IV and Some Basic Panel Model

Dr. Muhammad Shahadat Hossain Siddiquee

Professor, Department of Economics
University of Dhaka
Email: [email protected]
Cell: +8801719397749
IV
• Instrumental variable (IV) methods allow for endogeneity in individual participation,
program placement, or both.
• With panel data, IV methods can allow for time-varying selection bias.
• Measurement error that results in attenuation bias can also be resolved through this
procedure. Attenuation bias, also known as regression dilution or errors-in-variables
bias, is a form of bias in econometrics that occurs when measurement error in an
independent variable (X) leads to an underestimation of the true relationship between
that variable and the dependent variable (Y). Essentially, the estimated regression
coefficient is biased towards zero, meaning the effect of the independent variable on
the dependent variable appears smaller than it truly is.
• The IV approach involves finding a variable (or instrument) that is highly correlated
with program placement or participation but that is not correlated with unobserved
characteristics affecting outcomes.
• Instruments can be constructed from program design (for example, if the program of
interest was randomized or if exogenous rules were used in determining eligibility for
the program).
IV…
• Instruments should be selected carefully.
• Weak instruments can potentially worsen the bias even more than when
estimated by ordinary least squares (OLS) if those instruments are
correlated with unobserved characteristics or omitted variables affecting the
outcome.
• Testing for weak instruments can help avoid this problem.
• Another problem can arise if the instrument still correlates with
unobserved anticipated gains from the program that affect participation;
local average treatment effects (LATEs) based on the instruments can help
address this issue.
Instrumental Variable Estimation
• We turn to methods that relax the exogeneity assumption of OLS or PSM and
that are also robust to time-varying selection bias, unlike DD.
• For DD methods one cannot control for selection bias that changes over time.
• By relaxing the exogeneity assumption, the IV method makes different identifying
assumptions from the previous methods—although assumptions underlying IV
may not apply in all contexts.
• Consider the following equation:
Instrumental Variable Estimation (Contd..)
• If treatment assignment T is random in above equation, selection bias is not a problem at
the level of randomization.
• However, treatment assignment may not be random because of two broad factors.
• First, endogeneity may exist in program targeting or placement—that is, programs are
placed deliberately in areas that have specific characteristics (such as earnings
opportunities or social norms) that may or may not be observed and that are also correlated
with outcomes Y.
• Second, unobserved individual heterogeneity stemming from individual beneficiaries’ self-
selection into the program also confounds an experimental setup. In this case, selection
bias may result from both of these factors because unobserved characteristics in the error
term will contain variables that also correlate with the treatment dummy T.
• That is, cov(T, ε) ≠ 0 implies violation of one of the key assumptions of OLS in obtaining
unbiased estimates: independence of regressors from the disturbance term ε.
• The correlation between T and e naturally biases the other estimates in the equation,
including the estimate of the program effect β.
Panel-data Methods Overview (Contd..)
• Regression coefficient identification for some estimators can depend on
regressor type. Some regressors, such as gender, may be time invariant with xit =
xi for all t. Some regressors, such as an overall time trend, may be individual
invariant with xit = xt for all i. And some may vary over both time and individuals.
• Some or all model coefficients may vary across individuals or over time.
• The microeconometrics literature emphasizes the fixed-effects model. This
model permits regressors to be endogenous provided that they are correlated
only with a time-invariant component of the error.
• Most other branches of applied statistics instead emphasize the random-effects
model that assumes that regressors are completely exogenous.
• Finally, panel data permit estimation of dynamic models where lagged
dependent variables may be regressors. Most panel-data analyses use models
without this complication.
Some basic panel models
• There are several different linear models for panel data.
• The fundamental distinction lies in the fact of choosing fixed-effects and
random-effects models.
• The term ‘fixed effects’ is misleading because in both types of models
individual-level effects are random.
• Fixed-effects models have the added complication that regressors may be
correlated with the individual-level effects so that consistent estimation of
regression parameters requires eliminating or controlling for the fixed effects.
Individual-effects Model
• The individual-specific-effects model for the scalar dependent variable Yit
specifies that

• where Xit are regressors, αi are random individual-specific effects, and εit is an
idiosyncratic error. Two quite different models for the αi, are the fixed-effects
and random-effects models.
Fixed-effects Model (FEM)
• In the fixed-effects (FE) model, αi are permitted to be correlated with the regressors
Xit. This allows a limited form of endogeneity.
• We view the error uit = αi + εit and permit Xit to be correlated with the time-invariant
component of the error (αi), while continuing to assume that Xit is uncorrelated with
the idiosyncratic error εit·
• For example, we assume that if regressors in an earnings regression are correlated with
unobserved ability, they are correlated only with the time-invariant component of
ability, captured by αi .
• One possible estimation method is to jointly estimate α1 , ..., αN and β. But for a short
panel, asymptotic theory relies on N   , and here as N   so too does the
number of fixed effects to estimate. This problem is called the incidental-parameters
problem. Interest lies in estimating β, but first we need to control for the nuisance or
incidental parameters, αi;.
• Instead, it is still possible to consistently estimate β for time-varying regressors, by
appropriate differencing transformations applied to the regression equation mentioned
above that eliminate αi.