ECONOMETRICS

Report
Lecture 8:
Hypothesis
Testing
(Chapter 7.1–7.2, 7.4)
Distribution of Estimators
(Chapter 5.1–5.2,
Chapter 6.4)
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
Agenda for Today
• Hypothesis Testing (Chapter 7.1)
• Distribution of Estimators (Chapter 5.2)
• Estimating s 2 (Chapter 5.1, Chapter 6.4)
• t-tests (Chapter 7.2)
• P-values (Chapter 7.2)
• Power (Chapter 7.2)
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-2
What Sorts of Hypotheses to Test?
• To test a hypothesis, we first need to
specify our “null hypothesis” precisely, in
terms of the parameters of our
regression model. We refer to this “null
hypothesis” as H0.
• We also need to specify our “alternative
hypothesis,” Ha , in terms of our
regression parameters.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-3
What Sorts of Hypotheses to Test? (cont.)
• Claim: The marginal propensity to consume is
greater than 0.70 :
Ci  b0  b1Incomei  i
• Conduct a one-sided test of the null hypothesis
• H0 : b1 > 0.70 against the alternative,
• Ha : b1 = 0.70
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-4
What Sorts of Hypotheses to Test? (cont.)
• Claim: The marginal propensity to consume
equals the average propensity to consume:
Ci  b0  b1Incomei  i , with b0  0
• Conduct a two-sided test of
• H0 : b0 = 0 against the alternative,
• Ha : b0 ≠0
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-5
What Sorts of Hypotheses to Test? (cont.)
• The CAPM model from finance says that the
E(excess return on portfolio k)
 b ·(Excess return on market portfolio)
• Regress
E(excess return on portfolio k)
 b0  b1 (excess return on market portfolio)
for a particular mutual fund, using data over time.
Test H0 : b0 > 0.
• If b0 > 0, the fund performs better than expected, said
early analysts. If b0 < 0, the fund performs less well
than expected.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-6
What Sorts of Hypotheses to Test? (cont.)
• H0 : b0 > 0
• Ha : b0 = 0
• What if we run our regression and find
bˆ 0  0.012
• Can we reject the null hypothesis? What if
bˆ 0  0.012?
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-7
Hypothesis Testing: Errors
• In our CAPM example, we are testing
– H0 : b0 > 0, against the alternative
– Ha : b0 = 0
• We can make 2 kinds of mistakes.
– Type I Error: We reject the null hypothesis
when the null hypothesis is “true.”
– Type II Error: We fail to reject the null
hypothesis when the null hypothesis
is “false.”
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-8
Hypothesis Testing: Errors (cont.)
• Type I Error: Reject the null hypothesis when
it is true.
• Type II Error: Fail to reject the null hypothesis
when it is false.
• We need a rule for deciding when to reject a
null hypothesis. To make a rule with a lower
probability of Type I error, we have to have a
higher probability of Type II error.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-9
Hypothesis Testing: Errors (cont.)
• Type I Error: Reject the null hypothesis when
it is true.
• Type II Error: Fail to reject the null hypothesis
when it is false.
• In practice, we build rules to have a low
probability of a Type I error. Null hypotheses
are “innocent until proven guilty beyond a
reasonable doubt.”
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-10
Hypothesis Testing: Errors (cont.)
• Type I Error: Reject the null hypothesis when
it is true.
• Type II Error: Fail to reject the null hypothesis
when it is false.
• We do NOT ask whether the null hypothesis
is more likely than the alternative hypothesis.
• We DO ask whether we can build a
compelling case to reject the null hypothesis.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-11
Hypothesis Testing
• What constitutes a compelling case to
reject the null hypothesis?
• If the null hypothesis were true, would
we be extremely surprised to see the
data that we see?
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-12
Hypothesis Testing: Errors
• In our CAPM example
H 0 : b0  0
H a : b0  0
• What if we run our regression and find
bˆ 0  0.012?
• Could a reasonable jury reject the null
hypothesis if the estimate is “just a little lower”
than 0?
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-13
Hypothesis Testing: Errors (cont.)
• Type I Error: Reject the null hypothesis
when it is true.
• Type II Error: Fail to reject the null
hypothesis when it is false.
• In our CAPM example, our null
hypothesis is b0 > 0. Can we use our
data to amass overwhelming evidence
that this null hypothesis is false?
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-14
Hypothesis Testing: Errors (cont.)
• Note: if we “fail to reject” the null, it
does NOT mean we can “accept” the
null hypothesis.
• “Failing to reject” means the null has
“reasonable doubt.”
• The null hypothesis could still be fairly
unlikely, just not overwhelmingly unlikely.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-15
Hypothesis Testing: Strategy
• Our Strategy: Look for a Contradiction.
• Assume the null hypothesis is true.
• Calculate the probability that we see
the data, assuming the null hypothesis
is true.
• Reject the null hypothesis if this
probability is just too darn low.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-16
How Should We Proceed?
1. Ask how our estimates of b0 and b
are distributed if the null hypothesis
is true.
2. Determine a test statistic.
3. Settle upon a critical region to reject
the null hypothesis if the probability of
seeing our data is too low.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-17
How Should We Proceed? (cont.)
• The key tool we need is the probability
of seeing our data if the null hypothesis
is true.
• We need to know the distribution of
our estimators.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-18
Distribution of a Linear Estimator
(from Chapter 5.2)
If the true value of b 0  0, what is
the probability that we observe data
with estimated intercept bˆ ?
0
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-19
Distribution of a Linear Estimator (cont.)
• Perhaps the most common hypothesis test is
H0 : b = 0 against Ha : b ≠ 0
• This hypothesis tests whether a variable has
any effect on Y
• We will begin by calculating the variance of
our estimator for the coefficient on X1
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-20
Hypothesis Testing
• Add to the Gauss–Markov Assumptions
• The disturbances are normally distributed
i ~ N(0, s )
2
Yi ~ N(b Xi , s )
2
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-21
DGP Assumptions
Yi  b0  b1 X1i  b2 X 2i  bk X ki   i
E( i )  0
Var( i )  s
2
Cov( i , j )  0, for i  j
Each explanator is fixed across samples
 i ~ N (0, s )
2
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-22
How Should We Proceed?
• Ask how our guesses of b0 and b1
are distributed.
• Since the Yi are distributed normally,
all linear estimators are, too.
bˆ0 ~ N ( b 0 ,?)
bˆ1 ~ N ( b1 ,?)
What are the variances of bˆ0 and bˆ1?
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-23
What is the Variance of bˆ 1?
Var(wiYi )  Var(wiYi )  0
 wi2Var(Yi )  s 2 wi2
Let
xi  Xi - X
and
yi  Yi - Y
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-24
What is the Variance of bˆ 1? (cont.)
Var(wiYi )  s 2 wi2

ˆ
b1 
xiyi
x
i
wi 
2
xi
 xi 2
xi 2
2
2
2
ˆ
Var(b1)  s  wi  s  (
)
2
 xi
s
2

( x
xi 2
i
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
2 2
)

s2
x
i
2
8-25
What is the Variance of bˆ 1? (cont.)
Var ( bˆ 1) 
s
2
x
i
2
where xi  Xi  X
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-26
What is the Variance of bˆ 1? (cont.)
Var ( bˆ 1) 
s2
x
2
i
Thus...
bˆ 1 ~ N ( b 1,
s2
x
2
)
i
Suppose we have as our null hypothesis
H 0 : b 1  b 1*
Under the null hypothesis
bˆ 1 ~ N ( b 1 ,
*
s2
x
2
)
i
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-27
Distribution of a Linear Estimator
• We have a formula for the distribution
of our estimator. However, this formula
is not in a very convenient form. We
would really like a formula that gives a
distribution for which we can look up the
probabilities in a common table.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-28
Test Statistics
• A “test statistic” is a statistic:
1. Readily calculated from the data
2. Whose distribution is known (under the
null hypothesis)
• Using a test statistic, we can compute
the probability of observing the data
given the null hypothesis.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-29
Test Statistics (cont.)
bˆ 1 ~ N ( b 1 ,
*
s2
x
2
)
i
If we subtract the mean and divide by the
standard error, we can transform a
Normal Distribution into a Standard Normal
Distribution.
Z
bˆ 1 - b 1
*
s
2
 xi
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
~ N (0,1)
2
8-30
Test Statistics (cont.)
We can easily look up the probability
that we observe a given value of
Z
bˆ 1 - b 1
*
s2
x
2
i
on a Standard Normal table.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-31
Test Statistics (cont.)
Z
bˆ 1 - b 1
*
s
2
 xi
~ N (0,1)
2
Suppose we want to test
H 0 : b1  0.70, against the alternative
H a : b1  0.70
We could replace b1 * with 0.70 and calculate Z
If Z  -1.64, then we know there is less than a 5%
chance we would observe this data if b1 really
were greater than 0.70
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-32
Estimating s2
(from Chapter 5.1, Chapter 6.4)
One Problem: We cannot observe
bˆ 1 - b 1*
because we cannot observe s 2 .
s2
x
2
i
Solution: Estimate s .
2
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-33
Estimating s2 (cont.)
• We need to estimate the variance of the
error terms, i  Yi - b0 - b1 X1i - ...- bk Xki
• Problem: we do not observe i directly.
• Another Problem: we do not know
b0…bk, so we cannot calculate i either.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-34
Estimating s2 (cont.)
• We need to estimate the variance of the
error terms, i  Yi - b0 - b1 X1i - ...- bk Xki
• We can proxy for the error terms using
the residuals.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-35
Estimating s2 (cont.)
ˆ
ˆ
ˆ
ei  Yi  b 0  b1 X 1i  ...  b k X ki
 Y  Yˆ
i
i
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-36
Estimating s2 (cont.)
ei  Yi  bˆ0  bˆ1 X 1i  ...  bˆk X ki
 Y  Yˆ
i
i
• Once we have an estimate of the error
term, we can calculate an estimate of the
variance of the error term. We need to make
a “degrees of freedom” correction.
1
2
s 
e
i

n  k 1
2
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-37
Estimating s2 (cont.)
s2
Recall Var ( bˆ 1) 
 xi
2
Thus...
bˆ 1 ~ N ( b 1,
s2
x
2
)
i
Suppose we have as our null hypothesis
H 0 : b 1  b 1*
Under the null hypothesis
bˆ 1 ~ N ( b 1 ,
*
s2
x
2
)
i
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-38
Estimating s2 (cont.)
Suppose we have as our null hypothesis
H 0 : b1  b
*
1
Under the null hypothesis
2 

s
*
ˆ

b1 ~ N  b1 ,
2


x
i
 

Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-39
Estimating s2 (cont.)
Plug in our estimate for s :
2
2


s
*
ˆ


b1 ~ N b1 ,
2


x
i
 

2



e
*
i
ˆ


b1 ~ N b1 ,
2


(
n

k

1)
x
i
 

Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-40
Standard Error (from Chapter 5.2)
• Remember, the standard deviation of
the distribution of our estimator is called
the “standard error.”
• The smaller the standard error, the more
closely your estimates will tend to fall to
the mean of the distribution.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-41
Standard Error (from Chapter 5.2)
• If your estimate is unbiased, a low
standard error implies that your
estimate is probably “close” to the true
parameter value.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-42
t-statistic (from Chapter 7.2)
tˆ 
bˆ  b *
s2
2
x
i

~ tn  k  1
• Because we need to estimate the standard
error, the t-statistic is NOT distributed as
a Standard Normal. Instead, it is distributed
according to the t-distribution.
• The t-distribution depends on n-k-1. For large
n-k-1, the t-distribution closely resembles the
Standard Normal.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-43
t-statistic (cont.)
tˆ 
*
ˆ
b b
sbˆ 2
~ tn  k  1
• Under the null hypothesis H0 : b1 = b1*
t ~ tn-2
• In our earlier example, we could:
1. Replace b1* with 0.70
2. Compare t to the “critical value” for which the tn-2
distribution has .05 of its probability mass lying to the left,
3. There is less than a 5% chance of observing the data
under the null if t < “critical value.”
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-44
Figure 7.1 Critical Regions for Two-Tailed and
One-Tailed t-Statistics with a 0.05 Significance
Level and 10 Degrees of Freedom
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-45
Significance Level
• We can now calculate the probability of
observing the data IF the null hypothesis
is true.
• We choose the maximum chance we are
willing to risk that we accidentally commit a
Type I Error (reject a null hypothesis when
it is true).
• This chance is called the “significance level.”
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-46
Significance Level (cont.)
• We choose the probability we are willing to
accept of a Type I Error.
• This probability is the “Significance Level.”
• The significance level gives operational
meaning to how compelling a case we need
to build.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-47
Significance Level (cont.)
• The significance level denotes the
chance of committing a Type I Error.
• By historical convention, we usually
reject a null hypothesis if we have less
than a 5% chance of observing the data
under the null hypothesis.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-48
Significance Level (cont.)
• 5% is the conventional significance
level. Analysts also often look at the
1% and 10% levels.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-49
Critical Region
• We know the distribution of our test
statistic under the null hypothesis.
• We can calculate the values of the test
statistic for which we would reject the
null hypothesis (i.e., values that we
would have less than a 5% chance of
observing under the null hypothesis).
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-50
Critical Region (cont.)
• We can calculate the values of the
test statistic for which we would reject
the null hypothesis.
• These values are called the
“critical region.”
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-51
Figure 7.1 Critical Regions for Two-Tailed and
One-Tailed t-Statistics with a 0.05 Significance
Level and 10 Degrees of Freedom
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-52
Critical Region
• Regression packages routinely report estimated
coefficients, their estimated standard errors, and the
t-statistics associated with the null hypothesis that an
individual coefficient is equal to zero.
• Some programs also report a “p-value”
for each estimated coefficient.
• This reported p-value is the smallest significance
level for a two sided test at which one would reject
the null that the coefficient is zero.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-53
One-Sided, Two-Sided Tests
• t-tests come in two flavors: 1-sided and
2-sided.
• 2-sided tests are much more common:
– H0 : b = b*
– Ha : b ≠b*
• 1-sided tests look at only one-side:
– H0 : b > b*
– Ha: b = b*
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-54
One-Sided, Two-Sided Tests (cont.)
• The procedure for both 1-sided and
2-sided tests is very similar.
• For either test, you construct the same
t-statistic:
*
ˆ
b

b
tˆ 
ˆ
s.e.(b )
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-55
One-Sided, Two-Sided Tests (cont.)
• Once you have your t-statistic, you
need to choose a “critical value.” The
critical value is the boundary point for
the critical region. You reject the null
hypothesis if your t-statistic is greater
in magnitude than the critical value.
• The choice of critical value depends
on the type of test you are running.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-56
Critical Value for 1-Sided Test
• For a 1-sided test, you need a critical
value such that a of the distribution of
the estimator is greater than (or less
than) the critical value. a is our
significance level (for example, 5%).
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-57
Critical Value for 1-Sided Test (cont.)
• In our CAPM example, we want to test:
H 0 : b0  0
H a : b0  0
• We need a critical value t* such that a of
the distribution of our estimator is less
than t*
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-58
Figure 7.1 Critical Regions for Two-Tailed and
One-Tailed t-Statistics with a 0.05 Significance
Level and 10 Degrees of Freedom
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-59
Critical Value of a 1-Sided Test (cont.)
• For a 5% significance level and a large
sample size, t* = -1.64
• We reject the null hypothesis if:
ˆt  1.64
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-60
Critical Value for a 2-Sided test
• For a 2-sided test, we need to spread
our critical region over both tails. We
need a critical value t* such that
– a/2 of the distribution is to the right of t*
– a/2 of the distribution is to the left of –t*
• Summing both tails, a of the distribution
is beyond either t* or -t*
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-61
Figure 7.1 Critical Regions for Two-Tailed and
One-Tailed t-Statistics with a 0.05 Significance
Level and 10 Degrees of Freedom
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-62
Critical Value for 2-Sided Test
• For a large sample size, the critical value for a
2-sided test at the 5% level is 1.96
• You reject the null hypothesis if:
tˆ  1.96
or
tˆ  1.96
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-63
P-values
• The p-value is the smallest significance
level for which you could reject the
null hypothesis.
• The smaller the p-value, the stricter the
significance level at which you can
reject the null hypothesis.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-64
P-values (cont.)
• Many statistics packages automatically report
the p-value for a two-sided test of the null
hypothesis that a coefficient is 0
• If p < 0.05, then you could reject the null that
b= 0 at a significance level of 0.05
• The coefficient “is significant at the 95%
confidence level.”
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-65
Statistical Significance
• A coefficient is “statistically significant
at the 95% confidence level” if we could
reject the null that b = 0 at the 5%
significance level.
• In economics, the word “significant”
means “statistically significant” unless
otherwise qualified.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-66
Performing Tests
• How do we compute these test statistics
using our software?
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-67
Power
• Type I Error: reject a null hypothesis
when it is true
• Type II Error: fail to reject a null
hypothesis when it is false
• We have devised a procedure based
on choosing the probability of a
Type I Error.
• What about Type II Errors?
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-68
Power (cont.)
• The probability that our hypothesis test
rejects a null hypothesis when it is false
is called the Power of the test.
• (1 – Power) is the probability of a
Type II Error.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-69
Power (cont.)
• If a test has a low probability of rejecting
the null hypothesis when that
hypothesis is false, we say that the test
is “weak” or has “low power.”
• The higher the standard error of our
estimator, the weaker the test.
• More efficient estimators allow for more
powerful tests.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-70
Power (cont.)
• Power depends on the particular Ha
you are considering. The closer Ha is
to H0, the harder it is to reject the
null hypothesis.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-71
Figure 7.2 Distribution of bˆs for bs = -2, 0,
and a Little Less Than 0
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-72
Figure 7.3 Power Curves for Two-Tailed
Tests of H0 : bs = 0
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-73
Figure SA.12 The Distribution of the t-Statistic
Given the Null Hypothesis is False and = + 5
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-74
Figure SA.13 The t-Statistic’s Power
When the Sample Size Grows
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-75
Review
• To test a null hypothesis, we:
– Assume the null hypothesis is true;
– Calculate a test statistic, assuming the null
hypothesis is true;
– Reject the null hypothesis if we would be
very unlikely to observe the test statistic
under the null hypothesis.
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-76
Six Steps to Hypothesis Testing
1. State the null and alternative
hypotheses
2. Choose a test statistic (so far, we have
learned the t-test)
3. Choose a significance level, the
probability of a Type I Error
(typically 5%)
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-77
Six Steps to Hypothesis Tests (cont.)
4. Find the critical region for the test (for a
2-sided t-test at the 5% level in large
samples, the critical value is t*=1.96)
5. Calculate the test statistic
*
ˆ
b

b
tˆ 
s.e.(bˆ )
6. Reject the null hypothesis if the test statistic
falls within the critical region
Is tˆ  t * or tˆ  t *?
Copyright © 2006 Pearson Addison-Wesley. All rights reserved.
8-78

similar documents