10,259 reputation
31944
bio website
location England
age
visits member for 1 year, 7 months
seen 11 hours ago

Jul
19
reviewed Close Machine learning and statistical modeling
Jul
19
reviewed Close Interpreting lme results in R
Jul
19
reviewed Close how to test for fit of model for zeroinfl() models in R
Jul
19
reviewed Close Box Cox transformed my data, now how to use it in my mixed model?
Jul
19
reviewed Close gravity model of tourism
Jul
19
reviewed Close Imputing missing responses in test exam
Jul
19
reviewed Close Work And Hour Optimization
Jul
17
comment dichotomizing a predictor variable
Can't see why that couldn't happen. I think you'd need to ask another question with a lot more detail about what you're doing for someone to say any more than that.
Jul
17
comment Catagorical variables with very uneven distributions? Removal/modify/leave?
You have to use knowledge to define the bricks - why shouldn't the response be high for odd values of a predictor & low for even values? One handy tip from the book relevant to your question: you can use the coefficient of determination from a regression of the rank of the response on the dummy variables for a categorical predictor to measure how strongly that predictor relates to the response (marginally). Combining some categories for the ones that relate less strongly doesn't bias predictions much.
Jul
17
comment Catagorical variables with very uneven distributions? Removal/modify/leave?
I meant that many books deal with model fitting, interpretation, & diagnostics; but not with how to come up with a model in the first place. Though not dealing with it is better than the rather half-assed section on "Model Selection" or the like that you sometimes see in otherwise very good books.
Jul
17
comment dichotomizing a predictor variable
A lot's going to depend on the true relation between the predictor & response, as well as on the details of what you did. Suppose it's linear: you improve the accuracy of estimates by taking $n$ observations from the tails of the predictor distribution compared to $n$ from all over it; then worsen the accuracy by discarding information when you dichotomize.
Jul
17
comment Catagorical variables with very uneven distributions? Removal/modify/leave?
A lot of applied statistics books start with the supposition that the model is just known, & deal with these sort of issues poorly if at all. I'd highly recommend the two mentioned here.
Jul
17
comment Catagorical variables with very uneven distributions? Removal/modify/leave?
It certainly can be advisable to remove such variables from models, or to combine categories, where it makes substantive sense. (Depends rather on your modelling goals, which you haven't mentioned.) It's important to remember that the variables in a data-set aren't God-given "predictors"- it's up to you to form a useful model from them with one eye on their meaning & the other on the limitations of the data.
Jul
17
revised Catagorical variables with very uneven distributions? Removal/modify/leave?
added tag
Jul
16
reviewed Leave Open G*Power analysis -sample size calculation
Jul
15
comment Does free SPSS 16 actually exist?
There's PSPP.
Jul
15
reviewed No Action Needed Correlation between unordered categorical variable and numerical variables
Jul
15
reviewed Reviewed Fit nonlinear parameter
Jul
15
reviewed Reviewed Calibration in Statistics
Jul
15
comment Calibration in Statistics
What's this to do with calibration?