Module 5 Inference with Panel Data Measuring poverty Multidimensional poverty Poverty Dynamics Panel Data Inference with Panel Data International Poverty Comparisons Vulnerability Tackling Poverty Module 5 Inference with Panel Data Jonathan Haughton jhaughton@Suffolk.edu June 2017
JH: Poverty Measurement Course Objectives Explain the essential ideas and vocabulary of regression analysis. List and explain the main regression problems, including Measurement error Omitted variable bias Simultaneity bias Sample selectivity bias Multicolinearity Heteroskedasticity Outliers Show how panel data can help address some of these problems Outline the key ideas underlying impact evaluation Explain how impact evaluation can be more powerful with panel data, and illustrate this with an example from Thailand June 2017 JH: Poverty Measurement Course
JH: Poverty Measurement Course Regression 1 Fit a line to data As here, Vietnam 1998 Engel curves Linear Food spending/cap = 1.188 + 0.22 (spending/capita) Quadratic F/cap = 0.853 + 0.30 (spend/cap) – 0.0021 (spend/cap)^2 H&H chap 2, Fig 1.9 June 2017 JH: Poverty Measurement Course
JH: Poverty Measurement Course Vocabulary Intercept/constant Coefficients Dependent variable Independent variables/regressors/covariates Error (unobserved) Mean zero; identically & independently distributed True equation; estimated equation Residual (observed): June 2017 JH: Poverty Measurement Course
JH: Poverty Measurement Course Example R2 = 0.47. Goodness of fit. Coefficient signs as expected Consumption/capita: in logs. t-statistics shown. [>2 ≈ “significant at 5%”] Phnom Penh variable is binary Dependency ratio: (young + old)/(prime age) Femaleness: % aged 15-60 who are female. Positive coefficient June 2017 JH: Poverty Measurement Course
JH: Poverty Measurement Course Regression problems: Measurement error: y = a + bX + ε In Y: no bias, but poor fit In X: estimate of b biased toward 0 Omitted variable bias True: Child health = a + b MumEd + c MumAbility Estimated: Child health = a + b MumEd If ed, ability, correlated, estimate of b is too high June 2017 JH: Poverty Measurement Course
More regression problems Simultaneity bias E.g. Child health = a + b Nutrient intake But which comes first? Feed the weak or the healthy? Sample selectivity bias E.g. Only observe wages for those who work Solution: Two steps (Heckman) Multicolinearity Some X variables are intercorrelated Hard to get accurate coefficient estimates June 2017 JH: Poverty Measurement Course
Yet more regression problems Heteroskedasticity Errors do not have constant variance Coefficients not biased; but standard errors inefficient June 2017 JH: Poverty Measurement Course
Regression Problems Again Outliers Typo or truth? June 2017 JH: Poverty Measurement Course
Panel data can be helpful Rice output = a + b fertilizer + c ability + ε Ability unobservable, so estimate Rice output = a + b fertilizer + w But estimate of b is likely overstated. Use panel data: Differencing washes out “ability”, allows an unbiased estimate of b Very useful, but only works if ability is time-invariant. June 2017 JH: Poverty Measurement Course
JH: Poverty Measurement Course Impact Evaluation Purpose: quantify the impact of a project or policy E.g. Does microcredit raise incomes Complex; requires construction of a counterfactual Gold standard: Randomized Controlled Trial Randomly select beneficiaries (and controls) Collect baseline data After intervention, collect data again; compare outcomes between treated and controls June 2017 JH: Poverty Measurement Course
Powerful: Double Differences Panel data with pre- and post-treatment E.g. (30 – 10) – (21 – 14) = 13 June 2017 JH: Poverty Measurement Course
Can combine with regression Error: time-invariant and innovation components Sweeps away effects of variables, including unobservables, that do not vary over time Jalan and Ravallion (1998): Poverty alleviation program in China Double differencing biased; keep variables for initial conditions that influence program placement June 2017 JH: Poverty Measurement Course
Example: Thailand Village Fund Village-run microcredit; any impact on expenditure or income per capita? Rural panel, Socio-Economic Surveys, 2002 and 2004 T = treatment = borrow from Village Fund X variables include household variables (e.g. age of head, education of head, no. of adults). Y is log of expenditure (or income) per capita Estimates of γ: Expenditure: 3.5% (s.e. 1.5%; p-value 0.02) Income: 1.4% (s.e. 1.8%; p-value 0.44) Boonperm et al. 2013 June 2017 JH: Poverty Measurement Course
JH: Poverty Measurement Course Extensions There may be time-varying unobserved heterogeneity, biasing the coefficient on T E.g. Local conditions may change over time Solution 1: Interact treatment with time-varying variables Khandker (2006): Microcredit in Bangladesh Solution 2: Arellano-Bond dynamic lagged-dependent-variable approach See Jalan and Ravallion (1998), mentioned above June 2017 JH: Poverty Measurement Course
JH: Poverty Measurement Course Reading Haughton & Khandker (2009). Handbook on Poverty and Inequality. Chapter 14; 13 (pp. 256-270). World Bank, Washington DC. Khandker, Koolwal, & Samad (2010) Handbook on Impact Evaluation, World Bank, Washington DC. Haughton & Haughton (2011). Living Standards Analytics, Springer, New York. Chapters 2 (Regression) and 12 (Impact Evaluation) Boonperm et al. (2013) Does the Village Fund matter in Thailand? Evaluating the impact on incomes and spending. Journal of Asian Economics 25: 3-16. Cameron & Trivedi (2009) Colin Cameron & Pravin Trivedi. Microeconometrics Using Stata. Stata Press. June 2017 JH: Poverty Measurement Course