For ethical reasons, pregnant and breastfeeding women are rarely included in clinical trials. To close the knowledge gap regarding the safety of medicine use during pregnancy and breastfeeding, researchers rely on other sources of information. One of them is electronic healthcare data – sometimes referred to as Real World Data. These data are crucial for understanding health events in populations and making decisions in public health and regulation. A paper recently published in American Journal of Epidemiology suggests a method to improve the validity of studying health events in populations using Real World Data.
In database studies, different algorithms are used to measure the occurrence of events, like the occurrence of a disease or a negative effect of an exposure to a medicine in utero. But algorithms are not hard facts and are subject to errors, both false positives and false negatives. Validity indices like sensitivity, specificity, and positive predictive value (PPV) help measure these errors. Guidelines recommend estimating these indices to correct observed rates. The authors of a publication in the American Journal of Epidemiology propose using a screening algorithm on top of the chosen algorithm, and provide a way to test and correct biases in the study results.
Validation studies are often impractical due to cost, time, or ethical reasons. Estimating sensitivity is challenging due to the rarity of events, leading to underestimation of number of events and risk of occurrence of events. Moreover, misclassification of outcome algorithms can introduce bias in risk ratio and risk difference in studies of association.
The authors suggest using an auxiliary screening algorithm alongside the primary one. The primary algorithm should have high positive predictive value, and the screening one should capture additional cases. Estimating positive predictive value for both indicators allow to reduce bias and, if stratified by exposure groups, to test for differential misclassification and reduce bias in measures of association.
“We suggest a method to improve the validity of identifying health events in populations using electronic health data, especially when dealing with algorithms that might have errors. This will be important not only for studies on medicine safety for pregnancy and lactation but also for other epidemiological studies and will mitigate the impact of misclassification on study results,” says Rosa Gini, co-lead of ConcePTION WP7 and Head of the Pharmacoepidemiology Unit of ARS Toscana, Italy.
Do you want to know more? Read the study: Limoncella G, Grilli L, Dreassi E, Rampichini C, Platt R, Gini R. Addressing bias due to measurement error of an outcome with unknown sensitivity in database epidemiological studies. AJE, 2024. DOI: https://doi.org/10.1093/aje/kwae423
By Anna Holm Bodin & Märta Karlén