Multi-country analysis of the effects of diarrhoea on childhood stunting

Diarrhoea is an important cause of death and illness among children in developing countries; however, it remains controversial as to whether diarrhoea leads to stunting. We conducted a pooled analysis of nine studies that collected daily diarrhoea morbidity and longitudinal anthropometry to determine the effects of the longitudinal history of diarrhoea prior to 24 months on stunting at age 24 months. Data covered a 20-year period and five countries. We used logistic regression to model the effect of diarrhoea on stunting. The prevalence of stunting at age 24 months varied by study (range 21 – 90%), as did the longitudinal history of diarrhoea prior to 24 months (incidence range 3.6 – 13.4 episodes per child-year, prevalence range 2.4 – 16.3%). The effect of diarrhoea on stunting, however, was similar across studies. The odds of stunting at age 24 months increased multiplicatively with each diarrhoeal episode and with each day of diarrhoea before 24 months (all P < 0.001). The adjusted odds of stunting increased by 1.13 for every five episodes (95% CI 1.07 - 1.19), and by 1.16 for every 5% unit increase in longitudinal prevalence (95% CI 1.07 - 1.25). In this assembled sample of 24-month-old children, the proportion of stunting attributed to ≥5 diarrhoeal episodes before 24 months was 25% (95% CI 8 - 38%) and that attributed to being ill with diarrhoea for ≥2% of the time before 24 months was 18% (95% CI 1 - 31%). These observations are consistent with the hypothesis that a higher cumulative burden of diarrhoea increases the risk of stunting. © The Author 2008; all rights reserved. This analysis pooled longitudinal information from nine prospective studies that enrolled participants at or near birth and followed them with regular anthropometric measurements and daily records for diarrhoeal surveillance (Table 1). Details of the field methods of each study were published elsewhere.21,23–30 We aimed to combine multiple cohorts using original data from studies that collected longitudinal information on diarrhoea and growth. We used our literature search to guide us in identifying groups of investigators. Our PubMed search terms were ‘diarrhea’ AND ‘growth’ AND ‘longitudinal’ with the following limits: humans, English, Spanish, all infant (birth 23 months) and preschool children (2–5 years). This search identified 89 articles. We then searched for potential studies among the cited references. We selected studies that enrolled children in the first year of life, collected daily records for diarrhoeal surveillance and performed at least three anthropometric assessments. We excluded studies of HIV-infected children. This was not a systematic review, however, because we used expert opinion among the selected articles to identify groups of investigators with considerable experience in conducting longitudinal studies of diarrhoea and growth. We identified 14 investigators from this search, contacted them by electronic mail or telephone and obtained complete data for nine longitudinal studies. One investigator did not wish to participate, two investigators no longer had access to the original data, one investigator provided incomplete data and one investigator provided us with complete data, but we were unable to use this study because it had only two anthropometric assessments and it conducted diarrhoeal surveillance over a 2-week period only. Each study measured height in children at regular time intervals ranging from once monthly to every 4 months. Summary of nine studies included in a pooled analysis of the effect of diarrhoea on stunting at 24 months of age a(controls). We requested three separate sets of de-identified data from each participating investigator. For the serial anthropometrics, we requested a data file with the child's identification code, the date of measurement, weight in kilograms, length or height in centimetres, sex and the child's date of birth. For the daily records of diarrhoeal surveillance, we requested a data file with the date, the number of liquid and semi-liquid stools in that day or, alternatively, if the number of liquid and semi-liquid stools was not available then we requested investigators to provide maternal reporting of whether the child had diarrhoea or not without any further definition being required. For baseline information, we requested a data file with the number of years of maternal education, a measure of household sanitation, the number of people in the house and if available a measure of household income or a measure of household goods and services. We asked investigators to include other determinants of socioeconomic status (SES) in their community. We used data from children who had at least three height measurements and a recorded date of birth. We excluded measurements with a height-for-age Z-scores (HAZ) >4 and those with a HAZ <−7. We chose a HAZ <−7 as the lower cut-off because at least one quarter of children from the ‘Bangladesh 1978’ study had a HAZ <−4. All investigators collected socioeconomic data at baseline. Contributing investigators reported on the socioeconomic variables that were important in each study in published papers or by personal communication. All studies were conducted in rural or poor urban communities. Nonetheless, given that these studies were conducted in different countries with different cultural settings, different socioeconomic indicators and using different questionnaires, we created an aggregate SES variable. We used a combination of variables to construct the aggregate SES variable for each study. For ‘Brazil 1990’, we used a composite index based on the possession of household goods and services,23 maternal education and sanitation. For ‘Brazil 1989’, we used sanitation and household crowding.24 For ‘Bangladesh 1978’, we used maternal education, water source and household crowding.25 For ‘Peru 1995’, we used maternal educational, sanitation and household income.21 For ‘Peru 1985’, we used maternal education, sanitation and household crowding.27 For ‘Guinea-Bissau 1987’ study, we used maternal education, sanitation and household crowding.28 For ‘Guinea-Bissau 1994’, we used maternal education, household crowding and having a zinc roof.29 For ‘Ghana 1990’, we used maternal education, crowding and having a zinc roof.30 Socioeconomic data were not available for ‘Peru 1989’.26 Once we identified the above socioeconomic variables, we created rank scores for each of these variables. We used these rank scores to create an aggregate SES score, and stratified this aggregate score into low, intermediate and high categories (Appendix 1). We calculated HAZ by comparing length or height with the 2006 World Health Organization (WHO) international growth reference.31 We used an algorithm provided by the WHO in Stata 9 (StataCorp, College Station, TX) to calculate HAZ. We defined stunting as a HAZ less than or equal to 2 SD below the mean of the reference population (i.e., HAZ ≤−2). The international consensus definition for a day of diarrhoea is three or more liquid or semi-liquid stools in a day.15,16 Contributing investigators generally used this definition, although one study defined diarrhoea as four or more liquid stools per day,25 and three relied solely on mother's report of diarrhoea.28–30 Another study defined the onset of a diarrhoeal episode as three or more liquid or semi-liquid stools in a day and when the mother indicated that the child had diarrhoea.21 We used each investigators definition for a day of diarrhoea, but used a standard definition for an episode of diarrhoea for all studies, i.e. we defined the first day of a diarrhoeal episode as the onset of diarrhoea; and, defined the end of a diarrhoeal episode as the last day of diarrhoea that was followed by two consecutive days without diarrhoea. We calculated the cumulative diarrhoeal incidence per child-year as the number of diarrhoeal episodes divided by the number of days at risk over a specified time period and multiplied by 365. We calculated days at risk for diarrhoea as the total days of follow-up minus the days of diarrhoea but for the first day of an episode. We calculated the longitudinal prevalence of diarrhoea as the sum of the duration of all diarrhoeal episodes over a specified time period divided by the total number of follow-up days over that same time period and multiplied by 100. Stunting may be reversed over time. While the epidemiological definition of stunting is clearly established, there is no standard epidemiological definition for reversibility of stunting. A small change in HAZ between two points in time may be incorrectly defined as reversibility of stunting. For example, a change in HAZ from −2.01 to −1.99 between two time periods can be hardly considered recovery. Reversibility of stunting can also be affected by regression to the mean.32 To address these concerns, we defined recovery from stunting between ages t1 and 24 months for a child who was stunted at t1 months but not stunted at 24 months and for whom HAZ24 > r × HAZt1, where HAZ24 was the child’s HAZ at 24 months, HAZt1 was the child’s HAZ at t1 months and r was the correlation coefficient between HAZ24 and HAZt1 for the subset of children who were stunted at t1 months and not stunted at 24 months. That is, we did not include children for whom HAZ24 ≤ r × HAZt1 in the category of those who recovered. The objective of our analysis was to determine the effect of diarrhoea prior to 24 months of age on stunting at 24 months of age. The primary outcome in our analyses was the prevalence of stunting at 24 months. Because not all children were measured at exactly 24 months of age, we accepted the HAZ measurement at the oldest age in the interval between 18 and 24 months of age as the HAZ measurement at 24 months. We selected 24 months of age as the reporting age for this analysis because the majority (54%) of children from all nine studies contributed data at this age. In contrast, only 45% of children contributed data at 3 years of age and 28% of children contributed data at 5 years of age. We first conducted exploratory data analysis to determine the shape of the relationship between the cumulative burden of diarrhoea prior to 24 months of age and the log odds of stunting at 24 months. We then used logistic regression to model the prevalence of stunting at 24 months as a function of the cumulative burden of diarrhoea prior to 24 months. In our logistic regression model, the outcome was coded as 1 if a child was stunted at 24 months of age and coded as 0 if otherwise. We included the history of diarrhoea prior to 24 months as a continuous covariate. We required children to contribute at least 250 days of diarrhoeal surveillance to be included in the regression analysis. All studies contributed data on 48 or more children for this analysis. We constructed our regression model manually (Appendix 2). Because study and sex were important determinants of stunting at 24 months, we modelled the log odds of stunting at 24 months as a function of diarrhoea prior to 24 months, sex and study. In constructing our regression model, we began with three fixed-effects parameters for each study: an intercept, a parameter for the study-specific effect of diarrhoea on stunting and a parameter for the study-specific effect of sex on stunting. We compared nested models using the likelihood ratio test (LRT) to identify the model with the fewest number of parameters. We used the LRT to determine if we could pool studies to summarize the effect of diarrhoea on stunting. We first compared a regression model with only one parameter to explain the effect of diarrhoea on stunting vs a regression model with study-specific parameters to explain the effect of diarrhoea on stunting. We then compared a regression model with only one intercept vs a regression model with study-specific intercepts, and a regression model with only one parameter for a sex effect on stunting vs a regression model with study-specific parameters for a sex effect on stunting. We used the Hosmer-Lemeshow test to determine goodness-of-fit in logistic regression.33 We used Pearson residuals and deviance residuals to identify outliers,34 and used Pregibon’s delta–beta statistic to identify influential data points.34 In separate regression analyses, we estimated the odds ratio of stunting at 24 months of age by four categories of cumulative diarrhoeal incidence and four categories longitudinal diarrhoeal prevalence prior to 24 months. To calculate attributable risks, we categorized cumulative diarrhoeal incidence before 24 months (<5 episodes and ≥5 episodes) and longitudinal diarrhoeal prevalence before 24 months (<2% and ≥2%) into only two groups that represented a ‘low’ or ‘high’ cumulative burden. We calculated the proportion of stunting at 24 months of age attributed to having a high cumulative burden of diarrhoea prior to 24 months of age using parameter estimates obtained from logistic regression.35 We conducted biostatistical analyses in Stata and R (R Foundation for Statistical Computing, www.r-project.org). Fewer children had complete information on the requested socioeconomic variables. To determine whether socioeconomic status confounded the effect of diarrhea on stunting, we conducted a subset analysis using the data of children with complete SES data. To determine if the results of our regression model were affected by the exposure period, we modelled the effects of diarrhea prior to 23 months and the effects of diarrhea prior to 22 months on the prevalence of stunting at 24 months of age. We also conducted various subset analyses to exclude children who were stunted between birth and 6 months of age. Because not all children had an anthropometric measurement before 6 months of age, fewer children and fewer studies were included in these subset analyses. In the subset analyses that excluded children who were stunted at 6 months of age, we included HAZ at 6 months in our regression model. We accepted the HAZ measurement at the oldest date in the interval between 3 and 6 months of age as the HAZ at 6 months.