(251) Methodological Considerations for Matching Algorithms in Cohort Studies of COVID-19 Vaccine Effectiveness

Sunday, August 28, 2022

8:00 AM – 1:45 PM CEST

Location: Hall C

Publication Number: 1296

Presenting Author(s)

Wilhelmine Meeraus, PhD
– AstraZeneca, Cambridge, Cambridgeshire, United Kingdom

Co-Author(s)

Sudhir Venkatesan
– AstraZeneca, Cambridge, Cambridgeshire, United Kingdom
Mark Joy
– Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, UK, United Kingdom
Kathryn S. Taylor
– Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, UK, United Kingdom
Mario Ouwens
– Medical Evidence, BioPharmaceuticals Medical, AstraZeneca, Gothenburg, Sweden, Sweden
Trung N. Tran
– Medical Evidence, BioPharmaceuticals Medical, AstraZeneca, Gaithersburg, MD, USA, United States
John Dennis
– Momentum Data, Pendragon House, 65 London Road, St. Albans, Hertfordshire, AL1 1LJ, United Kingdom
Debasish Kar
– Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, UK, United Kingdom
Anna Forbes
– Clinical Informatics & Health Outcomes Research Group, Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, UK, United Kingdom
Sneha Anand
– Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, UK, United Kingdom
Anke Stuurman
– Medical Evidence, BioPharmaceuticals Medical, AstraZeneca, Cambridge, UK, United Kingdom
Sylvia Taylor, MPH, PhD
– Senior Director, Epidemiology, AstraZeneca, UK, United Kingdom
F D Richard Hobbs
– Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, UK, United Kingdom
Simon de Lusignan
– Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, UK, United Kingdom

Background: COVID-19 vaccine effectiveness (VE) studies must account for calendar time as an important proxy of time-dependent factors that influence VE (e.g. periods with social restrictions/other nonpharmaceutical interventions (NPIs), testing rates, dominant SARS-CoV-2 variant, seasonal factors). Proper matching of vaccinated and unvaccinated participants is one way to address these effects.

Objectives: To evaluate matching algorithms used in published VE cohort studies to inform development of a time dependent strategy for a retrospective cohort study of COVID-19 VE in English computerised medical records data.

Methods: We reviewed matching algorithms used in cohort studies of COVID-19 VE published from 1/1/2021-20/1/2022, considering strengths and limitations of each approach. For our COVID-19 VE study we selected the matching algorithm which best accounted for calendar time effects.

Results: Two broad strategies were used: (a) matching vaccinated group to unvaccinated controls over the entire study period with index dates randomly assigned to controls to match calendar time distribution of vaccination dates in the vaccinated group, or (b) matching by day/week/month all eligible vaccinated participants (including unvaccinated controls who are subsequently vaccinated) to individuals who were unvaccinated at that point in time, ensuring that each matched pair share the same index date (i.e. vaccination date). Option (a) allows conditional exchangeability on selected patient characteristics, but random assignment of index dates in controls does not fully control for effects of calendar time causing potential imbalances for time-dependent factors like NPI and roll-out policies. Option (b) allows for conditional exchangeability in terms of patient characteristics and calendar time. For our COVID-19 VE study we propose an approach where VE estimates are calculated per calendar week in a weekly-matched analysis population. Any decisions to pool VE estimates across time periods can be made following an evaluation of temporal heterogeneity.

Conclusions: The weekly-matching approach better accounts for important effects of calendar time, and it has the advantage that it naturally stratifies results by calendar time.