turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- SAS Programming
- /
- SAS Procedures
- /
- Interpreting Odds Ratios- and Omitted Variables

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

03-17-2009 02:55 PM

From SAS I get the following output: ( also attached) The dependent variable is binary, Y = 1 for pass, Y = 0 for fail.

Odds Ratio Estimates

Point 95% Wald

Effect Estimate Confidence Limits

SEX M vs F 1.162 1.004 1.345

RACE 2 vs 6 1.521 1.143 2.024

My interpretation is that with an odds ratio > 1 ( the point estimate) that implies that males are more likely to pass than females, actually between 1.004 and 1.345 times as 'likely' with 95% confidence. ( I'm not sure what I mean by 'likely' I'm just repeating what I've observed in literature)

The same goes for race, RACE 2 is more likely to pass than RACE 6 based on the odds ratios.

I know these results are significant because 1) the Wald Chi-square in the Type 3 analysis of effects ( not given above) is significant for RACE and SEX, implying differences exist and 2) the 95% wald confidence intervals above do not contian '1'. ( If this is correct I don't understand why )

If I'm correct about all of the above, I still have a problem. Just by observing the data, and theory, Females pass at greater percentages than males and RACE 6 passes with greater percentages than RACE 2. My results contradict theory and observation of raw data on both accounts.

I believe in linear regression, when you get the wrong sign on a coefficient it could be caused by an omitted variable. ( EX: if the sign is unexpectedly negative for Beta on X1 then you could be omitting a variable say X2 that is positively correlated to X1 but negatively correlated with Y in a model Y = B1X1 + B2X2) You can fix the problem by discovering the omitted variable and inserting it into the model.

Could an omitted variable be responsible for my odds ratio results? I'm not sure if it works like this. Can someone suggest what I'm doing /interpreting incorrectly?

Odds Ratio Estimates

Point 95% Wald

Effect Estimate Confidence Limits

SEX M vs F 1.162 1.004 1.345

RACE 2 vs 6 1.521 1.143 2.024

My interpretation is that with an odds ratio > 1 ( the point estimate) that implies that males are more likely to pass than females, actually between 1.004 and 1.345 times as 'likely' with 95% confidence. ( I'm not sure what I mean by 'likely' I'm just repeating what I've observed in literature)

The same goes for race, RACE 2 is more likely to pass than RACE 6 based on the odds ratios.

I know these results are significant because 1) the Wald Chi-square in the Type 3 analysis of effects ( not given above) is significant for RACE and SEX, implying differences exist and 2) the 95% wald confidence intervals above do not contian '1'. ( If this is correct I don't understand why )

If I'm correct about all of the above, I still have a problem. Just by observing the data, and theory, Females pass at greater percentages than males and RACE 6 passes with greater percentages than RACE 2. My results contradict theory and observation of raw data on both accounts.

I believe in linear regression, when you get the wrong sign on a coefficient it could be caused by an omitted variable. ( EX: if the sign is unexpectedly negative for Beta on X1 then you could be omitting a variable say X2 that is positively correlated to X1 but negatively correlated with Y in a model Y = B1X1 + B2X2) You can fix the problem by discovering the omitted variable and inserting it into the model.

Could an omitted variable be responsible for my odds ratio results? I'm not sure if it works like this. Can someone suggest what I'm doing /interpreting incorrectly?

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

Posted in reply to deleted_user

03-18-2009 09:11 AM

Just a note: I am using proc logistic with the descending option.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

Posted in reply to deleted_user

03-18-2009 09:17 AM

There is a good chance that the observation that you have made is called Simpson's Paradox.

http://en.wikipedia.org/wiki/Simpson's_paradox

http://en.wikipedia.org/wiki/Simpson's_paradox

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

Posted in reply to Doc_Duke

03-18-2009 04:28 PM

I read the link- and should look into it more in depth, but

This could be it! The overall model is giving me 88% correct predictions on validation data! If I take the mean prediced probability for race 6 it is higher than the mean predicted probability for race 2 ( which is what I would expect, but opposite the results the odds ratio/contrasts tell me) It is just when I look at the contrasts and odds ratios at the individual races or sexes that I get weird results. The same thing occurs when I look at the results by sex.

So in this case, in the aggregate my model seems like it predicts well, I just can't rely on the interpretation of the individual contrast statements due to 'Simpson's Paradox'?

Is there a way to correct for this?

This could be it! The overall model is giving me 88% correct predictions on validation data! If I take the mean prediced probability for race 6 it is higher than the mean predicted probability for race 2 ( which is what I would expect, but opposite the results the odds ratio/contrasts tell me) It is just when I look at the contrasts and odds ratios at the individual races or sexes that I get weird results. The same thing occurs when I look at the results by sex.

So in this case, in the aggregate my model seems like it predicts well, I just can't rely on the interpretation of the individual contrast statements due to 'Simpson's Paradox'?

Is there a way to correct for this?