0

I have a small problem with one of my analyses. I'm doing a multivariate logistic regression so some continuous variables have been discredited by quartiles.
Here are my results: OR and IC

Variable unit OR lower limit IC upper limit IC
Variable A 20 1.3437560  1.0814409 1.6757319
Variable B first quartile 1 2.0054150 0.8881467 4.8872483
Variable B second quartile  1  1.2150582 0.5014275  3.1025704
Variable B third quartile 1 1.4801394 0.6200311 3.7682139
Variable C first quartile 1 2.5960149 1.1427893 6.3318406
Variable C second quartile 1 2.7227333 1.2000598 6.6519756
Variable C third quartile 1 1.9318410 0.8255948 4.8292896
Variable D -200 1.0308258 0.8215131 1.3461578
Variable E -10 1.0478705 0.8473907 1.2948567
Variable F 1 1.0802896 0.5128503 2.1359383

Wald test

Variable Df Chisq p-value
Variable A  1  7.0270868  0.0080286
Variable B  3  3.4891983  0.3221658
Variable C  3  6.4613940  0.0911975
Variable D  1  0.0573684  0.8107047
Variable E  1  0.1876317  0.6648951
Variable F  1  0.0455897  0.8309234

For my variable C, which has been discredited, the confidence interval of the first and second quartile does not include 1 so significant, but my variable C is not significant according to the Wald test. As I understand (With my humble knowledge in statistics) the Wald test in the context of logistic regression is used to determine whether a certain predictor variable is significant or not. It rejects the null hypothesis of the corresponding coefficient being zero.

How can I interpret this result ? Should I reject the entire variable or just the third quartile?

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
BoudSTER
  • 1
  • 1
  • 2
    Breaking up the predictors into quartiles is [not a good approach](https://stats.stackexchange.com/q/68834/28500). Continuous modeling with restricted cubic splines is much more flexible and allows you to evaluate both the overall significance of the predictor and the significance of the nonlinear terms. If you are interested in prediction, the "statistical significance" of the individual coefficients doesn't really matter, anyway--there's no need to "reject" any variable. – EdM Sep 25 '21 at 16:05
  • 2
    Related to @EdM's great answer, the use of the term 'discredited' in the OP is very appropriate here. The result is arbitrary, power-losing, and is completely impossible to interpret for reasons detailed in [RMS](https://hbiostat.org/rms). – Frank Harrell Sep 26 '21 at 12:46
  • Thank you for your answer. Indeed using a modeling with restricted cubic splines is a good solution. But my thesis master wants us to discredit the non-linear continuous variables (old school I think). – BoudSTER Sep 26 '21 at 15:01
  • Thanks for the documentation ;). – BoudSTER Sep 26 '21 at 15:03
  • First: Multivariate regression is when you have multiple outcome variables, you probably have multiple regressors, which us *multiple regression*. Please edit/clarify! Then: It is known that Wald tests are not very adequate with logistic regression, see https://stats.stackexchange.com/questions/457851/logistic-regression-wald-statistic – kjetil b halvorsen Sep 27 '21 at 01:17

0 Answers0