Risk factors for mortality in patients with anti‐NMDA receptor encephalitis; Comment on data sparsity

    loading  Checking for direct PDF access through Ovid


Dear Editor‐in‐Chief,
We read the article authored by Chi et al.,1 appeared in the Acta Neurologica Scandinavica in 27 December 2016, with great interest. They aimed to examine predictors and causes of death in patients with Anti‐N‐methyl‐D‐aspartate (NMDA) receptor encephalitis. As one of the main findings of multivariable model, they found that admission to an intensive care unit (ICU) found to be very strongly associated with increased risk of mortality which is questionable (hazard ratio [HR]=70.15, 95% confidence interval [CI] =2.39‐2055.45). It has been discussed that large measure of association such as HR with remarkably wide CI can be obtained through the sparse data in which the number of the observations in different combinations of Exposure and Outcome variables is rare.2 In the multivariable analysis, the number of combinations between the variables are increased per each covariate being added to multivariable model and this process exacerbate the data sparsity.2 Hence, it is expected the data sparsity would be severe in the multivariable models than corresponding univariable models. The sparse data bias is unavoidable in the sparse data which inflate the measure of association and expand the CI.2 We clarified the association between the admission to ICU and mortality risk and found that the data sparsity is expected in the univariable model (n ICU+, Death=6; n ICU+, Survival=7; n ICU−, Death=5; n ICU−, Survival=78) and would be more exacerbated in the corresponding multivariable model. Penalization via Data Augmentation is one of the most effective methods which decreases the sparse data bias.2 This methods shrinks the inflated measure of association and narrows the wide CI to give more unbiased and precise measure of association.2 So, we respectfully suggest this effective model for Chi et al. to re‐estimate their crude and adjusted HR on the association between two aforementioned variables to obtain unbiased association.
Also, the authors conducted Cox proportional hazard (PH) regression model where its main assumption (PH assumption) has not been evaluated. We respectfully suggest authors to examine this important assumption and use Extended Cox regression model if it is violated to avoid any misleading findings.

Related Topics

    loading  Loading Related Articles