Technology Heterogeneity and Poverty Traps: A Latent Class Approach to Technology Gap Drivers of Chronic Poverty

Abstract

The analysis of household wealth dynamic remains an important methodology in the identification of poverty traps. To overcome measurement issues in survey data, livelihoods-based approaches of the dynamics of poverty are typically examined using panel regressions of a livelihoods regression on household assets and other socio-economic factors over time. In this paper, we characterise the livelihoods regression as a ‘livelihoods technology’, and use a latent class-technology approach to account for heterogeneity in how households generate a livelihood. We use a detailed dataset from rural India covering 213 households across 2001–2014, and control for selection issues through a Heckman Selection model. Our results are the first in the wealth dynamics literature to show that substantial heterogeneity exists in the technologies with which households generate their livelihoods. Importantly, we show that accounting for heterogeneity in household livelihoods ‘technologies’ more readily identifies different equilibria in wealth levels and provides previously foregone information on who is poor and why they remain poor.

Keywords:

• Livelihood dynamics
• asset indices
• poverty traps
• India
• latent class model

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data and analysis that support the findings of this study are openly available in the cloud.UNE data repository at https://cloud.une.edu.au/index.php/s/lJYrsgRVAlStBsO.

Notes

1 Mburu, Otterbach, Sousa-Poza, and Mude (2017) is an exception and only considers pastoralists in Northern Kenya.

2 A literature review of core publications on the topic of poverty traps and the consideration of heterogeneity within those studies is provided in the Supplementary Materials Table A1.

3 The natural log is approximated through an inverse hyperbolic transformation, as zero values in assets would be undefined using a natural log. Variables are expressed in per capita equivalence for the household (where appropriate). This follows weightings from Ryan, Bidinger, Rao, and Pushpamma (1984), assigning one adult equivalence for adults and 0.4 for children.

4 It is noted that this methodology does not allow households to move between technologies over time. The shorter cross section in the VDSA data would make it difficult to ensure reasonable power in the latent estimation. However, mean differences in time will be captured in the time dummy.

5 An ordered Probit is preferred was chosen over a multinomial Probit as there is a clear ordering of cohorts based on observed income and asset levels. A two-step procedure, compared to a FIML procedure, is more consistent for smaller sample sizes under practical applications following Chiburis and Lokshin (2007).

6 Some authors use non-parametric functions to estimate the dynamics of estimated livelihoods (for example, Kwak & Smith, 2013; Naschold, 2012). However, these are often close to linear or of a relatively simple polynomial form. Polynomial estimators benefit from having an analytical expression that allows direct comparison of trajectories behaviours, improved ability to forecast pathways, and provides greater local robustness to outliers.

7 This analysis uses only the second wave because of reliability concerns with the income and consumption data (Walker & Ryan, 1990) and only a small sample of 71 households available across both waves.

8 Previous studies of poverty dynamics in India have depended largely on cross-sectional National Sample Surveys (NSS) collected in five-year intervals (Thorat, Vanneman, Desai, & Dubey, 2017).

9 Badiani et al. (2007) shows there are significant attrition concerns between 1975 and 1984 survey wave and the second wave from 2001 onwards, but attrition rate within the waves remained low and resampling was undertaken in 2001 to restore representativeness at the village level in the second wave. This alleviates concerns around the representativeness of the sample when using only the second survey wave.

10 This poverty line is recommended for this sample in Badiani et al. (2007) and is used in subsequent studies using this data (for example, Naschold, 2012).

11 Using the standard benchmark of $1.90 USD/day in 2011 prices, 27 per cent of our sample is below the poverty line in 2001, falling to 9 per cent in 2014.

12 A summary of model statistics is provided in Table A2 in the Supplementary Materials. Full results are available from the authors.

13 A summary of model statistics is provided in Table A3 in the Supplementary Materials.

14 For example, the 0.410 coefficient for large livestock for Group 1 means that for a 1 per cent increase in large livestock value, on average Group 1 will have an increase in livelihood of 0.410 per cent, ceteris parabis.

15 As this term is the log efficiency constant, a positive constant occurs when a group has a technical efficiency constant greater than one, and negative implies a constant less than 1.

16 Statistical significance can be observed by the 95 per cent confidence bands around the convergence points for each trajectory with the 45-degree line.

17 Corresponding to a livelihood equal to one and log livelihood equal to zero.