10-years changes in lung function of cystic fibrosis patients in Europe: different statistical methods at work

Cystic fibrosis (CF) is the most common severe autosomal recessive disease in Europe, with pulmonary insufficiency as the main cause of death. For prognosis, forced expiratory volume in 1 second percent of predicted (FEV1pp), is regarded as the best generally available measure for assessing CF lung disease. Since FEV1pp has a slightly asymmetric distribution, it is often summarized using median and quartiles. However, when fitting regression models, the results are usually provided in terms of means. The aim of the current study is to explore changes in FEV1pp during the last decade, comparing results obtained with different statistical regression methods including random effects. To estimate the difference in FEV1pp values over 2011 and 2021, data of 18756 people with CF, homozygote for F508del mutation and included in the European Cystic Fibrosis Society Patient Registry, are used. Three regression models including a random effect for patients and with FEV1pp as response variable are fitted using R software: the classical generalized estimating equations (GEE) model with Gaussian family, a linear quantile mixed model (LQMM) [1], a Generalized Additive Models for Location, Scale and Shape (GAMLSS) [2] with Normal family distribution. Two different setting are explored: in the first one the year of follow-up is included as a continuous variable, in the second one it is included using dummy variables. The results of the different models are comparable in terms of coefficient estimates. GAMLSS provides the narrower confidence interval then GEE and LQMM when year is included as a continuous variable, LQMM gives the narrower CI then GEE and GAMLSS when year is included as dummy variables. The main problem in fitting models in R software on our big dataset, is the long computational time. To obtain coefficient estimates and standard errors: 50 minutes for GAMLSS and almost 6 hours for LQMM. In conclusion, these models need to be additionally compared in detail for diagnostic measures., Further research is needed to fulfill the unmet need of providing robust regression coefficient estimates on mixed effects models on big datasets, also simulation studies mimic real world practice are necessary. [1] M. Geraci, Linear quantile mixed models: The lqmm package for Laplace quantile regression. Journal of Statistical Software, 57(13), 2014,1-29. [2] D.M. Stasinopoulos, R.A. Rigby, Generalized additive models for location scale and shape (GAMLSS) in R. Journal of Statistical Software, 23(7), 2007, 1-46.