Background Assessing socioeconomic inequalities in health and nutrition in developing countries receives much attention from both researchers and policymakers. However, the assessment of socioeconomic status (SES) is not straightforward. Researchers have either used a single proxy indicator or applied principal component analysis (PCA) -a dimension reduction technique – to create an index variable from a range of indicators reflecting SES of a household. PCA creates components from correlated variables and the first component is usually used as a proxy of SES. It primarily relies on linear associations between continuous variables. The Demographic Health Survey commonly uses the Filmer-Pritchett PCA (FP-PCA) on ordinal and continuous variables which converts ordinal variables to dummy variables to create a wealth index. Researchers have found that these dummy variables can produce spurious correlations. A polychoric PCA (pPCA) is methodologically advanced, can handle dichotomous, ordinal and continuous variables, and is beginning to gain recognition in public health studies as an improvement over FP-PCA. This study used both FP-PCA and pPCA on two different data sets to assess their performance.

Methods Two datasets were used: 1) a nationally representative, Food Security Nutrition Surveillance Project (FSNSP) and 2) a large-scale intervention Suchana’s evaluation data. Both datasets were collected in Bangladesh. FSNSP covered 2011–2014, whereas Suchana represents two cross-sectional surveys in 2016 and 2019.

Results A total of 14 and 11 correlated variables were used from FSNSP and Suchana, respectively, to create a socioeconomic index applying FP-PCA and pPCA. Variables, such as parental education, occupation, ownership of assets and other household characteristics were used. The eigenvalue, representing the total variance explained by the first principal component was higher for pPCA than FP-PCA (FSNSP: 6.4 vs. 5.2; Suchana: 4.5 vs. 3.5). Similarly, pPCA resulted in a markedly larger proportion of the variance being explained by the first principal component, increasing from 21% to 45% for FSNSP and 18% to 40% for Suchana. The Cronbach’s alpha was >0.7 for both data sets, reflecting that the indices generated had good reliability.

Conclusion The pPCA is methodologically advanced and suitable for continuous, categorical and ordinal data. The SES index created by applying pPCA was reliable and explained larger variabilities in SES for both nationally representative surveillance and survey data. The pPCA-generated SES index can be useful for assessing socioeconomic inequalities in public health nutrition contexts.