Statistics 1: Problem Sheet Problems: page 1

Interpretations of Probability

1. A student is about to play a game against an opponent involving a standard 6-sided dice.
(a) She initially assigns the probabilities P (X = x) = 16 , with x = 1, . . . , 6 to each side of the
dice. What interpretation of probability might she be using?
(b) Her friend urgently tells her before the game starts that the opponent is a cheat, the dice is
loaded (i.e. unfair) and rolls a 6 more often than not. She thinks about this information and
then assigns new probabilities:
P (X = 6) = 0.6 and P (X = x) = 0.08, for x = 1, . . . 5 (1)
She is still prepared to play the game. What interpretation of probability might she now be
using?
(c) After playing the game for awhile, she diligently notes down the result of every roll of
the dice. After 1000 rolls she notes that the number 6 occurred 550 times. She discards
her previous assignment and now assigns P (X = 6) = 0.55 (and assigns the remaining
probabilities for x = 1, . . . , 5 from the observed counts using the same strategy). What
interpretation of probability might she be using here?
(d) Were her assignments of probability in parts (a), (b) and (c) correct? Discuss.
(e) In part (c), it seems reasonable that she discarded her previous assignment of P (X = 6) =
0.6 after 1000 rolls. Would it have been reasonable to discard this after 100 rolls, or 10 rolls
or 2 rolls? Discuss (without performing any detailed calculation).
Much of this course will aim to examine the kind of issues raised in question 1, to clarify the
many ambiguities, and to improve greatly upon the procedures used.
2. A six-sided die has four green and two red faces and is balanced so that each face is equally likely
to come up. The die will be rolled several times. You must choose one of the following three
sequences of colours; you will win £50 if the first rolls of the die give the sequence that you have
chosen.
Sequence 1: R G R R R
Sequence 2: R G R R R G
Sequence 3: G R R R R R
Without making any calculations, explain which sequence you choose. (In a psychological ex-
periment, 63% of 260 students who had not studied probability chose the second sequence. This
is evidence that our intuitive understanding of probability is not very accurate. This and other
similar experiments are reported by A. Tversky and D. Kahneman, Extensional versus intuitive
reasoning: The conjunction fallacy in probability judgment, Psychological Review 90 (1983),
pp. 293–315. The authors were psychologists: Kahneman was subsequently awarded the Nobel
Prize in economics (sadly Tversky died and hence missed out on the award). Kahneman’s famous
book “Thinking Fast and Slow”, available free online, describes many further cases.)
3. Recall De Finetti’s deliberately controversial statement:
“PROBABILITY DOES NOT EXIST”

(a) What did he mean by this statement: where did he suggest probability actually resides?
(b) What do you think of this statement: do you agree with him? (Any answer here is fine!)
Statistics 1: Problem Sheet Problems: page 2

Disease Testing Using Bayesian Methods

4. [2018 Exam Question] The following data classify 492 patients according to presence or absence
of a rare but serious disease as diagnosed by a “gold standard” and by the results of a cheaper but
less accurate diagnostic blood test.

Has disease?
Yes No Total
D+ D
Test result positive T+ 220 24 244
Test result negative T 30 218 248
Total 250 242 492

(a) Define and calculate the Sensitivity and Specificity of the test.
(b) Define and calculate the False Positive and False Negative rates. Which one is arguably
more of a problem in this example?
(c) A patient from the general UK population who is selected at random is tested and receives a
positive test result. In the UK it is known that about 1 in 1000 people have the disease. Cal-
culate the probability the patient has the disease given the positive test result, P (D+ |T + ).
Comment on your answer.
(d) The patient actually had the test done twice and received two positive results, represented by
the event T ++ . Assuming the test results are conditionally independent given disease sta-
tus, implying that P (T ++ |D+ ) = P (T + |D+ )2 and P (T ++ |D ) = P (T + |D )2 , calcu-
late the probability that they have the disease after receiving two positive tests, P (D+ |T ++ ).
(e) Derive a constraint on the number n of such positive test results that they would have to
receive in a row, in order to ensure the probability that they did indeed have the disease is
greater than p0 .
(f) Using the results of part (e) or otherwise, find the lowest possible n for p0 = 0.4.
Statistics 1: Problem Sheet Problems: page 3

5. [2016 Exam Question] A new blood test is developed for a rare, but often fatal disease. The test
is quicker and cheaper than existing methods (which are essentially perfect), but it is sometimes
inaccurate. Trials find that the average sensitivity of the test was 96% and the average specificity
was 93%.

(a) Explain in this context what is meant by sensitivity and specificity.

(b) Define and calculate the False Positive and False Negative rates, and explain why one is
worse in this context.
(c) A patient is randomly selected from the UK population and receives a positive test result.
He reads that the average incidence of this disease in the UK population is approximately
1 in 2000. Calculate the probability that he has the disease given the positive test result.
Comment on the result.
(d) There are plans for widespread screening of the UK population using this test (the plans
involve testing every adult in the UK). Comment on whether you think this is a good idea.
(e) The patient then reads that the incidence of the disease for men aged between 65-70, of
which he is a member, is suspected to be much higher, but is currently unknown. How high
would his prior probability P (D+ ) have to be before there was a 90% posterior probability
that he had the disease?
(f) Comment briefly on whether the prior found in part (e) is reasonable.
Statistics 1: Problem Sheet Problems: page 4

6. Genetic haemophilia is a genetic disorder caused by a defect in the activity of coagulation factor
VIII, also called AHG. The disease is sex-linked, so that the disease is overt amongst males
whose X chromosome has the abnormal gene (and rarely, and in more complicated fashion,
amongst females with abnormalities in both X chromosomes). Female patients with one affected
X chromosome are called carriers and do not have overt haemophilia, but can pass on the genetic
defect to their children with probability 0.5, independently for each child.
Around 1970, a test for presence of the genetic abnormality amongst females was established.
The test is based on the level of factor VIII activity. Applying the test to females known to have
the abnormality suggests that the test is correct in this situation with probability 0.94. Applying
the test to females known not to have the abnormality suggests that the test is correct in this
situation with probability 0.82.

(a) What is the sensitivity of the test?

(b) What is the specificity of the test?
(c) What is the probability of a false positive?
(d) What is the probability of a false negative?
(e) Patient X does not have haemophilia, but one of her brothers and a maternal uncle do.
However, another brother and her father do not have the disease.
i. What is the prior probability that patient X is a carrier?
ii. Patient X takes the test based on factor VIII activity, and this test is positive. What is
the posterior probability that patient X is a carrier, given the positive test result?
iii. Determine the prior and posterior odds in favour of patient X being a carrier.
(f) Patient Y was an adopted child and so has no known family history of haemophilia. How-
ever, she tests positive for the genetic defect. The background prevalence of carriers in the
population is about 3 women in every 10000.
i. What is the prior probability that patient Y is a carrier?
ii. What is the posterior probability that patient Y is a carrier, given the positive test result?
iii. Determine the prior and posterior odds in favour of patient Y being a carrier.
Statistics 1: Problem Sheet Problems: page 5

US Presidential Election Polling via Frequentist Methods: Proportions and Estimators

7. UK Local elections were held on May 2nd 2019. The previous day, an opinion poll was pub-
lished which was based on a sample of 1200 voters, of whom 348 indicated that they would vote
Conservative. (We will assume that the sample was random; in fact it wasn’t.) Imagine that it is
still May 1st 2019, and suppose that p is the proportion of voters who will vote Conservative on
May 2nd.

(a) Based on the results from this particular opinion poll, what would be your estimate for p?
(b) Let X be the random variable representing the total number of Conservative voters in a
sample of size n drawn from the UK voting population. Write down the distribution of
X, an approximate distribution for X, and an approximate distribution for the estimated
proportion, p̂ = X
n.
(c) In the worst case, if we require a margin of error of ±0.02 (i.e. 2%) in our opinion poll,
how large should we choose n?

In fact, the result of the election was that the Conservatives gained about 28% of the vote, so we
now know that p = 0.28 approximately. (Note: Labour also received about 28% of the vote.
Local elections for 2020 were postponed due to the Covid-19 pandemic.)

8. (a) We saw that if X is the sum of the successes in n Bernoulli trials with parameter p, then it
has Binomial distribution X ⇠ Bin(n, p), with p.m.f. given by
⇢ n x n x
f (x|n, p) = x p q for x = 0, 1, 2, . . . , n,
0 otherwise.

with q = 1 p. Assuming that x and n are fixed and can hence be treated as constants
(which incidentally would be the case if we had actually performed the trials and hence
know n and x), find the value of p that maximises f (x|n, p). Hint: you may want to
examine log f (x|n, p) as it is easier to differentiate.
(b) The value you found in part (a) can be considered an estimate for p and is usually denoted
p̂. Why do we use this notation: why not just call this value p?
(c) The value of p̂ you found in part (a) should agree with the expression for p̂ we found in
lectures, when we examined the p.m.f. of n Bernoulli trials f (x1 , x2 , . . . , xn |p) directly.
Why must these results be the same? Hint: as part of your answer, you may wish to discuss
the similarities and differences in the form of f (x1 , x2 , . . . , xn |p) compared to f (x|n, p).
Statistics 1: Problem Sheet Problems: page 6

9. [Exam standard] A US study in 1974-75 surveyed the attitudes of people to the role of women in
society. In particular, individuals were asked whether or not she or he agreed with the statement

Women should take care of running their homes and leave running the country up to
the men.

In all, 1566 women were surveyed. The number, X, of women in agreement with the proposition
were counted.

(a) Under what conditions is this a ‘binomial scenario’ (i.e. can it be thought of as the counts
from a series of Bernoulli trials)?
(b) Suppose that the proportion of US women in agreement with this proposition is 0.33. Find
the probability that we see between 500 and 560 (inclusive) women in favour of this propo-
sition.
(c) Do you believe that a correction for continuity is useful for this example? If not, when
might a continuity correction be more useful?
(d) The actual number of women in agreement for this sample was 555. What is the probability
of seeing a result as extreme (i.e. as far from the expected value in either direction, assuming
p = 0.33) as this?

10. Suppose that a random variable X has a binomial distribution with parameters n and p, and
consider the proportion p̂ = X
n . Prove that Var [p̂] is maximum at p = 0.5.

11. When p is small, the Normal approximation to the binomial distribution is not so useful. In-
stead, the Poisson distribution can be used. A random variable X has a Poisson distribution with
parameter when its probability distribution is
xe
f (x| ) = P (X = x) = , > 0, x = 0, 1, 2, . . .
x!
Here, X is a discrete random variable which can take the non-negative integers for its values.
The Poisson distribution is typically used to model the number of times a particular event occurs
over a specified period of time. The Poisson distribution is unusual in that it has the same mean
and variance, namely . The link between the Poisson distribution and the binomial distribution
is that, for large n and small p, X is approximately distributed Poisson with parameter = np.
A guide as to when the approximation is useful is that we need approximately n > 20 and
np = < 7.
People with certain kinds of stomach illness may be referred to a specialist who will carry out
a gastroscopy to identify whether or not the cause of the illness is gastric cancer. Amongst
individuals referred to the specialist, the diagnosis of cancer is made in 0.7% of cases. The
specialist sees about 150 patients per month. Find the probability that, in a given month, the
specialist will diagnose more than two cases of cancer.
Statistics 1: Problem Sheet Problems: page 7

The Frequentist Paradigm: Inference, the Sample Mean and the t-distribution

12. Under the standard Frequentist paradigm and before the sample data are seen, which of the
following quantities are random variables?

(a) The population mean.

(b) The population size.
(c) The sample size, n.
(d) The sample mean.
(e) The variance of the sample mean.
(f) The largest value in the sample.
(g) The population variance.
(h) The estimated variance of the sample mean.

Do the answers you have given seem natural to you? Discuss.

13. This question simply tests your ability to use t-tables.

(a) A random variable T has a t16 distribution (i.e. a t distribution with 16 degrees of freedom).
Find the number c such that P (T > c) = 0.025. What would c be if instead T had a
standard Normal distribution?
(b) A random variable T has a t7 distribution. Find the number c such that P (T  c) = 0.05.
What would c be if instead T had a standard Normal distribution?
(c) A random variable T has a t35 distribution. Find approximately the number c such that
P ( c < T  c) = 0.95. What would c be if instead T had a standard Normal distribution?
(d) A random variable T has a t12 distribution. Roughly, how likely is it to see a value of T
at least as large as 1.5? What would your answer be if instead T had a standard Normal
distribution?
(e) A random variable T has a t12 distribution. Roughly, how likely is it to see a value of T at
least as extreme as 1.5 (i.e. no larger than -1.5 or no smaller than +1.5)? What would your
answer be if instead T had a standard Normal distribution?
Statistics 1: Problem Sheet Problems: page 8

14. A dairy claims that the mean amount in their milk containers is 128 ounces. Let X be the
number of ounces of milk per container, and assume that X is Normally distributed with standard
deviation one ounce.

(a) If the claim is true, what percentage of containers will have

i. Less than 126 ounces?
ii. More than 129 ounces?
iii. Between 127.5 and 130.5 ounces?
(b) A random sample of 25 containers gave a sample mean of 127.4 ounces. Find P (X̄ < 127.4),
assuming again that the claim is true.
(c) Using the answer to the previous part, do you feel that there is evidence that the true mean
is less than 128 ounces? If so, why?

15. [Exam standard] An experimenter wishes to make an inference about the reaction time in sec-
onds for a psychological test which involves a person detecting a hidden image. The experi-
menter’s budget allows a sample size of n = 13.
Let the random variable X̄ represent the sample mean for n = 13, and suppose that the measure-
ments will be drawn independently from a distribution with mean µ and variance 2 .

(a) What, approximately, is the distribution of X̄, and why?

(b) When is the distribution of X̄ exactly Normal?
(c) Before taking any measurements, calculate a value c, using appropriate tables, such that
P ( c  X̄ µ  c) = 0.95 when it is known that 2 = 9.
(d) The experimenter now observes some data. The following data are reaction times in seconds
for 13 individuals:
21, 23, 21, 21, 17, 18, 23, 23, 25, 21, 15, 17, 13
i. The data are assessed using a Normal quantile plot and it is concluded that the data
seem to be reasonably Normal (see the solutions for the quantile plot and further dis-
cussion). Is the use of a t distribution justified here?
ii. Calculate a value c, using tables, such that P ( c  X̄ µ  c) = 0.95 when 2 is
not known, and must be estimated by s2 using the data supplied.
iii. Repeat the previous part, but find d such that P ( d  X̄ µ  d) = 0.50 when 2 is
estimated by s2 .

[Hint: In answering some of these questions, you will need to think about the distribution of the
standardised version of X̄ µ.]
Statistics 1: Problem Sheet Problems: page 9

The Frequentist Paradigm: Confidence Intervals

16. A candle making process is known to produce candles whose life times have a variance of 0.4
hours. However, the average lifetime is unknown. The lifetimes of a random sample of six
candles turned out to be 8.1, 8.7,9.2,7.8,8.4,9.4 hours respectively.

(a) Find a 98% confidence interval for the population mean, if you were told the population is
normally distributed.
(b) Comment on the accuracy of your answer for part (a) if you were not told the distribution
of the population.

17. Obtain a 99% confidence interval for the population mean for the logs of the F 3+ data:

0.79 1.08 1.12 1.25 1.41 1.60 1.64 1.71 1.74

1.83 1.98 2.07 2.18 2.48 2.74 2.77 2.91 2.92

You may assume that sample statistics for these data are x̄ = 1.90 and s = 0.66. State what
assumptions you make, and how you would go about checking their validity.

18. Find a 75% confidence interval for the population mean for the cuckoo egg length data:

22.5 20.1 23.3 22.9 23.1 22.0 22.3 23.6 24.7 23.7
24.0 20.4 21.3 22.0 24.2 21.7 21.0 20.1 21.9 21.9
21.7 22.6 20.9 21.6 22.2 22.5 22.2 24.3 22.3 22.6
20.1 22.0 22.8 22.0 22.4 22.3 20.6 22.1 21.9 23.0
22.0 22.0 22.1 22.0 19.6 22.8 22.0 23.4 23.8 23.3
22.5 22.3 21.9 22.0 21.7 23.3 22.2 22.3 22.8 22.9
23.7 22.0 21.9 22.2 24.4 22.7 23.3 24.0 23.6 22.1
21.8 21.1 23.4 23.8 23.3 24.0 23.5 23.2 24.0 22.4
23.9 22.0 23.9 20.9 23.8 25.0 24.0 21.7 23.8 22.8
23.1 23.1 23.5 23.0 23.0 21.8 23.0 23.3 22.4 22.4

You may assume that sample statistics for these data are x̄ = 22.57 and s = 1.07, with n = 100.

19. Suppose that last year’s volume of a certain internationally renowned medical journal contains
reported results from 200 separate clinical trials. Suppose that in each trial a 95% confidence
interval for an unknown parameter (such as a population mean µ for the effect of a drug) was
reported. About how many of these confidence intervals would not, in fact, contain the parameter
of interest? If X is the random variable representing the total number of confidence intervals
which do not contain the parameter of interest, write down the probability distribution for X,
together with any necessary assumptions. Write down E [X] and Var [X]. Comment on your
results.
Statistics 1: Problem Sheet Problems: page 10

20. The following data are measurements of fasting blood glucose levels in 17 Type I diabetics:

4.2 4.9 5.2 5.3 6.7 7.5 8.1 9.5 10.3

12.2 12.5 13.3 15.1 15.3 16.1 19.0 19.5

The measurements have sum 184.7 and sum of squares 2404.41. Use the t distribution to calcu-
late a 95% confidence interval for the mean of the population from which the measurements are
taken. State any assumptions necessary in doing so.

21. Draw a boxplot and, by its side, a Normal quantile plot of the measurements for the previous
question. Use your plots to assess the distribution of the measurements.

22. A random sample of 100 packs of apples with a professed weight of 1kg were found to have
mean 1020g and standard deviation 12g.

(a) Find a 95% confidence interval for the mean weight of all such packs of apples.
(b) Does the interval include 1kg? Can you think of a reason (which any retailer would be able
to tell you) why you might expect the interval not to include 1kg?
(c) What confidence level would have resulted in the interval just about including 1kg?
Statistics 1: Problem Sheet Problems: page 11

23. [Exam standard] [Hint: to answer this question, you will need to remember that P (aX + b > c) =
P (X > (c b)/a) for any constants a, b, c 6= 0.] One measurement of the wealth of a commu-
nity is mean disposable income. Crudely, suppose that a person’s disposable income is defined
to be gross income, less 25% income tax, then minus a further £3000 to cover some essentials.
Suppose now that a random sample of 400 people is taken and their gross incomes measured,
with the intention of estimating µ, the average disposable income in the overall population. You
may assume that previous studies have shown that the population standard deviation for gross
income is about £8000.

(a) Stating any assumptions that you make, find the probability that the gross income of one
person selected at random exceeds £30000 if truly (1) µ = £18000, (2) µ = £20000.
(b) Stating any assumptions that you make, find the probability that the disposable income of
one person selected at random exceeds £1000 if truly (1) µ = £18000, (2) µ = £20000.
(c) Suppose that the average gross income of 400 persons selected at random is £20000. Find
a 95% confidence interval for µ.
(d) Suppose that we repeated the entire process 1000 times, calculating a confidence interval
each time (notice that this means taking 1000 random samples of size 400 in 1000 separate
experiments). Roughly how many of the confidence intervals for µ that we calculate will
not, in fact, contain µ?
(e) For the remaining parts of this question, answer without making further calculations.
i. If = £8000 understates the actual value of the population standard deviation, how
reliable is the confidence interval calculated above?
ii. What happens to the confidence interval calculated above if instead we had taken a
sample of size of n = 100?
iii. What happens to the confidence interval calculated above if instead we had taken a
sample of size of n = 1600?
iv. Consider the shape of the distribution of income in a population such as in the UK. In
this context, discuss briefly the effect of distribution shape on the main measures of
central location and spread.
Statistics 1: Problem Sheet Problems: page 12

The Frequentist Paradigm: Hypothesis Testing

24. [Exam standard]

p Consider the candle data given in lectures. Summary statistics are x̄ = 8.6,
n = 6, and = 0.4.

(a) Test the hypothesis that the mean candle lifetime is µ = 8.0 hours at the 5% and 1% levels
of significance using the Confidence Interval approach.
(b) What is the corresponding p-value for the hypothesis test?
(c) Suppose that this experiment is repeated 1000 times and we count the number of experi-
ments, X, where we find statistically significant evidence against the null hypothesis at the
1% significance level. Assuming H0 was indeed true in each of the 1000 cases, what is
the probability distribution for the random variable X? Calculate E [X] and Var [X]. What
does this say about the number of mistaken conclusions reached by the 1000 experimenters?
[Hint: think about counting the total number of “successes” obtained in a given number of
independent trials. What is the scenario here?]

25. Consider the logs of the F 3+ data given in lectures. Summary statistics are x̄ = 1.90, n = 18,
and s = 0.66.

(a) Test the hypothesis that the underlying population mean µ = 2.0 at the 10% significance
level using the Confidence Interval approach.
(b) In this case, why is it difficult to obtain the corresponding p-value for the hypothesis test,
unless you have access to a computer?

26. Consider the cuckoo egg length data given in lectures. Summary statistics are x̄ = 22.57, n =
100, and s = 1.07.

(a) Carry out a test of the hypothesis that the underlying population mean for cuckoo egg length
is µ = 22.4 versus the hypothesis that µ 6= 22.4. Test the hypothesis at the 20% and 10%
levels of significance using the Confidence Interval approach.
(b) What is the corresponding p-value for the hypothesis test?

27. The level of calcium in the blood in healthy young European adults varies with mean about
9.5 mg/dl and standard deviation about 0.4. A clinic in rural Guatemala measures the blood
calcium level of 180 healthy pregnant women at their first visit for prenatal care. The mean is
x̄ = 9.57. We want to know if this suggests that healthy pregnant Guatemalans have a mean
calcium level which differs from 9.5.

(a) State H0 and Ha .

(b) Carry out a 1% significance test of the hypothesis. Report whether or not the null hypothesis
is rejected at this level.
(c) What does your answer say about mean calcium levels in healthy Europeans and pregnant
Guatemalans?
Statistics 1: Problem Sheet Problems: page 13

28. Poisoning by the pesticide DDT causes tremors and convulsions. In a study of DDT poisoning
(D.L. Shankland, Involvement of spinal cord and peripheral nerves in DDT-poisoning syndrome
in albino rats, Toxicology and Applied Pharmacology, 6 (1964), pp.97–213), researchers fed
several rats a measured amount of DDT. They then measured electrical characteristics of the
rats’ nervous systems that might explain how DDT poisoning causes tremors. One important
variable was the “absolutely refractory period”, the time required for a nerve to recover after a
stimulus; variation of the absolutely refractory period is known to follow a normal distribution.
Measurements on four rats gave the data below (in milliseconds).

1.6 1.7 1.9 1.8

(a) Find the average “absolutely refractory period” for the four rats and the standard deviation.
(b) Find a 90% confidence interval for the mean “absolutely refractory period” in all rats of this
strain when subjected to the same treatment. One researcher believes that this mean should
be 1.95: does the evidence support her view?

29. Historically, the mean yield of corn in the United States is about 120 bushels per acre. A survey
of 50 farmers this year gives a sample mean yield of x̄ = 123.6 bushels per acre. We want to
know if this is strong evidence that the national mean yield this year is not 120 bushels per acre.
Assume that the farmers surveyed were a random samples from the population of all commercial
corn growers and that the standard deviation of yield in this population is = 10 bushels per
acre.

(a) Find the p-value for the test of the hypotheses H0 : µ = 120 and Ha : µ 6= 120.
(b) Does that suggest that the population mean is not 120 bushels per acre?
(c) Does your answer depend on any particular shape for the population distribution?
Statistics 1: Problem Sheet Problems: page 14

30. [Exam standard] True, or false, or depends on context? You must explain your answer to receive
marks.

(a) We did an experiment where we measured many variables on a random sample of indi-
viduals. A careful exploration of the data indicated many interesting, possibly significant,
findings. We then carried out various hypothesis tests on the data to check these findings.
The conclusions that we draw from these tests are valid.
(b) We’ve measured the income of a large random sample of individuals. We’ve drawn a Nor-
mal quantile plot to assess the Normality of the data, and we’ve discovered that the distri-
bution is very skewed to the right. This means that we can’t use the Normal distribution in
making tests about the population mean µ using the sample mean.
(c) We’ve measured the income of a small random sample of individuals. We’ve drawn a
Normal quantile plot to assess the Normality of the data, and we’ve discovered that the dis-
tribution is very skewed to the right. We don’t know the population variance . Ordinarily
we could estimate by the sample standard deviation s and use the t distribution to make
tests about the population mean µ using the sample mean, but the non-Normality of the
distribution forbids the use of the t distribution in this case.
(d) We’ve calculated a 95% confidence interval for a population mean, µ. The interval is [12.1,
13.4]. This means that P (12.1  µ  13.4) = 0.95.
(e) We’ve carried out an experiment independently 200 times. In each experiment we tested
the same null hypothesis at the 1% level of significance. If the null hypothesis is actually
true, we should find ourselves mistakenly rejecting the null hypothesis in about two of the
experiments.
(f) We’ve carried out an experiment and tested a null hypothesis at the 1% level of significance.
The probability of a type II error is 0.01.
(g) For a two-sided hypothesis test at the 10% level of significance, we could construct a 90%
confidence interval, and reject the null hypothesis if the interval does not contain the pa-
rameter.
(h) For a two-sided hypothesis test at the 10% level of significance, we could construct a 95%
confidence interval, and fail to reject the null hypothesis if the interval does not contain the
parameter.
(i) We’ve independently repeated an experiment 200 times, and tested the same null hypothesis
at the 5% level of significance. If the null hypothesis is true, the total number (out of 200)
of type I errors is a random variable with a binomial distribution with parameters n = 200
and p = 0.05. Thus we expect about 10 type I errors, with a standard deviation of about
3.08.

31. You are engaged in designing a clinical trial for testing for side effects of a new drug aimed at
ameliorating nausea in pregnant women; for example the drug might or might not be harmful to
the unborn child.

(a) What are the two hypotheses in this case?

(b) What kinds of error might you make?
(c) Which kind of error would you try to minimise, and why? Explain how changing these
circumstances slightly could lead you to changing your answer.
Statistics 1: Problem Sheet Problems: page 15

32. The p-value, as used in several of the previous questions to decide whether to reject the null
hypothesis H0 , is defined in these cases as:

p-value = P (sample mean is at least as extreme as the observed sample mean | H0 is true)

An alternative probability was suggested in lectures for consideration:

P (H0 is true | observed sample mean)

(a) Discuss which of these two probabilities seems best suited for making decisions about
whether H0 is true.
(b) Use Bayes theorem to rewrite P (H0 is true | observed sample mean) in a more useful form,
involving “inverted” terms such as P (observed sample mean | H0 is true). Note that you
can just use probabilities here, and ignore the subtleties due to continuous random variables
for now.
(c) Examine the numerator of your answer to part (b). What additional term, that we have not
utilised in the entire Frequentist approach up to this point, does Bayes theorem suggest we
must inevitably consider if we do indeed want to find P (H0 is true | observed sample mean)?
Is this surprising or could it be considered natural?
(d) Say there is an alternative hypothesis Ha and that we are sure that either H0 or Ha is true.
Can you use the partition theorem to rewrite the denominator of your answer to part (b)
using terms that feature either H0 or Ha ?
(e) We see that the denominator now depends on the form of the alternative hypothesis, and
that therefore P (H0 is true | observed sample mean) depends directly on Ha . Comment on
whether this is a desirable feature or not.
(f) Write down a similar form for P (Ha is true | observed sample mean). At least up to the
level of detail we have used here, is there any major difference between how H0 and Ha
are treated in this second (Bayesian) framework? Again, does this seem natural/desirable?
Statistics 1: Problem Sheet Problems: page 16

The Frequentist Paradigm: Hypothesis Testing for Proportions

33. A poll of 1785 US adults asked whether people had attended a place of worship during the last
week. 750 respondents said that they had. Pretend that the 1785 people asked form a simple
random sample of US adults, and so can be taken to be representative of US adults. (Note: most
opinion polls are not simple random samples, and this was no exception.)

(a) Give a standard 99% confidence interval for p, the proportion of US adults attending a place
of worship during the previous week.
(b) Is it fair to conclude that about half of all US adults attended a place of worship during the
previous week?
(c) How large a sample would be required to obtain a margin of error of ±0.01 in a 99%
confidence interval for p, (i) at worst, when p is unknown; (ii) using the observed proportion
as a guess for p?

34. Studies are performed to estimate the percentage of the nation’s asthmatics who are allergic to
sulphites. Suppose that 500 asthmatics are randomly selected and that 38 of the 500 are found to
be allergic to sulphites.

(a) Estimate p, the proportion of all asthmatics who are allergic to sulphites.
(b) Find a standard 90% confidence interval for the proportion of all asthmatics who are allergic
to sulphites.
(c) Test approximately the hypothesis that p = 0.08.
(d) In the worst possible case, express the variance of the sample proportion as a function of
the sample size n. In this worst case, for what value of n do we obtain the 95% confidence
interval for p as y ± 0.01?

[Hint: for what value of p is the theoretical variance of Y largest? Notice that this part of
the question is about the theoretical sampling distribution of Y , and not about the sample
we actually obtained.]

35. A supplier of components to the motor industry makes a particular product which sometimes fails
immediately it is used. The manufacturing process is controlled so the the proportion of faulty
products is supposed to be only 4%. Out of 500 supplied in one batch, 28 prove to be faulty. Has
the process gone out of control to produce too many faulty products?