Instrumental Variable
What is IV?
Let us consider the following equation:
Yi = α + βTi + εi
Where Yi is the outcome variable of interest, say income or expenditure, Ti is
an indicator variable (i.e. dummy variable) that takes the value of 1 if
individual/household i has participated in a credit program and 0 if not. εi is
an error term.
If T is random, the cov(T, ε) = 0. This implies no violation of one of the key
assumptions of OLS in obtaining unbiased estimates: independence of
regressors from the disturbance term ε.
T may not always be random. If not, then cov(T, ε) ≠ 0. This implies
violation of one of the key assumptions of OLS in obtaining unbiased
estimates: independence of regressors from the disturbance term ε.
What is IV?
Suppose T is not random.
The IV aims to clean up the correlation between T and ε. That is, the
variation in T that is uncorrelated with ε needs to be isolated. To do so,
one needs to find an instrumental variable, denoted Z, that satisfies the
following conditions:
1. Correlated with T: cov(Z,T ) ≠ 0
2. Uncorrelated with ε: cov(Z, ε) = 0 i.e. Z affects the outcome variable
Y only through T (Z does not have a direct influence on Y which is
referred to as the exclusion restriction)
Estimating two equations (2SLS)
First regress the treatment (T) on the instrument Z and a disturbance,
say, u. This process is known as the first-stage regression:
Ti = c + σZi + ui
Then obtain the predicted value 𝑇 and estimate the following equation:
Yi = π + β𝑇i + τi
Here cov(𝑇i , τi)=0. This is called the second stage regression. β is an
unbiased/consistent estimate of the effect of T on Y.
Generalization
Suppose our regression is as follows:
Yi = α + σXi + βTi + εi
Where X is a continuous variable, say amount of landholding. Other
variables are as defined earlier. Suppose, cov(X, ε) ≠0 and cov(T, ε) ≠0.
Suppose, our interest is only to examine the effect of T on Y. So we
need an instrument for T not for X.
Say we find Z as an instrument of T.
Generalization
The first-stage regression is then:
Ti = a + πZi + ΩXi + ui
Then obtain the predicted value𝑇 and estimate the following second
stage equation:
Yi = π + β𝑇+ cXi + τi
Concerns with IV
• Implications of Weak Instruments on Estimates
• A drawback of the IV approach is the potential difficulty in finding a good
instrument
• When the instrument is correlated with the unobserved characteristics
affecting the outcome (that is, cov(Z, ε) ≠ 0), the estimates of the effect of T
will be biased.
• Furthermore, if the instrument only weakly correlates with the treatment
variable T, the standard error of the IV estimate is likely to increase because
the predicted impact on the outcome will be measured less precisely.
Concerns with IV
Testing for Weak Instruments
• One cannot test for whether a specific instrument satisfies the exclusion
restriction; as mentioned earlier, justifications can be made only through
direct evidence of how the program and participation evolved.
• With multiple instruments, however, quantitative tests (also known as tests
of over-identifying restrictions) exist.
Using in RCT with imperfect compliance
Suppose you want to implement an interest-free credit program in
the Karail Slum of Dhaka
Eligibility criteria: Income less than Tk. 2500 per capita per month
Identify eligible households, say there are 10,000 eligible households
Randomly select 1,000 from 10,000
Randomly divide them into two groups: Group A: 500 households and
Group B: 500 households
Offer credit to group A
Group A is the treatment and Group B the control group
Using in RCT with imperfect compliance
Assume that 300 of Group A has taken the credit (imperfect
compliance)
Then assuming that randomization is balanced, using endline survey
data you can estimate the following equation (using data on 1000
households):
𝒀𝒊 = 𝜶 + 𝜷𝑻𝒂𝒔𝒔𝒊𝒈𝒆𝒅𝒊 + 𝜺𝒊
Tassigned=1 if Group A (assigned to treatment), 0 if group B
where 𝜷 is the ITT effect of credit
Using in RCT with imperfect compliance
You can estimate TOT/LATE estimate using IV
First stage regression
𝑻𝒊 = 𝜶 + 𝜷𝑻𝒂𝒔𝒔𝒊𝒈𝒏𝒆𝒅𝒊 + 𝜺𝒊 ……………………..(2)
where 𝑻𝒂𝒔𝒔𝒊𝒈𝒏𝒆𝒅𝒊 takes the value of 1 if assigned to the treatment group
and zero otherwise; 𝑻𝒊 takes the value of 1 if participated in the credit
program and zero otherwise
Estimate predicted values using equation (2) and call it 𝑻𝒉𝒂𝒕𝒊 , and run the
following second stage regression
𝒀𝒊 = 𝒂 + 𝒃𝑻𝒉𝒂𝒕𝒊 + 𝜺𝒊
Where 𝒃 is the LATE/TOT estimate