Chapter 9 Lecture Notes: More on
Specification and Data Issues
Course: Introduction to Econometrics (Finance-Focused Version / Wooldridge)
Instructor Note: Comprehensive guide covering econometric intuition, model specifications,
finance applications, and practical research implications.
9.1 Functional Form Misspecification
Functional form misspecification occurs when the estimated regression model does not correctly
represent the true relationship between the dependent and explanatory variables. OLS assumes
that the conditional mean of the dependent variable is correctly specified. If the true relationship
is nonlinear but a linear model is estimated, coefficient estimates may be misleading even when
all relevant variables are included.
Suppose a researcher estimates:
INV_EFF = β₀ + β₁ FamilyFirm + β₂ TobinsQ + β₃ Leverage + u
The model assumes a linear relationship between Tobin's Q and investment inefficiency.
However, investment opportunities may have diminishing effects. The true model may be:
INV_EFF = β₀ + β₁ FamilyFirm + β₂ TobinsQ + β₃ TobinsQ² + β₄ Leverage
+ u
If the quadratic term is omitted, the estimated model suffers from functional form
misspecification. Researchers may incorrectly conclude that investment opportunities have a
constant marginal effect when the effect actually changes across firms.
9.1a RESET as a General Test for Functional Form
Misspecification
The Ramsey RESET test is a general specification test used to detect functional form errors.
The intuition is that if the model is correctly specified, nonlinear functions of the fitted values
should not provide additional explanatory power.
● Researchers first estimate the original model and obtain fitted values (ŷ).
● Additional terms such as fitted value squared (ŷ²) and fitted value cubed (ŷ³) are then
included in an auxiliary regression.
● The null hypothesis states that the original model is correctly specified. The alternative
hypothesis states that the model omits important nonlinearities or interactions.
9.2 Using Proxy Variables for Unobserved Explanatory
Variables
Many important concepts in finance cannot be directly observed. Examples include managerial
ability, corporate culture, investor sentiment, socioemotional wealth, and investment
opportunities. When such variables cannot be measured directly, researchers often use proxy
variables.
Suppose the true model is:
INV_EFF = β₀ + β₁ FamilyFirm + β₂ InvestmentOpportunities + u
Investment opportunities are unobservable. Researchers may therefore use Tobin's Q as a
proxy. Although Tobin's Q is imperfect, it is expected to be correlated with investment
opportunities.
The usefulness of a proxy variable depends on two conditions. First, the proxy must be highly
correlated with the unobserved variable. Second, the proxy should not contain substantial
independent effects unrelated to the omitted variable (so that OLS assumptions are not
violated).
9.4 Properties of OLS under Measurement Error
Measurement error occurs when observed data differ from the true underlying values.
Measurement error is pervasive in finance because accounting data, ownership measures, ESG
scores, and market variables are often imperfectly measured.
9.4a Measurement Error in the Dependent Variable
When measurement error occurs only in the dependent variable and satisfies the classical
assumptions, OLS coefficient estimates remain unbiased. However, the error variance
increases, making estimates less precise. Consequently, standard errors become larger and
statistical significance becomes harder to achieve. The primary effect is a loss of efficiency
rather than bias.
9.4b Measurement Error in an Explanatory Variable
Measurement error in explanatory variables is much more serious. The estimated coefficient
suffers from attenuation bias, meaning the coefficient is biased toward zero.
9.5 Missing Data, Nonrandom Samples, and Outliers
Empirical finance datasets rarely contain complete and perfectly representative information.
Missing observations, sample selection problems, and extreme values are common features of
real-world data.
9.5b Nonrandom Samples: A nonrandom sample occurs when the observed sample differs
systematically from the target population. If data are missing systematically, sample selection
problems arise. For example, if conclusions drawn from listed firms are applied to the broader
population, it may not generalize because listed firms tend to be larger and more transparent
than private firms.
9.5c Outliers and Influential Observations: Outliers are observations with unusually large or
small values. In Vietnamese data, firms such as Vingroup or Hoa Phat may have financial
characteristics very different from typical firms. Extreme values can dominate OLS estimation.
Researchers often winsorize variables at the 1st and 99th percentiles or at the 5th and 95th
percentiles to reduce the influence of extreme observations.
9.6 Least Absolute Deviations (LAD) Estimation
Ordinary Least Squares minimizes the sum of squared residuals. Because residuals are
squared, large errors receive disproportionate weight. Least Absolute Deviations (LAD)
estimation instead minimizes the sum of absolute residuals. Consequently, LAD is more robust
to outliers. The LAD estimator estimates the conditional median rather than the conditional
mean.