I am working in program R. I am modeling the incidence of flight in a seabird in relation to distance to the nearest ship (potential disturbance, range = 0 to 74 km from the bird). 1= flight during observation, 0 = no flight. The bird does fly with some unknown probability when no ships are present or really "far" away. I am trying to find this really far distance and associated probability of flight using binary logistic regression.
Model = Flight ~ ship distance. Other variables were explored but fell out with stepwise selection.
During exploratory analysis I truncated the data down only looking at smaller distances from the ship (20, 15, 10 km). These models are highly significant and predict that as the ship gets closer the probability of flight increases. However when I include all the data (out to 74 km) the intercept is significant (and predicts the true % of observed flight events) but the slope term is non-significant.
Can I use a weighting scheme to give more weight to observations when the ship was closer?
Edit: I am working through the suggestions made by @Scortchi and @Underminer. Here is a plot of a loess smooth on the observed data to better help visualize the pattern.
The distance to the ship data does not discriminate between approaching ships and departing ships it is just a straight line measure to the nearest ship. The dip in the probability of flight at 8.5 I believe can be attributed to "unaffected" birds that did not fly as the ship passed by them. So as the ship departs and gets further from the observation site we were more like to be observing birds that for whatever reason did not fly when the ship passed and are less likely to fly for "naturally occurring" reason. As additional birds fill back into the observation area the "baseline" flushing rate is resumed and birds start to fly at "normal" probabilities.