Scribd
Upload
22 views3 pages

Chapter_10.pdf

Uploaded by

lapel94551
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
22 views3 pages

Chapter_10.pdf

Uploaded by

lapel94551
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

1

WST322 2013
Chapter 10 Generalised Linear Models

Simple Linear Model vs. Generalised Linear Model (GLM): allows for data
to be non-normal (NB. for Actuarial work)

1. Identify the predictors: COVARIATES/INDEPENDENT VARIABLES

2. Quantify the relationship between the predictors and the risk via a
GLM: RESPONSE VARIABLES

Exponential Families

The distribution of a random variable Y belongs to an Exponential Family


if its density can be written in the form
{ }
yθ − b(θ)
fY (y; θ, ρ) = exp + c(y, ρ)
a(ρ)

for some functions a, b and c; θ is the natural parameter (function of E[Y ]


only), and ρ is the scale parameter.

Mean and Variance from the General Parameters

E[Y ] = b′ (θ), var(Y ) = a(ρ)b′′ (θ)

ˆ b′′ (θ) is the variance function NOT the variance


[ ∂ℓ ] [ 2 ] [( ) ]
∂ℓ 2
ˆ Know the proofs: use E ∂θ ∂ ℓ
= 0 and E ∂θ 2 + E ∂θ = 0 where
ℓ(y; θ, ρ) = lnfY (y; θ, ρ)

Link Functions and Linear Predictors

−−−→
covariate E[Y ] response

ˆ µ = E[Y ] where Y is from an exponential family

ˆ Model µ = β0 + β1 x (for example) is the linear predictor

This study source was downloaded by 100000849253938 from CourseHero.com on 05-23-2024 13:06:42 GMT -05:00

https://www.coursehero.com/file/32727650/Chapter-10pdf/
2

ˆ Link Function g(µ) = η so that E[Y ] = g −1 (η). Should be ⋆ differen-


tiable ⋆ one-to-one

– Canonical/natural link function: g(µ) = θ


– Normal Distribution: g(µ) = µ
– Poisson Distribution: g(µ) = log(µ)
( )
µ
– Binomial Distribution: g(µ) = log 1−µ
– Gamma: g(µ) = 1/µ

ˆ Covariates enter the model through the linear predictor

ˆ Covariates can be categorical or non-categorical: egs. β0 +β1 x, αi +β1 x,


αi + βi x, αi + βi + γij

Deviance of Model Fitting

1. Estimate the coefficients/parameters via MLE

ˆ Maximize ℓ(y; θ, ρ)
ˆ By the invariance property of MLEs the MLE of η gives the MLE
of µ via g(·)

2. Assessing the model fit (balance between too many parameters and a


better fit!):

ˆ Compare likelihood LM of a model to the likelihood of the satu-


rated model LS (i.e. the model where there are as many observa-
tions as parameters, and the fitted values are exactly the observed
values.)
– LIKELIHOOD RATIO STATISTIC LLMS : Model M is good if
LS ≈ LM and poor if the ratio is large
– LOGLIKELIHOOD RATIO STATISTIC log LLMS = ℓS − ℓM :

SM = scaled deviance = 2(ℓS − ℓM )


DM
=
ρ
deviance of the model M
=
ρ
Smaller DM implies a better model.

This study source was downloaded by 100000849253938 from CourseHero.com on 05-23-2024 13:06:42 GMT -05:00

https://www.coursehero.com/file/32727650/Chapter-10pdf/
3

Now if we have normal underlying data, DM ∼ χ2 (·) approx-


imately and χ20.05 (ν) ≈ 2ν. So that Model M1 is an improve-
ment over model M2 if

SM1 − SM2 > 2(p − q)

where p and q are the number of parameters in model M1 and


M2 respectively, and model M1 is a sub-model of M2 .
N.B. Do this for the normal distribution in detail, see
pages 24-28

Residuals Analysis and Assessment of Model Fit

1. Residuals: observed values-fitted values y − ˆ(µ)


ˆ
ˆ PEARSON RESIDUALS √y−(µ) (are skewed for non-normal data)
var(µ̂)

ˆ DEVIANCE RESIDUALS sign(y − µ̂)di where di is the contribu-


tion of yi to the scaled deviance (more symmetrical in general)
ˆ The above two are equal for normal data
ˆ Residuals should show no pattern, otherwise something has been
missed in the relationship between the response and covariates.

2. Parameter Significance: |β̂| > 2std. error(β̂) implies the parameter is


significant

3. Goodness-of-Fit Tests: Pearson’s χ2 and Likelihood Ratio Tests

This study source was downloaded by 100000849253938 from CourseHero.com on 05-23-2024 13:06:42 GMT -05:00

https://www.coursehero.com/file/32727650/Chapter-10pdf/
Powered by TCPDF (www.tcpdf.org)

You might also like