Hi @OmidS,

In the Methods section, it says "... identified ... nx as the value that reaches within 1 s.e.m. of the best possible neural self-prediction accuracy among all considered latent state dimensions." Say we consider 1 to 100 as potential neural latent state dimension, could you confirm if the following steps are correct or not?

1. For each potential nx value, call the PSID function with n1 = 0 to fit the model.
2. For each potential nx value, get Ypred (using the PSIDPredict function), and then evaluate the prediction (using the evalPrediction function) to have a scalar performance measurement, say CC, per neural dimension. Then calculate the mean CC value across neural dimensions.
3. Among the 100 mean CC values from all nx values under consideration, find the maximum (ccMax) and the standard deviation (ccSD). Then find the smallest nx value for which the mean CC value is >=(ccMax - ccSD). This nx value is decided to be the total number of neural latent state dimension.

Btw, were the above-mentioned procedure and functions used to assess the self-prediction of behavioral signals, too? If yes, were the behavioral data used as the y input to those functions? What should I do with the z input then?

I hope to try the PSID method on my own data, so I'd like to have the details figured out. I'd appreciate it if you could help me with this.

Regards,
Hongwei