The accuracy of death dates recorded in the Clinical Practice Research Datalink (CPRD)

Problem

CPRD provides general practice (GOLD) data linked to several other datasets. This can reveal discrepancies in key outcomes where similar data is available in two linked data sources. We explored discrepancies between death date recorded in the GOLD dataset and linked ONS national mortality data.

Approach

Analysis was conducted on data from 132,124 patients who had a GOLD death date between April 2009 and March 2014. Differences in death dates were examined by age and cause of death. We hypothesised that apparent clinical activity (prescription count, consultation count (including administrative activity), and recording of Read codes) would considerably decrease after the true death date. Clinical activity was determined at different time points (<1 week to >1 year) before and after each date.

Findings

47.4% of patients were male and the average age at death was 79 years. Cause of death reflected national statistics. Only 2 patients had no corresponding ONS death date. GOLD and ONS death dates agreed in 75% of cases, with GOLD >1 week later than ONS in 21% of cases, and >1 week earlier in nearly 3%. Of the observations with a discrepancy in death dates, GOLD was on average later (median 5 days, inter-quartile range (IQR) 1 to 15). The greatest discrepancies were observed in 18-29 year olds (median 15, IQR 2 to 148) and patients with an “external” cause of death (median 15, IQR 2-131). 6% of patients had a prescription issued after GOLD death date. 3.4-times more prescriptions were issued >1 month after GOLD death date than after ONS death date; this was most marked for larger discrepancies in death date where ONS date was later. A majority of patients had one or more Read code(s) (71%) or consultation(s) (94%) recorded after GOLD death date; differences between death dates were less marked than for prescriptions.

Consequences

GOLD death date is generally later than ONS death date, and clinical activity (specifically prescribing) decreases following ONS death date. This is consistent with administrative delays in GP data recording. Delays are most marked where there is likely to be coroner involvement. The accuracy of GOLD death dates should be questioned, and ONS death dates should be used whenever possible.