A comparison of weather variables linked to infectious disease patterns using laboratory addresses and patient residence addresses

Weather can contribute significantly to the occurrence of many infectious diseases, and high spatial resolution of exposure data has increasingly been recognized as a requirement in quantifying these relationships [1]. These analyses can help in understanding the likely changes in infectious diseases that may take place in the future through climate or other environmental change [2, 3], particularly in determining the drivers for change [4, 5], in the application of suitable models, and for designing interventions [6]. Variations in diseases by country have been used to examine the impact of weather that is separated from geography [7, 8]. For example, environmental weather parameters at low resolution (e.g. national average) have been used as a proxy for the weather experienced by an individual. The methods of collection for a range of different weather parameters differ, often based on comparatively few weather stations with limited interpolation of the data at other locations or because the exact location of the exposed individual is not known.

To understand the impacts of weather (and the general environment) on both communicable and non-communicable diseases, it is therefore desirable to have information on the weather parameters as close as possible to the location of people’s residence. This is not always possible for a variety of reasons. Datasets from different agencies are seldom linked together, although important efforts to integrate heterogeneous datasets have been made (see the MEDMI project at https://​www.​data-mashup.​org.​uk). With the movement of reported geographic data from the local to the national to the international infectious disease surveillance records [9], the geographic resolution can be ‘lost’ in this process. For example, records from diagnostic microbiology laboratories for England and Wales in the Public Health England Second Generation Surveillance System (SGSS) database have had limited patient postcode data before 2008. Furthermore, the computational resources needed for data linkage at the individual patient postcode level can be significant if large infectious disease datasets are being interrogated. As an example, a study found over a million Campylobacter infections recorded in England and Wales [10]. The linkage with patient postcodes also presents challenges to comply with ethical and legal requirements of confidentiality [11], with insufficient evidence about the efficacy of some anonymisation methods [12].

These concerns are particularly relevant for the increasing number of initiatives aiming to integrate data from different sources (e.g. big data mashups), particularly the linkage of human health data with environmental data needed to study the effects of climate and other environmental change on human diseases [13‐15]. This study sought to test whether the linkage of weather parameters in England and Wales to the diagnostic laboratory postcodes for daily Campylobacter and Cryptosporidium cases across England and Wales reported to Public Health England would be suitable for use as a proxy for the weather at the patient’s residence postcode. The linkage of interpolated local weather parameters supplied by the UK Met Office with the postcodes of the diagnostic laboratories instead of with patient residence postcodes would address many of the concerns above.

There were 441,203 Campylobacter infections, of which 11,381 had missing residence postcode information. Records with missing residence postcodes were removed from the study dataset, leaving 429,822 cases with both residence and laboratory postcodes. For Cryptosporidium infections, there were 30,970 cases with 785 missing residence postcode information, leaving 30,185 cases with both residence and laboratory postcodes.

The time series of average maximum weekly temperature and average weekly rainfall at laboratory postcodes are shown in Fig. 1 (a and b). The weekly number of cases of Campylobacter and Cryptosporidium are shown in Fig. 1 (c and d).

Figure 2 (panels a, b, c and d) shows strong associations between the daily measurements of rainfall (mm) and temperature (°C) obtained at residence postcodes and the daily measurements of rainfall (mm) and temperature (°C) obtained at laboratory postcodes for both pathogens (Campylobacter and Cryptosporidium). Because both temperature variables gave similar results for Campylobacter and Cryptosporidium cases, we plotted the figures for the maximum temperature.

For maximum and minimum daily temperatures at the laboratory postcodes, the estimates of the slope coefficients was 0.99 (p-value ≤0.001) and R-squared was 0.99 for both pathogens. For daily rainfall at the laboratory postcodes, the estimates of the slope coefficients was 0.99 (p-value ≤0.001) and R-squared was 0.96 for Campylobacter and 0.99 (p-value ≤0.001) and R-squared was 0.97 for Cryptosporidium as shown in Fig. 2.

The distribution of the distances between patient residence address and the corresponding laboratory address is shown in Fig. 3. Although the overall (i.e. averaged over the entire dataset) mean patient-laboratory distance is only 11.9 km (standard deviation = 19.8 km), the distribution exhibits a heavy tail with some rare situations where the patient-laboratory distance was larger than 500 km. This was confirmed by a detailed inspection of the data; for example, there were few instances where the diagnostic laboratory was located in South England and the patient residence in Northern Ireland or Scotland.

Figure 4 (a, c and e) shows the distributions of the differences in minimum temperature, maximum temperature and rainfall registered at the laboratory and patient addresses. The distributions exhibited a tail heavier than the corresponding exponential fit, with differences up to ±10 °C for the minimum and maximum temperature and ± 20 mm for rainfall. Exponential fits are commonly used as a benchmark for heavy tailed distributions, although other approaches have been proposed [21].

An important property of heavy tailed distributions is that they result in a higher probability of observing the values with many outliers. To test if the heavy tailed distributions arose from the exceptional instances of large patient-laboratory distances, we compared the distributions based on the entire dataset (red histograms in Fig. 4a, c, e), with the distribution based on the subset of cases when the patient-laboratory distances are lower than 20 km (blue histograms). The comparison confirmed that large discrepancies in the temperatures and rainfall between laboratory and patient residence tended to occur for only the large patient-laboratory distances.

Figure 4 (panels b, d and f) shows the standard deviation in rainfall, minimum and maximum temperatures difference versus the local mean patient-laboratory distances. As expected, when this distance was larger (e.g. because of a large laboratory catchment area), then the difference in the climatic variables between patient residence and laboratory were more scattered, resulting in a larger standard deviation.

Note that the weather stations used to link both laboratory and residence postcodes in the analysis were likely to be the same, but the interpolations used different weightings according to location and thus predominantly gave different results.

There is a growing need for a more ecological view of public health [22], and the associations linking data on weather with disease occurrence can be extremely useful in demonstrating the public health burden associated with climate change [23]. It is becoming evident that with many infectious diseases, the impacts of weather on seasonality can be teased out through the linkage of local weather parameters to cases [24]. There is a good rationale for developing methods to easily link disease data to weather and other environmental drivers, and to provide the tools to achieve this. For infectious diseases, linkage can be examined across pathogens; and this is being developed through the extraction of long-term data on individual local weather parameters that can easily be linked through the postcode of the laboratory where the case was diagnosed. In this study, the differences between the weather variables associated with the patient residence postcode and the laboratory postcode were comparatively small and were mostly as a result of a small number of cases whose domestic residence was at a great distance from the laboratory.

The use of the Inverse-Distance Weighted (IDW) interpolation tended to smooth the values spatially leading to a higher correlation between laboratory and residence locations (especially in urban areas where the two locations are likely to be close); more sophisticated methods would require altitude data and additional geo-processing. Gridded datasets that use the IDW method with statistical adjustments for known atmospheric effects have been used by some researchers [25, 26], but unadjusted IDW interpolation provides a more spatially localised estimate than taking a single value for a region or a country (e.g. the Central England Temperature) used by [27] or [28]), or simply by taking the measurements from nearby weather stations [29].

The analysis is straightforward and applicable to situations resembling the same circumstances in England and Wales, i.e. regions with same climatic zones such as regions in the same Koppen-Geiger classification [30], and with similar public health capacity and spatial density of diagnostic laboratories. The study can be relevant to other situations; for example, the findings are expected to be valid in countries in climatic zones exhibiting smaller spatial variance in temperature and rainfall and/or shorter patient-laboratory distances. However, it is possible that the relationship we found between weather parameters at the laboratory and residence postcodes may not transfer to other regions and countries where the health service administrative boundaries are on different scales and data distribution is different. In these cases, an assessment like the present study would be needed.

The analysis was done only for rainfall and temperature; however, some diseases may be affected by other environmental variables (not only climatic such as humidity and barometric pressure but also soil type, density of livestock etc.). Our analysis was based on outdoor weather variables and was not able to assess the seasonal indoor environmental variables that most people are exposed to. The underlying assumption was that most people spend longer time at home, however, it is estimated that people spend up to 40% of time at locations other than their residence address (schools, workplace, transport).

Using the weather parameters at laboratory postcode as a proxy for the weather parameters at the patient residence can be a valid approach in England and Wales and useful in exploring the relationships between weather and infectious diseases over time. However, large discrepancies between the weather parameters at the two different locations can occur in some cases. These typically arise in instances of large distances between patient residence address and the diagnostic laboratory, but also reflect the intrinsic spatial variability of both temperature and rainfall (e.g. localized episodes of rainfall). Data linkage at laboratory postcode level also ensures data confidentiality.