﻿Hi it's Margriet here.

And in this video we will talk
about multiple regression.

Now first we'll look at what the
regression looks like when you

have multiple predictors instead
of just a single predictor. And

we'll apply that to to an

example. And then there are
three kind of topics that we

need to get through around
multiple regression that is,

standardized coefficients, the
assumptions that are associated

with multiple regression and the
adjusted R ^2.

OK, now from

the previous video you might
remember that there are

several different types of
linear regression and we

already looked at simple
linear regression where you

have just one predictor in
your model. So today we

will talk about multiple
linear regression and that

involves models with multiple
predictors.

Now this is what the regression
line looked like for the

scenario where you had one
predictor. So remember you had

the dependent variable Y the
line that needs an intercept and

slope, so the Intercept is the
beta zero and the slope

determined by the weight
associated with your predictor.

So this is the mathematical
description of a line

and this is the scenario for
simple linear regression.

Now, what does that look
like in the context of

multiple regression?

It has the same form,
right? So the dependent variable

is again represented by Yand

then we have
multiple predictors. So we

have an X1 as in simple
linear regression. But we

also have an X2 and X3 and
you could have a whole range

of X-es so
represented by X n.

And then each of these
predictors has their own

beta or weight.

Of course you also have an
intercept and you have an

error term.

Now let's look at an example.
This example comes from the

Pirates Guide to R. Now pirates
like diamonds, and so this is an

example using the diamonds data
Set that comes with R.

And in that data set there is
information about various

characteristics of a whole
range of diamonds that we have

there. Weight, we have
clarity and we have the color

and various other things.

Now if we.

Put those

predictors into our regression
line, we get the following, so

we have the intercept

And so, let's say
that we want to

try to predict the value of a
diamond based on their weight,

their clarity and their color.

So here you can see

how do the beta of
weight multiplies by weight? So

that's the first predictor. Our
second predictor here is X2.

It's clarity. An our X3 is color.
Of course, we again we have the

intercept and this is what we
produce to predict our value

of the diamond and there will be
an error term as

well of course as.

As is always the case.

So when we then want to build
our model in R, we use the

lm() function for linear model
and the formula that

that function takes, mirrors the
regression line formula so.

Again, it has the form of Y as
a function of X1 plus X2 plus

all the other predictors that
you want to add, so where Y

is the dependent variable or
outcome variable, and X1 and

X2 etc. the independent variables
or the predictor variables.

That's just different names
for the same type of thing.

And of course, you need to tell
R what data you're using.

So if we use our

diamonds example then we could
say OK, let's build a linear

model for diamonds. That is just
the name of the object we're

assigning these to and
then we have lm for the function

and our formula equals the value
Y as a function of weight plus

clarity plus color.

And then we tell it OK, the data
That is this diamonds data set.

So this is how to
to read this,

kind of, formula notation in R
and value is being modelled as a

function of weight, clarity and
color. So y is modeled as a

function of X,

as a function of that thing.

It looks maybe a little bit
weird when you see thatthe

first few times, but when you
kind of read it out loud like

that, it helps you to make
sense of it and it is

important to remember that the
way it is coded in R

mirrors, the regression line
formula.

So when you look at the output,
you actually get that get that

whole thing again, just to
remind you, what you've actually

modeled. Now what are the other
important parts of the output?

So here at the bottom we have
our F statistic which tell us

whether the overall model is
significant or not. So we have

the F statistic and the P value
next to it.

Cause the P value if it is
smaller than .05, it tells

us that all the
predictors together in the

model

Significantly predict y, the model
overall is significant.

And then it also reports the
adjusted R-squared.

Sorryit reports the
R-squared that we talked about

in the previous video on simple
linear regression, but it also

reports the adjusted R-squared
and

As mentioned, I will. I'll talk
about that in in a little while.

And then of course there are
the coefficients and they

tell you whether or not each
predictor (and here we have

our weight and our clarity
predictor color predictor)

whether each one of them

by themselves (while kind of
keeping all the other ones

constant) contributes
significantly to the model, so

these values here on the right.

They tell you whether this
individual predictors and

whether weight is

significant or not, and in this
case a weight is significant and

this is the exact P value. But
there's also the three stars

next to it to indicate that it
is significant at the P is

smaller than .01 level.

And then we have. So weight is
significant and clarity is

significant. In the first
line here that that is your

intercept.

OK. So much about
outputs. Now over to

standardized coefficients.

Now it is important to keep the
metric of each variable, each

predictor in mind when
performing multiple regression.

So you need to ask yourself
what does one unit change

mean for each of the
different predictors?

Now let's look at an example.

So we will use the iconicity
study by Winter at al.

2017 for that.

Now the concept of iconicity
describes whether the the form

of words resembles their
meaning. So onomatopoeic words

such as boom or beeping or
buzzing are examples of words

with high iconicity.

And then in their study they
measured iconicity via rating

scale, asking listeners how
much does this word sound

like what it means?

In the study they modelled
iconicity as a function of

several and kind of linguistic
concepts, and one of them was

sensory experience, another
word frequency an another one

imagineability, and finally
systematicity.

Now don't worry about what
each of these things mean.

You can look at that
paper to learn more about that

if you're interested, But
let's here use it as an example

for why you need to think
about standardized

coefficients when you do
multiple regression.

So here we have our
regression formula, right? So

they modelled iconicity as a
function of sensory

experience, plus imagineability, second predictor,

plus systematicity, third
predictor, plus word

frequency, 4th predictor.

They built a
model and these are the

various coefficients for it.

No.

As you can see, hey, so let's
look at this value. So the

larger the coefficient, usually
the more important that

predictor is in the model.

So if you just look at the

different weights here, you'll see that

systematicity is way bigger
than the others. So you might

conclude from that, that systematicity
is a really important

predictor for iconicity.

But then if we go back to the
question of what does a

one unit change mean for that
predictor? So what does one unit

change in systematicity mean
for in the model,

in the context of the model and
we look at the values that are

actually associated with
systematicity, then we
NOTE Confidence: 0.7899729

identify a problem here.

So if you look at the

minimum value, you know do your
descriptives: you look at the

mean, you look at the standard
deviation and you also look at

the minimum and the maximum and
minimum value of

systematicity in this data set was

this. And so you need to go to
a lot of decimals to actually

find find a number. So this is
negative but very close to zero.

And similarly the maximum number
was also very close to 0.

So,

The variation in
systematicity basically varied from just

below 0 to just above 0, and so
the units were

tiny and if you then go up one
unit, it's a huge jump on the

Y axis basically.

So that that is when you use the

unstandardized
coefficient. If instead,

the numbers that you feed
into the model have been

standardized

(and if you can't remember that
means, please watch the video on

centuring and standardizing),

and so if you standardize

the numbers and you run the
model again, then you get

to this outcome here: the
standardized version. So these

are standardized coefficients
and you can see that now

that

the metric is expressed as a
standard unit across the

different predictors,

they are much more compareable,
so now we're talking about .5 for

sensory experience, minus .4 for
imagineability and O for systematicity,

and minus .3 for word

frequency. And so now you might
actually conclude that here

sensory experience is

the most important
predictor for iconicity.

So that is an example to kind of
illustrate why it is important

to use standardized values so
that you have standardized

coefficients in your model.

OK.

Over to assumptions.

As you know, statistical models
rely on assumptions, and that

is the case for multiple

regression too. So all the claims
made on the basis of a model are

contingent on satisfying its
assumptions to some reasonable

degree, and as mentioned as
mentioned in a previous lecture

for regression, the assumptions
are actually about the error

term, that is, they are related to the
residuals of the model.

Now, if the model satisfies
the normality assumption, its

residuals are approximately
normally distributed.

And you can see that

here (in the blue rectangle).

If the model satisfies the
constant variance assumption,

the spread of the residuals
should be about equal while

moving along the regression
line. And this is also known as

homoscedasticity.

We talked about those two in the
context of simple linear

regression as well.

Now, when it comes to
multiple linear regression,

there's a third assumption
that's important to look at,

and that is collinearity.

But first, let's just review
how we check for

normality and
homoscedasticity of the

residuals.

So it is generally recommended
to assess both normality and

homoscedasticity visually.

That is to assess whether the
residuals are normally

distributed, you draw a
histogram. You can see that here

in the blue box on the slide,

on the left.

And so in the graph on the left,
the residuals for iconicity

model have been plotted and they

look pretty good. Note that the
shape of this thing is is

Bell curve shaped.

It's pretty good.

It is

actually easier
to kind of graphically

explore the normality
assumption via quantile

quantile plot or QQ plot. You
see an example here in the

middle of the slide: this is
a QQ plot.

So when the sample quantiles
assemble around a

straight line, like there,
you know that

the residuals conform with
the normal distribution.

The function that

I usually use in R to look at
that (the qqPlot() function, and plot is

with a capital)

actually also gives
you kind of dashed lines

around that as a as a
confidence interval, and as

long as the individual data
points fall within that range,

so kind of in between those
dashed lines, you know that

you're fine; that the residuals
are normally distributed.

Now, according to the
constant variance

assumption, the error
should be equal across the

fitted values, right? And
that you can investigate in

a residual plot. So in the
green box here,

on the right.

So this plots the residuals on
the Y axis over here,

and fitted values on
the X axis over there.

If the constant variance
assumption is satisfied, the

spread of the residuals should
be approximately equal across

the range of fitted values. So
if you move your

eyes from this side to this
side, this spread should be

equal.

So the residual plot should
basically look like a big block.

And this is the case for
this model this iconicity model.

So maybe they funnel out
a little bit towards the

higher fitted values, but
there is really no drastic

violation of the constant
variance assumption.

Now you might find it a bit
disconcerting that the

assumptions are assessed
visually and there are other

options, such as the Shapiro
Wilk test of Normality, and we will

look at some of those in the
labs as well,but it is really

important to use graphs as well
because they they tell you more

about your model and about your

data. For example, the residuals
may reveal a hidden non-

linearity in the data or they
might reveal some extreme values

that need looking at.

So here we have some examples of

what 'good' residual plots
look like.

Mostly they look rather like
clouds of random dots, and even

though there are some apparent

patterns, they
just results from from a

chance process.

Now let's compare that to some
'bad' residual plots over here.

And indeed, you can clearly see.
these plots show non-constant

variance.

In other words, they violate

that assumption, these
are all examples of

heteroscedastic data rather than

homoscedastic. And as you
move along the X axis along

the fitted values from left
to right the residuals

progressively fan out.

So this is an example of 1 type
of pattern that might occur.

OK, now I already mentioned
that when doing multiple

regression, you also need to
check for collinearity.

Now, collinearity describes a
situation where one predictor

can be predicted by other
predictors. So it arises from

highly correlated predictors.
And then it makes regression

models harder to interpret.

So on the right we have an
example of no collinearity. If

we think of y; In the
middle, we have to circle that

represents y (our dependent or
outcome variable), and here we

have each of our individual
predictors. And as you can see

in this figure,

they each describe part of the
variation in in Y, but they

themselves don't overlap, so
this is a scenario where there

is no collinearity.

This is a scenario where you can

see that some of the variance
that is described by this X3 (the

blue bit), overlaps with variance
that is described by X2 our

second predictor. So it's kind
of that bit (the purple area) that is described

by both. And similarly
this orange area

here, that kind of reflects the
variance that is shared between

the Y variable,
the X1 and X2.

That is, it is a kind
of conceptual representation of

a moderate collinearity.

Now. Here, you have an example
of extreme collinearity. So here

you can see that X1 and X2

almost overlap in the extent
to which they describe the

variance in Y.

Now we can

demonstrate collinearity using
some simulated data.

Here, this gives us some
simulated data and now we create

a variable that is very similar.

So.

First, we've created the next
variable and just told R:

Please give me 50 random

numbers. So if we put that into
a regression formula and we say

OK in this simulated ('made up

data set') the regression
line is going to be described by

this. We have an intercept of 10
and we have here our X of random

numbers, and weight for that is
3 and we add an error term of

another 50 random numbers.

Now, why do we use simulated
data? Because it is sometimes

really helpful to know exactly
what should be in the data and

then see what models look like
to when, when, when you fit a

model to it. What does the model pull out?
Because you know what should be

in there, you can demonstrate
some things quite nicely.

OK, so that is our initial model
and then we create a second

variable X2. And we basically
tell it OK, X2 is X1.

So X2 is same as X1, but we can
just change one value.

So that would create something
like that on the right where X1 and X2

are extremely similar.

And indeed, if you look at the
correlation between the two,

it is very very high (point 98).

OK, so,

that's what we what we made up.

And now we're going to fit a
model to it and see what

So here we have a model with
just the X predictor

and you can see that it
tells us indeed that the

intercept is about 10 and the
weight the beta for X is 2.8.

Now we know that it is 3, this
is pretty close, you might think 'Why

is it not exactly 3?' That is
because we've added this error

term. So it's it can't
estimate it that precisely, but

overall it's pretty accurate.

So that's the model with one

predictor. If we then,

use X2 to predict Y. So
again, this is just one

predictor in here, instead of
X1 we now have X2.

We see that it predicts it pretty
well too, right and given the

setup; given how we have set
these things up, it comes as no

surprise that X2 also predicts

Y. Just as I was the case

with X1, right? OK, now it
gets interesting, with regard to

collinearity. So let's see what
happens when we enter both X and

X2 together into the same model.

Now, this is what you get,
so we have our intercept,

It's still about 10, that's
you know, close to that.

Then what happens to

slope for X2. So here
you can see that the slope for

X2 has changed dramatically.

If it's the only
predictor in the model, it's 2.7

and now it's minus .43.

It's negative, even though
the data has been set up so

that X2 and Y are
positively correlated!

So that that is an example of
what will happen to coefficients

when you're
dealing with strong collinearity

in a model. So the coefficients
can change dramatically

depending on which predictors
are in the model.

So to assess
whether you have to worry about

collinearity in your analysis,
you can use what are called

variance inflation factors.

And they measure the degree to
which one predictor can be

accounted for by other

predictors. So the VIF function
from the car

package in R can be used to
to compute variance inflation

factors and

There are various ideas about

how big a VIF value is deemed
problematic, so Zuur et al. are a

little bit more conservative and
say OK, any VIF value over

three or four, you should look
at that. Another suggest that

you need to worry about anything over

10. I tend to kind of

go with that more often.

So for the iconicity model, an
all VIF values are

close to 1, which is good.

But this is the situation for
the the model of the simulated

data that we looked at. So if
you have both the X and the X2

in the in the model of the
simulated data, the VIF values

are going through the roof

basically: correctly identifying
strong collinearity.

And here the VIF values for the
Iconicity model. You can see

they're all pretty close to
one. So that's all fine.

OK, so what what do you do
if you find that you are

dealing with a situation
with strong collinearity?

Now there's several, you've got
several options. The first one

is that you remove one of the
predictors with a high

VIR, and you need to
really use your

understanding, your
expertise, your knowledge

of the subject to decide,
and also to justify which

one you you remove.

A second solution,

can be to collect more data as that
will basically allow you to

estimate the regression
coefficients more precisely,

which should then help
or kind of decrease the

amount of collinearity.

You can also use another
approach than regression. So

random forests is an example of
that. Or, first do a principle

component analysis to

combine the predictor
variables before you do

regression. So principle

component analysis
identifies how strongly

correlated predictors can be
combined in an inappropriate

way, and then they give you
one new variable that you

can then use in your
regression analysis.

It's also really important to
think about this issue

actually at the planning stage
of your study and kind of

make theoretically motivated
choices as to which one of

possibly correlated measures
you want to include.

OK, finally

yeah, we need to talk about
R-squared and adjusted R ^2.

So, using the glance
function from the broom package,

if you run that on
your on your model summary,

it shows the model summary
output and in this case,

if you do that for the iconicity

model, the adjusted R-squared
and the, sorry, adjusted

R-squared here, and the

R-squared here, they are very
close together, which

suggests there is no problem
with with overfitting.

Um, So what

What adjusted R-squared does is that,

like like R^2,

it measures how much of the
variance in the outcome variable

is described by all the
predictors in the model

together, right? So that's R
squared. Adjusted R-squared

takes the number of predictors
in the model into account

because R-squared it is the

situation that, the more
variables you add,

the better it gets, the more
variance it describes.

But

that is kind of cheating, in
the sense that if you have

more more variables, yes, of
course you will describe it

more if you've got more and
more predictor variables, of

course you will describe more
variance in the outcome

variable, but you might get you to a situation

where

it is highly specific to that
particular data set rather than

more generic and generalizable
to other datasets as well.

So adjusted R-squared takes the
number of predictors in the

model into account, and so in
the case of multiple regression,

you should always report the
adjusted R ^2.

And as you can see here for the
iconicity model, adjusted

R-squared and R-squared are very
similar, which suggests that

there is an appropriate number

of variables in the model
and you haven't

been overfitting

your data with the model.

OK, in summary, we spoke about

the regression line with
multiple predictors and then

about why you need to

think about standardized
coefficients, and they make the

predictors more compareable by
converting them to standard

using standard units. So it really helps with
interpreting the coefficients.

Then we spoke about the
assumptions, so normality and

homoscedasticity of the
residuals and then in addition,

in multiple regression you need
to check for collinearity.

And finally we talked about R-
squared and adjusted R-squared

as a measure of an the overall
variance that's explained by

the by the model.

Thank you very much.
