Professional Documents
Culture Documents
(at the corresponding confidence and risk levels) DEMONSTRATED VIA A Z TEST
Semantically, the duality must be true, if these two concepts of statistics are to be consistent.
If we say that we accept the hypothesis H 0 : μ=μ0 , how can it not be in the confidence interval for the
unknown μ? In other words, how can we be not confident in μ=μ 0 .
Conversely, if μ1 is in the confidence interval, i.e., we being confident about μ1 being an estimate for
the unknown mean μ , how can we reject H 0 : μ=μ1 .
Both Confidence Interval (CI) and Hypothesis Testing (HT) are derived from the following same statistical
fact about the same statistic (i.e., Sample Average – a function of the random variables constituting the
random sample):
Assuming that Xi’s are independent and identically distributed with a Normal Distribution with a
KNOWN variance σ 2and a mean μ, which is assumed known in HT but is UKNOWN in CI.
( )
As formally defined, z 1−α/ 2 is such that P Z ≤ z 1− α =1−α /2 . It is a positive number depending on α .
2
( )
X−μ
P −z 1−α /2 ≤ =Z ≤ z 1−α/ 2 =1−α
σ (1)
√n
This is true for the random variable X regardless of what μ is.
In HT, μ is given and hence known. However, in CI, μ is UNKNOWN and must be estimated, with an
interval estimate.
Hypothesis testing is simple. Given a particular observed X = x , we accept H 0 : μ=μ0 if and only if
1
x−μ 0
−z 1−α / 2 ≤ ≤ z 1−α /2
σ
√n
To get a 95% CI for the UNKNOWN μ, we need to define the lower and the upper end point of an
interval such that regardless of what μ is, the probability that the interval covers the UNKNOWN μ is (at
least) 95%.
X−μ
−z 1−α/ 2 ≤ ≤ z 1−α/ 2
σ
√n
produces the ¿ ) confidence interval (when X =x is observed).
σ
Multiply all three sides by produces
√n
σ σ
−z 1−α / 2 ≤ X−μ ≤ z 1−α /2
√n √n
Subtracting X from all three produces
σ σ
−X −z 1−α / 2 ≤−μ ≤− X + z 1−α /2
√n √n
Multiplying all three by -1 produces
σ σ
X + z 1−α /2 ≥ μ ≥ X−z 1−α / 2
√n √n
This is the same as
σ σ
X −z 1−α/2 ≤ μ ≤ X + z 1−α / 2
√n √n
So, regardless of what μ is, we have
(
P X−z 1−α / 2
σ
√n
≤ μ ≤ X + z 1−α /2) σ
√n
=1−α (2)
2
So, when the random X is observed for (very) large number of times, the probability (i.e. the long-run
proportion) of μ being covered by the interval (defined by these two end points) is 1−α .
σ σ
x−z 1−α/ 2 ≤ μ ≤ x + z 1−α / 2 (3)
√n √n
Note: This particular CI may or may not cover the unknown μ . Again, all we know is that when we
conduct the same experiment and obtain the random X a (very) large number of times, the
corresponding CIs will cover the unknown μ 95% of the time.
Part 1: If we accept a null hypothesis H 0 : μ=μ0 , we show that μ0 must be in the CI.
x−μ 0
−z 1−α / 2 ≤ ≤ z 1−α /2
σ
√n
The same manipulations, but now with this particular given μ0 instead of a general μ , will lead to
σ σ
x−z 1−α / 2 ≤ μ0 ≤ x + z 1−α / 2
√n √n
So, μ0 is indeed in the confidence interval (obtained with the observed x ).
Part 2: For any μ1 in the confidence interval, we show that we accept H 0 : μ=μ1 . In fact, if we simply
work backward, any μ1 such that
σ σ
x−z 1−α/ 2 ≤ μ1 ≤ x + z 1−α /2
√n √n
would lead to
3
x−μ 1
−z 1−α / 2 ≤ ≤ z 1−α /2
σ
√n
And hence acceptance of H 0 : μ=μ1 .
4
THE DUALITY BETWEEN ONE-SIDED CONFIDENCE INTERVAL AND HYPOTHESIS TESTING
(at the corresponding confidence and risk levels) DEMONSTRATED VIA A Z TEST
(1) A LITTLE INTUITION: USE THE ALTERNATIVE HYPOTHESIS H 1 : μ> μ 0. THE CASE OF H 1 : μ< μ 0 IS
COMPLETELY SYMMETRICAL.
Semantically, the duality must be true, if these two concepts of statistics are to be consistent.
If we say that we accept the hypothesis H 0 : μ=μ0 vs. H 1 : μ> μ 0, how can it not be in the confidence
interval for the unknown μ? In other words, how can we be not confident in μ=μ 0 .
Conversely, if μ1 is in the confidence interval, i.e., we being confident about μ1 being an estimate for
the unknown mean μ , how can we reject H 0 : μ=μ1 .
Both Confidence Interval (CI) and Hypothesis Testing (HT) are derived from the following same statistical
fact about the same statistic (i.e., Sample Average – a function of the random variables constituting the
random sample):
Assuming that Xi’s are independent and identically distributed with a Normal Distribution with a
KNOWN variance σ 2and a mean μ, which is assumed known in HT but is UKNOWN in CI.
As formally defined, z 1−α is such that P ( Z ≤ z 1−α )=1−α . It is a positive number depending on α .
When α =0.05 , z 1−α =z 1−0.05=z 0.95 ≅ 1.645. ( −z 1−α ≅−1.645 .)
( )
X −μ
P =Z ≤ z 1−α =1−α
σ (1)
√n
This is true for the random variable X regardless of what μ is.
WHY ARE WE CONCERNED ABOUT THIS INEQUALITY, RATHER THE OTHER ONE? THIS IS WHY: UNDER
THE NULL HYPOTHESIS, HALF OF THE RANDOM OUTCOMES OF STANDARDIZED X WILL FALL TO THE
POSITIVE OR RIGHT SIDE OF 0; YES, HALF! THE QUESTION IS HOW FAR RIGHT FROM 0 IS TOO FAR, SO
FAR THAT WE SHOULD FAVOR THE NULL HYPOTHESIS H 1 : μ> μ 0 .
In HT, μ is given and hence known. However, in CI, μ is UNKNOWN and must be estimated, with an
interval estimate.
5
Hypothesis testing is simple. Given a particular observed X = x , we accept H 0 : μ=μ0 if and only if
x−μ0
≤ z 1−α
σ
√n
2.2 ‘Deriving’ the Confidence Interval:
To get a 95% CI for the UNKNOWN μ, we need to define the lower and the upper end point of an
interval such that regardless of what μ is, the probability that the interval covers the UNKNOWN μ is (at
least) 95%.
X−μ
≤ z 1−α
σ
√n
produces the ¿ ) confidence interval (when X =x is observed).
σ
Multiply all three sides by produces
√n
σ
X −μ ≤ z 1−α
√n
Subtracting X from all three produces
σ
−μ ≤− X+ z1−α
√n
Multiplying all three by -1 produces
σ
μ ≥ X−z 1−α
√n
This is the same as
σ
X −z 1−α ≤μ
√n
So, regardless of what μ is, we have
(
P X−z 1−α
σ
√n )
≤ μ =1−α (2)
6
So, when the random X is observed for (very) large number of times, the probability (i.e. the long-run
proportion) of μ being covered by the interval (defined by these two end points) is 1−α .
σ
x−z 1−α/ 2 ≤μ (3)
√n
Note: This particular CI may or may not cover the unknown μ . Again, all we know is that when we
conduct the same experiment and obtain the random X a (very) large number of times, the
corresponding CIs will cover the unknown μ 95% of the time.
Part 1: If we accept a null hypothesis H 0 : μ=μ0 , we show that μ0 must be in the CI.
x−μ0
≤ z 1−α
σ
√n
The same manipulations, but now with this particular given μ0 instead of a general μ , will lead to
σ
x−z 1−α ≤ μ0
√n
So, μ0 is indeed in the confidence interval (obtained with the observed x ).
Part 2: For any μ1 in the confidence interval, we show that we accept H 0 : μ=μ1 . In fact, if we simply
work backward, any μ1 such that
σ
x−z 1−α ≤ μ1
√n
would lead to
−x−μ1
≤ z 1− α
σ
√n
And hence acceptance of H 0 : μ=μ1 .
7
8