Endogeneity problem is one of the crucial complications in econometrics field.

In econometrics the
problem of endogeneity occurs when the independent variable is correlated with the error term in a
regression model. Endogeneity can arise as a result of measurement error, autoregression with
autocorrelated errors, simultaneity and omitted variables. And, there are many methods overcoming
this including instrumental variable. In this study, the issue of endogeneity will be presented and the
tests to identify and the methods to solve the problem will be analyzed. In the following sections,
the meaning and implications of the term will be briefly presented and the literature will be
surveyed thoroughly afterwards. The tests to identify the endogeneity and the methods to eliminate
the effects of endogeneity will follow in the following sections. 2 1.1. Classical Linear Regression
Model and Endogeneity In order to analyze the problems caused by the endogeneity, we should
start with the textbook definition of the econometric assumptions underlying the regression.
Endogeneity occurs when one of the independent variables of the regression equaiton is correlated
with the unknown random error term (Wooldridge, 2002). The reasons that cause the exogeneous
variables to become correlated with the error term is the core of this study. The classical linear
regression model consists 5 key assumptions about the sample is created (Kennedy, 1998), and
violations of these assumptions cause severe econometric problems which cause the estimation
biased in most cases. Endogeneity bias occurs because of a few violations in these assumptions. In
order to understand the nature of this bias, we will start with defining the key assumptions of an
econometric model: 1. Dependent variable (Y) can be measured as a linear function of a set of
independent variables (a vector X) and an unobservable random error term (). The coefficients of
this estimation, denoted by a vector , are assumed to be constant. 2. The expected value of the
error term is equal to zero. 3. The error term must have the same variance along the sample and is
supposed to be uncorrelated within the sample. 4. The values of independent variable are fixed in
repeated samples. 3 5. The number of observations is greater than the number of independent
variables and there are no exact linear relationships between independent variables. The violation of
these assumptions leads to different econometric problems: 1. One of the violations1 that can occur
in the first assumption is called “wrong regressors”. It includes either the omission of relevant
independent variables or inclusion of irrelevant independent variables (Kennedy, 1998). The
omission of relevant independent variables, or “omitted variable bias” is one of the causes of
endogeneity that we will discuss more deeply in the following pages. 2. When the expected value of
the error term is not equal to zero, then there is socalled “biased intercept problem”. 3. Two major
econometric problems arise because of the violation of this assumption, one of which is going to be
our concern in this study: heteroscedasticity and autocorrelated errors which the latter is proved to
be one of the reasons for endogenetiy. 4. When the values of independent variables are not fixed in
repeated samples, 3 related econometric problems arise: measurement error, autoregression and
simultaneity, all three of them are considered as the sources for endogenetiy. 5. Violation of the fifth
assumption is called multicollinearity, or the direct relationship among two or more independent
variables. 1 There are two more violations of the first assumption, namely nonlinearity and changing
parameters as it is shown in Table 1, but they are out of the scope of this study. 4 Table 1: The
Assumptions of the Classical Linear Regression Model2 Assumption Mathematical expression
Violations (1) Dependent variable is a linear function of a specific set of independent variables, plus
an error term Y = X +  Wrong regressors Nonlinearity Changing parameters (2) Expected value of
error term is zero E( = 0 Biased intercept (3) Errors have uniform variance and are uncorrelated
E(' =  2 I Heteroskedasticity Autocorrelated errors (4) Observations on independent variables can
be considered fixed in repeated samples X fixed in repeated samples Errors in variables
Autoregression Simultaneous equations (5) No exact linear relationships between independent
variables and there are more observations than independent variables Rank of X = K < T Perfect
multicollinearity After this brief introduction and definition of our topic, now we can dig more into
the subject by looking at the causes, tests and remedies for endogeneity. To do that, we will first
discuss the bias that is causing endogeneity, then we will give the tests to identify the bias, wherever
applicable, and finally we will present the solutions or methods to eliminate the bias for each of the
violations in the following sections. Therefore we will start with the violation of the first assumption,
omitted variable bias. 2 Table is taken from Kennedy (1998) 5 1.2. Omitted Variables and
Endogeneity According to Jargowsky (2002), “omitted variable bias is the difference between the
expected value of an estimator and the true value of the underlying parameter due to failure to
control for a relevant explanatory variable or variables.” In other words, when one or more variables
that is supposed to be included in a model is left out, our estimation is likely to be in error. Consider
our regression model: Y = X +  (Eq.1) where Y is the dependent variable, X is the vector of
independent (explanatory) variables,  is the vector of coefficients and  is the unobservable
random error term. As we briefly discussed above, our classical regression model in Equation 1
assumes that the regression equation represents the “population model” or sometimes called the
“true model” constructed by a random sample of n observations (Wooldridge, 2002). Therefore, the
assumption states that all the relevant variables are included in the regression equation to represent
the “true nature” of the model. However, in reality, there might be cases that we misspecify the
model by including an irrelevant variable to the model (overspecification), or more seriously,
excluding a relevant variable from the model. Including an irrelevant variable has no effect on the
unbiasedness of our estimator, whereas omitting a variable has serious consequences as it is shown
below. 6 Following Clarke (2005) and Hanushek and Jackson (1977), let’s suppose that our “true”
regression model takes the form: Yi =  XXi3+ Xi4+  (Eq. 2) While we misspecify the
model by omitting variable X4 in the following way: Yi =  XXi3+  (Eq. 3) Therefore,
our error term now becomes:  = Xi4+  (Eq.4) Suppose that we are interested with the
estimation results for . Since the expected value of the true error term is assumed to be zero, the
expected value of estimated coefficient on X2 will be: E() =  + b42 (Eq.5) where b42 is the
regression coefficient on X2 in the “auxiliary” regression of the excluded variable, X4, on the
included variables, X2 and X3 (Clarke, 2005). Therefore, the classical regression model results would
be biased and the estimation will be inconsistent in a way that our estimation independent variables
X2 will be correlated with the error term as it is shown above. And this, in turn, means that one of
our independent variables becomes endogeneous. Because of the bias, the variance covariance
matrix of the coefficients becomes smaller (unless the omitted variable is orthogonal to the included
variables (Kennedy, 1998)), and therefore inferences using these parameters will be inaccurate as
the variance of the error term will be biased upward. 7 There are other cases that omitted variable
bias may occur. One specific case, functional form misspecification, occurs when the omitted
variable is a function of another explanatory variable in the model (Wooldridge, 2002). Omitted
variable bias is a serious problem and theoretically it is very easy to determine and solve the
problem as it is assumed that true model is known. However in estimations with empirical data,
determining the omitted variable is not a straightforward task (Clarke, 2005). In the literature, some
tests and solutions for the omitted variables are offered. Ramsey (1969)’s seminal paper on the
Regression Specification Error Test (RESET) is very handy in most of the cases such as omitted
variables, incorrect functional form and correlation between dependent variables and the error term
(Thursby and Schmidt, 1977). Ramsey RESET test is basically and F-statistic that is calculated by
adding polynomials, such as square or cube, of the fitted dependent variable (that is the square of
the cube of the estimated values of the dependent variable) to the original regression equation in
order to detect different functional forms. If any non-linear combination of the independent
variables is found to be powerful in explaining the dependent variable, then there is a
misspecification in the model. When our general model takes the form in Equation 1, we can
construct RESET test as follows: Suppose that we decided to add squares and the cubic forms of the
estimated dependent variable to the model. Then we need to re-estimate Eq.1 by using: 8 Y = X
+k+1ŷ 2 k+2 ŷ 3 +  (Eq. 6) where ŷ represents the previous estimation of the dependent variable
with the original equation and it is assumed that there are k independent variables in the original
equation. That is, Equation 1 is re-estimated using square and the cube of the estimated values of
dependent variable. After this calculation, RESET test basically employs an F test to check whether
k+1 and k+2 are jointly meaningful or not. That is, under the null hypothesis: H0: k+1, k+2 = 0
HA: k+1, k+2 ≠ 0 and the test statistic uses the determination coefficient, or R2 , from both
equaiton and the following test statistic is computed: [(R2 1- R 2 0)/(k-1)]/[(1- R 2 1)/(n-k)], where R2
1 is the R2 from Ramsey’s auxiliary regression and R2 0 is the original regression’s R2 . If the null
hypothesis that all regression coefficients of nonlinear terms are zero is rejected, then the model
suffers from misspecification. One drawback of the RESET test is that it is a general specification test
and it is not possible to clearly distinguish between omitted variables and the functional
misspecification by using RESET (Wooldridge, 1995). Thursby and Schmidt (1977) improved the
RESET test by combining it with other tests to aid in specification (Kennedy, 1998). 9 One other
method to determine omitted variables bias is a variant of Hausman (1978) specification test.
Hausman (1978) basically employs a chi square distribution for testing model specification. Suppose,
in our regression model, for our vector of coefficients, , we have two different sets of coefficients
constructed using two different specification, namely   and   . Hausman test asks which set has
consistent and efficient estimators. Under the null hypothesis, both sets are consistent, but only  
is efficient, and under the alternative hypothesis   is consistent and   is not. Therefore the
Hausman test statistic is: H = (  -  ) [Var (  )-Var(  )]-1 (  -  ) (Eq. 7) where  denotes the
transpose of the difference matrix and this test is distributed as a chi-square with a degrees of
freedom equal to the number of elements in the coefficient vector. This test is used for many
specification purposes, not only with omitted variable case3. A specific version of Hausman test,
which employs instrumental variables in the estimation of the second specification (  ), is called
the OV, or omitted variables, version of the Hausman test. Equivalent forms of the OV version of
Hausman test use the errors from a regression of X on the instruments, or the estimated X from this
regression, as the omitted variable (Kennedy, 1998). One handicap with Hausman test is that it is
sensitive to several types of misspecification. Therefore, Godfrey and 3 We will return to Hausman
test when we are discussing Instrumental Variables approach in order to eliminate the endogeneity
problem. 10 Hutton (1994) recommend testing for general misspecification before applying the
Hausman test, and they develop a general misspecification test in their paper. Also, Wong (1996)
finds that bootstrapping the Hausman test improves its performance. If an omitted variable is
determined by using several methods, briefly presented above, we can include the relevant variable
to the equation by using several methods. This includes using economic intuition to add a new
variable, prior empirical research, using proxies for the missing variables if there is no available data
for the omitted variable (Wooldridge, 2002) or using more sophisticated methods such as
instrumental variables4 (IV) in order to eliminate the effects of the omitted variable bias. Wooldridge
(2002) suggests four different ways to deal with omitted variables: (1) we can ignore the problem
and suffer the consequences of biased and inconsistent estimators; (2) we can try to find and use a
suitable proxy variable for the unobserved variable; (3) we can assume that the omitted variable
does not change over time and use the fixed effects or first-differencing methods if we face the
omitted variable problem in panel data; and finally (4) we can use IV approach which leaves the
unobserved variable in the error term, but rather than estimating the model by OLS, it uses an
estimation method that recognizes the presence of the omitted variable. One should note however
that, in reality, due to lack of data, or measurement problems of variables and such, the omitted
variable bias is not 4 Instrumental Variables method requires more detailed explanation and
therefore it will be given specific attention in the following section. 11 easy to deal with. According
to Jargowsky (2002), the definitive solution for omitted variable bias is to conduct a classical
experiment in which individuals are randomly assigned to treatment and control groups. However,
social sciences do not always allow for such experiments and thus, omitted variable is a possibility
and probability in social sciences despite the sincere efforts of the researcher. A more sophisticated
method of IV can be used in some cases, which we will turn back in the following sections. 1.3.
Measurement Errors and Autocorrelation In the empirical studies, it is not always possible to collect
the data that is truly reflecting the underlying economic theory due to many reasons including data
fudging. Even worse than this, in some cases even the researcher may not be aware of the true
meaning of the issue at hand (Streissler, 1970). Therefore, researchers sometimes have to use an
imprecise measure of an economic variable in an econometric model (Wooldridge, 2002), and this
leads to a measurement error, or as known as “errors-in-variables problem”. Measurement errors, in
some situations, just like in the omitted variable bias case, may cause the explanatory variables to be
correlated with the random error term and therefore would inconsistent regression results. As we
stated in the beginning, the fourth assumption of the linear regression model specifies that the
observations of the explanatory variables are considered fixed in repeated samples, that is, we
assume fixed regressors. 12 However in reality, this is not the case due to either measurement
errors in the regressors, or due to the necessity of using lagged dependent variables in the model,
which leads to autocorrelation in the regression equation. According to Angrist and Krueger(2001),
measurement error can occur for many reasons, including the limited ability of statistical agencies to
collect accurate information and the deviation between the variables specified in economic theory
and practice. The measurement error problem is a very well and long-known issue in econometrics
that may be dated back as much as 1920s. A well-known quotation, by Josiah Stamp reflects the
nature of the measurement problem very nicely: The Government are very keen on amassing
statistics - they collect them, add them, raise them to the nth power, take the cube root and prepare
wonderful diagrams. But what you must never forget is that every one of those figures comes in the
first instance from the village watchman, who just puts down what he damn pleases (Stamp, 1929,
quoted in Kennedy (1998)). There are books and papers that investigate the problem of
mismeasurement and data accuracy. In other words, measurement errors are concerned with the
effects of using incorrectly measured variables on the econometric studies. Due to the nature of the
regression equation, the errors in the measurement of the dependent variable poses no threat on
the consistency of the regression due to the fact that the error term itself incorporates the
measurement errors in the dependent variable by definition (Kennedy, 1998; Wooldridge, 2002).
However, when there are measurement errors in the 13 independent variables, the fourth
assumption above will be violated and the regressors will no longer be fixed, they will rather be
stochastic. In other words, when there is random errors in measuring an explanatory variable, then
the OLS estimation of the coefficient on that variable will be biased toward zero (Angrist and
Krueger, 2001). Therefore, the distribution of the variable will be correlated with the disturbance
term and this will create a bias in the classical regression model as shown below. Suppose our true
regression equation is again in the form of Equation 1 with one explanatory variable: Yi = 
X+  (Eq. 8) but with a difference that, our explanatory variable X is measured with an error z
such that: zi = X+  this implies that our explanatory variable can be written as Xzi-  and
our original regression takes the form: Yi = zi- +   or Yi =  zi+ -  (Eq.9) as
can be seen from equation 8, now the original regression is biased and there is a correlation
between the error term and the explanatory variable, which would cause an endogeneity problem
that needs to be taken care of in order to produce consistent estimator. 14 Therefore the question
arises here that how can one determine if there is a contemporaneous correlation between the
error term and the explanatory variables due to a measurement error? A version of Hausman (1978)
test is again widely used for this purpose. Since the consistency of the Instrumental Variable
estimator is not affected by the measurement error5 , classical regression estimation and the IV
estimation should produce similar results under the null hypothesis given in equation 7. That is, if
there is no measurement error in the classical regression model, then its results will be identical to
those from the IV model, otherwise, the results will be different. Hausman test is a test of statistical
equality of the regression results and therefore it can be used to capture the measurement error in
the classical regression model. When there is a contemporaneous correlation between the
explanatory variable(s) and the error term, the solution is usually to use an alternative estimator
with desirable asymptotic properties as estimators with desirable small-sample properties are hard
to find in this case (Kennedy, 1998; Wooldridge, 2002). The most common estimator used in this
context is IV estimator. IV estimator provide a consistent estimate in the presence of measurement
error if it is possible to find an instrument that is correlated with the explanatory variable but
uncorrelated with the measurement error. As this estimator is used for many different reasons, we
need to discuss this estimator in more detail in the following section. 5 Technical details for this
proposition will be given in more detail in the following section. 15 1.4. Instrumental Variable
Estimation As it is discussed previous sections, IV estimation is used to correct for omitted variable
bias, as well as measurement error and as we will see in the following pages, it is also used to correct
for the simultaneity bias. The theory of the instrumental variables as a problem of identifying and
estimating one or more coefficients of a system of simultaneous equaitons was first developed by
Philip G. Wright in his 1928 book The Tariff on Animal and Vegetable Oils (Stock and Trebbi, 2003).
Later the method of IV is used in many different areas in econometrics including our subject matter.
The instrumental variable technique is a general estimation procedure applicable to situations where
the independent variable(s) are correlated with the error term. If an appropriate instrumental
variable can be found for each endogenous variable that appears as a regressor in a simultaneous
equation, the instrumental variable technique provides consistent estimates (Kennedy, 1998).
Therefore, it becomes very useful in solving such problems as endogeneity, omitted variables and so
on. To explain how the instrumental variable estimation works, let’s begin with its formal
mathematical derivation in the simultaneous equations context. Remembering that our classical
regression model given in Eq. 1 was, in matrix notation: Y = X +  where Y is the dependent variable,
X is the vector of independent or explanatory variables (x),  is the vector of coefficients and  is the
random error 16 term, assumed to be normally distributed with a zero mean and finite variance.
When we try to estimate this equation using ordinary least squares (OLS, classical reg) by using
number of n observations, one can write the OLS estimator for the vector of coefficients as follows:
βOLS = (XY/XX) = X(Xβ + )/XX = β + (X)/(XX) (Eq. 10) where βOLS denotes the estimated
coefficents vector and  denotes the transpose of the vector. With our initial assumptions about the
OLS, the X’s and are uncorrelated as given in the Table 1 and therefore, the OLS estimator βOLS is
unbiased and consistent given the 5 assumptions hold. However, so far we have seen that there are
several violations of these assumptions. Therefore let’s now define an “instrumental variable”. An
instrumental variable, z, is a variable that is correlated with the independent variable x, that is Cov
(z, x) ≠ 0 but not with the error term, , that is, Cov (z,) = 0. One should note that while the
covariance condition on z and x means that z must be related to the endogenous explanatory
variable x; the covariance condition between z and  is interpreted differently in different contexts.
In the case of omitted variables, for example, this means that z should have no partial effect on y and
z should not be correlated with other factors that impact the dependent variable, y (Wooldridge,
2002). In the case of simultaneous equations, we interpret this condition as an “exogeneity
condition” by saying z is exogenous to equation 1. Using our instrument and by employing method
of moments, the conditional expectation of the dependent variable Y, on z can be calculated as: 17
E[Y|z] = E[X|z] + E[|z] (Eq.11) by definition, the last term, conditional expectation of  on z is equal
to zero. If we solve Eq.11 for b and write the resulting expression in terms of sample moments, we
will get: βIV = (ZY/ZX) = β + (Z)/(ZX) (Eq. 12) since z and  are uncorrelated, the last term in eq.12
approaches to zero in the limit, which mathematically provides an asymptotically consistent
estimator for the coefficient vector. Kennedy (1998) states that the major drawback to IV estimation
is that the variance-covariance matrix of the IV estimator is larger than that of the OLS estimator, by
an amount that is inversely related to the correlation between the instrument and the regressor.
This is the price paid for avoiding the asymptotic bias of OLS. Another problem with IV approach is to
find appropriate instrumental variables which are uncorrelated with the error term. Bound [Link].
(1995) argues that the use of instrumental variable in the case of a weak relationship between
instruments and the endogenous explanatory variable causes large inconsistencies in the IV
estimates and in the finite samples IV suffers from the same bias as OLS does. Therefore they
suggest the use of partial R2 and F statistics on the excluded instruments in the first stage of the
regression. 18 1.5. Back to the Measurement: Using IV Method to correct for the Measurement Error
and Omitted Variable Bias After explaining the how IV is constructed and how it works, now we can
go back to our discussion on how to correct the endogeneity bias coming from omitted variables and
the measurement error by using IV methodology. In the case of omitted variables, after determining
the existence of an omitted variable by the tests explained above, one can use IV method to correct
the bias. As it is discussed above, in social sciences there are many factors relevant to an economic
behavior in question, and it is highly possible to omit one related variable from the model. If all
these variables can be measured and held constant in a regression, the omitted variable bias would
be eliminated. However, in practice, it is almost impossible to measure all of the relevant variables
(omitted variables); and measure them accurately (measurement error) even when they are
specified correctly (Angrist and Krueger, 2001). In order to see how IV corrects the omitted variables
bias, let’s consider our OLS equation given in equation 1 when there are two independent variables
to explain the variations in the dependent variable: Yi =  XXi3+  (Eq. 13) Suppose that
we have a third variable, the instrument, denoted by z, which is correlated with Xi2, but otherwise
unrelated to the dependent variable, Y. That is z is uncorrelated with the omitted variables and the
regression error, . Then an IV estimator estimate of the dependent variable is the sample analog of
19 Cov(Yi, zi) / Cov(Xi2, zi). IV method here allows us to estimate the coefficient of interest
consistently and free from asymptotic bias from omitted variables, without actually having data on
the omitted variables (Angrist and Krueger, 2001), because of the fact that we can now estimate our
coefficient of interest  . Cov (Yi , zi ) = Cov(Xi2, zi) + Cov(zi , ) (Eq.14) since the last term in
equation 14 is zero by definition, and the middle term is different from zero, we can solve for  as:
Cov (Yi , zi ) / Cov(Xi2, zi) after cancelling the sample sizes in the numerator and the
denominator, we get the instrumental variables estimator for : IV = [Σ(zi – z¯)(yi – y¯)]/ [Σ(zi –
z¯))(yi – x¯)] (Eq.15) where the bar denotes the mean of the variable in the sample. It can be easily
shown by using law of large numbers that IV estimator is consistent for our coefficient of interest
provided that our assumptions about correlations are satisfied (Wooldridge, 2002). Therefore,
instrumental variable approach is successfully solves the omitted variable bias. With IV approach the
most important potential problem arises if the instrument is “bad”, that is if the instrument is
correlated with the omitted variables (or the error term in the original equation) (Bound [Link], 1995).
Another 20 problem with this approach is the possibility of bias when instruments are weakly
correlated with the endogenous explanatory variable. In order to eliminate weak instruments
problem, several solutions are proposed in the literature, including Bound [Link] (1995)’s partial R2
and reducing the number of instruments in the equation as the bias is proportional to the degree of
overidentification (that is if there are K instruments and G endogenous variables, the bias is
proportional to K – G, and reducing K would decrease the bias.) (Angrist and Krueger, 2001). In terms
of the measurement error, the use of instrument variables follows almost the same steps as in the
case of omitted variables with some differences. When there is a measurement error, one can
include the lagged value of independent variable as it is not contemporaneously (pair-wise)
correlated with the error term, with the assumption that error term is not autocorrelated (Kennedy,
1998). Another method used in the literature is the “two-group method” as suggested by , in which
the observations are split into two equalsized groups on the basis of the size of the regressor and
then the slope coefficient is estimated by the line joining the arithmetic means of the two groups,
can be interpreted as an instrumental variables estimator with the instrumental variable taking the
value -1 if the regressor value is below its median value and +1 if above its median value. The
rationale behind this method is that by averaging the data in this way the measurement errors are
also averaged, reducing their impact (Johnston, 1984). After investigating the problem of the
omitted variables bias, and measurement errors, and the 21 solutions to them by using IV approach,
now we can further analyze the issue of endogeneity in the case of simultaneity bias, which will be
done in the next section. 1.6. Simultaneity Simultaneity occurs in many areas in economic research
due to the nature of the economics, as economic variables are highly interrelated with each other.
Simultaneity can be defined as the joint determination of one or more explanatory variables with the
dependent variable, typically through an equilibrium mechanism (Wooldrige, 2002). In order to deal
with the simultaneity issue intrinsic in the nature of economic data, simultaneous equation models
has long been developed in econometric theory. The development of the notion of simultaneity goes
back as early as the work of E. Working’s seminal paper “What do ‘Statistical Demand Curves’ Show”
where he gave an early account of identification problem, while Haavelmo is the first one who
realized that in the presence of jointly endogenous variables, a joint probability distribution is
necessary to analyze the data (Hausman, 1984). Koopmans (1949), Koopmans and Reiersol (1950)
and Koopmans, Rubin and Leipnik (1950) established conditions for identification in linear
simultaneous models. Later in the literature, many other models dealing with simultaneous
equations are developed and used in empirical research. 22 The joint endogeneity concept is the
principle behind the simultaneous equation models. Therefore, endogeneity is the key concept in
most simultaneous equation models, even though not all simultaneous equations systems suffer
from the simultaneous equation estimation bias6 . However in most cases, simultaneity is creating a
bias needs to be solved in order to attain consistent estimators. To understand the issue more
clearly, let’s begin with simultaneous equaitons. In a system of simultaneous equations, all the
endogenous variables are random variables since a change in any error term changes all the
endogenous variables since they are determined simultaneously. A most widely used example in
textbooks of econometrics to explain the simultaneity issue is the Keynesian system with a
consumption function7 : C = a + bY +  (Eq.16) and an equilibrium condition Y = C + I (Eq.17) 6 A good
example for this is Recursive Estimation. A recursive system is one in which there is unidirectional
dependency among the endogenous variables. In a recursive system, a change in the disturbance in
the fourth equation, for example, affects directly the fourth endogenous variable, which in turn
affects the higher-ordered endogenous variables in the system, but does not affect the lower-
ordered endogenous variables. Because only lower-ordered variables appear as regressors in the
fourth equation, there is no contemporaneous correlation between the disturbance and the
regressors in the fourth equation. If there is no correlation between disturbances in different
equations, OLS estimation is consistent, and if no lagged endogenous variables appear among the
exogenous variables in the equation, it is unbiased (Kennedy, 1998). However, this is only a special
case, and therefore it is not discussed in the text above. 7 This example is taken from Kennedy
(1998), p.158-159 23 where C, consumption and Y, income are endogenous variables and I,
investment is an exogenous variable. Consider the problem of estimating the consumption function,
regressing consumption on income. Suppose the error term in the consumption function increases.
This directly increases consumption, this in turn will increase income through the equilibrium
condition. But income is the independent variable in the consumption function. Thus, the
disturbance in the consumption function and the regressor are positively correlated. An increase in
the disturbance term (directly implying an increase in consumption) is accompanied by an increase
in income (also implying an increase in consumption). When estimating the influence of income on
consumption, however, the OLS technique attributes both of these increases in consumption
(instead of just the latter) to the accompanying increase in income. This implies that the OLS
estimator of the marginal propensity to consume is biased upward, even asymptotically. As this
example shows very clearly, the endogeneity bias is intrinsic in the system of simultaneous
equations with a few exceptions (see the previous footnote). In other words, simultaneity includes
endogeneity by its nature. Therefore it is important to present technical details of the bias to
understand the underlying structure that leads to inconsistency in the classical regression model
(OLS). To this end, we will use a basic two-equation structural model8 : Y1 =  Y2 X+ 
(Eq.18) 8 For simplicity we assume that there is no intercept in the equations. 24 Y2 =  Y1 X+
 (Eq.19) where Y’s are endogenous explanatory variables, and X’s are exogenous variables and 's
are the error terms for each equation. Now let’s turn our focus on estimating the first equation. The
variables X1 and X2 are exogenous, so that each is uncorrelated with  and . Now we want to
show that Y2 is generally correlated with . If we solve the two equations for Y2 in terms of the
exogenous variables and the error term and plug in plug the right-hand side of Eq.18 in for Y1 in
Eq.19, we get Y2 =  Y2+ X+  X +, or equivalently: (1-)Y2 =X1 +X2
+ (Eq.20) finally in order to solve for Y2 we need to make an assumption, which states:
 ≠ (Eq.21) and if we divide equation 20 by (1-we get: Y2 = π X+ π X+v , (Eq.22)
where π/(1- π/(1- and v2 = [+ /(1- (Eq.23) 25
Equation 23 is simply called the reduced form for Y2. π parameters are reduced form parameters,
which are the nonlinear functions of the structural parameters () in the structural
equations 18 and 19. When Y2 is correlated with  because of simultaneity, then it is said that OLS
suffers from simultaneity bias which leads to inconsistent estimation results. Since now we have a
notion of the endogenetiy bias, we can start presenting the test for endogeneity and the solution to
this bias. The solution to this bias is always to use an unbiased estimator such as IV, Two Stage Least
Squares, Maximum Likelihood and so on. But before reaching to the estimation with an alternative
estimator, we need to discuss the tests for identification, and for endogeneity and exogeneity within
the system of simultaneous equations. To begin with, let’s first discuss the tests of exogeneity used
in econometric literature. One of the most widely used test for exogeneity/endogeneity of the
variables is a variety of Hausman test, a test for contemporaneous correlation between the error
and regressors which is described in detail in the previous section. Suppose that one of our
equations has the form: y = δY + X + e where δ is the coefficient vector for endogenous explanatory
variables, Y, and we wish to test all the variables in Y for exogeneity. This is done exactly as in
previous section, via an omitted variable version of the Hausman test. Estimated Y values, Y*, are
formed from the instruments (all the exogenous variables in the system), y is regressed on Y, X and
Y*, and the coefficient 26 estimate of Y* is tested against zero using an F test, given in equation 7.
Instead of testing for all the variables in the endogenous variables vector, one may wish to test some
of the variables for exogeneity. Technical details of this derivation is skipped here but we should
note that, in this case, the Hausman testing methodology becomes more complicated, Spencer and
Berk (1981) constructed a variety of Hausman test for this purpose. After the tests for exogeneity,
one needs to “identify” the model in order to consistently estimate the regression. As we discussed
in the beginning of this section, OLS estimator is biased when there is a “joint endogeneity” among
the variables at hand which poses a problem of finding an unbiased estimator for the coefficients of
interest. A natural response to this estimating problem is to suggest that the simultaneous system be
solved and put into its reduced form. This means that every endogenous variable is expressed as a
linear function of all the exogenous variables (and lagged endogenous variables, which are
considered exogenous in this context). For the simple Keynesian example given at the beginning, the
structural equations in Equation 16 and 17 can be solved to give the reduced-form equations: Y =
(a /(1-b)) + (1/1-b)I + 1/(1-b) C = (a /(1-b)) + (b/1-b)I + 1/(1-b) which can be rewritten in more
general form as: Y = π1 + π2I + v1 27 C = π3 + π4I + v2 where the π are parameters that are
(nonlinear) functions of the structural form parameters and the  are the reduced-form
disturbances, functions of the structural form disturbances. Because no endogenous variables
appear as independent variables in these reduced-form equations, if each reduced-form equation is
estimated by OLS, these estimators of the reduced-form parameters, the p, are consistent (and if no
lagged endogenous variables appear among the exogenous variables, these estimators are
unbiased). Economic theory tells us that these reduced-form parameters are the long-run multipliers
associated with the model. If a researcher is only interested in predicting the endogenous variables,
or only wishes to estimate the size of these multipliers, he can simply use these estimators. If,
however, he is interested in estimating the parameter values of the original equations (the structural
parameters), estimates of the reduced-form parameters are of help only if they can be used to
derive estimates of the structural parameters. However, this is not always possible; this problem is
one way of viewing the identification problem. If an equation is identified, it may be either "just-
identified" or "overidentified." An equation is just-identified if the number of identifying restrictions
placed on the model is the minimum needed to identify the equation; an equation is over-identified
if there are some extra restrictions beyond the minimum necessary to identify the equation. After
the identification of the model 28 is done with certain conditions, we can apply an appropriate
method to solve the simultaneity bias. In the literature, there are many different methods to
estimate a simultaneous equation. Estimators that estimate the equations individually within the
system of simultaneous equations are called Single equation methods, while estimators that
estimate the equations in the system simultaneously are called system methods. Single equation
models are usually called “limited information” methods while the system methods. Since each of
these methods are the subject of a new paper, and they are widely used and known, we will just list
these variables and conclude this section. There are 5 very well-known single equation methods and
2 system methods in the literature. Single Equation Methods: 1. Ordinary Least Squares OLS 2.
Indirect Least Squares ILS 3. Instrumental Variables IV 4. Two-stage Least Squares 2SLS 5. Limited
information Maximum Likelihood LI/ML System Methods 1. 3 Stage Least Squares, 3SLS 29 2. Full
information Maximum Likelihood (FIML). 1.7. Further Studies About Correcting For Endogeneity We
can describe other methods correcting Endogeneity in choice models. Several methods have been
developed recently to estimate choice models in the presence of endogenous variables. In this
frame, we can discuss three approaches: BLP approach, Control Function approach and Maximum
Likelihood approach. In this chapter, these approaches will be considered briefly and we will give
some technical analyses as sub chapters. We first describe the BLP approach, developed by Berry,
Levinsohn and Pakes (hence the initials) through a series of publications. It was pointed out that
constants can be included in the choice model to capture the average effect of the product
attributes (both observed and unobserved). Then, the estimated constants can be regressed against
the observed attributes in a linear regression, where endogeneity is handled in the usual way by
instrumental variables estimation of the linear regression. Essentially, it has been showed that the
endogeneity could be taken out of the choice model, which is inherently non-linear, and put into a
linear regression model, where endogeneity can be handled through standard instrumental variables
estimation. To apply this method, it is often necessary to estimate a very large number of constants
in the choice model, which can be difficult using standard gradient-based methods for maximization.
30 The second procedure that we describe is the control function approach. The concepts motivating
this approach date back to Heckman (1978) and Hausman (1978), though the first use of the term
“control function” seems to have been by Heckman and Robb(1955). As we stated above
endogeneity arises when observed variables are correlated with unobserved factors. This correlation
implies that the unobserved factors conditional on the observed variables do not have a zero mean,
as is usually required for standard estimation. A control function is the variable that captures this
conditional mean, essentially “controlling” for the correlation. The procedure is implemented in two
steps. First, the endogenous explanatory variable (such as price) is regressed against exogenous
variables. The estimated regression is used to create a new variable (the control function) that is
entered into the choice model. And then, the choice model is estimated with the original variables
plus the new one, accounting appropriately for the distribution of unobserved factors conditional on
both this new and the original variables. The third procedure is maximum likelihood approach, as
applied by VillasBoas and Winer (1999) to a multinomial logit with fixed coefficients and generalized
by Park and Gupta (2008) to random coefficient choice models. The procedure is closely related to
the control function approach, in that it accounts for the non-zero conditional mean of the
unobserved factors. However, instead of implementing the two steps sequentially (i.e., estimate the
regression model to create the control function and then estimate the choice model with this
control function), the two steps are combined into a joint estimation criterion. 31 Additional
assumptions are required to allow the estimation to be performed simultaneously; however, the
procedure is more efficient when those assumptions are met.