Spatio-temporal clustering analysis and its determinants of hand, foot and mouth disease in Hunan, China, 2009–2015

Hunan province is one of the central inland provinces of China situated between 108°47′ to 114°15′ east longitude and 24°38′ to 30°08′ north latitude with a typically subtropical monsoon climate. Hunan, with a population of approximately 67.8 million, covers an area of about 211,800 km², and has 14 districts and 124 counties. High humidity, abundant heat, high population density and high levels of migration promote the development and spread of enteroviruses within the Province.

Surveillance data on HFMD in Hunan was obtained from the China Information System for Disease Control and Prevention (http://​www.​chinacdc.​cn). This data covered the study period (2009 to 2015), and included case types as well as pathogen types (only for laboratory diagnosed cases) for all the new cases identified during the study period. The diagnostic criteria were based on the 2010 HFMD Diagnosis and Treatment Guidelines by the National Health and Family Planning Commission of the People’s Republic of China. Cases without home addresses, clinical or laboratory diagnosis information were excluded from the study. Meteorological data was obtained from the Hunan Meteorological Administration. The Hunan map at county level was also provided by the Information department of the China Disease Control and Prevention Center.

Spatial autocorrelation is defined as the spatial dependence among the given attribute value of one geographic unit and its neighboring units, which lies in almost all spatial observations [16]. Global spatial autocorrelation is used to measure the overall clustering tendency in the study region while the local one can be further used to clarify the patterns and the exact location of the clusters among local counties [17‐19]. In this study, we used global Moran’s I statistic and local indicators of spatial autocorrelation (LISA) to investigate spatial association regarding incidence of HFMD in the Hunan province [20].

The global Moran’s I statistic ranges from −1 to 1. An I > 0 indicates positive autocorrelation and the closer it is to 1, the more aggregated the distribution is. An I < 0 indicates negative autocorrelation and the closer it is to −1, the more dispersed the distribution is. An I = 0 indicates no autocorrelation, suggesting that the disease is randomly distributed. The significance of Moran’s I is assessed by Monte Carlo test using Z statistic and P value (Z ≥ 1.96 or Z ≤ −1.96 indicates statistical significance at the 5% level) [21]. By using the LISA map, four patterns of spatial correlation can be detected: high-high clusters; low-low clusters; high-low clusters; and low-high clusters [22]. In reality, the high-high clusters (i.e. high-incidence districts surrounded by other high-incidence districts) are the most important patterns for the purpose of disease prevention and control.

The formula for global or local Moran’s I is:

Where n is the number of districts or counties; x _i and x _j are the incidence rates in districts or counties i and j;\( \overline{x} \) is the average incidence rate in the entire study area; and w _ij is the spatial weight between districts or counties i and j. In this study, by choosing the binary spatial weight matrix in Rook contiguity rule [23], Geoda 1.6 software was used for the spatial autocorrelation analysis whilst ArcGis 10.2 software was used to visualize the results.

Kulldorff’s method of retrospective space-time scan statistic was used to detect the geographical clusters of HFMD cases at the county level based on a discrete Poisson model [24]. In this approach, a dynamic cylindrical window was used to scan for different time and geographic area to detect all possible clusters. The circular geographic base moves with the radius varying with the population range of the area, while the height varying with the defined time interval [25]. For each scanning, the relative risk (RR) is calculated using the ratio of the observed case number to the expected case number within and outside the windows as well as the log likelihood ratio (LRR). The P values for clusters detected are calculated by Monte Carlo randomization method [26, 27]. The window with the maximum LLR is assumed to be the most likely cluster, and other windows with a statistically significant LLR are defined as secondary clusters ranked according to their LLR value.

The LLR for a given window can be calculated as follows:

Where N is the total number of cases; n is the observed number of cases within the scan window; E(n) and N − E(n) are the expected number of cases within and outside the scan window under the null hypothesis (H₀: The spatio-temporal clustering of the study area are caused by random factors.), respectively; and I ^′is an indicator function. It is equal to 1 when the window has more cases than expected under the null hypothesis, otherwise it is 0. In space-time scan analysis, the spatial unit is county, with 124 counties in Hunan Province; the temporal unit is month, with 84 months from 2009 to 2015. The spatial size of scanning window was limited to 30% of the total population at risk whilst the temporal size was set to 30% of the total study period in order to scan for small to large clusters. The number of Monte Carlo randomization was set at 999. By setting the time frame of the scan analysis to be one month, we can control the time trends and observe the cluster changes in the entire study. SaTScan 9.4 software was used to perform the spatiotemporal cluster analysis of HFMD and ArcGis 10.2 software was used to visualize the results.

Spatial autocorrelation is frequently present in the observations and it appears that neighboring units tend to have more similar value than those that are distributed randomly [28]. In considering spatial dependence, Besag et al. present the autologistic regression model which is a special type of ordinary logistic model introducing a spatial autocorrelation term (SAuto cov _i ) in the form of weighting coefficients [29]. Temporal autocorrelation commonly exists in infectious disease. In our study, we modified the autologistic regression by introducing TAuto cov _i as the temporal autocovariate, which is set with the incidence cases of last month in the same county. Meanwhile, many previous studies have found that population density is significantly associated with HFMD occurrence [30, 31], so we introduced the population density cov _i as the covariate to adjust for population density in different counties.

For the modified autologistic regression model, the form of the model is given by equation:

The probability of the occurrence of the outcome can be illustrated as follows:where y _i is the dependent variable which refers to the probability of spatiotemporal clusters occurrence, P _i is the mean of y _i , or equivalently, the probability of y _i =1, and x denotes the exposure factors in meteorology (the temporal resolution of independent variables is one month which coincide with the time frame of the SaTScan analysis). cov _i is the population density covariate. SAuto cov _i , TAuto cov _i are the spatial and temporal autocovariate, respectively. β, p, q and r are the regression coefficient. i is the index of the county.

SAuto cov _i is the weighted average of the probabilities of the geographic unit i which is surrounded by other k _i geographic units. w _ij is the spatial weight between county i and j, \( {w}_{ij}=\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{${h}_{ij}$}\right. \) and h _ij is the Euclidean distance between the centroids of county i and j. P _j is the probability of the event occurring in the neighbors of county i, which is equal to 1 when the event occurs, otherwise it is 0. In this study, P _j refers to the probability of spatiotemporal clusters occurrence in neighboring county j of county i. A county with the distance within 31,250 m from county i was defined as the neighbors of county i [31]. (see Additional file 1: Table S1–1, Table 2).

In this study, after calculating the variance inflation factor (VIF) and tolerance to decrease multicollinearity (see Additional file 1: Table S3–1), five independent variables were selected: monthly average rainfall, monthly average temperature, monthly average wind speed, monthly average relative humidity, monthly total sunshine. We also introduced the population density as the covariate. All variables were standardized. A forward stepwise method was used in the ordinary logistic regression model. We then introduced the SAuto cov _i and TAuto cov _i until each variable was statistically significant. Cox & Snell R² and Nagelkerke R² were used to test the goodness of fit of logistic or autologistic model. The higher the statistic value are, the better the model fits the observations. We used the Relative Operating Characteristic (ROC) to compare the probability predicted by the model with the reality (the occurrence of the spatiotemporal clusters). Closer to 1 means a better fit, which rages from 0.5 to 1. SPSS 22.0 software was used to conduct the autologistc regression model and ArcGIS 10.2 software was used to perform spatial analysis.

There were 895,429 HFMD cases reported in Hunan Province from 2009 to 2015, with the average annual incidence rate of 194.57 per 100,000 (ranged from 54.31 per 100,000 in 2009 to 318.05 per 100,000 in 2014). Of the total cases, 8263 (0.92%) were severe and 352 (0.39‰) were death.

A total of 559,033 male cases and 336,396 female cases were reported during the study period and the male-to-female incidence ratio was 1.66:1 (ranged from 2.01:1 in 2009 to 1.53:1 in 2014). Most patients were younger than 5 years old, accounting for more than 90% of all reported cases. Scattered children have the highest incidence rate, followed by nursery children and school students. Among 30,377 laboratory confirmed cases, CoxA16, EV71 and other EV accounted for 16.67%, 47.62% and 35.72%, respectively. The predominant pathogen was EV71 in most years except for 2013 and 2015 (Table 1).

The monthly distribution of HFMD cases in Hunan is shown in Fig. 1, which indicates that the major peak appeared between April and July whilst the minor peak observed between September and November. The occurrence of HFMD shows significant seasonality with the wave-like increasing tendency.

Global Moran’s I statistics was performed to investigate the presence of global autocorrelation, which were listed in Table 2. It clearly shows that the HFMD cases were not randomly distributed and that there was high global autocorrelation among HFMD number at county level in Hunan Province. The results of local autocorrelation analysis were mapped in Fig. 2. The LISA map shows that the high-high clusters (dark red color) were mostly in central and northern districts including Changsha, Yiyang, Loudi, Xiangtan, while southwestern districts of Hunan Province were low-low clusters (dark blue color). Nevertheless, contrary to other cities in western districts, Xiangxi showed a high-high pattern in the LISA map in recent years.

Table 3 and Fig. 3 show the scanning results of statistically significant spatial-temporal clusters of HFMD from 2009 to 2015. Seventeen (17) spatial-temporal clusters were detected each year during the study period, including 7 most likely clusters and 10 secondary clusters. All the clusters occurred in the period of the major peak of incidence except for 3 secondary clusters in 2011 and 2015. Clustering areas gathered in the northern regions in 2009, in the central regions from 2010 to 2012, moved to central-southern regions in 2013 and 2014, and central-western regions in 2015.

Considering the spatial and temporal autocorrelation which existed in Hunan HFMD cases, this study introduced the auto covariates SAuto cov _i and TAuto cov _i  on the basis of the ordinary logistic regression model (see Additional file 1: Table S1–4). The occurrence of the HFMD spatio-temporal cluster in the given county was the dependent variable (It is equal to 1 when the spatiotemporal cluster occurs, otherwise it is 0), while the selected exposure factors were the independent variables (they were continuous variables so should not to be assigned). Table 4 shows the results of the autologistic regression. The OR values for monthly average rainfall, monthly average temperature and monthly average relative humidity are all greater than 1, which indicates that these variables are risk factors that are positively related to the occurrence of HFMD spatial-temporal clusters. On the contrary, the OR value for monthly average wind speed is less than 1.

In addition, the goodness of the model was measured by the Cox & Snell R² and Nagelkerke R² statistics and the ROC value was used to compare the reality with the predicted outcome. All values were listed in Table 5 which indicated that the autologistic regression model had a better goodness of fit than the ordinary one.