[R] 'breackpoints' (package 'strucchange'): 2 blocking error messages when using for multiple regression model testing

[Michel Lutz]

Achim,

Thank you so much for this prompt answer. Really appreciated !

Anyway, I am still a bit lost... don't you mind if I ask you somme
additional questions?

* *one standard approach is to employ a HACcovariance matrix*
I did many researches but I never found this recommendation. The only paper
I know is Cadsby, Stengos (1985), proposing a transformation to use F-test
when AR(1) errors.
However as I'm not a statistics expert, for sure I missed something
important.
Are you aware of any reference paper recommending this standard approach?

** about the use of the F-test (I won't use gefp, because I have not studied
this method yet)*
Based on the example, I used the below code:
D <- data.frame(CPU=pred.cor2$CPU, PREP=PREP, BRG=BIZ$JOBPREPLOTRULE_BRG,
CLOG=res.WIP, WE=DUMMY)
model.mes <- CPU~PREP+BRG+CLOG+WE

stab.model <- Fstats(model.mes, data = D, from = 0.1,
         vcov = function(x, ...) vcovHC(x, type = "HC", ...))
plot(stab.model)

Here enclosed my result.
I am a bit scared because I am not knwoledgeable about F-Test with HAC (so
far: I need to read), and I've never seen so high F-statistics results. Does
this mean my model is poor?

** about the function breapoints*
I installed strucchange 1.4-5.
I used the below code:
bp.mes <- breakpoints(model.mes, data = D)

Unfortunately, the error occured again:
Erreur dans chol2inv(qr.R(fm$qr)) :
  l'élément (5, 5) est nul, donc l'inverse ne peut être calculé

Why such a chol2inv issue? No missing values in my data, I really don't know
what to do.

* *But the tests need to be adjusted*
Are such adjustements implement in breakpoints? (no mention in the "durab"
example, basic function settings are used).


In advance, thank you very much, and sorry for the disturbance.

Michel



On Fri, Jul 29, 2011 at 10:58 AM, Achim Zeileis <Achim.Zeileis at uibk.ac.at>wrote:

> Michel:
>
>
> I am encountering a blocking issue when using the function 'breackpoints'
>> from package 'strucchange'.
>>
>> *Context:*
>>
>> I use a data frame, 248 observations of 5 variables, no NA.
>> I compute a linear model, as y~x1+...+x4
>> x4 is a dummy variable (0 or 1).
>>
>> I want to check this model for structural changes.
>>
>
> If you want to _test_ for structural changes, then you should use a test,
> i.e., apply sctest() to an Fstats(), efp(), or gefp(). If your errors are
> correlated, one standard approach is to employ a HAC (heteroskedasticity and
> autocorrelation consistent) covariance matrix. There is a worked example
> with Fstats() using a HC matrix in example("durab"). An example with gefp()
> using a HAC matrix is in example("gefp"). See also the vignette("sandwich",
> package = "sandwich").
>
> The breakpoints() function is for _estimating_ (aka dating) structural
> changes, not for testing.
>
> *Process & issues:*
>>
>>
>> *First, I used function Fstats.* It works perfectly. However, this test is
>> not adapted because regression residuals are not independant.
>>
>> That is why *I used 'breackpoints', which works for depedant errors* (Bai,
>> 1997).
>>
>
> Yes, as for coefficient estimates in a regression model, the breakpoint
> estimates are still consistent. But the tests need to be adjusted. Note also
> the in the presence of autocorrelation, the standard information criteria do
> not perform well (Bai & Perron 2003).
>
>  Syntax:
>>
>> struc.test <- breakpoints(y~x1+x2+x3+x3+x4, data=D)
>>
>> *I get an error message:*
>>  Erreur dans chol2inv(qr.R(fm$qr)) :
>>  l'?l?ment (5, 5) est nul, donc l'inverse ne peut ?tre calcul?
>>
>> (sorry for the french version, I don't know how to get the message
>> english translation in R).
>>
>> My first assumption was this has *something to do with the dummy variable,
>> so I skipped it*:
>> struc.test <- breakpoints(y~x1+x2+x3+x3, data=D)
>>
>> *New error message:*
>> Erreur dans if (max(abs((betar - fm$coefficients)/fm$**coefficients)) <
>> tol)
>> check <- FALSE :
>>  valeur manquante l? o? TRUE / FALSE est requis
>>
>>
>> I really can't understand what is going wrong. What 'tol' stands for?
>> Seems
>> it is not a 'breackpoints' attributes.
>>
>
> The breakpoints() function needs to estimate the model on all possible
> subsets to determine the optimal breakpoints. This can be done via
> computation of recursive residuals and "tol" is an argument of the
> recresid() function. However, I recently enhanced the code trying to fix
> exactly this problem. Please try strucchange 1.4-5.
>
> Best,
> Z
>
>  Any help would greatly appreciated.
>>
>> Many thanks in advance,
>>
>> Regards,
>>
>> Michel
>>
>>        [[alternative HTML version deleted]]
>>
>>
>>