| Popis: |
Introduction Functional data analysis (FDA) methods have recently been developed to analyse several variables measured repeatedly and concurrently over a domain such as time in a cohort of individuals. However, many FDA methods require data to be measured regularly, with data being collected at the same fixed times for all individuals. Often, with studies in humans, there tends to be missing data. Our aim was to develop a Bayesian model version of a functional concurrent regression (FCR) model in the situation of irregular and missing functional data. Methods A review on modelling sensor data was performed and short comings were identified in current approaches. The Bayesian FCR model was developed for irregular missing functional data. A simulation study was performed to compare the Bayesian FCR model with other methods. A sensitivity analysis was performed to determine the best tuning parameters for the new model. In addition, a simulation study was performed to compare the Bayesian FCR as an imputation model with multiple imputation by chained equations (MICE). The Bayesian FCR model was then applied to real world data. Results The Bayesian FCR model was found to be competitive with other approaches for modelling irregular functional data with missingness. It is robust to the choice of smoothing penalty and prior distributions. Furthermore, it outperforms MICE in imputing both regular and irregular missing functional data. Conclusion The Bayesian FCR model is useful for estimating the parameter function for irregular functional data with missingness. It can be used to impute missing data for use with other methods that require fully observed functional data. With the ever growing use of wearable devices, the new model has excellent potential. |
