﻿1
00:00:00,366 --> 00:00:01,933
Dummy variable trap.

2
00:00:01,933 --> 00:00:04,066
Last time we learned
how to create dummy variables

3
00:00:04,066 --> 00:00:09,133
to replace our categorical predictor
estate in the model, we also discussed

4
00:00:09,133 --> 00:00:12,133
that you can never include
both dummy variables at the same time.

5
00:00:12,466 --> 00:00:15,466
In our example,
we omitted the California dummy.

6
00:00:15,733 --> 00:00:17,166
Now why is that?

7
00:00:17,166 --> 00:00:21,366
What will happen if we include the second
dummy variable in the model as well?

8
00:00:22,000 --> 00:00:23,133
Let's see.

9
00:00:23,133 --> 00:00:26,533
The intuition here is that you're
basically duplicating a variable.

10
00:00:26,733 --> 00:00:31,166
This is because D2
always equals to one minus d1.

11
00:00:32,000 --> 00:00:35,633
The phenomenon
where one or several independent variables

12
00:00:35,633 --> 00:00:39,166
in a linear regression predict
another is called multicollinearity.

13
00:00:39,700 --> 00:00:42,700
As a result of this effect,
the model cannot distinguish between

14
00:00:42,866 --> 00:00:48,300
the effects of D1 from the effects of D2,
and therefore it won't work properly.

15
00:00:48,766 --> 00:00:51,533
And this is called
the dummy variable trap.

16
00:00:51,533 --> 00:00:54,533
If you do the math behind the scenario,
you will see that

17
00:00:54,633 --> 00:00:58,200
the real problem is
that you cannot have these three elements

18
00:00:58,200 --> 00:01:02,366
in your model at the same time
the constant and both the dummy variables.

19
00:01:02,800 --> 00:01:05,666
I will leave it up to you
to figure this one out on your own,

20
00:01:05,666 --> 00:01:07,100
and let me know if you have any trouble.

21
00:01:07,100 --> 00:01:09,000
I'll definitely help you out.

22
00:01:09,000 --> 00:01:13,700
So to sum up, whenever you're building
a model, always omit one dummy variable.

23
00:01:14,133 --> 00:01:17,233
And this applies irrespective
of the number of dummy variables

24
00:01:17,533 --> 00:01:20,233
they are in that specific dummy set.

25
00:01:20,233 --> 00:01:22,533
If you have nine,
then you should only include eight.

26
00:01:22,533 --> 00:01:23,933
If you have 100.

27
00:01:23,933 --> 00:01:26,366
Then you should only include 99 of them.

28
00:01:26,366 --> 00:01:29,033
Also note that
if you have two sets of dummy variables,

29
00:01:29,033 --> 00:01:32,000
then you need to apply the same rule
to each set.

30
00:01:32,000 --> 00:01:34,900
For instance,
we could have had a column which specifies

31
00:01:34,900 --> 00:01:37,900
the industry
in which the companies operate.

32
00:01:38,000 --> 00:01:38,800
To build the model.

33
00:01:38,800 --> 00:01:42,466
In that case, we would have had to perform
exactly the same steps and create

34
00:01:42,466 --> 00:01:45,733
another set of dummy variables
specifically for that column.

35
00:01:46,200 --> 00:01:51,200
And then we would include all but one of
those dummy variables in our actual model.

36
00:01:51,633 --> 00:01:54,433
I hope this explanation was useful,
and you will never fall

37
00:01:54,433 --> 00:01:57,133
victim of the dummy
variable trap in your modeling.

38
00:01:57,133 --> 00:02:00,366
Next time we're going to cover
the different ways you can build a model.

39
00:02:00,600 --> 00:02:02,700
We will learn all about backward
elimination,

40
00:02:02,700 --> 00:02:06,300
forward selection, stepwise regression,
and much more.

41
00:02:06,633 --> 00:02:09,600
It's going to be an exciting tutorial
and I look forward to seeing you then.
