topic Proc GLMSelect Backward Selection With Many intereaction Terms in Statistical Procedures

Proc GLMSelect Backward Selection With Many intereaction Terms

s_nagelhout — Mon, 26 Oct 2015 21:15:20 GMT

proc glmselect data=data1;

by month;

class day weekday year ;

model percent_change = year| day | weekday dummy1 dummy2 dummy3 /

selection=backward slstay=0.10 details=all stats=all ;

output out=outdata predicted=p_expost residual=r_expost ;

ods output ParameterEstimates=paramest;

run;

ods output off;

I am trying to make a model using the year (2007 through 2015), the day of the month (1 through 28-31), and weekday (1 through7), as well as dummy variables, by month (so effectively 12 seperate models, one for each month). I want to start with all of the interaction terms and then use backward selection to end with only the most significant. However, when I run the model above, I get the note 'The SLSTAY= option has no effect with the selected options.' and all of my standard error and t-value columns for parameter estimates are empty. Is there a maximum number of interaction terms you can enter into the model (I'm at 182 total), or is there a problem with my code?

Thanks in advance for the help!

Re: Proc GLMSelect Backward Selection With Many intereaction Terms

PGStats — Tue, 27 Oct 2015 03:12:23 GMT

If you want the traditional approach for selecting which effect will leave the model based on significance, you must add SELECT=SL to the model statement. By default, SELECT=SBC which is incompatible with SLSTAY=. Check the documentation.

The absence of standard errors is probably due to overparameterization, i.e. some parameters cannot be estimated because there isn't enough data to support all the year*day*weekday interactions. Try grouping some of your class levels to reduce their numbers, i.e. weekday -> (workday, weekend), day -> (start, middle, end) of the month.

Re: Proc GLMSelect Backward Selection With Many intereaction Terms

s_nagelhout — Tue, 27 Oct 2015 16:42:54 GMT

Thanks! That works perfectly when I narrow down the number or variables.