# Ordinal regression shows negative effect when spearman correlation shows positive

#### micdhack

##### Member
I running a regression model where a variable is indicated to have a negative impact on the predicted variable:
Code:
		Estimate	Std. Error	Wald	df	Sig.	95% Confidence Interval
Lower Bound	Upper Bound
Threshold	[email_contact = 1]	-4.682	.664	49.713	1	.000	-5.983	-3.380
[email_contact = 2]	-2.401	.548	19.172	1	.000	-3.475	-1.326
[email_contact = 3]	.040	.526	.006	1	.940	-.992	1.071
Location	[unit_supportc=1]	-.162	.771	.044	1	.834	-1.674	1.351
[unit_supportc=2]	.170	.583	.085	1	.771	-.973	1.313
[unit_supportc=3]	-1.286	.497	6.699	1	.010	-2.260	-.312
[unit_supportc=4]	-.568	.267	4.533	1	.033	-1.090	-.045
[unit_supportc=5]	0	.	.	0	.	.	.
What I cannot understand is that the spearman correlation shows a .172 statistically significant effect that is obviously positive.

Why is that and which result should I trust?

#### noetsi

##### No cake for spunky
I am not sure Spearman's would even be appropriate (it has assumptions that may have been violated which you should consider). Ignoring that Spearman is a bivariate relationship while the relationships in logistic regression are marginal, the relationship between two variables controlling for other variables in your model. So the relationship is often significantly different.

I am no expert on this at all, but I think aspects such as moderation effect can cause the sign to be different between bivariate and multivariate models.

#### hlsmith

##### Less is more. Stay pure. Stay poor.
Ignore the Spearman, since you have a binary variable (your dependent variable). Also what was your independent variable and how was it formatted. Your output is very difficult to interpret (above).

One preliminary thing you can do is to either put data into a 2x2 table (if both variables are binary) or plot the distributions and means/sd if continuous. This way you will get an idea of their relationship. However, like Noetsi mention, these approaches would be crude in that they do not control for other variables.

#### micdhack

##### Member
Ignore the Spearman, since you have a binary variable (your dependent variable). Also what was your independent variable and how was it formatted. Your output is very difficult to interpret (above).

One preliminary thing you can do is to either put data into a 2x2 table (if both variables are binary) or plot the distributions and means/sd if continuous. This way you will get an idea of their relationship. However, like Noetsi mention, these approaches would be crude in that they do not control for other variables.
I should have clarified, both variables are ordinal scale. So the responses 1,2,3,4,5 are actually codes for likert scale. Hence, why I used spearman and why I am confused about why ordinal regression would show different results.

#### PeterFlom

##### New Member
First, the results you show have two independent variables and one dependent variable (total 3) and Spearman (or any) correlation is between only two variables. So, the two analyses are about different questions.

Second, e-mail contact has a positive relationship with the DV - the parameter estimate is negative when email contact is low.

Third the relationship between location and your DV is not monotonic.

#### micdhack

##### Member
First, the results you show have two independent variables and one dependent variable (total 3) and Spearman (or any) correlation is between only two variables. So, the two analyses are about different questions.

Second, e-mail contact has a positive relationship with the DV - the parameter estimate is negative when email contact is low.

Third the relationship between location and your DV is not monotonic.
Email contact is the DV, unit support is IV. It is an ordinal regression, so, the email contact has thresholds and the IV has the different levels. There is not third variable.

#### hlsmith

##### Less is more. Stay pure. Stay poor.
What happens when you plot them with a little jitter?

#### micdhack

##### Member
What happens when you plot them with a little jitter?
Excellent idea! It seems that mainly there is a positive correlation, but early categories seem to have a negative correlation. Does this make sense?

#### hlsmith

##### Less is more. Stay pure. Stay poor.
Good stuff. I would agree that you definitely don't see a bunch of 1:4 combinations and more 3:4 and 3:5, etc. Another question would be whether this all makes sense with your familiarity with the subject matter?