Copyright © 2006 The McGraw-Hill Companies, Inc. All rights reserved. McGraw-Hill/Irwin Statistical Inference: Estimation and Hypothesis Testing chapter five
Copyright © 2006 The McGraw-Hill Companies, Inc. All rights reserved.McGraw-Hill/Irwin
Statistical Inference: Estimation and Hypothesis
Testing
chapter five
5-2
Statistical Inference
Drawing conclusions about a population based on a random sample from that population
Consider Table D-1(5-1): Can we use the average P/E ratio of the 28 companies shown as an estimate of the average P/E ratio of the 3000 or so stocks on the NYSE?
If X = P/E ratio of a stock and Xbar the average P/E of the 28 stocks, can we tell what the expected P/E ratio, E(X), is for the whole NYSE?
5-3
Table D-1 (5-1)
Price to Earnings (P/E) ratios of 28 companies on the New York Stock Exchange (NYSE).
5-4
Estimation Is the First Step
The average P/E from a random sample of stocks, Xbar, is an estimator (or sample statistic) of the population average P/E, E(X), called the population parameter. The mean and variance are parameters of the normal
distribution A particular value of an estimator is called an estimate, say
Xbar = 23.
Estimation is the first step in statistical inference.
5-5
How good is the estimate?
If we compute Xbar for each of two or more random samples, the estimates likely will not be the same.
The variation in estimates from sample to sample is called sampling variation or sampling error.
The error is not deliberate, but inherent in a random sample as the elements included in the sample will vary from sample to sample.
What are the characteristics of good estimators?
5-6
Hypothesis Testing
Suppose expert opinion tells us that the expected P/E of the NYSE is 20, even though our sample Xbar is 23.Is 23 close to the hypothesized value of 20?Is 23 statistically different from 20?Statistically, could 23 be not that different from 20?
Hypothesis testing is the method by which we can answer such questions as these.
5-7
Estimation of Parameters
Point estimate Xbar = 23.25 from Table D-1 (5-1) is a point estimate of μX
the population parameter The formula Xbar = ∑Xi/n is the point estimator or statistic,
a r.v. whose value varies from sample to sample
Interval estimate Is it better to say that the interval from 19 to 24 most likely
includes the true μX, even though Xbar = 23.25 is our best guess of the value of μX?
5-8
Interval Estimates
If X ~ N(μX, σ2X), then the sample mean
Xbar ~ N(μX, σ2X/n ) for a random sample
Or Z = (Xbar- μX)/(σX/√n) ~ N(0, 1)And for unknown σX
2, t = (Xbar-μX)/(SX/√n) ~ t(n-1)
Even if X is not normal, Xbar will be for large nWe can construct an interval for μX using the t
distribution with n-1 = 27 d.f. from Table E-2 (A-2)P(-2.052 < t < 2.052) = 0.95
5-9
Interval Estimates
The t values defining this interval (-2.052, 2.052) are the critical t values t = -2.052 is the lower critical t value t = 2.052 is the upper critical t value
See Fig. D-1 (5-1). By substitution, we can get P(-2.052 < (Xbar-μX)/(SX/√n) < 2.052), OR P(Xbar-2.052(SX/√n) < μX < Xbar+2.052(SX/√n)) = 0.95 An interval estimator of μX for a confidence interval of
95% or confidence coefficient of 0.95 0.95 is the probability that the random interval contains the
true μX
5-10
Figure D-1 (5-1)
The t distribution for 27 d.f.
5-11
Interval Estimator
The interval is random because Xbar and SX/√n vary from sample to sample
The true but unknown μX is some fixed number and is not random
DO NOT SAY: that μX lies in this interval with probability 0.95
SAY: there is a 0.95 probability that the (random) interval contains the true μX
5-12
Example
For the P/E example23.25 – 2.052(9.49/√28) < μX < 23.25 +
2.052(9.49/√28)Or 19.57 < μX < 26.93 (approx.) as the 95%
confidence interval for μX
This says, if we were to construct such intervals 100 times, then 95 out of 100 intervals would contain the true μX
5-13
Figure D-2 (5-2)
(a) 95% and (b) 99% confidence intervals for μx for 27 d.f.
5-14
In General
From a random sample of n values X1, X2,…, Xn, compute the estimators L and U such that P(L < μX < U) = 1 – α
The probability is (1 – α) that the random interval from L to U contains the true μX
1- α is the confidence coefficient and α is the level of significance or the probability of committing a type I error
Both may be multiplied by 100 and expressed as a percent If α = 0.05 or 5%, 1 – α = 0.95 or 95%
5-15
Properties of Point Estimators
The properties of Xbar, compared to the sample median or mode, make it the preferred estimator of the population mean, μX: Linearity Unbiasedness Minimum variance Efficiency Best Linear Unbiased Estimator (BLUE) Consistency
5-16
Properties of Point Estimators
Linearity A linear estimator is a linear function of the sample
observations Xbar = ∑(Xi/n) = (1/n)(X1 + X2 +…Xn) The Xs appear with an index or power of 1 only
Unbiasedness: E(Xbar) = μX (Fig. D-3,5-3) In repeated applications of a method, if the mean value of an
estimator equals the true parameter (population) value, the estimator is unbiased.
With repeated sampling, the sample mean and sample median are unbiased estimators of the population mean.
5-17
Figure D-3 (5-3)
Biased (X*) and unbiased (X) estimators of populationmean value, μx.
5-18
Properties of Point Estimators
Minimum Variancea minimum-variance estimator has smaller variance
than any other estimator of a parameterIn Fig. D-4 (5-4), the minimum-variance estimator
of μX is also biased
Efficiency (Fig. D-5, 5-5)Among unbiased estimators, the one with the
smallest variance is the best or efficient estimator
5-19
Figure D-4 (5-4)
Distribution of three estimators of μx.
5-20
Figure D-5 (5-5)
An example of an efficient estimator (sample mean).
5-21
Properties of Point Estimators
Efficiency example Xbar ~ N(μX, σ2/n) sample mean
Xmed ~ N(μX, (π/2)(σ2/n)) sample median
(var Xmed)/(var Xbar) = π/2 ≈ 1.571 Xbar is a more precise estimator of μX.
Best Linear Unbiased Estimator (BLUE) An estimator that is linear, unbiased, and has the minimum
variance among all linear and unbiased estimators of a parameter
5-22
Properties of Point Estimators
Consistency (Fig. D-6, 5-6)A consistent estimator approaches the true value of
the parameter as the sample size becomes large.Consider Xbar = ∑Xi/n and X* = ∑Xi/(n + 1)
E(Xbar) = μX but E(X*) = [n/(n + 1)] μX.
X* is biased.As n gets large, n/(n + 1) → 1, E(X*) → μX .
X* is a biased, but consistent estimator of μX .
5-23
Figure D-6 (5-6)
The property of consistency. The behavior of the estimatorX* of population mean μx as the sample size increases.
5-24
Hypothesis Testing
Suppose we hypothesize that the true mean P/E ratio for the NYSE is 18.5Null hypothesis H0: μX = 18.5Alternative hypothesis H1
H1: μX > 18.5 one-sided or one-tailedH1: μX < 18.5 one-sided or one-tailedH1: μX ≠ 18.5 composite, two-sided or two-tailed
Use the sample data (Table D-1 (5-1), average P/E = 23.25) to accept or reject H0 and/or accept H1
5-25
Confidence Interval Approach
H0: μX = 18.5, H1: μX ≠ 18.5 (two-tailed) We know t = (Xbar - μX)/(SX/√n) ~ tn-1. Use Table (E-2) A-2 to construct the 95% interval
Critical t values (-2.052, 2.052) for 95% or 0.95 P(Xbar-2.052(SX/√n) < μX < Xbar+2.052(SX/√n)) = 0.95 23.25 – 2.052(9.49/√28) < μX < 23.25 + 2.052(9.49/√28) Or 19.57 < μX < 26.93
H0: μX = 18.5 < 19.57, outside the interval Reject H0 with 95% confidence
5-26
Confidence Interval Approach
Acceptance region19.57 < H0:μX < 26.93 interval for 95%
Critical region or region of rejectionH0:μX < 19.57 and 26.93 < H0:μX .
Accept H0 if value within acceptance regionReject H0 if value outside the acceptance regionCritical values are the dividing line between
acceptance and rejection of H0
5-27
Type I and Type II Errors
We rejected H0: μX = 18.5 at a 95% level of confidence, not 100%
Type I Error: reject H0 when it is trueIf we hypothesized H0: μX = 21 above, we
would not have rejected it with 95% confidenceType II Error: accept H0 when it is falseFor any given sample size, one cannot
minimize the probability of both types of error
5-28
Type I and Type II Errors
Level of Significance, αType I error = α = P(reject H0|H0 is true)
Power of the test, (1 – β)Type II error = β = P(accept H0|H0 is false)
Trade-off: min α vs. max (1 – β)In practice: set α fairly low (0.05 or 0.01) and
don’t worry too much about (1 – β)
5-29
Example
H0: μX = 18.5 and α = 0.01 (99% confidence)Critical t values (-2.771, 2.771) with 27 d.f.18.28 < μX < 28.22 is 99% conf. interval
Do not reject H0
See Fig. D-2 (5-2)Decreasing α, P(Type I error), increases β,
P(Type II error)
5-30
Test of Significance Approach
For one-sided or one-tailed tests Recall t = (Xbar - μX)/(SX/√n)
We know Xbar, SX, and n; we hypothesize μX .
We can just calculate the value of t for our sample and μX hypothesis
Then look up its probability in Table E -2 (A-2). Compare that probability to the level of significance,
α, you choose, to see if you reject H0
5-31
Example
P/E example Xbar = 23.25, SX = 9.49, n = 28 H0: μX = 18.5, H1: μX ≠ 18.5 t = (23.25 – 18.5)/(9.49/√28) = 2.6486
Set α = 0.05 in a two-tailed test (why?) Critical t values are (-2.052, 2.052) for 27 d.f. 2.65 is outside of the acceptance region Reject H0 at 5% level of significance
Reject null: test is statistically significant Do not reject: test is statistically insignificant The difference between observed (estimated) and hypothesized
values of a parameter is or is not statistically significant.
5-32
One tail or Two?
H0: μX = 18.5, H1: μX ≠ 18.5 two-tailed testH0: μX < 18.5, H1: μX > 18.5 one-tailed testTesting procedure is exactly the same
Choose α = 0.05Critical t value = 1.703 for 27 d.f.2.6 > 1.703, reject H0 at 5% level of significanceThe test (statistic) is statistically significant
See Fig. D-7 (5-7).
5-33
Figure D-7 (5-7)
The t test of significance: (a) Two-tailed;(b) right-tailed; (c) left-tailed.
5-34
Table 5-2
A summary of the t test.
5-35
The Level of Significance and the p-Value
Choice of α is arbitrary in classical approach 1%, 5%, 10% commonly used
Calculate the p-value instead A.k.a.: exact significance level of the test statistic For P/E example with H0:μX < 18.5, t ≈ 2.6 P(t27 > 2.6) < 0.01, p-value < 0.01 or 1% Statistically significant at the 1% level
In econometric studies, the p-values are commonly reported (or indicated) for all statistical tests
5-36
The χ2 Test of Significance
(n-1)(S2/σ2) ~ χ2(n-1) .
We know n, S2, and hypothesize σ2
Calculate χ2 value directly and test its significanceExample: n = 31, S2 = 12
H0: σ2 = 9, H1: σ2 ≠ 9, use α = 5%χ2
(30) = 30(12/9) = 40 P(χ2
(30) > 40) ≈ 10% > 5% = αDo not reject H0: σ2 = 9
5-37
Table 5-3
A summary of the x2 test.
5-38
F Test of Significance
F = SX2/SY
2
Or [(∑X-Xbar)2/(m-1)]/∑(Y-Ybar)2/(n-1)] follows the F distribution with (m-1, n-1) d.f.
IF σX2 = σY
2, so H0: σX2 = σY
2 . Example: SAT Scores in Ex. 4.15
varmale = 46.1, varfemale = 83.88, n = 24 for both F = 83.88/46.1 ≈ 1.80 with 23, 23 d.f. Critical F value for 24 d.f. each at 1% is 2.66 1.8 < 2.66, not statistically significant, do not reject H0
5-39
Table 5-4
A summary of the F statistic.