28.05.2013 Views

Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...

Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...

Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

tel-00703797, version 2 - 7 Jun 2012<br />

Chapitre 1. Sur la nécessité d’un modèle <strong>de</strong> marché<br />

which emphasizes the exponential family characteristic. Let us recall the first two moments<br />

are E(Yi) = πi and V ar(Yi) = πi(1 − πi) = V (πi). Assuming Yi is a Bernoulli distribution<br />

B(πi) implies that πi is both the parameter and the mean value of Yi. So, the link function<br />

for a Bernoulli mo<strong>de</strong>l is expressed as follows<br />

πi = g −1 (x T i β).<br />

Let us note that if some individuals have i<strong>de</strong>ntical covariates, then we can group the data<br />

and consi<strong>de</strong>r Yi follows a binomial distribution B(ni, πi). However, this is only possible if all<br />

covariates are categorical. As indicated in McCullagh and Nel<strong>de</strong>r (1989), the link function and<br />

the response variable can be reformulated in term of a latent variable approach πi = P (Yi =<br />

1) = P (x T i β − ɛi > 0). If ɛi follows a normal distribution (resp. a logistic distribution), we<br />

have πi = Φ(x T i β) (πi = Flogistic(x T i β)).<br />

Now, the log-likelihood is <strong>de</strong>rived as<br />

ln(L(π1, . . . , πn, y1, . . . , yn)) =<br />

n<br />

[yi ln(πi) + (1 − yi) ln(1 − πi)] ,<br />

plus an omitted term not involving πi. Further <strong>de</strong>tails can be found in Appendix 1.8.1.<br />

Link functions<br />

i=1<br />

Generally, the following three functions are consi<strong>de</strong>red as link functions for the binary<br />

variable<br />

<br />

π<br />

1. logit link: g(π) = ln 1−π with g−1 being the standard logistic distribution function,<br />

2. probit link: g(π) = Φ −1 (π) with g −1 being the standard normal distribution function,<br />

3. complementary log-log link: g(π) = ln(− ln(1−π)) with g −1 being the standard Gumbel<br />

II distribution function ∗ .<br />

On Figure 1.4 in Appendix 1.8.1, we plot these three link functions and their inverse functions.<br />

All these three functions are the inverses of a distribution function, so other link functions can<br />

be obtained using inverses of other distribution function. Let us note that the first two links<br />

are symmetrical, while the last one is not.<br />

In addition to being the ca<strong>non</strong>ical link function for which the fitting procedure is simplified,<br />

cf. Appendix 1.8.1, the logit link is generally preferred because of its simple interpretation as<br />

the logarithm of the odds ratio. In<strong>de</strong>ed, assume there is one explanatory variable X, the logit<br />

link mo<strong>de</strong>l is p/(1 − p) = e µ+αX . If ˆα = 2, increasing X by 1 will lead to increase the odds<br />

by e 2 ≈ 7.389.<br />

1.2.3 Variable selection and mo<strong>de</strong>l a<strong>de</strong>quacy<br />

As fitting a GLM is quick in most standard software, then a relevant question is to check<br />

for its validity on the dataset used.<br />

50<br />

∗. A Gumbel of second kind is the distribution of −X when X follows a Gumbel distribution of first kind.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!