This is the authors' abstract. We don't add key points for this paper.
Potential conflict of interest: Nothing to report. TO THE EDITOR: We have read with great interest the article by Dulai et al. published in Hepatology.1 The systematic review aimed to evaluate risk of mortality by fibrosis stage in nonalcoholic fatty liver disease and to present the results through meta‐analysis.1. The authors provide interesting data on fibrosis stage–specific all‐cause and liver‐related mortality. However, we noticed that the results are still biased because of data sparsity, which have not been addressed by an efficient way. The authors reported in the article that mortality rate ratio (95% confidence intervals [CIs])) by stage 4 is 42.30 (95% CI, 3.51‐510.34).1 The huge effect‐size estimates and impossibly large confidence intervals is a marker of sparse data bias in the estimating of the effect size where there are inadequate data such as 0 number in the strata of Exposure‐Outcomes.2 The authors tried to reduce the bias in the risk estimation attributed to 0 mortality for any stages through adding 0.05 to each cell of number of deaths and person‐years of follow‐up. However, in the presence of data sparsity for tabulated data, conventional bias reduction method is the Firth bias adjustment as half will be added to each cell3; but the method, in some scenarios, can produce unacceptably large bias in the estimates.2 To offset the probable bias attributed to Firth bias adjustment, penalization using data augmentation have been suggested.2 In penalization methods, for example, it is assumed that the risk ratio falls between 1:40 and 40, a common range for effect of risk factors on chronic outcomes. Hence, the risk of mortality by fibrosis stage in nonalcoholic fatty liver disease may be different from those reported in the article1 after adjusting for sparse data bias. A take‐home message for the reader is that sparse data bias is a common mistake in biomedical research4 that can be avoided by efficient and recently introduced statistical methods.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Safiri et al. (2017) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: