01.06.2015 Views

Actuarial Modelling of Claim Counts Risk Classification, Credibility ...

Actuarial Modelling of Claim Counts Risk Classification, Credibility ...

Actuarial Modelling of Claim Counts Risk Classification, Credibility ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

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

<strong>Risk</strong> <strong>Classification</strong> 69<br />

2.3.8 Wald Confidence Intervals<br />

The asymptotic variance-covariance matrix ̂<br />

<strong>of</strong> the maximum likelihood estimator ̂ <strong>of</strong><br />

the regression coefficients vector is the inverse <strong>of</strong> the Fisher information matrix. This<br />

matrix can be estimated by<br />

̂̂<br />

=<br />

( ) n∑ −1<br />

˜x i˜x T ̂ i i where ̂ i = d i expŝcore i <br />

i=1<br />

We know from (1.49) that provided the sample size is large enough ̂− is approximately<br />

or0 ̂̂ distributed. It is thus possible to compute confidence intervals at level 1− for<br />

each <strong>of</strong> the j s. These intervals are <strong>of</strong> the form<br />

[̂j − z /2̂̂j<br />

̂<br />

]<br />

j + z /2̂̂j<br />

(2.8)<br />

where ̂ 2̂j<br />

is the estimated variance <strong>of</strong> ̂ j , given by the element j j <strong>of</strong> ̂̂.<br />

Remark 2.3 (Confidence Intervals for the j s: the Likelihood Ratio Method) The<br />

confidence interval (2.8) is based on the large sample properties <strong>of</strong> the maximum likelihood<br />

estimator ̂. Other methods for constructing such a confidence interval are available. One<br />

such method is based on the pr<strong>of</strong>ile likelihood for j that is defined as<br />

j j = max <br />

0 j−1 j+1 p<br />

If ̂ is the maximum likelihood estimator <strong>of</strong> , we have that 2 ( L̂ − L j j ) is<br />

approximately 1 2 provided j is the true parameter value. A confidence interval at level<br />

1 − for j is then given by<br />

{ ∣ ∣∣Lj<br />

j j ≥ L̂ − 1 }<br />

2 2 11−<br />

<br />

where 2 11− is the 1 − th quantile <strong>of</strong> the 2 1 distribution.<br />

2.3.9 Testing for Hypothesis on a Single Parameter<br />

It is <strong>of</strong>ten interesting to check the validity <strong>of</strong> the null hypothesis H 0 : j = 0 against the<br />

alternative H 1 : j ≠ 0. If the jth explanatory variable is dichotomous (think for instance<br />

<strong>of</strong> gender), then failing to reject H 0 suggests that this variable is not relevant to explaining<br />

the expected number <strong>of</strong> claims. If the jth explanatory variable is coded by means <strong>of</strong> a set<br />

<strong>of</strong> binary variables, then the nullity <strong>of</strong> the regression coefficient associated with one <strong>of</strong> the<br />

binary variables means that the corresponding level can be grouped with the reference level.<br />

In such a case, equality between the regression coefficients should also be tested to decide<br />

about the optimal grouping <strong>of</strong> the levels. Hypotheses involving a set <strong>of</strong> regression parameters<br />

will be examined in Section 2.3.13.

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

Saved successfully!

Ooh no, something went wrong!