wuzzy wrote:
>
> How do you dichotomize a continuous variable, like age, into 3 decades
> of life.
> And is this a good thing to do in linear regression?
No. Reasons:
(a) by clustering into decades in *any* way you throw away significant
information [I would say roughly 10% in this case, based on squared
variation, but this is a handwave.]
(b) by dichotomizing ("ISUNDER10={0,1}", "ISTEEN={0,1}") rather that
trichotomizing ("DECADE={1,2,3}") you throw away further information
that teenagers lies *between* kids & adults in age.
> Is it just
>
> V1: ages0-10=1 else =0
> V2: ages11-20=1 else =0
> v3: ages21-30=1 else =0
This is definitely wrong as the third variable is redundant (V3 =
1-V1-V2); this will mess up primitive regression software & get you
snickered at by more advanced routines.
> I want to do this because I think this would reduce my variation
> within groups,(measurement error)
You may (mind you, I said "may") get a lower standard error; however,
this is illusory and more than balanced by the reduction in numbers. If
r^2 is big, the conversion of genuine age variation within decades into
apparent "error" [=variation not accounted for by the model] may
outweigh this.
also I think that the effect of the
> different decades is not linear within the groups.
If this is the case you would do better to consider one of the
following:
(a) find a transformation that linearizes the effect
(b) fit a polynomial model or other ad-hoc nonlinear model
(c) use nonparametric techniques.
-Robert Dawson
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
