Abstract:
|
Missing data is a common analytical problem, particularly in longitudinal studies. In studies of chronic disease, differential drop-out may occur if patients cease participation due to a lack of improvement, or conversely, if the improvement is so large they no longer feel the need to continue treatment. We conducted a simulation study to investigate potential biases that may arise when using a random-effects mixed model. Patients' slopes (rates of decline) were sampled from a normal distribution, and the number of observations per patient was calculated such that there was a relationship between the slope and the number of observations. For validation and comparison, datasets with no missing data and data missing at random were also generated. A mixed model was used to compute an estimate of the slope which was compared to the "true" slope used to generate the data. For correlations between the number of observations and the slope between 0.2 and 0.3, a 0.05 level test of the null hypothesis was inappropriately rejected 50% of the time. When there was no relationship between the number of observations and the slope, no Type I error inflation was noted.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.