More information about text formats
Thank you for raising this important point. Actually, the Poisson regression is usually used for count data with the variance equal to the mean. And Likert scale data is not suited to this statistic method directly. However, we standardized the scale and the data acquisition for the disability status within 30 days. We assumed that the standardized scores of each domain and summary score (from 0 to 100) as the count of disability status event in 30 days. For analysis the association between the variables of demographic data and standardized WHODAS 2.0 score, we choose the Poisson regression analysis, which could not be perfect for this study. (And the data is near to 1 even statistical significant) Therefore, we didn’t mention the outcome of table 3 in discussion part and conclusion part (merely, mentioned in result part). Our study finding is based on table 2 and we discussed this finding (lower disability status in the WHODAS 2.0 domains of getting along and social participation for patients with dementia with formal education compared with those without formal education) in discussion and conclusion part.
Thank you again for your precious suggestion. We agree that Multi-level IRT could be an appropriate way to analyze multiple Likert scales. The following studies of original Likert scales of WHODAS 2.0 will be analyzed as your suggestion and this could lead our study to be more convincing.
Poisson regression is unsuitable for analysing data from Likert scales, even in aggregate (see http://rcompanion.org/handbook/E_01.html).
Summing enough Likert scales (as when summing enough random variables) might result in summary data which are suitable for least squares regression, via the central limit theorem. But, Poisson regression is suitable for count data where the variance is equal to the mean (count data that violate this equality may require negative binomial regression).
Since the statistical analysis is inappropriate, the Results and Conclusions may be unsound.
Multi-level IRT is probably an appropriate way to analyse multiple Likert scales (e.g. Luo & Wang, Stat Med. 2014 Oct 30; 33(24): 4279–4291)