Chapter 5 Regression with a Single Regressor: Hypothesis Tests and Confidence Intervals Regression with a Single Regressor: Hypothesis Tests and Confidence Intervals (SW Chapter 5) Overview Now that we have the sampling distribution of OLS estimator, we are ready to perform hypothesis tests about 1 and to construct confidence intervals about 1 Also, we will cover some loose ends about regression: Regression when X is binary (0/1) Heteroskedasticity and homoskedasticity (this is new) Efficiency of the OLS estimator (also new) Use of the t-statistic in hypothesis testing (new but not surprising) 2 But first a big picture view

(and review) We want to learn about the slope of the population regression line, using data from a sample (so there is sampling uncertainty). There are four steps towards this goal: 1. State precisely the population object of interest 2. Derive the sampling distribution of an estimator (this requires certain assumptions) 3. Estimate the variance of the sampling distribution (which the CLT tells us is all you need to know if n is large) that is, finding the standard error (SE) of the estimator using only the information in the sample at hand! 4. Use the estimator ( 1 ) to obtain a point estimate and, with its SE, hypothesis tests, and confidence intervals. 3 Object of interest: 1 in, Yi = 0 + 1Xi + ui, i = 1,, n 1 = Y/X, for an autonomous change in X (causal effect) The Least Squares Assumptions: 1. E(u|X = x) = 0. 2. (Xi,Yi), i =1,,n, are i.i.d.

3. Large outliers are rare (E(X4) < , E(Y4) < . The Sampling Distribution of 1 : Under the LSAs, for n large, is approximately distributed, 1 2 1 ~ N 1 , v4 n X , where vi = (Xi X)ui 4 Hypothesis Testing and the Standard Error of 1 (Section 5.1) The objective is to test a hypothesis, like 1 = 0, using data to reach a tentative conclusion whether the (null) hypothesis is correct or incorrect.

General setup Null hypothesis and two-sided alternative: H0: 1 = 1,0 vs. H1: 1 1,0 where 1,0 is the hypothesized value under the null. Null hypothesis and one-sided alternative: H0: 1 = 1,0 vs. H1: 1 < 1,0 5 General approach: construct t-statistic, and compute p-value (or compare to N(0,1) critical value) In general: estimator - hypothesized value t= standard error of the estimator where the SE of the estimator is the square root of an estimator of the variance of the estimator. Y Y ,0 For testing the mean of Y: t= sY / n

1 1,0 For testing 1, t= , SE ( 1 ) where SE( 1 ) = the square root of an estimator of the variance of the sampling distribution of 1 6 Formula for SE(1) Recall the expression for the variance of 1 (large n): 2 var[( X ) u ]

v i x i var( 1 ) = = , where vi = (Xi X)ui. 4 2 2 n( X ) n X The estimator of the variance of replaces the unknown 1 population values of 2 and X4 by estimators constructed from the data: 1 n 2 vi 2 1 1

estimator of v n 2 i 1 2 = = 2 2 2 1 n (estimator of X ) n 1 n 2 n ( Xi X ) i 1 where vi = ( X i X )ui . 7 1 n 2 vi n 2 i 1

1 = , where vi = ( X i X )ui . 2 n 1 n 2 n ( Xi X ) i 1 SE( ) = 2 = the standard error of 2 1 1 1 1 OK, this is a bit nasty, but: It is less complicated than it seems. The numerator estimates

var(v), the denominator estimates var(X). Why the degrees-of-freedom adjustment n 2? Because two coefficients have been estimated (0 and 1). SE( ) is computed by regression software 1 STATA has memorized this formula so you dont need to. 8 Summary: To test H0: 1 = 1,0 v. H1: 1 1,0, Construct the t-statistic 1 1,0 1 1,0 t= = SE ( 1 ) 2 1 Reject at 5% significance level if |t| > 1.96 The p-value is p = Pr[|t| > |tact|] = probability in tails of normal outside |tact|; you reject at the 5% significance level if

the p-value is < 5%. This procedure relies on the large-n approximation; typically n = 50 is large enough for the approximation to be excellent. 9 Example: Test Scores and STR, California data Estimated regression line: TestScore = 698.9 2.28 STR Regression software reports the standard errors: SE( 0 ) = 10.4 SE( 1 ) = 0.52 1 1,0 2.28 0 t-statistic testing 1,0 = 0 = = = 4.38 0.52

SE ( 1 ) The 1% 2-sided significance level is 2.58, so we reject the null at the 1% significance level. Alternatively, we can compute the p-value 10 The p-value based on the large-n standard normal approximation to the t-statistic is 0.00001 (105) 11 Confidence Intervals for 1 (Section 5.2)) Recall that a 95% confidence is, equivalently: The set of points that cannot be rejected at the 5% significance level; A set-valued function of the data (an interval that is a function of the data) that contains the true parameter value 95% of the time in repeated samples. Because the t-statistic for 1 is N(0,1) in large samples, construction of a 95% confidence for 1 is just like the case of the sample mean:

95% confidence interval for 1 = { 1.96 SE( )} 1 1 12 Confidence interval example: Test Scores and STR Estimated regression line: TestScore = 698.9 2.28 STR SE( 0 ) = 10.4 SE( 1 ) = 0.52 95% confidence interval for 1 : { 1 1.96 SE( 1 )} = {2.28 1.96 0.52} = (3.30, 1.26) The following two statements are equivalent (why?) The 95% confidence interval does not include zero; The hypothesis 1 = 0 is rejected at the 5% level 13

A concise (and conventional) way to report regressions: Put standard errors in parentheses below the estimated coefficients to which they apply. = 698.9 2.28 STR, R2 = .05, SER = 18.6 TestScore (10.4) (0.52) This expression gives a lot of information The estimated regression line is = 698.9 2.28 STR TestScore The standard error of 0 is 10.4 The standard error of is 0.52 1 The R2 is .05; the standard error of the regression is 18.6 14 OLS regression: reading STATA

output regress testscr str, robust Regression with robust standard errors Number of obs = 420 F( 1, 418) = 19.26 Prob > F = 0.0000 R-squared = 0.0512 Root MSE = 18.581 ------------------------------------------------------------------------| Robust testscr | Coef. Std. Err. t

P>|t| [95% Conf. Interval] --------+---------------------------------------------------------------str | -2.279808 .5194892 -4.38 0.000 -3.300945 -1.258671 _cons | 698.933 10.36436 67.44 0.000 678.5602 719.3057 ------------------------------------------------------------------------- so: = 698.9 2.28 STR, , R2 = .05, SER = 18.6 TestScore

(10.4) (0.52) t (1 = 0) = 4.38, p-value = 0.000 (2-sided) 95% 2-sided conf. interval for 1 is (3.30, 1.26) 15 Summary of Statistical Inference about 0 and 1: Estimation: OLS estimators 0 and 1 and have approximately normal sampling distributions 0 1 in large samples Testing: H0: 1 = 1,0 v. 1 1,0 (1,0 is the value of 1 under H0)

t = ( 1,0)/SE( ) 1 1 p-value = area under standard normal outside tact (large n) Confidence Intervals: 95% confidence interval for 1 is { 1.96 SE( )} 1 1 This is the set of 1 that is not rejected at the 5% level The 95% CI contains the true 1 in 95% of all samples. 16 Regression when X is Binary (Section 5.3) Sometimes a regressor is binary: X = 1 if small class size, = 0 if not

X = 1 if female, = 0 if male X = 1 if treated (experimental drug), = 0 if not Binary regressors are sometimes called dummy variables. So far, 1 has been called a slope, but that doesnt make sense if X is binary. How do we interpret regression with a binary regressor? 17 Interpreting regressions with a binary regressor Yi = 0 + 1Xi + ui, where X is binary (Xi = 0 or 1): When Xi = 0, Yi = 0 + ui the mean of Yi is 0 that is, E(Yi|Xi=0) = 0 When Xi = 1, Yi = 0 + 1 + ui the mean of Yi is 0 + 1 that is, E(Yi|Xi=1) = 0 + 1 so: 1 = E(Yi|Xi=1) E(Yi|Xi=0) = population difference in group means 18

Example: 1 if STRi 20 Let Di = 0 if STRi 20 = 650.0 + 7.4 D TestScore (1.3) (1.8) Tabulation of group means: Class Size Average score (Y ) Std. dev. (sY) Small (STR > 20) 657.4 19.4 Large (STR 20) 17.9 650.0 OLS regression:

N 238 182 Difference in means: Ysmall Ylarge = 657.4 650.0 = 7.4 Standard error: ss2 sl2 19.42 17.92 SE = = = 1.8 238 182 ns nl 19 Summary: regression when Xi is binary (0/1)

Yi = 0 + 1Xi + ui 0 = mean of Y when X = 0 0 + 1 = mean of Y when X = 1 1 = difference in group means, X =1 minus X = 0 SE( ) has the usual interpretation 1 t-statistics, confidence intervals constructed as usual This is another way (an easy way) to do difference-in-means analysis The regression formulation is especially useful when we have additional regressors (as we will very soon) 20 Heteroskedasticity and Homoskedasticity, and Homoskedasticity-Only Standard Errors (Section 5.4) What? Consequences of homoskedasticity Implication for computing standard errors What do these two terms mean?

If var(u|X=x) is constant that is, if the variance of the conditional distribution of u given X does not depend on X then u is said to be homoskedastic. Otherwise, u is heteroskedastic. 21 Example: hetero/homoskedasticity in the case of a binary regressor (that is, the comparison of means) Standard error when group variances are unequal: ss2 sl2 SE = ns nl Standard error when group variances are equal: SE = s p 1 1 ns nl 2 2

( n 1) s ( n 1) s s l l where s 2p = s (SW, Sect 3.6) ns nl 2 sp = pooled estimator of 2 when l2 = s2 Equal group variances = homoskedasticity Unequal group variances = heteroskedasticity 22

Homoskedasticity in a picture: E(u|X=x) = 0 (u satisfies Least Squares Assumption #1) The variance of u does not depend on x 23 Heteroskedasticity in a picture: E(u|X=x) = 0 (u satisfies Least Squares Assumption #1) The variance of u does depends on x: u is heteroskedastic. 24 A real-data example from labor economics: average hourly earnings vs. years of education (data source: Current Population Survey): Heteroskedastic or homoskedastic? 25 The class size data:

Heteroskedastic or homoskedastic? 26 So far we have (without saying so) assumed that u might be heteroskedastic. Recall the three least squares assumptions: 1. E(u|X = x) = 0 2. (Xi,Yi), i =1,,n, are i.i.d. 3. Large outliers are rare Heteroskedasticity and homoskedasticity concern var(u|X=x). Because we have not explicitly assumed homoskedastic errors, we have implicitly allowed for heteroskedasticity. 27 What if the errors are in fact homoskedastic? You can prove that OLS has the lowest variance among estimators that are linear in Y a result called the GaussMarkov theorem that we will return to shortly. The formula for the variance of and the OLS standard 1 error simplifies (pp. 4.4): If var(ui|Xi=x) = u2 , then

2 2 E [( X ) ui ] var[( X ) u ] i x i x i var( 1 ) =

= 2 2 n( X ) n ( X2 ) 2 u2 = n X2 Note: var( 1 ) is inversely proportional to var(X): more spread in X means more information about 1 - we discussed this earlier but it is clearer from this formula. 28 Along with this homoskedasticity-only formula for the variance of , we have homoskedasticity-only standard 1 errors: Homoskedasticity-only standard error formula: SE( 1 ) =

1 n 2 ui n 2 i 1 1 . n n 1 2 ( X X ) i n i 1 Some people (e.g. Excel programmers) find the homoskedasticity-only formula simpler. 29

We now have two formulas for standard errors for 1 Homoskedasticity-only standard errors these are valid only if the errors are homoskedastic. The usual standard errors to differentiate the two, it is conventional to call these heteroskedasticity robust standard errors, because they are valid whether or not the errors are heteroskedastic. The main advantage of the homoskedasticity-only standard errors is that the formula is simpler. But the disadvantage is that the formula is only correct in general if the errors are homoskedastic. 30 Practical implications The homoskedasticity-only formula for the standard error of and the heteroskedasticity-robust formula differ so in 1 general, you get different standard errors using the different formulas.

Homoskedasticity-only standard errors are the default setting in regression software sometimes the only setting (e.g. Excel). To get the general heteroskedasticity-robust standard errors you must override the default. If you dont override the default and there is in fact heteroskedasticity, your standard errors (and wrong tstatistics and confidence intervals) will be wrong typically, homoskedasticity-only SEs are too small. 31 Heteroskedasticity-robust standard errors in STATA regress testscr str, robust Regression with robust standard errors Number of obs = 420 F( 1, 418) = 19.26 Prob > F = 0.0000

R-squared = 0.0512 Root MSE = 18.581 ------------------------------------------------------------------------| Robust testscr | Coef. Std. Err. t P>|t| [95% Conf. Interval] --------+---------------------------------------------------------------str | -2.279808 .5194892 -4.39 0.000 -3.300945 -1.258671 _cons | 698.933 10.36436 67.44

0.000 678.5602 719.3057 ------------------------------------------------------------------------- If you use the , robust option, STATA computes heteroskedasticity-robust standard errors Otherwise, STATA computes homoskedasticity-only standard errors 32 The bottom line: If the errors are either homoskedastic or heteroskedastic and you use heteroskedastic-robust standard errors, you are OK If the errors are heteroskedastic and you use the homoskedasticity-only formula for standard errors, your standard errors will be wrong (the homoskedasticity-only estimator of the variance of is inconsistent if there is 1 heteroskedasticity). The two formulas coincide (when n is large) in the special

case of homoskedasticity So, you should always use heteroskedasticity-robust standard errors. 33 Some Additional Theoretical Foundations of OLS (Section 5.5) We have already learned a very great deal about OLS: OLS is unbiased and consistent; we have a formula for heteroskedasticity-robust standard errors; and we can construct confidence intervals and test statistics. Also, a very good reason to use OLS is that everyone else does so by using it, others will understand what you are doing. In effect, OLS is the language of regression analysis, and if you use a different estimator, you will be speaking a different language. 34 Still, some of you may have further questions: Is this really a good reason to use OLS? Arent there other estimators that might be better in particular, ones that might

have a smaller variance? Also, what ever happened to our old friend, the Student t distribution? So we will now answer these questions but to do so we will need to make some stronger assumptions than the three least squares assumptions already presented. 35 The Extended Least Squares Assumptions These consist of the three LS assumptions, plus two more: 1. E(u|X = x) = 0. 2. (Xi,Yi), i =1,,n, are i.i.d. 3. Large outliers are rare (E(Y4) < , E(X4) < ). 4. u is homoskedastic 5. u is distributed N(0,2) Assumptions 4 and 5 are more restrictive so they apply to fewer cases in practice. However, if you make these assumptions, then certain mathematical calculations simplify and you can prove strong results results that hold if these additional assumptions are true. We start with a discussion of the efficiency of OLS

36 Efficiency of OLS, part I: The Gauss-Markov Theorem Under extended LS assumptions 1-4 (the basic three, plus homoskedasticity), has the smallest variance among all linear 1 estimators (estimators that are linear functions of Y1,, Yn). This is the Gauss-Markov theorem. Comments The GM theorem is proven in SW Appendix 5.2 37 The Gauss-Markov Theorem, ctd. 1 is a linear estimator, that is, it can be written as a linear function of Y1,, Yn: n 1 1 =

(X i X )ui i 1 n 2 ( X X ) i 1 n = wi ui , n i 1 i 1

where wi = (Xi X ) n 1 2 ( X X ) i n i 1 . The G-M theorem says that among all possible choices of {wi}, the OLS weights yield the smallest var( ) 1

38 Efficiency of OLS, part II: Under all five extended LS assumptions including normally distributed errors 1 has the smallest variance of all consistent estimators (linear or nonlinear functions of Y1,,Yn), as n . This is a pretty amazing result it says that, if (in addition to LSA 1-3) the errors are homoskedastic and normally distributed, then OLS is a better choice than any other consistent estimator. And because an estimator that isnt consistent is a poor choice, this says that OLS really is the best you can do if all five extended LS assumptions hold. (The proof of this result is beyond the scope of this course and isnt in SW it is typically done in graduate courses.) 39 Some not-so-good thing about OLS The foregoing results are impressive, but these results and the OLS estimator have important limitations. 1. The GM theorem really isnt that compelling: The condition of homoskedasticity often doesnt hold

(homoskedasticity is special) The result is only for linear estimators only a small subset of estimators (more on this in a moment) 2. The strongest optimality result (part II above) requires homoskedastic normal errors not plausible in applications (think about the hourly earnings data!) 40 Limitations of OLS, ctd. 3. OLS is more sensitive to outliers than some other estimators. In the case of estimating the population mean, if there are big outliers, then the median is preferred to the mean because the median is less sensitive to outliers it has a smaller variance than OLS when there are outliers. Similarly, in regression, OLS can be sensitive to outliers, and if there are big outliers other estimators can be more efficient (have a smaller variance). One such estimator is the least absolute deviations (LAD) estimator: n min b0 ,b1 Yi (b0 b1 X i ) i 1

In virtually all applied regression analysis, OLS is used and that is what we will do in this course too. 41 Inference if u is Homoskedastic and Normal: the Student t Distribution (Section 5.6) Recall the five extended LS assumptions: 1. E(u|X = x) = 0. 2. (Xi,Yi), i =1,,n, are i.i.d. 3. 4. Large outliers are rare (E(Y4) < , E(X4) < ). u is homoskedastic 5. u is distributed N(0,2)

If all five assumptions hold, then: 0 and 1 are normally distributed for all n (!) the t-statistic has a Student t distribution with n 2 degrees of freedom this holds exactly for all n (!) 42 Normality of the sampling distribution of 1 under 15: n 1 1 = (X i X )ui i 1 n

2 ( X X ) i i 1 (Xi X ) 1 n = wi ui , where wi = . n 1 n i 1 2 ( X X )

i n i 1 What is the distribution of a weighted average of normals? Under assumptions 1 5: n 1 2 2 (*) 1 1 ~ N 0, 2 wi u n i 1 Substituting wi into (*) yields the homoskedasticity-only variance formula. 43 In addition, under assumptions 1 5, under the null hypothesis the t statistic has a Student t distribution with n 2 degrees of

freedom Why n 2? because we estimated 2 parameters, 0 and 1 For n < 30, the t critical values can be a fair bit larger than the N(0,1) critical values For n > 50 or so, the difference in tn2 and N(0,1) distributions is negligible. Recall the Student t table: degrees of freedom 10 20 30 60 5% t-distribution critical value 2.23 2.09 2.04 2.00 1.96 44 Practical implication:

If n < 50 and you really believe that, for your application, u is homoskedastic and normally distributed, then use the tn2 instead of the N(0,1) critical values for hypothesis tests and confidence intervals. In most econometric applications, there is no reason to believe that u is homoskedastic and normal usually, there is good reason to believe that neither assumption holds. Fortunately, in modern applications, n > 50, so we can rely on the large-n results presented earlier, based on the CLT, to perform hypothesis tests and construct confidence intervals using the large-n normal approximation. 45 Summary and Assessment (Section 5.7) The initial policy question: Suppose new teachers are hired so the student-teacher ratio falls by one student per class. What is the effect of this policy intervention (treatment) on test scores? Does our regression analysis answer this convincingly? Not really districts with low STR tend to be ones with

lots of other resources and higher income families, which provide kids with more learning opportunities outside schoolthis suggests that corr(ui, STRi) > 0, so E(ui|Xi) 0. So, we have omitted some factors, or variables, from our analysis, and this has biased our results. 46