﻿1
00:00:00,300 --> 00:00:01,633
Hello. Welcome back to the course.

2
00:00:01,633 --> 00:00:03,633
Super excited to have you on board today.

3
00:00:03,633 --> 00:00:06,600
We've got a very interesting
and important tutorial.

4
00:00:06,600 --> 00:00:08,566
Assumptions of linear regression.

5
00:00:08,566 --> 00:00:10,000
So let's have a look.

6
00:00:10,000 --> 00:00:11,266
Here we've got a data set.

7
00:00:11,266 --> 00:00:13,133
With a linear regression applied.

8
00:00:13,133 --> 00:00:16,900
And this linear regression seems
to be serving its purpose really well.

9
00:00:17,366 --> 00:00:20,233
However if we look at the following
three data sets,

10
00:00:20,233 --> 00:00:23,366
we can see that
a linear regression is applied each time.

11
00:00:23,366 --> 00:00:26,800
And in fact it's exactly the same linear
regression as in the first case.

12
00:00:27,133 --> 00:00:30,133
However, those linear regressions
are not serving their purpose.

13
00:00:30,166 --> 00:00:32,666
In fact, they are misleading.

14
00:00:32,666 --> 00:00:36,433
So we shouldn't be using linear
regressions in those situations.

15
00:00:36,966 --> 00:00:39,733
These four data sets are called
the Anscombe Quartet,

16
00:00:39,733 --> 00:00:43,900
and they illustrate that you can't just
simply blindly apply linear regression.

17
00:00:44,100 --> 00:00:49,300
You have to make sure that your data set
is fit for using linear regression.

18
00:00:49,300 --> 00:00:52,300
And that's where assumptions of linear
regression come in.

19
00:00:52,600 --> 00:00:53,733
So let's have a look at them.

20
00:00:53,733 --> 00:00:56,733
There's going to be five
assumptions in total plus an extra check.

21
00:00:57,000 --> 00:00:58,966
The first assumption is linear.

22
00:00:58,966 --> 00:01:01,766
We want to make sure
that there is a linear relationship

23
00:01:01,766 --> 00:01:05,200
between our dependent variable
and each independent variable.

24
00:01:05,366 --> 00:01:06,900
And if you look at the chart

25
00:01:06,900 --> 00:01:10,833
here on the right, you'll see that
the linear regression is misleading.

26
00:01:10,966 --> 00:01:11,400
It is.

27
00:01:11,400 --> 00:01:14,666
There is actually no linear relationship
between the two variables.

28
00:01:14,666 --> 00:01:17,666
So we wouldn't use
this kind of model there.

29
00:01:18,100 --> 00:01:20,766
The second assumption is homoscedasticity.

30
00:01:20,766 --> 00:01:21,300
And even though

31
00:01:21,300 --> 00:01:25,866
it sounds like a complex term,
it actually simply means equal variance,

32
00:01:26,666 --> 00:01:30,166
meaning that you don't want to see
a cone type shape

33
00:01:30,166 --> 00:01:33,166
on your chart with an increasing cone
and decreasing cone,

34
00:01:33,233 --> 00:01:37,600
which would mean that variance
is dependent on the independent variable.

35
00:01:37,900 --> 00:01:40,900
So in this case we wouldn't
use a linear regression either.

36
00:01:41,433 --> 00:01:44,466
The third assumption
is multivariate normality

37
00:01:44,500 --> 00:01:47,400
or normality of error distribution.
If you look at the chart

38
00:01:47,400 --> 00:01:49,733
here on the right,
you can feel that something is off.

39
00:01:49,733 --> 00:01:52,700
The best way to intuitively
think about it is

40
00:01:52,700 --> 00:01:55,833
if you look along the line of the linear
regression,

41
00:01:55,833 --> 00:01:59,200
you want to see a normal distribution
of your data points.

42
00:01:59,566 --> 00:02:02,933
In the case on the right here,
we can see something different.

43
00:02:03,166 --> 00:02:06,600
And so again we wouldn't apply a linear
regression there.

44
00:02:07,166 --> 00:02:10,266
The fourth assumption
is independence of observations.

45
00:02:10,266 --> 00:02:13,300
And this includes the term
no autocorrelation.

46
00:02:13,300 --> 00:02:16,566
Sometimes you'll see this assumption
titled as no autocorrelation.

47
00:02:17,200 --> 00:02:19,900
And what that means is that we don't want

48
00:02:19,900 --> 00:02:23,200
to see any kind of pattern
in our data pattern in the data

49
00:02:23,200 --> 00:02:27,733
like we see here indicates
that our rows are not independent,

50
00:02:27,733 --> 00:02:31,800
that some rows are affecting other rows
and other rows, etc..

51
00:02:32,066 --> 00:02:35,166
A classic example of
this would be the stock market, where

52
00:02:35,866 --> 00:02:40,000
previous prices of future prices
which affect future prices and so on.

53
00:02:40,133 --> 00:02:43,666
So in this case, we wouldn't
apply a linear regression model.

54
00:02:44,500 --> 00:02:47,400
The fifth assumption is
lack of multicollinearity.

55
00:02:47,400 --> 00:02:50,633
Basically we want our independent
variables or predictors

56
00:02:50,966 --> 00:02:54,166
not to be correlated with each other
if they're not correlated

57
00:02:54,333 --> 00:02:56,433
then we can build a linear regression.

58
00:02:56,433 --> 00:02:58,000
If they are correlated

59
00:02:58,000 --> 00:03:01,833
then if we do proceed and build a linear
regression model, then the coefficient

60
00:03:01,833 --> 00:03:05,366
estimates that we get in
the model will become unreliable.

61
00:03:06,133 --> 00:03:08,766
And the sixth point is the outlier check.

62
00:03:08,766 --> 00:03:11,433
This is not an actual assumption,
but rather an extra check.

63
00:03:11,433 --> 00:03:15,366
That is important to keep in mind
when building linear regression models.

64
00:03:15,766 --> 00:03:18,366
If you look at the chart here
on the right, you can see that the outlier

65
00:03:18,366 --> 00:03:22,700
is significantly affecting the linear
regression line that we get.

66
00:03:22,900 --> 00:03:25,900
So something that we want to consider is

67
00:03:26,200 --> 00:03:30,100
should we remove the outliers
before building a linear regression.

68
00:03:30,100 --> 00:03:33,600
Or do we want to build a linear regression
with the outliers included?

69
00:03:33,833 --> 00:03:36,900
This will depend on your business
knowledge and knowledge of the data set.

70
00:03:37,066 --> 00:03:37,833
So there we go.

71
00:03:37,833 --> 00:03:39,966
Those are the assumptions of linear
regression.

72
00:03:39,966 --> 00:03:42,800
In this course
we're going to assume by default

73
00:03:42,800 --> 00:03:45,800
that they are correct
for all the data sets that we work with.

74
00:03:45,933 --> 00:03:48,800
However, when you're working
with your own data sets, it's important

75
00:03:48,800 --> 00:03:52,533
to do these checks to make sure you're
building a linear regression.

76
00:03:52,733 --> 00:03:55,533
When it is fit for the data set.

77
00:03:55,533 --> 00:03:58,700
And to finish off this tutorial,
I have a special bonus for you.

78
00:03:58,700 --> 00:04:02,333
If you enjoyed this explanation,
you can download the PDF

79
00:04:02,333 --> 00:04:07,166
version of this slide and keep it at home
printed out as a poster.

80
00:04:07,200 --> 00:04:09,033
Keep it somewhere handy

81
00:04:09,033 --> 00:04:12,933
for those times when you do need to check
assumptions of linear regression.

82
00:04:13,300 --> 00:04:14,966
If that's something
that you would like to have.

83
00:04:14,966 --> 00:04:16,000
Head on over to Super Data

84
00:04:16,000 --> 00:04:19,366
science.com/assumptions
and you can download your poster there.

85
00:04:19,833 --> 00:04:21,600
And I look forward to seeing you back here
next time.

86
00:04:21,600 --> 00:04:23,400
Until then, enjoy machine learning.
