BACKGROUND AND RELATED LITERATURE

In the middle of the nineteenth century, when our study begins, Londoners experienced what are by modern standards extremely high mortality rates, comparable to the levels faced by the very poorest urban residents in developing countries today.^{Footnote 5} Much of this mortality was concentrated in infants and young children, and it is useful to focus on this group for comparison purposes because this has been the main focus of existing research and, as we will see, this group plays a central role in the relationship between temperature, mortality, and the disease environment. In 1860-1899, at the beginning of our study period, out of every thousand children born in London, 158 died before age one in an average year (Williams and Mooney, 1994). While high by modern standards, this put London just above the national average (149 per thousand) and below more industrialized cities such as Birmingham and Manchester (Williams and Mooney, 1994).^{Footnote 6} Starting from this high level, infant mortality dropped to 132 per thousand births in 1900-1910, nearly halved to 66 per thousand in the 1920s, and fell to 24 per thousand in the 1950s, at which point infant mortality accounted for just 2.5 percent of all deaths.^{Footnote 7} In this pattern, London was relatively similar to the average across British cities, though there has been some dispute about the similarity in the exact timing of the onset of the infant mortality decline.^{Footnote 8}

Our data show that, in the 1870s, most deaths were due to infectious diseases. The largest single cause of all-age mortality was tuberculosis (14 percent of all deaths), followed by bronchitis (13 percent). Cardiovascular diseases accounted for only about 6 percent of deaths, followed by pneumonia (5.4 percent) and whooping cough (3.6 percent). Cancer accounted for less than 3 percent of deaths. Among infants, the most important category was digestive diseases, which accounted for 13 percent of all deaths from 1876 to 1885. By the 1950s, infectious diseases had become much less important, and overall mortality, which was concentrated among older adults, was driven by cardiovascular factors (28 percent of deaths), cancer (22 percent), and stroke (9 percent). Even for infants, digestive diseases were much less important, accounting for just under 3 percent of infant deaths (65 percent of infant deaths were attributable to either prematurity, birth injuries, or congenital defects).

A substantial amount of scholarly work has been focused on understanding the mortality transition in London, its causes, and how it compares with other parts of the country. Of this extensive literature, the most relevant for our study is work looking at the seasonality of deaths. A remarkable early paper in this area is Buchan and Mitchell (1875), written for the Scottish Meteorological Society, which calculated the average death contribution by each week of the year from 1845 to 1874 using the weekly mortality statistics produced by the Registrar General’s office. These weekly data seem to have been largely overlooked by studies of historical mortality patterns in London until a recent paper by Carson et al. (2006), which examines mortality patterns in the twentieth century. Relative to their paper, our study includes several decades of data from the nineteenth century. This difference is important, because it means that our analysis begins before the sharp decline in infant mortality in London began (see Online Appendix Figure A.3). We also offer a richer analysis framework that captures the importance of lagged mortality effects and allows us to calculate counterfactual mortality patterns.

Buchan and Mitchell were not the only contemporary observers to recognize the connection between season and mortality in British cities in the nineteenth and early twentieth centuries. Arthur Newsholme discussed the seasonality of infant diarrheal deaths extensively in his Presidential Address to the Incorporated Society of Medical Officers of Health in 1899.^{Footnote 9} He continued to write on the subject when he joined the Medical Office of the Local Government Board (e.g., Newsholme Reference Newsholme1901). E.W. Hope, the Medical Office of Health of Liverpool, undertook similar studies (Hope 1899a, 1899b), as did George Newman, the Medical Office for the district of Finsbury (Newman Reference Newman1906). These and other contemporary observers, many of whom were local Medical Officers, were acutely concerned about the causes of the summer diarrhea patterns that they observed. They proposed a number of possible causes, including poor sanitation (particularly sewage removal), overcrowded living conditions, insufficient breastfeeding leading infants to be fed with contaminated milk or other food, and generally poor childcare among the working classes, often because the mother went back to work soon after giving birth. Contemporary observers generally agreed that all of these factors likely mattered, though opinions differed on whether the bulk of the blame should be placed on mothers, for a failure to adequately breastfeed or care for children, or local authorities, for their inability to provide adequate sanitation and healthy living conditions.^{Footnote 10}

In the intervening years, a number of other authors have studied the seasonality of mortality or how mortality was affected by temperature variation using either annual or quarterly data. Woods, Watterson, and Woodward (1988, p. 360), for example, run regressions comparing mean temperature to infant mortality in England and Wales for 1870-1911 (N = 42) and find evidence that temperature in the third quarter of the year is positively related to mortality. After documenting substantially higher mortality in cities than in rural areas, they speculate that “the most likely reason for this ‘urban effect’ is that climatic conditions, especially during the third quarter of the year, interacted with poor urban sanitary environments which resulted in high levels of diarrhea and dysentery among infants…” The same quarterly data have been used to study the seasonality of infant mortality by a number of other authors, including Mooney (1994) and Williams and Galley (1995). Wrigley and Schofield (Reference Wrigley and Schofield1989, ch. 9) also examines the relationship between their mortality data, based on Anglican parish registers, and monthly temperature. One notable feature of most of the recent work on the seasonality of mortality is that almost no attention is paid to the impact on older age groups or the effect of unusually cold weather, two topics that we tackle in this paper.

One interesting discovery in existing work is that the seasonal nature of mortality in London appears to have been absent before the second quarter of the nineteenth century. While data are scarce before the beginning of general death registration in 1837, findings from the Bills of Mortality compiled by Landers and Mouzas (1988) show no evidence of a summer mortality peak in the late eighteenth or early nineteenth century (though the summer peak does appear in the early eighteenth century). A similar pattern is identified by Huck (1997) for a sample of industrial parishes in Northern England. Thus, it appears that the summer peak in infant mortality that plays such an important role in our results only emerged in British cities sometime in the first half of the nineteenth century.

Despite extensive research into the mortality decline that took place in British cities in the decades around 1900, no clear consensus has emerged as to the causes. It is clear that this decline was concentrated among infants and driven by the reduction in diseases of the digestive system, but why these diseases declined remains debated. McKeown (Reference McKeown1976) emphasized the importance of improving overall nutrition among the British population. McKeown’s hypothesis finds some support in the work of Millward and Bell (1998), which also emphasizes the importance of improvements in the housing stock.^{Footnote 11} However, the importance of nutrition has been challenged by Szreter (Reference Szreter1988), among others, who instead emphasized the role of public health measures.^{Footnote 12} Recent evidence from Chapman (2019) and Harris and Hinde (2019) provides support for the role of public health infrastructure. Chapman, for example, finds that between 30 and 60 percent of the mortality decline that took place in British cities from 1861 to 1900 can be explained by infrastructure investments.^{Footnote 13} However, the case for sanitary infrastructure investments playing a key role in reducing infant mortality in London is hard to make. Most of the major improvements, including water filtration and other water quality improvements, continuous water supply, and sewer access, generally took place years or decades before the decline in infant digestive disease mortality that appears in our data (e.g., Online Appendix Figure A.3).^{Footnote 14}

One aspect of nutrition that has received particular attention, both by contemporary observers (e.g., Peters 1910) and in later research (e.g., Beaver 1973), is the availability of uncontaminated milk supplies. The importance of milk has found support in a recent study of the United States (Komisarow 2017) but not in others (Reference Anderson, Charles and ReesAnderson, Charles, and Rees forthcoming), and the contribution of improvements in milk quality to the decline in British mortality in the decades around 1900 remains disputed (Atkins 1992). A closely connected factor is the level of maternal care, which was thought by some contemporaries to have a large influence on infant mortality during our study period (Newsholme 1899; Newman Reference Newman1906), particularly among the many poorer mothers who returned to work soon after giving birth. In response, by the early twentieth century, many local areas employed Health Visitors who helped educate mothers on the feeding and care of children (Fildes 1992; Mooney 1994). Evaluating the efficacy of this and other similar interventions remains challenging.

In summary, a variety of factors, including improved sanitation, water and milk quality, better general nutrition (especially for nursing mothers), improved maternal care and breastfeeding practices, and better housing conditions, have been put forward as potential explanations for the decline in mortality in British cities in the decades around 1900. Many of these suggestions have found some empirical support, though almost all have also had their critics. It is not our purpose in this paper to attempt to settle this debate. What is clear, however, is that there was a general decline in mortality in London and other British cities commencing in the years around 1900, that this decline was concentrated among infants and young children, and that it was principally due to a decrease in digestive diseases. Taking these facts as given, in the remainder of this paper, we consider the role of weather in contributing to mortality patterns in London and how the relationship was altered as the underlying disease environment changed.

Finally, this study connects to a literature in epidemiology using time- series methods to assess the impact of temperature on mortality. The large literature in this area using modern data includes papers such as Hajat et al. (2002), which studies the mortality effects of heat using daily data for London from 1976 to 1996. There is also a smaller historical literature in this area, which includes, in addition to the paper by Carson et al. (2006) on London cited above, papers by Petkova, Gasparrini, and Kinney (2014) looking at New York City, Äström et al. (2013) looking at Stockholm, and Ekamper et al. (2009) using data from the Dutch province of Zeeland. Of these, only the Ekamper et al. paper uses data stretching into the nineteenth century, though their data come from a mainly rural/ agricultural setting. Our analysis approach differs substantially from the one used in these papers. In particular, we do more to assess the lag structure of temperature effects, which are ignored in many existing studies. We also provide counterfactual estimates that allow us to quantify the impact of temperature and how this impact changed over time.

DATA

The weekly mortality data used in this study were digitized from printed reports produced by the Registrar General’s office collected mainly from the London School of Economics Library and the New York Public Library. The collection of birth, marriage, and mortality data by the Registrar’s Office, which commenced in the 1830s and was well established by the 1870s, was an enormous undertaking. The system was registering as many as a million events per year in the 1840s and 1.7 million by the early twentieth century, which “considering the detail of reporting and the standard of accuracy obtained…was a very considerable achievement’’ (Woods Reference Woods2000, p. 36). To undertake this task, a new office was established within the civil service, headed by the Registrar General and employing a professional staff (Woods Reference Woods2000, pp. 31-32). These oversaw a network of local registrars and superintendent registrars drawn from among local officials or leading citizens (Woods Reference Woods2000, p. 31). Within this system, it was to the collection of mortality statistics “that most attention was given and from which the greatest detail was sought,’’ making them the “shining star of Victorian civil registration.”

The 300,000 mortality observations in our main data (Reference Hanlon, Hansen and KantorHanlon, Hansen, and Kantor 2020) cover 4,540 weeks stretching from 1866 to 1965, with breaks in 1915-1918 (WWI) and 1940-1948 (WWII). We end our main study period in 1965 because the geographic area for which our data are reported changed at that point, though in Online Appendix A.2, we present supplementary results for 1981-2006.^{Footnote 15} One benefit of ending our analysis in 1965 is that we avoid the impact of air conditioning, which began to be used in some locations in London around that time.

In each week, we observe deaths broken down by cause and age group for the city as a whole, so our analysis data set is structured as a single time series. Our mortality data come from a consistent geographic area that encompasses all of modern Central London. This area is somewhat smaller than the current Greater London administrative area, which was established in 1966, but much larger than the City of London. Administrative changes in 1966 resulted in a change in the geographic area covered by the weekly data series to the modern Greater London. This change motivates our decision to end the study period in 1965.

Our analysis focuses mainly on either all-age mortality, which is available for the full study period, or infant mortality, which is available starting in 1874. The availability of weekly mortality data with complete breakdowns by age and cause of death is unique to London. While some weekly mortality figures are reported for other cities, or for neighborhoods within London, only for London as a whole do we observe the fully detailed breakdown of age group by cause of death, so this is our focus.

The cause-of-death categories require some standardization. In our data set, causes of death are reported in as many as 130 different categories in some years and as few as 57 in others, with substantial changes in the reported categories over time. To deal with these changes and generate series that are reasonably consistent across the study period, we collapse these causes of death into more aggregated categories such as digestive diseases (including diarrhea, dysentery, cholera, typhoid, and so on) or respiratory diseases (bronchitis, asthma, and so on).^{Footnote 16} While historical cause-of-death data must be treated with some caution, within these broadly defined categories, it is likely that most diseases are correctly categorized, particularly deaths due to the most common causes. Importantly, digestive diseases, the most important category for this study, typically show clear defining features, so the classification of these diseases should be reasonably accurate, even early in our study period. The causes of death categories change substantially after WWII. For digestive diseases, which includes diarrhea, dysentery, enteritis, gastritis, typhoid, and cholera, we are able to generate a consistent series up to 1965, but when we look at other causes of death, we focus only on data before WWI.

The existing literature provides a vigorous discussion of potential issues in the use of the Registrar General’s mortality data (e.g., Luckin 1980; Hardy 1994). One common concern is the completeness of the records. However, for London in the period we study, this is unlikely to be a major concern; Hardy (1994) reports that registration rates reached 98 percent in London by 1870. Another issue is the classification of causes of death, which naturally evolved over time as medical knowledge advanced.^{Footnote 17} Certain disease categories, such as tuberculosis, which were difficult to diagnose, have been highlighted as particularly problematic (McKeown and Record 1962; Hardy 1994). However, the most important category for our analysis, diarrheal diseases, is “one of the least troublesome’’ according to Hardy (1994, p. 486) because the fact that it represented a symptom, rather than a disease, made it relatively easy to recognize. This is particularly true in our main analysis, since we have combined diarrheal diseases with those other categories (cholera, enteric fever) that were most likely to be confused with one another.

Our temperature data come from the Radcliffe Observatory in Oxford, 80 km outside of central London. We have two primary reasons for preferring the temperature data series from Oxford over other alternatives. First, they come from a single location, while the weather data for London reported in the Registrar General’s report come partially from Greenwich Observatory and partly (after WWII) from Kew Gardens. Second, the Oxford data provide both average maximum and average minimum temperatures for each week, rather than just weekly mean temperatures. This additional detail is useful for focusing specifically on heat- and cold- related deaths. Another advantage of using data from Oxford is that it is far enough outside of London that the temperature there is unlikely to be substantially influenced by urban heat-island effects.^{Footnote 18} Dealing with the potential endogeneity created by urban heat-island effects is important if we want to identify the causal impact of temperature on mortality. Finally, we use some additional weather data tracking precipitation and humidity from the Greenwich Observatory and Kew Gardens.^{Footnote 19}

In our main analysis, we assess the impact of temperature non-parametrically by dividing weeks into temperature bins. Given London’s relatively mild climate, we examine heat effects by looking at bins with maximum temperatures (in Fahrenheit) from 65 to 70, 70 to 75, 75 to 80, and above 80. This division provides sufficient observations in each bin to estimate effects. For cold effects, we focus on bins in which minimum temperatures were between 30 and 35, 25 and 30, or below 25. Online Appendix Figure A.1 shows the distributions of the minimum and maximum weekly temperatures during our sample period.

Table 1 provides statistics on the temperature bins included in our analysis for the full sample as well as several sub-periods. The subperiods that we consider are naturally defined by the breaks in the data. The first covers all of the years before the onset of WWI in 1914. The second covers the interwar period, with data from 1918 to 1939, while the third period covers the years just after WWII, from 1949 to 1965. In the analysis, we pool the data after WWI to increase the sample size and because we observe similar patterns in both the interwar and post-WWII periods.

Table 1

TEMPERATURE BINS USED IN THE ANALYSIS

Source: Authors’ calculations using temperature data from the Radcliffe Observatory, Oxford, UK.

EMPIRICAL SPECIFICATION

Our main analysis uses a lead-lag model in which we estimate the non-parametric relationship between the occurrence of unusually high or low temperatures in a week on mortality in that week as well as several previous and subsequent weeks. Estimates of the mortality response in weeks before a high- or low-temperature event is observed provide a check on our identification strategy; we should expect these estimates to be very close to zero, since temperature today should not impact mortality in the past (controlling for past temperature). Our estimates of the effect of high- or low-temperature events in subsequent weeks allow us to understand whether relatively extreme temperature events have lagged mortality effects. To estimate the effect of temperature non-parametrically, we classify weeks into the temperature bins shown in Table 1, with weeks where the high temperature was never above 65°F and the low temperature never below 35°F treated as the reference category.

Our baseline empirical specification is

$$\[\ln ({y_{wt}}) = \sum\nolimits_{j = - m}^k {\sum\nolimits_{q = 1,q \ne 4}^8 {\alpha _j^qTEMP_{wt}^q[w = j] + {\delta _w} + {\delta _t} + {X_{wt}}\beta + {\varepsilon _{wt}},} } \]$$

where ln y_wt is the log number of total or infant deaths (or such deaths due only to certain causes) in week w of year t in London. The $TEMP_{wt}^q$ term is a set of indicator variables, taking on the value one if the weekly temperature is in the qth temperature bin, with q = 4 corresponding to the reference bin. The estimated $\alpha _{j}^q$ quantifies the non-parametric relation-ship between temperature and mortality for each m (week) “leads” and each k (week) “lags.” Note that the leading or lagged effects are estimated while also estimating the direct effect of contemporaneous temperature. Our specification includes controls for week-of-the-year fixed effects (δ_w), year fixed effects (δ_t), and a vector of week-by-year varying weather controls (X_wt) (precipitation and an indicator for weeks with heavy fog).^{Footnote 20}

To be clear, y_wt is the number of deaths, not a death rate. However, because we are using a log specification with year fixed effects, the coefficients we estimate are identical to what would be obtained if we instead replaced y_wt with a death rate calculated using population measured annually.

Because our data are structured as a time series, serial correlation is a potential concern. However, we find that allowing for serial correlation using Newey-West standard errors results in smaller confidence intervals, indicating negative serial correlation. This most likely reflects that, when there are many deaths in one week, there are fewer people at risk of dying in the next week. Thus, in the main results, we present more conservative robust standard errors, though results obtained using Newey-West standard errors are provided in the Online Appendix.

Our analysis approach, which relies on high-frequency variation in time-series data, is similar to the approach commonly used in the public health literature (reviewed by Deschênes (2014)) but differs from most existing studies in the economics literature, which instead uses lower- frequency panel data, often at the monthly level. This difference comes with both advantages, such as an ability to examine the structure of lagged effects in more detail, and disadvantages, such as limitations on the variety of climate conditions observed. Given this, we view our approach as complementary to the panel data approach used in studies such as Barreca et al. (2016).

One final point to note regarding our estimation strategy is how our ability to control flexibly for time effects compares to studies using lower- frequency panel data such as Barreca et al. (2016) and Geruso and Spears (2018). Typically, panel studies have an advantage in that it is possible to more flexibly control for time-varying factors by including time-period fixed effects. In studies using panel data at monthly frequency, such as the two cited above, this takes the form of month-by-year effects. However, note that our data are sufficiently rich that we are also able to estimate results while including month-by-year effects, just as in those studies. These results, which can be found in Online Appendix Figures A.8 and A.15, show that the inclusion of these flexible time controls has very little impact on our results. This highlights the fact that we are able to control for time effects as flexibly as any existing panel data study reliant on monthly data.

COUNTERFACTUALS

The results presented above have implications for the impact of rising temperatures on mortality in London. In Table 3, we conduct some simple counterfactual exercises in order to assess the magnitude of these effects. Panel A of Table 3 describes our estimates of the actual number of excess deaths due to relatively hot and relatively cold weeks in different periods. These are based on the estimated coefficients obtained from applying our regression approach to the data from each period. Heat effects reflect the number of deaths associated with maximum temperatures above 80°F, while cold effects are those associated with minimum temperatures below 35°F, the bins for which we observe statistically significant effects before WWI. Panel A shows that we estimate that warm weeks, those falling into our top temperature bin, are associated with 59,773 excess deaths in 1876-1914, or 1.9 percent of all deaths. However, after WWI, we observe few deaths in weeks when temperatures are in the top bin, and by the post-WWII period, weeks in the top bin are, on average, relatively healthy. This shift reflects the patterns described in the previous section. Unlike warm weeks, cold weeks remain associated with substantial numbers of excess deaths throughout the study period: equal to 9.2 percent of total deaths before WWI, 13.5 percent in the interwar period, and 8.4 percent after WWI. Overall, temperature-related deaths were a major component of mortality, accounting for 11.2 percent of all deaths in the pre-WWI period, 13.7 percent in the interwar period (when influenza deaths spiked), and 7.7 percent after WWII.

Table 3

ESTIMATED AND COUNTERFACTUAL EFFECTS OF TEMPERATURE ON MORTALITY BY PERIOD

Source: Authors' estimates using Registrar General's data.

Next, we ask: What would mortality have looked like if the temperature-mortality relationship had not improved after WWI? To answer this, we calculate mortality in the interwar and post-WWII periods while imposing, for each age group, the temperature-mortality relationship estimated on pre-WWI data. We do this by age group in order to account for changes in the age composition of the population. These estimates are shown in Panel B of Table 3. It is important to note that this delivers conservative counterfactual estimates, since our estimates incorporate the baseline reduction in overall mortality observed during each period. That is, we are not holding overall mortality rates at the pre-WWI period. Instead, we are simply applying the estimated percentage increase in mortality associated with warm weeks, relative to weeks with moderate temperature, from the pre-WWI period to the baseline mortality rates observed in the interwar and post-WWII periods. Since we are allowing baseline mortality rates to change, this counterfactual incorporates broad health improvements that occurred across these periods.

The estimates in Panel B suggest that, had the pre-WWI temperature-mortality relationship persisted into the later periods, there would have been an additional 10,098 heat-related deaths in the interwar period and 9,451 deaths in the decades after WWII, equal to a 0.9-1.4 percent increase in overall mortality. This provides a direct estimate of how many heat-related deaths were averted as a result of the changes in the temperature-mortality relationship that took place in the early twentieth century.

Finally, we provide an assessment of the impact that rising temperature would have had under the different mortality regimes we have observed.

In Panel C of Table 3, we present counterfactual mortality given the mortality patterns observed in each period but imposing an increase in average temperatures of 1.5°C, a common reference level used in the climate change literature.^{Footnote 26} These results show that had this temperature increase occurred in the pre-WWI period, we would have expected an additional 33,011 heat-related deaths, equal to a 1 percent increase in total mortality. This would have been more than offset by a reduction of 86,317 cold-related deaths. However, in the later periods, the increase in heat-related deaths resulting from this temperature increase was quite small, or even negative. This provides a direct illustration of how much the impact of a rise in temperature can vary across environments with different underlying disease burdens.

Another benefit of being able to generate results using counterfactual temperature patterns is that we can examine the impact of temperature on the timing of the mortality transition. A number of papers, such as Cutler and Miller (2005) and Anderson, Charles, and Rees (forthcoming), compare the timing of public health interventions to annual mortality data to see if they correspond to the timing of infant mortality declines. In the context of London, however, it has been argued that the timing of the decline in infant mortality may have been substantially affected by several hot summers that occurred in the 1890s (Woods, Watterson, and Woodward 1988; Woods Reference Woods2000, p. 296). Our data allow us to assess this argument.

To do so, we use our estimates of the impact of temperature on mortality to generate two counterfactual infant mortality series. In the first, we completely remove the (contemporaneous and lagged) impact of weeks with temperatures falling into the top bin. In the second, we replace the pattern of high-temperature weeks in all years with that observed in an average high-temperature year.

These counterfactual s are compared to the actual pattern of infant deaths in Figure 7. The main takeaway from this figure is that, under either counterfactual scenario, the decline in infant mortality in London begins several years earlier than the true infant mortality decline. Whereas actual infant mortality in London peaks in 1899, under the counterfac- tuals, the decline begins after 1895. The difference is due to the unusually high number of warm weeks that occurred in the late 1890s. This result confirms the suspicions of demographers such as Woods (Reference Woods2000). It also tells us that we have to be very careful to control for the impact of temperature in studies that compare the timing of an intervention, such as water filtration or chlorination, to annual mortality patterns, since temperature variation can alter the timing of mortality declines by several years.

Figure 7

ACTUAL AND COUNTERFACTUAL INFANT DEATHS

Sources: Authors' calculations based on Registrar General's data. See text for further details.

Figure 7 also provides a sense of the extent to which temperature variation influenced mortality in different years. Specifically, the vertical distance between the “Actual deaths” line and the “Without heat deaths (q8)” line describes the number of deaths in a year due just to the effect of temperatures falling into the top temperature bin. It is clear that this varies substantially over time, with hot weeks making a particularly large contribution to total deaths in the late 1890s.