Improper reporting of multivariable logistic regression.

Sir,

The application of multivariable regression models have increased over the past years in Indian medical journals. The main reason behind this is availability of menu-driven statistical packages. Our review suggests non-statisticians and medical professionals are applying these packages without adequate knowledge of these models. Various medical institutes and other organization are organizing the training programs or work-shops to increase the awareness about statistical methods and statistical modeling in India. Apart from these, various journals are issuing the Tutorial, Guidelines to spread the awareness. These efforts improved the basic requirement like mentioning exact P-value, confidence intervals of estimates, detail of software used, multiple comparison adjustment, etc.,[1] But model building and reporting of multivariable regression models is still inadequate in medical journals not only in Indian journals but also in foreign journals.[2] Multivariable logistic regression (MLR) is most commonly applied multivariable regression method in medical journals.[3] Multivariable regression models require in-depth understanding of model building, its related assumptions, proper interpretation of regression coefficient, and correct reporting to get a valid and reliable conclusion from model results.

The study published by Mathew and Nanoo[4] to assess the risk factors associated with adolescent suicide attempts using MLR and reported their results, which are presented in Table 1.

Table 1

Multivariate logistic regression analysis to assess risk factors

Author reported log-odds (regression coefficient) and its standard error, most of the standard error of log-odds are in hundred, which seems to be impossible because P-value of the regression coefficient are evaluated by Wald test, which is equal to square of regression coefficient divided by its standard error.

The MLR results of the study showed Distancing and PSLES score as significant risk factors, but 95% confidence intervals of these risk factors did not include 1. Thus, P-value cannot be significant. On the other hand, accepting responsibilities had significant preventive effect (OR < 1) and also 95% CI did not include 1, but P-value was not significant.

It seemed author did not consult the competent statistician/epidemiologist in compiling and in reporting of study data. Peer reviewers also did not raise these issues. This manuscript had other model building and reporting deficiencies. Improper reporting of results and not reporting of model assumptions explicitly wastes the researcher valuable efforts. This also raised question mark on reliability of results to the reader. Furthermore, these types of errors affect the quality of journal also.

The author using MLR should report model results completely and accurately. The non-statistician and medical professional should use these models after proper understanding of model assumptions and its reporting. Karkouti et al.[5] provides a good example for reporting and model building of MLR model. In addition, manuscript using multivariable regression models should be peer reviewed by the competent statistician to avoid such type of errors.