Winter 2024: Cyclone Yaku and Dengue

Mean temperature and total precipitation reported hourly in the ECMWF ERA5-Land Hourly reanalysis dataset were extracted using Google Earth Engine. Given evidence of temperature biases in this dataset relative to Global Historical Climatology Network (GHCN) weather stations, particularly at high elevations, we debiased hourly temperature by comparison to monthly average temperatures reported by the high resolution climatology WorldClim between 1970 and 2000 following (Childs et al., n.d.), using the equation:

\(\widehat{ERA5_{ihmy}} = ERA5_{ihmy} - \overline{ERA5_{im}}+\overline{WorldClim_{im}}\)

where \(\widehat{ERA5_{ihmy}}\) is the debiased ERA5 hourly temperature in a given province (where the subscripts i, h, m, and y indicate the mean temperature in a given province, hour, month, and year respectively), \(ERA5_{ihmy}\) is the raw ERA5 hourly temperature in the corresponding province, \(\overline{ERA5_{im}}\) and \(\overline{WorldClim_{im}}\) are the mean monthly temperatures for a given province from 1970 - 2000 from ERA5 and WorldClim respectively. We then calculated the average temperature and precipitation across a given spatial region by taking a population-weighted average (e.g., weighting by the proportion of the total population living in a given 100m x 100m grid cell). Population estimates are based on WorldPop residential population estimates for 2020.

Ecuador and Peru have three broad climatic regions from west to east: the dry Pacific coastline, the cool Andes mountains, and the warm and wet Amazon rainforest (Figure1A). Tropical Cyclone Yaku, which was first formed on March 7th 2023 and dissipated on March 20th, primarily impacted the northern coast of Peru along with parts of Ecuador (Figure1B). Although some northeastern regions in the Amazon basin also experienced anomalously heavy precipitation during Cyclone Yaku, we focus our analysis on those in the first percentile for precipitation (i.e., particularly dry provinces that experience less than 5.4 mm of rain per day on average) that experienced anomalous precipitation during Cyclone Yaku (exceed the 95th percentile of their distribution for biweekly cumulative precipitation across thirty years) (Figure1C). Five provinces meet these criteria: three northwestern coastal regions in Peru (Lambayeque [PE14], Piura [PE20], and Tumbes [PE24]) and two coastal provinces in Ecuador (Manabí [EC13] and Santa Elena [EC24]). These selection criteria for the treated provinces reflect the hypothesized mechanism whereby regions with a dry climate that experience heavy precipitation may be particularly vulnerable to dengue outbreaks, as water storage practices and lack of infrastructure to prevent flooding may promote formation of new vector breeding habitat.

Figure 1: Climate conditions across regions in Peru and provinces in Ecuador from 1993 - 2023 compared to during Cyclone Yaku. In all panels, the bolded outline encompasses locations included in treated group. (A) A bivariate plot where locations are shaded according to mean temperature and mean precipitation compared to other regions from 1993 - 2023. Redder tones indicate warmer locations and bluer tones indicate locations with more precipitation. (B) Locations are shaded according to the total amount of precipitation received during Cyclone Yaku, from March 7th - March 20th, 2023. (C) Locations are shaded according to the percentile for biweekly cumulative precipitation experienced during Cyclone Yaku, compared to the distribution of biweekly cumulative precipitation from 1993 - 2023. Darker shades indicate locations that experienced more precipitation than usual during Cyclone Yaku.

Dengue cases are predominantly reported on the Pacific coast and in the Amazon basin, with many provinces in the Andes highlands not reporting any cases from 2010 onward. Several of the provinces most affected by Cyclone Yaku in Peru have a relatively high mean weekly dengue incidence compared to other provinces (Figure 2A). During Cyclone Yaku, these provinces had a mean incidence of XX compared to an average incidence of XX across other dengue-endemic provinces, but we do not observe elevated incidence in treated provinces in Ecuador (Figure 2B). As shown in Figure 2C, many of the provinces with the most anomalously high precipitation also had the greatest dengue incidence in the weeks following the outbreak.

Figure 2: weekly dengue incidence across provinces in Peru from 2013 - 2022 compared to the 2023 outbreak. In all panels, the bolded outline encompasses the group of northwestern provinces most affected by Cyclone Yaku, used as the treatment group in the synthetic control analysis. Provinces in white had no reported dengue from 2013 - 2023. (A) Provinces are shaded according to their mean weekly dengue incidence across the time series prior to Cyclone Yaku (2013-2022). (B) Provinces are shaded according to their mean weekly dengue incidence in the weeks following Cyclone Yaku (week 10 - 26 of 2023). (C) A bivariate plot where provinces are shaded according to their precipitation anomaly during Cyclone Yaku and dengue incidence in the weeks following the cyclone. More pink tones indicate provinces with greater dengue incidence and more blue tones indicate provinces with a greater anomaly in precipitation (e.g., high precipitation compared to their typical weather).

Here, we demonstrate using a synthetic control method to evaluate the impact of the cyclone on the treated provinces (those impacted by the cyclone) compared to the synthetic control, a set of untreated provinces selected by matching on dengue cases (logged), temperature, and precipitation prior to Cyclone Yaku and logged population size (in 2020), aggregating weekly variables to four-week periods. The pool of untreated units used to construct the synthetic control includes locations in Mexico, Brazil, and Colombia.To avoid including provinces that also may have been impacted by Yaku, we remove provinces in Peru and Ecuador that experienced precipitation above their 80th percentile during the biweek of Cyclone Yaku, which may experience some treatment effect and bias our analysis toward underestimating the cyclone’s impact.

We use two different approaches. First, we apply the microsynth package to estimate a treatment effect across all the treated provinces, essentially taking an average across the five units to collapse them into a “macroprovince.” Each control unit is then assigned a weight such that the weighted average of time-varying mean temperature, precipitation, and cases (logged) across these control regions approximates those in the treated group prior to the cyclone. We additionally match on population (logged). The effect of the Cyclone is then estimated as the difference in cases between the treated provinces and their synthetic control (Abadie and Gardeazabal 2003) .

The synthetic control is able to match most of the outbreaks in the treated provinces, particularly since 2020 (Figure3A). The synthetic control is generally a bit cooler than the treated provinces, which is generally hotter than the mean temperature across all locations in the untreated pool (Figure3B). Similarly, the synthetic control generally matches the relative dryness of the treated provinces, although there is a large mismatch in 2017 when synthetic control was considerably more dry than the treated provinces (Figure3C). Cases in the treated provinces increased by a factor of 4.3 following the cyclone, and this increase was statistically significantly (p = 0.008, estimated using permutation tests)3D).

Figure 3: Synthetic control results where three treated Peru regions were matched to locations in Peru, Colombia, Mexico, Ecuador, and Brazil. (A) Dengue incidence aggregated to four-week periods from 2010 - 2023 across the treated provinces (black), the synthetic control (maroon), and all untreated province (grey). The dashed vertical line indicates when Cyclone Yaku occured. Panels B and C similarly display temperature and precipitation over time in the treated, synthetic control, and all untreated province. (D) The difference in cases between the treated and synthetic control provinces over time. Values above the dashed vertical line indicate greater incidence in the treated compared to synthetic control provinces. Each grey line is the result of a permutation test for comparison. Panels E and F similarly display the difference in temperature and precipitation over time. (G) Weighting of provinces in constructing the synthetic control. Treated provinces are indicated in pink. Excluded provinces are indicated in grey. [NOTE: I know the color scheme is bad for now, allows more flexibility until we have actual results]

WORK IN PROGRESS BELOW When each provinces in the synthetic control are disaggregated and the workflow is repeated for each province separately, we find evidence of a significant increase in cases following the cyclone in Peru but not Ecuador.

Figure 4: This figure is a work in progress, but the top panel mirrors 3A-3C for each of the treated provinces, labeled on the left hand side. The bottom panel is similar to Fig 3D, but each line indicates the treatment effect of logged cases for individual provinces. The dark lines are treated provinces and the gray lines are untreated provinces (placebo). The treated provinces are labeled on the right side of the figure with their ADM1 pcode. Note that provinces are arranged from north to south: EC24, EC13, PE24, PE20, PE14.

Abadie, Alberto, and Javier Gardeazabal. 2003. “The Economic Costs of Conflict: A Case Study of the Basque Country.” American Economic Review 93 (1): 113–32. https://doi.org/10.1257/000282803321455188.

Childs, Marissa, Kelsey Lyberger, Mallory Harris, Marshall Burke, and Erin Mordecai. n.d. “Climate Warming Is Expanding Dengue in the Americas and Asia.”