alm.R Line 1137

[Steviey]

R 4.2, greybox 0.6.7.41009

Hello,

While using ro(), I could locate the root of the error below, to alm.R Line 1137

simpleError in if (any(corHigh)) { removexreg <- unique(which(corHigh, arr.ind = TRUE)[, 1]) matrixXreg <- matrixXreg[, -removexreg, drop = FALSE] nVariables <- ncol(matrixXreg) variablesNames <- colnames(matrixXreg) if (!occurrenceModel && !CDF) { warning("Some exogenous variables were perfectly correlated. We've dropped them out.", call. = FALSE) }}: missing value where TRUE/FALSE needed>

'value where TRUE/FALSE needed' suggests that there will be an empty variable proofed- under certain circumstances. However, since this is highly correlated to individual input data- we could fail to reproduce this error with random data.

For now, I avoid this to be a 'tragic code break' by using a tryCatch(), while searching for the right model.

[config-i1]

Thanks for this! Unfortunately, it's difficult to understand why this happens without even knowing how alm() was called. If NAs appear in correlation matrix, then there are some missing observations in the data. I can make this a bit more robust by including use="complete.obs" in cor(), but this won't guarantee the fix of the issue.

[config-i1]

Implemented in greybox v0.6.7.41010. Can you, please, install from github and see whether the issues was fix or not?

[Steviey]

OK Ivan, I will test this. For now, I can mention that all my data is na-imputed by 0, and that it happens while using auto.adam() + ro().

[config-i1]

This is a useful information. Do you use occurrence in this case?

[Steviey]

Here is an example paramter-setting.

model      regressors         ic    distribution         initial 
"FFF"        "select"           "AIC"       "default"   "backcasting" 
lags          bounds      occurrence          
"5"         "usual"          "none"          

[config-i1]

Okay. This looks mysterious. Not sure how to track the error. Let's see if the updated version helps...

[Steviey]

R 4.2, greybox v. 0.6.7.41010

It will take some time...it's running... :-)

[Steviey]

Update: This did not do the trick. What I can say is the following. from grid search: In iteration A it has model='MAN' ic='AICc' result: no error In iteration B it has model='FFF' ic='AIC' result: the mentioned error...

<simpleError in if (any(corHigh)) { removexreg <- unique(which(corHigh, arr.ind = TRUE)[, 1]) matrixXreg <- matrixXreg[, -removexreg, drop = FALSE] nVariables <- ncol(matrixXreg) variablesNames <- colnames(matrixXreg) if (!occurrenceModel && !CDF) { warning("Some exogenous variables were perfectly correlated. We've dropped them out.", call. = FALSE) }}: missing value where TRUE/FALSE needed>

I use an external multicollinearity check and separate correlation check beforehand. The methods: auto.adam() + ro() (with xRegs). The data is 0-imputed/has no NA's. xRegs are dimension reduced by PCA and mostly boxCox-transformed. Activated/deactivated boxCox-transform (external), could play a role... I have to check this further.

[Steviey]

Let me know, if there is something new to test out.

[config-i1]

Hi, I got distracted by my job... :) Can you paste a reproducible example, so that I could investigate and fix?

[Steviey]

At the moment I can't reproduce it anymore Ivan. There are so many changes in your and in my code. Maybe I was wrong on any code part. I let you know, if this happens again.

[config-i1]

I've been fixing so many bugs recently, that I could have fixed it by accident :) Thanks!

[config-i1]

I've been fixing so many bugs recently, that I could have fixed it by accident :) Thanks!