Analyses

The numbers of individuals in the cohort and in each sub-group were calculated and summarised graphically.

The classification of records as either paediatric or adult was summarised by age and split by condition group and inpatient admissions and outpatient appointments.

The percentage of individuals in the cohort for whom a transition point could be determined (i.e. had at least one paediatric and one adult record) was calculated, for the whole cohort and the Frequent Care group, by year of birth.

Age at transition was estimated for the whole cohort and the Frequent Care group, under the methods outlined above. These were presented graphically by method and density distributions of differences in transition age were compared pairwise between methods. Individuals for whom a transition point could not be estimated were excluded.

Finally, a simulation was used to understand the possible impact of using different methods to estimate transition on an outcome that varied between paediatric and adult care.

Many healthcare outcomes of interest - for example, numbers of A&E visits, inpatient admissions, GP consultations or inpatient bed days are count data. Poisson distributions were used in the simulations as the source of notional outcome data pre- and post-transition, assumed to be counts of a healthcare event. A negative binomial distribution may be more realistic in many circumstances, to account for over dispersion, but adds complication by requiring not only specification of means for the pre- and post-transition distributions, but also their dispersion [47] The pre-transition Poisson distribution mean was set to 2 (as a realistic mean for a healthcare event - GP consultations - in the population [48]) and the post-transition Poisson distribution mean was set to 2.4 (20% higher - clinically significant and also plausible at post-transition ages for GP consultations [48]).

Individuals in the cohort were each assigned five binary transition variables in each year, with 1 indicating adult and 0 indicating child, as follows:

0 in years prior to transition year as estimated by the Last Paediatric estimation method and 1 otherwise

0 in years prior to transition year as estimated by the First Adult estimation method and 1 otherwise

0 in years prior to transition year as estimated by the Fitted estimation method and 1 otherwise

0 in years prior to reaching age 16 years and 1 otherwise (i.e. transition set to age 16 years)

0 in years prior to reaching age 18 years and 1 otherwise (i.e. transition set to age 18 years)

Three outcome variables were assigned each year, one for each of the transition estimation methods. These were populated with counts drawn at random from the pre-transition Poisson distribution if the corresponding transition variable was 0 for that year and drawn from the post-transition distribution if the corresponding transition variable was 1 for that year (Table 1).

Data are shown for a single individual, present from 2006 to 2013 aged 14 to 21 years. In each year the person has five binary transition variables, for the three estimation methods and transition set to age 16 and age 18 years. For this person, the First Adult approach estimates transition at 15 years, Last Paediatric approach estimates transition at 19 years and Fitted approach estimates transition at 17 years - indicated by 0 for paediatric care and 1 for adult care in the Trans_FA, Trans_LP and Trans_Fit variables, respectively. The three outcome variables have values drawn from the pre-transition Poisson distribution (P_2.0) where the corresponding transition variable is 0 and from the post-transition distribution (P_2.4) where the corresponding transition variable is 1. As a visual guide, post-transition observations are in bold type.

Poisson regressions were then used to estimate associations between the count outcomes and the binary transition indicators, using only observations while aged 12 to 23 years and from the final two years of paediatric care and the first two years of adult care (as defined by the transition method used to assign outcomes - e.g. when comparing against the Last Paediatric method, years 2009–2012 would be used in Table 1). Used observations were restricted to this four year window as being of the most interest for identifying changes at transition (for example, data from ages 12 and 20 years might be of little interest for assessing impacts of a transition occurring at age 16 years, but data from ages 14, 15, 16 and 17 years would be more relevant). For each of the three outcome variables, five regressions were run, one each for each of the binary transition variables.

Individuals for whom a transition point could not be estimated were excluded from the simulation.

The sets of models were stratified by demographic variables (sex, deprivation category and ethnic group - the last collapsed to White and non-White due to small numbers) and condition group to explore the potential for systematic bias in using a fixed transition age. The Frequent Care group was also included as a sub group.

The process described above was repeated 10,000 times (with random draws each time from the appropriate Poisson distributions). The change in predicted events associated with transition according to the models was calculated as the mean across the 10,000 runs and the 95% confidence interval as the 2.5 and 97.5 percentiles.