Transcript Through the looking glass, and What OLS found there: On Growth

David Roodman (2008)
Presentation by
Faraharivony Rakotomamonjy and Estelle Zemmour
Outline
 Introduction
 Literature review
 Empirical strategy
 Specifications issues on the aid-growth literature
 Conclusion
Introduction
 Originality of this paper:
-use of non-instrumental techniques to
examine the nature of endogeneity in aidgrowth literature (sign-strength-causality)
-discussion on a number of common
specification problems
 He concludes the relationship goes to the
looking-glass : while AEL supports a positive
relationship from aid to growth, growth
appears to negatively Granger-cause aid .
Literature review (1)
 Up to the late 1990s : attempts to provide
evidences of the overall effect of aid on
growth=> no consensus
 Structural break in AEL with Burnside and
Dollar (2000): analyze the conditionality of
aid effectiveness
Literature review (2)
 This coincides with advances in econometrics:
- from cross-section to panel: Hansen and Tarp
(2001)
- from OLS to 2SLS and GMM: Michaelowa and
Weber (2006); Mishra and Newhouse (2007)…
- from a single linear aid regressor to interaction
terms and aid subcomponents: Burnside, Dollar
(2000); Dalgaard, Hansen, Tarp(2004); Neanidis
and Varvarigos (2007)…
Empirical strategy (1)
 Adapt Hansen and Tarp unconditional aid-growth
litterature(2001) with 2SLS.
 Regress per-capita economic growth on aid/GDP and
(aid/GDP)² , panel:1974-1993
 Same controls as Burnside and Dollar(2000) but
introduce different instruments: among which one
period- lagged value of aid/GDP.
Empirical strategy (2)
 Before giving a unified theory, take and increase BLZ
framework with some tests of causal relationship
between aid and growth.
 Expand table 1, larger time-data than HT study: 19622001 (4-year period) replacing HT controls. In
particular, he drops the quadratic aid term.
Issues of the aid-growth literature
(1) Autocorrelation (of the error terms)
(2) Instrument proliferation
(3) Multicollinearity
Context
 Ideally: instrumentation corrects endogeneity
 In practice: estimates on aid’s impact present
autocorrelation in the errors, there is proliferation of
instruments, and multicollinearity
 These problems bring pessimism on the ability of
demonstrating aid effectiveness with cross-country
econometrics, thus suggesting that the average effect
of aid on growth is too small to be detected statistically
(1) Autocorrelation
 Pooled OLS and 2SLS regression in the aid and growth
literature face serial correlation in the errors : if lagged
aid is endogenous to lagged growth and the lagged
growth innovation is correlated with contemporary
growth innovation, then lagged aid can be correlated
with it too
=> possibility of endogeneity bias
=> lagged variables are not valid instruments
(1) Autocorrelation
 Example:
Dropping the lags of aid and aid^2 from the
instrument sets in the 2SLS regressions in HT and
Clemens, Radelet, and Bhavnani (2004) eliminates the
significance of the coefficients on the aid terms
=> this suggests that identification depends on these
instruments
(2) Instrument proliferation
 Diff-in-Diff and GMM estimation, that
dominate this literature since the early 2000s,
led to instrument proliferation
However, the assumptions necessary for the
validity of the instruments (Blundell & Bond,
1998) are non-trivial while rarely checked (with
difference-in-Hansen tests) to protect the
power of the tests
=> risk of overfitting of endogenous variables
(2) Instrument proliferation
 Example:
- in Michaelowa and Weber (2006), the significance of
aid terms appears to go hand-in-hand with the
instrument count
- in Mishra and Newhouse (2007), the coefficient on
the lagged dependent variable is 1.0, which invalidates
the GMM instruments (Blundell and Bond 1998)
(3) Multicollinearity
 Adding a nearly collinear term to the regression of
growth on aid/GDP allows a much better fit by
inflating the t stat
=> 2 collinear aid terms have an inherent propensity to
generate seemingly strong results, thus implifying
endogeneity bias
(3) Multicollinearity
 Example:
- Burnside and Dollar (2000) include both
(aid/GDP)*policy and (aid/GDP)^2 *policy in their
OLS regression, thus providing huge t stat.
But, if one eliminates the collinearity by dropping
(aid/GDP)^2 *policy, the large t stat disappear
Conclusion (1)
 The main aid-growth relationship is causal and
negative from growth to aid
 AEL fails in finding significant effect of aid on growth
which are robust and free of methodological problems
 Econometric sophistication has clouded rather than
sharpened AEL
Conclusion (2)
 However this does not end the quest of evidence on aid
effectiveness but shifts it to smaller questions such as
Chen, Mu, and Ravallion (2006) which study how
much the placement of WB-financed rural
development projects in China can explain subnational variation in household income ten years later