To the Editor:
Administrative health databases are becoming a common source of data to measure disease occurrence. However, database-specific sources of bias (including data inaccuracy, missing data, or misclassification)1–3 may affect the estimates of disease occurrence. Therefore, researchers frequently use case-identifying algorithms, which are validated through medical chart review, surveys, or linkage with other sources of data.
Despite such validation, however, the methodological approach to data analysis itself may also introduce bias for diseases where multiple physician contacts or prescriptions over relatively extended time periods are needed to fulfill the criteria for case diagnosis. The factors leading to bias arise from temporal aspects of administrative databases, such as the time and number of physician visits required to fulfill the criteria, the timing of disease diagnosis, and the time span of the database observation period.
We used the provincial administrative health databases of Québec, Canada (1996–2009) to illustrate and quantify the impact of time-related bias on estimates of the incidence of inflammatory bowel disease, including Crohn disease and ulcerative colitis, in this Canadian province of 7 million people. Cases were identified using 3 case definitions requiring increasing numbers of physician visits with an inflammatory bowel disease diagnosis (from 2 to 4 to 6 visits) (see also eAppendix, http://links.lww.com/EDE/A832).
The Table shows that improper timing of disease diagnosis as the first of the series of visits rather than the case-defining one, caused biases of up to 53% in inflammatory bowel disease incidence. The magnitude of the bias varied with the time span of the observation period and with the number of visits required to fulfill the case definition criteria. The span of the observation period caused biases of up to 21% in incidence when disease diagnosis was timed at the first inflammatory bowel disease contact, but had no impact when diagnosis was timed at the case-defining contact, when the number of visits required was reached.
Bias from timing of disease diagnosis was reduced to approximately 3%–5% when a time period to meet the number of visits was specified. (see eTable 1, http://links.lww.com/EDE/A832 illustrating bias when the criteria were met in specified 2-year period.)
Differentiating between incident and prevalent cases may be difficult when using administrative databases, because information before the start of the study period is not available. Nevertheless, bias in annual incidence estimates can in turn cause bias in prevalence estimates for the same period. Indeed, Büsch et al4 showed that inflammatory bowel disease prevalence varied with the span of observation period and the number of events required to meet the criteria. The use of a disease-free period before the first disease contact helps avoid an overestimation of incidence rates. We found a small decrease in inflammatory bowel disease incidence when a 2-year disease-free period was used, and the magnitude of bias from timing of disease diagnosis and span of observation period changed accordingly (see eTables 2 and 3, http://links.lww.com/EDE/A832 illustrating bias from time-related factors using a mandatory 2-year disease free period before first inflammatory bowel disease contact).
In conclusion, time-related biases in estimating incidence rates can be minimized if the case diagnosis is considered when all criteria are met and if case definitions involve a specified time span. It is important to avoid these biases in studies of disease incidence because they will inherently introduce immortal time bias in subsequent studies of disease prognosis.5 The time required to fulfill the criteria needs to be considered as unexposed.6 As administrative health databases are more often used to estimate disease occurrence for the assessment of the burden of disease and for projecting healthcare expenditures, proper account for the described time-related issues can reduce bias.
McGill University Health Centre
Division of Gastroenterology
Faculty of Medicine
Lady Davis Institute for Medical Research
Centre for Clinical Epidemiology
Jewish General Hospital
1. Glasson EJ, Hussain R. Linked data: opportunities and challenges in disability research. J Intellect Dev Disabil. 2008;33:285–291
2. Iezzoni LI. Assessing quality using administrative data. Ann Intern Med. 1997;127(8 Pt 2):666–674
3. Suissa S, Garbe E. Primer: administrative health databases in observational studies of drug effects–advantages and disadvantages. Nat Clin Pract Rheumatol. 2007;3:725–732
4. Büsch K, Ludvigsson JF, Ekström-Smedby K, Ekbom A, Askling J, Neovius M. Nationwide prevalence of inflammatory bowel disease in Sweden: a population-based register study. Aliment Pharmacol Ther. 2014;39:57–68
5. Suissa S. Immortal time bias in pharmacoepidemiology. Am J Epidemiol. 2008;167:492–499
6. Rothman KJ, Greenland SRothman KJ, Greenland S, Lash TL. Section II study design and conduct. Modern Epidemiology. 20083rd ed Philadelphia, PA Lippincott Williams & Wilkins In