ECON Introductory Econometrics. Lecture 2: Review of Statistics

Transcription

1 ECON415 - Introductory Econometrics Lecture 2: Review of Statistics Monique de Haan (moniqued@econ.uio.no) Stock and Watson Chapter 2-3

2 Lecture outline 2 Simple random sampling Distribution of the sample average Large sample approximation to the distribution of the sample mean Law of large numbers central limit theorem Estimation of the population mean unbiasedness consistency efficiency Hypothesis test concerning the population mean Confidence intervals for the population mean

3 Simple random sampling 3 Simple random sampling means that n objects are drawn randomly from a population and each object is equally likely to be drawn Let Y 1, Y 2,..., Y n denote the 1st to the nth randomly drawn object. Under simple random sampling: The marginal probability distribution of Y i is the same for all i = 1, 2,.., n and equals the population distribution of Y. because Y 1, Y 2,..., Y n are drawn randomly from the same population. Y 1 is distributed independently from Y 2,..., Y n knowing the value of Y i does not provide information on Y j for i j When Y 1,..., Y n are drawn from the same population and are independently distributed, they are said to be i.i.d random variables

4 Simple random sampling: Example 4 Let G be the gender of an individual (G = 1 if female, G = if male) G is a Bernoulli random variable with E (G) = µ G = Pr(G = 1) =.5 Suppose we take the population register and randomly draw a sample of size n The probability distribution of G i is a Bernoulli distribution with mean.5 G 1 is distributed independently from G 2,..., G n Suppose we draw a random sample of individuals entering the building of the physics department This is not a sample obtained by simple random sampling and G 1,..., G n are not i.i.d Men are more likely to enter the building of the physics department!

5 The sampling distribution of the sample average 5 The sample average Ȳ of a randomly drawn sample is a random variable with a probability distribution called the sampling distribution. Ȳ = 1 n (Y 1 + Y Y n) = 1 n Y i n Suppose Y 1,..., Y n are i.i.d and the mean & variance of the population distribution of Y are respectively µ Y & σ 2 Y The mean of Y is E ( Ȳ ) = E ( 1 n ) n Y i = 1 n i=1 i=1 n E (Y i ) = 1 n ne(y ) = µ Y The variance of Y is ( ) Var Y = Var ( 1 n n i=1 Y ) i i=1 = 1 n n 2 i=1 Var (Y i) n n n 2 i=1 j=1,j i Cov(Y i, Y j ) = = 1 n 2 nvar (Y ) + 1 n σ2 Y

6 The sampling distribution of the sample average:example 6 Let G be the gender of an individual (G = 1 if female, G = if male) The mean of the population distribution of G is E (G) = µ G = p =.5 The variance of the population distribution of G is Var (G) = σ 2 G = p(1 p) =.5(1 5) =.25 The mean and variance of the average gender (proportion of women) G in a random sample with n = 1 are ( ) E G = µ G =.5 Var ( ) G = 1 n σ2 G = 1.25 =.25 1

7 The finite sample distribution of the sample average 7 The finite sample distribution is the sampling distribution that exactly describes the distribution of Y for any sample size n. In general the exact sampling distribution of Y is complicated and depends on the population distribution of Y. A special case is when Y 1, Y 2,..., Y n are i.i.d draws from the N ( ) µ Y, σy 2, because in this case ( ) Y N µ Y, σ2 Y n

8 The finite sample distribution of average gender G 8 Suppose we draw 999 samples of n = 2: Sample 1 Sample 2 Sample 3... Sample 999 G 1 G 2 G G 1 G 2 G G 1 G 2 G G 1 G 2 G Sample distribution of average gender 999 samples of n=2.4 probability sample average.

9 The asymptotic distribution of Y 9 Given that the exact sampling distribution of Y is complicated and given that we generally use large samples in econometrics we will often use an approximation of the sample distribution that relies on the sample being large The asymptotic distribution is the approximate sampling distribution of Y if the sample size n We will use two concepts to approximate the large-sample distribution of the sample average The law of large numbers. The central limit theorem.

10 Law of Large Numbers 1 The Law of Large Numbers states that if Y i, i = 1,.., n are independently and identically distributed with E (Y i ) = µ Y and large outliers are unlikely; Var (Y i ) = σ 2 Y < Y will be near µ Y with very high probability when n is very large (n ) Y p µ Y

11 Law of Large Numbers Example: Gender G Bernouilli (.5,.25) 11.5 Sample distribution of average gender 999 samples of n=2.25 Sample distribution of average gender 999 samples of n=1.4.2 probability.3.2 probability sample average sample average.1 Sample distribution of average gender 999 samples of n=1.6 Sample distribution of average gender 999 samples of n=25 probability sample average probability sample average

12 The Central Limit theorem 12 The Central Limit Theorem states that if Y i, i = 1,.., n are i.i.d. with E (Y i ) = µ Y and Var (Y i ) = σ 2 Y with < σ 2 Y < The distribution of the sample average is approximately normal if n ( ) Y N µ Y, σ2 Y n The distribution of the standardized sample average is approximately standard normal for n Y µ Y σ 2 Y N (, 1)

13 The Central Limit theorem Example: Gender G Bernouilli (.5,.25) 13.5 Sample distribution of average gender 999 samples of n=2.25 Sample distribution of average gender 999 samples of n=1 probability probability sample average sample average Finite sample distr. standardized sample average Standard normal probability densitiy Finite sample distr. standardized sample average Standard normal probability densitiy probability Sample distribution of average gender 999 samples of n=1 probability Sample distribution of average gender 999 samples of n= sample average sample average Finite sample distr. standardized sample average Standard normal probability densitiy Finite sample distr. standardized sample average Standard normal probability densitiy.

14 The Central Limit theorem 14 How good is the large-sample approximation? If Y i N ( µ Y, σ 2 Y ) the approximation is perfect If Y i is not normally distributed the quality of the approximation depends on how close n is to infinity For n 1 the normal approximation to the distribution of Y is typically very good for a wide variety of population distributions

15 Estimation

16 Estimators and estimates 16 An Estimator is a function of a sample of data to be drawn randomly from a population An estimator is a random variable because of randomness in drawing the sample An Estimate is the numerical value of an estimator when it is actually computed using a specific sample.

17 Estimation of the population mean 17 Suppose we want to know the mean value of Y (µ Y ) in a population, for example The mean wage of college graduates. The mean level of education in Norway. The mean probability of passing the econometrics exam. Suppose we draw a random sample of size n with Y 1,..., Y n i.i.d Possible estimators of µ Y are: The sample average Y = 1 n n i=1 Y i The first observation Y 1 The weighted average: Ỹ = 1 n ( 1 2 Y Y Y n Yn )

18 Estimation of the population mean 18 To determine which of the estimators, Y, Y 1 or Ỹ is the best estimator of µ Y we consider 3 properties: Let ˆµ Y be an estimator of the population mean µ Y. Unbiasedness: The mean of the sampling distribution of ˆµ Y equals µ Y E (ˆµ Y ) = µ Y Consistency: The probability that ˆµ Y is within a very small interval of µ Y approaches 1 if n ˆµ Y p µ Y Efficiency: If the variance of the sampling distribution of ˆµ Y is smaller than that of some other estimator µ Y, ˆµ Y is more efficient Var (ˆµ Y ) < Var ( µ Y )

19 Example 19 Suppose we are interested in the mean wages µ w of individuals with a master degree We draw the following sample (n = 1) by simple random sampling i W i The 3 estimators give the following estimates: W = i=1 W i = W 1 = ( W = 1 1 W W W W ) 2 1 =

20 Unbiasedness 2 All 3 proposed estimators are unbiased: ( ) As shown on slide 5: E Y = µ Y Since Y i are i.i.d. E (Y 1 ) = E (Y ) = µ Y ) E (Ỹ = E ( ( 1 1 Y n Y Y )) 2 n Yn ( 1 2 E(Y 1) E(Y 2) E(Y n 1) E(Yn)) = 1 n = 1 n [( n ) E (Yi ) + ( n ) E (Yi ) ] E (Y i ) = µ Y

21 Consistency Example: mean wages of individuals with a master degree with µ w = 6 21 By the law of large numbers W p µ W which implies that the probability that W is within a very small interval of µ W approaches 1 if n Sample average as estimator of population mean 999 samples of n=1.4 Sample average as estimator of population mean 999 samples of n=1.3 probability sample average probability sample average

22 Consistency Example: mean wages of individuals with a master degree with µ w = 6 22 W = 1 n ( 1 2 W W W n Wn ) is also consistent Weighted average as estimator of population mean 999 samples of n=1.3 Weighted average as estimator of population mean 999 samples of n=1.4 probability weighted average probability However W 1 is not a consistent estimator of µ W : 5 weighted average First observation W1 as estimator of population mean 999 samples of n=1.4 First observation W1 as estimator of population mean 999 samples of n=1.4 probability first observation W probability first observation W

23 Efficiency 23 Efficiency entails a comparison of estimators on the basis of their variance The variance of Y equals Var ( ) Y = 1 n σ2 Y The variance of Y 1 equals The variance of Ỹ equals Var (Y 1 ) = Var (Y ) = σ 2 Y ) Var (Ỹ = n σ2 Y For any n 2 Y is more efficient than Y 1 and Ỹ

24 BLUE: Best Linear Unbiased Estimator 24 Y is not only more efficient than Y 1 and Ỹ, but it is more efficient than any unbiased estimator of µ Y that is a weighted average of Y 1,..., Y n Y is the Best Linear Unbiased Estimator (BLUE) it is the most efficient estimator of µ Y among all unbiased estimators that are weighted averages of Y 1,..., Y n Let ˆµ Y be an unbiased estimator of µ Y ˆµ Y = 1 n n a i Y i i=1 with a 1,...a n nonrandom constants then Y is more efficient than ˆµ Y, that is ( ) Var Y < Var (ˆµ Y )

25 Hypothesis tests concerning the population mean

26 Hypothesis tests concerning the population mean 26 Consider the following questions: Is the mean monthly wage of college graduates equal to NOK 6? Is the mean level of education in Norway equal to 12 years? Is the mean probability of passing the econometrics exam equal to 1? These questions involve the population mean taking on a specific value µ Y, Answering these questions implies using data to compare a null hypothesis H : E (Y ) = µ Y, to an alternative hypothesis, which is often the following two sided hypothesis H 1 : E (Y ) µ Y,

27 Hypothesis tests concerning the population mean p-value 27 Suppose we have a sample of n i.i.d observations and compute the sample average Y The sample average can differ from µ Y, for two reasons 1 The population mean µ Y is not equal to µ Y, (H not true) 2 Due to random sampling Y µ Y = µ Y, (H true) To quantify the second reason we define the p-value The p-value is the probability of drawing a sample with Y at least as far from µ Y, given that the null hypothesis is true.

28 Hypothesis tests concerning the population mean p-value 28 [ ] p value = Pr H Y µ Y, > Y act µ Y, To compute the p-value we need to know the sampling distribution of Y Sampling distribution of Y is complicated for small n With large n the central limit theorem states that ( ) Y N µ Y, σ2 Y n This implies that if the null hypothesis is true: Y µ Y, σ 2 Yn N (, 1)

29 Computing the p-value when σ Y is known p value = Pr H Y µ Y, σ 2 Y n > Y act µ Y, σ 2 Y n = 2Φ Y act µ Y, σ 2 Y n 29 For large n, p-value = the probability that Z falls outside Y act µ Y, σ 2 Y n

30 Estimating the standard deviation of Y 3 In practice σ 2 Y is usually unknown and must be estimated The sample variance s 2 Y is the estimator of σy 2 = E [(Y i µ Y ) 2] s 2 Y = 1 n 1 n i=1 ( ) 2 Y i Y division by n 1 because we replace µ Y by Y which uses up 1 degree of freedom if Y 1,..., Y n are i.i.d. and E ( Y 4) <, sy 2 p σy 2 (Law of Large Numbers) The sample standard deviation s Y = s 2 Y is the estimator of σ Y

31 31 Computing the p-value using SE ( Y ) = σ Y The standard error SE(Y ) is an estimator of σ Y SE ( Y ) = s Y n Because sy 2 is a consistent estimator of σy 2, we can (for large n) replace ( ) by SE Y σ 2 Y n = s Y n This implies that when σy 2 is unknown and Y 1,..., Y n are i.i.d. the p-value is computed as p value = 2Φ Y act µ Y, ( ) SE Y

32 The t-statistic and its large-sample distribution 32 ) ( ) The standardized sample average (Y act µ Y, /SE Y plays a central role in testing statistical hypothesis It has a special name, the t-statistic t = Y µ Y, ( ) SE Y t is approximately N (, 1) distributed for large n The p-value can be computed as ( ) p value = 2Φ t act

33 The t-statistic and its large-sample distribution 33 95% 2.5% 2.5% t

34 Type I and type II errors and the significance level 34 There are 2 types of mistakes when conduction a hypothesis test Type I error refers to the mistake of rejecting H when it is true Type II error refers to the mistake of not rejecting H when it is false In hypothesis testing we usually fix the probability of a type I error The significance level α is the probability of rejecting H when it is true Most often used significance level is 5% (α =.5) Since area in tails of N (, 1) outside ±1.96 is 5%: We reject H if p-value is smaller than.5. We reject H if t act > 1.96

35 4 steps in testing a hypothesis about the population mean 35 H : E (Y ) = µ Y, H 1 : E (Y ) µ Y, Step 1: Compute the sample average Y Step 2: Compute the standard error of Y ( ) SE Y = s Y n Step 3: Compute the t-statistic t act = Y µ Y, ( ) SE Y Step 4: Reject the null hypothesis at a 5% significance level if t act > 1.96 or if p value <.5

36 Hypothesis tests concerning the population mean Example: The mean wage of individuals with a master degree 36 Suppose we would like to test H : E (W ) = 6 H 1 : E (W ) 6 using a sample of 25 individuals with a master degree Step 1: W = 1 n n i=1 W i = ( ) Step 2: SE W = sw n = Step 3: t act = W µ W, SE(W) = = 1.48 Step 4: Since we use a 5% significance level, we do not reject H because t act = 1.48 < 1.96 Note: We do never accept the null hypothesis!

37 Hypothesis tests concerning the population mean Example: The mean wage of individuals with a master degree 37 This Thursday is howjanuary to do the 12 12:17:46 test in Stata: 217 Page 1 (R) / / / / / / / / / / / / Statistics/Data Analysis 1. ttest wage=6 One-sample t test Variable Obs Mean Std. Err. Std. Dev. [95% Conf. Interval] wage mean = mean( wage) t = Ho: mean = 6 degrees of freedom = 249 Ha: mean < 6 Ha: mean!= 6 Ha: mean > 6 Pr(T < t) =.932 Pr( T > t ) =.1396 Pr(T > t) =.698

38 Hypothesis test with a one-sides alternative 38 Sometimes the alternative hypothesis is that the mean exceeds µ Y, H : E (Y ) = µ Y, H 1 : E (Y ) > µ Y, In this case the p-value is the area under N (, 1) to the right of the t-statistic p value = Pr H (t > t act) = 1 Φ (t act) With a significance level of 5% (α =.5) we reject H if t act > 1.64 If the alternative hypothesis is H 1 : E (Y ) < µ Y, p value = Pr H (t < t act) = 1 Φ ( t act) and we reject H if t act < 1.64 / p value <.5

39 Hypothesis test with a one-sides alternative Example: The mean wage of individuals with a master degree 39 Thursday January 12 12:17: Page 1 (R) / / / / / / / / / / / / Statistics/Data Analysis 1. ttest wage=6 One-sample t test Variable Obs Mean Std. Err. Std. Dev. [95% Conf. Interval] wage mean = mean( wage) t = Ho: mean = 6 degrees of freedom = 249 Ha: mean < 6 Ha: mean!= 6 Ha: mean > 6 Pr(T < t) =.932 Pr( T > t ) =.1396 Pr(T > t) =.698

40 Confidence intervals for the population mean 4 Suppose we would do a two-sides hypothesis test for many different values of µ Y, On the basis of this we can construct a set of values which are not rejected at a 5% significance level If we were able to test all possible values of µ Y, we could construct a 95% confidence interval A 95% confidence interval is an interval that contains the true value of µ Y in 95% of all possible random samples. Instead of doing infinitely many hypothesis tests we can compute the 95% confidence interval as { ( ) ( )} Y 1.96 SE Y, Y SE Y ( Intuition: a value of µ Y, smaller than Y 1.96 SE Y 1.96 SE ( Y ) will be rejected at α =.5 Y ) or bigger than

41 Confidence intervals for the population mean Example: The mean wage of individuals with a master degree 41 When the sample size n is large: 95% confidence interval for µ Y = 9% confidence interval for µ Y = 99% confidence interval for µ Y = { ( )} Y ± 1.96 SE Y { ( )} Y ± 1.64 SE Y { ( )} Y ± 2.58 SE Y Using the sample of 25 individuals with a master degree:. 95% conf. int. for µ W is { ± } = { , } 9% conf. int. for µ W is { ± } = { , } 99% conf. int. for µ W is { ± } = { , }