<!doctype html public "-//w3c//dtd html 4.0 transitional//en">

<html>

Hello,

<p>I am try to perform a modeling which is relevant in a strongly heteroscedastic

context.

<br>So I perform a dual modeling (modeling of both mean and variance of

a response) in using the following loop:

<p>jointmod &lt;- function(formula, data, itercrit=10,devcrit=0.0001)

<br>{

<br>#

<br># Init step

<br>#

<br>init &lt;- glm(formula=formula,family=gaussian, data=data)

<br>response &lt;- resid(init,type="response") + predict(init,type="response")

<br>mu &lt;- predict(init,type="response")

<br>iter &lt;- 1

<br>dev &lt;- init$deviance

<br>if (dev != 0) vardev &lt;- 1

<br>else stop(message="Null deviance in initialisation step")

<br>endformula_ strsplit(formula,"~")[[3]]

<br>formulavar_paste("d","~")

<br>formulavar_as.formula(paste(formulavar,endformula))

<br>#

<br># Modeling loop

<br>#

<br>while ((iter &lt;= itercrit)| (vardev > devcrit)) {

<br>d_(response - mu)^2

<br>modvar_glm(formula=formulavar,family=Gamma(link=log),data=data)

<br>sig &lt;- predict(modvar,type="response")

<br>weights &lt;- 1/sig

<br>modesp &lt;- glm(formula=formula, family=gaussian,data=data,weights=weights)

<br>mu &lt;- predict(modesp,type="response")

<br>devold &lt;- dev

<br>dev &lt;- modesp$deviance

<br>vardev &lt;- (devold -dev)/devold

<br>iter &lt;- iter + 1

<br>}

<p>list(modesp=modesp, modvar=modvar, iter=iter, vardev=vardev)

<br>}

<p>This program evaluation always stops when evaluating the "modesp" line

:

<br>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; modesp &lt;- glm(formula=formula,

family=gaussian,data=data,weights=weights)

<br>with the following error:

<br>error in model.frame(formula,rownames,variables,varnames,extras,extranames,

:

<br>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; invalid

variable type.

<p>The same program without specifying weights runs.

<br>I don't understand this behavior since when I execute the glm command

with weights out of the program, it runs.

<p>I run R1.3.0 under Windows.

<p>If someone can help me ...

<p>Isabelle Zabalza-Mezghani

<br>&nbsp;

<br>&nbsp;

<br>&nbsp;

<br>&nbsp;

<br>&nbsp;

<pre>--&nbsp;

Isabelle Zabalza-Mezghani&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; Tel : 01 47 52 61 99

Institut Fran&ccedil;ais du P&eacute;trole&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; E-mail : isabelle.zabalza-mezghani@ifp.fr

1-4 Av. Bois Preau - Bat Lauriers

92852 Rueil Malmaison Cedex, France</pre>

&nbsp;</html>