Point and Interval Estimates

Point and Interval Estimates
Suppose we want to estimate a parameter, such as p or , based

on a finite sample of data. There are two main methods:
1. Point estimate: Summarize the sample by a single number
that is an estimate of the population parameter;
2. Interval estimate: A range of values within which, we believe,
the true parameter lies with high probability.
Example. Suppose I wanted to estimate the mean height of all

female students at UNC. I took a sample in this class and the
sample mean was x
= 65.5 (inches). So the obvious thing to
do is to take that as an estimate for the population mean. But
I didnt have to use the sample mean. I could have taken the
sample median (65) or the sample mode (63). It makes sense
to ask which is better.
1
What properties make a good point estimator?

1. Its desirable that the sampling distribution be centered around
the true population parameter. An estimator with this property is called unbiased.
2. Its desirable that our chosen estimator have a small standard
error in comparison with other estimators we might have
chosen.
The sample mean is exactly unbiased (whereas the sample median may not be), and also, if the true population is normal,
the sample mean has a smaller standard error than the sample
median. Both of these would indicate that the sample mean is
preferable to the sample median as an estimator of the population mean. However there are other properties that could nevertheless make the median preferable (e.g. its more resistant to
outliers).
2
In the case of a binomial proportion, the obvious point estimator is the sample proportion. For example, consider our example about President Obamas popularity rating (class posted
03/05/09 Chapter 6 material).
In this example, 68% of respondents gave Obama a positive rating after he had been in office for one month (the answer could
be different if we repreated the poll now). The most natural interpretation of this is that 68% or 0.68 is a statistic which serves
as an estimator of the true but unknown proportion of people
who would have approved of Obama if the whole population had
been surveyed. It seems obvious that we would use the sample
proportion as an estimator of the population proportion, but we
dont have to.
3
Now lets turn to interval estimates. The simplest way to introduce this is through an example.
Example. In a college of 25,000 students, the administration
would like to know for what proportion of students both parents
had completed college. A sample of 350 students was drawn at
random and in that sample, 276 of the students said that both
their parents had completed college.
276 = 0.789 (or 78.9%), so by the
The sample proportion is 350
same logic as in the last example, it makes sense to use that
number also as an estimator of the population proportion (in
this case the population is all 25,000 students at this college).
But, how accurate is that?
The idea: Consider the interval

Sample Proportion 1.96 Standard Error
()
Why would this work?
Sample Proportion 1.96 Standard Error
()
First lets calculate Pr{1.96 < X < 1.96} when X is a standard

normal random variable (mean 0, standard deviation 1).
From the normal table, for z = 1.96 we have left-tail probability
0.9750. For z = 1.96 we have left-tail probability 0.0250. The
difference is 0.9750.025=0.95.
But given that the sample proportion has an approximately normal distribution, this means that the probability that the sample
proportion lies within 1.96 standard errors of the true mean is
also 0.95. Or in other words, the probability that the interval (*)
includes the true mean is 0.95. This is what we mean by saying
that the interval we calculated is a 95% confidence interval.
6
A side comment. Earlier in the course, we said that there is

a 95% chance that a normal random variable lies within two
standard deviations of the mean (this is part of the empirical
rule first discussed in Chapter 2, because although at that time
we didnt use the words normal distribution, thats actually what
the empirical rule refers to). So why have we now replaced the
number 2 with 1.96? Actually, 1.96 is more accurate. If we
repeat the above normal probability calculations with z = 2
instead of z = 1.96, the probability becomes .9772 .0228 =
.9544. Thats still quite close to 95%, and 1.96 is quite close
to 2, so in practice, it doesnt make much difference whether
we use 2 or 1.96 standard deviations. But at this stage of the
course, were trying to be more precise about things than we
were earlier on, hence the change.
Confidence interval for a population proportion

To construct a confidence interval to measure the proportion p of
a population that has a particular characteristic (e.g. supporters
of President Obama):
Step 1: Take a sample of size n, calculate p (pronounced p-hat)
as the sample proportion of people who have that characteristic
(e.g. saying they support Obama in a survey)
r
Step 2: Calculate the standard error SE =

r
Note: The formula should really be

p, so we use p instead.
p(1
p)
n .
p(1p)
n , but we dont know
Step 3: The 95% confidence interval (

p1.96SE, p+1.96SE).
8
Example from text. In one question of the GSS in 2000, 1154

people were asked whether they would be willing to pay higher
prices to support the environment. 518 said yes.
Find a 95% confidence interval for p, the true proportion in the
whole population who would be willing to pay higher prices to
support the environment.
518 = 0.449 to 3 decimal places.

Step 1: p = 1154
Step 2: The standard error is
0.4490.551 = 0.0146.
1154
Step 3: 1.96 0.0146 = 0.029 to 3 decimal places.

The 95% confidence interval is (0.420,0.478).
In practice, we wouldnt usually express this to three decimal
places and simply say that we believe the true proportion of
people who support the proposition (i.e. who would be willing
to pay higher prices to protect the environment) is between 42%
and 48%.
10
A side comment. In another GSS survey people were asked

whether they would support legislation to force industry to adopt
more environment-friendly policies. This time close to 80% answered yes. Yet it seems likely that tighter regulation on industry
will result in higher prices for consumers (this will certainly be
true if you ask the industry representatives). Another case where
the wording of a question arguably influences the answer to a
much greater extent than standard error calculations indicate.
11
Sample size condition

For these calculations to be valid (standard error formula including p in place of p, normal approximation for the distribution of p)
we require the sample size to be reasonably large. In practice,
it is sufficient that n
p 15 and n(1 p) 15.
12
Confidence intervals with other confidence coefficients

So far we have worked with 95% confidence intervals, signifying
that there is supposed to be a 95% probability that the interval
includes the true population parameter. However, theres nothing special here about the probability 95% we could equally
well work with 90%, or 99%, or any other probability we care to
specify.
If we want a 99% confidence interval, we make the same calculation but replace 1.96 by 2.58. If we want a 90% confidence
interval, we make the same calculation but replace 1.96 by 1.645.
See Table 7.2, page 328.
13
Example. In an earlier example, we considered a sample of

350 students from a college (with a total student population
of 25,000) and asked for what fraction of students it was true
that both their parents had been to college. In that case, the
sample proportion was .789. Suppose we wanted a 90% confidence interval.
r
.789(1.789)
The standard error is
= .0218 multiply by
350
1.645, the margin if error is 0.036 to three decimal places. Thus
the 95% confidence interval is (.753, .825).
14
Interpretation of a confidence interval

Continue the previous example: Does the answer mean there is a
90% chance that the true proportion is between .753 and .825?
Strictly speaking, such a statement doesnt make sense were
talking about a finite population of parents; either they went
to college or they didnt; what does it mean to talk of a 90%
chance?
What the 90% confidence statement really means is that in many
repetitions of the procedure, the interval will cover the true value
90% of the time.
15
Example: Lets suppose the true proportion is 80%. I ran 10

simulations experiments where I generated a binomial random
variable with n = 350, p = 0.8. The results were: 272 285 266
289 287 278 269 277 285 271.
I now constructed a 90% confidence interval for each of the ten
hypothetical samples. The results were:
X
272
285
266
289
287
Lower
Bound
0.741
0.780
0.722
0.792
0.786
Upper
Bound
0.814
0.848
0.798
0.859
0.854
X
278
269
277
285
271
Lower
Bound
0.759
0.731
0.756
0.780
0.738
Upper
Bound
0.830
0.806
0.827
0.848
0.811
The interval is slightly different each time, and in fact, in 9 out

of the 10 cases the interval covered the true value 0.8.
16
Conclusion: The interval is random. the confidence coefficient

(in this case, 90%) represents the long-run probability that the
interval would cover the true value in many repetitions of the
sampling procedure.

Point and Interval Estimates

Uploaded by

Copyright:

Available Formats

Point and Interval Estimates

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Point and Interval Estimates

Uploaded by

Copyright:

Available Formats

Point and Interval Estimates

Suppose we want to estimate a parameter, such as p or , based

Example. Suppose I wanted to estimate the mean height of all

What properties make a good point estimator?

The idea: Consider the interval

Why would this work?

Sample Proportion 1.96 Standard Error

First lets calculate Pr{1.96 < X < 1.96} when X is a standard

A side comment. Earlier in the course, we said that there is

Confidence interval for a population proportion

Step 2: Calculate the standard error SE =

Note: The formula should really be

Step 3: The 95% confidence interval (

Example from text. In one question of the GSS in 2000, 1154

518 = 0.449 to 3 decimal places.

Step 2: The standard error is

Step 3: 1.96 0.0146 = 0.029 to 3 decimal places.

A side comment. In another GSS survey people were asked

Sample size condition

Confidence intervals with other confidence coefficients

Example. In an earlier example, we considered a sample of

Interpretation of a confidence interval

Example: Lets suppose the true proportion is 80%. I ran 10

The interval is slightly different each time, and in fact, in 9 out

Conclusion: The interval is random. the confidence coefficient

You might also like