0
$\begingroup$

enter image description here

I have a question. I want to conduct a binary logistic regression analysis, with a dependent categorical variable of "20 or higher score on an addiction severity questionnaire" and predictors. But, i want to understand the following that I have included as an image as well in this posting.

1) when I include 4 predictor variables, I do understand that I get different data and exp (b) per predictor than when I use 6 predictors. But, what i do not understand is, when i use extra predictors, that a predictor that earlier revealed a lower predictive value, somehow shows a higher predictor value suddenly. And 2) in the picture you can see the "Drinker + me association score using all blocks", which shows a strange exp (b) value when I use it in the analysis with 6 predictors. >>>>>>>> i would like to understand how this is possible, what i am doing wrong, and if i can prevent this from occurring or solve it.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

Logistic regression requires ten events (i.e., 1s) per predictor entered in the model. If this is violated, parameter estimates and hypothesis tests are invalid. That is the case here. You are getting wild estimates because the data cannot support the model you proposed. You should not trust any of the coefficient estimates.

In general, massive coefficients arise due to perfect separation, which means that the variable perfectly predicts whether a zero or one was observed. This is a frequently asked-about topic here, so you can search for it to find causes and solutions. Some of the solutions (e.g., Firth's bias-corrected logistic regression) may not be available in SPSS.

The solution to your problem is to get more data. Logistic regression is a large-sample method. If you can't do that, you can't test your hypotheses.

Improve this answer
$\endgroup$
3
  • $\begingroup$ thank you for you constructive comment! $\endgroup$
    – JUDITH BEERTEN
    Commented Apr 25, 2020 at 16:09
  • 1
    $\begingroup$ I am very sorry, but I have to disagree with almost everything in this answer. I can agree that this model probably uses too many variables for the 50 observations in the dataset. However, your initial statements are not generally correct and are stated with far more certainty than they deserve. Moreover, they aren't even supported by the statistical output, which indicates the particular coefficient in question differs significantly from zero but is not due to "perfect" separation. Logistic regression is not limited to large samples. In this case the hypotheses can be tested. $\endgroup$
    – whuber ♦
    Commented Apr 25, 2020 at 21:20
  • $\begingroup$ I agree I spoke with more certainty than deserved without evidence. I'm relying on the results of Peduzzi et al. (1996) who find that parameter estimates and inferences are unpredictably biased with fewer than 10 EPV. My understanding is also that logistic regression parameters estimated with MLE are only asymptotically unbiased and this sample size is meaningfully far from that. $\endgroup$
    – Noah
    Commented Apr 25, 2020 at 22:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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