Is omission of free text records a possible source of data loss and bias in Clinical Practice Research Datalink studies? A case–control study

Abstract

Objectives

To estimate data loss and bias in studies of Clinical Practice Research Datalink (CPRD) data that restrict analyses to Read codes, omitting anything recorded as text.

Design

Matched case–control study.

Setting

Patients contributing data to the CPRD.

Participants

4915 bladder and 3635 pancreatic, cancer cases diagnosed between 1 January 2000 and 31 December 2009, matched on age, sex and general practitioner practice to up to 5 controls (bladder: n=21 718; pancreas: n=16 459). The analysis period was the year before cancer diagnosis.

Primary and secondary outcome measures

Frequency of haematuria, jaundice and abdominal pain, grouped by recording style: Read code or text-only (ie, hidden text). The association between recording style and case–control status (χ ² test). For each feature, the odds ratio (OR; conditional logistic regression) and positive predictive value (PPV; Bayes’ theorem) for cancer, before and after addition of hidden text records.

Results

Of the 20 958 total records of the features, 7951 (38%) were recorded in hidden text. Hidden text recording was more strongly associated with controls than with cases for haematuria (140/336=42% vs 556/3147=18%) in bladder cancer (χ ² test, p<0.001), and for jaundice (21/31=67% vs 463/1565=30%, p<0.0001) and abdominal pain (323/1126=29% vs 397/1789=22%, p<0.001) in pancreatic cancer. Adding hidden text records corrected PPVs of haematuria for bladder cancer from 4.0% (95% CI 3.5% to 4.6%) to 2.9% (2.6% to 3.2%), and of jaundice for pancreatic cancer from 12.8% (7.3% to 21.6%) to 6.3% (4.5% to 8.7%). Adding hidden text records did not alter the PPV of abdominal pain for bladder (codes: 0.14%, 0.13% to 0.16% vs codes plus hidden text: 0.14%, 0.13% to 0.15%) or pancreatic (0.23%, 0.21% to 0.25% vs 0.21%, 0.20% to 0.22%) cancer.

Conclusions

Omission of text records from CPRD studies introduces bias that inflates outcome measures for recognised alarm symptoms. This potentially reinforces clinicians’ views of the known importance of these symptoms, marginalising the significance of ‘low-risk but not no-risk’ symptoms.