Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
110,676 characters · 9 sections · 47 citation commands
Testing the Exogeneity of Instrumental Variables and Regressors in Linear Regression Models Using Copulas
Suppose that we want to estimate the following linear regression model: $ Y_t=\beta_0+\beta^T X_t+\alpha P_t+\epsilon_t,~t=\{1,...,T\},$ where $X_t$ is the $k \times 1$ vector of exogenous regressors that are uncorrelated with the error term $\epsilon_t$, while $P_t$ is a scalar variable that is endogenous; i.e., it is potentially correlated with the error term. We are particularly interested in estimating the coefficient of the endogenous variable $\alpha$. Using instrumental variables is a popular approach to address this endogeneity issue (angrist1996identification). Suppose that we have access to $m \times 1$ vector of instrumental variables denoted by $Z_t$, where $Z_t$ does not include any variable in $X_t$. There are two fundamental conditions in the approach of instrumental variables: first, the instruments $Z_t$ should be correlated with the exogenous variable (relevance), and second, the instruments $Z_t$ should not be correlated with the error term (exogeneity or validity). Evaluation of the relevance condition is relatively straightforward, but testing the instrument exogeneity condition is not (wooldridge2010econometric). Usually, researchers justify this condition using economic-theoretical arguments. However, in many situations, these arguments could be considered subjective beliefs that could not be supported by the data. On the other hand, it has been shown in the existing literature (kiviet2020testing, dufour2003identification and bound1995problems) that a violation of this non-testable condition can not only lead to loss of efficiency, but can also result in even a higher estimation bias in comparison to the regular OLS.
To alleviate these concerns, there have been two streams of literature. The first stream does not address testing the instrument exogeneity condition. It explores the effect of violation of this condition on the estimate of interest and the resulting inference (ashley2009assessing, kraay2012instrumental and nevo2012identification). In other words, these works assess the robustness of different inferential conclusions to uncertainty in the validity of instruments. The second stream attempts to test the exogeneity of instruments with some caveats. sargan1958estimation and hansen1982large study formal testing of over-identification restrictions contingent on the validity of the initial just-identifying set of instruments. In other words, these works' methods rely on a non-testable assumption that the initial set of just-identifying instruments satisfies the exogeneity condition. As a result, as illustrated in deaton2010instruments and parente2012cautionary, these overidentification tests cannot reliably test the exogeneity for all instruments.
Taking a different direction, a more recent stream of work evaluates the instrument exogeneity condition using an indirect approach (kiviet2020testing and kripfganz2021kinkyreg). In kiviet2020testing, the authors adopt nonorthogonal moment conditions in regression by developing an instrument-free approach to estimate the parameters. The authors assume that the correlation between the endogenous variable and the error term is known. Then, they show that the instrument exogeneity condition is satisfied if the correlation value between the endogenous variable and the error term, which rejects the instrument exogeneity condition, is not in an admissible range. However, the issue is that the admissible range for the correlation coefficient of the endogenous variable and the error term is unknown from the data. Therefore, a researcher must rely on expert knowledge of the admissible degree of endogeneity. Relying on expert knowledge induces subjectivity to the approach in kiviet2020testing. In kitagawa2015test, the author tests a necessary condition for the exogeneity of the instruments based on ordering the conditional joint distribution of the outcome and the treatment depending on the instrument. However, the method in kitagawa2015test applies only to cases with binary treatments and discrete instruments, which limits its applicability.
Despite the lack of reliable methods to test instrument exogeneity in the existing literature, the issue remains critical due to the vast number of papers that leverage the instrumental variable approach. To address this issue, we propose a Copula-based method to test the exogeneity of instruments from the error term. Copulas have been used as an instrument-free approach to address endogeneity issues in the estimation of linear regression models (park2012handling and yang2022addressing). However, our work is the first to use this approach to test the exogeneity of instruments from the error term.
To model the endogeneity of $P_t$, we consider the reduced-form equation for it given by $ P_t=\delta_0+\delta^T X_t+\gamma^T Z_{t}+\eta_t$. Given that $X_t$ is exogenous, the endogeneity of $P_t$ implies the potential endogeneity of $Z_t$ and the error term $\eta_t$. Therefore, we model the endogeneity of $P_t$ by characterizing the joint distribution of the error term in the original regression equation $\epsilon_t$, $Z_t$ and $\eta_t$ by a Gaussian copula. The Gaussian copula is a multivariate normal distribution on the normal transformations of variables. The normal transformation of a variable in $\epsilon_t$, $Z_t$, and $\eta_t$ is a variable from the standard normal distribution with the same CDF value.
Next, using the Gaussian copula as the joint distribution of $\epsilon_t$, $Z_t$ and $\eta_t$, we decompose the error term $\epsilon_t$ into two components. The first is a linear combination of the standard normal transformations of $Z_t$ and $\eta_t$, to capture the endogeneity of $P_t$. The second component is a normal, independent error term essential for the consistent estimation of the regression model. We add the normal transformations of $Z_t$ and $\eta_t$ as added regressors to the original regression equation and estimate the model. We establish that the test for the exogeneity of the instruments is equivalent to a Wald test, where a linear transformation for the coefficients of the standard normal transformations of the instruments $Z_t$ should be equal to zero.
Next, by simplifying the reduced-form equation for $P_t$ and setting $\eta_t=P_t$, we establish an instrument-free approach to test for the exogeneity of a regressor. This is a valuable approach given that in the existing literature, the regressor endogeneity test boils down to comparing the coefficient of the regressor under OLS and TSLS or testing if the coefficient of the error term from the first stage is not statistically significant in the second stage (durbin1954errors, hausman1978specification, wu1973alternative, wooldridge2010econometric). However, both of these methods rely on having an exogenous instrument from the error term. In other words, the traditional approach to test for the endogeneity of a regressor is not helpful if the researcher does not have access to an instrument. In addition, this approach can lead to a wrong conclusion if the instrument used in the test is actually not exogenous to the error term. In particular, if the instruments used in the Hausman endogeneity test are not exogenous to the error term, this test will reject the exogeneity hypothesis even if the regressor is actually exogenous. In a recent work, caetano2015test develops an instrument-free approach to test the endogeneity of regressors in models with bunching. However, their method is based on two assumptions: 1) the outcome variable is continuous in the endogenous regressor, and 2) there is at least one unobservable confounder that is discontinuously distributed with respect to the endogenous variable. These two assumptions may not hold in various situations. Hence, the applicability of the method in caetano2015test could be limited.
Through a series of simulation studies, we evaluated our test's efficacy, yielding the following key findings:
A shortcoming of our instrument exogeneity test is that if an instrument is normally distributed and highly correlated with the endogenous variable $P_t$, its normal transformation will be highly correlated with $P_t$ creating the multicollinearity issue. Moreover, similar to park2012handling, if the endogenous variable $P_t$ has a normal distribution, its normal transformation will be $P_t$ divided by its standard deviation, creating the multicollinearity issue in the regressor endogeneity test. In these cases, collecting a data set with a higher number of observations can mitigate the efficiency loss caused by multicollinearity.
Furthermore, we apply our test empirically in a Two-Stage Least Squares (TSLS) setting in Angrist's study on compulsory education, angristcompulsory. Our test successfully validates the endogeneity of the education variable. Additionally, it confirms the exogeneity of the instruments used in this study, which include interactions between the year of birth and the quarter of birth. We also show that after adding the demographic variables as regressors the endogeneity of the education variable disappears. This empirical application underscores the practical effectiveness of our test in real-world TSLS scenarios.
To the best of our knowledge, our work is the first one that provides a rigorous approach to test instruments and regressor exogeneity with minimal assumptions and straightforward implementation. Our method holds significant promise for practitioners who previously had to depend on untestable arguments to substantiate their methodologies and findings. We are confident that our work will greatly improve the rigor and reliability in the field of IV regression, offering substantial benefits to those in applied econometrics and related disciplines.
We aim to estimate the following linear regression model:
where $X_t$ represents a $k\times 1$ vector of exogenous regressors and $P_t$ is a scalar endogenous regressor. The term $\epsilon_t$ is the error component of the model. Here, $t$ indexes either time or cross-sectional units. We have at our disposal $m$ instrumental variables excluded from the original regression model to address the endogeneity of $P_t$. Denote by $Z_{t}$ the $(m\times 1)$ vector of instrumental variables, where $Z_{it}$ is the $i$th instrument. It is hypothesized that these instrumental variables may exhibit correlation with the error term $\epsilon_t$. Furthermore, for any variable $\kappa_t$, we denote its marginal CDF, its mean, and its standard deviation by $F_{\kappa}(\cdot)$, $\mu_{\kappa}$ and $\sigma_{\kappa}$, respectively.
Consider the reduced-form equation for the endogenous variable expressed as follows:
Given the correlation between $P_t$ and the error term in the main regression equation, and assuming the exogeneity of $X_t$, the potential correlation of $Z_t$ and $\eta_t$ with the error term $\epsilon_t$ becomes the focal point of consideration. We operate under the assumption that $Z_t$ and $\eta_t$ (the error term in the reduced-form equation) are uncorrelated.
According to Sklar's Theorem (sklar1959fonctions), any multivariate joint distribution can be expressed by a copula function alongside the marginal distributions of the involved variables. Based on this theorem, we model the endogeneity of $Z_t$ and $\eta_t$ by assuming that a Gaussian copula can represent the joint distribution of $Z_t$, $\eta_t$ and $\epsilon_t$.
Suppose that $H(Z_t, \eta_t,\epsilon_t)$ denotes the joint Cumulative Distribution Function (CDF) of the instrumental variables, the error term in the reduced-form equation for the endogenous variable, and the error term in the original regression equation. We can express this relationship as:
where $\Psi$ represents the CDF of a standard multivariate normal distribution. The variables $Z_t^* = [Z_{it}^*]_{i=1, \ldots, m}$, $\eta_t^*$, and $\epsilon_t^*$ are the standard normal transformations of $Z_t$, $\eta_t$, and $\epsilon_t$, respectively. For any variable $\kappa_t \in \{Z_t, \eta_t, \epsilon_t\}$, its standard normal transformation $\kappa_t^*$ is defined as a standard normal random variable corresponding to the same CDF value. The standard normal transformation of a variable $\kappa_t$ follows a standard normal distribution. The normal transformation $\kappa_t^*$ is defined as follows:
We represent the Pearson correlation coefficient between two variables, $\kappa_t$ and $\chi_t$, from the set $\{Z_t, \eta_t, \epsilon_t\}$ as $\rho_{\kappa\chi}$. Similarly, the Pearson correlation between their corresponding standard normal transformations, denoted by $\kappa_t^*$ and $\chi_t^*$, is expressed as $\rho_{\kappa^*\chi*}$.
We assume that the error term \(\epsilon_t\) has a normal distribution, which is a widely used assumption in the literature (kleibergen2003bayesian, ebbes2005solving, rossi2003bayesian, park2012handling and yang2022addressing). This assumption leads to the following relationships: \(F_{\epsilon}(\epsilon_t) = \Phi(\sigma_{\epsilon} \epsilon_t^*)\) and \(\epsilon = \sigma_{\epsilon} \epsilon_t^*\).
In the subsequent analysis, we establish that for any variable $\kappa_t$, the condition $\rho_{\kappa \epsilon} = 0$ holds if and only if $\rho_{\kappa^* \epsilon^*} = 0$. Put differently, the exogeneity of $\kappa_t$ is synonymous with the absence of correlation between its standard normal transformation $\kappa_t^*$ and $\epsilon_t^* = \epsilon_t / \sigma_{\epsilon}$. This result will be the basis for our exogeneity test.
We emphasize the non-triviality of this result, particularly in distinguishing between exogeneity and independence conditions. The former merely necessitates the absence of correlation with the error term, as opposed to the latter's requirement of independence. Consider a continuous variable $\kappa_t \in \{Z_t, \eta_t, \epsilon_t\}$. We have $\kappa_t^*=\Phi^{-1}(F_{\kappa}(\kappa_t))$, indicating that $\kappa_t^*$ is a monotone transformation of $\kappa_t$. Consequently, if $\kappa_t$ is independent of $\epsilon_t$, it directly follows that $\kappa_t^*$ is independent of the transformed error term $\epsilon_t^*=\epsilon_t/\sigma_{\epsilon}$, leading to a correlation $\rho_{\kappa^* \epsilon^*} = 0$. However, the mere absence of correlation, $\rho_{\kappa \epsilon} = 0$, does not imply independence. Moreover, the standard normal transformation of $\kappa_t$ given by $\Phi^{-1}(F_{\kappa}(\kappa_t))$ is a non-linear transformation. Therefore, $\rho_{\kappa \epsilon} = 0$ resulting in $\rho_{\kappa^* \epsilon^*} = 0$ is not a trivial argument.
This result will be demonstrated first for variables $\kappa_t$ that are continuous, followed by an examination of discrete $\kappa_t$. We begin by presenting a proposition that articulates the relationship between $\rho_{\kappa \epsilon}$ and $\rho_{\kappa^* \epsilon^*}$ for a continuous variable $\kappa_t$.
The following proposition follows from Proposition (ref).
Proposition (ref) establishes the same result as Proposition (ref) for discrete $\kappa_t$.
Building upon the findings of Propositions (ref), (ref), and (ref), we formulate a methodology to test the exogeneity of the instrumental variables \(Z_t\). This test hinges on examining the correlation between the standard normal transformation of \(Z_t\), denoted as \(Z_t^*\), and the normalized error term \(\epsilon^*\). Specifically, the test seeks to ascertain whether \(\rho_{Z_i^* \epsilon^*} = 0\). The details and formal structure of this exogeneity test are outlined in Theorem (ref).
Theorem (ref) demonstrates that testing the exogeneity of the instrumental variables essentially reduces to a Wald test for \(\Sigma_{Z^*} \theta_{Z_t^*} = 0\). Furthermore, the exogeneity of the error term in the reduced-form equation for the endogenous variable \(\eta_t^*\) can be assessed by verifying \(\theta_{\eta_t^*} = 0\). Notably, \(\theta_{\eta_t^*} = 0\) implies \(\rho_{\eta^* \epsilon^*} = 0\), which subsequently leads to \(\rho_{\eta \epsilon} = 0\).
A potential scenario that may adversely affect the performance of our exogeneity test involves multicollinearity in the estimation of (ref). Specifically, if \(P_t\) is highly correlated with \(Z_t^*\), it leads to multicollinearity, which in turn causes inefficiency in regression estimates and the Wald test. Such high correlation may emerge under two conditions:
A larger sample size may be required to mitigate the inefficiency issue in scenarios satisfying these conditions,.
Furthermore, considering that \(\Sigma_{Z^*} \theta_{Z_t^*} = \sigma_{\epsilon}\rho_{Z^* \epsilon^*}\), we can utilize the estimated coefficients ($\widehat{\theta_{Z_t^*}}$) and the Root Mean Square Error (RMSE) from the regression equation ($\widehat{\sigma_{\epsilon}}$) in (ref) to derive an estimate for \(\widehat{\rho_{Z^* \epsilon^*}} = [\widehat{\rho_{Z_i^* \epsilon^*}}]\). Subsequently for continuous instrumental variables, employing the results from Proposition (ref), we are able to calculate an estimate \(\widehat{\rho_{Z \epsilon}} = [\widehat{\rho_{Z_i \epsilon}}]\). This approach enables us to gauge the extent of endogeneity for each instrument, thereby aiding in more informed instrument selection. Instruments exhibiting higher levels of correlation can be excluded, and we also gain insight into the potential bias introduced in the Two-Stage Least Squares (TSLS) approach. This methodology not only refines instrument selection but also enhances the overall reliability of the econometric analysis.
In the next section, we delineate the application of Theorem (ref) for testing the endogeneity of regressors in the absence of instrumental variables. This approach leverages the theorem's results to test for endogeneity directly, bypassing the traditional reliance on instruments.
Suppose that we want to estimate the regression model in (ref), where \(P_t\) is a potentially endogenous variable. Traditional methods to test endogeneity, such as those proposed by durbin1954errors, hausman1978specification, wu1973alternative, and wooldridge2010econometric, require access to an exogenous instrument. In scenarios devoid of such instruments, these methods offer no reliable means to ascertain the exogeneity of \(P_t\). This section introduces an instrument-free approach, leveraging Theorem (ref), to test the exogeneity of a regressor.
Theorem (ref) provides a framework to test the exogeneity of instruments. Since \(\theta_{\eta_t^*}\) in (ref) equals \(\sigma_{\epsilon}\rho_{\eta^* \epsilon^*}\), and in light of Propositions (ref) and (ref), testing the exogeneity of \(\eta_t\) is analogous to verifying \(\theta_{\eta_t^*} = 0\). To assess the exogeneity of \(P_t\), we modify the reduced-form equation in (ref) by setting \(P_t^* = \eta_t\). Given that \(P_t^*\) follows a standard normal distribution, \(\eta_t^* = P_t^*\). This allows us to apply the results of Theorem (ref) under the assumption that there are no instruments and \(\eta_t^* = P_t^*\). The ensuing corollary formalizes this test:
In this section, we conduct simulation studies to demonstrate the efficacy of our copula-based tests in assessing the exogeneity of both instruments and regressors.
We examine a scenario where both \(X_t\) and \(P_t\) are \(1 \times 1\) vectors, with three available instrumental variables for the endogenous variable \(P_t\). The data-generating process adheres to the Gaussian copula model described in (ref). Initially, \(Z_t^*\), \(\eta_t^*\), \(\epsilon_t^*\), and \(X_t^*\) are generated using a multivariate normal distribution with a correlation matrix described below. Next, for each variable \(\kappa\) in the set \(\{Z_t, \eta_t, \epsilon_t, X_t\}\), we define \(\kappa_t = F_{\kappa}^{-1}(\kappa_t^*)\), where the CDF function $F_{\kappa}(\cdot)$ for each variable corresponds to the distributions specified below . The simulation parameters are as follows:
For each scenario, we conduct 100 simulations and apply the Wald test for the correlation factor of each instrument as stipulated in Theorem (ref). The proportion of instances where the Wald test rejects the null hypothesis (that the correlation between the instrument and error term is zero) is recorded. Additionally, the correlation between \(P_t\) and \(\epsilon_t\) is analyzed to assess the degree of endogeneity. The results are summarized in Tables (ref) and (ref) at the 5% significance level for $T=200$ and $T=1,000$, respectively. We summarize the main insights as follows:
We present the simulation results for the 1% significance level in Appendix (ref). The insights are similar to the case with the 5% significance level.
To evaluate the performance of our test for regressor exogeneity, we generate the standard normal transformations of \(P_t\), \(X_t\), \(\epsilon_t\), and \(Z_t\) based on a specified correlation matrix described below using a multivariate normal distribution, subsequently creating the variables using their inverse CDF functions. We consider only one instrumental variable. This instrumental variable may have a potential correlation with the error term. We utilize this instrument to conduct the Hausman endogeneity test, comparing its outcomes with our instrument-free copula-based endogeneity test. The simulation setup is detailed as follows:
Tables (ref) and (ref) show the regressor endogeneity results for the case with $\rho_{Z^* \epsilon^*}=0$ at the 5% significance level for $T=200$ and $T=1,000$, respectively, for various scenarios using both the instrument-free copula-based method and the Hausman approach. For brevity, we present the results for the case with $\rho_{Z^* \epsilon^*}=0$ at the 1% significance level for $T=200$ and $T=1,000$, and the case with $\rho_{Z^* \epsilon^*}=0.2$ at the 5% significance level for $T=1,000$ in Appendix (ref).
Key insights drawn from these simulation results are summarized below:
We emphasize that even though the type-II error of the copula based test is high for low levels of endogeneity, still our copula based method outperforms the Hausman endogeneity test approach.
This section evaluates the performance of our exogeneity tests in the context of Angrist's seminal paper on instrumental variables (angristcompulsory). The paper examines the impact of education duration on earnings, addressing the endogeneity issue arising from the unobserved ability of a student influencing both education duration and wage. The authors employ a two-stage-least-square (TSLS) method to assess the effect of education (\(E_i\)) on the log of wage (\(\ln W_i\)), using quarter of birth dummies interacted with year of birth dummies as instruments.
We apply our exogeneity tests to the endogenous variable (education length, \(EDUC\)) and the 30 instrumental variables (\(QTR120-QTR129\), \(QTR220-QTR229\), and \(QTR320-QTR329\)) for men born between 1920 and 1929. Our analysis mirrors the TSLS results presented in Table IV of angristcompulsory, specifically for Cases 2, 4, 6, and 8, described as follows:
Considering the discrete nature of \(EDUC\) and the instrumental variables, and the randomness in the normal distribution transformation for discrete variables, we execute our test 100 times with different random draws for the normal distribution transformations. We then calculate the frequency, out of 100 trials, at which the exogeneity hypothesis is rejected at the 5% significance level. The results are shown in Table (ref). We provide the results for the 1% significance level in Appendix (ref).
As evidenced in Table (ref), for cases 2 and 4, the exogeneity hypothesis of the endogenous variable \(EDUC\) is rejected in over 77% of the simulations, strongly indicating the endogeneity of the education length variable. However, this apparent endogeneity measure diminishes to 6% and 2% in cases 6 and 8, where demographic variables are included, suggesting that these demographic factors account for the portions of the error term correlated with \(EDUC\).
Furthermore, Table (ref) indicates that for all instrumental variables across all cases, the exogeneity hypothesis is rejected on average 5% of the time. This consistently low rejection rate provides robust evidence supporting the exogeneity of the instruments.
In conclusion, our study addresses a pivotal challenge in causal inference using instrumental variables – the testing of the exogeneity condition, which has long been a contentious and unresolved issue in applied econometrics. We propose a novel Copula-based method, a significant departure from traditional practices that largely relied on economic-theoretical justifications or untestable assumptions. Our approach, grounded in modeling the joint distribution of error terms and variables using a Gaussian copula, marks a substantial advancement in verifying the exogeneity of both instruments and regressors.
Our findings, derived from extensive simulation studies, demonstrate the robustness and accuracy of our test. Notably, the instrument exogeneity test shows high efficacy, correctly rejecting exogenous or endogenous instruments in most cases, even under non-normal error term distributions. Furthermore, our regressor exogeneity test outperforms the Hausman test, providing more reliable results without the need for exogenous instruments.
The empirical application of our test in a TSLS setting, using Angrist's study on compulsory education, further validates its practical utility. We successfully confirmed the endogeneity of the education variable sing an instrument-free approach. We also confirmed and the exogeneity of the birth year and quarter interaction instruments, exemplifying the test's applicability in real-world scenarios.
Our work makes a significant contribution to the field of applied econometrics by introducing a reliable and easy-to-implement approach for testing exogeneity. This advancement enhances the rigor and reliability of econometric analyses and offers substantial benefits to researchers and practitioners in the field. It opens new avenues for conducting more accurate and credible empirical research, paving the way for informed econometric analysis across various applied contexts.