How to choose the Best Reference Group? Please Help, Thanks

[Kanyange]

Hi All,

Below is my Logistic Regression output results. I had a question, how to choose a group of reference? Does our choice of group of reference improve the model? For example in my case below, VAR1 has a negative Estimate, which means that customers in group(1,2,5,9,11,.) are less likely to respond. My group reference is the group that is more likely to respond. Is it better to keep the group that is more likely to respond as  a Group reference or shall I swap? Your help will be much appreciated.

Many Thanks

Analysis of Maximum Likelihood Estimates
Parameter		DF	Estimate	Standard	Wald	Pr > ChiSq
				Error	Chi-Square
Intercept		1	-6.994	0.0779	8060.583	<.0001
VAR1	1,2,5,9,11,.	1	-0.0645	0.0204	10.0047	0.0016
VAR2	-2,2,1,8	1	0.2587	0.0222	135.8357	<.0001
VAR3	<7145.5	1	0.2042	0.0295	47.89	<.0001
VAR4	784.5 To 894.5 or Miss	1	-0.2501	0.035	51.186	<.0001
VAR4	<784.5	1	0.3376	0.0278	147.6797	<.0001
VAR5	1,.	1	0.2385	0.0277	74.1201	<.0001
VAR6	<1106.5	1	0.3637	0.0509	51.01	<.0001
VAR7	0,.	1	0.1021	0.0357	8.1737	0.0043
VAR8	11,.	1	-0.3077	0.0431	50.8802	<.0001
VAR8	3,4,7,8	1	0.133	0.0303	19.2755	<.0001
VAR8	9,10,12,13	1	-0.084	0.0353	5.6456	0.0175
VAR9	1	1	-0.3008	0.0875	11.811	0.0006
VAR9	2 or Miss	1	0.1234	0.0414	8.8989	0.0029
VAR9	3	1	0.0998	0.0388	6.6322	0.01
VAR10	0,.	1	0.3695	0.047	61.7212	<.0001

[lryan · Accepted solution]

Good afternoon - in general, here are some things to keep in mind when choosing a reference group:

1.  sample size (you don't want your reference group to have a much smaller sample size in comparison to the non-reference group(s))

2.  the definition of the group: it should be a well-defined group (that is, not something like "other")

3. utility (maybe it is the control group, or the most important for some reason, etc.)

[Kanyange]

Thank you Laura for your response...So the sign of the estimate doesn't matter then....What I mean, if the estime sign is negative, I shouldn't need to change the reference group and make the estimate positive...

[lryan]

You are welcome; I am glad it was helpful.

The sign of the estimate matters in that it tells you how that category changes the probability of success (or the probability of a particular level of the response).  The relationship between the level of the response for a given category is not going to change based on what level you select to be the reference group.  The perspective is all relative.