Professional Documents
Culture Documents
Section 1
Evaluation and the Synthetic Control Method
Have traditionally used the Before/After Design frequently. (e.g. See RBIs State-Wise Analysis of Fiscal Performance) Issues Easy to collect time-series data Identification rests on pre-intervention acting as a counterfactual for post-intervention Concern: Does the pre-intervention status really give a good sense of what would have happened in the absence of the intervention?
Program Works
FD/GSDP
T= Pre T= Post
Time (T)
T= Pre T= Post
Time (T)
Our best estimate of what would have happened in the absence of the intervention C Impact = T -C
T= Pre
T= Post
Time (T)
Goal is
reproduce pre-intervention outcome trends Statistical balance on key predictors across Basque and Synthetic Basque
8
Let (YitI,YitN) denote an outcome in state i in year t; I, N indicate if the state received PRMA or not. T0 is the intervention year thus, for all t in {1, , T01} and for i in {1, , N}, we want YitI = YitN. Goal is identify a set of weight for each control state such that
is minimizing: where is the set of predictors of Y for Assam and for other states and V is a diagonal matrix with specific weights for X.
9
An Example
Suppose there are three states Assam (Program state), Maharashtra and Tamil Nadu There are two predictors of the outcome variable (Y) : X1 and X2 The Minimization exercise is: Min: V1(X1A- X1MWM)2 + V2(X2A- X2TWT)2 Subject to WT,WM>0 and WT + WM = 1
But we still need to determine Vs Vs are determined by minimizing the prediction error of the outcome variable (Y) for Assam from the synthetic Assam during the pre-intervention period.
W vs. V
W gives us which states are to contribute to the synthetic control unit V ensures that the chosen synthetic unit matches Assam in it its pre-intervention outcomes. W is a function of V we have a two stage simultaneous optimization problem computationally intensive Currently exists as a package in R a statistical programming language.
11
Main Advantages
It is a data-driven process- reduces selection bias. Unlike Propensity Score Matching, does not require large data set. Ideal for Macro interventions. This method provides safeguard against extrapolation.
Section 2:
Case Study: Public Resource Management in Assam
14
15
Key Evaluation Challenge: What would have happened in the absence of the PRMPA?
16
Section 3
Findings
17
Two Outcomes
Own Tax Revenues/GSDP Pre- Intervention Indicators
Variable Log(GSDP) Non-Agriculture/GSDP Per Capita # of Factories Synthetic Assam Assam 10.594 0.661 0.057 10.587 0.663 0.059
18
19
20
21
Broad Inferences
With the three years of follow-on data Short-term gains have been achieved with respect to own tax to GSDP ratio - an increase in the order of 0.35% to 1.1% in the post-intervention period, averaging about 0.71% No evidence of short-term gains achieved with respect to a decline in interest payments to GSDP ratio Future trends will confirm however, it would be interesting to learn about the pattern and nature of debt swap Longer term horizon needed for full evaluation.
22
Thank You
23