Linear Regression and Correlation

60
Linear Regression and Correlation

description

Linear Regression and Correlation. Fitted Regression Line. Equation of the Regression Line. Least squares regression line of Y on X. Regression Calculations. Plotting the regression line. Residuals. Using the fitted line, it is possible to obtain an estimate of the y coordinate. - PowerPoint PPT Presentation

Transcript of Linear Regression and Correlation

Page 1: Linear Regression and Correlation

Linear Regression and Correlation

Page 2: Linear Regression and Correlation

Fitted Regression Line

54 56 58 60 62 64 66 68 7080

100

120

140

160

180

200

Length (cm)

Y=W

eigh

t(g)

Page 3: Linear Regression and Correlation

Equation of the Regression Line

XbbY 10 Least squares regression line of Y on X

2)())((

1 xxyyxx

i

iib

xbyb 10

Page 4: Linear Regression and Correlation

Regression Calculations

Page 5: Linear Regression and Correlation
Page 6: Linear Regression and Correlation
Page 7: Linear Regression and Correlation

Plotting the regression line

Page 8: Linear Regression and Correlation

Residuals

Using the fitted line, it is possible to obtain an estimate of the y coordinate

The “errror” in the fit we term the “residual error”

ii xbby 10ˆ

ii yy ˆ

Page 9: Linear Regression and Correlation

54 56 58 60 62 64 66 68 7080

100

120

140

160

180

200

Length (cm)

Y=W

eigh

t(g)

Residual

Page 10: Linear Regression and Correlation

Residual Standard Deviation

2)ˆ( 2

|

nyys i

XY

54 56 58 60 62 64 66 68 7080

100

120

140

160

180

200

Length (cm)

Y=W

eigh

t(g)

Page 11: Linear Regression and Correlation

Residuals from example

Page 12: Linear Regression and Correlation

Other ways to evaluate residuals

Lag plots, plot residuals vs. time delay of residuals…looks for temporal structure.

Look for skew in residualsKurtosis in residuals – error not distributed

“normally”.

Page 13: Linear Regression and Correlation

0

20

40

0

20

400

20

40

0

20

40

0 0.3-0.3

Model Residuals: constrained

Pairwise model

Independent model

0 0.3-0.3

Model Residuals: freely moving

Pairwise model

Independent model

-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15-0.2

-0.15

-0.1

-0.05

0

0.05

0.1

0.15

-0.4 -0.3 -0.2 -0.1 0 0.1-0.4

-0.3

-0.2

-0.1

0

0.1

Pairwise model

Independent model

-0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08-0.08

-0.06

-0.04

-0.02

0

0.02

0.04

0.06

0.08

-0.35 -0.3 -0.25 -0.2 -0.15 -0.1 -0.05 0 0.05-0.35

-0.3

-0.25

-0.2

-0.15

-0.1

-0.05

0

0.05

Pairwise model

Independent model

Page 14: Linear Regression and Correlation

Parametric Interpretation of regression: linear models

Conditional Populations and Conditional DistributionsA conditional population of Y values associated

with a fixed, or given, value of X.A conditional distribution is the distribution of

values within the conditional population above

XY |

X|Y

Population mean Y value for a given X

Population SD of Y value for a given X

Page 15: Linear Regression and Correlation

The linear model

Assumptions:LinearityConstant standard deviation

XY

X

Y

10

10X|Y

X|Y

Page 16: Linear Regression and Correlation

Statistical inference concerning

You can make statistical inference on model parameters themselves

1

XY |1b

0b estimates 0estimates

XYs | estimates

1

Page 17: Linear Regression and Correlation

Standard error of slope

95% Confidence interval for

2

|

)(1 xx

sSE

i

XYb

1

1025.01 bSEtb where

Page 18: Linear Regression and Correlation

Hypothesis testing: is the slope significantly different from zero?

1 = 0Using the test statistic:

1

1

bs SE

bt

df=n-2

Page 19: Linear Regression and Correlation

Coefficient of Determination

r2, or Coefficient of determination: how much of the variance in data is accounted for by the linear model.

2

2

)()ˆ(

1yyyy

i

i

Page 20: Linear Regression and Correlation

Line “captures” most of the data variance.

Page 21: Linear Regression and Correlation

Correlation Coefficient

R is symmetrical under exchange of x and y.

X

Y

ssrb *1

22 )()(

))((

yyxx

yyxxr

ii

ii

Page 22: Linear Regression and Correlation

What’s this?It adjusts R to compensate for the factThat adding even uncorrelated variables tothe regression improves R

Page 23: Linear Regression and Correlation

Statistical inference on correlations

Like the slope, one can define a t-statistic for correlation coefficients:

21

11

1r

nrSEbtb

s

Page 24: Linear Regression and Correlation
Page 25: Linear Regression and Correlation

Consider the following some “Spike Triggered Averages”:

-5 0 5 10 15-500V

0500V

-5 0 5 10 15-2mV

02mV

-5 0 5 10 15-500V

0500V

-5 0 5 10 15-500V

0500V

-5 0 5 10 15-5mV

05mV

-5 0 5 10 15-5mV

05mV

-5 0 5 10 15-10mV

010mV

-5 0 5 10 15-5mV

05mV

-5 0 5 10 15-1mV

01mV

-5 0 5 10 15-2mV

02mV

-5 0 5 10 15-500V

0500V

-5 0 5 10 15-2mV

02mV

-5 0 5 10 15-500V

0500V

-5 0 5 10 15-500V

0500V

-5 0 5 10 15-5mV

05mV

-5 0 5 10 15-5mV

05mV

-5 0 5 10 15-10mV

010mV

-5 0 5 10 15-5mV

05mV

-5 0 5 10 15-1mV

-5 0 5 10 15-500V

0500V

-5 0 5 10 15-2mV

02mV

-5 0 5 10 15-500V

0500V

-5 0 5 10 15-500V

0500V

-5 0 5 10 15-5mV

05mV

-5 0 5 10 15-5mV

05mV

-5 0 5 10 15-10mV

010mV

-5 0 5 10 15-5mV

05mV

-5 0 5 10 15-1mV

01mV

-5 0 5 10 15-2mV

02mV

Page 26: Linear Regression and Correlation

STA example

-10 -5 0 5-6

-4

-2

0

2

4

6

R2=0.25.Is this correlation

significant?

N=446, t = 0.25*(sqrt(445/(1-0.25^2))) = 5.45

21

11

1r

nrSEbtb

s

Page 27: Linear Regression and Correlation

When is Linear Regression Inadequate?

CurvilinearityOutliersInfluential points

Page 28: Linear Regression and Correlation

Curvilinearity

60 61 62 63 64 65 66 67 68 69 70-25

-20

-15

-10

-5

0

Page 29: Linear Regression and Correlation

OutliersCan reduce correlations and unduly influence the

regression lineYou can “throw out” some clear outliersA variety of tests to use. Example? Grubb’s test

SDvaluemean

Z

Look up critical Z value in a table Is your z value larger? Difference is significant and data

can be discarded.

Page 30: Linear Regression and Correlation

Influential pointsPoints that have a lot of influence on

regressed modelNot really an outlier, as residual is small.

50 55 60 65 70 75 80 85 90 95 100 10550

100

150

200

250

300

Page 31: Linear Regression and Correlation

Conditions for inference

Design conditions Random subsampling model: for each x observed, y is viewed as

randomly chosen from distribution of Y values for that X Bivariate random sampling: each observed (x,y) pair must be

independent of the others. Experimental structure must not include pairing, blocking, or an internal hierarchy.

Conditions on parameters X10X|Y

XYs | is not a function of X

Conditions concerning population distributions Same SD for all levels of X Independent Observatinos Normal distribution of Y for each fixed X Random Samples

Page 32: Linear Regression and Correlation

Error Bars on Coefficients of Model

Page 33: Linear Regression and Correlation

MANOVA and ANCOVA

Page 34: Linear Regression and Correlation

MANOVA

Multiple Analysis of VarianceDeveloped as a theoretical construct by

S.S. Wilks in 1932Key to assessing differences in groups

across multiple metric dependent variables, based on a set of categorical (non-metric) variables acting as independent variables.

Page 35: Linear Regression and Correlation

MANOVA vs ANOVA

ANOVAY1 = X1 + X2 + X3 +...+ Xn

(metric DV) (non-metric IV’s)

MANOVA Y1 + Y2 + ... + Yn = X1 + X2 + X3 +...+ Xn

(metric DV’s) (non-metric IV’s)

Page 36: Linear Regression and Correlation

ANOVA Refresher

SS Df MS FBetween SS(B) k-1

Within SS(W) N-k

Total SS(W)+SS(B) N-1

1)(

kBSS

)()(

WMSBMS

kNWSS)(

Reject the null hypothesis if test statistic is greater than critical F value with k-1Numerator and N-k denominator degrees of freedom. If you reject the null,At least one of the means in the groups are different

Page 37: Linear Regression and Correlation

MANOVA Guidelines

Assumptions the same as ANOVAAdditional condition of multivariate

normalityall variables and all combinations of the

variables are normally distributedAssumes equal covariance matrices

(standard deviations between variables should be similar)

Page 38: Linear Regression and Correlation

Example The first group receives technical

dietary information interactively from an on-line website. Group 2 receives the same information in from a nurse practitioner, while group 3 receives the information from a video tape made by the same nurse practitioner.

User rates based on usefulness, difficulty and importance of instruction

Note: three indexing independent variables and three metric dependent variables

Page 39: Linear Regression and Correlation

Hypotheses

H0: There is no difference between treatment group (online learners) from oral learners and visual learners.

HA: There is a difference.

Page 40: Linear Regression and Correlation

Order of operations

Page 41: Linear Regression and Correlation

MANOVA Output 2

Individual ANOVAs not significant

Page 42: Linear Regression and Correlation

MANOVA output

Overall multivariate effect is signficant

Page 43: Linear Regression and Correlation

Post hoc tests to find the culprit

Page 44: Linear Regression and Correlation

Post hoc tests to find the culprit!

Page 45: Linear Regression and Correlation

Once more, with feeling: ANCOVA

Analysis of covarianceHybrid of regression analysis and ANOVA

style methodsSuppose you have pre-existing effect

differences between subjectsSuppose two experimental conditions, A and

B, you could test half your subjects with AB (A then B) and the other half BA using a repeated measures design

Page 46: Linear Regression and Correlation

Why use?Suppose there exists a particular variable that *explains* some of

what’s going on in the dependent variable in an ANOVA style experiment.

Removing the effects of that variable can help you determine if categorical difference is “real” or simply depends on this variable.

In a repeated measures design, suppose the following situation: sequencing effects, where performing A first impacts outcomes in B. Example: A and B represent different learning methodologies.

ANCOVA can compensate for systematic biases among samples (if sorting produces unintentional correlations in the data).

Page 47: Linear Regression and Correlation

Example

Page 48: Linear Regression and Correlation
Page 49: Linear Regression and Correlation

Results

Page 50: Linear Regression and Correlation

Second Example

How does the amount spent on groceries, and the amount one intends to spend depend on a subjects sex?

H0: no dependenceTwo analyses:

MANOVA to look at the dependenceANCOVA to determine if the root of there is

significant covariance between intended spending and actual spending

Page 51: Linear Regression and Correlation

MANOVA

Page 52: Linear Regression and Correlation

Results

Page 53: Linear Regression and Correlation

ANCOVA

Page 54: Linear Regression and Correlation

ANCOVA Results

So if you remove the amount the subjects intend to spend from the equation,No significant difference between spending. Spending difference not a resultOf “impulse buys”, it seems.

Page 55: Linear Regression and Correlation

Principal Component Analysis

Say you have time series data, characterized by multiple channels or trials. Are there a set of factors underlying the data that explain it (is there a simpler exlplanation for observed behavior)?

In other words, can you infer the quantities that are supplying variance to the observed data, rather than testing *whether* known factors supply the variance.

Page 56: Linear Regression and Correlation

0 500 1000 1500 2000 2500

Example: 8 channels of recorded EMG activity

Page 57: Linear Regression and Correlation

PCA works by “rotating” the data (considering a time series as a spatial vector) to a “position” in the abstract space that minimizes covariance.

Don’t worry about what this means.

Page 58: Linear Regression and Correlation

0 500 1000 1500 2000 2500

0 500 1000 1500 2000 2500

Note how a single component explainsalmost all of the variance in the 8 EMGsRecorded.

Next step would be to correlatethese components with some other parameter in the experiment.

Page 59: Linear Regression and Correlation

0 500 1000 1500 2000 2500

Largest PCN

eura

l firi

ng ra

tes

Page 60: Linear Regression and Correlation

Some additional uses: Say you have a very large data set, but

believe there are some common features uniting that data set

Use a PCA type analysis to identify those common features.

Retain only the most important components to describe “reduced” data set.