How handling of missing data impacts trial results: a meta-epidemiological study

Article type
Authors
Døssing A1, Tarp S1, Furst D2, Gluud C3, Beyene J4, Brandt Hansen B1, Bliddal H1, Christensen R1
1The Parker Institute, Department of Rheumatology, Frederiksberg and Bispebjerg Hospital, Copenhagen, Denmark
2University of California, Los Angeles, David Geffen School of Medicine, Los Angeles, USA
3Copenhagen Trial Unit, Centre for Clinical Intervention Research, Rigshospitalet, Copenhagen University Hospital, Copenhagen, Denmark
4Department of Clinical Epidemiology & Biostatistics, Faculty of Health Sciences, McMaster University, Hamilton, Ontario, Canada
Abstract
Background:
In randomised clinical trials (RCTs) dropouts cause missing data [1] and the efficacy analyses may be conducted using imputation. Two common imputation methods are last observation carried forward (LOCF) and non-responder imputation (NRI). NRI is recommended and presents a conservative method for dichotomous or categorical variables. NRI(all) assumes treatment failure for all dropouts, and NRI(LOE) assumes treatment failure for participants who dropout due to lack of efficacy [2]. No empirical evidence has prior distinguished between the two NRI methods.

Objective:
To investigate how the treatment effect differs between imputation methods.

Methods:
A systematic search in MEDLINE, EMBASE, and CENTRAL was performed. Inclusion criteria: blinded RCTs investigating approved biological agents for rheumatoid arthritis (RA); reporting the American College of Rheumatology 20 response (ACR20) as an outcome. A meta-epidemiological random-effects analysis was used to calculate the odds ratio (OR) of each imputation method. Statistical analyses were based on GLIMMIX procedure in SAS [3].

Results:
Seventy-two RCTs were included and analysed (incl 23,842 analysed patients) from 5237 references. Figure 1 illustrates the OR for each imputation method reflecting the pooled estimate of treatment effect – all in favour of the intervention. Equal treatment effect was seen in NRI(all) and NRI(LOE) trials. LOCF provided the highest treatment effect and Unclear handling of missing data provided the lowest treatment effect. We were unable to assess the treatment effect associated with multiple imputation.

Conclusion:
The present study supports current recommendations; NRI(LOE) and NRI(all) are preferable for securing a conservative estimate of treatment effect, LOCF was associated with less a conservative estimate and should be avoided. More stringent consensus of missing data handling should be addressed.
References: [1]Boers M. Arthritis & Rheum.(2008) [2]EMA. Guideline on Missing Data in Confirmatory Clinical Trials(2010) [3]Dossing A PROSPERO CRD42013006702 (2013).

This work was supported by grants from the Oak and the Michaelsen Foundation.