The conundrum of labour shortage in a labour surplus economy: an investigation of Nepal

Abstract

Nepal continues to be an agrarian economy with agriculture providing primary employment to 64% of the workforce and contributing 21.3% in the value-added according to the most recent figures available (World Bank, 2020a, 2020b). The pressure on agricultural land is also more than that in its neighbouring countries China and India, which are still regarded as surplus labour economies. While this cursory picture indicates that Nepal is also a labour surplus agrarian economy in the Lewisian sense (Lewis, 1954), recent literature has consistently referred to labour shortage in agriculture (e.g. Maharjan et al., 2013b; Pant, 2013; Sunam & Adhikari, 2016; Tuladhar et al., 2014 to mention few) and indicated the exhaustion of surplus labour (Pant, 2013; Tuladhar et al., 2014)—an apparent conundrum that warrants further analysis. The paper starts by statistically confirming the presence of surplus labour in Nepal based on the most recent household survey available for the country (Nepal Living Standards Survey 2010/11) despite recent reports of labour shortage in the sector. In the next part, the paper investigates the role of migration in explaining labour shortage following relevant literature on Nepal and other countries (Rozelle et al., 1999; Tuladhar et al., 2014). Towards this end, Propensity Score Matching estimation is carried out. The results indicate that though migration led to reduction in agricultural output, this reduction cannot be interpreted as labour shortage, thereby rejecting the hypothesis that migration can systematically explain the recent labour shortage in Nepal.

Appendices

Appendix

Appendix 1: Descriptive statistics of covariates used in PSM exercise and Cobb–Douglas production function estimate

Table 5 presents the descriptive statistics of some important control variables used in regression analysis, specifically for the migration equation. The mean value of harvest is around NPR 43,309 over the agriculture calendar year. There is a negligible difference in the value of harvest between migrant and non-migrant households. This difference is not statistically significant. The migrant household has an older head than the non-migrant household before migration.  The head of the migrant household has more formal education and slighter bigger household size. Migrant households have less share of children and elderly, while more share of men and women than non-migrant households. Caste/ethnicity division of migrant and non-migrant households are almost the same.

Table 5 Mean of Sample over Migrant and non-Migrant Household, the difference in mean and statistical significance

Regarding ownership, on average, migrant households have 0.77 hectares of land under control while non-migrant households have 0.70 hectares, with more plots and better quality of the land. Migrant households’ local residence area seems more electrified than that of non-migrant households. Similarly, migrant households seem to reside closer to the agriculture input market than non-migrant households. Wage employment is more prevalent in non-migrant communities than in migrant communities at the district level, and the rest of the variables have negligible differences (Table 5).

Table 6 Cobb–Douglas Production function estimation

Appendix 2: Balance, overlap, sensitivity analysis for PSM and Nearest Neighbourhood matching estimates

The matching on the estimated propensity score balance in our data as the density plot for the matched samples is nearly indistinguishable (Fig. 1).

The box plot (Fig. 2)  also indicates that our matched data are properly balanced. The median, 25th percentiles, and 75th percentile appear to be the same for matched sample.

The graph (Fig. 3) displays the estimated density of the predicted probabilities that a migrant household do not send migrant and the estimated density of the predicted probabilities that non-migrant household do not send migrant.

Neither plot indicates too much probability mass near 0 or 1. There is no evidence that the overlap assumption is violated (Fig. 3).

The matched sample results in Table 7, indicate that matching on the estimated propensity score balanced the covariates. We can see in Table 7 that Standardized differences for matched sample is less than 0.10 in absolute term, which is the rule of thumb value. So, our matching estimate balances the covariates. Most of the ratio of variance in the matched sample is around 1.

Table 7 Covariate Balance Summary

We also calculated Rubin’s B and Rubin’s R to check the balance. We found Rubin’s B 18.2 and Rubin’s R 1.33. Rubin (2001) recommends that B be less than 25 and R be between 0.5 and 2 for the samples to be considered sufficiently balanced.

Though balanced in observed covariates, criticisms remain about unobserved covariates which might lead to hidden bias. “In an observational study, one could never assert with warranted conviction that the naïve model is precisely true” (Rosenbaum 2020, p. 85). “The sensitivity of an observational study to bias from an unmeasured covariate is the magnitude of the departure from the naïve model that would need to be present to materially alter the study’s conclusions.” (ibid.)

The first row, in Table 8, Γ = 1, is the usual randomization inference. To alter our conclusion that migration has negative impact on value of harvest, Γ should be 2.4, which permits a substantial departure from random treatment assignment. It, Γ = 2.4, implies that two households with the same observed covariates may not have the same chance of migration. One such household is almost two and half times more likely to be a migrant because they differ in terms of unobserved covariates. Till Γ = 2.3, our conclusion does not alter. So, our study is not very sensitive to unobserved or ‘hidden bias’ (Table 8).

Table 8 Sensitivity analysis

Table 9 Propensity Score and Nearest Neighbourhood matching using different algorithms