How to code a fixed additive residual error without freeze (and keep FOCE-ELS available)

Phoenix Certara Forums Pirana and RsNLME
1

If you need one component of a combined (proportional + additive) residual error to be a fixed constant — the natural instinct is error(CEpsAdd(freeze) = value). Worth knowing before you do that: freezing an error()/sigma term has a side effect beyond just fixing that value — it restricts engineParams(method=...) to "Laplacian", "QRPEM", "IT2S-EM", or "Naive-Pooled", and if you don’t specify a method explicitly, the default silently changes from FOCE-ELS to LAPLACIAN.

Here’s a way to get a fixed additive SD without touching freeze at all, so FOCE-ELS stays available.

The standard combined-error pattern has the estimated error() term represent the additive SD, with the proportional coefficient as a separate fixef():

error(CEpsAdd = 1)
fixef(tvPropErr = c(0, 0.1, ))
observe(CObs = C + CEpsAdd * sqrt(1 + C^2 * (tvPropErr/sigma())^2))

(A fixed proportional coefficient — fixef(tvPropErr(freeze) = ...) — is fine here and does not trigger the restriction; it’s specifically freezing something in the error()/sigma block that does.)

If the additive component is the one that needs to be fixed, invert the roles and let the estimated error() term’s own SD (sigma()) directly represent the proportional coefficient, and declare the fixed additive SD as a plain double()or real() /sequence{} constant:

double(CAddFix)
sequence{
  CAddFix = 0.5   // the fixed additive SD - a true constant, never estimated
}

error(CEpsProp = 0.1)   // estimated - its own SD IS the proportional coefficient

observe(CObs = C + CEpsProp * sqrt(C^2 + (CAddFix/sigma())^2))   # <units>

Why this works: Var(CObs) = sigma()² · [C² + (CAddFix/sigma())²] = sigma()²·C² + CAddFix². Since sigma() estimates to the proportional coefficient, this is exactly PROP²·C² + ADD_fixed² — the usual combined-error variance, just with the fixed and estimated roles swapped relative to the familiar template.

2 Likes