewide planning is needed to adequately prepare for these types of events, which can be devastating in terms of health impacts and can greatly strain resources designated for emergency response.

On a regional basis, the strongest health impacts were seen in the Central Valley and the north and south coasts. Although the largest impact in terms of patient numbers is greatest for the south coast owing to its larger population, the north coast is disproportionately affected by extreme heat. During heat waves, the south coast has a 5.6% increase in hospitalizations on peak heat-wave days while the north coast experiences an increase of 10.5%.

In the Central Valley, temperatures in the range of 33°–42°C (6.1°C above normal, on average) were associated with a health impact, whereas for both coastal regions we detected an impact for temperatures in the range of 27°–36°C (9.1° and 4.5°C above normal, on average, for the north and south coasts, respectively). In general, these temperatures are above the 90th percentile in the Central Valley and north coast whereas the south coast is more vulnerable, with impacts occurring when temperatures are at or below the 85th percentile. The north coast appears to be most vulnerable in terms of increased hospitalizations (10.5% increase on peak heat waves), but the south coast appears to be more vulnerable to lower temperatures. This contrast could be due to differences in demographics or access to care. In addition, there are factors other than high temperatures that are possibly contributing to the observed health effects, such as air pollution, Santa Ana winds, or smoke from wildfires that often accompany dry, late season heat waves on the south coast.

The relative impact of dry versus humid heat waves was investigated using regional anomalies in specific humidity. The results showed that humid heat waves have a stronger impact on human health in the Central Valley and north coast, accounting for 66% of heat-related excess hospitalizations in both regions. In the south coast there was an approximately equal impact seen during humid and dry heat waves. In terms of seasonal timing, midsummer events have the strongest impact on health in the Central Valley, but early season heat waves have the strongest impact in the coastal regions. Early in the season, coastal California experiences many cloudy and cool days that result from the prominent marine layer, often referred to as “May Gray” or “June Gloom.” Therefore, heat waves during this part of the season would come as a stark contrast to typical conditions.

These results suggest that local percentile thresholds that consider seasonal timing would be more appropriate for use in issuing heat warnings than is the current system, which uses a single threshold throughout the summer and regional baselines that are based on only very limited health-impact information. New criteria developed by NWS San Diego use a temperature curve that is based on departures from normal for different climate zones, therefore incorporating seasonality and local acclimatization. This approach will address some of the geographic and population differences in vulnerability. California could also benefit from a multitiered system that accounts for the vulnerabilities of different populations such as outdoor agricultural workers, the elderly, and those with preexisting conditions who have been shown to be especially vulnerable to heat (e.g., Trent 2007). Lower threshold warnings could be issued for these vulnerable populations. This kind of analysis is beyond the scope of this study, but future work will take a more localized focus and consider local differences in outcome that are based on demographic and other risk factors or exasperating conditions such as air quality, occurrence of Santa Ana winds, or marine-layer conditions. Given that heat waves are expected to become more frequent and more severe, it is crucial to understand the impact on human health now so that public health officials can respond effectively and plan adequately for the future. This is especially true for California, which has a population of nearly 40 million, with the majority living along the coast, where heat acclimation is poor, air conditioners in homes are sparse (especially in Northern California), and research shows that heat waves will continue to become more intense and more humid.

Acknowledgments

This work was supported by the University Corporation for Atmospheric Research (UCAR) Postdocs Applying Climate Expertise (PACE) fellowship (32947252), by the U.S. Department of the Interior via the Southwest Climate Science Center, by NOAA via the RISA program through the California and Nevada Applications Center, and by the National Science Foundation Awards ANT-1043435 and DUE-1239797. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the funding sources. We thank Mary Tyree for data retrieval and handling. We thank two anonymous reviewers for helpful comments during the evaluation of this paper.

REFERENCES

REFERENCES
Agence France-Press
, cited
2013
: Russian heat wave caused 11,000 deaths in Moscow: Official. Press release, 17 September 2010. [Available online at http://www.google.com/hostednews/afp/article/ALeqM5ih6EGeelXvvrhgivI66OKFJ1MFLA.]
Alfaro
,
E.
,
A.
Gershunov
, and
D. R.
Cayan
,
2006
:
Prediction of summer maximum and minimum temperature over the central and western United States: The role of soil moisture and sea surface temperature
.
J. Climate
,
19
,
1407
1421
.
Allen
,
S. K.
, and
Coauthors
,
2012
: Summary for policymakers. Managing the Risks of Extreme Events and Disasters to Advance Climate Change Adaptation, C. B. Field et al., Eds., Cambridge University Press, 1–19.
Barnett
,
T. P.
, and
R.
Preisendorfer
,
1987
:
Origins and levels of monthly and seasonal forecast skill for United States surface air temperatures determined by canonical correlation analysis
.
Mon. Wea. Rev.
,
115
,
1825
1850
.
Basu
,
R.
,
2009
:
High ambient temperature and mortality: A review of epidemiological studies from 2001 to 2008
.
Environ. Health
,
8
,
40
,
doi:10.1186/1476-069X-8-40
.
Basu
,
R.
, and
J. M.
Samet
,
2002
:
Relation between elevated ambient temperature and mortality: A review of the epidemiological evidence
.
Epidemiol. Rev.
,
24
,
190
202
.
Basu
,
R.
, and
B.
Malig
,
2011
:
High ambient temperature and mortality in California: Exploring the roles of age, disease, and mortality displacement
.
Environ. Res.
,
111
,
1286
1292
.
Basu
,
R.
,
W.-Y.
Feng
, and
B. D.
Ostro
,
2008
:
Characterizing temperature and mortality in nine California counties
.
Epidemiology
,
19
,
138
145
.
Basu
,
R.
,
D.
Pearson
,
B.
Malig
,
R.
Broadwin
, and
S.
Green
,
2012
:
The effect of elevated ambient temperature on emergency room visits
.
Epidemiology
,
23
,
813
820
.
Bergman
,
L.
,
Narrator
,
2001
: “Blackout.” Frontline (television program), originally aired 15 June, PBS.
Bumbaco
,
K. A.
,
K. D.
Dello
and
N. A.
Bond
,
2013
:
History of Pacific Northwest heat waves: Synoptic pattern and trends
.
J. Appl. Meteor. Climatol.
,
52
,
1618
1631
.
Cleveland
,
W. S.
,
1979
:
Robust locally weighted regression and smoothing scatterplots
.
J. Amer. Stat. Assoc.
,
74
,
829
836
.
Ebi
,
K. L.
,
T. J.
Teisberg
,
L. S.
Kalkstein
,
L.
Robinson
, and
R. F.
Weiher
,
2004
:
Heat watch/warning systems save lives: Estimated costs and benefits for Philadelphia 1995–98
.
Bull. Amer. Meteor. Soc.
,
85
,
1067
1073
.
Gershunov
,
A.
, and
D.
Cayan
,
2003
:
Heavy daily precipitation frequency over the contiguous United States: Sources of climatic variability and seasonal predictability
.
J. Climate
,
16
,
2752
2765
.
Gershunov
,
A.
, and
K.
Guirguis
,
2012
:
California heat waves in the present and future
.
Geophys. Res. Lett.
,
39
, L18710,
doi:10.1029/2012GL052979
.
Gershunov
,
A.
,
D. R.
Cayan
, and
S. F.
Iacobellis
,
2009
:
The great 2006 heat wave over California and Nevada: Signal of an increasing trend
.
J. Climate
,
22
,
6181
6203
.
Gershunov
,
A.
,
Z.
Johnston
,
H.
Margolis
, and
K.
Guirguis
,
2011
:
The California heat wave 2006 with impacts on statewide medical emergency: A space–time analysis
.
Geogr. Res. Forum
,
31
,
53
69
.
Hajat
,
S.
, and
Coauthors
,
2006
:
Impact of high temperatures on mortality. Is there an added heat wave effect?
Epidemiology
,
17
,
632
638
.
Hotelling
,
H.
,
1935
:
The most predictable criterion
.
J. Educ. Psychol.
,
26
,
139
142
.
Hotelling
,
H.
,
1936
:
Relations between two sets of variables
.
Biometrica
,
28
,
321
377
.
Kaiser
,
R.
,
A.
Le Tertre
,
J.
Schwartz
,
C. A.
Gotway
,
W. R.
Daley
, and
C. H.
Rubin
,
2007
:
The effect of the 1995 heat wave in Chicago on all-cause and cause-specific mortality
.
Amer. J. Public Health
,
97
,
158
162
.
Knowlton
,
K.
,
M.
Rotkin-Ellman
,
G.
King
,
H. G.
Margolis
,
D.
Smith
,
G.
Solomon
,
R.
Trent
, and
P.
English
,
2009
:
The 2006 California heat wave: Impacts on hospitalizations and emergency department visits
.
Environ. Health Perspect.
,
117
,
61
67
.
Le Tertre
,
A.
, and
Coauthors
,
2006
:
Impact of the 2003 heatwave on all-cause mortality in 9 French cities
.
Epidemiology
,
17
,
75
79
.
Martiello
,
M. A.
, and
M. V.
Giacchi
,
2010
:
Review article: High temperatures and health outcomes: A review of the literature
.
Scand. J. Public Health
,
38
,
826
837
.
Mastrangelo
,
G.
,
U.
Fedeli
,
C.
Visentin
,
G.
Milan
,
E.
Fadda
, and
P.
Spolaore
,
2007
:
Pattern and determinants of hospitalization during heat waves: An ecologic study
.
BMC Public Health
,
7
, 200,
doi:10.1186/1471-2458-7-200
.
Maurer
,
E. P.
,
A. W.
Wood
,
J. C.
Adam
,
D. P.
Lettenmaier
, and
B.
Nijssen
,
2002
:
A long-term hydrologically-based data set of land surface fluxes and states for the conterminous United States
.
J. Climate
,
15
,
3237
3251
.
Meehl
,
G. A.
, and
C.
Tebaldi
,
2004
:
More intense, more frequent, and longer lasting heat waves in the 21st century
.
Science
,
305
,
994
997
.
Mesinger
,
F.
, and
Coauthors
,
2006
:
North American Regional Reanalysis
.
Bull. Amer. Meteor. Soc.
,
87
,
343
360
.
NCDC
,
2009
: Data documentation for data set 3200 (DSI-3200): Surface land daily cooperative summary of the day. National Climatic Data Center Tech. Doc., 19 pp. [Available online at http://www1.ncdc.noaa.gov/pub/data/documentlibrary/tddoc/td3200.pdf.]
Ostro
,
B. D.
,
L. A.
Roth
,
R. S.
Green
, and
R.
Basu
,
2009
:
Estimating the mortality effect the July 2006 California heat wave
.
Environ. Res.
,
109
,
614
619
.
Reid
,
C. E.
,
M. S.
O’Neill
,
C. J.
Gronlund
,
S. J.
Brines
,
D. G.
Brown
,
A. V.
Diez-Roux
, and
J.
Schwartz
,
2009
:
Mapping community determinants of heat vulnerability
.
Environ. Health Perspect.
,
117
,
1730
1736
.
Robine
,
J. M
,
S. L.
Cheung
,
S.
Le Roy
,
H.
Van Oyen
, and
F. R.
Herrmann
,
2007
: Report on excess mortality in Europe during summer 2003. EU Community Action Programme for Public Health Rep., 15 pp. [Available online at http://ec.europa.eu/health/ph_projects/2005/action1/docs/action1_2005_a2_15_en.pdf.]
Robinson
,
P. J.
,
2001
:
On the definition of a heat wave
.
J. Appl. Meteor.
,
40
,
762
775
.
Sailor
,
D.
, and
A.
Pavlova
,
2003
:
Air conditioning market saturation and long-term response of residential cooling demand to climate change
.
Energy
,
28
,
941
951
.
Sheridan
,
S. C.
, and
A. J.
Kalkstein
,
2010
:
Seasonal variability in heat-related mortality across the United States
.
Nat. Hazards
,
55
,
291
305
.
Sheridan
,
S. C.
,
M.
Allen
,
C. C.
Lee
, and
L. S.
Kalkstein
,
2012
:
Future heat vulnerability in California, Part II: Projecting future heat-related mortality
.
Climatic Change
,
115
,
311
326
.
Solomon
,
S.
,
D.
Qin
,
M.
Manning
,
Z.
Chen
,
M.
Marquis
,
K.
Averyt
,
M.
Tignor
, and
H. L.
Miller
Jr.
, Eds.,
2007
: Climate Change 2007: The Physical Science Basis. Cambridge University Press, 996 pp.
Son
,
J.-Y.
,
J.-T.
Lee
,
G. B.
Anderson
, and
M. L.
Bell
,
2012
:
The impact of heat waves on mortality in seven major cities in Korea
.
Environ. Health Perspect.
,
120
,
566
571
.
Sweeney
,
J. L.
,
2002
. The California Electricity Crisis. Hoover Institution Press, 291 pp.
Trent
,
R. B.
,
2007
: Review of July 2006 heat wave related fatalities in California. California Dept. of Health Services Rep., 10 pp. [Available online at http://www.cdph.ca.gov/HealthInfo/injviosaf/Documents/HeatPlanAssessment-EPIC.pdf.]
Vaneckova
,
P.
, and
H.
Bambrick
,
2013
:
Cause-specific hospital admissions on hot days in Sydney, Australia
.
PLoS One
,
8
, e55459,
doi:10.1371/journal.pone.0055459
.
Vaneckova
,
P.
,
M. A.
Hart
,
P. J.
Beggs
, and
R. J.
de Dear
,
2008
:
Synoptic analysis of heat-related mortality in Sydney, Australia, 1993–2001
.
Int. J. Biometeor.
,
52
,
439
451
.
Wilks
,
D. S.
,
2006
: Statistical Methods in the Atmospheric Sciences. Academic Press, 648 pp.