Meeting in class. Present: Scott, Sean, Tom. Brett and Sarah joined for parts as well.
Tom fit correllograms for species abundances, taking the mean abundance over time. Found an interesting relationship; high local variance (low distance), declining to negative covariance for intermediate distance, then going up above zero for larger distances.
This could be reflective of the biology of the saddle; points on the edge of the saddle may be more similar due to their slope. Brett says periodic structure sometimes occurs when there's some sort of spatial pattern like this.
But, are these correlations spurious? Correlogram has white and black circles - are black circles statistically significant correlations?
Tom will continue looking into spatial variance and covariance looking forward. This could be useful in variance partitioning.
We had been thinking previously about using vegetation classes in analysis. Is there a good way to do this? Part of the concern is that it's circular; use abundances of a species to determine the vegetation class, then use the vegetation class to determine abundances. It's not statistically rigorous.
Brett says that we could consider determining vegetation classes by removing observations of each of our focal species. This is still a little bit strange - there's some conflation of cause and effect here. But, for building predictive models, cause and effect is less important.
Here, the surrounding community type could even be a predictor. E.g., see what the mean change/vegetation type is per each of the vegetation classes. Alternatively, these could be used as a unit of prediction, e.g., what is the change in Deschampsia within dry meadow communities compared to others?
Could abiotic factors be used to fit these? We aren't sure how this would be done; Marko's paper did not do this.
See Sean's idea about vegetation classes in analysis below.
We had considered fitting a single multinomial model rather than several indiviual species models. This handles the fact that when one species goes up in abundance, other species are likely due to go down due to the fact that there are only 100 samples per plot.
According to Brett, this would be a good thing to focus on in the future, but not now. It would be time-intensive to hand-code a flexible multinomial model (e.g., one including spatial random effects).
Nitrogen deposition could be useful here. Cliff says:
"From other experiments we know that deschampsia like warming and nitrogen. geum and kobresia both decline with nitrogen. geum and deschampsia dominate moist meadows while kobresia and carex rupestris dominate dry meadow"
Through Emily Farris's et al. (2015, Journal of Ecology) paper, he found the National Atmospheric Deposition Program:
http://nadp.slh.wisc.edu/data/
They have a sampling site up on the Niwot saddle:
http://nadp.slh.wisc.edu/data/sites/siteDetails.aspx?net=NTN&id=CO02
Sean downloded this data and it should be in the repo now.
Sarah also says there hould be some nutrient sampling done off the saddle. Surprisingly, there isn't any overlap of these samples with the vegetation plots. There may be samples at the sensor nodes... these maybe could be used for some sense of spatial variance (e.g., if there is little variation between the measurements here, this would suggest there isn't much variation in space).
Amazingly, NADP has annual measurements. These could cut down on temporal variance.
Scott also found this: https://portal.edirepository.org/nis/mapbrowse?packageid=knb-lter-nwt.50.1 worth looking into and exploring. Could also look at the Niwot "friends of Niwot" page.
Sarah is not a huge fan of using the maximum temperature observed in a given year; it throws out so much data. What if the instrument is acting up on a given day? Very risky.
One thing to consider is the "extended summer" metric that Cliff and Katie have developed. It's an index which incorporates a number of abiotic measures.
GDD up until the sampling date is a logical choice of temperature metric.
In a limited number of studies, Sarah and others have found that the average temperature (over the summer, i.e., June - August) over the last four or five years is a good temperature metric to use. Do we have temperature data from the saddle for five years before the first sampling date? This was supported by a GLORIA study. We could get more sophisticated that June - August measurements, as the decision to use these metrics was more a result of data availability than anything else.
For annual measurements (temperature and Nitrogen), we could use packages to "downscale" the measurments of various metrics, i.e., to get plot-level measurements for each year.
Related to temperature, solar radiation on August 1st is a good annual metric of solar radiation, as it's the middle of the growing season. However, early in the growing season may also be good, as it catches the early stages of plant growth, which may have outsized influences on dynamics.
What are achievable pieces of work before the semester ends? Sarah recommended that rather than adding very sophisticated features, we should focus on building a framework for incorporating variables into the model. Focus on four or five variables we want to include, think of what a model based on these would look like, and come up with a way to evaluate these predictors.
Some predictors we thought of:
-
Temperature (Scott can work through this)
-
Nitrogen (Sean can look into this)
-
Spatal trends (Tom can do this)
-
Snow measurements (Scott has been looking into this)
This is what we'll work on for he next several weeks.
Advantages to temperature and nitrogen: we can do forecasting based on future estimates of these variables. Disadvantages is there is less purely-temporal variance in the model, so these variables may not explain much.
Sean had a good idea on framing a multi-level model: look at within- and among-community type changes. What are the drivers of among-community change? I.e., are plots switching between community types, and if so, what is determining these changes? This is similar to, but could be made more sophisticated than, Marko's analysis. It makes sense, because trends in species abundance could be solely due to increases in certain plot types. It would also be interesting to see what the dynamics within a community type are - e.g., are dry meadows becoming more dominated by Kobresia? This seems like it could be answered with a multi-level model.