WEBVTT

00:00:00.107 --> 00:00:03.926
♪ [music] ♪

00:00:21.040 --> 00:00:22.077
- [Thomas Stratmann] Hi!

00:00:22.077 --> 00:00:24.268
In the upcoming series of videos

00:00:24.268 --> 00:00:26.858
we're going to give you
a shiny new tool

00:00:26.858 --> 00:00:30.414
to put into your
Understanding Data toolbox:

00:00:30.414 --> 00:00:31.981
linear regression.

00:00:32.885 --> 00:00:34.668
Say you've got this theory.

00:00:34.668 --> 00:00:37.249
You've witnessed
how good-looking people

00:00:37.249 --> 00:00:39.067
seem to get special perks.

00:00:39.642 --> 00:00:40.878
You're wondering,

00:00:40.878 --> 00:00:43.798
"Where else might we see
this phenomenon?"

00:00:44.132 --> 00:00:45.637
What about for professors?

00:00:45.637 --> 00:00:48.259
Is it possible
good-looking professors

00:00:48.259 --> 00:00:50.010
might get special perks too?

00:00:50.350 --> 00:00:53.899
Is it possible
students treat them better

00:00:53.899 --> 00:00:57.209
by showering them
with better student evaluations?

00:00:57.866 --> 00:01:00.467
If so, is the effect of looks

00:01:00.467 --> 00:01:03.573
on evaluation score
big or [inaudible]?

00:01:04.349 --> 00:01:08.143
And say there is a new professor
starting at a university.

00:01:08.619 --> 00:01:11.810
What can we predict
about his evaluation

00:01:11.810 --> 00:01:13.371
simply by his looks?

00:01:13.940 --> 00:01:17.216
Given that these evaluations
can determine pay raises,

00:01:17.671 --> 00:01:21.709
if this theory were true
we might see professors resort

00:01:21.709 --> 00:01:24.980
to some surprising tactics
to boost their scores.

00:01:25.471 --> 00:01:27.461
Suppose you wanted to find out

00:01:27.461 --> 00:01:30.801
if evaluations really improve
with better looks.

00:01:31.441 --> 00:01:34.450
How would you go about
testing this hypothesis?

00:01:34.956 --> 00:01:36.552
You could collect data.

00:01:36.761 --> 00:01:40.025
First you would have students rate
on a scale from 1 to 10

00:01:40.025 --> 00:01:42.076
how good-looking a professor was,

00:01:42.076 --> 00:01:44.807
which gives you
an average beauty score.

00:01:45.229 --> 00:01:48.552
Then you could retrieve
the teacher's teaching evaluations

00:01:48.552 --> 00:01:50.421
from twenty-five students.

00:01:50.421 --> 00:01:53.273
Let's look at these two variables
at the same time

00:01:53.273 --> 00:01:54.738
by using a scatterplot.

00:01:54.981 --> 00:01:57.419
We'll put beauty
on the horizontal axis,

00:01:57.852 --> 00:02:00.589
and teacher evaluations
on the vertical axis.

00:02:01.463 --> 00:02:05.514
For example, this dot
represents Professor Peate,

00:02:06.173 --> 00:02:08.811
who received a beauty score of 3

00:02:08.811 --> 00:02:11.866
and an evaluation of 8.425.

00:02:12.084 --> 00:02:14.958
This one way out here
is Professor Helmchen.

00:02:14.958 --> 00:02:16.797
- [Ben Stiller, "Zoolander"]
Ridiculously good-looking!

00:02:16.797 --> 00:02:18.721
- [Thomas] Who got
a very high beauty score,

00:02:18.721 --> 00:02:20.872
but not such a good evaluation.

00:02:21.101 --> 00:02:22.283
Can you see a trend?

00:02:22.283 --> 00:02:25.533
As we move from left to right
on the horizontal axis,

00:02:25.533 --> 00:02:27.963
from the ugly to the gorgeous,

00:02:27.963 --> 00:02:31.186
we see a trend upwards
in evaluation scores.

00:02:31.870 --> 00:02:35.174
By the way, the data
we're exploring in this series

00:02:35.174 --> 00:02:38.923
is not made up --
it comes from a real study

00:02:38.923 --> 00:02:40.897
done at the University of Texas.

00:02:41.337 --> 00:02:46.023
If you're wondering, "pulchritude"
is just the fancy academic way

00:02:46.023 --> 00:02:47.880
of saying beauty.

00:02:48.405 --> 00:02:51.474
With scatterplots
it can sometimes be hard

00:02:51.474 --> 00:02:55.594
to make out the exact relationship
between two variables --

00:02:55.594 --> 00:02:59.104
especially when the values
bounce around quite a bit

00:02:59.104 --> 00:03:01.318
as we go from left to right.

00:03:02.000 --> 00:03:04.908
One way to cut through
this bounciness

00:03:04.908 --> 00:03:08.144
is to draw a straight line
through the data cloud

00:03:08.144 --> 00:03:10.775
in such a way that this line
summarizes the data

00:03:10.775 --> 00:03:12.613
as closely as possible.

00:03:13.295 --> 00:03:17.181
The technical term for this
is "linear regression."

00:03:17.669 --> 00:03:20.888
Later on we'll talk about
how this line is created,

00:03:20.888 --> 00:03:24.278
but for now we can assume
that the line fits the data

00:03:24.278 --> 00:03:26.456
as closely as possible.

00:03:27.087 --> 00:03:29.536
So, what can this line tell us?

00:03:30.067 --> 00:03:32.596
First, we immediately see

00:03:32.596 --> 00:03:35.358
if the line is sloping
upward or downward.

00:03:36.107 --> 00:03:39.827
In our data set we see
the [fitted] line slopes upward.

00:03:40.794 --> 00:03:43.807
It thus confirms what
we have conjectured earlier

00:03:43.807 --> 00:03:45.587
by just looking at the scatterplot.

00:03:46.070 --> 00:03:50.237
The upward slope means
that there is a positive association

00:03:50.237 --> 00:03:53.026
between looks
and evaluation scores.

00:03:53.544 --> 00:03:55.907
In other words, on average,

00:03:55.907 --> 00:03:59.469
better-looking professors
are getting better evaluations.

00:03:59.768 --> 00:04:03.939
For other data sets we might see
a stronger positive association.

00:04:04.377 --> 00:04:07.420
Or, you might see
a negative association.

00:04:07.857 --> 00:04:10.764
Or perhaps no association at all.

00:04:11.158 --> 00:04:13.903
And our lines
don't have to be straight.

00:04:14.389 --> 00:04:17.304
They can curve to fit the data
when necessary.

00:04:17.770 --> 00:04:21.262
This line also gives us
a way to predict outcomes.

00:04:21.579 --> 00:04:25.569
We can simply take a beauty score
and read off the line

00:04:25.569 --> 00:04:28.429
what the predicted
evaluation score would be.

00:04:28.609 --> 00:04:30.546
So, back to our new professor.

00:04:31.097 --> 00:04:34.109
We can precisely predict
his evaluation score.

00:04:34.683 --> 00:04:36.749
"But wait! Wait!" you might say.

00:04:37.019 --> 00:04:38.749
"Can we trust this prediction?"

00:04:39.233 --> 00:04:41.665
How well does
this one beauty variable

00:04:41.665 --> 00:04:43.515
really predict evaluations?

00:04:44.844 --> 00:04:47.890
Linear regression gives us
some useful measures

00:04:47.890 --> 00:04:49.770
to answer those questions

00:04:49.770 --> 00:04:52.039
which we'll cover
in a future video.

00:04:52.838 --> 00:04:55.439
We also have to be aware
of other pitfalls

00:04:55.439 --> 00:04:58.340
before we draw
any definite conclusions.

00:04:58.833 --> 00:05:00.430
You could imagine a scenario

00:05:00.430 --> 00:05:03.639
where what is driving
the association we see

00:05:03.639 --> 00:05:06.900
is really a third variable
that we have left out.

00:05:07.344 --> 00:05:09.965
For example,
the difficulty of the course

00:05:09.965 --> 00:05:12.456
might be behind
the positive association

00:05:12.456 --> 00:05:15.645
between beauty ratings
and evaluation scores.

00:05:16.052 --> 00:05:18.956
Easy intro. courses
get good evaluations.

00:05:19.228 --> 00:05:22.972
Harder, more advanced courses
get bad evaluations.

00:05:23.660 --> 00:05:27.668
And younger professors might
get assigned to intro. courses.

00:05:28.080 --> 00:05:32.095
Then, if students judge
younger professors more attractive,

00:05:32.095 --> 00:05:34.335
you will find
a positive association

00:05:34.335 --> 00:05:37.383
between beauty ratings
and evaluation scores.

00:05:37.861 --> 00:05:40.388
But it's really
the difficulty of the course,

00:05:40.388 --> 00:05:43.537
the variable that we've left out,
not beauty,

00:05:43.537 --> 00:05:45.848
that is driving evaluation scores.

00:05:46.346 --> 00:05:49.807
In that case, all the primping
would be for naught --

00:05:50.289 --> 00:05:54.441
a case of mistaken correlation
for causation,

00:05:54.900 --> 00:05:58.166
something we'll talk about further
in a later video.

00:05:58.922 --> 00:06:02.069
And what if there were
other important variables

00:06:02.069 --> 00:06:05.781
that affect both beauty ratings
and evaluation scores?

00:06:06.626 --> 00:06:09.575
You might want to add
considerations like skill,

00:06:09.846 --> 00:06:14.577
race, sex, and whether English
is the teacher's native language

00:06:14.577 --> 00:06:18.994
to isolate more cleanly the effect
of beauty on evaluations.

00:06:19.408 --> 00:06:21.758
When we get
into multiple regression

00:06:21.758 --> 00:06:24.477
we will be able to measure
the impact of beauty

00:06:24.477 --> 00:06:26.219
on teacher evaluations

00:06:26.219 --> 00:06:28.368
while accounting
for other variables

00:06:28.368 --> 00:06:30.737
that might confound
this association.

00:06:31.762 --> 00:06:35.509
Next up, we'll get our hands dirty
by playing with this data

00:06:35.509 --> 00:06:39.070
to gain a better understanding
of what this line can tell us.

00:06:41.169 --> 00:06:42.445
- [Narrator] Congratulations!

00:06:42.445 --> 00:06:45.247
You're one step closer
to being a data ninja!

00:06:45.568 --> 00:06:47.139
However, to master this

00:06:47.139 --> 00:06:48.700
you'll need
to strengthen your skills

00:06:48.700 --> 00:06:50.404
with some practice questions.

00:06:50.865 --> 00:06:53.976
Ready for your next mission?
Click "Next Video."

00:06:54.313 --> 00:06:55.364
Still here?

00:06:55.598 --> 00:06:58.325
Move from understanding data
to understanding your world

00:06:58.325 --> 00:07:01.642
by checking out MRU's
other popular economics videos.

00:07:01.892 --> 00:07:04.406
♪ [music] ♪