Many studies, even studies published in impact journals include the two eyes of the same patient. This is not a good practice, since these observations are not independent observations, because both eyes are usually similar. Including the two eyes of the same patient increases the size of the sample, and narrows confidence intervals, making easier the task of achieving statistical significance. However I think that some methods can correct to a certain extent the bias generated by including both eyes. Anyone knows where can I find information about how to address inter-ocular correlation?
Julio: There are a number of references that directly address your question. Some dealing with the problem with setting up the initial studies, e.g.:
- Ray, W. A., & O'Day, D. M. (1985). Statistical analysis of multi-eye data in ophthalmic research. Investigative ophthalmology & visual science, 26(8), 1186-1188.
- Murdoch, I. E., Morris, S. S., & Cousens, S. N. (1998). People and eyes: statistical approaches in ophthalmology. British Journal of Ophthalmology, 82(8), 971-973.
These two are dedicated articles to statistical adjustment techniques and when to apply them:
- Rosner, B. (1982). Statistical methods in ophthalmology: an adjustment for the intraclass correlation between eyes. Biometrics, 105-114.
- Glynn, Robert J., and Bernard Rosner. "Accounting for the correlation between fellow eyes in regression analysis." Archives of ophthalmology 110.3 (1992): 381-387.
One more recent review that includes many references (Table 3 and the flowchart found in Figure 1) is particularly useful:
- Armstrong, R. A. (2013). Statistical guidelines for the analysis of data obtained from one or both eyes. Ophthalmic and Physiological Optics, 33(1), 7-14.
I used the term inter-ocular correlation to perform the search and I only found two of the references you propose.
Thank you so much for such valuable information.
Julio.
Julio,
About 60 years ago, mixed-effects models were developed (e.g. see Wikipedia). That approach allows the inclusion of data from both eyes of a subject with the relationship dealt with appropriately when the "clustering" is declared (e.g. error terms, DoF). Note that there is value in having data from two eyes, in that the high correlation (in healthy eyes) provides a better estimate than data from only one eye. Such models can also account for other relationships (clusters) such as between siblings and repeated measures over space or time. Most statistical packages include mixed-effects models (there are a variety of other names for the same thing).
Russell
Thanks` very much for your orientation. It is not an easy topic. I suppose SPSS includes mixed-effects models.
Best regards from sunny Spain.
Julio.
Julio,
I do not use SPSS (too expensive, and instead use Stata, which has a much easier syntax). With a quick web search using "SPSS mixed model" I found links to material from SPSS and from the UCLA IDRE (which is a really great resource). Note that mixed models can be applied to many research questions and expt designs. For example, there are versions for linear (regression), logistic and poisson models. As an example, it seems that the command MIXED is used for linear mixed models.
Russell
- Badges
- Science method