Brettschneider JA and Burgess M
Using a frailty model to measure the effect of covariates on the disposition effect
Abstract: Proportional hazards regression models have previously been used to measure the effect of covariates on the disposition effect (DE) in datasets of individual investor trading records. Correlation between trades made by the same investor can be problematic for models of this kind. Past approaches have used a marginal model, where standard errors robust to investor-level correlation are computed after the model has been estimated. This report explores the use of frailty models as an alternative, where the correlation is modelled explicitly through the use of latent variables. Using a dataset of trading records from a large discount brokerage, a frailty model is shown to provide a significantly improved fit relative to the corresponding marginal model. The frailty model is able to significantly estimate a greater number of effects, and it adheres more closely to the important proportional hazards assumption. Results from the frailty model provide some new evidence on experience and learning; the number of trades an investor has made does not have a significant effect on the DE when the investor's self-assessed experience level is included in the model, and the length of time an investor has held an account does not appear to be a reliable measure of experience in this dataset, as those who opened an account most recently exhibit the weakest DE. Graphical checking of the proportional hazards assumption adds nuance to the interpretation of some variables. For example, the weakening of the DE in December is much larger for positions that have already been held for a long period of time, and differences in the DE between positions in small and large cap-size stocks only start to materialize after they have been held for roughly 100 days.