Study design and Setting
A nationwide, population based retrospective cohort study in UK primary care, 2000-2014.
Data sources
The data were extracted from the Clinical Practice Research Datalink (CPRD) – one of the world’s largest longitudinal primary care databases[16]. By the mid-year of 2013, the CPRD contained anonymized primary care health records for 4.4 million active (alive, currently registered) patients from 7.2% (n=674) of all UK general practices. Patients are broadly representative of the UK general population in terms of age, sex and ethnicity. The CPRD contains data generated during the process of health care in general practices, with demographics, and longitudinal information on clinical aspects (e.g. diagnoses, symptoms, comorbidity), GP contacts, prescriptions, management and referrals and so on. CPRD has been widely applied to health service research.[16]
Patient cohort
The inclusion criteria were all patients who: 1) were diagnosed with a common cancer, as ascertained by Read codes[17] (Lung – B22; Colorectal – B13, B14, B1z0.11; Female breast – B34, B36..00; Prostate – B46); and 2) had at least 6 months of registration with the practice before the cancer diagnosis; and 3) died between 01/01/2000 and 30/04/2014 inclusive; and 4) were registered at a general practice with acceptable data quality.
Study variables
Outcome variables were the service use in the last three months of life in three main categories: a) GP consultations (primary), b) medicines prescribed (using the CPRD unique product codes selected by GPs) and c) referral to secondary care or other care services. We included consultations involving patient contacts, either through face-to-face or telephone, irrespective of where the consultations taking place. We excluded the palliative care referral as this was used to identify patients having palliative care needs recognised (detailed as below).
The explanatory variables were: 1) socio-demographics – age (<50, 50-59, 60-69, 70-79, 80-89, 90+), gender (female, male), year of death, the region where the patient registered general practice was based; 2) clinical variables – cancer site (lung, colorectal, breast, prostate), number of comorbid conditions (0,1, 2, 3, 4+), time (months) from cancer diagnosis to death (0-5, 6-12, 13-36,37-60, 61-120, 121+), the status of having palliative care needs recognised (PC group) or not (non-PC group) following the cancer diagnosis. Comorbid conditions were the 17 conditions included in a modified Charlson Comorbidity Index proposed by Khan et al. [18]. The time period of counting the comorbid conditions was between the diagnosis of the concerned cancer and the death of the patient. A patient who was either on the palliative care register or had a referral record of palliative care after their cancer diagnosis was categorised as having palliative care needs recognised. The Quality and Outcomes Framework (QOF) was first introduced as part of the new General Medical Services contract in 2004.[19] The QOF incentivises general practices to identify and register patients with palliative care needs, regularly review, assess their needs and preferences and proactively planning care. Palliative care was endorsed as a new clinical area for improvement from 2006. The needs of palliative care were identified by the recommended Read codes for palliative care QOF (See Appendix 1 QOF_codes.txt, the download link). A similar approach was successfully employed in a previous CPRD-based study which identified inequity in recognition of palliative care needs for people with heart failure[20]. The palliative care referral was identified using the National Health Service (NHS) specialty field. This contains detailed information about the referring speciality but its completion by general practice staff is not compulsory.
The socio-economic status, measured by the quintile of the index of multiple deprivation (IMD2010) (1=least deprived to 5=most deprived) score of the area where the practice located, was also available as an extra explanatory variable to the patient data from practices in England. The IMD score is an UK government’s official measure.[21] It is a composite score derived from seven domains: income, employment, health and disability, education skills and training, barriers to housing and services, crime and disorder, and living environment. The linkage was done by the CPRD through the postcodes of the practices with which patients were registered.
Statistical analysis
Data were described using count and percentage for categorical variables and mean (standard deviation, SD) for continuous data where applicable. Temporal patterns of service use (consultation, prescribing and referral) in the last 12 months of life were explored using the line chart by plotting month to death against the following statistics: 1) the proportion of patients using a specific type of service with a pre-defined intensity; 2) mean number of a specific type of service. The service use pattern was explored by the status of having palliative care needs recognised or not.
To facilitate interpretations of the findings and the comparability with other studies, we categorised all continuous variables. The generalised estimating equation (GEE) was used to account for the clustering effect within the practice, meaning that the patients from the same practice tended to have similar server use patterns. Three GEE regression models were constructed to evaluate variables independently associated with the outcomes. The GEE model was built with log link function, Poisson distribution and an exchangeable working correlation matrix.
The candidate explanatory variables were selected using a combination of prior clinical knowledge and statistical criteria. The important demographical variables (e.g. age, gender) and clinical variables (e.g. cancer site, number of comorbid conditions) were forced to stay in the model regardless their statistical significance. The multiple adjusted risk ratios (aRRs) were derived from the constructed multiple regression models to quantify the association strength between the explanatory and outcome variables. Two-way interactions of explanatory variables were explored.
A similar multiple regression modelling framework was applied to identify the patient demographic and clinical characteristics associated with if a patient having palliative care needs recognosied or not.
We conducted four sensitivity analyses: 1) the service use patterns where outcome variables (number of consultations, number of prescriptions and number of referrals) were derived from the services used in the last 3 months of life; 2) using the data from practices in England only, it allowed to include IMD2010 as an extra explanatory variable; 3) using the post-2006 data only, as GP practices were incentivised from April 2006 to register patients with palliative care needs; 4) using all GP consultations involving direct patient contact only.
All analyses were performed with the Statistical Analysis Software, version 9.4 (SAS Institute, Cary, NC, USA). A two-sided p value of 0.05 was considered statistically significant.