[RsR] use of lmrob() on ecological time series

[Martin Maechler]

>>>>> Stahel Werner A 
>>>>>     on Sat, 24 Aug 2019 20:53:22 +0000 writes:

    > Dear Emily

    > This is a late answer to your message from July 15.
    > The first issue is the use of robust linear regression of log(abundance) on the year.
    > I think that this is a very reasonable way to summarize the time series -- as long as
    > a log-linear trend appears appropriate.
    > The trend is then measured by the slope. A great advantage of using the log is that
    > slopes are then on a common scale for rare and abundant species, as a certain slope
    > corresponds to a certain percent increase or decrease per year.

    > You then go on to question the classification into significantly increasing, and so on.
    > This classification is common but unreasonable.
    > A 5% increase per year can be significant for one species and insignificant for another,
    > just because the former shows less random fluctuations than the latter.
    > We should focus on estimation and supply confidence intervals for characterizing the
    > (im-)precision.
    > (You may have read about the controversy about "null hypothesis significance testing"
    > and p values.)

    > Since the slopes are used for further analysis, the classification is not needed nor helpful
    > at all.
    > In any case, I have not read in detail what is done with the slope. In one paper, it is used
    > as the target variable in further regression models.
    > I wonder if such regressions make sense when different species are used in the same
    > regression. I thought it was a basic paradigm of biology that species have different
    > ways to react to environments.
    > If one simply want to show that management is helpful, one might compare managed and
    > non-managed regions in terms of the number of species (within taxonomic groups?) that
    > have recovered -- or directly in terms of average slopes for individual species or taxonomic
    > groups.

    > Nevertheless, let me add a thought about (2).
    > I think the expression "non-significant change" is quite appropriate since a change of 0
    > does not exist in real life. It is likely small (unless fluctuations are big and/or the time series
    > short, which causes in-significance), but never precisely 0.
    > Again, a confidence interval says it all: It contains all plausible values of the true slope.

Indeed! ... and the generic R function for confidence intervals,
confint() also has a good methods for "lmrob" objects,  e.g.,

    > (RlmST <- lmrob(log.light ~ log.Te, data = starsCYG))

    Call:
    lmrob(formula = log.light ~ log.Te, data = starsCYG)
     \--> method = "MM"
    Coefficients:
    (Intercept)       log.Te  
	 -4.969        2.253  

    > confint(RlmST)
		      2.5 %   97.5 %
    (Intercept) -11.8375506 1.898775
    log.Te        0.7041937 3.802129
    > 


  > Are these thoughts helpful?

    > Werner Stahel
    > M +41 79 784 9330 | P +41 44 364 6424

    > ________________________________
    > Von: R-SIG-Robust <r-sig-robust-bounces using r-project.org> im Auftrag von Emily Klein <emily.klein04 using gmail.com>
    > Gesendet: Montag, 15. Juli 2019 21:34:45
    > An: r-sig-robust using r-project.org
    > Betreff: [RsR] use of lmrob() on ecological time series

    > Dear all,

    > I am using the lmrob() function from the robustbase package, and I have a
    > few questions. To keep the threads clear, I have a general inquiry here,
    > and will ask more specific Qs in a second thread. NB: I don't typically
    > update in the middle of a project, so am running on R version 3.4.1.

    > (1) I am curious the community's thoughts on our approach: We have several
    > hundred ecological time series and we're using robust linear models to
    > determine if the time series are increasing, decreasing, or not changing,
    > by looking at the modeled slope. This approach follows several others,
    > including Lotze et al. 2017 (doi: 10.1111/cobi.12957) and Magare et al.
    > 2013 (doi:10.1371/journal.pone.0077908). I don't have much experience with
    > RLMs, so any thoughts on this approach would be very welcome.

    > More specifically, following the work noted above, we are running (with the
    > time series indexed with "DBx"):
    > lm_test<-lmrob(log(pop_status+1)~observation_year,DBx)

    > (2)  Previous use of RLMs to identify the direction of ecological time
    > series was asked in peer review to use "non-significant change" to
    > reference time series with a slope of zero within the 95% confidence
    > intervals. I can see excluding time series where there is no agreement on
    > the direction of slope, but I think that slope=0 is more "stable" or "no
    > change" and is not necessarily "non-significant". Any thoughts?

    > Thank you all very much for any feedback you may have. I will start a
    > second thread on a few warnings I am getting.

    > Emily

    > --
    > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    > Emily S. Klein, Senior Postdoctoral Associate (she / her / hers)
    > The Frederick S. Pardee Center for the Study of the Longer-Range Future |
    > Boston University
    > *Co-Chair*, ICES Working Group on the History of Fish & Fisheries (WGHIST)
    > esklein04 using gmail.com

    > http://www.bu.edu/pardee/
    > http://www.ices.dk/community/groups/Pages/WGHIST.aspx


    > _______________________________________________
    > R-SIG-Robust using r-project.org mailing list
    > https://stat.ethz.ch/mailman/listinfo/r-sig-robust