Page 188 - Computational Retinal Image Analysis

4 On choosing the right statistical analysis method 183

the millions of comparisons. However, there are scenarios where we do not need
to adjust for multiple testing. If the study is clearly stated as exploratory and an
acknowledgment made that the research is being conducted in order to generate hy-
pothesis for further testing, corrections for multiple testing may not be needed. It is
best however to avoid multiple testing.

4 On choosing the right statistical analysis method
4.1 The most common statistical methods

Here we aim to give a quick glimpse into the most commonly used methods in oph-
thalmology. This list is not exhaustive, by any means, and it is based on our research
practice. Our introduction is very short, however, a keen medical reader, who is new
to statistics, is welcome to read a gentle intuitive introduction to the statistical meth-
ods by Altman [4]. Those would like a more mathematical overview and understand-
ing of statistics are advised to consult a monograph by Wasserman [3] which was
written as a textbook for statisticians or computer scientists. For a good ophthalmic
introduction to diagnostic studies, sensitivity and specificity we recommend [15]. A
discussion of multivariate versus multivariable statistical methods is in [22].
It is important to note that statistics is an evolving scientific discipline with new
methods being developed to better analyze real world data of increasing complexity.
It is impossible to give a guide to all the statistical methods. Instead, we attempt to
give a guide to simpler statistical methods used in ophthalmology.
We created a table (Table 3) of the most used methods and in doing so we divided
them by their goal, specific objective, design of data collection, and type of data.
Some more complex modern methods that we do not list in the table are: longitudinal
modeling, joint-modeling and predictive modeling.

4.2 How to decide what method to use?

“What statistical method should I use?” This is a question that statisticians are often
asked when approached by researchers. The answer is often not straightforward. It
depends on several factors, some factors are common to all studies some are specific
to a type of study (see Table 3). It is fair to say that the answer to this question is only
found after a thorough discussion between the statistician and the researcher.
The main factors that determine the data analysis method are:
• Goal of the analysis together with specific objectives e.g. Is the goal to describe
the sample?, Is the goal to test a hypothesis?, Is the goal to derive and evaluate a
diagnostic method?
• Type of data being collected e.g. Are the collected data continuous skewed or
normally distributed, are they categorical?
• Design of the study e.g. Is the study a prospective study or randomized study?
Are there repeated measurements taken on eyes? Is it a longitudinal study?