Spatiotemporal Error in Rainfall Data: Consequences for Epidemiologic Analysis of Waterborne Diseases

Abstract

The relationship between rainfall, especially extreme rainfall, and increases in waterborne infectious diseases is widely reported in the literature. Most of this research, however, has not formally considered the impact of exposure measurement error contributed by the limited spatiotemporal fidelity of precipitation data. Here, we evaluate bias in effect estimates associated with exposure misclassification due to precipitation data fidelity, using extreme rainfall as an example. We accomplished this via a simulation study, followed by analysis of extreme rainfall and incident diarrheal disease in an epidemiologic study in Ecuador. We found that the limited fidelity typical of spatiotemporal rainfall data sets biases effect estimates towards the null. Use of spatial interpolations of rain-gauge data or satellite data biased estimated health effects due to extreme rainfall (occurrence) and wet conditions (accumulated totals) downwards by 35%-45%. Similar biases were evident in the Ecuadorian case study analysis, where spatial incompatibility between exposed populations and rain gauges resulted in the association between extreme rainfall and diarrheal disease incidence being approximately halved. These findings suggest that investigators should pay greater attention to limitations in using spatially heterogeneous environmental data sets to assign exposures in epidemiologic research.

Keywords: bias; environmental epidemiology; exposure misclassification; extreme weather; measurement error; precipitation; waterborne diseases.

Figure 1.

Regions and rainfall distributions for a simulation study. Mean annual precipitation from 1985–2013 in each of the 2 simulation study regions: the Pacific Northwest (NW, top left) and the Atlantic Gulf Coast (GC, bottom right) of the United States. Data represent daily precipitation from the Parameter-elevation Relationships on Independent Slopes Model (PRISM) (33, 34).

Figure 2.

Rainfall data and sources in the Pacific Northwest (NW) region of the United States, 1985–2013. Rainfall data sources are reference (A), satellite (B), and interpolated in situ from 25 (C) and 100 (D) rain gauges. Reference precipitation is from the Parameter-elevation Relationships on Independent Slopes Model (PRISM) (4-km resolution) (33, 34); satellite precipitation is from the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks–Climate Data Record (PERSIANN-CDR) (28-km resolution) (35); and “rain gauges” used in interpolations are samples from PRISM, which are interpolated (shown at a 4-km resolution) using ordinary kriging.

Figure 3.

Rainfall data and sources in the Atlantic Gulf Coast (GC) region of the United States, 1985–2013. Rainfall data sources are reference (A), satellite (B), and interpolated in situ from 25 (C) and 100 (D) rain gauges. Reference precipitation is from the Parameter-elevation Relationships on Independent Slopes Model (PRISM) (4-km resolution) (33, 34); satellite precipitation is from the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks–Climate Data Record (PERSIANN-CDR) (28-km resolution) (35); and “rain gauges” used in interpolations are samples from PRISM, which are interpolated (shown at a 4-km resolution) using ordinary kriging.

Figure 4.

Relative location of rain gauges and villages in coastal Esmeraldas Province, Ecuador, 2004–2007. The map shows the location of the study region, within which there are 19 villages (points) and both original (plusses, from a previous study (2)) and new, supplementary rain-gauge data (crosses) from the same period. Shading indicates the distance of each village to the closest rain gauge; gray lines are the river network along which villages are situated; and the black outline is the coastal-draining river basin within which villages are located.

Figure 5.

Bias in the estimated incidence rate ratio (IRR) measuring the effect of extreme rainfall and wet conditions on disease incidence in the simulation study based on data from the Pacific Northwest (NW) and the Atlantic Gulf Coast (GC) of the United States, 1985–2013. Bias in the extreme rainfall exposure (A) and wet conditions exposure (B) are shown with respect to region, indicated by color, and data type, where solid lines and dotted lines are results from interpolated and satellite data, respectively. The number of rain gauges (horizonal axis) does not pertain to satellite data. Bias is quantified as the percent of the true value of the IRR ($\exp (\mathrm{\beta })=2.72$); negative bias indicates bias towards the null. Results are the mean of 1,000 simulation iterations; line thickness includes simulation uncertainty error.

Figure 6.

Bias in the estimated incidence rate ratio measuring the effect of extreme rainfall and wet conditions on disease incidence as a function of distance between an individual community and rain gauge in the simulation study based on data from the Pacific Northwest (NW) and the Atlantic Gulf Coast (GC) of the United States, 1985–2013. Bias in the extreme rainfall exposure (A) and wet conditions exposure (B) are shown with respect to region, indicated by color. Bias is quantified as the percent of the true value of the incidence rate ratio ($\exp (\mathrm{\beta })=2.72$); negative bias indicates bias towards the null. Results are the mean of 5,000 simulation iterations; line thickness includes simulation uncertainty error.

Figure 7.

Changes in the incidence rate ratio (IRR) measuring the association between extreme rainfall and disease incidence according to village and rain-gauge proximity, using data from previous studies in Ecuador (, –38). A) Points and lines are IRR estimates made using data from groups of villages that include those closest to (left) and those farthest from (right) a rain gauge; the horizontal axis is the maximum distance from any village in the group to a rain gauge. The points indicate values (IRR, maximum distance) for groups of villages (increasing in count from 2 to 19) that were used for regression model fitting. B) The same black line as in panel A, plotted along the number of villages (instead of distances) and alongside the mean IRR from 1,000 random sequencing of same-sized groups of villages (gray line). The shaded regions are the 95% confidence intervals.