Measuring the Overall Burden of Early Childhood Malnutrition in Ghana: A Comparison of Estimates From Multiple Data Sources

Background: Childhood malnutrition contributes to nearly half (45%) of all deaths among children under 5 globally. The United Nations’ Sustainable Development Goals (SDGs) aims to end all forms of malnutrition by 2030; however, measuring progress towards these goals is challenging, particularly in countries with emerging economies where nationally-representative data are limited. The primary objective of this study was to estimate the overall burden of childhood malnutrition in Ghana at national and regional levels using 3 data sources. Methods: Using data from the long-standing Ghana Demographic and Health Surveys (GDHS), Ghana Multiple Indicator Cluster Survey (GMICS), and the emerging Ghana Socioeconomic Panel Survey (GSPS), we compared the prevalence of malnutrition using the extended composite index of anthropometric failure (eCIAF) for the period 2008-2011. This study included data for children aged 6-59 months and calculated all anthropometric z-scores based on the World Health Organization (WHO) Growth Standards. We tested for differences in malnutrition subtypes using two-group configural frequency analysis (CFA). Results: Of the 10 281 children (6532 from GMICS, 2141 from GDHS and 1608 from GSPS) included in the study, the only demographic difference observed was the children included in the GSPS were slightly older than those included in the GDHS and GMICS (median age of 36 vs 30 vs 33 months, P <.001). Based on the eCIAF, the overall prevalence of malnutrition at the national level was higher among children in the GSPS (57.3%, 95% CI: 53.9%–60.6%), followed by the GDHS (39.7%, 95% CI: 37.0%–42.5%), and then those in the GMICS (31.2%, 95% CI: 29.3%–33.1%). The two-group CFA showed that the 3 data sources also estimated different prevalence rates for most of the malnutrition subtypes included in the eCIAF. Conclusion: Depending on the data source adopted, our estimates of eCIAF showed that between one-third and half of all Ghanaian children aged 6-59 months had at least one form of malnutrition over the period 2008-2011. These eCIAF estimates should complement the commonly reported measures such as stunting and wasting when interpreting the severity of malnutrition in the country to inform policy decisions. This study relied on data obtained from 3 nationally representative surveys in Ghana, namely the 2008 GDHS, the 2009-2010 GSPS, and the 2011 GMICS. The 3 surveys were selected because they were the closest in years to one another between the 3 data sources (Table 1). Secondly, they were selected because they represent a meaningful baseline for tracking progress towards SDG 2.2. As part of efforts to meet the SDGs, the World Health Assembly Resolution 65.6 endorsed a comprehensive implementation plan on maternal, infant and young child nutrition in 2012, which specified a set of 6 global nutrition targets that should be met by 2025. The 2008 GDHS, 2009-2010 GSPS and 2011 GMICS data used in this study are likely to have been the most updated data sources on childhood malnutrition in Ghana at the time these targets were adopted, and can serve as useful baselines for ascertaining progress. The 2008 GDHS was carried out by the Ghana Statistical Service and the Ghana Health Service, with technical assistance from MEASURE DHS program.26 The 2009-2010 GSPS was conducted and managed by the Economic Growth Center at Yale University and the Institute of Statistical, Social, and Economic Research at the University of Ghana.27 The 2011 GMICS was carried out by the Ghana Statistical Service, with technical support from the United Nations Children’s Fund (UNICEF) and other international organizations.28 Abbreviations: GMICS, Ghana Multiple Indicator Cluster Survey; GDHS, Ghana Demographic Health Survey; GSPS, Ghana Socioeconomic Panel Survey. aData not publicly available at the time this study was conducted. We used the children sub-sample from each of the data sources, focusing on those aged 6-59 months who had their anthropometric (ie, weight and height) measurements recorded. Similar to previous studies,29,30 we excluded children aged less than 6 months because of the possibility of a significant decrease in adiposity during this period. We also excluded children based on the World Health Organization (WHO) flagging criteria31,32 for unusual or biologically implausible anthropometric measurements, as these height and/or weight values and their z-score derivatives are usually considered to be errors in measurement and/or recording. The 2008 GDHS, 2009-2010 GSPS, and 2011 GMICS all used a 2-stage stratified sample design. At the first stage, 411, 334, and 810 clusters were selected for the 2008 GDHS, 2009-2010 GSPS, and 2011 GMICS, respectively. These clusters were selected using a simple random sampling technique from an updated master sampling frame constructed from the 2000 Ghana Population and Housing Census. For the second stage, 30 households were selected from each cluster selected at the first stage for the 2008 GDHS whereas 15 households were selected for both the 2009-2010 GSPS and 2011 GMICS. In total, 11 913 occupied households were selected in the 2008 GDHS, out of which 11 778 were successfully interviewed, yielding a household response rate of about 99%.26 For the 2009-2010 GSPS, 5041 occupied households selected and 5009 were successfully interviewed, resulting in a household response rate of over 99%.22 The household response rate was close to 100% in the 2011 GMICS, given that 11 925 households were successfully interviewed out of a total of 11 970 occupied households selected.28 The 3 surveys conducted face-to-face interviews with all participants. The GDHS requires an authorization from the custodians before the data can be accessed. The authorization for access and access to the whole DHS database can be completed through the DHS program website at the address https://dhsprogram.com/. The 2009-2010 GSPS data is publicly available, but its usage is subject to terms and conditions of the data custodians. The data can be accessed from The World Bank Microdata Library website at the address https://microdata.worldbank.org/index.php/catalog/2534. Similar to the GDHS, the GMICS requires an authorization from the data custodians before the data can be accessed. The authorization for access and access to the whole MICS database can be completed through the MICS program website at the address https://mics.unicef.org/surveys. The 3 surveys fulfilled all ethical requirements and participants’ records in the databases were anonymized, fully concealing the identities of all participants. No further approvals are required for the retrospective use of these data. The study variables included age (in months), sex (male vs female), region of residence (Ashanti, Brong Ahafo, Central, Eastern, Greater Accra, Northern, Volta, Western, Upper East, and Upper West), height (in cm) and weight (in kg). Except for region of residence, the other variables were used to calculate a range of z-scores, including HAZ, WAZ, WHZ, and BMI-for-age (BMIZ) using the 2006 WHO growth standards.33 The following standard case-definitions were applied to each record to classify children as: wasted: WHZ <-2.0, stunted: HAZ <-2.0, underweight: WAZ <-2.0, or overweight: WAZ >2.0. We also used the CIAF and eCIAF to measure the overall prevalence of malnutrition (ie, over- and under-nutrition). The CIAF identifies 7 mutually exclusive subtypes of undernutrition, including: (a) no failure: normal WAZ, HAZ, and WHZ, (b) wasting only: WAZ <-2.0 but normal HAZ and WHZ, (c) wasting and underweight: WAZ and WHZ <-2.0 but normal HAZ, (d) stunting, wasting, and underweight: HAZ, WAZ, and WHZ <-2.0, (e) stunting and underweight: HAZ and WHZ <-2.0 but normal WAZ, (f) stunting only: HAZ <-2.0 but normal WAZ and WHZ, and (y) underweight only: WHZ <-2.0 but normal HAZ and WAZ. The CIAF sums all the subtypes (b) to (y) of undernutrition to estimate the overall prevalence of early childhood undernutrition in a population. The eCIAF accounts for over-nutrition by adding 2 more groups to the CIAF, namely (g) stunting and overweight, and (h) overweight only. Thus, the eCIAF can estimate the overall prevalence of early childhood malnutrition in a population. We compared the distributions of sex, age, height, weight, HAZ, WAZ, WHZ, BMIZ, stunting, wasting, underweight, and overweight between the 3 data sources; using the chi-square test or analysis of variance, as appropriate. Further, we used quantile regression models to make sex- and age-specific comparisons of the entire distributions of HAZ, WAZ, WHZ, and BMIZ between the GDHS, GMICS, and GSPS data sources. Quantile regression allows the estimation and inference on quantiles of a distribution without assumptions of normality and equality of variance required for standard regressions on mean values.34 Similar to Seirs et al,35 we generated quantile regression fits for every 2.5th percentile from the 0.05th to 0.95th quantiles of the 4 z-score values for each sex and age group. Estimates for each quantile were connected with lines and polygons were used to represent the 95% confidence intervals for the estimates, with non-overlapping confidence polygons indicating statistically significant differences. We used the eCIAF to calculate and plot the overall prevalence of early childhood malnutrition for each of the 10 regions in the country. These prevalence estimates were weighted to be regionally representative and reflect the regional variations in malnutrition across the country. We applied the appropriate weights (ie, PERWEIGHT for the GDHS, hhweight3 for the GSPS, and chweight for the GMICS) for each of the surveys. Given the 2-stage stratified sample design of the survey, we included information on both the primary and secondary sampling units, as well as using the region of residence as strata in specifying the survey design effects in the function svydesign of the R survey package. We used 2 methods to test for differences between the GDHS, GMICS, and GSPS for each of the subtypes of malnutrition captured in the CIAF and eCIAF. First, we calculated the prevalence of each of the subtypes of malnutrition in the eCIAF with their 95% confidence intervals, with non-overlapping confidence intervals indicating statistically significant differences. Similar to the weighting procedure explained above, these prevalence estimates were weighted to be nationally representative. Second, we used a two-group configural frequency analysis (CFA) to identify the subtypes of CIAF that are significantly different between the GDHS, GMICS, and GSPS. Unlike using confidence intervals to ascertain statistical significance, the CFA is a nonparametric method useful for analyzing the frequencies in multi-way contingency tables.36 In our two-group CFA, the observed frequencies were compared with expected frequencies in order to identify the “discrimination types” (ie, subtypes of undernutrition which are significantly different between each combination of 2 data sources). We made comparisons between the GSPS and GDHS, the GSPS and GMICS, and the GDHS and GMICS. With regards to the secondary objective of the study, we compared the changes in the prevalence of stunting, wasting, and overweight across the data sources. For the GDHS, we compared our estimates from the 2008 survey to the latest prevalence estimates from the 2014 survey.37 For the GMICS, we compared our estimates from the 2011 survey to the latest prevalence estimates from the 2017-2018 survey.28 Although the second wave of data collection for the GSPS took place in 2014-2015, the data are currently not publicly available, thus we were unable to report on the changes in prevalence estimates from this data source. For all tests conducted in this study, statistical significance was set at 0.05 alpha level. All data restructuring and analyses were conducted in R version 3.6.0.38 Specifically, we used the R-packages data.table39 for data preparation, confreq40 for the CFA, quantreg41 for the quantile regression, ggmap42 for mapping regional malnutrition prevalence rates, survey43 for weighting prevalence estimates to be nationally representative, zscorer44 for calculating anthropometric z-scores in the GSPS data, and ggplot245 for plotting the anthropometric z-scores.