Case inclusion criteria were as previously described for the EURODIAB registers : new diagnoses of type 1 (insulin-dependent) diabetes mellitus among children aged under 15 years resident in the geographically defined region. The completeness of registration was estimated separately in each of the 10-year periods using capture–recapture methodology , which requires that independent primary and secondary sources of ascertainment are available. In most centres, the primary source of ascertainment was through hospital records or notifications by paediatricians and family doctors, whereas secondary sources varied depending on local circumstances and included social insurance schemes, diabetes associations and prescription data.
Annual estimates of the population resident in each centre’s geographically defined area were used as denominators for the calculation of directly standardised incidence rates using a standard population consisting of equal numbers of children in each of six subgroups defined by age group (0–4, 5–9 and 10–14 years) and sex.
Poisson regression was used to estimate the trends in incidence rate within centres. For each centre, a model with terms for age group, sex and an age group × sex interaction was first fitted. Then either a categorical variable representing the 5-year subperiods or a linear term testing for trend across individual years was added to the model to provide comparisons of incidence rates over time that took account of changes in the age structure of the population. For three centres whose registers changed coverage during the period, separate estimates of trend were fitted for each 10-year subperiod, and the two estimates were compared by likelihood ratio test. For the remaining centres, a similar model was fitted but with the added constraint that the fitted lines should meet between 1998 and 1999. Models were fitted using Stata Release 11 (Stata, College Station, TX, USA).
The Joinpoint regression program (Version 3.5 – April 2011; Statistical Methodology and Applications Branch and Data Modeling Branch, Surveillance Research Program National Cancer Institute, Bethesda, MD, USA) specifically designed for surveillance of trends in cancer incidence, was also used to see how sensitive conclusions were to the arbitrary division of the period into two 10-year subperiods. Joinpoint provides greater flexibility by accommodating the fitting of two or more linear segments that join at time points that are estimated from the data. The program provides a permutation test to assess the number of linear segments and the times at which they join, while taking into account the multiple testing issues inherent in the approach. A less conservative Bayesian information criterion for model selection was also employed. In order that Joinpoint should mimic as closely as possible the Poisson regression approach, the log-linear model option was chosen, and heteroscedasticity was taken into account by using the standard error of the annual standardised rates.
Further details are provided in the electronic supplementary material [ESM] Statistical methods.