Analysis with R: Assignment 2

Distance sampling
1

Submit your assignment for the Covariates module in a reply to this topic

Remember to include these five elements:

  1. The results :1234: from your best model
  2. Estimates :dart: for density, detectability and ESHW, including confidence intervals
  3. Interpretation :bulb: What conclusion do your draw from your results?
  4. A brief reflection :thinking: on your next step, challenge you solved, problem you still have or leap of understanding
  5. Give specific and useful feedback :thumbsup: to a coursemate

Remember that you can always ask for help or share a draft of your assignment in our course discussion area if you want feedback or assistance

2

Using the water deer data, I computed the following 2 models:

  • Model using uniform function for detectability that only computes Land cover type as a covariate (uniLC_.)
  • Same as previous but also including Distance to coast as covariate (uniLC_DTC)

To choose which of them fits best the data I analysed the model likelihoods and weights:

nPars AIC delta AICwt cumltvWt
uniLC_DTC 2 385.66 0.00 1.0e+00 1.00
uniLC_. 1 406.98 21.32 2.4e-05 1.00

Clearly the one only computing LC does not to a good job modeling density in our case. Still, I wanted to contextualise these compared to the one which fitted best our data previously:

                    `nPars    AIC  delta   AICwt cumltvWt`

hnLC_. 3 271.39 0.00 1.0e+00 1.00
uniLC_DTC 2 385.66 114.27 1.5e-25 1.00
uniLC_. 1 406.98 135.59 3.6e-30 1.00

It is clear, in fact, that both have a very poor performance compared to the model using a half normal function and containing LC only.

At this point I got a little bit stuck on the following reflection. Are models using a uniform function for detectability suited for the introduction of covariates? Because detectabilty decrease uniformly as distance increases (it feels like it could go against the assumption of the function; I even read that in the case of RStudio package “Distance” -Marshall et al 2019- covariates cannot be included in the model if uniform function is being used for the specific model). What do you think, would you agree?

Having this in mind, I moved on to do the other parts with the 3rd and 4th best models we produced in the module (both using hazard rate function but one containing both covariates and the other only LC) and obtained the following:

  • Detectability was 153 (CI: 109 to 215) in grassland compared to 42 (CI: 35 to 51) in wetland, showing once again how, as expected, it is clearly higher in the former (not just by the estimate but also by witnessing no overlapping of the CIs).
  • Density was 1.3 deer/km^2 (CI: 0.8-1.9) at a distance of 25 m to the coast compared to 4.6 deer/ km^2 (CI: 3.7– 5.6) at 400 m, a trend much “steeper” to the one we observed with the models used in the course (but once again, this would not be very trustworthy).
  • As for the Effective Strip Half Width, this was 34.7 in wetland compared to 147 in grassland.
3

Great work interpreting the AIC model comparison table, Javi!

I can see that your results from the hazard-rate models are in the right ball-park, but I wasn’t able to replicate your precise results. I’m no’t sure which model you used for each estimate, and we might have used different truncation distances too. If you would like me to double-check your calculations, please let me know the truncation distance and model you used for each estimate. Thanks

See this post for my thoughts on the uniform detection function and covariates