Ask When—Not Just Whether—It's a Risk: How Regional Context Influences Local Causes of Diarrheal Disease

Abstract

Contemporary epidemiology is enriched when it incorporates ecological concepts about systems and dependencies. With regard to diarrheal disease, the causes of which are many and interacting, the dynamics of within- and between-community disease transmission have distinct components but are also linked in important ways. However, few investigators have studied how regional-scale disease dynamics affect local patterns of diarrheal disease transmission. Characterizing this dependence is important for identifying local- and regional-level transmission pathways. We used data from active surveillance of diarrheal disease prevalence gathered from February 2004 through July 2007 in 21 neighboring Ecuadorian villages to estimate how disease prevalence in spatially and temporally proximate villages modulates the influences of village-level risk and protective factors. We found that the impact of local, village-level interventions such as improved latrines and water treatment can be quite different under conditions of high and low regional disease prevalence. In particular, water treatment was effective only when regional disease prevalence was low, suggesting that person-to-person spread, not waterborne spread, is probably responsible for most between-village transmission in this region. Additional regional-scale data could enhance our understanding of how regional-scale transmission affects local-scale dynamics.

The dynamics of enteric pathogen transmission operate at multiple scales. Within a community, the socioecological structure—which entails person-to-person contact rates, rates of contact with pathogens in the environment, and community infrastructure—has been noted to affect transmission dynamics (1–3). Between communities, transmission acts through human travel patterns (4), as well as by movement of pathogens through the environment via, for example, water flows (5) and food transportation (6). While there are clear distinctions between these 2 scales, they are also linked. For example, a regionally remote village may have a different social structure than a less remote village (1). Dependencies between neighboring villages in terms of prevalence levels have been studied (7), but to our knowledge no research has focused on how regional-scale dynamics affect local transmission dynamics. Characterizing this impact is essential for pinpointing the pathways through which local- and regional-level diarrheal disease exposures operate. Progress on this front is crucial, as diarrhea continues to be a major cause of disability and mortality in the developing world (8) and is the second-leading cause of childhood mortality worldwide (9). In this paper, we address this gap by empirically studying how regional prevalence levels affect local drivers of diarrheal disease.

A number of studies have established that multiple modes of enteric pathogen transmission contribute to diarrheal disease burden and that these modes interact (10). However, a majority of observational studies of diarrheal disease have examined only a single transmission pathway (2). Of those that have evaluated multiple pathways, most investigators have made analytical choices which tacitly assume that each transmission route acts independently of the others (2). Such an approach reflects a view that risk-factor effects are inherently static—that is, that these risk factors do not depend on other factors. Here, we conceptualize risk-factor effects as being inherently dynamic. Specifically, we envision disease rates in neighboring villages as defining a context that changes how within-community disease dynamics operate.

The importance of modeling diarrheal disease processes as context-dependent has been noted in the literature (4, 11, 12) and has been shown empirically in several studies. Examples include the effects of poor community sanitation on household risk in households with improved sanitation versus those with unimproved sanitation (13, 14) and how the determinants of disease transmission differ 1) in public household domains versus private domains (15), 2) in rural settings versus urban settings (13), and 3) on the basis of proximity to specific locales (16). Demographic information can also be important; Fink et al. (17) found that risk factors for childhood mortality can act differently depending on the age group of the children being studied. Similarly, environmental information can change how disease dynamics work; Reiner et al. (7) found that risk-factor effects for cholera and dependencies between neighboring communities can be modified by weather conditions.

In the current study, we analyzed data collected through active surveillance of diarrheal disease in 21 communities in rural coastal Ecuador to study how regional and local disease dynamics interact. We built a spatiotemporal model, envisioning each village's diarrhea prevalence as a Markov chain whose transition probabilities depended on 1) water, sanitation, and socioeconomic status variables as well as 2) the disease prevalences of the other villages in the region. Within this framework, we estimated the influence that neighboring communities had on a given community, with particular focus on how diarrheal disease prevalence in neighboring communities modulated the effects of village-level risk and protective factors.

METHODS

Data description

This study took place in the northern coastal Ecuadorian province of Esmeraldas in the cantons of Eloy Alfaro and San Lorenzo, in an area containing approximately 150 small villages. While most villages in the area lie upstream from Borbón, the main population center in the region, some villages are located downstream from Borbón. Additionally, some villages have direct access to a newly constructed road, while others do not. While the endemic diarrhea prevalence in the study villages is low, there is notable temporal fluctuation (see Web Figures 1–5, available at http://aje.oxfordjournals.org/). Over 60% of the households in the study region report not treating their drinking water (defined as chlorinating or boiling water prior to consumption), and over 60% report disposing of human waste in the open; both practices vary notably by village and over time. More details about the study population can be obtained elsewhere (4, 18).

Twenty-one villages were selected through stratified sampling. Five strata were used: road villages, villages lying along each of 3 river basins (Onzole, Santiago, and Cayapas), and villages lying downstream from Borbón (“Baja Borbón”). All households in these 21 villages were recruited for the study. All consenting households were visited on a weekly basis from February 18, 2004, through July 4, 2007, by 25 community health workers employed by the study investigators. The self-identified head of each household was questioned about whether anyone in the household had felt sick during the previous week; if the person said yes, he or she was asked more detailed questions about symptoms and treatment. A diarrhea episode was defined as at least one 24-hour period during the past week in which an individual experienced 3 or more loose stools. Institutional review board committees at the University of Michigan (Ann Arbor, Michigan), Trinity College (Hartford, Connecticut), and Universidad San Francisco de Quito (Quito, Ecuador) approved all protocols.

Explanatory variables used in the analysis included individual, household, and village characteristics. Data were collected from 2 sources: an ongoing census of the villages, which was conducted annually during the study period, and a case-control study of diarrhea (the Ecologica, Desarrollo, Salud, y Sociedad (EcoDeSS) Project) being conducted concurrently in the same villages between August 2003 and October 2008 (4). Data on hygiene, water sources, and water treatment were collected as part of the case-control study. Demographic data (including age, sex, education of household head, and job type of household head) and household information (including household size, type of sanitation facility, and various indices of asset ownership) were collected through the census surveys. We were interested in the associations of water/sanitation and infrastructure variables (e.g., improved sanitation, improved water source, water treatment, asset ownership, quality of housing construction) with diarrheal disease prevalence. We controlled for group-level factors (e.g., household size, community size, community remoteness) as well as individual-level factors (e.g., age <5 years, job stability (any household member; usually the household head), >6 years of education, and sex). For a more complete description of the variables and how data from different sources were synchronized, see Markovitz et al. (18).

The unit of analysis was the village; that is, data on all variables were averaged at the village level (or converted to a percentage, as appropriate). This was done primarily because the variables for which information was collected through the case-control study were not available for all study participants and had to be aggregated at the village level using inverse-sampling-probability-weighted averages. We divided the prevalence outcomes into 4 ordinal categories—“low” (<0.6%), “low-medium” (0.6%–<2.2%), “high-medium” (2.2%–<5.2%), and “high” (5.2%–100%). The average prevalence values in the 4 groups were 0.1%, 1.3%, 3.1%, and 7.4%, respectively. This particular split into 4 categories was chosen to minimize the within-group variance of the prevalence values, effectively creating the most homogeneous groups possible. We found that 4 categories provided a substantially better categorization than 3 categories and that going beyond 4 categories did not improve resolution substantially (see Web Appendix and Web Figure 6).

Statistical modeling

Let Y_k(t) ∈ {0,1,2,3} be the village k ordinal diarrhea prevalence category at time t with corresponding covariates X_k(t). Similar to the model of Reiner et al. (7), this analysis focuses on the 1-lag (1-week-lagged) transition probabilities, conditional on covariate values and the states of the other villages, Y_{_k}(t):

(1)

A schematic description of this model is given in Figure 1.

Figure 1.

Map of the study area showing the locations of villages, roads, and rivers, overlaid with a graphical depiction of the transition probability model, Ecuador, February 2004–July 2007. p_ij(X) refers to the probability of moving from state i to state j for a specified value of the covariates and the regional risk index, denoted compactly as X. For simplicity, this diagram shows the transition of a single village's diarrhea prevalence state while the others remain fixed, but the model supposes that mass transition occurs at each time point.

In principle, one could include the time t − 1 prevalence status of each village as a predictor, but this approach would risk overfitting and results could be difficult to interpret. To produce a tractable model for the transition probabilities, we introduce the “regional risk index,” R_k(t), for village k at time t as a compact way of quantifying the effect of surrounding villages:

(2)

where W_kj is interpreted as the effect of village j on village k. Our parameterization of W is similar in spirit to a “gravity model”:

(3)

where P_k is the population size of village k, δ_kj is the distance (along the road/river, in this case) between villages k and j, and θ₁, θ₂, and θ₃ are parameters defining the relative contribution of each part. The intuition here is that connections involving larger villages and pairs of villages that are closer together may be stronger. These weights are applied to village pairs that fall into the same stratum (see Eisenberg et al. (4) for a graphical depiction of the strata). From knowledge of the area, village pairs that lie in different strata are unlikely to contact each other, and their inclusion presents a practical statistical modeling issue. Specifically, the weighted average can only be calculated when all villages with nonzero weights are observed and only about 22% of weeks have observations for all 21 villages, leading to substantial data loss if all weights are allowed to be nonzero. For this reason, we constrain all between-strata links to be zero and expect this to have minimal substantive impact. We define g(y) as the average diarrhea prevalence (scaled by the maximum) seen in the data set for villages that are in state y. Because the number of villages enrolled was in rough proportion to the density of villages within a given area, our regional risk index is approximately proportional to the regional-level exposures incurred by village k at time t. As Bertuzzo et al. (5) noted, the propagation of pathogens can differ depending on village orientation (upstream vs. downstream). We experimented with this dynamic in this model and found that it made little difference, so we opted for the simplified specification above.

A convenient method for analyzing ordinal data is the use of the proportional odds model (19). We model the transitions using a separate proportional odds model for each possible time t state:

(4)

In narrative terms, the linear predictor——places the observation on a risk continuum (we call this the “logistic risk index”), and its position relative to the threshold parameters, ψ_ij, determines the response probabilities. The linear predictor and threshold parameters depend on the current state (the subscript i), reflecting the autoregressive nature of this model. Regional-scale spatial dependence is captured by R_k(t). In our fitted models, for simplicity, we equate covariate effects across levels; that is, we constrain for all r.

We chose to model within- and between-village dependencies in this way (using the mean structure) for several reasons. First, a typical approach to controlling within-village temporal autocorrelation and between-village spatial autocorrelation is to use random effects. Such models are not well developed for non-Gaussian continuous measurements like prevalences and, we argue, are not as cleanly interpretable for categorical measurements as the model used here. Second, by operating on the means, our approach avoids the necessity to integrate out the high-dimensional random effects that induce spatial and temporal dependence, which, under all but the simplest structures, presents a significant computational burden. Third, using this strategy, it is straightforward to allow the spatial dependencies (the regional risk index) to interact with village-level predictors, which was the primary purpose of this work.

We used the above framework to assess the influence of neighboring villages as well as the water, sanitation, and infrastructure factors of interest, while controlling for demographic variables, village characteristics, and temporal dependence. Next we tested, one at a time, each of the risk factors of interest for a significant interaction with R_k(t), which would have indicated that the effect depended on the regional context.

RESULTS

A total of 3,400 village-weeks were observed; 2,437 of these were “low” weeks, 712 were “low-medium,” 213 were “high-medium,” and 38 were “high.” Model 1 contained all predictors, with no interaction terms. For model 1, the estimated structure for the W matrix was effectively the same as if k, j were in the same village stratum. For computational reasons, this structure for was used in subsequent models and was not re-estimated. Model fit was assessed graphically by breaking observations into “bins” based on the value of the logistic risk index and comparing the observed transition probabilities with those estimated under the model. This was done separately for each prevalence state, and results are displayed in Figure 2. The model fitted reasonably well, although there was evidence that it underestimated the probability of transitioning from “high-medium” diarrhea prevalence to “low” diarrhea prevalence (Figure 2C). Figure 2 also displays the temporal autocorrelation in the data. For example, the figure shows that, for all values of the logistic risk index, a village currently in the “low” state has a higher probability of moving into a “low” state at the next time point than a village that is currently in the “low-medium” state.

Figure 2.

Probability of moving into each state of diarrhea prevalence as a function of the logistic risk index when a village is currently in the “low” state (diarrhea prevalence <0.6%; panel A), the “low-medium” state (0.6%–<2.2%; panel B), or the “high-medium” state (2.2%–<5.2%; panel C), Ecuador, February 2004–July 2007. The solid lines show the model-fitted transition probabilities as a function of the logistic risk index. Individual points are the observed transition probabilities in subsets of the data defined by splitting the logistic risk index into 7 groups based on the quantiles; the x-axis location of the point in the plot is the median value of the logistic risk index within that split. The “high” prevalence state (5.2%–100%) is not included in this plot because there were not enough data to support the kind of stratification required.

Summaries of our model fits can be found in Table 1. The coefficients in the second column of Table 1 are interpreted as how much a 1-unit change in the predictor changes the value of the logistic risk index, whose implications in terms of the probability distribution can be seen visually in Figure 2. In the model with no interactions, both “percent educated” and “decreasing village remoteness” were found to be significantly associated with risk, conditional on the other predictors in the model, while other village demographic characteristics were not. Diarrhea prevalence levels were found to be lowest in villages located downstream from Borbón, which is also visually apparent in Web Figures 1–5. The regional risk index and all water/sanitation indicators were found to be significant drivers of prevalence, as well as the ownership index.

Table 1.

Results From Probability Modeling of Transitions Between Diarrhea Prevalence States, Ecuador, February 2004–July 2007

	Model 1a		Models 2–6b
	β (SE)	βRk(t) (SE)	βX (SE)	βRk(t)X (SE)
Village characteristicsc
% under age 5 years	1.07 (1.34)
% female	0.24 (1.25)
Population size (in hundreds)	−0.08 (0.04)
Average household size	−0.07 (0.06)
% with >6 years of education	1.14 (0.47)*
% of adults with a stable job	0.66 (0.53)
Remoteness from other villages	−5.82 (1.77)**
Village location
Baja Borbón	Reference
Road	0.64 (0.37)
Santiago	0.63 (0.36)
Cayapas	1.01 (0.28)***
Onzole	0.72 (0.33)*
Regional risk index, Rk(t)	0.48 (0.16)**
Standard of living
Asset ownership index (average score)d	−3.80 (1.22)**	−0.47 (1.05)	−4.53 (1.45)**	2.47 (2.68)
Housing construction quality (average score)e	−0.56 (0.57)	−0.26 (0.73)	0.88 (0.60)	−0.67 (1.00)
Water/sanitation
% with improved latrine	0.71 (0.27)**	1.30 (0.34)***	1.45 (0.34)***	−1.87 (0.63)**
% with improved water source	−0.34 (0.16)*	−0.41 (0.23)	0.43 (0.21)*	0.19 (0.42)
% treating water	−0.65 (0.25)**	−1.10 (0.32)***	0.08 (0.24)	1.60 (0.64)*

Abbreviation: SE, standard error.

* P < 0.05; **P < 0.01; ***P < 0.001.

^a Model 1 included all predictors but no interactions.

^b Models 2–6 included each of 5 variables (ownership index, housing quality, improved latrine, improved water source, and treated water) as an interaction term with R_k(t), one at a time. The table shows the main effect coefficients for the predictor and for R_k(t) and the interaction term . All other variables were controlled in models 2–6, but for brevity those estimates are not displayed here.

^c Data were collected at the individual level but were summarized at the village level for this analysis.

^d The ownership index was based on the number of assets owned by the household. Higher weights were assigned to items that cost more. Respondents were asked about many different types of assets. “Minor” assets (e.g., a blender) were coded 1 and more “major” assets (e.g., farming tools and equipment) were coded 3; items in between were coded as 2. The total asset index was the sum of these numerical values. The index was scaled by the maximum possible score (i.e., owning every possession asked about) to yield a measure between 0 and 1.

^e The housing quality score was based on the quality of materials used to construct the house and roof. The housing material was coded as 2 for any cement, 1 for wood only, and 0 for any other combination. The roofing material was coded as 2 for zinc and 0 for any other combination. The total score was calculated as (housing + roof)/4.

Using models 2–6, which included each of 5 variables as an interaction term with R_k(t) one at a time, only the proportion of inhabitants who drank treated water (P = 0.013) and the proportion with an improved latrine (P = 0.006) showed evidence of interaction with the regional risk index (Table 1, column 4). Water treatment was shown to be less protective as the regional risk index increased. Specifically, when regional risk is low, increasing the percentage of inhabitants consuming chlorinated/boiled water from 0% to 100% decreases the logistic risk index from about −0.8 to −2.0 (Figure 3). This increase in water treatment corresponds to a sharp increase in the probability of moving into a low state, indicating a lower-risk situation (Figure 2, parts A–C). When regional risk is high, increasing the proportion consuming treated water does little to change the logistic risk index, indicating that the distribution of the village prevalence at the next time point is only slightly affected by water treatment. On the other hand, having an improved latrine is a risk factor when regional risk is low, and this risk dissipates as the regional risk index increases (Figure 3).

Figure 3.

Interaction plots for each significant interaction with R_k(t)—percentage of inhabitants using an improved latrine (A) and percentage of inhabitants treating their water (B), coastal Ecuador, February 2004–July 2007. Each of 5 variables (asset ownership, housing quality, improved latrine, improved water source, and treated water) was entered into the model as an interaction term one at a time (models 2–6). All other predictors were set to their mean values in order to produce this plot. Higher values on the y-axis indicate an increased probability of residing in a higher state of diarrhea prevalence. Low, medium, and high regional risk were defined as the minimum, 67th percentile, and 90th percentile of the observed values for R_k(t).

DISCUSSION

The transmission of infectious pathogens depends on the social, environmental, and biological context, which all operate at different time scales. In this article, we have presented empirical data on how the level of disease in a region affects transmission dynamics within a localized community. Analysis of 4 years of active diarrheal-disease surveillance data across 21 communities suggested that community disease dynamics are influenced by the disease status of neighboring communities, both in terms of a direct increase in risk and in terms of potentially modulating the strength of risk-factor effects.

Transmission of disease between communities can occur through environmental processes or through social processes. In our study region in northwestern Ecuador, the study communities lay alongside one of 3 river basins. Transport of pathogens via the river is one potential pathway. The plausibility of this pathway's operating as a means of pathogen transport depends on a variety of environmental factors, such as water flow, water temperature and pH, the depth and width of the river, and types of human exposure to the river. Another potential pathway is travel between these neighboring communities by infected individuals. These infected individuals can either transmit pathogens directly or can contaminate different environmental media such as food, water, soil, or fomites. Determining which of the multiple possible pathways regional and local transmission act through is important for the optimal design of disease prevention strategies.

The manner in which between-village transmission affects local disease dynamics can be important for identifying regional- and local-level transmission pathways. In our data, the protective effect of water treatment was diminished considerably when the diarrheal prevalence in the surrounding villages was high. This may suggest that when there is little disease exposure from outside villages, water acts as a key local transmission pathway, and therefore, treating the water confers protection. The fact that there was a diminished protective effect of water treatment when there was high regional risk suggests that this regional exposure does not act through water pathways; rather, transmission is probably acting through person-to-person contact or person-to-environment contact. Understanding which pathways are operating in a given context is important given the finding that persons using treated water may neglect other behavior-based preventative measures that may be more important when there are higher levels of pathogens in neighboring communities (20).

The other contextual dependency identified in our analysis was that related to the role of improved sanitation (latrines). When regional-level risks are low, improved sanitation is a risk factor for diarrheal disease. The observation that improved sanitation in rural regions may be associated with increased risk of disease is consistent with the fact that latrines are often located close to the house, where family members spend large amounts of time—especially young children, who are at higher risk for diarrheal disease than older age groups. When improved latrines are not used, people may travel farther from the house to defecate; therefore, children are less exposed to the site of defecation, conferring less overall risk to the community. The potential for this transmission pathway to be operating will depend on both the quality of the latrines and the population density of the community. Additionally, when people in this region use latrines, they are often shared among several families, effectively concentrating more pathogens in a smaller area. When there are lower levels of pathogens in the environment (i.e., when regional risk is low), this concentration of pathogens at a single sanitation location may be a primary source of local transmission. However, when regional risk is high, transmission may be occurring through other pathways as well, such as person-to-person contact and food-sharing, making pathogen concentration at the sites of latrines comparatively less important. Few data on these nuances of sanitation interventions exist. As such, further studies on sanitation are needed to better understand how risks may change with differing qualities of latrines under differing user population densities.

The ability to more precisely understand disease dynamics requires proper characterization of spatial and temporal trends. In particular, spatially and temporally proximate conditions can affect which risk factors are important. The Markov chain model used here can account for these spatial and temporal trends by allowing the probability distribution of disease prevalence at the next time point to be modulated by 1) covariates, 2) the current state of the village, and 3) the states of the nearby villages, as well as their interactions. This work focused primarily on how temporally proximate spatial trends interact with covariates, showing that 1) the recent prevalence in nearby villages affects local prevalence and 2) this interdependence can change which local risk factors are important, giving evidence about which pathways are driving regional transmission. Specifically, this model indicates that regional-level disease dynamics are not primarily acting through water or sanitation pathways, providing evidence that person-to-person transmission or food-sharing is more likely to be responsible.

Our work was not without limitations. First, the modeling framework used here relies on the assumption that the future is conditionally independent of the past, given present conditions and all measured covariates. Therefore, we assumed that the measured covariates we incorporated accounted for any long-range dependence in village-level diarrhea rates. Second, an important unmeasured variable in this analysis was the rates of travel to and from each village. For example, if a particular village had a local attraction (e.g., festivals), this village would be more susceptible to cycling of pathogens as tourists came and went on a regular basis, possibly changing how temporally and spatially proximate conditions affected disease. Other important contact-based factors could not be incorporated into this analysis; for example, some villages had schools that served children in neighboring villages. Related to this point, the proxy measure we used for “distance” was the best information we had available, but it did not completely reflect the connectivity between villages. Finally, it is possible that unmeasured temporally varying factors were affecting regional risk patterns. For example, weather patterns may modulate individual behavior, as well as the ability of pathogens to survive in the environment, changing the way villages confer risk on each other. The results of this work must be interpreted as having “averaged over” these kinds of effects, but we admit the need to study such dynamics in situations where relevant temporally varying covariates are available.

Most researchers define the risk factors for diarrheal disease as static and model the effect of multiple risk factors as a combination of their marginal effects. This should be reevaluated. As we have shown here, the magnitude and even direction of particular risk-factor effects can be very different under different circumstances. Ignoring contextual variables in this setting is analogous to leaving an important interaction out of a regression model. To properly understand which variables are most relevant to prevention of disease, we suggest assessing not only what variables are important but also when they are important. Doing so can elucidate previously unapparent information about how regional and local disease dynamics are operating. To continue the progress that has been made in decreasing the burden of diarrheal disease, researchers should identify which contextual variables are likely to operate across a variety of environments and examine their effects, thereby improving both the accuracy and generalizability of results.