Background
The presence of a person’s record in an electronic health record database is not random but is usually a result of presenting to medical services for some condition or illness. People in electronic health records are therefore systematically different from those not in electronic health records (Goldstein 2016). Health records contain people with more medical encounters than the general population. When examining the electronic health records for associations between different conditions, this bias can lead to spurious associations.
Example
In electronic health records, the prevalence of depression among pregnant women might be seen to be greater than that of non-pregnant women. However, pregnant women attend medical services throughout and after their pregnancy, giving them more opportunities than women not attending medical services to be assessed and diagnosed with mental health conditions. This might distort a relationship between pregnancy and depression. There might be a relationship between the number of times each woman was seen by a healthcare worker, and the likelihood of receiving a diagnosis of depression. Caution must thus be taken when extrapolating results to women in the general population (rather than the electronic health record group).
Impact
The impact of informed bias is illustrated in an example in which researchers looked at the relationship between depression and diabetes in a collection of hospital health records (see Figure). Adjusting for the number of healthcare encounters the diagnosis of depression in people with diabetes compared to non-diabetics decreased by 37% (from an odds ratio of 2.15 unadjusted to 1.36 adjusted odds).
Controlling for informed presence bias due to the number of health encounters in an electronic health record. Am J Epidemiol. 2016 Dec 1;184(11):847-855.
Preventive steps
Analyses of electronic health records could look at numbers of contacts with health professionals and how this relates to diagnoses recorded. Studies utilising electronic health records should be aware of this bias, note it as a limitation, and, where applicable undertake sensitivity analyses to adjust for the number of consultations per patient.
