R help - Aug 2015 - Non linear regression - Von Bertalanffy Growth Function - "singular gradient matrix at initial parameter estimates"

Dear all,

I am trying to estimate VBGF parameters K and Linf using non linear 
regression and nls(). First I used a classic approach where I estimate 
both parameters together as below with "alkdyr" being a subset per
year
of my age-length-key database and running in a loop.

vbgf.par <- nls(Lgtcm ~  Linf *(1 - exp(-K * (Age - tzero))), start = 
c(K= 0.07, Linf = 177.1), data=alkdyr)

I obtain an estimation of both parameters that are strongly correlated. 
Indeed after plotting Linf ~ K and fitting a linear regression I obtain 
a function (Linf = a + b*K) with R2= 0.8 and a = 215, b = -763.

In this context, to take into account explicitly correlation between 
parameters, I decided to fit a new non linear regression derivate from 
VBGF but where Linf is expressed depending on K (I am most interested in 
K). To do so, I tried this model:
vbgf.par <- nls(Lgtcm ~  (a + (b*k)) *(1 - exp(-k * (Age - tzero))), 
start = c(k= 0.07, a= 215, b=-763), data=alkdyr)

Unfortunately at this point I cannot go further as I get the error 
message "singular gradient matrix at initial parameter estimates".

I tried to use alg= plinear (which I am not sure I understand properly 
yet). If I give a starting value for a and b only, I have an error 
message stating "step factor below minFactor" (even when minFactor is 
set to 100000000000).

Any help will be more than welcome as this is quite urgent....

Best,

Xochitl C.




-- 

<>< <>< <>< <><

Xochitl CORMON
+33 (0)3 21 99 56 84

Doctorante en ?cologie marine et science halieutique
PhD student in marine ecology and fishery science

<>< <>< <>< <><

IFREMER
Centre Manche Mer du Nord
150 quai Gambetta
62200 Boulogne-sur-Mer

<>< <>< <>< <><

These appear to be primarily statistics/nonlinear optimization issues
that are off topic here, which is about R programming. Post on a
statistics list like stats.stackexchange.com instead.


Cheers,
Bert




Bert Gunter

"Data is not information. Information is not knowledge. And knowledge
is certainly not wisdom."
   -- Clifford Stoll


On Tue, Aug 18, 2015 at 9:08 AM, Xochitl CORMON
<Xochitl.Cormon at ifremer.fr> wrote:
> Dear all,
>
> I am trying to estimate VBGF parameters K and Linf using non linear
> regression and nls(). First I used a classic approach where I estimate both
> parameters together as below with "alkdyr" being a subset per
year of my
> age-length-key database and running in a loop.
>
> vbgf.par <- nls(Lgtcm ~  Linf *(1 - exp(-K * (Age - tzero))), start =
c(K> 0.07, Linf = 177.1), data=alkdyr)
>
> I obtain an estimation of both parameters that are strongly correlated.
> Indeed after plotting Linf ~ K and fitting a linear regression I obtain a
> function (Linf = a + b*K) with R2= 0.8 and a = 215, b = -763.
>
> In this context, to take into account explicitly correlation between
> parameters, I decided to fit a new non linear regression derivate from VBGF
> but where Linf is expressed depending on K (I am most interested in K). To
> do so, I tried this model:
> vbgf.par <- nls(Lgtcm ~  (a + (b*k)) *(1 - exp(-k * (Age - tzero))),
start > c(k= 0.07, a= 215, b=-763), data=alkdyr)
>
> Unfortunately at this point I cannot go further as I get the error message
> "singular gradient matrix at initial parameter estimates".
>
> I tried to use alg= plinear (which I am not sure I understand properly
yet).
> If I give a starting value for a and b only, I have an error message
stating
> "step factor below minFactor" (even when minFactor is set to
100000000000).
>
> Any help will be more than welcome as this is quite urgent....
>
> Best,
>
> Xochitl C.
>
>
>
>
> --
>
> <>< <>< <>< <><
>
> Xochitl CORMON
> +33 (0)3 21 99 56 84
>
> Doctorante en ?cologie marine et science halieutique
> PhD student in marine ecology and fishery science
>
> <>< <>< <>< <><
>
> IFREMER
> Centre Manche Mer du Nord
> 150 quai Gambetta
> 62200 Boulogne-sur-Mer
>
> <>< <>< <>< <><
>
> ______________________________________________
> R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.

Packages nlmrt or minpack.lm use a Marquardt method. minpack.lm won't 
proceed if the Jacobian singularity is at the starting point as far as 
I'm aware, but nlxb in nlmrt can sometimes get going. It has a policy 
that is aggressive in trying to improve the sum of squares, so will use 
more effort than nls when both work.

JN

On 15-08-18 12:08 PM, Xochitl CORMON wrote:
> Dear all,
>
> I am trying to estimate VBGF parameters K and Linf using non linear
> regression and nls(). First I used a classic approach where I estimate
> both parameters together as below with "alkdyr" being a subset
per year
> of my age-length-key database and running in a loop.
>
> vbgf.par <- nls(Lgtcm ~  Linf *(1 - exp(-K * (Age - tzero))), start >
c(K= 0.07, Linf = 177.1), data=alkdyr)
>
> I obtain an estimation of both parameters that are strongly correlated.
> Indeed after plotting Linf ~ K and fitting a linear regression I obtain
> a function (Linf = a + b*K) with R2= 0.8 and a = 215, b = -763.
>
> In this context, to take into account explicitly correlation between
> parameters, I decided to fit a new non linear regression derivate from
> VBGF but where Linf is expressed depending on K (I am most interested in
> K). To do so, I tried this model:
> vbgf.par <- nls(Lgtcm ~  (a + (b*k)) *(1 - exp(-k * (Age - tzero))),
> start = c(k= 0.07, a= 215, b=-763), data=alkdyr)
>
> Unfortunately at this point I cannot go further as I get the error
> message "singular gradient matrix at initial parameter
estimates".
>
> I tried to use alg= plinear (which I am not sure I understand properly
> yet). If I give a starting value for a and b only, I have an error
> message stating "step factor below minFactor" (even when
minFactor is
> set to 100000000000).
>
> Any help will be more than welcome as this is quite urgent....
>
> Best,
>
> Xochitl C.
>
>
>
>

Thank you for the tip. Indeed, nlxb in nlmrt works and results are not 
crazy.

I would like however to assess goodness-of-fit (gof) and ultimately to 
compare it with gof from linear regression (fitted with same variables).

Before I used AICc to compare the nls() and lm() fit, however I get now 
an error message concerning the method loglike and its non compatibility 
with nlmrt class object. I guess it is because we use now Marquardt 
method to minimise sum-of square instead of Gauss-Newton? I am right? Or 
this is just an incompatibility coming between AICc function and nlmrt 
objects? Is there an R function to do that?

Best,

Xochitl C.


<>< <>< <>< <><

Xochitl CORMON
+33 (0)3 21 99 56 84

Doctorante en ?cologie marine et science halieutique
PhD student in marine ecology and fishery science

<>< <>< <>< <><

IFREMER
Centre Manche Mer du Nord
150 quai Gambetta
62200 Boulogne-sur-Mer

<>< <>< <>< <><



Le 19/08/2015 15:11, ProfJCNash a ?crit :
> Packages nlmrt or minpack.lm use a Marquardt method. minpack.lm won't
> proceed if the Jacobian singularity is at the starting point as far as
> I'm aware, but nlxb in nlmrt can sometimes get going. It has a policy
> that is aggressive in trying to improve the sum of squares, so will use
> more effort than nls when both work.
>
> JN
>
> On 15-08-18 12:08 PM, Xochitl CORMON wrote:
>> Dear all,
>>
>> I am trying to estimate VBGF parameters K and Linf using non linear
>> regression and nls(). First I used a classic approach where I estimate
>> both parameters together as below with "alkdyr" being a
subset per year
>> of my age-length-key database and running in a loop.
>>
>> vbgf.par <- nls(Lgtcm ~  Linf *(1 - exp(-K * (Age - tzero))), start
>> c(K= 0.07, Linf = 177.1), data=alkdyr)
>>
>> I obtain an estimation of both parameters that are strongly correlated.
>> Indeed after plotting Linf ~ K and fitting a linear regression I obtain
>> a function (Linf = a + b*K) with R2= 0.8 and a = 215, b = -763.
>>
>> In this context, to take into account explicitly correlation between
>> parameters, I decided to fit a new non linear regression derivate from
>> VBGF but where Linf is expressed depending on K (I am most interested
in
>> K). To do so, I tried this model:
>> vbgf.par <- nls(Lgtcm ~  (a + (b*k)) *(1 - exp(-k * (Age - tzero))),
>> start = c(k= 0.07, a= 215, b=-763), data=alkdyr)
>>
>> Unfortunately at this point I cannot go further as I get the error
>> message "singular gradient matrix at initial parameter
estimates".
>>
>> I tried to use alg= plinear (which I am not sure I understand properly
>> yet). If I give a starting value for a and b only, I have an error
>> message stating "step factor below minFactor" (even when
minFactor is
>> set to 100000000000).
>>
>> Any help will be more than welcome as this is quite urgent....
>>
>> Best,
>>
>> Xochitl C.
>>
>>
>>
>>