Prior to conducting a logistic regression of the 0/1 likelihood of a nest hatching or failing as a function of 9 continuous predictors, I plotted each of the standardized (mean = 0, SD = 1) predictors against the logit of the outcome (see below). An assumption of logistic regression is a linear relationship between continuous predictors and the logit of the outcome.

The assumption appears satisfied for covariates: diffRobel, DistAspen, DistPonds and maybe meanrobel. Ignore Hatch, that's the response variable.

It is obviously not satisfied for: aspen_m, Count_ppncon, DistRoad, ponds_EDmha, and upland_ha.

1) How can I properly meet this assumption using splines? I've read that it's better to apply splines than transform the data, but I'm looking for clarification on the use of splines. For example, ponds_EDmha (x) appears linear after -0.5 and DistRoad (d) appears linear until -0.5 . Thus, should I write y~(x-0.5)+(d+0.5) instead of y~x+d? This follows instructions on how to center raw data using that method, but it seems very strange.

The data need to be standardized because the covariates are in a number of different units; should I instead apply a spline prior to standardizing data, and if so how?

2) How to apply a spline where a bimodal relationship exists? Like with aspen_m and upland_ha

I'm looking for a simplified explanation of splines or other methods to address non-linearity in the context of standardized predictors, and perhaps some R code Thanks in advance

enter image description here

Code to generate similar plots is found here:http://www.sthda.com/english/articles/36-classification-methods-essentials/148-logistic-regression-assumptions-and-diagnostics-in-r/

Edit: in response to comment I plotted logit ~ predictor values instead, and got the following plot. I'm now thinking that the relationships are linear enough not to worry about it. Count_ppncon isn't great but I intend to replace that covariate with a more relevant one anyway

enter image description here

  • $\begingroup$ Would it not be more informative to plot logit as a function of predictor not vice versa as you seem to have done? $\endgroup$ – mdewey Dec 6 '18 at 21:52
  • $\begingroup$ Thanks for the suggestion, I was following the tutorial linked in the question but swapping the axes was helpful, I'll post the new plot $\endgroup$ – Emily Dec 6 '18 at 22:53

Given the distribution of some of the predictors it might be hard to tell whether a linear term is all that is needed. My main concern if I was fitting this model would be that some of the predictors are quite skewed and if it happens, for example, that high values for diffRobel and DistAspen belong to the same observation that observations might be quite influential in the model so I would suggest looking at regression diagnostics carefully after fitting the model.


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.