In this paper, we examine the utility of second-stage stratification within the first-stage units (FSUs) of large-scale sample surveys, focusing on India’s official labour force surveys, namely the NSSO Employment-Unemployment Survey 2011-12, the Labour Bureau Employment-Unemployment Survey 2013 – 14, and the Periodic Labour Force Survey 2022 – 23. These three surveys stratify households within the FSU in different ways, and we examine the implications of these choices for the statistical precision of the resulting estimates. Using Monte Carlo simulations on FSUs constructed from PLFS 2022 – 23 unitlevel data, we estimate the variance of each design across three labour market outcomes: labour force participation rate (LFPR), unemployment rate (UR), and labour income. Our primary finding is that the household-level stratification designs employed in these surveys do not necessarily improve precision. For labour force participation and unemployment, simple random sampling outperforms stratification in most FSUs. In general, gains depend critically on the correlation between the stratification variable and the outcome of interest. We propose an alternative education-based scheme that combines lower secondary (10^th pass) and graduate attainment. This design raises the median efficiency gain over the PLFS by about 14% for unemployment and 17% for income. We conclude with some reflections on the factors that are relevant to the choice of stratification design.

*We thank Anmol Somanchi and Prof G.C. Manna for helpful discussions. Views are authors’ and do not necessarily represent those of their institutions

Authors: 

Pratyush Priyadarshi, Anand Shrivastava, & Amit Basole