Ammar Kuti Nasser In my limited experience working with medical statistics, almost everything fundamental I learned as a statistician, including basic probability theory, inference and distributions became useful. Some specific topics that come to mind are general and generalised models, experimental design, sampling, longitudinal data analysis, epidemiology, survival analysis, clinical trials, non-parametric methods and meta analysis.

In my opinion, understanding how to use the battery of statistical tests that are designed to differentiate between normal and uniform distributions of data and those that are non-normal and non-uniform is essential and very important to avoid the error of using a mathematically inappropriate statistical method, given the normality-uniformity status of the data distribution being analyzed.

It is well understood that our favorite General Linear Models (multivariate multiple linear regression, Pearson correlation, et.) and other Parametric methods, all have model assumptions that are severely violated when used on non-normal or non-uniform distributions. Ignoring this step when making a decision to use a method in ANY branch of statistical science (especially biostatistics and biomedical statistics) has unfortunately resulted in printed retractions when published results are challenged and it is discovered that the results and findings were based upon an invalid mathematical method.

The skill and ability to accurately manage the critical decision to use the appropriate and valid parametric or non-parametric model that is driven by an initial assessment of the distribution of the data (normal versus not normal) is a core skill for ANY mathematician.

An example of a simple test that can instantly determine if a multivariate linear regression (parametric) or BLR (binary logistic regression) (non-parametric) should be used is the Kolmogerov-Smirnov Test of normality and uniformity of a distribution. Among the most useful non-parametric methods in addition to BLR are Chi-Square tests such as (Pearson, Fishers Exact, Likelihood Ratio, etc) and CHAID (Chi-Square-Automatic-Interaction-Detection).

In my 50 years of experience in health care, with the last 30 years devoted to biostatistical nursing research, the majority of the data distributions that are commonly streaming from the health care delivery systems that surround us is decidedly non-normally and non-uniformly distributed. Not all, but most biostatistical research mathematical methods subsequently require a non-parametric approach in order to have any hope of being valid upon editorial peer review and eventually capable of third party replication using similar methods on similar populations at risk.

This is a very important topic in not just biostatistics but all statistics: was the appropriate method used given the nature of the data?

No one ever wants to be involved in having to print a retraction after publication due to discovery that the wrong method has been used that render the findings invalid and requiring retraction. Fortunately that kind of fundamental error can be completely avoided when the guidelines are known and observed. Such errors escape even rigorous peer review as retractions involving methodological violations of the assumptions for mathematical models used to generate the published findings continue to occur. This is rare, but it continues. Researchers need to supply proof that the asserted mathematical model method is valid for the type of data it is being used upon and does not violate assumptions of the model regardless of the model.

Just my opinion, but you asked for it. Thank you for posting this discussion. Valuable and actionable information for all researchers here.

Richard E. Gilder, MS RN-C

Bioinformatics Scientist

The Gilder Company

TTUHSC Adjunct Faculty

[email protected]

www.thegildercompany.com