advertisement
Editors Selection IGR 15-2
Structure and Function: Methodological caveats and pitfalls
Comment by Shaban Demirel on:
53678 Choice of statistical method influences apparent association between structure and function in glaucoma, Marín-Franch I; Malik R; Crabb DP et al., Investigative Ophthalmology and Visual Science, 2013; 54: 4189-4196
Marín-Franch et al. describe pitfalls encountered when exploring the association between structural and functional test results in glaucoma. In this paper, the authors explore two datasets from the literature, examining the results and conclusions that were made in the papers that originally presented them. They recreate the original analyses and then perform other equally valid, or invalid, analyses. For example, when they simply swapped the independent and dependent variables in a linear regression model (structure is independent vs. function is independent), very different coefficients arose from the regression. A similar outcome was seen when using segmented regression, where simply swapping independent and dependent variables shifted the breakpoints between linear segments and significantly altered the slopes of the segments themselves. Something as simple as designating a dependent and independent variable can drastically alter conclusions made about the association between structure and function in eyes with glaucoma. These warnings apply whenever examining the association between any two variables if both contain noise and there is no obvious order of dependence. The authors remind us that ordinary linear regression assumes that there is an independent variable that is a known quantity without noise. The authors also explain that part of the problem is that structural and functional measures correlate poorly. If the association between them was stronger, there would be less risk of inappropriate conclusions being reached even if the assumptions underlying linear regression were not met. Another problem that the authors hint at is the fact that performing these types of analyses 'correctly' can be difficult.
In science we strive to make progress and perfection should not be the enemy of the good. But data analysts, authors, editors and ultimately readers must remember to revisit the assumptions that underlie the analysis tools being used. More importantly, when assumptions are violated, we must be mindful of the potential impact on conclusions. Marín-Franch et al. demonstrate this quite clearly indeed.
Comments
The comment section on the IGR website is restricted to WGA#One members only. Please log-in through your WGA#One account to continue.Log-in through WGA#One
