3
$\begingroup$

I have the following contingency table

      D    not-D
E     980    122
not-E 2420  6439

where D may stand for disease, and E for exposure.

This leads to the odds ratio (OR) $$ \frac{odds(D|E)}{odds(D|\bar{E})} = 21.4 $$

[1] I hope that is correct?

[2] This ratio says the odds of having D given E is about 21 times the odds of having D given you do not have E i.e. having E greatly increases the chance of D?

[3] Therefore, in a logistic regression where D is the response and where E is included (amongst several other predictors) as a binary predictor variable, with $\bar{E}$ as the reference level, I would expect its model coefficient to be positive, so that if E is present the odds of D is greater than if E is not present?

[4] The problem I have is that the model coefficient for E is negative which I interpret to mean having E reduces the odds of having D, which contradicts what I understand the OR is telling me?

EDIT:

Using R, with some variable name changes,

fit <- glm(D ~ x1 + x2 + x3 + x4 + xE, family = binomial, data = df)

$\endgroup$
  • $\begingroup$ Welcome to CV, PM.! Can you edit your question to indicate your software and command for your logistic regression? $\endgroup$ – Alexis Sep 27 '18 at 16:40
4
$\begingroup$
  1. Yes
  2. Yes
  3. No, after adjustment for other variables, it's possible for the association to change direction. The above table is a crude odds ratio, so may be subject to bias of confounding.
  4. To verify you haven't made a coding issue, fit the logistic model without adjustments and verify that the log odds ratio is log(21.4).
$\endgroup$
  • $\begingroup$ Your succinctness and specificity are laudable. :) $\endgroup$ – Alexis Sep 27 '18 at 16:41
  • $\begingroup$ @AdamO If I understood your suggestion correctly, I ran the fitting with xE as the only predictor variable, and obtained the coefficient 3.06214. Now, log(21.4)=3.063391 so they are equal ( subject to a bit of rounding i.e. 21.4 is the OR rounded to 1 d.p. ) $\endgroup$ – PM. Sep 27 '18 at 16:59
  • $\begingroup$ @PM. great, it sounds like your other adjustments have a VERY strong correlation with both E and D if the crude OR is 21 and the adjusted OR is less than 1. I would just be sure to describe very carefully what you did, and how you selected those adjustments, when you report the findings. $\endgroup$ – AdamO Sep 27 '18 at 17:51
  • $\begingroup$ @AdamO Many thanks, very helpful. I clearly have a few things to ponder and some further reading to do. $\endgroup$ – PM. Sep 28 '18 at 8:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.