1

Chapter 6: Inference

Read Chapter 6 first omit the power section,

then read the notes and try the WEBCT
assignment questions. If you need more
practice, try the practice questions with answers
available on the web.

Chapter 6 - Introduction

Statistical inference is a mathematical technique

used by scientists to make conclusions about a
population from the sample data they have collected.
Two common forms of inference are introduced in
this chapter, confidence intervals and tests of
significance. We introduce a simple form of
inference concerning the mean of a normal
population with known variance.

Dr. Robert Gebotys Winter 2006

 2
Background

It is rare that we can know exactly the distribution of

a variable X over a population π.

Based on a sample X1,X2,…,Xn from this

distribution we want to make inferences about the
true values of various characteristics of this
distribution. For example: µx, σx, etc.

There are 3 Forms of Inference

(1) estimation
(2) confidence intervals
(3) tests of significance

(1) Estimation

For example: Consider quantitative variables

Let X1, X2,...Xn be a sample from a distribution with

unknown mean µ and unknown variance σ2. We
want to estimate:
µ by x

Dr. Robert Gebotys Winter 2006

 3
and estimate

σ2 by s2 = 1 ∑ (xi – x) 2
n–1
We have seen previously (chapter 3 and 5) that the
sampling distribution of X is centered at the mean
of the population:

µx = µ
and the variability of the sampling distribution about
µ is measured by
2 2
σ x = σ
n
Hence, we estimate the accuracy of x as an estimate
of µ by the standard error of the estimate below:
S
√n

Example: Here is a simulation to help you

understand estimation. Estimation is illustrated here
with a computer simulation with the population

Dr. Robert Gebotys Winter 2006

 4
given below for a variable Y. Suppose Y ~ N (50,
15) and we wish to take 40 samples each of size 25.
We want to estimate the mean.

SPSS EXAMPLE ONE

In my simulation we obtain:

X = 419.866

S = .4359
√n
which is close to the population mean with small
standard error. The researcher usually knows what
accuracy would be required for the estimate. Greater
accuracy can be achieved by increasing the sample
size.
View the Video – Confidence Intervals

(2) Confidence Intervals

How do we interpret the standard error of an
estimate?

Recall the Central Limit Theorem from Chapter 5. If

X1, X2,...,Xn is a sample from a distribution_with
Dr. Robert Gebotys Winter 2006
 5
mean µ and variance σ2 then the mean “x” is
distributed approximately:

N µ, σ
√n

We then can derive the confidence interval as

follows:

P (X ≤ X) = P (X – µ ≤ X – µ)

= P x –µ ≤ x –µ
σ√n σ√n

by the CLT we have

= P Z ≤ x –µ
σ√n

where Z ~ N(0,1). We now choose α, which is the

total amount of probability in the tails of the
distribution and 1- α is the probability in the center
of the distribution. A common value used in
research is .05. Graphically we have N(0,1):

Dr. Robert Gebotys Winter 2006

 6
δ δ

-Zδ Zδ

Note that α is between 0 and 1.

We have that δ + δ = α or 2δ = α. We can write
δ=α/2. Graphically we have:

α/2 α/2

-Zα/2 1-α Zα/2

Lower (L) Upper (U)
We assume σ is known and construct an interval
such that the center probability is 1 - α.

This is called a (1-α) confidence interval (C.I.) for

µ or P (L ≤ µ ≤ U) = 1 – α where L = Lower CI
and U = Upper CI.

After observing X1, X2,...,Xn our inference takes

form. We say µ is contained in [L,U] with
confidence (1-α).
Dr. Robert Gebotys Winter 2006
 7

If (1-α) is close to 1, and the length U-L is small we

have very precise inference about µ.

Now by CLT

= P Zα/2 ≤ x –µ ≤ Zα/2 = 1-α

σ√n

for known σ and fixed α is a (1-α). C.I. for µ of

length 2 σ/√n Zα/2 is:

[x – σ/√n Zα/2 , x + σ/√n Zα/2]

Note a common choice is 1 - α = .95

If we have a 95% C.I. then α/2 = .025. Using

standard Z tables Z.025 = 1.96. The choice of Z
depends on the choice of confidence interval (90%,
95%, 99%, etc.).

Dr. Robert Gebotys Winter 2006

 8

Figure 6.1 x(bar) lies within +/- 9 of µ in 95% of all

samples, so µ also lies within +/- 9 of x(bar) in those
samples.

Example: Simulation

We will use SPSS to simulate how a confidence

interval behaves and consequently better understand
this concept. We generate 20 samples, each of size
25 from a population which is N (50,15). For each
sample calculate a 90% confidence interval. How
many of the intervals contain the true value 50?

Dr. Robert Gebotys Winter 2006

 9
In my simulation of the 20 intervals, 3 do not cover
the mean of 50. We expect approximately 10% (or
2) not to cover mean since we have a 90% C.I. If we
were to run the simulation over many samples in the
long run, 10% would not contain the mean.

SPSS EXAMPLE TWO

Click above to view the C.I. simulation. If you wish

you can repeat the simulation for a (1 - α) = 95%
C.I. using the instructions at the end of the above
example.

X
µ

0
0

0
0

Dr. Robert Gebotys Winter 2006

 10
in the above figure 3 confidence intervals capture the
population mean however one does not. In the long
run if we calculate a 95% confidence interval for
each sample, 95% of them will contain the mean of
the population.

Figure 6.2 Twenty-five samples from the same

population gave these 95% confidence intervals. In
the long run, 95% of all samples give an interval that
covers µ.

Dr. Robert Gebotys Winter 2006

 11

The above picture shows how a confidence interval

behaves. The population distribution of a variable x
has a mean (µ) which is unknown to the researcher.

The researcher, however, wants to know the mean of

the population. The confidence interval gives the
researcher a quantification of uncertainty with
respect to the location of the mean.

For each of 4 samples(hand drawn figure above) the

sample mean is presented with the corresponding
C.I. The 0’s denote the sample means and the solid
line the confidence interval. Note the first three
samples CI contain the population mean whereas the
4th does not.

Note most intervals will contain the population mean

but some will not. In the long run, over many
samples, a 90% C.I. will contain the population
mean 90% of the time. Remember in real life you
will usually calculate only one interval. In the above
context we calculate many to give you a better
understanding of the CI concept.

Dr. Robert Gebotys Winter 2006

 12
The confidence interval is not the probability that the
true mean lies in the interval but the percent of the
time the method gives correct answers. For example:
95% of all samples give an interval that covers µ for
a 95% confidence interval for the mean. Examine
the simulations and convince yourself this is true.

Example

A researcher makes 3 independent measurements on

the concentration of a drug. They are:
.8403 .8363 .8447
You are given σ =.0068, find a 99% CI for µ, the
mean concentration. Calculate x =.8404, z =2.576
for .005 in each tail (2 x .005 = 0.01). A 99% CI for
µ is:
x ± Z σ/√n
.8404 ± 2.576 (.0068) / √3
.8404 ± .0101
We see that (.83030, .85057) is a 99% confidence
interval for the mean drug concentration. 1 - α =
.99, therefore α = .01. The confidence interval is

Dr. Robert Gebotys Winter 2006

 13
symmetric, therefore α/2 = .005. The Z for .005
probability in the tail is 2.576. α is divided by two
because the C.I. is symmetric.
How does the sample size and % confidence effect
the length of the interval?

Figure 6.4 Confidence intervals for n = 4 and n = 1.

Figure 6.5 Confidence intervals for different confidence.

View the Video – Significance Tests

Dr. Robert Gebotys Winter 2006

 14

TEST OF SIGNIFICANCE
X quantitative variable
Suppose X1, X2,...,Xn independent with unknown σ2.
Suppose some theory prescribes a value µo for the
unknown.

Based on the observed data we want to test the

null hypothesis: Ho: µ=µo

against the
alternative hypothesis: Ha: µ=/µo

By 'test' we mean we want to decide whether or not

the data X1, X2,...,Xn provide evidence that Ho is not
true. How should we go about testing Ho? If Ho is
true then CLT says:

Z = x – µ ~ N (0,1)
σ√n
If Ho is true the observed value Zo of Z should be a
reasonable value for the N(0,1) distribution.

Dr. Robert Gebotys Winter 2006

 15

-Zo Zo
If Zo is far out in the tails of the N(0,1) distribution
then we have a value which provides evidence
against Ho.

We measure how far Zo is out in the tails by

computing the approximate p-value or observed
level of significance (OLS). SPSS calls this SIG.

This is the probability that Z is bigger/as big as Zo

(use Z tables) or smaller/as small as -Zo.

If the P-value is small (e.g., p ≤ .05) then Zo is

providing evidence against Ho since this is a rare
event given the mean value. Note, the mean value in
this case comes from the null hypothesis; researchers
assume the null hypothesis is true and they calculate
the chance of obtaining the data they have collected.

The above test is called a 2 sided Z test.

Dr. Robert Gebotys Winter 2006

 16
Example: A developmental psychologist wants to
see if handling during infancy affects infant weight.
n=16 infants and their caregivers are involved in a
special handling course.

The infants are weighed at 2 yrs old x = 28.5 lbs.

We know the weight distribution is N (26,4) for
untreated infants. Does the new handling course
have an impact on the weight of the infants?
Test this hypothesis:

Ho: µ = 26 (mean is equal to 26)

Ha: µ =/ 26 (mean can be greater than
26 or less than 26)
We assume the mean of the population is 26 (that H0
is true). Does the data support this assumption?

X = 28.5 σx = σ = 4 =1
µ = 26 √n √16
We know the location and spread of the sampling
distribution for the mean using the statistics above.
Since we know the location and spread of the
sampling distribution under the assumption µ = 26,

Dr. Robert Gebotys Winter 2006

 17
we can find the probability of obtaining the sample
mean under this assumption. Substitute into Z.

Z = x – µ = 28.5 – 26 = 2.5
σ√n 1

X 26 28.5
Z 0 2.5

The sample mean of 28.5 is 2.5 standard deviations

above the mean. The probability of getting a value
greater than or equal to 2.5 is:

P (Z > 2.5)= .0062

but note Ha is not equal, because the alternative

states that the mean could be bigger or smaller than

Dr. Robert Gebotys Winter 2006

 18
26, we double value to include the possibility of less
than (P (Z < -2.5)).

Therefore: p – value = 2(.0062) = .0124.

This is a small probability or rare event. In other

words, if the population mean is 26 the chance of
obtaining the observed mean (28.5) is small (p-
value=.0124). We conclude that there is strong
evidence against Ho. Handling during infancy affects
the weight of the child.

One-Sided Test

If we are only concerned with alternatives to Ho: µ =

µo that are directional. For example:

Ha : µ > µo (greater than)

Then the value of our test statistic:

Z0 = x – µ
σ√n

Dr. Robert Gebotys Winter 2006

 19
will tend to lie far into the right tail when Ho is false.
In this case the p-value is: P (Z > Zo) and this is
called the one sided Z-test.

Example:
Consider our previous example with children and
handling. The researcher is only concerned with
gains in weight (µ>26).

In other words, the test is one-sided.

Ho: µ = 26
Ha: µ > 26
The calculation of the statistic is identical to our
previous example.

Z = x – µ = 28.5 – 26 = 2.5
σ√n 1

σx = 1

X 26 28.5
Z 0 2.5

Dr. Robert Gebotys Winter 2006

 20

The mean 28.5 is 2.5 standard deviations above the

mean.

P (Z > 2.5)= .0062 (very small probability)

Since test is one-sided (mean is greater than 26) we

do not double the probability to take into account the
left side(mean is less than than 26). The p value is
.0062 and provides very strong evidence against Ho.
Handling does increase weight again.

Example: DRP (Degree of Reading Power scores).

The data for 44 students who have taken the DRP
test is given below.

Data 40 26 ....19
47 19 ....46
52 25 ....52
47 35 ....45

There are n= 44 students. The standard deviation is

given in the question as σ=11 as the population of all
students in the district. The national average is 32 on

Dr. Robert Gebotys Winter 2006

 21
this test. We want to test if this group of students
comes from a population with a higher average.

Ho: µ = 32
Ha: µ > 32 (directional test)

X = 35.091 so we substitute into out

test statistic

Z = 35.091 - 32 = 1.86
11/ √44
We use the Z tables to find the probability to the
right of 1.86 and obtain p-value .031 (small
probability).

Conclusion:
There is strong evidence that the children have a
mean score higher than the national average. A 95%
confidence interval is given below:

35.091 ± 1.96 11/√44

(32.84, 38.35)

Dr. Robert Gebotys Winter 2006

 22

The researcher now knows with 95% confidence the

population mean of the scores is between 32.84 &
38.35.

Tests with a Fixed Significance Level

Suppose we test: Ho: µ = µo

vs
Ha: µ =/ µo

and decide to reject whenever

P(Z > Zo) or P(Z < -Zo) ≤ α

α/2 α/2

-Zα/2 1-α Zα/2

reject Ho reject Ho

critical regions
The regions in the tails of the distribution are called
the critical regions since, if the statistic falls in them,

Dr. Robert Gebotys Winter 2006

 23
the decision is to reject the null hypothesis.
Researchers usually use α = .05 therefore α/2 = .025.
The corresponding Z from tables Z = 1.96. If the Z
statistic is greater than 1.96 or less than -1.96 we
reject the null hypothesis.

-1.96 0 1.96
reject Ho reject Ho

Note: With statistical software readily available the

fixed level test is not used often by researchers since
the p-value is easily computed for an exact
probability rather than the reject/accept at
alpha=some value language used in a fixed test. The
fixed level test was much more popular in the past
however the mechanics of conducting the test are
important for the calculation of power.

Example: Use the previous infant data (continued)

and perform a fixed level test at α=.05.
Ho: µ = 26
Ha: µ =/ 26
Dr. Robert Gebotys Winter 2006
 24

The calculation of the Z statistic is identical to our

previous example.

Z = x – µ = 28.5 – 26
σ√n 1

= 2.5

X 26 28.5
.025 .025
Given α=.05 and the test is 2-sided, then α/2=.025.
We look up Z for .025 probability in the tail and find
Z=1.96 as the critical value.

Z -1.96 0 1.96

Dr. Robert Gebotys Winter 2006

 25
Now is the Z statistic of the data larger than 1.96 or
smaller than -1.96? Does it fall in the rejection
region? Yes! (Z = 2.5). Note you do not have an
exact probability(p-value) only that it is significant
at the alpha specified.

We reject Ho and conclude that handling affects

weight. An equivalent way to carry out a 2-sided test
at level α for µ is to compute a (1-α) CI for µ and
reject if it does not contain µo.

For example: A 95% CI for the infant data is given

below:

X + σ Zα/2
√n
= 28.5 + 4 1.96
√16
= 28.5 + 1.96

The 95% C.I. is (26.54,30.46). Does this interval

contain µ=26?

NO!!! THEREFORE, we reject Ho at α=.05, and

conclude handling affects weight.

Dr. Robert Gebotys Winter 2006

 26

Practical Significance vs Statistical Significance

When we test Ho: µ=µo at level α and have evidence

against Ho we say the results are statistically
significant. Note, if we take n large enough we will
reject Ho where:
Ho: µ = µo + ε
no matter how small ε

ε may not be a difference of practical significance.

We look at the (1-α) CI for practical significance.
Both the p-value/OLS and fixed-level tests of
hypothesis are reported by researchers. With the
advent of readily available computers and software,
the most popular method of reporting is the p-value
or OLS.

Chapter 6 Summary

The confidence interval is used to estimate an

unknown population parameter. The interval gives
the researcher an idea of the accuracy of the
estimate. The confidence level (95%, 99%, etc.)

Dr. Robert Gebotys Winter 2006

 27
gives the probability that the method will provide
the correct answer.
A test of significance weighs the evidence in the
data against a null -hypothesis (Ho) and indirectly in
favor of an alternative hypothesis (Ha). Usually Ho
is a statement that no effect is present and Ha says
the parameter differs from the null value in one
direction (one-sided test) or both (two-sided test).
The p-value is the probability assuming that Ho is
true that the test statistic will take a value as extreme
as that which is observed. Small p-values indicate
strong evidence against H0. However a statistically
significant effect need not be practically important.

Remember: Tips for Success

1) Read the text.

2) Read the notes.
3) Try the assignment.
4) If needed, try the exercise questions.
5) Try the simulations and view the videos if you
need more help with a concept.
6) Try the self tests for practice on each chapter of
the text at www.whfreeman.com/ips.
7) Steady work = Success

Dr. Robert Gebotys Winter 2006