Chapter 5 - Random Sapling

Statistics Pre-Term Program – Chapter Five
Random Sampling and Sampling Distributions
CHAPTER 5
RANDOM SAMPLING AND
SAMPLING DISTRIBUTIONS
TABLE OF CONTENTS
Page
1 Selection Bias and Sampling 144
Key Definitions, Representative Samples, Simple Random Samples
2 The 1970 Draft Lottery Example and More Sampling Concerns 145
Problems with Mechanical Selection Procedures, Miscellaneous Sampling Issues
3 Disastrous Sampling Stories: The 1936 Literary Digest 147
4 Disastrous Sampling Stories: The New Hite Report 143
5 Disastrous Sampling Stories: A Meaningless Poll 149
6 Sampling Distributions 152
Sampling Distribution of Sample Mean from a Normal Population, Probable Error
7 The Central Limit Theorem 158
Sampling from ANY population, How Large a Sample Do We Need?, Examples
8 Cents and the Central Limit Theorem – Ages of 800 Pennies 159
9 CLT through Simulation 162
10 Four (Straightforward) Questions on Sampling Distributions 165
11 Two Exercises in Sampling Distributions 165
Answer Sketches
Statistics Pre-Term Program - Chapter 5 - Page 143

Selection Bias and Sampling
We now return to the primary objective of statistics -- namely, the use of sample information to
infer the nature of a population. Predicting your company’s annual sales from its sales records
of the last 10 years, the annual return on an investment from the results of a number of similar
investments, and an evening’s winnings at blackjack based on your previous (successful and
unsuccessful) treks to the casino are examples of statistical inferences, and each involves an
element of uncertainty.
Defining a Common Language about Sampling
• The population (or universe) is the entire collection of units, (individuals or objects or the
list of measurements) about which we would like information.
• The sample is the collection of units we will actually measure or the collection of
measurements we will actually obtain.
• The sampling frame is a list of units from which the sample is chosen. Ideally, it includes
the whole population.
• In a sample survey, measurements are taken on a sample from the population.
• A census is a survey in which the entire population is measured.
A parameter is a numerical descriptive measure of a population. A sample statistic is a

numerical descriptive measure of a sample. Since it is almost always too costly and/or time-
consuming to conduct a census of a population and compute its parameters, we use the
information contained in a sample to make inferences about the parameters of a population.
How a sample is selected from a population is of vital importance. Since statistics vary from
sample to sample, any inferences base on them will necessarily be subject to some uncertainty.
How, then, do we judge the reliability of a sample statistic as a tool in making an inference about
the corresponding population parameter?
First, some basic notions of sampling. We’d like the sample to be representative of the
population. If we select n elements as a sample from a population, how can we tell whether the
sample is representative? Sometimes it is hard to judge whether a sample is representative.
What other criteria could be used to select valid samples from a population of interest?
One of the simplest and most frequently used sampling procedures produces what is known as
a random sample. Suppose we take a sample of n elements from a population. If some
samples are more likely to be selected than others, then the sampling is not random. If our
method of sampling ensures that every possible combination of n elements in the population has
an equal chance of being selected, the n elements are a random sample. Strictly, speaking, this
is a simple random sample.

Example 5.1. 1970 Draft Lottery1 [LOTTERY data file]
From 1948 through about 1969, men were drafted by age -- oldest first, starting with 25-year
olds. On December 1, 1969, the Selective Service System conducted a lottery to determine
the order of selection for 1970. 366 possible days in a year were written on slips of paper and
placed in egg-shaped capsules. The capsules were then drawn one by one to obtain the order
of induction. So the lottery assigned a rank to each of the 366 birthdays. Does this seem like a
fair method of determining the order of selection? Below is the summary of the data for the
resulting order. Does it appear to provide a fair order of selection?
Time Series (Overlay) Plot of 1970 Lottery Data
350
300
250
200
150
100
50
0
0 50 100 150 200 250 300 350
Order
Month Mean SD Min 1st Q Median 3rd Q Max

January 201.16 99.71 17 118 211 305 355
February 202.97 103.96 4 117.5 210 294.5 365
March 225.81 95.83 29 166 256 300 362
April 203.67 109.37 2 88.25 225 282.75 351
May 207.97 114.97 31 103 226 313 364
June 195.73 117.87 20 82 207.5 309.5 366
July 181.55 109.61 13 88 188 284 350
August 173.45 112.73 11 61 145 291 352
September 157.30 87.15 1 79.25 168 232.25 315
October 182.45 96.78 5 94 201 244 359
November 148.73 94.40 9 73.5 131.5 209.75 348
December 121.55 95.06 3 43 100 163 328
1
Adapted from McClave & Benson, case study 3.2.

Even though the process itself seems to be fair, there is a general downward trend in monthly
mean and median selection orders (thus disproportionate number of men born later in the year
were selected early in the draft). To see this clearly, consider the scatter plot below of the mean
rank by month.
1970 Draft Lottery
250
Average selection number
200
183.5
150
Monthly Averages
100
50
0
0 1 22 34 4 56 6 87 8 10 9 1012 11 14
12
month
If the lottery were truly random than one would expect that the monthly means would be fairly
close to the average value of the selection order =
(1 + 2 + 3 + 4 + … +365+366)/366 = 183.5
and they would show no pattern from month to month.
Why did method that seemed so fair produce such a bias? The problem was with the bowl.
The January capsules had all been put in first, the February capsules were added next, and so
on. The capsules were then mixed but not thoroughly enough. There was still a tendency for
dates early in the year to be near the bottom of the bowl. Compounding the problem was that
the bowl’s compactness made it physically difficult to reach in and select a capsule from beneath
the topmost layers. As a result capsules were more or less chosen from the top down. Since
composition of the layers was not random because of the earlier inadequate mixing, the final
sequence contained the bias evident in the picture above.

Why Randomize?
Results from polls and other statistical studies reported in newspapers or magazines often
emphasize that the samples were randomly selected. Why the emphasis on randomization?
Couldn’t a good investigator do better by carefully choosing respondents to a poll so that
various interest groups were represented? Perhaps, but samples selected without objective
randomization tend to favor one part of the population over another. For example, polls
conducted by sports writers tend to favor the opinions of sports fans. This leaning toward one
side of an issue is called sampling or selection bias. In the long run, random samples seem to
do a good job of producing samples that fairly represent the population. In other words,
randomization reduces sampling bias. Note that random sampling can be tough. A random
sample is not a casual or haphazard sample. The target population must be carefully identified,
and an appropriate sampling frame must be selected.
SELECTION BIAS A systematic tendency to over-represent some segment of the

population. (Some subject or group of subjects is more likely to be selected by the sampling
method than another.)
Example 5.2. Literary Digest 1936

In 1936, the Literary Digest mailed 10 million questionnaires to U.S. voters, asking them
whether they preferred Franklin Delano Roosevelt or Alf Landon. The Digest had been
predicting elections since 1916 and boasted of their accurate projections. Received 2.4 million
responses -- a sample size approximately 800 times larger than the Gallup Poll.
Sample Results: Landon 57% FDR 43% Actual Election: FDR 62% Landon 38%
What happened?
• Bad sampling frame -- Literary Digest subscribers were not representative of the
American voting population.
• Nonresponse bias -- 76% of subscribers who didn’t return the survey were much more
likely to support FDR.
• Also, Landon’s supporters, desiring change, tended to be more vocal in their support of
their candidate than did FDR’s supporters.
Difficulties and Disasters in Sampling (Utts, p. 62)

Difficulties:
1. Using the wrong sampling frame
2. Not reaching the individuals selected
3. Getting no response or getting a volunteer response
Disasters:

1. Getting a volunteer sample
2. Using a convenience or haphazard sample
Example 5.3. The New Hite Report -- Controversy over the Numbers
[Reference: Streitfeld, D. (1988). “Shere Hite and the trouble with numbers.” Chance, 1, 26-31.]
In 1968, researcher Shere Hite shocked conservative America with her now-famous “Hite
Report” on the permissive sexual attitudes of American men and women. Twenty years later,
Hite was surrounded by controversy again with her book, Women and Love: A Cultural
Revolution in Progress (Knopf Press, 1988). In this new Hite report, she reveals some
startling statistics describing how women feel about contemporary relationships:
• 84% of women are not emotionally satisfied with their relationship

• 95% of women report “emotional and psychological harassment” from their men
• 70% of women married 5 years or more are having extramarital affairs
• Only 13% of women married more than 2 years are “in love”
Hite conducted the survey by mailing out 100,000 questionnaires to women across the country
over a 7-year period. Each questionnaire consisted of 127 open-ended questions, many with
numerous subquestions and follow-ups. Hite’s instructions read: “It is not necessary to answer
every question! Feel free to skip around and answer those questions you choose.”
Approximately 4,500 completed questionnaires were returned for a response rate of 4.5%, and
they form the data set from which these percentages were determined. Hite claims that these
4,500 women are a representative sample of all women in the United States, and therefore, the
survey results imply that vast numbers of women are “suffering a lot of pain in their love
relationships with men.” Many people disagree, however, saying that only unhappy women are
likely to take the time to answer Hite’s 127 essay questions, and thus her sample is
representative only of the discontented.
The views of several statisticians and expert survey researchers on the validity of Hite’s
“numbers” were presented in an article in Chance magazine (Summer 1988). A few of the
more critical comments follow.
• Hite used a combination of haphazard sampling and volunteer respondents to collect her
[data]. First, Hite sent questionnaires to a wide variety of organizations and asked them to
circulate the questionnaires to their members. She mentions that they included church groups,
women’s voting and political groups, women’s rights organizations and counseling and walk-in
centers for women. These groups seem not representative of women in general; there is an
over-representation of feminist groups and of women in troubled circumstances.... The use of
groups to distribute the questionnaires meant that gatekeepers had the power of assuring a
zero response rate by not distributing the questionnaire, or conversely of greatly stimulating
returns by endorsing the study in some fashion. Second, Hite also relied on volunteer
respondents who wrote in for copies of the questionnaire. These volunteers seem to have
been recruited from readers of her past books and those who saw interviews on television and

in the press. This type of volunteer respondent is the exact opposite of the randomly selected
respondent utilized in standard survey research and even more potentially unrepresentative
than the group samples cited above. (Tom Smith, National Opinion Research Center)
• So few people responded, it’s not representative of any group, except the odd group who
agrees to respond. Hite has no assurance that even her claimed 4.5% response rate is
correct. How do we know how many people passed their hands over these questionnaires.
You don’t want to fill it out, you give it to your sister, she gives it to a friend. You’ll get one
response, but that questionnaire may have been turned down by five people. (Donald Rubin,
Harvard University)
• When you get instructions to answer only those questions you wish to, you’re likely to skip
some. Isn’t it more likely that, for example, a woman who feels strongly about affairs would
more likely answer questions on that subject than a woman who does not feel as strongly?
Thus, her finding that 70% of all women married over five years are having affairs is
meaningless because she does not report how many people answered each question. I cannot
tell whether this means 70% of 1000 women or 70% of 10 women. (Judith Tanur, State
University of New York at Stony Brook)
• Even in good samples, where you have a 50% or 70% response rate, you usually have some
skews -- say, with income, race, or region. If she can do a sample like this, she’s got the
Rosetta Stone, and I’ll come study with her. (Martin Frankel, Baruch College,
commenting on Hite’s claim that her sample matches that of the U.S. female population in
terms of demographic balance.)
• According to Hite, whether you’re 18 or 71, you’re going to answer the questions the same
way. Whether white, black, Hispanic, Middle Eastern, or Asian American, you’re going to
answer the same way. Whether you make $5,000 a year or over $75,000, you’ll answer the
questions the same way. I’ve never seen anything like this in my career -- and the Kinsey
Institute collects data from everybody. (June Reinisch, Kinsey Institute, Indiana
University, commenting on Hite’s numbers showing that no matter what the demographic
breakdown of the women married 5 years or more, about 70% are having extramarital
affairs.)
Example 5.4. A Meaningless Poll (Utts, pp. 64-65)
On February 18th, 1993, shortly after Bill Clinton became President of the United States, a
television station in Sacramento, California asked viewers to respond to the question: “Do you
support the President’s economic plan?” The next day the results of a properly conducted
study asking the same question were published in the newspaper.
Television Poll Survey

Yes 42% 75%
No 58% 18%
Not sure 0% 7%
As you can see, those who responded to the television poll were more likely to be those who
were dissatisfied with the President’s plan. Trying to extend those results to the general

population is misleading. It is irresponsible to publicize such studies, especially without a
warning that they result from an unscientific survey and are not representative of general public
opinion. You should never interpret such polls as anything other than a count of who bothered
to go to the telephone and call.
Haphazard/Convenience Sampling
A few years ago, the student newspaper at a California university announced as a front page
headline: “Students ignorant, survey says.” The article explained that a “random survey”
indicated that American students were less aware of current events than international students
were. However, the article quoted the undergraduate researchers, who were international
students themselves, as saying that “the students were randomly sampled on the quad.” The
quad is an open air area where students relax, play frisbee, eat lunch, and so on. There is
simply no proper way to collect a random sample of students by selecting them in an area like
that. In such situations, the researchers are likely to approach people whom they think will
support the results they intended for their survey. Or, they are likely to approach friendly
looking people who look like they will cooperate. This is called a haphazard sample, and it
cannot be expected to be representative.
Nonprobability Sampling
• Judgment Sample: Asking various newspaper food critics about which restaurants they
would recommend
• Quota Sample: If a school has 30% off-campus and 70% on-campus students, attempting
to get 30 on- and 70 off-campus students when doing a survey of 100 students
• Chunk Sample: If I want to know what students think of a textbook quickly, I could use
this class to represent all students using the book.
• The only way to make correct statistical inferences from a population to a sample is to use
probability sampling...
Investigators Led Astray By Nonrandom Samples2
Wainer, Palmer and Bradlow [WPB] (1998) presented a series of situations in which the use of
nonrandom samples led or could lead to seriously incorrect conclusions.
• The Most Dangerous Profession In 1835, H. C. Lombard published a study on the

mean longevity of professions, based on death certificates gathered over more than a half
century in Geneva. Each certificate contained the deceased’s profession, and age of death.
Lombard found that the average age of death for the professions ranged from the early 50s
2
This material is drawn from Wainer, H., Palmer, S. and Bradlow, E. (1998). “A selection of selection
anomalies.” Chance, 11(2): 3-7.

to the mid 60s. There was one surprise: the most dangerous profession – with the lowest
longevity – was “student” with an average age of 20.7!
• The Twentieth Century Is A Dangerous Time WPB collected 204 birth and death
dates from the Princeton (NJ) Cemetery. When age at death was plotted as a function of
birth year, they found that age of death stayed relatively constant until 1920, when longevity
begins to decline rapidly. The average age of death decreases from around 70 years of age
in the 1900s to as low as 10 in the 1980s. There must be a reason for this anomaly in the
data (what WPB call the “Lombard Surprise”) but what?
As WPB indicate, “data obtained by nonrandom sampling occurs often in practice…. We

address here … data that have been obtained through self-selection; that is, inclusion in the
sample is determined by the units themselves and not the data gatherer. In such cases, valid
inferences require careful thought about the character of the selection process…. Ignoring self-
selection can lead to flawed, often ridiculous findings.”
Asking Questions
Suppose we want an honest reading on the number of American citizens who cheat on their
taxes? How do we ask people if they cheat?
How you ask the question can seriously affect the answer…
1. Do you agree that unions can cause inconvenience and bad labor-management
relations?
2. Do you agree that unions have been important in securing employee rights and decent
pay?

Sampling Distributions
A sample statistic is a random variable, and thus must be compared and judged on the basis of
its probability distribution. The probability distribution of a sample statistic is called its sampling
distribution.
The sampling distribution of a sample statistic (based on n observations) is the relative frequency
distribution of the values of the statistic theoretically generated by taking repeated random
samples of size n and computing the value of the statistic for each. To illustrate the difference
between distribution of some random variable and sampling distribution of statistic of this
random variable lets consider following example.
Example 5.5
Consider population created by tossing a fair coin infinitely many times. Let random variable X
take value 1 if coin shows tails and 2 if coin shows heads. The probability distribution of X is
x 1 2
p(x) 0.5 0.5
Probability distribution of a coin flip
0.6
0.5
probability
0.4
0.3
0.2
0.1
0
1 2
Series1 0.5 0.5

coin faces
After reading Chapter 3 of this bulkpack you should be able to compute that expected value of
X is 1.5 and variance of X is 0.25.
Now suppose that you did not know the mean value of X . You could draw a sample (say of
size n = 4) from the population (by flipping coin 4 times) and use sample mean to estimate the
population mean. The following table summarizes all possible samples.

Sample Sample mean

1,1,1,1 1
1,1,1,2 1.25
1,1,2,1 1.25
1,1,2,2 1.5
1,2,1,1 1.25
1,2,1,2 1.5
1,2,2,1 1.5
1,2,2,2 1.75
2,1,1,1 1.25
2,1,1,2 1.5
2,1,2,1 1.5
2,1,2,2 1.75
2,2,1,1 1.5
2,2,1,2 1.75
2,2,2,1 1.75
2,2,2,2 2
Each of these samples has equal chance to come up. Therefore, probability that any one of
them occurs is 1/16. Note, however that some of the sample mean values appear more than
once. For example, sample mean 1.25 appears 4 times in this table. Hence, it is 4 times more
likely to happen compared to sample mean 1 that appears only once. Therefore, probability
that sample mean of 1.25 will occur is 4*1/16 =1/4 =0.25. Repeating this argument for other
values of sample mean we obtain the following probability distribution of sample mean.
Sample mean 1 1.25 1.5 1.75 2
Probability 0.0625 0.25 0.375 0.25 0.0625

Distribution of sample mean
0.4
0.35
0.3
0.25
probability
0.2
0.15
0.1
0.05
0
1 1.25 1.5 1.75 2
Series1 0.0625 0.25 0.375 0.25 0.0625
value of mean
Note that, first of all, sample mean is a random variable itself. Second, distribution of sample
mean is very different from the distribution of X. Third, note the bell shape of the sample mean
distribution. In fact, the larger the sample size n is the more sample mean distribution will
resemble normal distribution (this is the essence of the Central Limit Theorem that we will
discuss later in the chapter).
Having the distribution for sample mean we can calculate expected value and variance of the
sample mean.
E(sample mean) = 1*0.0625 + 1.25*0.25 + 1.5*0.375 + 1.75*0.25 + 2*0.0625 = 1.5

Var(sample mean) = 0.0625
So, E(sample mean) = E(X), and Var(sample mean) < Var(X). Note that variance of sample
mean of X is less than the variance of X itself. This is not a coincidence, in fact, we will see
later that sample of size n, Var(sample mean)=Var(X)/n.

The Sampling Distribution of the Sample Mean
Regardless of the shape of the population’s distribution:

1. The mean of the sampling distribution of X will equal µ, the population mean.
2. The standard deviation of the sampling distribution of X is called the standard error of the
sample mean, and will equal σ, the population standard deviation, divided by the square
root of the sample size n.
σ
Standard Error of the Sample Mean = σ X = pop
n
Example 5.6. Telephone Company3
A telephone company knows that during non-holidays the number of calls that pass through the
main branch office each hour has a Normal distribution with µ = 80,000 and σ = 35,000.
Describe the mean, standard deviation and shape of the sampling distribution of X , the mean
number of incoming calls per hour for a random sample of 60 non-holiday hours.
Solution of example 5.6:

E( X ) = µ = 80,000
σ 35,000
σ X = pop = = 4518.481
n 60
In order to describe shape of the distribution of X , we will use the following theorem.
Theorem One: Sampling from a Normally Distributed Population
What else might we want to know about the sampling distribution? What can we say about the
shape of the sampling distribution of the sample mean?
Theorem: If a random sample of n observations is taken from a population with a Normal

distribution, the sampling distribution of X will also be a Normal distribution.
Solution of example 5.6 (continues)

Since population has normal distribution we conclude that X is normally distributed too. This
along with the values of E( X ) and σ X that we computed earlier allows us to make the
following statement: X has normal distribution with mean 80,000 and standard deviation
4518.481.
3
Adapted from Sincich, exercise 7.24.

Sample Statistics and Sampling Distributions
Probable Error for a Sample Mean
There is no systematic tendency to over- or underestimate µ with Y . This wouldn’t be much of

a consolation if we knew that half the time we made a huge overestimate and the other half an
equally huge underestimate. The standard error of the sample mean, σ Y , in conjunction with
the Empirical Rule, can be used to give a good indication of the probable deviation of a
particular sample mean from the population mean.
Example 5.7. (H&O, BSIFM, Example 6.3)
Suppose that a supermarket manager is interested in estimating the mean checkout time for the
non-express checkout lanes. An assistant manager obtains a random sample of 25 checkout
times. If previous data suggest that the population standard deviation is 1.10 minutes, describe
the probable deviation of Y from the unknown population mean µ.
Solution to Example 5.7. The Empirical Rule indicates that approximately 95% of the time Y
( )
is within two standard errors 2σ Y of the population mean µ. For n = 25,
2σ pop 2(110 . )
2σY = = =.44
n 5
The probable error for Y is no more than .44 minute.
The probable accuracy of a sample mean, as measured by its standard error, is affected by the
sample size. Because the standard error of the sample mean is the population standard
deviation divided by the square root of the sample size, the standard error decreases as the
sample size increases. For example, if the sample size had been either 50 or 100 instead of 25
( )
in Example 5.8., the probable errors 2σ Y would have been, respectively, .31 or .22.
Example 5.8. (H&O, Exercise 6.16)
The average demand for rental skis on winter Saturdays at a particular area is 148 pairs, which
has been quite stable over time. There is variation due to weather conditions and competing
areas; the standard deviation is 21 pairs. The demand distribution seems to be roughly Normal.
a. The rental shop stocks 170 pairs of skis. What is the probability that demand will exceed
this supply on any one winter Saturday?

b. The shop manager will change the stock of skis for the next year if the average demand over
the 12 winter Saturdays in a season (considered as a random sample) is over 155 or under
135. These limits aren’t equidistant from the long-run process mean of 148 because the
costs of oversupply and undersupply are different. If the population mean stays at 148,
what is the probability that the manager will change the stock?

The Central Limit Theorem: Sampling from Any Population
Basically, the CLT states that if we have a large enough sample taken from any population, not
just a Normally distributed one, then the sampling distribution of the sample mean will also be
Normally distributed.
More formally, if a random sample of n observations is selected from essentially any population,
then, when n is sufficiently large, the sampling distribution of X will be approximately Normal.
The quality of the Normal approximation to the sampling distribution of X depends on two
things: the size of the sample, and the shape of the underlying distribution of the population.
• The larger the sample size n, the better the Normal approximation to the sampling
distribution of X
• Generally speaking, the greater the skewness of the underlying population, the larger the
sample size must be to obtain an adequate Normal approximation.
How Large is Large Enough?

A good rule of thumb is based on plots of the sample data.
• If a plot of the sample data shows severe skewness, it is reasonable to assume that the
underlying population is severely skewed, and the Normal approximation to the sample
mean’s sampling distribution is not appropriate unless n is at least 100.
• For mild skewness, n = 30 should generally be sufficient to make the Normal
approximation to the sampling distribution appropriate.
• For symmetric but outlier-prone data, n = 15 sample should be enough for the CLT effect
to make the Normal approximation reasonable.
• For normalish data, n = 1 is sufficient.
Example 5.9. (H&O, Example 6.7)
In the supermarket checkout time situation of Example 5.8., the following actual times in minutes
were observed (n = 25): .4, .4, .5, .5, .5, .6, .6, .7, .8, .9, 1.1, 1.2, 1.4, 1.5, 1.8, 2.0, 2.3, 2.6,
2.9, 3.4, 4.2, 5.0, 6.6, 9.2, 16.3 ( y = 2.70, median = 140 . , s = 3.56) . Does it appear that a
normal approximation to the sampling distribution of Y (for future samples of size n = 25, for
instance) would be satisfactory?

Checkout histogram
20
18
16
14
Frequency
12
10 Frequency
8
6
4
2
0
0.4 3.58 6.76 9.94 13.12 More
Bin
Solution to Example 5.9. The sample data suggest that the population distribution of checkout
times is likely to be highly skewed. Most times are quite brief, but there are a few people who
really slow things up. (See histogram) A sample of 25 is not enough to deskew the sampling
distribution. Therefore, the Empirical Rule probabilities (which are based on the Normal
approximation) in Example 5.8. are most likely inaccurate for n = 25.
Example 5.10. Cents and the CLT – A Lesson worth $8.00
As a consistent example, I collected a sample of 800 pennies in the summer of 1996, and
stored the data in the CENTS file available on the Course Volume. From the original 800
measurements, samples of size 5, then of size 10, 25 and 40 were drawn (in files
CENTS1..CENTS4), and descriptive statistics calculated:
VARIABLE N MEAN SD MINIMUM MAXIMUM DATA FILE

AGE 800 10.706 9.0771 0.0000 43.000 CENTS
MEAN of 5 160 10.706 4.1451 1.8000 23.800 CENTS1
MEAN of 10 80 10.706 3.0946 4.1000 17.400 CENTS2
MEAN of 25 32 10.706 1.9389 7.4000 14.640 CENTS3
MEAN of 40 20 10.706 1.4406 7.7250 13.100 CENTS4

An exploratory analysis of the 800 penny ages in the CENTS file is shown below.
AGE
0 5 10 15 20 25 30 35 40 45
Below are histograms and outlier boxplots of the sample means of sizes 5, 10, 25 and 40. Note
that they have been drawn on the same x-axis as the individual ages were above. Note the
CLT effect on the shape, the center and the spread of the distribution.
MEAN5
0 5 10 15 20 25 30 35 40 45

MEAN10
0 5 10 15 20 25 30 35 40 45

MEAN25
0 5 10 15 20 25 30 35 40 45
MEAN40
0 5 10 15 20 25 30 35 40 45
Simulation Demonstrations: CLT: How Large is Large Enough?
• We begin with histograms of 1000 means sampled from an exponential population.

• Note that sample means have a distribution, different than the population. Even for
n = 4, difference in shape is clear.
• The average value of the sample means is about 1.0 = population mean for standard
exponential.
• Comparison of scales of the four histograms indicates that variability of means
decreases with increasing sample size, but not in proportion to the sample size. (In
particular, width of scale for n = 30 is not half the width of the scale for n = 60).
• Finally, the CLT effect on the shape of the distribution is obvious.

Next, we summarize (means and standard deviations of sample means) to reinforce ideas.
• The mean of means is a simulation approximation to the expected value. It’s
approximate because we have 1000 samples, not an infinite number.
• The standard deviation of means is a simulation approximation to the standard error
of the mean. The values are quite close to the theoretical values; note that the
population standard deviation in this case is 1.
1. Summary Results for Mean of Exponential Sample
Simulation of expected value and standard error (1,000 samples)

Population shape is skewed. µ = 1.000, σ = 1.000
Summary statistics for sampling distribution of the mean:
n Mean Std Dev
4 0.994 0.4747
10 0.998 0.3101
30 1.003 0.1830
60 1.003 0.1278
• Next, we use samples from a squared exponential (highly skewed) distribution to

demonstrate that the rule n > 30 won’t work for extreme skew.
• Finally, we use samples from a Laplace distribution to show rapid convergence to Normal
for outlier-prone, but symmetric populations
2. Summary Results for Mean of Squared Exponential Sample

Population shape is highly skewed. µ = 2.000, σ = 4.472
n Mean Std Dev
4 2.017 2.2875
10 2.079 1.5147
30 2.037 0.8121
60 2.023 0.5783
3. Summary Results for Mean of Laplace Sample

Population shape is outlier prone. µ = 0.000, σ = 1.4142
n Mean Std Dev
4 -0.023 0.7029
10 0.010 0.4366
30 0.003 0.2639
60 -0.002 0.1845

Example 5.11. Automobile Batteries
A manufacturer of automobile batteries claims that the distribution of the lifetimes of its best
battery has a mean of 54 months and a standard deviation of 6 months. A consumer group
purchases a sample of 50 batteries and determines their lifetimes.
a. Assuming the manufacturer’s claim is true, describe the sampling distribution of the
mean lifetime of a sample of 50 batteries.
A sample of 50 batteries will produce a sampling distribution of the sample mean which follows
• a Normal distribution, (assuming that the population of batteries is not grossly
skewed), with
• Expected value E( X ) = µpop= 54
σ 6
• and Standard error of the mean SE( X ) = pop = = 0.85
n 50
b. Assuming the manufacturer’s claim is true, what is the probability that the group’s
sample has a mean lifetime of 52 hours or less?
We need to find the probability that the sample mean of 50 batteries, X , is 52 or less. But,
assuming the manufacturer’s claim is true, we know that the sample mean X follows a Normal
distribution with mean E( X ) = 54 and standard error SE( X ) = 0.85. So we have
X−E X
Pr( X < 52) = Pr 
[ ]
<
52 − E X [ ]
[ ]
 SE X SE X [ ] 
 52 − 54 
= Pr  Z < 
 0.85 
= Pr ( Z < −2.35)
Now Pr(Z < -2.35) can be found using Table 2, and is equal to .0094, which is the probability
that the group’s sample has a mean lifetime of 52 hours or less, assuming the manufacturer’s
claim is true.
Other Versions of the CLT
Versions of a CLT also apply to sample statistics other than sums and means, but the Normal
distribution does not always apply. Mere largeness does not imply Normality, unless a sum or
average is involved. The CLT guarantees that the distribution of the sample mean will be
Normally distributed, even though the individual values may be quite skewed. The best advice
is to draw pictures...

Four Questions on Sampling Distributions
1. Suppose we draw a sample of n = 64 observations from a new population, with µ = 980
and σ = 300.
a. Find Pr( X > 1,050) b. Find Pr( X < 960) c. Find Pr( X > 1,100)
2. An automatic machine in a manufacturing process is operating properly if the lengths of an

important subcomponent are Normally distributed, with mean µ = 117 cm and standard
deviation σ = 2.1 cm.
a. Find the probability that a randomly selected unit has a length greater than 120 cm.
b. If 3 units are randomly selected, what is Pr(their mean length exceeds 120 cm.)?
3. The manufacturer of cans of salmon that are supposed to have a net weight of 6 ounces tells
you that the net weight is actually a random variable with mean 6.05 ounces and standard
deviation of 0.18 ounce. Suppose you take a random sample of 36 cans.
a. Find the probability that the sample mean will be less than 5.97 ounces.
b. Suppose your random sample of 36 cans produces a mean of 5.95 ounces. Comment on
the statement made by the manufacturer.
4. The sign on the elevator in a large skyscraper states “Maximum capacity 2,500 pounds, or
16 persons.” A professor of statistics wonders what the probability is that 16 people would
weight more than 2,500 pounds. If the weights of the people who use the elevator are
Normally distributed, with a mean of 150 pounds and a standard deviation of 20 pounds, what
is the probability that the professor seeks?
Two Exercises in Sampling Distributions

Example 5.6. Telephone Company (Continued)
Suppose the telephone company wishes to determine whether the true mean number of
incoming calls per hour during holidays is the same as for non-holidays. To accomplish this, the
company randomly selects 60 hours during a holiday period, monitors the incoming phone calls
each hour, and computes y , the sample mean number of incoming phone calls. If the sample
mean is computed to be y = 91,970 calls per hour, do you believe that the true mean for
holidays is µ = 80,000 (the same as for non-holidays)? Assume that the standard deviation of
the number of incoming calls per hour for holidays is 35,000.

Example 5.12. Risk-Taking In Successful Entrepreneurs 4
A distinguishing characteristic of entrepreneurs is their propensity for taking risks. R. H.

Brockhaus used a choice dilemma questionnaire (CDQ) to measure the risk-taking propensities
of successful entrepreneurs (Academy of Management Journal, September 1980). He found
that the CDQ scores of entrepreneurs had a mean of 71 and a standard deviation of 12.
(Lower scores are associated with a greater propensity for taking risks.) Let y be the mean
CDQ score for a random sample of n = 50 entrepreneurs.
a. Describe the sampling distribution of y .

b. Find Pr( 69 ≤ y ≤ 72) .
c. Find Pr( y ≤ 67) . Would you expect to observe a sample mean CDQ score of 67 or
lower? Explain.
Answer Sketches for Sampling Distribution Exercises
Telephone Company
To assess whether 80,000 seems reasonable, we need to calculate the probability of obtaining a
sample mean of 91,970 or larger when µ = 80,000. So we assume that µpop = 80,000 and that
σpop = 35,000 and take a sample of n = 60 hours. Again, with a sample of size n = 60, we
need only to assume that the population of holiday hour incoming calls are not grossly skewed
in order to allow the CLT effect to occur.
If this is true, then the sampling distribution of y is Normal, with Expected value E( y ) = µpop=
σ 35,000
80,000 calls and Standard error SE( y ) = pop = = 4,518 . Now, we want to find the
n 60
probability of obtaining a sample mean of 91,970 or larger:
 y − E[ y ] 91,970 − E[ y ]
Pr( y ≥ 91,970) = Pr  ≥ 
 SE[ y ] SE[ y ] 
 91,970 − 80,000 
= Pr  Z ≥ 
 4,518 
= Pr ( Z ≥ 2.65)
and, from Table 2, Pr(Z ≥ 2.65) = 1 - .9960 = .0040, which seems awfully small, and thus
suggests that the mean number of calls on holiday hours is higher than 80,000.
Risk-Taking In Successful Entrepreneurs
4
Adapted from Sincich, exercise 7.18.

a. Assuming that the distribution of entrepreneur’s CDQ scores in the population is not
grossly skewed, the sampling distribution of y should be Normal, thanks to the CLT.
The expected value of y should be equal to the population mean, 71, and the standard
12
error of y should be = 1.70 .
50
b.
 69 − E[ y ] y − E[ y ] 72 − E[ y ] 
Pr( 69 < y < 72) = Pr  < < 
 SE[ y ] SE[ y] SE[ y ] 
 69 − 71 72 − 71
= Pr  <Z< 
 17 . . 
17
= Pr ( −118
. < Z < 0.59)
and, from Table 2, Pr(-1.18 < Z < 0.59) = .8810 - .2776 = .6034
c.
 y − E[ y ] 67 − E[ y ] 
Pr( y ≤ 67) = Pr  ≤ 
 SE[ y ] SE[ y ] 
 67 − 71
= Pr  Z ≤ 
 . 
17
= Pr ( Z ≤ −2.35)
and, from Table 2, Pr(Z ≤ -2.35) = .0094. Thus, we’d expect a sample mean CDQ
score of 67 or lower to occur less often than 1 in 100 times. It would be a real surprise
to observe such a result.

Chapter 5 - Random Sapling

Uploaded by

Copyright:

Available Formats

Chapter 5 - Random Sapling

Uploaded by

Document Information

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

Chapter 5 - Random Sapling

Uploaded by

Copyright:

Available Formats

Statistics Pre-Term Program – Chapter Five

Random Sampling and Sampling Distributions

Statistics Pre-Term Program - Chapter 5 - Page 143

Defining a Common Language about Sampling

A parameter is a numerical descriptive measure of a population. A sample statistic is a

Statistics Pre-Term Program - Chapter 5 - Page 144

Month Mean SD Min 1st Q Median 3rd Q Max

Statistics Pre-Term Program - Chapter 5 - Page 145

1970 Draft Lottery

Statistics Pre-Term Program - Chapter 5 - Page 146

SELECTION BIAS A systematic tendency to over-represent some segment of the

Example 5.2. Literary Digest 1936

Difficulties and Disasters in Sampling (Utts, p. 62)

Statistics Pre-Term Program - Chapter 5 - Page 147

• 84% of women are not emotionally satisfied with their relationship

Statistics Pre-Term Program - Chapter 5 - Page 148

Example 5.4. A Meaningless Poll (Utts, pp. 64-65)

Television Poll Survey

Statistics Pre-Term Program - Chapter 5 - Page 149

Investigators Led Astray By Nonrandom Samples2

• The Most Dangerous Profession In 1835, H. C. Lombard published a study on the

Statistics Pre-Term Program - Chapter 5 - Page 150

As WPB indicate, “data obtained by nonrandom sampling occurs often in practice…. We

Statistics Pre-Term Program - Chapter 5 - Page 151

Probability distribution of a coin flip

Series1 0.5 0.5

Statistics Pre-Term Program - Chapter 5 - Page 152

Sample Sample mean

Statistics Pre-Term Program - Chapter 5 - Page 153

Distribution of sample mean

E(sample mean) = 1*0.0625 + 1.25*0.25 + 1.5*0.375 + 1.75*0.25 + 2*0.0625 = 1.5

Statistics Pre-Term Program - Chapter 5 - Page 154

Regardless of the shape of the population’s distribution:

Example 5.6. Telephone Company3

Solution of example 5.6:

Theorem One: Sampling from a Normally Distributed Population

Theorem: If a random sample of n observations is taken from a population with a Normal

Solution of example 5.6 (continues)

Statistics Pre-Term Program - Chapter 5 - Page 155

Probable Error for a Sample Mean

There is no systematic tendency to over- or underestimate µ with Y . This wouldn’t be much of

Example 5.7. (H&O, BSIFM, Example 6.3)

Example 5.8. (H&O, Exercise 6.16)

Statistics Pre-Term Program - Chapter 5 - Page 156

Statistics Pre-Term Program - Chapter 5 - Page 157

How Large is Large Enough?

Example 5.9. (H&O, Example 6.7)

Statistics Pre-Term Program - Chapter 5 - Page 158

Example 5.10. Cents and the CLT – A Lesson worth $8.00

VARIABLE N MEAN SD MINIMUM MAXIMUM DATA FILE

Statistics Pre-Term Program - Chapter 5 - Page 159

Statistics Pre-Term Program - Chapter 5 - Page 160

Statistics Pre-Term Program - Chapter 5 - Page 161

Simulation Demonstrations: CLT: How Large is Large Enough?

• We begin with histograms of 1000 means sampled from an exponential population.

Statistics Pre-Term Program - Chapter 5 - Page 162

1. Summary Results for Mean of Exponential Sample

Simulation of expected value and standard error (1,000 samples)

• Next, we use samples from a squared exponential (highly skewed) distribution to

2. Summary Results for Mean of Squared Exponential Sample

E(sample mean) = 10.0625 + 1.250.25 + 1.50.375 + 1.750.25 + 2*0.0625 = 1.5