Re: GeneralizedLinearModelFit and offsert for poisson
- To: mathgroup at smc.vnet.net
- Subject: [mg113044] Re: GeneralizedLinearModelFit and offsert for poisson
- From: Darren Glosemeyer <darreng at wolfram.com>
- Date: Tue, 12 Oct 2010 04:24:47 -0400 (EDT)
On 10/9/2010 5:33 AM, Ester Vilapriny=F3 wrote:
> Hello, I would like to solve a poisson regression with the
> GeneralizedLinearModelFit function.
> I have a matrix with the following form:
>
> {{x11,x12, x13, y1},
> {x21,x22, x23, y2},
> ...
> }
>
> The GeneralizedLinearModelFit for an exponential family Poisson will retu=
rn
> e^(b0+b1 f1+ b2 f2+ b3 f3)
> Can I fix that b3==1?
>
> In my case y refers to death counts and x3 to population. In other
> programs such as Stata this would be solved by offset(x3). I tried to
> solve this using LinearOffsetFunction but I did not succeed with the
> following steps: 1) created a new matrix for data == {{x11,x12,
> y1},{x21,x22, y2},...}. 2) Create a function that will return the offset
> for each input f[x1_,x2_]:==x3 3)
> GeneralizedLinearModelFit[data,{x1,x2},{x1,x2},ExponentialFaily->"Poisson=
",
> LinearOffsetFunction->(f[#1,#2]&)]. Why it does not work?
>
> Thanks!
>
>
>
The slots (#1, #2, etc.) refer to the predictor variables and
LinearOffsetFunction is a function of the predictors, so you will want
to include x3 as a predictor variable but not a basis function. Then you
can define the offset as #3& to get b3==1.
Here is an example:
In[1]:== data == Flatten[
Table[{x1, x2, x3,
RandomInteger[PoissonDistribution[Exp[.4 + .8 x1 +
.15 x2 + x3]]]},
{x1, .1, .5, .05}, {x2, .1, .5, .05}, {x3, .1,
.5, .05}], 2];
In[2]:== GeneralizedLinearModelFit[data, {x1, x2}, {x1, x2, x3},
ExponentialFamily -> "Poisson", LinearOffsetFunction
-> (#3 &)] //
Normal // InputForm
Out[2]//InputForm== E^(0.541887051809523 + 0.6416041403513674*x1 -
0.14021042180307658*x2 + x3)
Darren Glosemeyer
Wolfram Research