Discussion
These results show that the eight depression time series used in this study cannot be distinguished from their linear surrogates using nonlinear and linear in-sample forecasting methods. This result contrasts with the claim in Bonsall et al. ([2012]) that weekly time series from patients with bipolar disorder are described better by nonlinear than linear processes. Could the divergence between the studies be a result of selection: that is, that the Bonsall et al. study tended to select nonlinear series while this study selected linear series? An earlier paper ([Moore et al. 2012]) reported the prediction error for 100 patients from the same monitoring scheme used by both Bonsall et al. and this study. For 100 patients, the interquartile range of prediction errors (SES) is between 2 and 4 in units of the QIDS rating scale. It can be seen that the most of the results in Table 3 lie within this range. Further, the median RMSE forecast value over 100 patients is 2.7 (0.1 normalised) and the median error in Table 3 is 2.65 (0.1). The median forecast errors in Bonsall et al. are reported as 5.7 (0.21) for the stable group and 4.1 (0.15) for the unstable group (Bonsall et al. [2012], Data supplement). We note that the data set used in the present study might not be directly comparable with that used by Bonsall et al.: for example, the time series lengths are unlikely to be the same in each set. However, for the reasons given earlier in this paper, we suggest that high prediction errors in Bonsall et al. arise from the analysis rather than the selection of time series.
The question remains as to what kind of stochastic process best describes the weekly data. The relatively better performance of the linear methods suggests a low-order autoregressive process or a random walk plus noise model ([Chatfield 2002, S2.5.5]), for which simple exponential smoothing is optimal ([Chatfield 2002, S4.3.1]). However, the identification of system dynamics, which might be high dimensional and include unobserved environmental influences would be difficult using the data available.