Interaction coefficient causing tdpreds error

[rohitbatra29]

Hey Charles,

I'm estimating the following CT model which has a time-varying predictor and the interaction term with this predictor:

$$ dy = (\beta_{y} \cdot y + \mu_{y} + \beta_{pred} \cdot Pred + \beta_{y-int} \cdot y \cdot Pred) \cdot dt + \sigma \cdot dW $$

Following is the code for estimating this model with individually-varying terms:

model_3_trial <- ctModel(
  type = "stanct"
  , manifestNames = c("Y")
  , n.manifest = 1
  , n.latent = 1
  , n.TDpred = 1
  , latentNames   = c("y")
  , LAMBDA        = matrix(c(1), nrow = 1, byrow = T)
  , DRIFT         = matrix(c("beta_y + beta_y_int * pred"), nrow = 1, byrow = T)
  , CINT          = matrix(c("mu_y"), nrow = 1, byrow = T)
  , MANIFESTMEANS = matrix(c(0), nrow = 1)
  , MANIFESTVAR   = matrix(c(0), nrow = 1)
  , DIFFUSION     = matrix(c("sigma_y"), nrow = 1, byrow = T)
  , TDpredNames = c("pred")
  , TDPREDEFFECT = matrix(c("beta_pred"))
  , id            = "SID"
  , time          = "time"
  , PARS = c("beta_y", "beta_y_int")
)

model_3_trial$pars$indvarying[which(model_3_trial$pars$param %in% c("beta_y","mu_y", "beta_pred", "beta_y_int"))] <- TRUE

ctModelLatex(model_3_trial)

model3_fit <- ctStanFit(
  data = small_data
  , model_3_trial
)

However, whenever I run the ctStanFit function, it gives me the following error:

Maximum likelihood estimation requested
Error in eval(parse(text = e)) : object 'tdpreds' not found

After running it through the debugger, it looks like the error is coming from these two nested functions: ctStanModelMatrices and ctModelTransformsToNum. Do you have any suggestions on dealing with this error?

[cdriveraus]

Thanks for the report, there was indeed a little bug in the tdpred reference detection, fixed now!

[cdriveraus]

Btw -- interactions between latents and the time dependent predictors directly are a little bit incoherent with the ct framework, as you're essentially saying that the drift value at occasion u, which determines how the process transitions from u-1 to u (however long that may be) is a function of the time dependent predictor at u -- so the time dependent predictor is influencing the past. There can be good reasons to do such things, just warning you that some careful thought is usually needed. It's possible that modelling an extra latent state that is affected by the tdpreds, and interacting with that, would be a better representation.

[rohitbatra29]

Great, thank you!

Yeah, I agree: the interpretation of this interaction term is not straightforward and we're exploring the future direction of treating the TVP as its own dynamic process for better interpretation.