Geographical distribution of typhoid risk factors in low and middle income countries

In order to identify risk factors for typhoid candidate risk factor variables were validated against typhoid disease burden data. A systematic literature review was previously conducted to establish a typhoid disease burden database [3]. To include more surveillance data than in the previous database, a second round of additional search was conducted with more relaxed criteria: extension of publication years from 1990–2013 to 1980–2013, inclusion of hospital-based studies, and community-based studies which did not meet the criteria in the previous search (e.g., blood culture, refusal rates). Thirty-eight studies were selected after considering the availability of matching data sources (Fig. 1), and these provided 84 disease burden data points by age group, as shown in Table 1.

For candidate risk factors it was critical to find data sources available over time and across countries. Demographic and Health Surveys (DHS) and National Geographical Data Center (NGDC) satellite lights data were used for this analysis. Both DHS and NGDC provide not only longitudinal datasets, but also geo-coordinates of DHS clusters and NGDC lights data. Thus, each incidence rate estimated in a specific year and site was compared to candidate variables obtained from the same year of the surveillance (or close approximation) around the area where the surveillance was conducted. Figure 2 shows an example of Pakistan and demonstrates how both DHS and NGDC night lights data were paired with the disease burden data from the surveillance site. For example, clusters within 100 km from a site were included by using coordinates for the DHS dataset. For countries where there is no GPS information available, the next smallest geographical unit, which is the state-level in DHS, was used to select data in the state where the site was located. It is arbitrary whether the radius should be greater or less than 100 km. However, the distribution of DHS clusters is dispersed in some countries or states, whereas some clusters are densely gathered in other countries depending upon the size of countries or other logistical issues. Because this study includes multiple countries across continents, it is important to ensure that certain numbers of DHS clusters were included for all the surveillance sites. The 100 km radius ensures at least over 400 households for all the sites. This is a better way than pairing up with available country-level indicators ignoring the time and site location where surveillance was conducted.

It should be noted that although DHS carries out standardized surveys there are some differences across countries and survey periods in terms of questionnaire types and data availability. For this reason, variables which are most relevant to typhoid fever and exist consistently over time and across places were selected. As water and sanitation issues are the main risk factors for typhoid infection [19, 21‐24], drinking water sources (hv201) and toilet facility types (hv205) were considered from the DHS dataset. Since the major categories of the variables are standard but individual codes are country-specific [25], the variables were sub-categorized as shown in Table 2. Each category was converted into a proportion out of total households who responded to each question. Prior to creating TRF indices and making formal comparisons, the sub-categories were first screened by simple scatter plots and correlations with the disease burden data by age group. The sub-categories which indicated unexpected signs or weak correlations (ρ < 0.4 for hv201, ρ < 0.3 for hv205) were excluded from further analysis.

In addition population density was considered because highly crowded areas tend to have high rates of typhoid infection in developing countries [4, 6, 18]. In order to measure population density at the small geographical level over time, satellite lights data were obtained from NGDC. The satellite lights data are available from 1992 to 2012 in smaller geographical units: 30 arc second grid spacing (approximately 1 km²), and the same radius was used as DHS. The higher the lights were observed, the higher the population density was assumed. Prior to applying this assumption, correlations between lights data and census data were examined to ensure that lights data can be used as an appropriate proxy (overall ρ = 0.7).

The sub-categories filtered through the first screening (2 7, 9, 10, 11, 12 for hv201 and 1, 2, 3, 6, 7 for hv205) were used to create a composite index called the typhoid risk factor (TRF) index. To identify the most suitable index which explains the trend in typhoid incidence rates, five TRF indices were created based on different combinations of the sub-categories for the DHS variables and NGDC lights data. The variables in each set were first standardized individually by subtracting the mean and dividing by its standard deviation. The standardized values were then averaged across the variables. Some of the variables which go to the opposite direction were converted by subtracting from one, so all variables go towards the same underlying concept which is the typhoid risk level. For example, while a high proportion of the households who use river or lake as their water source would be at high risk, a high proportion of those who drink private tap water would be at low risk. The values were converted into a range from zero (low risk) to one (high risk) by using the max-min method and categorized into three percentiles (cTRF1: 0–25%, cTRF2: 25–75%, cTRF3: 75–100%). This method is more commonly used in the field of social science, and more details were extensively discussed elsewhere [26].

Incidence rates (/1000) can be considered as a non-negative integer value. Count models are suitable for our risk-factor validation because the count model estimates non-negative integer values and specifies the incidence rate with a mean that is dependent on exogenous variables [27, 28]. The Poisson or its variants (e.g., negative binomial) typically takes the exponential form for expected demand, and the Poisson probability density function can be written aswhere n is observed demand and λ_i is the mean, λ_i = exp(z_iβ). Overdispersion may occur when the variance is greater than the mean of the distribution [29, 30]. Additional file 1: Appendix 1 provides more details on how the overdispersion issue was managed. Care must be taken when dealing with counts of events observed in small geographical areas. While spatial correlation was less of a concern due to the data points obtained from dispersed geographical locations in this study, Moran I test was carried out to confirm that there is no significant spatial autocorrelation in the dataset [31, 32] (see Additional file 1: Appendix 1).

Because hospital-based incidence rates are likely to be higher than population-based incidence rates a dummy variable was created to treat them separately. This dummy variable was then multiplied by an age group categorical variable, and used as an interaction variable. Each of the five TRF indices was regressed against typhoid incidence rates separately, and the most statistically significant TRF index was chosen for the mapping analysis.

To understand the model fit the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) fit tests were used. Considering the relatively small sample size and generalizability of the model, a Hausman test (1978) on the equality of coefficients was carried out with 50% of randomly selected data from the main model [29]. The most statistically significant TRF index was chosen and used for further analyses.

While a specific year of the datasets was considered to match the same period of each surveillance data for the validation model above the most recent datasets were obtained from the same data sources for the mapping analysis to show the current states of typhoid risk levels across countries. The final TRF index was re-estimated at two different levels: sub-national boundary level and grid-cell level. All countries where relevant information was available from both data sources were selected. Although NGDC provides lights data for all countries, DHS data were only available for the selected countries. As shown in Fig. 3, 66 out of 88 countries where drinking water sources and toilet facility types in the DHS household recode data exist were chosen for the mapping analysis.

The most recent years of DHS surveys vary country by country ranging from 1996 (Brazil) to 2013 (Nigeria). Because it would not be sensible to compare data in 1996 with the ones in 2013 the 66 DHS countries were divided into three groups (Table 3), and the same TRF index was estimated separately for each group.

Figure 4 demonstrates how the TRF index was re-estimated by the sub-national boundary level and the grid-cell level. DHS provides sub-national boundaries which are usually the state or province level. This is the smallest geographical unit in which DHS sample weights were based. The proportions of sub-categories of the risk factor variables validated above were calculated by the DHS sub-national boundary level. The sample weights provided by DHS were applied so that the outcomes are representative at the population level. The mean value of lights data was also estimated by the sub-national boundary level for population density. The TRF index was then estimated by the same level. In addition, the grid-cell analysis was conducted in order to look at typhoid risk levels in the smaller geographical unit than the sub-national boundary level.. DHS provides geo-coordinates for the groups of households that participated in the survey, known as clusters for 46 of 66 countries. Considering the different sizes of countries and computational power, the size of grid-cells in each country was determined based on three categories: large (radius = 20.5 km), medium (radius = 10.5 km), and small (radius = 5 km). Centroids were created for all grid-cells. DHS clusters and NGDC lights data within a radius defined above from a centroid were included for each grid-cell. The risk factors were calculated in the same manner by the grid-cell level. It should be noted that unlike the sub-national boundary level, there were no sample weights applied in the grid-cell level analysis.