Page 196 - Computational Retinal Image Analysis

6 Designing an ophthalmic study 191

• If we believe that there may be a loss to follow up of the patients (e.g. because
the patient decides not to participate, or patient dies) the sample size needs to be
increased to allow for the expected loss during the follow-up. Such an increase
in the sample size must be done at the planning stage of the study.
• Some useful references to understand the topic further and to do the simpler
calculations by hand can be found in [4, 40]. For more complex sample size
calculations, a dedicated sample size software is needed. For even more
complex scenarios, a skilled statistician or mathematician may be able to do
computer simulations.
While we strongly recommend to work with statistician on sample size (or power)
calculations, some recommended software for simpler sample size calculations are:
• Commercial software available e.g. nQuery Advisor or GPower
• Most of statistical packages (such as SPSS, SAS, Minitab, STATA) do offer
some functionality for sample size calculations.
• There are also several online sample size calculators such as PS Power and
Sample Size, http://biostat.mc.vanderbilt.edu/wiki/Main/PowerSampleSize (free
Windows program).

6.2 Words of caution for two eyes: What to do and what not to do?

In clinical studies the focus of interest is very often on the patient e.g. we want to
know if the quality of life of the patient has improved. On the contrary, in ophthalmic
studies very often the focus is on the eye of the patient e.g. we want to know if the
vision improved in a treated eye rather than in patient. If an ophthalmologist wants
to evaluate an effectivity of a treatment of AMD, it may be tempting to use both eyes
of a patient, whenever possible, with an illusion that this will allow to test the treat-
ment on less number of patients, but this is not so obvious. The issue is that the eyes
may be correlated in the primary outcome that is being used for the treatment evalu-
ation (e.g. visual acuity) because they come from the same patient. Many statistical
methods (such as t-test or regression analysis) assume that the units of analysis are
independent (i.e. unrelated) of each other (Table 3).
The fact that patients have two eyes presents challenges in the design, analysis
and interpretation of ophthalmic research [12]. This problem is also known as unit
of analysis issue [41] i.e. the unit of analysis can be the patient or the eye. This issue
still frequently gives rise to statistical errors and it is not uncommon for studies that
are brilliant in terms of methodology and clinical trial design to ignore this issue.
What to do when designing an ophthalmic study and deciding on unit of
analysis? First of all, one needs to be clear what the unit of analysis is, and write
this into the data analysis section of the report or published paper. Then one
needs to decide how the data will be analyzed and write this in the data analysis
section of the report. Bunce et al. (see Fig. 2 in Refs. [12, 41]) provides a brief
overview of when ocular unit of analysis issues may arise and illustrates ways
that these can be dealt with.