Publication bias and the human albumin review

Article type
Authors
Preston C, Smyth R, Ashby D
Abstract
Background: The aim of this review was to quantify the effect on mortality of administering human albumin of plasma protein fraction to critically ill patients. Before this review it had been thought that this was a safe and effective treatment. However the results of the meta-analysis showed that contrary to prior beliefs use of human albumin increased the risk of death. Based on a test for funnel plot asymmetry the authors concluded that publication bias was unlikely to have affected the conclusions of this review.

Objectives: To consider the potential impact of publication bias on the albumin review using weighted distribution methodology that adjusts combined effect size estimates for publication bias. Data: The human albumin review taken from the CDSR.

Results: The combined estimate from the meta-analysis gave an odds ratio of 1.88 with 95% confidence interval (1.32 , 2.64). The funnel plot appeared to be visually symmetrical and this was not contradicted by tests for asymmetry. Estimates of the 'file drawer N' ranged from 7 to 77 missing studies required to overturn the statistically significant result while estimates of the actual number of missing studies in existence ranged from 3 to 9. Weighted distribution methodology was used to derive effect size estimates adjusted for publication bias. When the selection process was modelled on p-values only these adjusted estimates ranged between 2.00 and 2.77. Modelling the selection process on p-values and precision gave adjusted estimates between 2.05 and 2.38.

Conclusions: Funnel plot asymmetry resulting from publication bias is only likely to be seen if a simplistic and extreme selection mechanism has been used. Methods that are used to test for publication bias are based on the assumption of funnel plot asymmetry, while methods that estimate the number of missing studies do not allow the impact of publication bias to be assessed. The influence of publication bias on the human albumin review is likely to have under estimated the detrimental effect of this treatment. However in contrast to other examples of publication bias the selection mechanism has moved the point estimate closer to the null value while adjusting for bias makes the result more extreme.