Mortality in the British Panel Household Survey: a Test of a Standard Treatment for Non-Response

Mortality rates computed from data reported in the British Household Panel Survey are lower than those computed from registration of deaths; the main source of this error is likely to be a failure to distinguish non-response due to unreported death from other forms of non-response. Here we model the interaction between state of health, mortality and non-response of men aged 65 and over in a trivariate probit model so as to correct for non-response. We then explore whether the restrictions on the model coefficients required to produce the observed aggregate mortality rates can be accepted statistically. We find that the required restrictions are rejected suggesting that the standard treatment of non-response does not fully address the problem. A possible explanation is that the original sample is not fully representative of the population.