0:21

And we're now starting the fourth unit, a unit on a topic that,

in the traditional treatment, would be labeled stratification.

Stratification of your sampling units as a process that can,

for almost no investment of cost, some investment of cost,

but typically not a very large one, give us more efficient samples.

That is, samples that have smaller standard errors than we can get

by just randomization alone.

And certainly by randomization and clustering, as we've talked about before.

0:55

So, welcome.

Welcome to Unit 4 on Being more efficient, the fourth of our six units.

And here we're going to have a series of six lectures about the basic ideas of

stratified sampling.

Stratification is the idea, stratified sampling is the process.

And in our discussions we're going to talk about how to form groups.

This is a grouping exercise, this is one where we form the groups ourselves.

And the ways of forming the groups through an illustration,

something about sampling variance, the precision of our estimates.

Having understood something about sampling variance then, more on how we might do

grouping to achieve greater gains in efficiency.

Gains in precision,

smaller sampling variances, through alternative ways of forming groups.

1:59

In this lecture, the first of our six lectures on Unit 4,

we're going to talk about forming groups.

And there are four basic things that we want to talk about here.

The procedure, the procedure of forming the groups.

2:26

For our purposes, let's go back to an illustration we've had before

where we have a list of individuals.

A list of individuals,

in this case, persons, who are members of the faculty of a university.

And we have in that list, only a portion of which is shown here,

the identification number, a sequence number for each element in the file.

An ID number, in this case an eight-digit ID number.

Their Division, their Sex, and their Rank.

That’s what we know about them, and in our past sampling of activity

when we work with this particular list of frame actually.

We didn't use any of this background information, all we did was use

the labels, the sequence number as a way to draw the sample.

Here what we're going to do is use, deliberately and specifically,

the additional information that’s available for every element in the frame.

3:20

We're not going to let it go to waste.

We'll call those additional variables, such as the Division, the Sex,

and the Rank, the auxiliary information that we're going to use to form

groups that we will create before we make sample selection.

But, let me describe the process, then.

The stratification procedure, we'll go through this in a series of steps,

starts first with the finding of the population, in this case,

it's the population of faculty.

Now I've altered the list a little bit from what we looked at before just to give

us more round numbers,

even numbers, so that it makes things come out a little bit better.

So I've actually added 30 elements to the list.

There's 400 faculty on our list now rather than, I think, before it was 370.

4:02

But that's the population.

It's been redefined slightly, but it's the faculty at this university.

And we have a frame, a list of the faculty members.

We know that that frame may be deficient, it may be missing some faculty who have

joined recently and not gotten into the list before it was produced for us.

Or there may be some faculty who are on the list who are no longer at

the university, they've retired or they're deceased.

So there are some flaws in the frame.

But that is our list.

And on that list we have auxiliary variables, things that we know about each

element, each member of the frame before the sample is drawn.

4:44

And in our case we know a sequence number, an ID, a rank, a sex, and a division.

Now this corresponds to our diagram.

If you recall our scheme,

our seven steps to understanding the statistical estimation process.

The first two steps the finding the population and the frame.

Except what we're going to do is take that frame and as shown here,

we're going to divide into groups.

Now, not like we did with clusters, where there were lots and lots of clusters and

then we sampled clusters to go into the sample.

Here what we're going to do is divide that frame, implicitly dividing the population

as well, but the frame into two groups, three groups, or more.

And from each of the groups, we're going to draw a sample.

All of the groups are going to be in our sample, unlike with cluster sampling,

where only a subset of the groups are going to be in our sample.

5:42

Well then, what do we going to do?

We're going to divide the listed groups based on the auxiliary variables.

The division mean that the variables that we're using for

the subdivision, the auxiliary variables, have to be discrete categories,

or categorical if you will.

They must be known for every element on the list.

If there are some elements on the list for

which we don't know the value of a particular auxiliary variable, we may need

to assign them to a missing data category and count that as one of the categories.

Everything has to be grouped.

Everything has to be in a group, and everything must be in only one group.

And then what we're going to do is count up the number of elements in each group.

And we're going to call that capital N, because that refers back to

the population, but capital N sub something, capital N sub group.

And a fairly typical subscript for this kind of thing is the letter h, and

I'll mention more about why h a little bit later in the course.

6:38

So here's our list,

which has now been divided into three groups on the basis of rank.

The Assistant, the Associate, and the Full professors.

And we see in the last column of this table,

how many of each category there are.

They sum to 400, and this is our starting point.

If we want, we could form three separate files, rather than having it in one file.

Or we could keep them all in one file and only filter out the ones that we need for

any given operation, but it's the same basic idea.

7:07

There's another piece of information that we need for this, and

that is, we not only need to know the number in each group,

we also need to know the fraction of the population in each group.

Do we need to know it?

Well, maybe we don't need to know it, but it's a very useful part of the process

utilized in a lot of the statistical notation that we're going to see.

So we're going to follow that particular approach.

It's actually not needed, but it's one way of dealing with this.

So we also need the fraction of the population that's in each of the groups.

So going back to our list for our three strata, Assistant,

Associate and Full professors.

With the counts, the capital N sub h, there's also a capital W sub h.

W sub h being the fraction in each of the groups, so

the Assistant professors are almost 30% of the population of the frame, 28.75%.

The Associate professors are the smallest group, 18.75%.

And the Full professors the are the largest group, just slightly over half.

That W sub h, W for weight.

We're actually going to use this as a weighing factor for

combining things across groups when we're done.

It's not the same thing as the weight that we're going to talk about at

the very end of our six lectures here, as well as in Unit 6.

It's a different kind of weighting scheme, but

this is a weighting to combine the groups.

And we're going to combine them on the basis of their relative size.

That is, we want to see in the end, anything we come out with, that a little

over of 50% of what we say about the population comes from the Full professors.

A little under 20% comes from the Associate professors, and so on.

8:50

All right, so, what we're going to do then, now that we've formed the groups.

We know how large they are, we have their relative sizes,

we're going to draw a sample from each group.

Looking all the way down to the third bullet from the bottom.

Draw a sample from each group, lowercase n sub h.

This is step 3 in our 7-step process, and we're going to keep track in this

particular case not only of that sample size, but also the sampling rates.

These will be important for us for estimation as well.

Their sampling rates or their sampling fractions,

we'll use both terms interchangeably.

The sampling rate, perhaps is a little bit more descriptive.

Its lowercase n sub h, the number that are in the stratum that are in the sample,

divided by capital N sub h, the number that are in the population or

the frame, that are in the strata.

And there will be three of those in our particular case, an f one,

an f two and an f three.

Three different sampling fractions, three different sampling rates.

So in our particular case I've allocated the sample of about 80.

We're doing a 20% sample, we're taking 80 from the 400, one in five.

And I've made the decision to allocate the sample the same way across each of

the strata as it is for the total population, the total frame.

In this case, 20% Of the first Stratum is 23.

23 of the 115 would be in the sample.

10:13

For the second Stratum, 15 of the 75, again 20% would be in the sample.

And for the third stratum, 42, or 20% of 210 would be in the sample.

So this is another part of the process.

We have to decide how many in each of the strata, and

the sampling rates help us keep track of what we're doing.

In this case, we see we've got a very interesting design.

This would be an EPSEM design.

If you recall that term, equal probability of selection method.

We have the same sampling rate in each of the strata.

We're also going to refer to this when we talk about allocations

as the proportionate allocation.

We'll come back to that, let's not dwell on that right now, but

this is just the illustration.

What we're then going to do is take our sample and sort of our step 4,

remember our step 4 was one of estimation.

But there are going to be two steps to the estimation process here.

What we're going to do is calculate a mean for each of the groups.

We draw our sample of 23 from the assistant professors, we obtain our data,

and perhaps what we're doing is measuring income.

We're asking them, what was your total earnings from the university last year?

Not your family earnings, not the earnings from all sources,

just, how much did you earn from the university?

Because our record keeping system may not show that exactly.

Sometimes it comes from multiple sources, we just want to collect that information.

And so here are the data for the three groups, for assistant, associate, and

full professors.

We see that the mean amount that they reported increases as the rank increases,

and that's not unexpected.

The associate, full professors have been there longer,

they also have accumulated merit.

For their sake of their promotion, that merit has led to increases in salary.

And so we see differences across these groups, at least in the sample.

12:07

Here's that data, just to reinforce it now.

This is everything together.

We keep squeezing things in from the right.

The number in the strata, capital N sub h, the fraction in each of the strata,

capital W sub h.

The sample size in each of the strata, lowercase n sub h.

The sampling fraction in each of the strata, 0.2.

And by the way, W sub h and f sub h, we can see them both there,

they're not the same thing are they?

They're both fractions.

The first one is the fraction of the population, capital W sub h,

in each of the strata.

The second one, f sub h, is the fraction of the Stratum that is in the sample.

And those are two different things we're going to have to keep track of, and

it will be confusing at times as we work with this kind of a design.

12:51

Okay, part B is then to take that result and

combine them to get an estimate, let's say in our case of a mean.

What we want to do is take the means for each of the groups and combine them.

And we're going to combine them by the relative sizes of the groups,

the capital W sub h.

So we have a formula here in which we're saying, let's take the mean from group 1,

$50, multiply it by its fraction, 0.2875 as I recall from that group.

We're going to multiply those two together, and then add to that the $70 for

the second group, times its fraction, 0.1875.

And then finally, add to that now accumulating sum,

the third mean, but multiplied by its fraction, 0.525.

That will be our estimate of the mean.

And what we're doing is taking the separate groups and

giving them a relative contribution to the overall estimate that is

reflected in terms of the population distribution across the strata.

So in this particular case, if we do the calculation,

we see that we end up with a mean of about 74,000.

These are in thousands of dollars, I'm sorry, I should have said that earlier.

$74,750 is our mean.

Which is very close to the actual population mean, in this particular case.

So this is our average,

this is our estimate, it's obtained through this process.

Now, of course, this is only step 4.

We had steps 5, imagining the sampling distribution.

Step 6, calculating a standard error from that sampling distribution.

And step 7, looking at a confidence interval.

So we're not quite done with this process even though we've pretty well filled

out our table now.

And have, in the very bottom right hand corner cell for

the total, the estimate of the population mean on the basis of our sample.

14:40

Now this particular sample has a very nice property from

the credibility point of view.

This has more credibility than a simple random sample,

from the point of view that we have controlled,

in our sample, the distribution across three important groups.

We know those groups differ in salary,

we would have anticipated this from the beginning.

But now, what we've done is deliberately control it.

So that our sample looks like the population,

with respect to three important categories of our population,

where we think that that salary rate would vary.

We're not allowing that to just fluctuate by chance in simple random sampling.

A simple random sample of size 80 could have been only Assistant professors.

It could have been only Full professors.

Not only Associate professors, because there's fewer than 80 of them.

But there are some peculiar distributions across

these three groups that are possible in simple random sampling, not likely, but

possible, that we have eliminated.

We are only looking at sample designs that

the sample follows the population distribution.

This is what some people would think of,

we've used the term representative sampling before.

This is probably as close to representative sampling as we're going to

get with the techniques we're using.

But we're not calling it that,

we're calling is stratified random sampling with proportionate allocation.

And now you see what that proportionate allocation means,

the sample distribution looks like the population distribution because we're

using the same sampling rates across the groups.

16:15

Okay, so let's turn to those last two steps, two more steps to go,

the standard error and the confidence interval in terms of computation.

But now, we're poised to imagine all possible stratified random samples

that have exactly the same characteristics as those that we just saw.

Imagine all possible samples that had three groups, and

from those three groups we drew exactly the same sample sizes that we were talking

about in the previous displays.