OLS Asymptotics YSS2211 Introductory Econometrics Lecture 10 Ran Song Yale-NUS College, Singapore Semester 1, 2022/23 Semester 1, 2022/23 1 / 11 OLS Asymptotics Asymptotics - Introduction In the last few sections we covered finite sample (a.k.a. small sample, or exact) properties of the OLS estimators. Now interested in what happens when n → ∞. When we added the assumption that the error terms were normally distributed (MLR 6), we were able to derive exact sampling (for finite n) distributions of the OLS estimators. If the error is not normally distributed, we can no longer say that the distribution of the t statistic is exactly t or that the F statistic has an exact F distribution for any sample size. Semester 1, 2022/23 2 / 11 OLS Asymptotics Consistency In estimation, we are typically constrained by the amount of information (in the form of a random sample) we have to compute β̂ –we make a sample estimate of an unknown parameter β. When looking at consistency, we are interested in the thought experiment - if you are not constrained by the availability of data (n → ∞), what happens to the sampling distribution of our estimator. Semester 1, 2022/23 3 / 11 OLS Asymptotics Consistency -plim Definition An estimator θ̂n = f (Y1 , . . . , Yn ) is a consistent of θ if for every ϵ > 0, lim P{|θ̂n − θ | > ϵ} = 0 n→ ∞ If θ̂n is not consistent for θ, then we say it is inconsistent. When θ̂n is consistent, we also say that θ is the probability limit of θ̂n , written as plim(θ̂n ) = θ. Unlike unbiasedness which is a feature of an estimator for a given sample size - consistency involves the behavior of the sampling distribution of the estimator as the sample size n gets large. Semester 1, 2022/23 4 / 11 OLS Asymptotics Consistency -plim continues We want the limn→∞ Var(θ̂n ) = 0, and limn→∞ Bias(θ̂n ) = 0. But Var(θ̂n ) does not always exist. We want the sampling distribution to become more concentrated on the true value θ as n increases. Semester 1, 2022/23 5 / 11 OLS Asymptotics Consistency - plim continues The earliest and simplest demonstration of consistency is the Weak Law of Large Numbers. Let Y1 , Y2 , . . . , Yn be independent, identically distributed random variables with mean µ and Var(Yi ) = σ2 < ∞. Then, plimȲ = µ Notice that in the case of the sample mean, Bias(Ȳ) = EȲ − µ = 0, Var(Ȳ) = σ2 /n Hence easy to show that limn→∞ Var(Ȳ) = 0, and limn→∞ Bias(Ȳ) = 0. Semester 1, 2022/23 6 / 11 OLS Asymptotics Consistency - OLS Estimator Theorem 5.1 If MLR1 to MLR 4 holds, β̂ j is a consistent estimator for β j for all j = 1, 2 . . . , k Consistency is a minimum quality/property we expect of our estimators. “If obtaining more and more data does not generally get us closer to the parameter value of interest, then we are using a poor estimation procedure.” - page 169, Wooldridge. Recall the case of omitted variable bias - we can show that β̂ 1 = β 1 + 1 n n ∑i=1 (x1i − x̄1 )ui 1 n 2 n ∑i=1 (x1i − x̄1 ) cov(x ,u So plim( β̂ 1 ) = β 1 + var(x1i ) i . that is, OLS is inconsistent 1 when a relevant regressor is omitted from the regression model. Semester 1, 2022/23 7 / 11 OLS Asymptotics Some properties of plim Supposed plim(Wn ) = θ, and g(·) is some continuous function. 1. plim g(Wn ) = g(plim(Wn )). 2. If plim(Tn ) = α, and plim(Un ) = β, where β ̸= 0 then, Tn a) plim U = αβ n b) plim(Tn + Un ) = α + β c) plim(Tn Un ) = αβ Semester 1, 2022/23 8 / 11 OLS Asymptotics Applying plim rules Examples Consider the sample variance estimator s2n = n−1 1 ∑(Yi − Ȳ)2 . We know that plim s2n = σ2 and Es2n = σ2 . Suppose we are interested in an estimator for σ, we know that q q 2 E sn ̸= Es2n = σ But, we know from Rule 1 of plim that q q √ plim s2n = plim s2n = σ2 = σ Semester 1, 2022/23 9 / 11 OLS Asymptotics Applying plim rules continues Examples Consider Ȳ as the sample proportion of successes in n trials. It is an estimator of θ, the probability of success in each trial. Suppose we are interested in the odds ratio 1−θ θ , a natural estimator is γ̂ = 1−ȲȲ . This is a biased estimator since Eγ̂ ̸= θ . 1−θ But it is a consistent estimator plimγ̂ = Semester 1, 2022/23 plimȲ θ = 1−θ plim(1 − Ȳ) 10 / 11 OLS Asymptotics Asymptotic Normality of the OLS Estimator continues Even when u are not normally distributed (MLR6 does not hold), we can show that the OLS estimators satisfy asymptotic normality Theorem Under the Gauss-Markov Assumptions MLR. 1 - MLR. 5: β̂ j − β j s.e.( β̂ j ) ∼ N (0, 1) where j = 1, 2, . . . , k and s.e.( β̂ j ) is the usual OLS standard error. The asymptotic normality theorem says that even if we do not know the distribution of the error terms, β̂ j is still normally distributed in large samples, and that the standardized β̂ j goes to a standard normal distribution in large samples . Semester 1, 2022/23 11 / 11