Impact of missing data on analysis of postoperative cognitive decline (POCD)

摘要

Background: There are a variety of techniques to handle missing data, such as removing observations with missing data from the analyses or estimating the missing values using various imputation algorithms. Dropping subjects from standard regression models and analyzing only completers, however, may bias results from the true value of reality. Likewise, 'last-observation-carried-forward' may not be an appropriate technique for studies measuring a particular variable over time. Methods: This dataset was part of a larger prospective cohort study that examined postoperative cognitive decline (POCD) after surgery in older adults. Data collectors had provided the reasons for data being missing using adjectives including 'confused', 'incapable', 'stuporous', 'comatose', and 'intubated'. Data having these qualitative notations were re-coded as 'incapable' and trial scores of zero were recorded. This value of '0' indicated that the patient was cognitively incapable of performing the neuropsychological test. Results: Missing data varied by cognitive test and postoperative day. Re-coding word list scores from missing to zero when a patient was too cognitively impaired to complete the tests improved sample size by 13.5% of postoperative day (POD) 1 and 12.8% on POD 2. Recoding missing data to zero for the digit symbol test resulted in 29.3% larger sample size on POD 1 and 22.7% on POD 2. Verbal fluency gained 15.7% sample size with re-coding for POD 1 and 13.7% for POD 2. Re-coding of each cognitive test reduced missing data sample size to 20-32% in all cognitive tests for each day. Discussion: Our data suggest that using a scoring system that enters a value of '0' when patients are unable to perform cognitive testing did significantly increase the number of patients that met the diagnosis of postoperative cognitive decline using the criteria that were determined a priori and may lessen chances of type II error (failure to detect a difference).