1/243
Looks like no tags are added yet.
Unbiased estimator
An estimator whose expected value equals the true parameter, E(β̂) = β, across repeated samples.
Consistent estimator
An estimator that converges to the true parameter as the sample size grows, plim β̂ = β.
Efficient (BLUE) estimator
The linear unbiased estimator with the smallest variance, which OLS is under the Gauss-Markov assumptions.
OLS slope after substituting the true model
β̂₁ = β₁ + Σ(xᵢ − x̄)uᵢ / SSTₓ, the starting point for nearly every bias derivation.
Probability limit of the OLS slope
plim β̂₁ = β₁ + Cov(x,u)/Var(x), so OLS is inconsistent whenever x is correlated with the error.
Why Σ(xᵢ − x̄)ū drops out
Deviations from a mean always sum to zero, so ū times Σ(xᵢ − x̄) equals zero.
MLR.4 zero conditional mean
E(u | x₁,…,xₖ) = 0, nothing in the error is systematically related to the regressors.
What MLR.1–MLR.4 deliver
Unbiasedness of the OLS coefficients.
What MLR.1–MLR.5 deliver
OLS is BLUE and the usual variance formula σ²/[SSTⱼ(1 − Rⱼ²)] is correct.
Omitted variable bias formula
E(α̂₁) = β₁ + β₂b₂₁, where b₂₁ is the slope from regressing the omitted x₂ on the included x₁.
Intercept bias with an omitted variable
E(α̂₀) = β₀ + β₂(x̄₂ − b₂₁x̄₁), which stays biased even when b₂₁ = 0.
When omitted variable bias disappears
When the omitted variable has no effect (β₂ = 0) or is uncorrelated with the included regressor (b₂₁ = 0).
OVB sign: β₂ > 0 and Corr(x₁,x₂) > 0
Upward bias, so OLS overstates the effect of x₁.
OVB sign: β₂ > 0 and Corr(x₁,x₂) < 0
Downward bias, so OLS understates the effect of x₁.
Standardized coefficient relationship
γ₁ = β₁·σₓ/σᵧ, and the standardized intercept equals zero because standardized means are zero.
Problems that break only the standard errors
Heteroskedasticity, serial correlation with exogenous regressors, and within-cluster correlation leave β̂ unbiased but SEs wrong.
Problems that break the coefficients
Omitted variables, unit heterogeneity correlated with x, measurement error in x, simultaneity, and lagged y with AR(1) errors.
Heteroskedasticity
The error variance depends on the regressors, Var(u | x) = σᵢ², violating MLR.5.
Effect of heteroskedasticity on OLS coefficients
OLS remains unbiased and consistent, and R² is unaffected.
Effect of heteroskedasticity on inference
OLS is no longer BLUE and conventional standard errors, t, and F statistics are invalid.
OLS variance under heteroskedasticity (8.2)
Var(β̂₁) = Σ(xᵢ − x̄)²σᵢ² / SSTₓ², which reduces to σ²/SSTₓ when variances are equal.
White robust variance (8.3)
Replace each unknown σᵢ² with the squared residual ûᵢ², giving Σ(xᵢ − x̄)²ûᵢ² / SSTₓ².
Why White's estimator works with noisy ûᵢ²
Only the weighted sum is needed, and averaging over many observations washes out the noise in each ûᵢ².
Breusch-Pagan test
Regress squared residuals on the regressors and test joint significance with LM = nR² ~ χ²ₖ or an F test.
Null hypothesis of heteroskedasticity tests
H₀: Var(u | x₁,…,xₖ) = σ², meaning homoskedasticity (MLR.5 holds).
White test
Regress squared residuals on the regressors, their squares, and cross products, then use nR² as a χ² statistic.
Number of regressors in a White test with k = 3
Nine: three levels, three squares, and three pairwise cross products.
Why a dummy reduces White test degrees of freedom
A dummy squared equals itself, so its squared term duplicates the level and is dropped.
Power tradeoff between BP and White tests
BP has high power against linear forms, White catches any form but has lower power against a specific one.
Logging the dependent variable
Often removes heteroskedasticity because proportional errors become additive errors in logs.
Generalized least squares for heteroskedasticity
Divide every term by σᵢ so the transformed error has variance one, making OLS on it BLUE.
Weighted least squares
When Var(u|x) = σ²h(x), divide by √h so noisier observations receive less weight.
Intercept in the WLS-transformed equation
β₀ becomes the coefficient on 1/√hᵢ, so the transformed model has no ordinary intercept.
Feasible GLS (FGLS) for heteroskedasticity
Estimate the variance function from squared residuals, then run WLS with the estimated weights.
Properties of FGLS
Not unbiased, but consistent and asymptotically more efficient than OLS if the variance model is correct.
OLS and FGLS estimates differ greatly
A red flag for misspecification, since both should converge to the same β if the model is right.
Linear probability model variance
Var(y | x) = P(x)[1 − P(x)], so the LPM is heteroskedastic by construction.
Clustered standard errors
SEs allowing any variance and correlation within a group while assuming independence across groups.
Assumption ladder for Var(u|X)
Conventional assumes σ²I, White allows diagonal unequal variances, cluster allows block-diagonal within-unit correlation.
What clustering adds to White's formula
It keeps the within-unit cross terms ûᵢₜûᵢₛxᵢₜxᵢₛ' for t ≠ s that White's method drops.
Moulton factor
V_cluster ≈ [1 + (T − 1)ρᵤρₓ]·V_conventional for equal-size clusters.
Moulton example with T = 10 and ρᵤ = ρₓ = 0.6
The variance rises by a factor of 4.24, so the standard error roughly doubles.
Clustering's asymptotics
Clustered SEs are consistent as the number of clusters grows, not as total observations grow.
Few clusters problem
With fewer than about 30–50 clusters, clustered SEs are biased downward and tests over-reject.
Remedy for few clusters
Use t critical values with n − 1 degrees of freedom or the wild cluster bootstrap.
Correct clustering level for a state policy
Cluster by state, the level at which the key regressor varies and errors are correlated over time.
Why clustering by year fails in a state-year panel
It still treats each state's errors as independent over time, which is the actual problem.
Bertrand, Duflo and Mullainathan placebo laws
Conventional SEs called up to 45% of fake state laws significant, while clustering by state restored about 5%.
Time series data
One unit observed in a meaningful fixed order, a single realization of a stochastic process.
Stochastic process
A random process unfolding over time, {yₜ: t = 1,2,…}.
TS.3 strict exogeneity
E(uₜ | X) = 0 for all t, the error is uncorrelated with regressors in every past and future period.
Contemporaneous exogeneity
E(uₜ | xₜ) = 0, the error is uncorrelated only with same-period regressors.
Strict exogeneity versus contemporaneous exogeneity
Strict exogeneity is needed for unbiasedness, contemporaneous exogeneity is enough for consistency.
Feedback violation of strict exogeneity
A crime shock today raises next year's police hiring, so uₜ is correlated with policeₜ₊₁.
Lagged dependent variable and strict exogeneity
Strict exogeneity always fails because uₜ is part of yₜ, which is a regressor in period t + 1.
TS.5 no serial correlation
Corr(uₜ, uₛ | X) = 0 for all t ≠ s, errors from different periods are uncorrelated.
Why TS.5 has no cross-sectional counterpart
Random sampling delivered independence automatically in cross sections, but ordered time series rarely have it.
What TS.1–TS.6 deliver
Unbiasedness, BLUE, and exact t and F distributions in any sample size.
Static model
yₜ = β₀ + β₁zₜ + uₜ, where β₁ measures only the immediate same-period effect.
Okun's law
Δunemₜ = β₀ + β₁growthₜ + uₜ, with β₁ ≈ −0.5 by Okun's original rule of thumb.
Okun speed limit
The growth rate that holds unemployment steady, g* = β̂₀/(−β̂₁), about 3.15% for the U.S.
Finite distributed lag model
yₜ = α₀ + δ₀zₜ + δ₁zₜ₋₁ + … + δqzₜ₋q + uₜ, letting z affect y over several periods.
Impact propensity
δ₀, the immediate change in y from a one-unit change in z.
Lag coefficient δⱼ
The change in y j periods after a temporary one-unit increase in z.
Long run propensity
LRP = δ₀ + δ₁ + … + δq, the total change in y from a permanent one-unit increase in z.
Why individual lag coefficients are imprecise
Lags of the same variable are highly collinear, so OLS cannot tell which period an effect arrives in.
Why the LRP is precisely estimated
Collinear lag estimates have negative covariances that cancel when summed, shrinking Var(LRP).
LRP reparameterization
Substitute δ₀ = θ₀ − δ₁ − δ₂ so y = β₀ + θ₀zₜ + δ₁(zₜ₋₁ − zₜ) + δ₂(zₜ₋₂ − zₜ) + u reports θ₀ directly.
Spurious regression from trends
Two series trending for unrelated reasons show a large t and high R², with time as the omitted variable.
Fix for trending series
Include a time trend t as a regressor, which by Frisch-Waugh-Lovell is equivalent to detrending all variables.
Exponential trend coefficient
In ln(yₜ) = α₀ + α₁t + eₜ, α₁ is the average growth rate per period.
Detrended R²
1 − SSR/SST of detrended y, measuring how much non-trend variation the regressors explain.
Seasonal dummies
Indicators for all but one month or quarter, equivalent to deseasonalizing every variable.
Seasonal dummies on seasonally adjusted data
They accomplish nothing because the seasonal pattern was already removed.
Exact percent effect of a dummy in a log model
100(e^δ − 1), with 100δ only an approximation for small δ.
Stationary process
A process whose joint distributions are unchanged when shifted in time.
Covariance stationary
Constant mean, constant variance, and covariances that depend only on the lag h, not on t.
Weakly dependent series
A series where Corr(xₜ, xₜ₊ₕ) → 0 as h → ∞, so it eventually forgets its past.
Why weak dependence matters
It replaces random sampling so the law of large numbers and central limit theorem still apply.
MA(1) process
xₜ = eₜ + α₁eₜ₋₁ with white-noise e, which is weakly dependent for any α₁.
MA(1) variance
Var(xₜ) = (1 + α₁²)σₑ².
MA(1) first autocorrelation
Corr(xₜ, xₜ₊₁) = α₁/(1 + α₁²), and correlations at lags two and higher equal zero.
AR(1) process
yₜ = ρ₁yₜ₋₁ + eₜ, today's value is a geometrically weighted sum of all past shocks.
AR(1) stability condition
|ρ₁| < 1, required for the AR(1) to be stable, covariance stationary, and weakly dependent.
AR(1) variance
Var(yₜ) = σₑ²/(1 − ρ₁²), derived from the geometric series 1 + ρ² + ρ⁴ + ….
AR(1) autocorrelation at lag h
Corr(yₜ, yₜ₊ₕ) = ρ₁ʰ, decaying geometrically.
Half-life of a shock in an AR(1)
h = ln(0.5)/ln(ρ), the periods until a shock's effect is halved.
Random walk
yₜ = yₜ₋₁ + eₜ = y₀ + e₁ + … + eₜ, an AR(1) with ρ₁ = 1.
Random walk variance
Var(yₜ) = tσₑ², growing without bound, so the series is not covariance stationary.
Random walk correlation
Corr(yₜ, yₜ₊ₕ) = √(t/(t + h)), which decays so slowly the series is not weakly dependent.
Random walk with drift
yₜ = α₀ + yₜ₋₁ + eₜ, with a linear trend in the mean and variance growing with t.
Adding a trend to a random walk with drift
Removes the drift but not the persistence, so the series is still not weakly dependent.
I(1) process
A series that is not weakly dependent but whose first difference is, also called a unit root process.
I(0) process
A weakly dependent series that is already suitable for OLS.
Typical I(1) series
Price levels, nominal interest rates, exchange rates, GDP, and the money supply.
Typical I(0) series
Growth rates, returns, the unemployment rate, and changes in inflation.
Informal unit root warning sign
A first-order autocorrelation near one, with 0.9 already alarming because ρ̂ is biased toward zero.
Cost of estimating I(1) series in levels
Spurious regression with meaningless t statistics, which is worse than the efficiency loss of overdifferencing.
Spurious regression symptoms with random walks
Too many rejections, a decent R², and a Durbin-Watson statistic near zero.
TS.1′ in the asymptotic assumptions
The series are stationary and weakly dependent and follow a linear model.