BookmarkSubscribeRSS Feed
LanMin
Fluorite | Level 6

Hi, all

Y is binary variable .

X1 is my primary interest . First, I did a simple T-test to compare sample observations that have X1 values below sample median vs. above sample median, what their Y value is, I found that below sample median firms are more likely to have Y=1,  e.g. 32% of firms in the below sample median group  have Y=1, In contrast, 27% of firms in the above sample median group have Y=1. That is to say Y and X1 is inversely related.

Next, I use proc logistic procedure

Y=X1 X2 X3 X4

proc logistic data= comb1 descending ;

model Y=X1 X2 X3 X4 ;

run;

now, my coefficient on X1 is positive and significant, that suggests Y and X1 are positively related. I checked VIF for multicollinearity, it does not seem to be an issue.

How should I reconcile this? Would this suggest X1 is positively related to Y?

Thanks !

Lan

11 REPLIES 11
Reeza
Super User

Is SAS modeling the p(y=1) or p(y=0)? not sure how it treats the Y variable here. 

usually i explicitly specify with model y (event='1') but the descending may already take care of that.

MichelleHomes
Meteorite | Level 14

FYI, by default PROC LOGISTIC models the probability of response levels with lower Order Value. So you are Modeling P(Y=0).


http://support.sas.com/documentation/cdl/en/statug/63962/HTML/default/viewer.htm#statug_logistic_sec...



//Contact me to learn how Metacoda software can help keep your SAS platform secure - https://www.metacoda.com
PGStats
Opal | Level 21

Are X2, X3 and X4 coefficients significant as well?

PG
PGStats
Opal | Level 21

If you don't have too many tied values in X1--X4, you might get an idea of what's going on from the table :


proc rank data=comb1 out=testr groups=2;
var x1-x4;
ranks rx1-rx4;
run;

proc tabulate data=testr format=4.0;
class y rx1-rx4;
table rx1*rx2*rx3*rx4, y;
run;

PG

PG
LanMin
Fluorite | Level 6

To PG:

yes, X2, X3 and X4 coefficients ARE significant.

question: what is the purpose of doing proc rank based on your last post? Please enlighten me.

To Reeza,

using descending option, SAS models the p(y=1). It is the same as using (event='1') option.

looking forward to everyone's response to my original post !!!

Lan

PGStats
Opal | Level 21

Proc rank with groups=2 creates two groups for each variable. Group=0 for values less than the median and group=1 for values above. The same idea you had tried, extended to all four variables and combined in a table with proc tabulate.

PG

PG
LanMin
Fluorite | Level 6

Thanks, PG.

By doing what you suggested, would it help explain why the coefficient on X1 is opposite of the expected relation between X1 and Y using T-test, that is, logit regression gives a positive coefficient on X1 , but T-test suggests that high X1 values are correlated with low probability of Y=1 events.

Lan

JohnW_
Calcite | Level 5

The initial check you did was a worthwhile investigation, but it may be the source of your confusion.  Y=1 32% of the time below X1's median and 27% above it.  That is not a huge difference, and depending on your sample size may not be signficant in itself.  Any of X2-X4 could still provide a much stronger signal than X1, causing this reversal.  Can you post the parameter estimates table with significance tests?  Also, try running logistic with X1 as the only explanatory variable...what is the result/significance?

Similarly, does your prior knowledge of the problem give you any insight into the relationship between the two variables?  Would you expect any of X2-X4 to exhibit a stronger relationship with Y than X1?

Vince28_Statcan
Quartz | Level 8

It is also possible, since your X1 appears to be continuous, that all extreme values of X1 have Y=1 whereas the 32% below the median is distributed moreso randomly accross the n/2 cases. Maybe I need to review my log regression litterature but it appears to me as though if all your Y=1 above the median occur with your highest values of X1 with a few extreme values, that it would be reasonable, for the log odds of a shift of 1 unit in X1 to have a positive effect in a model.

If you output the proc logistic data (or maybe that is only availible with proc genmod) but there is an "influence" statistics. Try a model with only X1, sort the resulting dataset by X1 and take a look at the influence statistics column for the records with high values of X1

LanMin
Fluorite | Level 6

Per JohnW’s request, I am posting results below(see attached).

-Lan

JohnW_
Calcite | Level 5

Interesting.  The interaction is new and may help explain what's going on.  On the univariate level, your model predicts as follows:

oneway.jpg

Your full model includes an interaction between X1 and X2 and therefore the impact of X1 depends on the value of X2.  At the point where all other variables are equal to 0, your prediction is as follows (X2 is the left axis and X1 is the right axis, sorry for the small axis titles):

twoway.jpg

Here you can visualize that in certain parts of the predictor space (X2 negative), X1 has the inverse relationship with Y, while in others (X2 positive), X1 has a direct relationship with Y.

I hope this helps.  To PGStats' point above, if you group your data (100 points in each group, or 1000...depending on how many observations), sorted by X1 (and X2), calculate the average response (between 0 and 1), and plot them, it may help visualize the relationship.

sas-innovate-2024.png

Join us for SAS Innovate April 16-19 at the Aria in Las Vegas. Bring the team and save big with our group pricing for a limited time only.

Pre-conference courses and tutorials are filling up fast and are always a sellout. Register today to reserve your seat.

 

Register now!

How to Concatenate Values

Learn how use the CAT functions in SAS to join values from multiple variables into a single value.

Find more tutorials on the SAS Users YouTube channel.

Click image to register for webinarClick image to register for webinar

Classroom Training Available!

Select SAS Training centers are offering in-person courses. View upcoming courses for:

View all other training opportunities.

Discussion stats
  • 11 replies
  • 1587 views
  • 1 like
  • 6 in conversation