Noncommunicable disease among adults with cerebral palsy

Study design and data source

A matched cohort study was conducted using data from the UK Clinical Practice Research Datalink (CPRD). The CPRD contains primary care data routinely collected via electronic health records from general practices across all regions of the United Kingdom. The CPRD covers ≈7% of the UK population, and data are largely representative of the UK population in terms of age and sex.^24,25 The CPRD includes data on clinical events, prescriptions, preventive care, referrals, hospital admissions, and lifestyle and anthropometric measures. Clinical diagnoses are recorded in general practices using the Read clinical classification system, which is a hierarchical system of recording clinical data. Read codes, consisting of letters and numbers, are recorded for each clinical diagnosis and are accompanied by Read terms, which are a corresponding written description of the condition. Data collection began in 1987, and we used data obtained during the period from January 1, 1987, to November 30, 2015.

Participants

Patients ≥18 years of age with at least 1 record of CP occurring during the study period and research-standard follow-up (i.e., follow-up meeting criteria developed by the CRPD to determine whether data are of sufficient quality to be used in research) were included. A list of 23 Read codes, which included cerebral palsy, spastic cerebral palsy, and congenital cerebral palsy, was used to identify patients with CP. Start of follow-up (i.e., the index date) for patients with CP was the latest of the following: the date of registration with the general practice, the date data were considered research standard, or January 1 of the year in which the patient turned 18 years of age. Patients with a record of CP were matched to 3 control patients without CP on age (within 3 years in either direction), sex, and practice. Matching on practice controlled for practice-level socioeconomic status and between-practice variations in general practitioner (GP) diagnostic behaviors. The index date for control patients was set to the index date of their matched case. Control patients had a complete history for the duration of their case's follow-up.

Outcomes

We examined the incidence of cancer, cardiovascular disease, type 2 diabetes mellitus, and respiratory disease because these 4 NCDs account for 82% of NCD-related premature mortality.¹² Cancer was identified as a first event of breast, colorectal, lung, or prostate cancer. Breast, colorectal, lung, and prostate cancer are consistently the most common cancer sites, accounting for 52.8% of cancer registrations in England in 2016.²⁶ Cardiovascular disease was identified as a first event of heart failure, hypertensive disease, ischemic heart disease, cerebrovascular disease, or other heart disease. Respiratory disease was identified as a first event of chronic obstructive pulmonary disease (COPD) or asthma. Asthma was identified using a previously validated list of Read codes.²⁷ For other diagnoses, when a validated list was not available, the Read code dictionary was searched to identify Read terms pertaining to all outcomes. Final code lists were reviewed by a consultant cardiologist.

In addition, the following criteria were applied to identify COPD and type 2 diabetes mellitus. A Read code for COPD, first recorded after the patient turned 35 years of age, was used to identify COPD. Diabetes mellitus was identified by (1) a code for a definite diabetes diagnosis, (2) a code for a possible diabetes diagnosis with a subsequent prescription for oral antidiabetics, or (3) ≥2 diabetes drug prescriptions after the start of follow-up. These criteria have been used to identify diabetes mellitus sin a previous study using the CPRD.²⁸ Because it is not always possible to distinguish between type 1 and type 2 diabetes mellitus from Read codes for diabetes mellitus and when the type of diabetes mellitus is stated this information may not be reliable,²⁹ age at first diagnosis, age at first treatment, and treatment received were used to classify type of diabetes mellitus.^28,30 Patients were identified as having type 2 diabetes mellitus if they were >35 years of age at the first diabetes diagnosis or they received exclusively oral antidiabetics after the age of 35 years.

Covariates

Smoking status (current smoker, ex-smoker, nonsmoker), alcohol consumption (current drinker, ex-drinker, or nondrinker), and body mass index (BMI) were identified. BMI was identified using values for height, weight, and BMI taken by the GP.³¹ When multiple records of BMI, smoking status, and alcohol consumption were available, the earliest record after the patient turned 18 years of age was used for the analyses. Although the record may have been taken before the start of research-standard follow-up, it was deemed preferable to include these data to maximize the available data. Records taken before research-standard follow-up were removed in a sensitivity analysis. We calculated the mean yearly GP consultation rate for each patient by dividing the total number of face-to-face or telephone consultations during follow-up by the total number of years of follow-up (i.e., start of follow-up to first event of any NCD or censoring).

Statistical analysis

For each outcome, patients were followed up to the earliest of the following: transfer out of CPRD, end of study period, practice last collection date, death, or first event of the outcome. Baseline characteristics of patients with and without CP were described.

In the primary analysis, a separate Cox model with an underlying age time scale, stratified by matched set, was fitted for each outcome (i.e., cancer, cardiovascular disease, type 2 diabetes mellitus, and respiratory disease).³² A stratified model assumes that patients in each strata have a different baseline hazard but that all other exposures within each strata satisfy the proportional hazards assumption. To determine the hazard of experiencing any NCD, a Cox model stratified by outcome type and matched set was fitted to allow the baseline hazard to vary by outcome type. Robust standard errors were calculated for this model to account for the within-patient dependency between failure times.³³ In a secondary analysis, the incidence of heart failure, hypertensive disease, ischemic heart disease, cerebrovascular disease, other heart disease, COPD, and asthma was compared between patients with and without CP using Cox models stratified by matched set.

Initially, unadjusted Cox models were fitted before adjustment for smoking status, alcohol consumption, BMI, and mean yearly GP consultation rate by including these variables as predictor variables in the Cox model. We adjusted for consultation rate because people with CP may be more likely to have a higher consultation rate and a higher consultation rate may be associated with diagnosis of an NCD. Multiple imputation by chained equations was used to impute missing data for BMI from a linear regression model. The procedure was performed with the ice command in Stata (StataCorp, College Station, TX). To avoid bias in the analysis models, the imputation model included all predictor variables in the outcome model, matching variables, outcomes, and additional predictors of BMI.³⁴ In addition to the event indicator, we included the untransformed event or censoring time because it was found to be one of the best of several approaches when multiply imputing a single normal covariate.³⁵ Forty imputed datasets were generated.³⁴ Boxplots of the observed and imputed data over the imputations were generated to identify discrepancies between the observed and imputed datasets. Estimates were combined with the use of the Rubin³⁶ rules. Because the mechanism underlying missing records for smoking status and alcohol consumption in primary care data is likely to be missing not at random as opposed to missing at random,^37,38 multiple imputation is not appropriate for these data. Missing data for smoking status are most likely pertaining to ex-smokers and nonsmokers, and missing data for alcohol consumption are most likely pertaining to current drinkers.^37,38 Therefore, smoking status was dichotomized as current smokers and noncurrent smokers, with missing data assumed to relate to non–current smokers as proposed by Marston et al.³⁷ Alcohol consumption was dichotomized as current drinkers and non–current drinkers with missing records assumed to pertain to current drinkers. Because this is potentially a strong assumption, we conducted sensitivity analyses to assess the impact of imputation of missing data on effect estimates and inference.

Subgroup analysis

People with ID have established risk factors for NCDs such as obesity, physical inactivity, unhealthy eating behaviors, and antipsychotic medication.^39,–,42 Therefore, the association between CP and NCD may be explained by ID. Furthermore, people with CP and ID are more likely to have epilepsy and sensory impairment and to be nonambulatory compared to those with CP without ID.⁴³ Therefore, ID may be considered a crude proxy for severity of CP or complexity of the individual's overall medical condition. We conducted a subgroup analysis to examine whether the association between CP and NCD was explained by presence of ID. Participants with ID were identified by at least 1 Read code for ID recorded during the study period. The primary analysis was then repeated stratified by ID status to compare the hazard of cancer, cardiovascular disease, type 2 diabetes mellitus, and respiratory disease between patients with CP and ID and their matched controls and between patients with CP without ID and their matched controls.

Model checking and sensitivity analyses

The validity of the assumption of proportional hazards was assessed by plotting scaled Schoenfeld residuals against time for all models and including an interaction term between time and each predictor variable in each model. Martingale residuals were plotted against BMI and mean yearly GP consultations to ensure that there was a linear association between BMI, GP consultations, and each outcome. Three sensitivity analyses were conducted: (1) missing data for smoking status, alcohol consumption, and BMI were excluded (complete case analysis); (2) missing data for smoking status were recategorized as current smokers, and missing data for alcohol consumption were recategorized as non–current drinker; and (3) records of BMI, smoking status, and alcohol intake taken before research-standard follow-up were excluded.

Standard protocol approvals, registrations, and patient consents

The CPRD has obtained research ethics approval from a National Research Ethics Service Committee for observational research using anonymized CPRD data. The protocol was approved by the Independent Scientific Advisory Committee for Medicines and Healthcare Products Regulatory Agency Database Research (protocol No. 16_077 R2A).

Data availability

As a result of CPRD data governance, data obtained from the CPRD are not available from the study authors. Applications to access data may be made to the CPRD at cprd.com.