Forced expiratory volume in 1 second variability in cystic fibrosis—has the clinical utility been lost in statistical translation?

To the Editor

Accurate prediction of timing and degree of rapid lung function decline are essential to improving cystic fibrosis care. Although forced expiratory volume in 1 second (FEV₁) variability has been examined in cystic fibrosis epidemiologic studies,¹ little has been done to distill its meaning outside of statistical models for utility in clinical trials and practice. In a report published in The Journal, Morgan et al studied the extent to which clinically intuitive calculations of FEV₁ variability are associated with its subsequent decline.² However, simplicity often comes at the expense of scientific or statistical rigor. The authors performed aggregate calculations of longitudinal FEV₁ rather than longitudinal modeling, which would have efficiently accounted for correlated predictors and patient-specific decline. Throughout the report there are repeated assertions of ‘prediction’, but association does not causation make. The reported R² (in each case, <0.4) and small P-values (in a very large dataset) do not make a strong argument for patient-level prediction. The authors do not provide uncertainty intervals for the mean FEV₁ drops, or prediction intervals for an individual projection, which could be quite wide given the variability of the outcome–change in best FEV₁ from one period to the next. A patient with highly erratic lung function in one period will likely experience the same in the next; the authors’ key finding is that highly variable FEV₁ in one period is met with large reductions in the next. Therefore, one could deduce that uncertainty around those big drops is quite large itself–making individual predictions conceivably very poor. Additionally, if this mean–variance relationship exists, a key regression assumption is violated: constant variance. The analysis also includes multiple observation intervals from single patients without accounting for their correlation, thereby violating the independence assumption of simple regression.

As the accompanying editorial suggests,³ more research is needed to address clinical utility of these measures; however, we recommend leveraging existing longitudinal methods for prediction purposes. Coupling statistically efficient and unbiased models with computational feasibility in the clinic could pave the way to timelier, targeted interventions for lung function decline.