Topic 5: Confidence Intervals (Chapter 9)

1. Introduction
• The two general areas of statistical inference are:
1) estimation of parameter(s), ch. 9
2) hypothesis testing of parameter(s), ch. 10

• Let X be some random variable with unknown mean µ . Suppose we take a

sample of size n . One could find the sample mean, x , and use it to estimate µ .
Such a single number used to estimate a parameter is called a point estimate.

• Question: Are there other reasonable point estimates (other than x ) for µ ?

• Claim: The “best” point estimate of µ is x .

• The point estimate does not give us a sense of how close x might be to µ , i.e. it
doesn’t tell us about the inherent uncertainty in the process. Therefore, interval
estimates are used, which provide a range of values that contain the parameter,
say µ , with some specified degree of confidence. This range is called a
confidence interval.

2. Two-Sided Confidence Interval

• Consider now some random variable X with mean µ and standard deviation σ .
Suppose we want to estimate µ (the estimation of σ follows later). Using the
central limit theorem
σ
X ~ N (µ , )
n
Therefore,
(X − µ)
Z= (1)
σ
n
is a standard normal.

• Recall that by definition

P ( − zα < Z < zα ) = 1 − α (2)
2 2

e.g. because z.025 = 1.96

P(−1.96 < Z < 1.96) = 1 − .05 = .95

1
• Substituting (1) into (2), and using some algebra, one can show that (2) implies
σ σ
P ( X − zα < µ < X + zα ) = 1−α (3)
2 n 2 n
e.g. letting α = .05
σ σ
P ( X − 1.96 < µ < X + 1.96 ) = .95
n n

• The interpretation of (3) is tricky. Notice that we calculate a random interval

σ σ
( X − 1.96 , X + 1.96
).
n n
If we do this repeatedly, we expect that 95% of the intervals contain the
unknown parameter µ . It does not mean that µ assumes a value within the
interval with probability 0.95. An illustration of this is given in Figure 9.1.

• In general, a 100 (1 − α ) % confidence interval for µ is

σ σ
( X − zα , X + zα ). (4)
2 n 2 n

• The confidence interval could be written as

X ±E
σ
where E = zα . (5)
2 n
E is also called the bound on error (BOE).

• Suppose one wants to reduce E. This could obviously be done by increasing α

or increasing n . In fact, n may be determined from E. Notice from (5),
σ
n = zα .
2 E
Therefore,
σ
n = [ zα ]2 . (6)
2 E

• Consider an example. Let X be the cholesterol level of a U.S. male who

smokes. Suppose µ is unknown, but that we know σ = 46 (we actually
assume this value because we presumably know σ = 46 for population of all
U.S. adult males). Suppose we take a sample of n = 12 , and find
x = 217 .

Then the best point estimate of µ is 217. Also, a 95% confidence interval for
µ is (using X = 217 )

2
 1.96(46)
217 ±
12
= 217 ± 26
= (191, 243).
A 99% confidence interval for µ is
2.58(46)
217 ±
12
= (183, 251)
Why is it wider?

• Suppose we are designing the study, and we want to know µ within 10 units.
(Note that E=10 and the interval’s width is 20). Substituting σ , E, and α = .01
into (6), one has
2.58(46) 2
n=( ) = 140.8 .
10
Therefore our recommendation is to obtain a sample for n = 141 males.

3. One-Sided Confidence Intervals

• Sometimes (though not often), one desires just an upper or a lower bound on µ .
Using algebraic manipulations as before, one has:

An 100(1- α )% upper (one-sided) confidence bound on µ is

σ
X + zα , (7)
n
and a corresponding lower confidence bound is
σ
X − zα . (8)
n
e.g. for U.S. males, to find a 95% upper bound, recall that z.05 = 1.65 . Therefore
the upper bound from the sample is
1.65(46)
217 ±
12
= 217 + 21.9
= 238.9

4. Students’ t distribution

• Consider a random variable X which is known to be normally distributed, but

where neither µ nor σ are known. To make inferences about µ , we could use
the Z transformation, provided σ is known. However, in case of unknown σ ,
which is the common case, we replace σ by s to find

3
 (x − µ)
t= .
s
n
This statistic has the t distribution with n-1 “degrees of freedom” (df).

• Several facts of interest are

1) The t-distribution is also symmetric about 0, but it has “thicker tails”, i.e.
it’s a bit flatter. This is because there is added variability introduced by
using the random variable s rather than a constant σ in the denominator
(see Figure 9.2).
2) The t-distribution approaches Z as n becomes large. This is because as
n increases, we have more information and s approaches σ . Why?

• Recall we had an extensive Z table, but it’s impractical to give a table for each
tn −1 statistic. Therefore, tables typically give only upper percent points. For
example, from Table A-4, one has
t10,.025 = 2.228 .
By symmetry,
t10,.975 = −2.228 ,
and thus one can find upper and lower percentiles.

• One can construct confidence intervals based on the t-distribution in the same
way as before. The expression comparable to (3) is
s s
P ( X − tα < µ < X + tα ) = 1−α .
2 n 2 n
In general, a 100(1- α )% confidence interval for µ , with approximately normal
x , is
s s
( X − tα , X + tα ). (9)
2 n 2 n

• For example, let X denote an infant’s plasma aluminum level, which is

assumed to be approximately normal. Suppose we take a random sample of
n = 10 infants, and find x = 37.2 and s = 7.13 .

To find a 95% confidence interval, we find

t9,.025 = 2.262 .
Hence, the interval estimate is
2.262(7.13)
37.2 ±
10
= 37.2 ± 5.1
= ( 32.1, 42.3)

4
 • Notice that in this problem, the interval width is a random variable. Hence, if
two experimenters took different samples, their interval widths would differ,
unlike the case of known σ . In general, the interval based on t would be wider
than that based on Z. Why? Though the widths differ, we expect 100(1- α )%
of the intervals to contain µ . See Figure 9.3.

5. Case Study

• This has a nice case study on the efficacy of using a particular drug for treating
ADD children.