Wooldridge, Introductory Econometrics, 4th ed.

Chapter 9: More on specication and data problems

Functional form misspecication We may have a model that is correctly specied, in terms of including the appropriate explanatory variables, yet commit functional form misspecicationin which the model does not properly account for the relationship between dependent and observed explanatory variables. We have considered this sort of problem when discussing polynomial models; omitting a squared term, for instance, and constraining y/x to be constant (rather than linear in x) would be

a functional form misspecication. We may also encounter diculties of this sort with respect to interactions among the regressors. If omitted, the eects of those regressors will be estimated as constant, rather than varying as they would in the case of interacted variables. In the context of models with more than one categorical variable, assuming that their eects can be treated as independent (thus omitting interaction terms) would yield the same diculty. We may, of course, use the tools already developed to deal with these problems, in the sense that if we rst estimate a general model that allows for powers, interaction terms, etc. and then test down with joint F tests, we can be condent that the more specic model we develop will not have imposed inappropriate restrictions along the way. But how can

we consider the possibility that there are missing elements even in the context of our general model? One quite useful approach to a general test for functional form misspecication is Ramseys RESET (regression specication error test). The idea behind RESET is quite simple; if we have properly specied the model, no nonlinear functions of the independent variables should be signicant when added to our estimated equation. Since the tted, or predicted values () of the estimated model are linear in the iny dependent variables, we may consider powers of the predicted values as additional regressors. Clearly the y values themselves cannot be added to the regression, since they are by construction linear combinations of the x variables. But their squares, cubes,... are not. The RESET formulation reestimates the original equation, augmented by powers of y (usu ally squares, cubes, and fourth powers are sucient) and conducts an F-test for the joint null

hypothesis that those variables have no signicant explanatory power. This test is easy to implement, but many computer programs have it already programmed; for instance, in Stata one may just specify estat ovtest (omitted variable test) after any regression, and the Ramsey RESET will be produced. However, as Wooldridge cautions, RESET should not be considered a general test for omission of relevant variables; it is a test for misspecication of the relationship between y and the x values in the model, and nothing more. Tests against nonnested alternatives The standard joint testing framework is not helpful in the context of competing models, or nonnested alternatives. These alternatives can also arise in the context of functional form: for instance,

y = 0 + 1x1 + 2x2 + u y = 0 + 1 log x1 + 2 log x2 + u

(1)

are nonnested models. The mechanical alternative, in which we construct an articial model that contains each model as a special case, is often not very attractive (and sometime will not even be feasible). An alternative approach is that of Davidson and MacKinnon. Using the same logic applied in developing Ramseys RESET, we can estimate each of the models in (1), generate their predicted values, and include them in the other equation. Under the null hypothesis that the rst form of the model is correctly specied, a linear combination of the logs of the x variables should have no power to improve it, and that coefcient should be insignicant. Likewise, one can reestimate the second model, including the predicted values from the rst model. This

testing strategyoften termed the DavidsonMacKinnon J testmay indicate that one of the models is robust against the other. There are no guarantees, though, in that applying the J test to these two equations may generate zero, one, or two rejections. If neither hypothesis is rejected, then the data are not helpful in ranking the models. If both are rejected, we are given an indication that neither model is adequate, and that a continued specication search should be conducted. If one rejection is received, then the J test is denitive in indicating that one of the models dominates (or subsumes) the other, and not vice versa. However, this does not imply that the preferred model is well specied; again, this test is against a very specic alternative, and does not deliver a clean bill of health for the preferred model should one emerge.

Proxy variables So far, we have discussed issues of misspecication resulting from improper handling of the x variables. In many economic models, we are forced to employ proxy variables: approximate measures of an unobservable phenomenon. For instance, admissions ocers use SAT scores and high school GPAs as proxies for applicants ability and intelligence. No one argues that standardized tests or grade point averages are actually measuring aptitude, or intelligence; but there are reasons to believe that the observable variable is well correlated with the unobservable, or latent, variable. To what extent will a model estimated using such proxies for the variables in the underlying relationship be successful, in terms of delivering consistent estimates of its parameters? First, of course, it must be established that there is a correlation between the observable variable and the latent variable. If we consider the

latent variable as having a linear relation to a measurable proxy variable, the error in that relation must not be correlated with other regressors. When we estimate the relationship including the proxy variable, it should be apparent that the measurement error from the latent variable equation ends up in the error term, as an additional source of uncertainty. This is an incentive to avoid proxy variables where one can, since they will inexorably inate the error variance in the estimated regression. But usually they are employed out of necessity, in models for which we have no ability to measure the latent variable. If there are several potential proxy measures, they might each be tested, to attempt to ascertain whether bias is being introduced to the relationship. In some cross-sectional relationships, we have the opportunity to use a lagged value of the dependent variable as a proxy variable. For instance, if we are trying to explain cities crime

rates, we might consider that there are likely to be similaritiesirregardless of the eectiveness of anti-crime strategiesbetween current crime rates and last years values. Thus, a prior value of the dependent variable, understandably independent of this years value, may be a useful proxy for a number of factors that cannot otherwise be quantied. This approach might often be used to deal with factors such as business climate, in which some states or municipalities are viewed as more welcoming to business; there may be many aspects to this perception, some of them more readily quantiable (such as tax rates), some of them not so (such as local ocials willingness to negotiate infrastructure improvements, or assist in funding for a new facility). But in the absence of radical changes in localities stance in this regard, the prior years (or decades) business investment in the locality may be a good proxy for those factors, perceived much more clearly by the business decisionmakers than by the econometrician.

Measurement error We often must deal with the issue of measurement error: that the variable that theory tells us belongs in the relationship cannot be precisely measured in the available data. For instance, the exact marginal tax rate that an individual faces will depend on many factors, only some of which we might be able to observe: even if we knew the individuals income, number of dependents, and homeowner status, we could only approximate the eect of a change in tax law on his or her tax liability. We are faced, therefore, with using an approximate measure, including some error of measurement, whenever we might attempt to formulate and implement such a model. This is conceptually similar to the proxy variable problem we have already discussed, but in this case it is not a latent variable problem. There is an observable magnitude, but we do not necessarily observe it. For instance, reported income is

an imperfect measure of actual income, while IQ score is only a proxy for ability. Why is measurement error of concern? Because the behavior were trying to modelbe it of individuals, rms, or nationspresumably is driven by the actual measures, not our mismeasured approximations of those factors. To the extent that we fail to capture the actual measure, we may misinterpret the behavioral response. If measurement error is observed in the dependent variablefor instance, if the true relationship explains y , but we only observe y = y + , where is a meanzero error process, then becomes a component of the regression error term: yet another reason why the relationship does not t perfectly. We assume that is not systematic, in particular, that it is not correlated with the independent variables X. As long as that is the case, then this form of measurement error does no real harm; it

merely weakens the model, without introducing bias in either point or interval estimates. If the magnitude of the measurement error in y is correlated with one or more of the x variables, then we will have a problem of bias. Measurement error in an explanatory variable, on the other hand, is a far more serious problem. Say that the true model is y = 0 + 1x + u 1 (2)

but that x is not observed; instead, we ob1 serve x1 = x + 1. We can assume that E( 1) = 1 0 with generality. But what should be assumed about the relationship between 1 and x ? First, let us assume that 1 is uncorre1 lated with the observed measure x1 (that is, larger values of x1 do not give rise to systematically larger (or smaller) errors of measurement). This can be written as Cov( 1, x1) = 0. But if this is the case, it must be

true that Cov( 1, x ) = 0 : that is, the error 1 of measurement must be correlated with the actual explanatory variable x , so that we can 1 write the estimated equation (in which x is 1 replaced with the observable x1) as y = 0 + 1x1 + (u 1 1) (3)

Since both u and 1 have zero mean and are uncorrelated (by assumption) with x1, the presence of measurement error merely inates the 2 2 error term: that is, V ar (u 1 1) = u +1 21 , given that we have assumed that u and 1 are uncorrelated with each other. Thus, measurement error in x does not negatively aect the 1 regression of y on x1; it merely inates the error variance, like measurement error in the dependent variable. However, this is not the case that we usually consider under the heading of errors-invariables. It is perhaps more reasonable to

assume that the measurement error is uncorrelated with the true explanatory variable: Cov( 1 , x1 ) = 0. If this is so, then Cov( 1 , x1 ) = Cov( 1, x + 1 ) = 0 by construction, and the 1 regression (3) will have a correlation between its explanatory variable x1 and the composite error term. The covariance of (x1, u 1 1) = 1Cov( 1, x1) = 1 21 = 0, causing the OLS regression of y on x1 to be biased and inconsistent. In this simple case of a single explanatory variable measured with error, we can determine the nature of the bias:

plim(b1) = 1 +

= 1 2 x1 + 21

Cov (x1, u 1 1) V ar(x1) 2 x1

(4)

demonstrating that the OLS point estimate will be attenuatedbiased toward zerosince

the bracketed expression must be a fraction. Clearly, in the absence of measurement error, 21 0, and the OLS coecient becomes unbiased and consistent. As 21 increases relative to the variance in the (correctly measured) explanatory variable, the OLS coecient becomes more and more unreliable, shrinking toward zero. What can we conclude in a multiple regression equation, in which perhaps one of the explanatory variables is subject to measurement error? If the measurement error is uncorrelated to the true (correctly measured) explanatory variable, then the result we have here applies: the OLS coecients will be biased and inconsistent for all of the explanatory variables (not merely the variable measured with error), but we can no longer predict the direction of bias in general terms. Realistically, more than one explanatory variable may be subject to measurement

error (e.g. both reported income and wealth may be erroneous). We might be discouraged by these ndings, but fortunately there are solutions to these problems. The models in question, in which we suspect the presence of serious errors of measurement, may be estimated by techniques other than OLS regression. We will discuss those instrumental variable techniques, which may also be used to deal with problems of simultaneity, or two-way causality, in Chapter 15.