Surveillance on the endemic of Zika virus infection by meteorological factors in Colombia: a population-based spatial and temporal study

Colombia, the third-most populous country in Latin America is located in the northwest of South America, and has 32 departments (see Fig. 1) [21]. The population of Colombia is concentrated in the Andean highlands and the Caribbean coast. The equator is across the southern Colombia. The climate in Colombia is characterized as tropical, so its warm and isothermal weather condition is a favorable environment for breeding mosquitoes. The Amazon Rainforest, located in the southeastern Colombia, is one of the main habitats of mosquitoes.

We collected data on ZIKV infection records and meteorological information in Colombia from 2015 to 2016. These data were collected at the department level, which was analogous to the state level in the United States. Weekly data on ZIKV infection counts of each department were reported to the Colombian National Institute of Health through the public health surveillance system. These data were publicly available at Epidemiological Bulletin in Colombia (http://​www.​ins.​gov.​co/​buscador-eventos/​Paginas/​Vista-Boletin-Epidemilogico.​aspx). Note that Bogota, the capital of Colombia, only has available ZIKV infection records in 2015 and the first 2 weeks of 2016. A report mentioned that ZIKV infection in Bogota is not endemic because it is at 2640 m above the sea level, and most cases were immigrant cases from people travelling to Bogota from Cundinamarca [22]. Hence, we combined the data from Bogota and Cundinamarca before the 2nd week of 2016. Meteorological factors were obtained from 42 monitoring stations across 32 departments. These data were retrieved from the Weather Underground (https://​www.​wunderground.​com/​).

Weather conditions were quantified by 15 meteorological measurements from 42 monitoring stations in Colombia (Fig. 1), including temperature (minimum; maximum; average), humidity (minimum; maximum; average), dew point temperature (minimum; maximum; average), sea level pressure (minimum; maximum; average), wind speed (maximum; average), and rainfall. Because the missing data rate was as high as 29.3% among those monitoring stations, we chose to apply kriging technique because spatial autocorrelation (measured by Moran’s I) is strong, spatial stratified heterogeneity is weak (measured by q-statistic [23]), and Pearson’s correlation coefficient is small. Evidence in details can refer to Additional file 1: Table S1, Table S2 and Figure S1. Therefore, we especially adopted a local time-space kriging method to impute missing data for each department according to collected measurements along with calendar day and geographic information of monitoring stations (i.e., latitude and longitude) [24]. For matching the weekly ZIKV case data, we first selected the weekly maximum values of maximum temperature, maximum dew point temperature, maximum humidity, maximum sea level pressure, and maximum weed speed from monitoring stations in each of the ten departments with multiple monitoring stations. The weekly minimum values of the previous meteorological factors were determined similarly. The weekly average values for the department of all previous indicated meteorological factors were estimated by average across monitoring stations. For the other 19 departments with only one monitoring station, weekly measurements were derived from maximum, minimum, and average values of corresponding meteorological variables per week. For the three departments without monitoring station, we applied the local time-space kriging method again to impute their meteorological measurements from the other departments. Afterward, we applied a geodetector method in those imputed data, which is a technique to preliminarily detect whether covariates are responsible for an outcome measurement [23].

In order to take both nonlinear lag effects of meteorological measurements and spatial correlation into account simultaneously, we applied the distributed lag nonlinear model (DLNM) for assessing the nonlinear associations among ZIKV infection, weather, and temporal lagged effect [25]. We assumed that the number of ZIKV infection cases at time t in department d, denoted by Y_dt, followed a Poisson distribution POI(μ_dt), where E(Y_dt) = φμ_dt, and φ is a scale parameter. The DLNM could be regarded as a generalized linear model with additional cross-basis functions, which are interactions constituted by a nonlinear function for a covariate and a nonlinear function for a lag variable; hence, this study established the DLNM in a quasi-Poisson model framework to take possible over-dispersion into account:where α is the intercept, and the offset term is the logarithm of population in each department. A natural cubic spline f(t) with respect to the calendar time for 104 weeks from 2015 to 2016 is a time smoother with seven degrees of freedom (df) to control long-term temporal autoregressive correlations. Weather conditions and lagged effects were analyzed in the cross-basis function CB(Weather, lag), which is an interaction between a b-spline of a meteorological factor and another b-spline of the lag factor. The spatial function f _spat (d) describes spatial local heterogeneity of ZIKV incidence by applying Markov random fields, which follow a normal distribution with mean \( \sum \limits_{d^{\prime}\in {\Omega}_d}{f}_{spat}\left({d}^{\prime}\right)/{N}_d \) and variance \( {\sigma}_d^2/{N}_d \)[26]. In details, we defined a neighborhood as an adjacent department d^′ sharing a part of boundaries of another department, and a neighborhood set containing all adjacent departments around the department d was denoted by Ω _d . The number of adjacent departments in Ω _d was denoted by N _d , and f _spat (d^′) is the spatial estimated effect in each adjacent department. Thus, the mean can be regarded as the average neighborhood effect. In addition, \( {\sigma}_d^2 \) is the original variance in the department d, whereas the number of neighbors diluted its spatial variance. The spatial function can evaluate the remaining variation of ZIKV incidence not explained by meteorological factors. Each department can obtain a spatial estimate from the spatial function, which can be interpreted as the logarithm of relative risk (logRR) in ZIKV of a department compared to the average of all departments. In the other words, the exponential spatial estimate can account for the excessive relative risk (RR) of ZIKV incidence in each department over the whole country. Moreover, the spatial function can detect each department as a high-risk area of ZIKV infection after controlling meteorological factors as the 95% confidence interval (CI) of the spatial estimate significantly greater than zero. We adopted a 3-step model selection to choose the best model. In details, the first step is to determine the best model with only one cross-basis function from 14, 16, 18, and 20 lags by comparing the quasi Akaike information criterion (QAIC) among them. The second step is to determine the best cross-basis function with different degrees of freedoms from 3 to 7 in each meteorolgocial measurement. The third step is selecting the best model with one, two, and three cross-basis functions. Eventually, the best model has three cross-basis functions for average humidity, total rainfall, and maximum temperature with 20 lags. The chosen criterion in each step depends on the smallest QAIC (data not shown but available upon request from the authors). Each estimated cross-basis function was transformed into RR, while the reference levels varied in different meteorological factors. Diagnosis of the residuals were conducted via plots of autocorrelations and partial autocorrelations, and distribution of residuals over time.

Sensitivity analyses were performed by increasing the length of lags to 22 and 24 in the final model. We calculated the similarity of estimated cross-basis function with the first 21 lags (including lag 0) by using the RV-coefficient, which is a correlation measurement between two matrices [27]. Like the general correlation coefficient, a higher RV-coefficient close to one reflects more similarity between two matrices with the same dimension. We also applied the analysis of variance to compare the spatial estimates to see whether the change of lag in cross-basis functions could affect the spatial function.

Data management of this study was carried out in SAS V9·3 (SAS Institute Inc., Cary). Data imputation and model fitting were conducted in the R software, version 3·3·3 (R Development Core Team, 2011). All maps were generated in QGIS v2·18·2. Statistical significance was determined by a type I error level of 0·05.