I have fitted a logistic regression model showing the probability of flight given distance to a disturbance event. My sample size is 140 observations of which 45 were observed to fly. Distance to the disturbance is the only significant predictor among many explored (biologically relevant interactions as well). Here is the out from the R console

glm(formula = flight ~ approach_km, family = binomial)

Deviance Residuals: 
   Min       1Q   Median       3Q      Max  
-1.1233  -0.9317  -0.7460   1.3538   1.9400  

        Estimate Std. Error z value Pr(>|z|)  
(Intercept)  -0.1233     0.3438  -0.359   0.7199  
approach_km  -0.4530     0.2225  -2.036   0.0418 *

Null deviance: 175.82  on 139  degrees of freedom
Residual deviance: 171.30  on 138  degrees of freedom
AIC: 175.3

Number of Fisher Scoring iterations: 4

Here is a plot of the predicted probabilities of flight.

binary logistic regression, flight~distance

As you can see there is no point along the distance continuum in which the probability of flight exceeds 0.5. Since this is the case the classification tables will never predict a flight (it predicts 67.9% correctly, ALL THE NO FLIGHT EVENTS!). McFadden's psuedo R2 = 0.026. I am happy with the fit of the model but I am under the belief that I need to include these types of goodness of fit statistics when publishing. I would rather show that within a certain range of the data, say 1 km, there is roughly a 0.40 probability of flight. If there were 20 observations associated with this estimate there would be 8 flight events and 12 non-flight events. Is there a way to show this? How can I write up these results to show my findings without referee's balking at my lack of fit tests? Any thoughts or references would be greatly appreciated.

UPDATE: In response to @Drew75.

Here is the ROC curve. Does this mean that the model does not perform as well I has thought? ROC curve

  • $\begingroup$ Have you thought about an ROC curve? I'd guess that this model only performs slightly better than a random guess. $\endgroup$ – Drew75 Feb 26 '14 at 19:00
  • $\begingroup$ @Drew75 I added the plot of the ROC curve as an edit to the original post $\endgroup$ – marcellt Feb 26 '14 at 23:38
  • $\begingroup$ @marcelit, do you know how to interpret the ROC curve? It supports the previous graph, saying that your model doesn't predict much better than a random guess. $\endgroup$ – Drew75 Feb 27 '14 at 6:49

You should think of when your predicted probability crosses .301 (the share of flight events) not 0.5. If the probability is higher than .301 then it is higher than a random guess provided the distribution of the data, even if it is not as high as 0.5. i.e. you learn something.


Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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