This notebook introduces instrumental variable analysis. We look the conditions that must be satisfied to perform an IV analysis, how the two-stage-least-squares approach works, and how to interpret the results.
Author
Marina Adshade, Paul Corcuera, Giulia Lo Forte, Jane Platt
Published
29 May 2024
Prerequisites
Run OLS regressions.
Learning Outcomes
Understand what an instrumental variable is and the conditions that must be satisfied to address the endogeneity problem.
Implement a Two Stage Least Squares (2SLS) regression-based approach using an instrument.
Describe the weak instrument problem.
Interpret the first stage test of whether or not the instrument is weak.
17.1 The Linear Instrumental Variable Model
Consider a case where we want to know the effect of education on earnings. We may want to estimate a model like the following:
\[
Y_{i} = \alpha + \beta X_i + \epsilon_i,
\]
where \(Y_i\) is earnings of individual \(i\) and \(X_i\) is years of education of individual \(i\).
A possible issue with this model comes from omitted variable bias: it is possible that the decision to attend school is influenced by other individual characteristics that are also correlated with earnings. For example, think of individuals with high innate ability. They may want to enroll in school for longer and obtain higher-level degrees. Moreover, their employers may compensate them for their high ability, regardless of their years of schooling.
Instrumental variables (IVs) can help us when there are hidden factors affecting both the treatment (in our case, years of education) and the outcome (in our case, earnings). The instrumental variable approach relies on finding something that affects the treatment and affects the outcome, but that affects the outcome solely through the treatment. In short, the instrument must satisfy two assumptions:
Relevance: the instrument should be correlated with the explanatory variable; in our case, it should be correlated with the years of education \(X_i\);
Exclusion restriction: the instrument should be correlated with the dependent variable only through the explanatory variable; in our case, it should be correlated with \(Y_i\) only through its correlation with \(X_i\).
Let’s say we have found an instrumental variable \(Z_i\) for the variable \(X_i\). Then, using an IV analyis implies estimating the following model: \[
\begin{align}
Y_i &= \alpha_1 + \beta X_i + u_i \quad \text{(Structural Equation)}\\
X_i &= \alpha_2 + \gamma Z_i + e_i \quad \text{(First Stage Equation)}
\end{align}
\]
where the two conditions we have seen above imply that:
\(\gamma \neq 0\);
\(Z_i\) is uncorrelated with \(u_i\).
In practice, using an IV analysis often implies using a Two-Stages Least Square (2SLS) estimator. The two steps of 2SLS are:
Estimate the first stage equation by OLS and obtain the predicted value of \(X_i\). In this way, we have effectively split \(X_i\) into \[
X_i = \underbrace{\hat{X}_i}_\text{exogenous part} + \underbrace{\hat{e}_i}_\text{endogenous part}
\]
where \(\hat{X_i} \equiv \hat{\alpha_2} + \hat{\gamma} Z_i\).
Plug \(\hat{X_i}\) instead of \(X_i\) into the structural equation and estimate via OLS. We are then using the “exogenous” part of \(X_i\) to capture \(\beta\).
Warning: We can run 2SLS following the steps above, but when we want to do inference we need to be sure we’re using the true residuals in the structural equation \(\hat{u}_i\). The built-in Stata commands ivregress and ivreg2 automatically give us the right residuals.
Let’s see how to estimate this in Stata. Once again, we can use our fictional data set simulating wages of workers in the years 1982-2012 in a fictional country.
clear* *cd ""use fake_data, cleardescribe, detail
Contains data from fake_data.dta
Observations: 138,138
Variables: 11 16 Jul 2023 17:25
Width: 28
-------------------------------------------------------------------------------
Variable Storage Display Value
name type format label Variable label
-------------------------------------------------------------------------------
workerid long %12.0g Worker Identifier
year int %8.0g Calendar Year
sex str1 %9s Sex
age byte %9.0g Age (years)
start_year int %9.0g Initial year worker is observed
region byte %9.0g group(prov)
treated byte %8.0g Treatment Dummy
earnings float %9.0g Earnings
sample_weight float %9.0g
quarter_birth float %9.0g Quarter of birth
schooling float %9.0g Years of schooling
-------------------------------------------------------------------------------
Sorted by: workerid
In Stata, we can perform IV analysis with a 2SLS estimator by using one of the following two commands: ivregress or ivreg2. They have a similar syntax:
where instead of <Y>, <X>, and <Z>, we write the names of the corresponding dependent, independent, and instrument variables of our model.
We now have to choose an IV that can work in our setting. A well-known example for an instrument for years of schooling is studied by Angrist and Krueger (1991): they propose using \(Z\), the quarter of birth. The premise behind their IV is that students are required to enter school in the year they turn 6 but not necessarily when they are already 6 years old, creating a relationship between quarter of birth and schooling. At the same time, the time of the year one is born shouldn’t affect one’s earnings aside from its effect on schooling.
Let’s see how to estimate a simple IV in Stata using our data and each one of the commands ivregress and ivreg2.
IV (2SLS) estimation
--------------------
Estimates efficient for homoskedasticity only
Statistics consistent for homoskedasticity only
Number of obs = 138138
F( 1,138136) = 0.03
Prob > F = 0.8691
Total (centered) SS = 8.82816e+15 Centered R2 = -36.2867
Total (uncentered) SS = 9.80603e+15 Uncentered R2 = -32.5684
Residual SS = 3.29173e+17 Root MSE = 1.5e+06
------------------------------------------------------------------------------
earnings | Coefficient Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
schooling | 714972.8 4339032 0.16 0.869 -7789374 9219319
_cons | -1.09e+07 6.68e+07 -0.16 0.870 -1.42e+08 1.20e+08
------------------------------------------------------------------------------
Underidentification test (Anderson canon. corr. LM statistic): 0.026
Chi-sq(1) P-val = 0.8719
------------------------------------------------------------------------------
Weak identification test (Cragg-Donald Wald F statistic): 0.026
Stock-Yogo weak ID test critical values: 10% maximal IV size 16.38
15% maximal IV size 8.96
20% maximal IV size 6.66
25% maximal IV size 5.53
Source: Stock-Yogo (2005). Reproduced by permission.
------------------------------------------------------------------------------
Sargan statistic (overidentification test of all instruments): 0.000
(equation exactly identified)
------------------------------------------------------------------------------
Instrumented: schooling
Excluded instruments: quarter_birth
------------------------------------------------------------------------------
Both Stata functions give us a standard output: the values of the coefficients, standard errors, p-values, and 95% confidence intervals. From the regression output, years of schooling does not seem to have any effect on earnings. However, before trusting these results, we should check that the two IV assumptions are met in this case.
Notice that ivreg2 gives us more details about tests we can perform to assess whether our instrument is valid. We will talk more about these tests, especially the weak identification test, in the paragraphs below.
17.2 Weak Instrument Test
While we cannot really test for the exclusion restriction, we can check whether our instrument is relevant. We do that by looking directly at the coefficients in the first stage.
In Stata, we only need to add the option first to get an explicit output for the first stage.
ivregress 2sls earnings (schooling = quarter_birth), first
ivreg2 earnings (schooling = quarter_birth), first
First-stage regressions
-----------------------
First-stage regression of schooling:
Statistics consistent for homoskedasticity only
Number of obs = 138138
------------------------------------------------------------------------------
schooling | Coefficient Std. err. t P>|t| [95% conf. interval]
-------------+----------------------------------------------------------------
quarter_bi~h | .0009984 .0061896 0.16 0.872 -.0111332 .01313
_cons | 15.3971 .0165699 929.22 0.000 15.36463 15.42958
------------------------------------------------------------------------------
F test of excluded instruments:
F( 1,138136) = 0.03
Prob > F = 0.8719
Sanderson-Windmeijer multivariate F test of excluded instruments:
F( 1,138136) = 0.03
Prob > F = 0.8719
Summary results for first-stage regressions
-------------------------------------------
(Underid) (Weak id)
Variable | F( 1,138136) P-val | SW Chi-sq( 1) P-val | SW F( 1,138136)
schooling | 0.03 0.8719 | 0.03 0.8719 | 0.03
Stock-Yogo weak ID F test critical values for single endogenous regressor:
10% maximal IV size 16.38
15% maximal IV size 8.96
20% maximal IV size 6.66
25% maximal IV size 5.53
Source: Stock-Yogo (2005). Reproduced by permission.
NB: Critical values are for Sanderson-Windmeijer F statistic.
Underidentification test
Ho: matrix of reduced form coefficients has rank=K1-1 (underidentified)
Ha: matrix has rank=K1 (identified)
Anderson canon. corr. LM statistic Chi-sq(1)=0.03 P-val=0.8719
Weak identification test
Ho: equation is weakly identified
Cragg-Donald Wald F statistic 0.03
Stock-Yogo weak ID test critical values for K1=1 and L1=1:
10% maximal IV size 16.38
15% maximal IV size 8.96
20% maximal IV size 6.66
25% maximal IV size 5.53
Source: Stock-Yogo (2005). Reproduced by permission.
Weak-instrument-robust inference
Tests of joint significance of endogenous regressors B1 in main equation
Ho: B1=0 and orthogonality conditions are valid
Anderson-Rubin Wald test F(1,138136)= 1.01 P-val=0.3143
Anderson-Rubin Wald test Chi-sq(1)= 1.01 P-val=0.3143
Stock-Wright LM S statistic Chi-sq(1)= 1.01 P-val=0.3143
Number of observations N = 138138
Number of regressors K = 2
Number of endogenous regressors K1 = 1
Number of instruments L = 2
Number of excluded instruments L1 = 1
IV (2SLS) estimation
--------------------
Estimates efficient for homoskedasticity only
Statistics consistent for homoskedasticity only
Number of obs = 138138
F( 1,138136) = 0.03
Prob > F = 0.8691
Total (centered) SS = 8.82816e+15 Centered R2 = -36.2867
Total (uncentered) SS = 9.80603e+15 Uncentered R2 = -32.5684
Residual SS = 3.29173e+17 Root MSE = 1.5e+06
------------------------------------------------------------------------------
earnings | Coefficient Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
schooling | 714972.8 4339032 0.16 0.869 -7789374 9219319
_cons | -1.09e+07 6.68e+07 -0.16 0.870 -1.42e+08 1.20e+08
------------------------------------------------------------------------------
Underidentification test (Anderson canon. corr. LM statistic): 0.026
Chi-sq(1) P-val = 0.8719
------------------------------------------------------------------------------
Weak identification test (Cragg-Donald Wald F statistic): 0.026
Stock-Yogo weak ID test critical values: 10% maximal IV size 16.38
15% maximal IV size 8.96
20% maximal IV size 6.66
25% maximal IV size 5.53
Source: Stock-Yogo (2005). Reproduced by permission.
------------------------------------------------------------------------------
Sargan statistic (overidentification test of all instruments): 0.000
(equation exactly identified)
------------------------------------------------------------------------------
Instrumented: schooling
Excluded instruments: quarter_birth
------------------------------------------------------------------------------
From both methods, we can see that the IV we have chosen is not relevant for our explanatory variable \(X\): quarter_birth is not correlated with schooling. Another indicator of the lack of relevance is given by the F-statistic reported by Stata in the “Weak Identification test” row: as a rule of thumb, every time its value is less than 10, the instrument is not relevant.
Whenever the correlation between \(X\) and \(Z\) is very close to zero (as in our case), we say we have a weak instrument problem. In practice, this problem will result in severe finite-sample bias and large variance in our estimates. Since our instrument is not valid, we cannot trust the results we have obtained.
17.3 Wrap Up
In this module, we studied the linear IV model and how to estimate it using the 2SLS Method using ivregress or ivreg2. We learned that we can overcome the endogeneity problem when we have access to a different type of variable: an instrumental variable. A good instrument must satisfy two important conditions:
It must be uncorrelated with the error term (also referred to as the exclusion restriction).
It must be correlated, after controlling for observables, with the variable of interest (there must be a first stage).
While the second condition can be checked using the regression results of the first stage, the first condition is inherently not testable. Therefore, any project that uses IVs must include a discussion, using contextual knowledge, of why the first condition may hold.
Finally, do not forget that for every endogenous variable in our regression, we require at least one instrument. For example, if we have a regression with two endogenous variables, we require at least two IVs!
17.4 Wrap-up Table
Command
Function
ivregress 2sls
It performs Instrumental Variable analysis using a Two-Stage Least Squares estimator.
ivreg2
It performs Instrumental Variable analysis using a Two-Stage Least Squares estimator by default.
, first
This option shows the results for the First Stage regression in the IV analysis.