Dual photon microscopy based quantitation of fibrosis‐related parameters (q‐FP) to model disease progression in steatohepatitis: Methodological issues

We read the valuable article authored by Wang et al. published in HEPATOLOGY in 2017, with great interest. The aim of their study was to examine hepatic fibrosis in nonalcoholic fatty liver disease (NAFLD). In the study, they used dual photon microscopy based quantitation of fibrosis-related parameters (q-FPs). In addition, the univariate and multivariate analyses were conducted on 50 patients with NAFLD and the full spectrum of fibrosis (fibrosis stage I [n 5 9], stage II [n 5 12], stage III [n 5 12], stage IV [n 5 7], and stage V [n 5 10]). The main finding of their study was that both in the univariate (odds ratio [OR] 5 7.16; 95% confidence interval [CI], 3.10-16.53) and multivariate models (OR 5 11.98; 95% CI, 1.47-97.30), strand length was significantly associated with fibrosis stage. It seems that the finding of their study may be problematic and cannot be valid. Generally, it is argued that large effect estimates such as OR with remarkably wide CI should not be regarded as a large effect. It is because of the fact that they may result from sparse data in which the number of events in different combinations of Exposure and Outcome variables is scarce. In other words, in multivariate models, when the number of combination increases, data sparsity is aggravated. While, as shown, in Wang et al.’s study, the adjusted OR on the association between the strand length and fibrosis stage was inflated. Furthermore, in comparison with the corresponding univariate association, its CI was expanded, which resulted in high sparsity on the multivariate model. Indeed, this phenomenon is expected given that the number of events in different stages of fibrosis in the univariate model is low and would be very low in the multivariate model. Fortunately, there are a number of methods that have been suggested to control sparse data bias. Probably, the method introduced by Greenland et al. is one of the most robust ones that effectively removes or decreases the aforementioned bias. Greenland et al.’s method is known as Penalization via Data Augmentation. We aimed to utilize Greenland et al.’s method, to reanalyze the univariate and multivariate association between the strand length and fibrosis stage of Wang et al.’s study. However, the individualized data were required to carry out the whole procedure. Hence, we suggest Wang et al. to use the introduced method to remove sparse data bias and present more valid and precise estimates. Also, there is an important methodological issue in the model building, which should have been regarded by Wang et al. The more detailed information can be found in Supporting File S1.