Why opposing results occured with categorical and continuous predictor ?

[Minhtrang]

Hi all,

I am running a Cox model with outcome as mortality, and predictor as Magnesium (Mg).

When I treat Mg as quintiles, p-value for this predictor is significant.
However, when Mg was entered into the model as continuous, p-value is >0.05.

I can not find an explanation for this difference. What are next steps to be considered?

I would love to hear from your experience about this problem.

Thank you.

[Doc_Duke]

You didn't tell us the actual p-values.  For instance, .04 and .06 are not that different in strength of evidence.

Other things to look at: 

--  Is the effect of mg linear?  A curvilinear relationship can show up as non-significant in a continuous setting.  Looks at your graphical diagnostics.

--  There may be an interaction that you are not modeling.

There are a couple of good SAS Books-By-Users that can lead you through exploring the relationships.

Doc Muhlbaier

[PaigeMiller · Accepted solution]

---

@Minhtrang wrote:

Hi all,

I am running a Cox model with outcome as mortality, and predictor as Magnesium (Mg).

When I treat Mg as quintiles, p-value for this predictor is significant.
However, when Mg was entered into the model as continuous, p-value is >0.05.

I can not find an explanation for this difference. What are next steps to be considered?

I would love to hear from your experience about this problem.

Thank you.

---

There's no reason to think that a model based upon quintiles (in which you are throwing away data) will produce the same result as a model based upon the continuous variable (where you are not throwing away data). In fact, if you have the continuous variable values, I can't really think of a reason why you'd even want to use quintiles in place of the continuous data ... I doubt you could justify this to a reviewer or professor.