Application to case study
6.2 Aims and objectives
6.3.4 Analysis definition
( )
i.e. the timeframe starts on patient’s first registration date at general practice or the date when practice continuously keeping up-to-standard ( ) with GPRD quality control, whichever is later. Timeframe ends when patient transferred out of the practice ( ), on the last collection date of data from the practice, when patient died ( ), or when patient turns 18 years old ( ), whichever is earliest.
6.3.4 Analysis definition
In line with objectives i-v, I define the analysis to be undertaken to address the underpinning issues.
Chapter 6 | 138 i. Trends of overall prescribing rates
We characterised time trends in the prescribing of cisapride and domperidone from 1990 to 2006 by number of prescription, number of patients, age at starting, and age at prescription. The time trend for age was broken into those under the age of two years, and those over two years old. For those under two years old, age were given in months, and in years for over two. We calculated the mean number of prescription per patient per year to investigate the trend in the amount of prescription for individual patient. We also described the drug use by year in terms of prevalent (current in that year) users, and incident (new in that year) users.
ii. Trends of overall prescribing rates by geographical regions within GPRD
The overall trends of prescribing were investigated further by broken them down into geographical regions within the GPRD where the prescriptions originated. The rates of prescribing per million children were given by region over time. Regional variation is investigated to find any evidence whether prescribing trends were different in different regions and may need to be addressed separately in analysis.
iii. Trends of prescribing rates by age and sex
The overall rates of prescriptions were then calculated separately for boys and girls.
Age was given in years. The rates were given as rates per million from the entire GPRD children population by year. This approach estimates the ‘true’ prevalence in GPRD children population by age and sex.
iv. Trends of individual prescribing characteristics
Individual prescribing trends for cisapride and domperidone were characterised in terms of number of prescriptions per patient, mean therapy duration per prescription, and doses prescribed per patient. Doses were characterised by mg/kg/day, which were then compared to the recommended amount of prescription for each drug.
v. Temporal effect from the withdrawal of cisapride on domperidone prescriptions Due to cisapride withdrawal in July 2000 in the UK, we expect an influx in the prescription of domperidone around the same time as those who would have been prescribed with cisapride may have been given domperidone instead. I describe number of children who switched to domperidone. This is carried out for the subgroup of children who had been prescribed both cisapride and domperidone within the study time period.
Chapter 6 | 139 6.3.5 Missing data
As with any longitudinal dataset, missing data is always a problem – GPRD data is no exception. A high rate of missingness is anticipated. Most of the times, data that were missing could be ignored because they were not used in analysis. We acknowledge that some data were missing simply because the GPs did not record them in the first place, but there is no way to find out about this type of missingness therefore was also ignored. There were three scenarios where missing data mattered critically: missing date of birth, missing weight, and missing duration and doses of prescriptions.
GPRD actively anonymise day of birth to protect patient confidentiality. We imputed day of birth as first of the month for the person-time to cover the entire month. We imputed missing month of birth as “January” to cover the entire year. Figure 6.2 shows the improvement in the proportions of missing month of births recorded in GPRD over time.
Weight of patients at the time of prescribing was far from complete. Because we anticipated to characterise doses as a function of body weight, we imputed missing weights based on age in months using data from Growth Online (Harlow Healthcare, 2003). This is assuming that children with missing weight were on average weight for their age. Furthermore, we assumed that weight remained constant for the duration of a single prescription.
Some doses are also anonymised by GPRD to govern patients’ confidentiality. Consequently, prescription durations are unknown. We impute missing durations by simple mean imputation of the non-missing durations for the same drug code on the data set. Where we were unable to impute by drug-specific non-missing mean durations, we impute using the overall mean of non-missing durations. This is assuming that on average the same drugs are prescribed for the same amount of time for children of any age and sex.
Missing doses of prescriptions were imputed using a stratified median imputation approach.
We calculated the median doses stratified by drug codes, age, and sex initially; and imputed missing doses within each stratum. We then relaxed the sex criterion, and imputed median doses by drug codes and age. Finally, we relaxed the drug codes criterion, and imputed median doses by age. This is recognising that the distribution for dosage of a prescription is skewed and that the median of the distribution for a particular drug characterises the dosage
Chapter 6 | 140 prescribed well for children of the same age and sex. This assumes that different GPs prescribe cisapride and domperidone according to the same guideline.
Figure 6.2 Change in the proportion (%) of missing month of birth in the background population