1/29
Looks like no tags are added yet.
Strict Exogeneity
The condition that the regression disturbance has a conditional mean of zero given the entire covariate matrix across all individuals and time periods: E(Ui∣X)=0, or equivalently E(Yi∣X)=E(Yi∣Xi)=Xi′β. It is the core requirement for the finite-sample unbiasedness of OLS.
Unbiasedness vs. Consistency
Unbiasedness is a finite-sample property stating that the mean of the sampling distribution equals the true parameter across repeated samples of fixed size N (E(β^)=β). Consistency is an asymptotic property stating that the estimator converges in probability to the true parameter as sample size approaches infinity (plimN→∞β^=β). Neither property implies the other.
Downwards Bias of OLS in AR(1) Models
In stationary autoregressive time series (Yt=β0+β1Yt−1+Ut), conditioning on the entire covariate matrix X includes future lags where Yt appears as an explanatory variable (Xt+1=(1,Yt)′). Thus, E(Yt∣X)=Yt=Xt′β, violating strict exogeneity and inducing finite-sample downward bias.
Role of N1/2 in OLS Asymptotic Normality
The scaling factor N1/2=N balances the rate of convergence of the standardized sum N1∑i=1NXiUi. Without N, the difference (β^−β) collapses to a degenerate point mass at zero; scaling by N causes the variance to diverge to infinity. The exponent 1/2 stabilizes the variance to yield a non-degenerate Gaussian limit distribution.
Heteroskedasticity-Consistent (HC) Covariance Matrix
The sandwich variance estimator V^HC=(X′X)−1(∑i=1NU^i2XiXi′)(X′X)−1 where U^i=Yi−Xi′β^. It yields standard errors that are asymptotically valid under both conditional homoskedasticity and arbitrary forms of conditional heteroskedasticity without changing the coefficient estimates.
Instrument Relevance Condition
The requirement that excluded instruments are correlated with the endogenous regressors after partialling out all included exogenous covariates. Formally, the matrix of population regression coefficients from the first-stage equation must have full column rank: rank(ΣZX)=K. It is assessed via a joint first-stage F or Wald test on the excluded instruments.
2SLS HC Residual Computation Rule
When constructing the heteroskedasticity-consistent variance-covariance matrix for 2SLS, residuals must be computed using the actual structural explanatory variables: U^i=Yi−Xi′β^2SLS. Using first-stage fitted values X^i yields invalid residuals and inconsistent standard errors because the actual structural disturbance is Ui=Yi−Xi′β.
Over-Identifying Restrictions (OIR / Sargan / Hansen J) Test
A diagnostic test available when the number of excluded instruments exceeds endogenous regressors (M>K). It evaluates H0:E(ZiUi)=0 by regressing structural residuals U^i on all instruments Zi (J=NRU^2→dχM−K2). Failure to reject supports over-identifying restrictions but does not prove exogeneity of the baseline just-identified model.
Regression-Based (Wu-Hausman) Endogeneity Test
A test for regressor endogeneity conducted by estimating the first-stage regression, extracting the residual vector V^i, and including it as an additional explanatory variable in the structural regression estimated via OLS: Yi=Xi′β+δV^i+εi. A rejection of H0:δ=0 indicates endogeneity, requiring 2SLS over OLS.
Indirect Least Squares (ILS)
A method to recover structural parameters in just-identified simultaneous equations systems by estimating the reduced-form equations via OLS and algebraically solving for the structural parameters as ratios and combinations of reduced-form coefficients (e.g., α^12=π^12/π^22).
Cluster-Robust (HAC) Panel Covariance Matrix
A panel variance estimator V^cluster=(X′X)−1(∑i=1NXi′U^iU^i′Xi)(X′X)−1 that groups all T residuals for individual i into a vector U^i. The cross-product matrix U^iU^i′ accounts for arbitrary heteroskedasticity across individuals and arbitrary serial correlation (autocorrelation) over time within each cross-sectional unit.
Long-Run Propensity (LRP) in FDL Models
In a Finite Distributed Lag model of order q (Yit=β0+∑j=0qδjXi,t−j+Uit), the Long-Run Propensity or long-run multiplier measures the total cumulative change in the outcome resulting from a permanent, one-unit increase in X, given by LRP=∑j=0qδj.
First-Differences (FD) Estimator
An estimator that subtracts period (t−1) observations from period t: ΔYit=ΔXit′β1+ΔVit. It completely eliminates unobserved, time-invariant individual heterogeneity αi, but drops all time-invariant observed regressors (ΔZi=0) and reduces the effective sample size from NT to N(T−1).
Fixed Effects (Within) Estimator
An estimator obtained by demeaning all variables by their entity-specific time averages: (Yit−Yˉi)=(Xit−Xˉi)′β+(Vit−Vˉi). It is algebraically identical to the Least Squares Dummy Variable (LSDV) estimator and relies exclusively on within-unit variation over time to identify β, conditioning out individual intercepts αi.
Incidental Parameters Problem
In short panels where N→∞ with fixed T, estimating individual intercepts αi directly (as in LSDV) requires estimating N separate parameters. Because each αi is estimated using only T observations, α^i is unbiased but inconsistent, even though the within slope estimator β^FE is consistent.
Identification Failure in Two-Way Fixed Effects (TWFE)
Double-demeaning in TWFE models (Yit∗=Yit−Yˉi⋅−Yˉ⋅t+Yˉ⋅⋅) eliminates any regressor that can be decomposed additively into an individual component and a time component (Xit=ai+bt). Examples include Ageit=t−BirthYeari and potential experience, which become exactly zero after transformation.
Nickell Bias
The severe downward finite-sample bias of OLS applied to first-differenced dynamic panel models (ΔYit=ρΔYi,t−1+ΔVit). Because Yi,t−1 contains the disturbance shock Vi,t−1, the differenced regressor is correlated with the differenced error: Cov(ΔYi,t−1,ΔVit)=−Var(Vi,t−1)=0.
Anderson-Hsiao and Arellano-Bond Estimators
Instrumental variable solutions for dynamic panels. Anderson-Hsiao instruments the differenced lag ΔYi,t−1 using lagged levels (Yi,t−2) or lagged differences (ΔYi,t−2). Arellano-Bond generalizes this via GMM, utilizing all available linear orthogonality conditions from deeper historical lags (Yi1,…,Yi,t−2) as time progresses.
Optimality of the Conditional Expectation Function (CEF)
Under expected quadratic loss (E[(Y−m(X))2]), the CEF m(X)=E(Y∣X) is the Minimum Mean Squared Error (MMSE) predictor over all possible functions. Population OLS is the Best Linear Predictor (BLP), representing the MMSE approximation constrained to linear functions X′b.
Decomposition of MSPE
Mean Squared Prediction Error on out-of-sample data decomposes mathematically into three additive terms: MSPE=σU2+[Bias(Y^i)]2+Var(Y^i), representing irreducible random noise, squared model approximation bias, and parameter estimation variance, respectively.
Conditional Mean CI vs. Individual Prediction Interval
A 95% confidence interval for the conditional mean (C′β^±1.96⋅se(C′β^)) captures only parameter estimation uncertainty. A 95% prediction interval for an individual (C′β^±1.96⋅se(C′β^)2+RMSE2) also incorporates the irreducible variance of the idiosyncratic individual shock U0, making it substantially wider.
Jensen's Inequality in Log Retransformations
Because the exponential function is strictly convex, E(exp(Ui)∣Xi)>exp(E(Ui∣Xi))=exp(0)=1. Consequently, the naive retransformed predictor Y^i=exp(Xi′β^) systematically underestimates the true conditional expectation E(Yi∣Xi), requiring a retransformation adjustment.
Duan's Smearing Estimator
A non-parametric retransformation adjustment that does not assume normality of disturbances. It estimates the multiplicative correction factor as the sample average of exponentiated residuals: a~0=N1∑i=1Nexp(U^i), yielding level predictions Y^i=a~0exp(Xi′β^).
Ridge vs. LASSO Regularization
Ridge (L2 penalty: λ∑bk2) shrinks coefficients continuously toward zero via closed-form solution (X′X+λI)−1X′Y, handling multicollinearity but never setting coefficients to zero. LASSO (L1 penalty: λ∑∣bk∣) features a non-differentiable kink at zero, driving small coefficients exactly to zero to produce sparse models and automatic variable selection.
Selection of Shrinkage Tuning Parameter (\lambda)
λ cannot be selected by minimizing in-sample SSR because in-sample SSR is mechanically minimized at λ=0 (OLS). It must be chosen by evaluating out-of-sample predictive performance over a grid of candidate values using M-fold cross-validation to find the value that minimizes MSPECV(λ).
Residual Autocorrelation Theorem
By the Law of Iterated Expectations, prediction errors from a correctly specified conditional expectation function cannot be autocorrelated at any lag: Cov(Ut,Ut−j)=E[Ut−jE(Ut∣Ft−1)]=0. Statistically significant autocorrelation in regression residuals indicates lag underspecification.
Akaike Information Criterion (AIC)
A model selection metric defined as AIC(p)=log(nSSRp)+n2(p+1). It balances in-sample goodness-of-fit against a linear parameter penalty. Rather than targeting structural model-selection consistency, AIC targets predictive optimality by minimizing out-of-sample forecast error variance.
Multi-Step Forecasting Failure of ARDL Models
While an ARDL(p,q) model can produce one-step-ahead forecasts using lagged covariates, it cannot generate iterated forecasts for horizons h>1 because it lacks structural equations describing the future evolution of the auxiliary predictors E(Xn+1∣Fn). A Vector Autoregression (VAR) solves this by modeling all variables simultaneously.
Granger Causality
Variable X Granger-causes variable Y if historical lags of X contain statistically significant information that helps predict future values of Y beyond the information contained in past lags of Y alone. It evaluates predictive forecasting content via a joint Wald/F test, not structural or economic causality.
Empirical Size vs. Power in Monte Carlo Experiments
Empirical size is the rejection frequency of a test evaluated when the null hypothesis is true (b=β0); an unbiased test should produce a rejection rate within sampling error of nominal size (α=0.05). Empirical power is the rejection frequency when the null hypothesis is false (b=β0); for a consistent estimator, power converges toward 1.0 (100%) as sample size grows.