tho for logit or probit, you could do the things I described in class,
putting the probs in bins or doing an ROC curve. whatever model you use,
you should include some demonstration that the model fits or fix the model
so we know that it does.
Gary
On Sat, 10 May 2003, Kosuke Imai wrote:
In general, you don't want to over-predict or
under-predict, but for many
models, the distribution of residuals is not specified. Consider the logit
model. The model doesn't tell you the distribution of residuals. Even in
the case of probit model, since we don't observe the latent outcome
variable, the difference between the observed binary outcome variable and
the linear predictor does not correspond to the residuals.
Kosuke
On Sat, 10 May 2003, Traci Burch wrote:
For any type of model you estimate, is it true
that ideally, plotting
the residuals shows n points scattered about the y=0 line with no
apparent pattern or is this only true for linear regression?
_______________________________________________
gov2001-l mailing list
gov2001-l(a)fas.harvard.edu
http://www.fas.harvard.edu/mailman/listinfo/gov2001-l