[R] logLik == -Inf in gls

[Prof Brian Ripley]

Please do not repeatedly post the same thing.  This is the same as

https://stat.ethz.ch/pipermail/r-help/2006-February/086381.html

(except you remembered to sign that one).

You are fitting a weighted not a generalised least squares model: lm() 
will do that.

On Thu, 9 Feb 2006, nhy303 at abdn.ac.uk wrote:

> I am trying to fit a generalised least squares model using gls in the nlme
> package.
>
> The model seems to fit very well when I plot the fitted values against the
> original values, and the model parameters have quite narrow confidence
> intervals (all are significant at p<5%).
>
> The problem is that the log likelihood is always given as -Inf.  This
> doesn't seem to make sense because the model seems to fit my data so well.
> I have checked that the residuals are stationary using an adf test.  I
> can't work out whether
>  - the model really doesn't fit at all
>  - there is something in my data that stops the implementation of logLik
> working correctly (the -Inf value says the calculation hasn't worked)
>
> Possible causes are:
>  - There are lots of NAs in my data (model and response variables)
>  - There is some autocorrelation in the data that is not accounted for by
> the model (most is accounted for).
>
> But, I've tried recreating the problem using a simpler data set, and have
> never found the same problem.

Well, how then do you expect us to be able to recreate it?

As a pure guess, look at your weights.  Are any numob4150 zero?


> The command I use to fit the model is...
>
>
>
> result2 <- gls(lci4150 ~ propCapInStomachs +
>                        temperature +
>                        as.factor(monthNumber) +
>                        lagLci1 +
>                        lagcap1 +
>                        lagcap2,
>              data = monthly,
>              subset = subset1985,
>              na.action = na.approx,
>              weights = varFixed( ~ 1/numob4150)
>             )
>
>
>
> The output I get is...
>
>
>
> Generalized least squares fit by REML
>  Model: lci4150 ~ propCapInStomachs + temperature +
> as.factor(monthNumber) +      lagLci1 + lagcap1 + lagcap2
>  Data: monthly
>  Subset: subset1985
>  AIC BIC logLik
>  Inf Inf   -Inf
>
> Variance function:
> Structure: fixed weights
> Formula: ~1/numob4150
>
> Coefficients:
>                              Value Std.Error   t-value p-value
> (Intercept)              -0.3282412 0.5795665 -0.566356  0.5717
> propCapInStomachs         0.0093283 0.0039863  2.340107  0.0202
> temperature               0.4342514 0.1526104  2.845490  0.0048
> as.factor(monthNumber)2   0.3990717 0.3869991  1.031195  0.3036
> as.factor(monthNumber)3   1.3788334 0.3675690  3.751223  0.0002
> as.factor(monthNumber)4   1.4037195 0.3857764  3.638686  0.0003
> as.factor(monthNumber)5   0.9903316 0.3436177  2.882074  0.0043
> as.factor(monthNumber)6   0.3453741 0.3043698  1.134719  0.2577
> as.factor(monthNumber)7   0.3948442 0.3035142  1.300909  0.1946
> as.factor(monthNumber)8   0.5021812 0.3532413  1.421638  0.1565
> as.factor(monthNumber)9  -0.0794319 0.3598981 -0.220707  0.8255
> as.factor(monthNumber)10  0.3536805 0.3790538  0.933061  0.3518
> as.factor(monthNumber)11  0.7874834 0.3557116  2.213826  0.0278
> as.factor(monthNumber)12  0.1854279 0.3178320  0.583415  0.5602
> lagLci1                   0.5488437 0.0576144  9.526151  0.0000
> lagcap1                   0.0110994 0.0043669  2.541714  0.0117
> lagcap2                  -0.0088080 0.0041099 -2.143127  0.0332
>
>
>
> Does anyone have any suggestions of how I can get a meaningful value for
> logLik?  Or some other way that I can compare models.
>
> Thankyou,
>
> Lillian.
>

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595