Data Filter vs.Stratifying

85 views
Skip to first unread message

kate.wi...@gmail.com

unread,
Jan 4, 2018, 9:49:52 AM1/4/18
to distance-sampling
Hi all,

I have a question about an issue I've experienced with stratifying versus filtering my data in program Distance. 

Research summary: 
I am trying to obtain abundance estimates of grassland bird species before and after bison reintroduction and between bison-grazed, cattle-grazed, and reference (no cattle or bison) treatments. I conducted point count surveys at 20 point count stations within each treatment. I had three field seasons from May-June 2015, 2016, and 2017. Within each season, I visited each point count station 5 times. For my first analysis, I would like to compare abundance between years (before and after bison reintroduction) and between only the bison and reference treatments . 

Analysis issue:
I am performing a Half-normal Cosine, MCDS analysis with covariates that include observer, wind, rain, and cloud cover. For my most observed species (>100 observations per treatment included in analysis after truncation), when I stratify by treatment (bison or reference) in the "Model Definition" window and place covariates on detection probability, program Distance cannot estimate abundance for the reference site. I get a zero for abundance, and do not get an estimate of the effective detection radius or detection probability for the reference site. However, when I filter by treatment in the "Data Filter" window, I am able to place covariates on detection probability and get results for the reference site.

Why would I be able to filter the data by treatment but not stratify the data by treatment?

Thanks so much for any insights on this!

Best,
Kate

Eric Rexstad

unread,
Jan 5, 2018, 9:07:24 AM1/5/18
to kate.wi...@gmail.com, distance-sampling

Kate

Sounds like an interesting experiment.  I'll make several comments.

Your inclusion of four covariates are likely causing you difficulty.  You aren't explicit, but I suspect most or all of the covariates are factor rather than continuous.  If they are all factors, then you should query your data to discover how many detections you have in each of the possible categories of covariates.  Pretend you have but two levels for each of your covariates--that results in 2^4 different combinations (16).  That shrinks your 100 observations per treatment down to ~8 detections per combination (assuming all combinations are equally represented).  That will make for a challenging parameter estimation problem.

However, the situation is likely not that rosy.  There are likely combinations that have no detections (presence of rain in the absence of cloud cover); that make the estimation problem even more difficult (likely leading to convergence failure in the estimation routine).  To top it off, your covariates are likely collinear (correlation between your meteorological variables).  You are also incorporating adjustment terms along with the covariates.  Not a surprise that you suffer failure of model fitting. 

I suggest you work with a less ambitious models.  There is a section in the users guide called "MCDS analysis guidelines" that discusses these difficulties.  In part, here is a portion of that section:

Fitting the detection function with multiple covariates is significantly harder, computationally, than in the case where there is only one covariate.  This has
several consequences:
• the analysis engine takes longer to run
• the algorithm will fail to converge more often
Because of this, it is important to be careful and thoughtful when setting up and running MCDS analyses.  Here are some recommendations:
• Rather than including numerous covariates at once, start by including one covariate at a time, and selecting the covariate that gives the best model fit/lowest AIC/AICc/BIC. You can then carry out forward stepwise selection by adding one additional covariate at a time, while retaining the one(s) already selected, until there is no decrease in the AIC/AICc/BIC value (depending on the criteria you are using).
• Bear in mind that factor covariates are usually harder to fit than non-factor covariates, especially as the number of factor levels increases (see Factor and Non-factor Covariates in MCDS earlier in this Chapter for information on factor covariates). If you encounter problems while trying to fit a factor covariate, try reducing (condensing) its number of levels, if possible. 
• Avoid using the feature that allows automatic selection of adjustment terms – at least to start with.  Instead, start by using a model with no adjustments, and if this converges try one with one adjustment.  Gradually work up to more adjustments if required.

Two final thoughts.  Recognise distance sampling has the property of pooling robustness--meaning abundance estimates are unbiased even when all forms of heterogeneity in detection probability are not modelled.  Another reason to simplify the detection function model.  Lastly, you note you are still in the preliminary states of your analysis, have a look at this paper

Buckland, Stephen T., Robin E. Russell, Brett G. Dickson, Victoria A. Saab, Donal N. Gorman, and William M. Block. 2009. “Analyzing Designed Experiments in Distance Sampling.” Journal of Agricultural, Biological, and Environmental Statistics 14 (4):432–42. https://doi.org/10.1198/jabes.2009.08030.

for a situation perhaps similar to yours.
--
You received this message because you are subscribed to the Google Groups "distance-sampling" group.
To unsubscribe from this group and stop receiving emails from it, send an email to distance-sampl...@googlegroups.com.
To post to this group, send email to distance...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/5ff403ee-7684-4550-b3b2-50623c1a4ace%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

-- 
Eric Rexstad
Research Unit for Wildlife Population Assessment
Centre for Research into Ecological and Environmental Modelling
University of St. Andrews
St. Andrews Scotland KY16 9LZ
+44 (0)1334 461833
The University of St Andrews is a charity registered in Scotland : No SC013532

kate.wi...@gmail.com

unread,
Jan 10, 2018, 10:39:47 AM1/10/18
to distance-sampling
Hi Eric,

Thanks so much for your detailed thoughts- I really appreciate it!

Best,
Kate
Reply all
Reply to author
Forward
0 new messages