Data Source
Current study utilized data from multiple rounds of the Cambodia Demographic Health Survey (CDHS). Up to the present time, five rounds of CDHS have been conducted, spanning from 2000 to 2022. The National Institute of Statistics (NIS), operating under the Ministry of Planning, spearheaded the implementation in conjunction with the Ministry of Health. This demographic health survey received financial support from the United States Agency for International Development (USDA) and technical support from ICF.
The CDHS, a nationally representative survey based on the population, employed a two-stage stratified cluster sampling design for participant selection. For first-stage sampling, a specific number of enumeration areas (EAs or clusters) were chosen in a manner that was proportionate to their size, while considering urban and rural distinctions. In the subsequent stage, a defined number of households (usually 25–30) were randomly selected from the listed households within each EA, utilizing a simple random selection process. The survey covered eligible women (aged 15–49), men (aged 15–54), and children under the age of 5 years residing in the selected households. The objective of the survey was to furnish estimates related to fertility, mortality, maternal and child health care services, and reproductive health facilities at both national and sub-national levels. Comprehensive details concerning the study’s design, sampling method, framework, and non-response rate are available in specific reports for each round 12,16–18.
We extracted data from the last four rounds of the CDHS Kid’s records for current research. The initial round of the survey was excluded due to the absence of information on the diet diversity variable, which stands as the focal point of interest. The guidelines from the Demographic Health Survey pertaining to diet diversity calculations are intended for the youngest child aged between 6 and 23 months, who resides with their mother. Accordingly, final sample encompassed 9,056 youngest children falling within the age bracket of 6 to 23 months, currently living with their mothers. This amalgamation includes data from CDHS 2005 (2,020 participants), CDHS 2010 (2,372 participants), CDHS 2014 (2,143 participants), and CDHS 2021-22 (2,321 participants).
Dependent & Independent Variables
The outcome variable of interest is inadequate intake of minimum diet diversity. In the survey, the mother or caregiver of the child was inquired about the types of food the child had eaten in 24 hours leading up to the interview. These food groups are: (i) breast milk, (ii) dairy products (milk, yogurt, cheese, infant formula), (iii) grains, roots, and tubers, (iv) vitamin A-rich fruits and vegetables, (v) other fruits and vegetables (vi) eggs, (vii) flesh foods (viii) legumes and nuts. Aligned with the most recent guidelines from the WHO (WHO, 2017) and the Demographic and Health Surveys (as outlined in the Guide to DHS Statistics DHS-7). 19 assigned a score of 1 for each food group consumed, while non-consumption was denoted by a score of 0. The total score was then calculated by summing the scores across all food groups. Subsequently, a binary outcome variable was generated to compute the inadequate intake of MDD. Children who consumed five or more groups were assigned a score of “0”, while those who consumed fewer than five groups were assigned a score of “1”. Until 2017, adequate MDD intake was defined as consuming four out of seven food groups. The revised indicator acknowledges breast milk as an additional food group and changes the criterion for achieving MDD. It now requires the consumption of a minimum 5 food groups out of 8, in contrast to the previous requirement of a minimum 4 food groups out of 7 20.
To measure the disparity in inadequate MDD, we computed the inequality for all eight food groups that make up the MDD separately. Considering this, we ended with a total of nine outcomes: (1) not meeting the MDD, (2) not currently breastfeeding, (3) did not consume dairy, (4) did not eat grains, roots, and tubers, (5) did not eat vitamin A-rich fruits and vegetables, (6) did not eat other fruits and vegetables, (7) did not eat eggs, (8) did not eat flesh foods, (9) did not eat legumes and nuts 21.
To establish the conceptual framework for comprehending the variations in failing to achieve Minimum Dietary Diversity (MDD), distinct determinants were selected at both child and maternal levels. These selections were guided by insights gleaned from the literature review, primarily due to their established associations with MDD 10,21–24. Child-related factors encompassed the child’s sex, age recategorized from a continuous scale into three groups (6–11, 12–17, and 18–23 months), and birth order reclassified as “first,” “second or third,” and “fourth or higher.”
Maternal factors included maternal age, categorized into three groups (15–29, 30–39, and 40–49 years old), employment, educational, and marital status, preceding birth interval categorized into three groups (First birth, < 36 months, and > = 36 months), media exposure combining newspapers, radio, and television into levels: no exposure, partial (two media), and full (all three), and wealth quintiles derived through asset indices based on household attributes like amenities and materials. A wealth score, a pivotal variable reflecting data variability, was computed through principal component analysis. Urban/rural households were assessed separately, and the indices were standardized. Households were divided into quintiles based on wealth scores, ranging from the poorest (20%) to the richest (20%) 25,26.
The study also considered place of residence (rural vs. urban), household gender composition, and size categorized into three groups (< 4, 5–9, and > = 10 members). Paternal factors, including father’s education and occupation, were included as well. Health service utilization factors encompassed the number of antenatal care (ANC) visits during pregnancy (categorized into < 4 visits and > = 4 visits), place of delivery, and visits to healthcare facilities in the last 12 months (yes vs. no).
Analytical Approach
The data from the CDHS demonstrate a hierarchical structure, wherein children aged 6–23 months form the first level nested within clusters at the second level and provinces at the third level. This hierarchical arrangement could potentially violate standard logistic regression assumptions, such as independence and equal variance 27. Given the presence of regional heterogeneity, a single-level model proves insufficient, leading to inaccuracies in parameter estimation. To evaluate how various layers of explanatory variables impact insufficient minimum dietary diversity and account for variations at the cluster and province levels, we utilized a multilevel binary regression model. This approach allows for a simultaneous examination of effects at both group levels (clusters and provinces) and individual levels on outcomes, while addressing the lack of independence among observations within groups 28. Hierarchal analysis enables the investigation of both inter-group and intra-group variability, as well as the relationship between variables at both group and individual levels. To achieve this, we employed a three-level variance component model. This model initially decomposes the overall geographic variation into clusters and provinces, with respect to the probability of a child “i” in cluster “j” and province “k” for inadequate MDD or inadequate consumption of each of the food groups, utilizing the Eq. (1)
\(log\frac{\left({\pi }_{ijk}\right)}{\left({1-\pi }_{ijk}\right)}= \alpha +{X}_{ijk}\beta +{\mu }_{jk}+{\varOmega }_{k}\) Eq. (1)
Where, subscript \(i,j, k\) denote children, cluster, and province, respectively.
\({\pi }_{ijk}\) is the probability of ith children of cluster j and province k inadequate MDD.
\(\alpha\) is the intercept that is the effect of feeding inadequate minimum DD when the effect of all explanatory variables is absent
\({X}_{ijk}\) selected socio-economic and demographic characteristics for ith children of cluster j and province k
\(\beta\) vector of constants giving the log odds resultant from one unit change in variable \({X}_{ijk}\)
\({\mu }_{jk}\) & \({\varOmega }_{k}\) are the random effect for cluster \(j\) and province \(k\)
Each of the residual differentials is assumed to be normally distributed with a mean of zero and variances of \({\sigma }_{\mu }^{2}\) and \({\sigma }_{\varOmega }^{2}\), variances quantify the between-cluster (\({\sigma }_{\mu \varOmega }^{2}\)), between-district (\({\sigma }_{\varOmega }^{2}\)) variation. The variance at level one (children) is assumed to be a constant in binary models 29,30. Furthermore, we investigated the proportion of geographic variation attributed to clusters and provinces for each of the nine outcomes within the hierarchical model. This was achieved by dividing the variance at a specific level by the total geographic variation (i.e., for the cluster level, \(\frac{{\sigma }_{\mu }^{2}}{{\sigma }_{\mu }^{2}+{\sigma }_{\varOmega }^{2}}\)).
The initial and final survey years, 2005 and 2021-22, were divided into two groups to assess factors contributing to inadequate MDD disparities between the two periods, and non-linear Binder Oaxaca decomposition analysis employed 31–33. The Blinder- Oaxaca- technique was first introduced by 34,35. This approach segregates differences into three parts: explained endowment, coefficient, and unexplained interaction. Endowment signifies variation due to variable changes, while coefficient results from variable composition shifts. For instance, in our study, focusing on inadequate MDD, child’s age influences it. The difference in inadequate MDD attributed to child’s age change is explained, while unexplained accounts for variations in age effects. The analysis draws from the prevalence of not meeting MDD and coefficients from multivariate binary logistic regression models for each survey year. Compliant with DHS guidelines, analyses incorporated survey weights, clustering, and stratification to ensure national representativeness 36.