I'm working with logistic regressions for the first time, and I'm trying to wrap my head around my results. I've read tutorials and other posts here on CV quite extensively, but I'm still unsure that I have this straight in my head.

I have used the logistic regression to determine how rate of change of a landscape (%) influences the climate-growth relationship (the binary variable - positive or negative) of individual trees.

Here is one of the logistic regression models, the summary, the CI, and the odds ratio that I calculated using exp(coefficient):

> glm_wintermin_neg<-glm(wintermin$Neg~wintermin$thaw, family="binomial")
> summary(glm_wintermin_neg)

glm(formula = wintermin$Neg ~ wintermin$thaw, family = "binomial") 

Deviance Residuals: 
    Min       1Q   Median       3Q      Max  
-0.6391  -0.5989  -0.5967  -0.3720   2.3261  

               Estimate Std. Error z value Pr(>|z|)    
(Intercept)    -2.80287    0.55190  -5.079  3.8e-07 ***
wintermin$thaw  0.03136    0.01535   2.044    0.041 *  
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

    Null deviance: 290.64  on 349  degrees of freedom
Residual deviance: 285.28  on 348  degrees of freedom
AIC: 289.28

Number of Fisher Scoring iterations: 5

> confint(glm_wintermin_neg)
Waiting for profiling to be done...
                      2.5 %      97.5 %
(Intercept)    -4.067980576 -1.85254625
wintermin$thaw  0.004366896  0.06591151

> exp(glm_wintermin_neg$coef)
       (Intercept) wintermin$thaw 
    0.06063549     1.03185810 

The odds ratio I calculated as 1.03, and this is significant (p = 0.041). My concern is that I have other non-significant (p > 0.1) models that have odds ratios that are very close to this odds ratio, but both slightly higher and slightly lower (OR = 1.019, 1.047) which set off some alarm bells. My interpretation is that what my results mean is that the odds of a negative response are 1.04x larger than a non-negative response. The reason for this particular model being significant vs. the others that have similar ORs I believe is because for the significant model, the CI doesn't encompass 1. If someone would be able to confirm my interpretations, and/or clarify areas that I'm off track, I would be very grateful.

Thanks in advance, and if there is any other information, code, or results that I can provide, please let me know.


There are already many answers to this question: 1, 2, 3, 4 to name a few. You should have a look through these and perhaps look into a book on the topic. I recommend Applied Logistic Regression by Hosmer, Lemeshow for a rather comprehensive take on the subject, or this site this tutorial from IDRE for a quick brief.

However, to give you a few starting pointers, the interpretation of the odds ratio, strictly, is that a one unit change in the predictor, here wintermin$thaw, increases the risk of a negative respone by 1.03 times.

The confidence intervals that you have computed so far are in fact the intervals of the estimates of the coefficients, not the odds ratios. Even so, you should reach a similar conclusion given that these do not cross 0. To get the confidence intervals for the odds ratios, you type:

  • $\begingroup$ I agree with Johan. An odds ratio of 1.03 would likely be too close to 1 to be statistically significant. I agree that the output you showed contained a model coefficient of 1.03 but no odds ratio. $\endgroup$ – Michael Chernick Nov 28 '16 at 2:17

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.