> If you think y is a function of N predictors, x1, x2, etc. you can try
> to esimate a model
>
> y = f(x1,x2,...,xN) + noise
>
> When will combining models that use subsets of predictors
>
> y = c1*f1() + c2*f2() + noise
>
> work better? I see many emprical papers nowadays using model
> combination nowadays, but I'd like to see a theoretical justification.
> Thanks.

> On Jan 23, 12:26 pm, Beliavsky aol.com> wrote:
>

>> If you think y is a function of N predictors, x1, x2, etc. you can try
>> to esimate a model

>

>> y = f(x1,x2,...,xN) + noise

>

>> When will combining models that use subsets of predictors

>

>> y = c1*f1() + c2*f2() + noise

>

>> work better? I see many emprical papers nowadays using model
>> combination nowadays, but I'd like to see a theoretical justification.
>> Thanks.

>
> Are f, f1, f2 fully specified, or do they have free parameters that
> must be fit to the data? Is one equation a special case of the other?