The overall goal of this project is to improve measurements of poverty dynamics, and generally measurements of economic mobility, in data-scarce environments, as well as to evaluate the impact of programs and policies to alleviate poverty.
The study of mobility in and out of poverty has increasingly captured the attention of policy makers and researchers. The types of policies needed to counteract persistent poverty may be quite different from those required to address transient poverty or movements in higher parts of the income distribution. When measuring mobility and poverty dynamics, longitudinal (panel) datasets, which follow individuals or households over time, represent the most appropriate source of information. Unfortunately, such data are often not available, or of low quality, not only in low-income countries but also in most high-income ones. These data-scarce environments are therefore common in contexts where dynamics of income and poverty would be more important to be assessed.
Approaches to the measurement of poverty transitions and economic mobility in absence of panel data have been recently proposed in the literature. They use repeated cross-sectional surveys and are commonly referred to as pseudo-panels.
The project first seeks to validate these approaches, providing robust evidence about accuracy of pseudo-panel estimates relative to benchmarks based on estimates directly derived from genuine panel data. As an alternative to the existing literature on pseudo-panels, a class of dynamic multilevel models will be proposed to track the temporal (and the spatial) dynamics of the data. Alternative ways to build pseudo-panels will be also explored following a statistical matching approach. Within this framework, it will be possible to evaluate the performance of poverty-alleviating programs. New methods for evaluating the distributional (as opposed to just average) impact of those policies will be finally investigated.
The project will complement theoretical advances with empirical evaluations, to assess welfare dynamics in the context of limited information. In this respect, the expected results of this research are highly innovative and can contribute to the state of knowledge in several fields.
First (i), our project analyses whether pseudo-panels are suitable substitutes for "true" panels for estimating economic mobility. At this regard, there is paucity of literature, which reached rather mixed conclusions (see Cruces et al. 2015; Hérault and Jenkins, 2019). We intend to provide robust evidence about accuracy of pseudo panels estimates relative to benchmarks based on estimates directly derived from genuine longitudinal survey data. Moreover, we want evaluate if the pseudo-panels have good performances not only with respect to the specific questions they want to answer, but also with respect to a broader set of mobility and poverty transition measures. Our validation exercise will not be confined to the most commonly evaluated case of just two time points.
Second (ii), it extends the dispersion-based approach using dynamic multilevel models, able to evaluate dynamic effects of both micro and macro predictors on individual poverty risk, overcoming also the problem of excluding community-specific (contextual) shocks that can potentially affect household income. To implement this model, it is necessary to fully exploit the spatial-temporal information in the repeated independent cross sectional surveys. A dynamic multilevel model explicitly accounts for the hierarchical nature of the data and for their different levels of variation: individuals (households), region-years, regions and time. Our proposal treats individual outcome (income, poverty status) as a function of individuals' characteristics and circumstances, possibly varying over time, in interaction with time-varying and time-constant features of their economic contexts. The underlying idea is that, albeit different households would have been measured at different time points, they share common shocks at macroeconomic level which can be modeled and used to leverage information on spatial and temporal trends. The complexity of these dynamic hierarchical models has prevented their use so far because of their well-known problems of convergence (Carpenter et al., 2017). Casting them in the Bayesian statistical framework offers a reasonable solution and recent developments of simulation techniques such as Markov chain Monte Carlo (MCMC) facilitates fitting these models.
Third (iii), it aims at proposing alternative methodologies to "synthetic panels" by framing the construction of such ¿panels¿ in the more general and rigorous environment of the micro-approach of the statistical matching (e.g. D'Orazio et al., 2006) and relating it to the literature on matching of longitudinal data sets. The extension to three or more waves will be presented and empirically exemplified in the project.
Synthetic panels allows to evaluate poverty reduction oriented public policies. Actually, the evaluation of such policies requires data before and after the intervention over a time horizon sufficiently long to manifest its effect. The existence of panel data, which are ideally preferred to pure cross-sections since they allow to flexibly control for unobserved individual heterogeneity, may not be lengthy enough to capture the impact of the policies, and therefore repeated cross-sections/synthetic panels could be a valid substitute.
Program evaluations mainly concern with the estimation of average treatment effects (ATE), the difference between the average level of outcomes in the presence of treatment and the average level of outcomes in its absence. However ATE miss important aspects of policy evaluation such as impacts on inequality and whether treatment harms some individuals. In the evaluation of an income policy it would be even more important to consider its effect in terms of income concentration, by comparing, e.g., Gini concentration indices for treated and untreated units. As a step forward, the comparison of Lorenz curves for treated and untreated units would be of special interest. The goal of this part of the project is to develop new estimators of the Lorenz curves for both treated and untreated units, based on preliminary results in Conti and De Giovanni (2019). The main idea is to reweight observed outcomes on the basis of appropriate estimates of propensity scores. It is possible to obtain the weak convergence of the estimated Lorenz curves (for both treated and untreated subjects), appropriately rescaled, to an appropriate Gaussian process. In order to make comparison, through appropriate statistical tests, between treated and untreated, the estimation of corresponding variance requires a special attention. Methods based on resampling proves to be promising and will be implemented in this project.