I have sales data for different locations. My outcome variable is making a sale (1,0) and my IVs are about 200 different locations, handling time etc.. I ran a logistic regression using the class statement for locations in SAS. Some of the locations appear to be significant others don’t. My problem is that with the class statement there is always a reference group. So any odd ratio for a particular location provides information how the odds for this location to make a sale compare to the reference group. Can I avoid the comparison to the reference group, by creating dummies for only the locations that appeared to be significant when using the class statement? I want to provide a more general overview of how a location performs instead of comparing it to a specific location.
5
-
$\begingroup$ Can you provide more detail about your data? For example, besides locations, what other predictors do you have for sales? Also, what is your sample size, and how much sales data do you have per location? Is your goal to identify best locations, or is it to determine if location makes a difference in sales? $\endgroup$– BryanCommented Mar 9, 2018 at 20:24
-
$\begingroup$ This sounds like a textbook case to use multilevel modeling/mixed effects models. $\endgroup$– AlexisCommented Mar 10, 2018 at 15:37
-
$\begingroup$ @Alexis Can you please provide more info why you would suggest multilevel modeling and what this is from a low level perspecitve. $\endgroup$– breeCommented Mar 11, 2018 at 16:50
-
$\begingroup$ en.wikipedia.org/wiki/Multilevel_model $\endgroup$– AlexisCommented Mar 11, 2018 at 18:48
-
$\begingroup$ Duncan, C., Jones, K., and Moon, G. (1998). Context, composition and heterogeneity: Using multilevel models in health research. Social Science & Medicine, 46(1):97–117. $\endgroup$– AlexisCommented Mar 11, 2018 at 18:49
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
4
difficult to understand what you're asking: how can you assess performance without using some benchmark (reference group)? perhaps you just want to estimate proportions, as follows:
ods output lsmeans=pred1;
proc genmod data=... plots=none descending ;
class location;
model sale=location/link=logit dist=binomial type3;
lsmeans location / ilink cl;
run;
quit;
-
$\begingroup$ The issue is that there are many locations and some of them don't seem to be significantly predicting the outcome. I would like to exclude the locations that don't matter. If I do so, I just want to make sure that it is still informative and I can say something like: if the agent is from Berlin, for the agent to make the sale the odds are 2 times more likely as if the agent is not from Berlin. I don't want to compare it to a specific location, because it is hard to pick a reference group. I am missing information about the locations, to make the choice of a reference location. $\endgroup$– breeCommented Mar 11, 2018 at 16:54
-
$\begingroup$ it seems to me you're trying to impose odds ratios and logistic regression onto your data, and talking about 'significantly predicting' etc. Make your analysis suit your data, rather than vice versa. Eg for each location you have what? people making sales? ie what does the 0/1 outcome correspond to? sales people? it's not clear that you're handling the data properly $\endgroup$ Commented Mar 12, 2018 at 5:58
-
$\begingroup$ yes, I am running a logistic regression and I tested the correlation of the variables prior to running it. I have about 20 locations and 200 add ons. If you can suggest another pre-selection method for variables let me know. And the outcome variable is: agent making a sale vs. not making a sale. $\endgroup$– breeCommented Mar 12, 2018 at 14:23
-
$\begingroup$ i would begin by estimating the proportion (with confidence interval) for each location (using the sas code i give above) and then plot these proportions in descending sequence (with location on the x-axis). then you will see immediately which locations are performing best, and you will also see very large confidence intervals for some locations i imagine. you may think about discarding sites that are uninformative. That could be a starting point for your analysis $\endgroup$ Commented Mar 12, 2018 at 20:25