Table 1.

Heterogeneity in the analytic approaches in CF epidemiologic research of FEV₁.

Role of FEV1
Predictor/Marker	Longitudinal Outcome
Survival	Rate of Decline	Modeling Covariance*
Cox Regression(2) Estimates HR or comparative “risk” of dying over follow-up period between cohorts with higher and lower FEV1 Accommodates missing data due to loss to follow up (right censoring) Logistic Regression Estimates odds of dying between cohorts in the form of an OR (2, 14) Requires complete follow up	Marginal Models Separately model mean FEV1 progression and within-patient association over time Marginal models fit to FEV1 data using GEE(17) Mixed Effects Models Most widely used approach to model FEV1 progression as a function of time and other covariates Intercept and linear terms used to model FEV1 progression(59) Piecewise linear change-point models before and after treatment(23, 60) (24) (25) (26)or infection(17) Quadratic term added to depict FEV1 curvature over time(27) Splines used for nonparametric modeling of FEV1 over time(28) (15) Aggregate measures of rate of decline in FEV1 Year-to-year changes(61) Calculate patient-level slopes and use as response(60)	Unstructured covariance Allows pairwise FEV1 observations to have unique correlation Sometimes difficult to estimate all pairwise correlations due to large number of model parameters Exchangeable Pairwise FEV1 observations have the same correlation, regardless of time between measurements Not realistic for encounter-level data from patient registries Random effects May include center-specific effects to account for how FEV1 progression varies from center to center(27) Accounts for how each patient’s FEV1 progression varies from the population mean FEV1 progression Often incorporated through patient-specific intercepts(22), slopes or quadratic terms(27) Serial Correlation Uses a relatively small number of parameters to model within-subject variation in FEV1 over time Some structures do not require regularly observed FEV1(22)

Abbreviations: generalized estimating equations (GEE); hazard ratio (HR); odds ratio (OR). *Extent of correlation between temporal FEV₁ measurements should be examined a priori, in order to determine an appropriate covariance model.