Stephen Richards

 Every good statistician knows that correlation does not imply causation.  Just because two things appear linked does not mean they are. However, with historical data we often don't have the luxury of carrying out controlled, scientific experiments to see if A really does cause B.

Liabilities within any given portfolio are rarely equal, and they usually differ widely in size. Typically, a large proportion of liabilities is concentrated in a relatively small number of lives, so this should always be checked.

Public-health officials typically allocate their resources using evidence-based methods. They know their annual budget for spending on health measures, and they typically want to save as many lives as they can with that fixed budget.

During an analysis of a large annuity portfolio we took some time out to look at the pattern of mortality by season as well as the overall time trend. We fitted a model for age, gender and season, where the definition of season is that used by the ONS: each season covers three months, and where winter covers December, January and February.

In any model-fitting exercise you will be faced with choices. What shape of mortality curve to use? Which risk factors to include? How many size bands for benefit amount? In each case there is a balance to be struck between improving the model fit and making the model more complicated.

We discussed how we use the AIC to choose between models.

Following my entry on postcode profiling, I have been asked two further questions. The first is whether you have to use Experian's Mosaic system.

Deduplication is an essential part of data preparation for statistical modelling. The phenomenon of multiple policies per person is a major issue for annuity portfolios, and arises from life companies' policy-orientated view of the world.

Or do you know someone who is? Some people are uncomfortable with the idea of statistical models, especially ones with parameters.