Introduction to Statistical Hypothesis Testing

Prof. Arun K Tangirala

Department of Chemical Engineering
Indian Institute of Technology, Madras

Lecture - 08
Statistics for Hypothesis Testing - Part 2

Welcome to this lecture. This is the second and final lecture on sampling distributions, I should say.
Well, there is a part of the sampling distribution with respect to sample correlation that will discuss
later, but more or less we would be covering all the sampling distributions that we require for
hypothesis testing through this lecture. And again, in this lecture, what we are going to look at is a
continuation of what we discussed previously.

(Refer Slide Time: 00:47)

We were going to look at sampling distributions of difference in sample means, sample variance, ratios
of sample variances, sample proportions and also difference in sample proportion. Now, all of these are
kind of in line with examples that we discussed in the motivation lecture. And let me also tell you up
front that in this lecture mostly it is a statement of results, neither we will try to prove anything nor use
many standard results that are available and so may be for a few will use some standard results, but will
not attempt to prove anything. And, anyway in 10 hour course there is not much time to prove
anything, and we are not interested in that we want quickly get to hypothesis testing. And of course,
just to reiterate it is important to study the sampling distributions for hypothesis testing because after all
the entire test of hypothesis rests on the knowledge of the sampling distribution of the test statistics that
we use. In the previous lecture, we looked at sampling distribution of mean under different conditions
known variance unknown variance and so on.

(Refer Slide Time: 02:04)

Now, we look at difference as in sample means again Ii would advise you to go back and relate to the
examples that we discussed in the motivation lecture. At least one example involved comparison of
means, right, where we were looking at the nylon connector example. So, of course, this situation arises
in many, many other applications as well. So, let us come to point here when I have 2 populations of
different means, let us say mu 1 and mu 2, I would like to know for example, if indeed they are
different or not, and for this purposes we setup a test statistic. And this test statistic has to be
commensurate with what I want to test, what I want to test is whether mu 1 which is the mean of
population 1 and mu 2 which is that of population two are different from each other or not. So naturally
the test statistic involves differences in the estimates of means and once again as I said in the previous
lecture, we are going to use only the sample mean as an estimator of the mean; however, that should
not preclude you from or prevent you from using other estimators.
So, coming back to the sample mean as an estimator. If X 1 bar and X 2 bar are sample means
computed from samples of the respective populations of sizes n 1 and n 2 and known variances. So, we
assume in the beginning a simpler situation, where the variances of the populations are known, namely
sigma 1 square and sigma 2 square then this test statistic X 2 bar minus X 2 bar minus mu 1 minus mu
2 divided by square root of sigma 1 square by n 1 plus sigma 2 square by n 2 is approximately standard
normal, if the conditions of the central limit theorem apply. What are the conditions of central limit
theorem, well the observations have to follow out of an identical and independent distribution, right.

Now, as usual we usually talk of the distribution of the standardized one and z is also the standardized
one; recall in the previous lecture why we talk of this standardized test statistic. One advantage is we
can use a standard distribution. Now, what if the variances are not known or the populations are normal
and so on, this results that you seen equation one is of it for a general population that is X 1 and X 2
that is the population one and population two can be characterized by any distribution in this results.
However, if the populations are characterized by Gaussian distributions then Z follows an exact
standard normal distribution.

(Refer Slide Time: 05:12)

Suppose, I do not know the variances which is the more realistic case then what do I do, well I adopt
the same approach that I adopt in the case of sample mean that is testing of single sample mean. We
replace the theoretical variances with the respective estimates denoted by S 1 square and S 2 square;
once again we have not talked about how to estimate sample, how to estimate variances yet we will do
so shortly. So, the result is again more or less the same; again this is for the large sample case, you
should remember. The test statistic that is given in equation two is approximately standard normal. The
only difference between the test statistic in 1 and 2 is that, we have replaced the theoretical variances
with the respective estimates. And, what will do next is look at the small sample case, but under two
different situations; variances are equal and variances are unequal. This is possible that 2 populations
can have same variances.

(Refer Slide Time: 06:19)

How does it make a difference? Let us look at this. So, now as you must have guessed like in the
sample mean case when the variance was unknown, and when the samples were small, the test statistic
followed T distribution there. Here also the test statistic based on the differences in sample means of
course, the standardized test statistic follows T distribution with n 1 plus n 2 minus 2 degrees of
freedom. So, you can say well, we lose one degree of freedom when we estimate sample mean form
population one and another degree of freedom when we are estimating sample mean for another
population. So, in total, we have two degrees of freedom loss, therefore the total degrees of freedom
that we have when it comes to estimating the variance is n 1 plus n 2 minus 2.
We have assumed the variance to be equal or variances to be equal, and therefore, we construct what is
known as a pooled variance. What happens is when the variances is equal, we are exploiting this
situation to our advantage; from population one, we have sample of size n 1; from population 2, we
have a sample of size n 2. More the number of observations to estimate a variance (Refer Time: 07:42)
situation. Therefore, since we are already given that the variances are equal; we pooled the data
together to estimate the common variance and that is the expression given to you in equation 4 for the
sample variance. Where you can see in the denominator we have n 1 plus n 2 minus 2 that is nothing
but your degrees of freedom, alright.

So, this n 1 minus 1 and n 2 minus 1 we already know. What about S 1 square and S 2 square? S 1
square is a sample variance estimated from the first population; S 2 square is a sample variance
calculated from the second population. And of course, again I say here, we have not yet seen the
expression for calculating S 1 square and S 2 square will do so very soon. Just to tell you that we are
going at probably slightly rapid base because as I said early on in the lecture, we are looking at
statement of results more than trying to prove anything. And, at some point this may get boring to you,
but it is an inevitable devil you can say that we have to deal with. So, what about the small sample size,
unequal variances case; again the same story, because it is a small sample size the test statistic would
follows T distribution.

(Refer Slide Time: 09:20)

And in this case, here again we are looking at a non-Gaussian case and that is why it is an
approximately T distributed. The difference between this and a previous cases the variances are
unequal, therefore, I cannot pool the variance estimates to come with the common variance. And, the
calculation is a bit more involve as you can see particularly in the calculation of degrees of freedom
denoted by n 1 2 minus 2, the n 1 2 calculation is quite complicated compare to the previous case. But,
we have to leave it there were no worries I mean most of the times the computer calculates set for you
nevertheless if you have to do it by hand then you have to use this expression no other choice. But in
the end, what is important is to know also the type of distribution it is a T distribution that does not
change fortunately. And, as usual we have X 1 bar minus X 2 bar as the estimate of the difference in
means.

So, in all of this something should evolve for you that what we are doing is we first identify the
parameter of interest in this case the difference in means then identify the test statistic before between
identifying the parameter and identifying test statistic there is an intermediate step which is choosing
the statistic or the estimator itself. We can say the estimator here is X 1 bar minus X 2 bar for mu 1
minus mu 2, and the test statistic is what you see in each of this, so that is the routine that we will see in
the remaining examples, therefore, you should not have any difficulty in terms of understanding the
procedure. If any difficulty that you may have that would be in remembering the distributions, but that
is all right in most of the hypothesis testing exercises. Typically, we do not have the necessarily
remember the sampling distribution let us say in an exam I have to remember I do not have to I can be
actually given that information on a simple sheet and perhaps we will do that for you in the exam, it is
not a difficult thing to do.

What is important is to understand the concepts of hypothesis testing that there is a sampling
distribution of the statistic that you dealing with and that goodness of the entire hypothesis test realize
on the sampling distribution apart from a few other factors such as sample sizes and variances and so
on. So, with this we kind of close the means case, but there is one more example, I am sorry, there is
one more example that is going to come up and with that will close the means and that has got to do
with the paired difference of means.
(Refer Slide Time: 11:58)

In the previous 2 or 3 situations, we had looked at differences in means from 2 different populations;
assuming that these 2 populations are kind of independent of each other. But now, there are we are
looking at a situation where that is not the case; where the 2 populations are not independent, and when
thus the situation arise. A standard example that is given is a weight-loss program where I am looking
at a set of individuals that is I want to test a weight-loss program and I want to test that it has been
effective. How would I test that? Well, I look at the average weight before implementing the weight-
loss program and then the average weight after the implementing the weight loss program. So, what I
do is I randomly sample a set of individuals, before the weight loss-program, take their weights and get
an average estimate of the average weight, and then put them through the weight-loss program and then
again do the same calculation, take the weights and compute the average weight.

So, now, obviously, these are the same individuals that are going to be screened for testing right. And
there is a high level of dependency and there better be a dependency otherwise, what is the meaning of
a weight-loss program. When we conduct test like this, in these kinds of situations, it could be a
weight-loss program, it could be a training program to see if there is an improvement in the
performance of teachers or students and so on, there also you will end up with this pair difference of
means case; where these obviously, the sample sizes are identical for both population, because we are
looking at the same set of individuals or same set of specimens you may say. So, there is a one-on-one
correspondence, and you are looking at a pair you are looking at a pair difference therefore. In this
case, the independence assumption is violated, and therefore it warrants a separate analysis.

So, what is the result here the first thing is you a define this variable d I as X 1 i minus X 2 i that is
think of X 1 i as a weight of the i th individual before the program and X 2 i as a weight after the
program. So, we are not actually testing, it basically amounts to testing the zero mean you can say for d
or d i, we are not individually testing for X 1 and X 2 you can think of it that way. So, now, having
defined a variable d i as X 1 i minus X 2 i, we can now more or less think of this as a sing single
sample mean calculation or mean hypothesis testing case. But now in terms of d, so as we did in the
single sample case; that means, single population case where we were testing for means here we
compute the average mean of the differences, so that is your D bar and then we calculate the variance
in a way that we would calculate for a single population.

So, what we have done is essentially taken 2 populations, but because they are dependent on each other
they; they have a strong correlation between them. We have clubbed that into single population now
described in terms of this difference variable D i that is all. Now, the result that is given here is for the
small sample case; there is no need to specify the large sample case because we know the T distribution
tends to Gaussian distribution. So, we do not need to state that separately. The rest of the story is the
same. Now, the test statistic D minus mu 1 minus mu 2 by your S over root n follows T distribution
with n minus 1 degrees of freedom where n is the sample size of the individual population. Here as we
said earlier, both populations have to be of identical sorry, both samples have to be of identical size and
that is your n, very good. Please keep these distributions handy of course, I will also bring up the
respective distributions, when we take up the eight examples that we had discussed in the motivation
lecture for hypothesis testing - illustrating the hypothesis test, alright.
(Refer Slide Time: 16:42)

Now, let us move on to the estimation of variance. In all of this, we have seen wherever the variance
was unknown I have to estimate it form the sample. Once again, what is the best estimator for
variance? There are many different variance of estimating variance, I could look at sample variance as
you see on the screen right now or I could use range for example, as an estimator of variance and so on.
Again given many estimators which estimated do I pick is always the question that has been of interest
for several decades in estimation theory and continues to be so. For some situations, there are results
that are established, but in general, there are set of criteria as we discussed in the previous lecture, for
choosing an estimator. One criteria is that we want an accurate or an unbiased estimator and two that
we want an efficient estimator; an estimator which gives me as low error as possible, and thirdly
consistent estimator which is that as the sample size increases the estimate converges to the truth. So,
these are some of the 3 very important criteria in choosing an estimator.

And a sample variance satisfies all of this it is an accurate estimator which means it is unbiased the
expectation of S square is nothing but sigma square of course, under some conditions. And then it is an
efficient estimator, it is not the most efficient estimator, but it is fairly efficient; it would not be give me
estimates with large errors. And thirdly, it is a consistent estimator as n grows large, the variance
estimates will convert to the truth very good. So, with these points in mind, we will stick to the sample
variance as an estimator of the variance and will work with those with this estimator only.
Now of course, the question of interest is what is a distribution of this sample variance; can we derive
this theoretical? Yes, we can, if you recall one of the standard results that we discussed yesterday, there
was one result where we said when I take random variables - standardized random variables that fall
out of a Gaussian distribution and square them and add them up, then the resulting random variable
follows a chi square distribution. If I am adding up n such variables then the resulting random variable
has a chi square distribution with n degrees of freedom. How does that result apply here? Well, look at
the expression for the sample variance. What am I doing there I am actually adding up squared
variables of course, there are deviations; are these deviation variables normally distributed? Well, yes,
if you assume that the distributions are following out of a Gaussian distribution. If they are not, what
about it will discuss that shortly, but assume that the samples are following out of a Gaussian
distribution.

Then we know X bar also follows a Gaussian distribution, difference of two Gaussian distributed
variables also follows a Gaussian distribution. Therefore, I can apply that result here and come to the
conclusion that S square follows a chi square distribution with here with n minus 1 degrees of freedom
that is the only point that we have to watch out for why n minus 1 degrees of freedom, why not n
degrees of freedom. Because, the result that we had seen earlier says that the resulting random variable
follows a chi square distribution the n degrees of freedom, but the catch here is that although we are
adding n terms here, we do not really have n independent terms. And, let me actually explain that to
you quickly.
(Refer Slide Time: 20:53)

Adding up n such terms and the point of contention is whether all these terms are independent. What
this mean says let us look at linear dependence part from linear algebra view point, we say that n
variables are linearly independent; if you cannot find linear combination that yields 0, or in other words
if you can express one as a linear combination of the other. Now, if you take X i minus X bar, we know
that X bar itself has been computed form X i right, which means that there is a result what this means is
there is a relation between X i and X bar. How does that come about from this we can show that these
deviations variable are not completely independent; there is a linear dependence between them. How
do we do that? We have n x bar equals sigma X i, right and I have n such X bars and straight away I
can see that sigma minus X bar is 0. So, which means there is a non trivial combination of X i minus X
bars that gives me 0, which means truly this n deviation quantities are not linearly independent. Simply
said if I know n minus 1 deviation quantities, I can construct the nth one; it does not matter it whether it
is last one or it is a first one and so on.

So, this is another way of looking at the loss of decrease of freedom and that also explains why a need
an n minus 1 here although I have n terms, so that I get a unbiased estimator. And on your calculators,
normally we see S n minus 1, and also an S n, this S n minus 1 is nothing but the positive square root of
S square based on n minus 1 in the denominator. Very often, you will also see the literature a one over n
in place of n minus 1. So, we call as S square subscript n. The rest of the expression looks the same.
And the positive square root of this leads to S n. Clearly if this is unbiased, this is the biased estimator,
because expectation s square n is n minus 1 over n sigma square or you can say that this is 1 minus 1
over n sigma square. So, for finite n, this is a biased estimator and you can see the bias here. However
when n becomes very large then this a 1 over n goes to 0; as result, it becomes what is non as
aesthetically unbiased estimator of variance. And therefore, for a large n whether you use s square n
minus 1 or S square n it should not make so much of a difference, but for small n let say 20, 30 and so
on, it can make a bit of a difference.

In hypothesis testing, typically we work with this, because we do not know whether we have large
samples or small samples. But if I know for sure that I have large samples typically the preference is
for this. The reason is you can show theoretically that although this has a bias, this is more efficient S
square n although it is biased; it is more efficient than S square n minus 1, which means it has a lower
error than the S square n minus 1. And for large n, typically does not matter whether you use this or
not, so we prefer to use this. There are other reasons also which have beyond this scope of this course;
the bottom line is we will use this, but there is also this competing estimator and on you are calculators
and an a good calculator, you will see both of this, alright.

Now, we are convinced that there are n minus 1 degree of freedom only, although we have n terms in
the expression for the sample variance. So under the Gaussian distribution assumption for the
observation, we can now say that the sample variance follows as a chi square distribution or typically
we work with standardized statistics. So, we say that the n minus 1 times S square by sigma square
follows a chi square distribution with n minus 1 degree of freedom. Now in a general case, when x size
fall out of some arbitrator distribution non Gaussian distribution what can be said well nothing can be
said theoretically about the distribution, may be for the large sample case we can say it follows a
Gaussian distribution or something like that. But now you can resort to the Monte Carlo simulation
approach and check what kind of distributions S square would have for a non Gaussian kind of
distribution. Again, you can go back here and ask this question through simulations.
(Refer Slide Time: 26:35)

So, let say I do this, say replicate it is a 1000 times, what do I want to replicate? The distribution of the
calculation of variance, so, we use the variance command here and I leave it to you to check in R
whether this variance command uses S square n or S square n minus 1 in R. At the moment, we do not
worry about it. Now what do I have to supply, I will let us give a non Gaussian distribution let say
uniform distribution. And let us take about 20 samples, so that we are looking at a small sample case
and standard uniform distribution all right. So, you see seems to follow Gaussian distribution. What
about even fewer samples may be about 8, so easy slowly it is getting skewed; more or less, it is a chi
square distribution, but perhaps with the different degrees of freedom. So, the distribution stills seems
to be chi square.

Now chi square distribution has this feature that it is defined only over non-negative values. Clearly,
because as you can see from the constructional of the chi square variable it has some square, so some
squares can only be a non zero, non-negative value, right. Therefore, always chi square distributed
variables are defined over the intervals 0 to infinity. We can go back to the large sample case and you
should expect a Gaussian distribution. So, we can take here 200 samples and see what happens you get
a nice Gaussian distribution. Everything in his world seems to be standing to Gaussian right; at least
when it comes to test statistics, so that is a nice thing at least we know in the large sample case
everything tends to the Gaussian one fix distribution that I can always assume, good.
So, now let us get back to the discussion although nothing can be said theoretically about the
distribution, we have to use simulations to do that. What we can definitely say is that this S square that
we have defined here on the screen is always an unbiased estimator, of course, assuming that the x is a
fall in out a random sample, they belong to a random sample that means, they are all independent of
each other. When x size are correlated then it becomes a complicated thing to analyze, and we will not
worry about that in this course.

(Refer Slide Time: 29:20)

The next situation of interest is ratio of variances and I think we talked about this again whether in the
oxide layer thickness for the semi conducted manufacturing process. We wanted to see if one mixture
of gases gave me a lower variance than another mixture of gases. So, typically, we do not compute
differences here, we compute ratios of variances. And this ratio of variances follow us and f distribution
again with the standard assumption in place; we assume that we have taken to random samples
meaning collection of observations of sizes n 1 and n 2 respectively, and also assume that the samples
are falling out of a Gaussian distribution. So, you see as compare to the mean the result available for
sample variances and ratio of variances are quite restrictive, whereas the sample mean always seems to
follow Gaussian distribution or T distribution more or less depending on other sample size is large or
small. But, when it comes to the variance we have theoretical results available predominantly only for
the Gaussian case. There are of course, may be a few research papers that you can look up depending
on you are needs. What kind of distribution this ratio of variances follows; when they fall out of that
the samples fall out of a non Gaussian distribution.

We will in this course assume that is samples fall out of a Gaussian distribution. This F distribution is
characterized by 2 degrees of freedom. Again, let me tell you what degrees of freedom means in
statistics; degrees of freedom means in how many independent ways the variability of the random
variable is affected. When it comes to the F variable, because it is a ratio of 2 random variables S 1
square and S 2 square, we say that the F distribution is characterized by 2 degrees of freedom nu 1 and
nu 2. And, nu 1 is n 1 minus 1, nu 2 is a n 2 minus 1. The reason for n 1 minus n 2 minus 1 is now
obvious from the previous discussion; S 1 square has n 1 minus 1 degrees of freedom; S 2 square has n
2 minus 1 degrees of freedom. So, you can expect therefore, the F distribution to have n 1 minus 1
comma n 2 minus 1 degrees of a freedom.

The analytical expression for the F distribution is quite intimidating it can actually be really
intimidating. However, we do not have to worry about it; normally, we do not give out the analytical
expressions, either we look up a table in a book or the more modern way of doing that would be to just
use the computer software package like R, where you turn to the R F to generate a randomly sample
from an F distribution or P F to compute the probability of a random variable which as F distribution
and so on. So, go ahead and actually generate a probability density function the probability density
function of a random variable that has F distribution with some pre specified numerator and
denominator degrees of freedom. Again, this is a result we will use this result in hypothesis testing.
(Refer Slide Time: 32:44)

Now, the final thing to be discussed in this lecture is that of the sample proportion. Again recall the
example that we discussed in the motivation lecture, there is an manufacture of a controllers for
automobile application, and the manufacture claims that the proportion of defective items is not more
than a particular value and we want to test that hypothesis. So, what do we do here, of course, now the
random variable that we are looking at is a binomial distributed random variable, because now we are
looking at proportions were of the success or failure? And once again, you should go back to the
definition of a binomial random variable, and recap all the conditions under which a random variable
follows a binomial distribution. And one of the parameters, the only parameter that characterized as a
binomial distribution is the proportion or the probability of success, which is denoted by p.

So, assume that now I actually take a sample of size n from a population which follows a binomial
distribution the probability success being P, X is a number of observation belonging to the success class
than P would be X by n. So, what I do is, I collect the sample of size n and count the number of
success. So, in the case of defective items, I would randomly sample from a lot and through an
experiment or through inspection and determine what is non-defective and call that as X. Then the
proportion and estimate of the proportion of success probability success is given by X by n naturally,
where X is the number of success cases by n which is the total number of cases that you have in your
sample. We call that has P hat as an estimate of the proportion sorry probability of success. The small p
is true probability of success.

We will also make an assumption that population is fairly large that is important. And sample size is
also going to satisfy certain conditions given at the bottom of the slide. Now the result of interest was is
the on the sampling distribution of P hat, and the results says that P hat also follows a Gaussian
distribution which is very nice; provided certain conditions are made. And one of the key conditions
sees that the true probability of success is not to close to 0 or 1; that means, here population should not
have too many success cases or too few success cases. There should be some reasonable balance
between success and failure; in that case, the P hat follows a Gaussian distribution. In fact, I would say
that go to you r package, and check if this result holds good by means of simulation again using the
same idea repeat experiments using replicate and plot a histogram.

What about the mean and variance of the Gaussian distribution the mean is n p, if you recall from the
discussion on the statistics that we had, we said that the expected value of a binomially distributed
random variable is n p. And variance is given by n p times 1 minus p that is for that case. But here as
you can see X follows a binomial distribution, but p had therefore, also follows a binomial distribution,
but it is X by n and therefore, you would have here normal distribution with variance p times 1 minus p
over n.

So, now we move on to the last item in this lecture, where we discuss the case of a sample proportion.
And again, you should recall the corresponding example that we used in the motivating lecture, where
we were looking at the number of defective controllers in controllers applied by manufacturer for
automobile applications. And in general, of course, they supplies to all those cases where we are
looking at the defective, non-defective or success, failure, head and tail in a toss - coin toss and so on.
So, we the setting as is as follows, we have a population characterized by binomial distribution and
probability of success being p; and from which, we randomly sample observation of size n. The natural
estimator of the proportion of success - the number of success in a sample of size n is denoted by X in
the slide. And an estimate of the probability of success is naturally X by n that is very much easy to
understand.

If I want to know what is a probability of getting a head in a toss of a coin. What do I do, I repeatedly
conduct those trials the coin toss trials and count the number of heads in those trials let say I conduct
about hundred trials and I count and I get about let us say 55 heads then the probability of getting a
head in any coin toss in general would be 55 by 100 of course, that would give you 0.55. And, as you
conduct more and more trials that is as a number of observation increase or the sample size increase,
we should expect to the estimate to converts to the truth. So, here an unbiased estimator of the
probability of success is P hat given by X by n. Why is it unbiased? It is very easy to see expectation of
P hat is expectation X by n and we know from theory that if X has is a binomially distributed random
variable, its averages n p, therefore, expectation of P hat is p, alright.

Now with that estimator question is, what is the sampling distribution of P hat under the assumption
that the population is fairly large, and that the probability of success p remains constant and that a
couple of more assumptions P hat approximately follows a Gaussian distribution with expected value p
showing that it is an unbiased estimator and the variance being p times 1 minus p by n. How do you get
this? Well, expectation of P hat is expectation of X by n, so that any ways establish that the mean of P
hat is p which is the probability of success. Variance of P hat is variance of X by n square; right.

And, we know once again from the lecture go back to the lecture on statistics that variance of a
binomially distributed random variable is n p times 1 minus p. Therefore, variance of P hat would be n
p times 1 minus p by n square, where you get p, therefore, the variance of P hat to be p times 1 minus p
by n. Now, this result is valid for large n; that means, in the coin toss example, you would have
conducted many trials or in the defective controller example, you would have collected enough items
not 5 or 10 items in your random sample lot and also that the probability of success is neither to low
not to large. That means, the in the original population, you should not have 2 few defective items in
terms of proportion or too large as number of defective items or non-defective items as well. And, there
is a typical guideline that is given here this that is approximation is valid when n p is greater than 5 and
the other condition that is given as well.
(Refer Slide Time: 41:26)

So, we assume that more or less this condition hold good at least in this course for the general case I
urge you to refer to the literature. Now as usual we work with standard or standardized statistics, since
P hat follows a Gaussian distribution, the variable Z, which is P hat minus p by sigma, should also
follow a standard a normal distribution. And that is what your equation tells you here that P hat minus p
by square root of the variance of P hat follows an approximately standard normal distribution, because
P hat itself follows an approximate Gaussian distribution. We use this in the hypothesis testing of the
proportions.

Now in the next lecture of course, now will focus on hypothesis testing, but before I conclude in this
lecture, I would like to reiterate that it is not just important to know these results, but what is equally
important is to know the assumptions under which we have derived this sampling distribution, because
when you have a case of hypothesis testing, you should first look at the assumptions and see whether
those assumptions actually match with the assumptions that we are made here. If not, then you may
have to turn to a literature or some other text book to see what test statistic is appropriate and what is
the sampling distribution of the test statistic under the assumptions in which you have going to conduct
the hypothesis test.

In this course, on the reminder of this course, we will assume that whatever assumptions we have made
here are the assumptions under that will apply to the hypothesis test that examples that we are going to
look at.

(Refer Slide Time: 43:30)

So, the final thing that is of interest towards differences in proportions. And again, the story is the same
very much similar to the difference in the samples mean case. We have two different populations;
remember the soft drink example, we have 2 campuses and we wanted to test, if the proportion of
students on each campus is same when it comes to preference for a soft drink. So, here we have two
populations; from which, I have randomly drawn samples of sizes n 1 and n 1, and as usual all the other
assumptions holding good. I estimate the proportion of has P 1 hat and P 2 hat and the difference that is
P 1 hat minus P 2 hat approximately follows a normal distribution. As a result of which, Z which is a
standardized statistics corresponding to difference in proportions also follows approximately standard
Gaussian distribution.

Now, Z is not truly a statistic, because p is unknown in the denominator, we do not know p. So, what
we do is as for as a denominator is concerned, any way p 1 minus p 2 is something that we will
postulate in a hypothesis test. So, in the numerator it is not of a concern; the denominator it is a concern
why have the p 1 and p 2 appeared in the denominator because if you can see p 1 times 1 minus p 1
over n 1 is nothing but the sigma square, the variance or the you can say variance of P 1 hat and
likewise, the other term being variance of P 2 hat. They have appeared there because we are
standardizing the P 1 hat minus P 2 hat. The point it is now I need to know the denominator to be able
to calculate the observed statistic p 1 minus p 2 is specified by the user in a hypothesis test, but the
denominator has to be supplied as well. If the denominator is not supplied, typically that is the case
then what you do is you use the same approach that you use in the sample mean case. When we did not
know the variance, we used the estimates of variance.

Here the variance is in terms of a proportions, so we use the respective estimates of proportions in this
terms to arrive at the estimates of the respective variances so that is the approach that is used here. At
this point, do not get confused that if I know p 1 and p 2; that means, if I am supplying p 1 and p 2 in
the denominator then where the question of hypothesis is testing. We are not supplying the true p 1 and
true p 2; we are saying I do not know the true p 1 and p 2. One option is to say, I will not solve this
problem and go home, and sleep happily. The other option is to be practical and say well let me get an
estimate and one of the easy ways of doing that is to use the estimates of the proportion from the
respective samples and plug that in to the expression for Z, calculate the observed statistics. P 1 hat and
P 2 hat have calculated from the respective samples; p 1 minus p 2 is specified by the user in the
hypothesis test, and the denominator is again calculated based on your estimates that is P 1 hat and P 2
hat that is the story.

(Refer Slide Time: 47:11)

Again, you should remember that, this result is only valid under certain assumptions all right and that
brings us to the end of this lecture. In fact, the end of the topic of sampling distributions as for as the
univariate case is concerned, for the bivariate case, when we look at correlation we will talk about the
sampling distribution of correlation in that respective lecture that is the respective lecture on hypothesis
testing of correlations. The sampling distributions are useful in many different ways, but the 2 uses for
us in this course is in hypothesis testing and two is in construction of confidence intervals that we will
demonstrate later on.

So, the next lecture that is in order is; obviously, hypothesis testing the crux of this course; well, where
we will introduced some key terminology type I error, type II error, p value, significance level, power
of a hypothesis test and so on. And then, take the sampling distributions take the 8 examples that we
took up in the motivation lecture, quickly go through all those examples. And we now have all the
paraphernalia that is required for hypothesis testing, now it is just a matter of implementing the
standard procedure, therefore, when it comes to using this informational example it is going to be fairly
very easy thing and a breeze thing, alright.

So, see you in the next lecture. Hopefully, you are enjoying all these lectures.

Bye for now.