TWINSPAN is perhaps one of the most popular clustering methods (at least among vegetation ecologists), which is not implemented in R. R-sig-eco forum has several posts (mostly from Jari Oksanen and Dave Roberts) on the topic of TWINSPAN in R, where they described difficulties with importing the original TWINSPAN code (written in FORTRAN) into R. Seems that both Jari and Dave spent considerable effort trying to implement the method into R, but seems like there is some problem which is not easy to crack.
From my experience (and I guess also from experience of many others, not only vegetation ecologists), TWINSPAN does sometimes give rather nice and ecologically meaningful results, since it is based on cutting the data along the main compositional gradients. There is yet another divisive method, somewhat analogous to TWINSPAN, called DIANA (DIvisive ANAlysis clustering), which was proposed by Macnaughton-Smith et al. (1964), described in detail by Kaufman & Rousseeuw (1990), and made available in series of Fortran written programs with poetic names like AGNES, CLARA, DAISY, DIANA, or FANNY (later implemented into R package cluster). But it seems to me that this method for some reason never gained so much popularity like TWINSPAN did, mostly perhaps due to its sensitivity to outlying samples (which will be separated first into one-item clusters). This is why I think it still would be nice to have TWINSPAN handy in R, also because of its modified version, which we introduced couple years ago (Roleček et al. 2009) and which up to now is available only in JUICE, program for editing and analysis of vegetation data developed by Lubomír Tichý (Tichý 2002). Modified TWINSPAN is basically a sequence of divisions calculated by standard TWINSPAN, each time applied on the most heterogeneous group – here is again a similarity with DIANA, which also in each step divides the group which is the most compositionally heterogeneous.
Recently, I got a simple idea, how to make TWINSPAN work in R, and how to extend this implementation also for modified TWINSPAN. The way I used is somewhat clumsy, but it seems to work. I used the twinspan.exe file, which is an executable program based on original FORTRAN library written by Mark O. Hill (author of the algorithm and the original FORTRAN code) and compiled in 2003 by Stephan M. Hennekens into a stand-alone program for use in MEGATAB, software which before was used together with TURBOVEG for editing and classification of vegetation data (Hennekens & Schaminée 2001, Schaminée & Hennekens 2001). I created R package twinspanR, which includes twinspan.exe, and added bunch of supporting functions to maintain import and export of data to twinspan.exe and back to R. So basically the TWINSPAN is calculated by the original Hill’s algorithm, and R functions in twinspanR package are for handling the whole thing conveniently in R (see the notes below for more technical details how the library exactly works).
For details how to install the library from GitHub see the Readme.md file. Some more examples how to use twinspan package will be (hopefully) made soon available on my website for analysis of community ecology data in R.
The following is an example code, using example dataset danube with Ellenberg’s meadow dataset (used also as an example in the first publication of TWINSPAN and DCA by Hill 1979 and Hill & Gauch 1980, respectively). In this example I used modified TWINSPAN (Roleček et al. 2009) with division into 4 groups and heterogeneity of clusters measured by Bray-Curtis dissimilarity measure. I projected the results into DCA ordination diagram, along to the original tabular classification made manually by Ellenberg (from Mueller-Dombois & Ellenberg 1974):
## Modified TWINSPAN on traditional Ellenberg's Danube meadow dataset,
## projected on DCA and compared with original classification into
## three vegetation types made by tabular sorting:
res <- twinspan (danube$spe, modif = TRUE, clusters = 4)
k <- cut (res)
dca <- decorana (danube$spe)
par (mfrow = c(1,2))
ordiplot (dca, type = 'n', display = 'si', main = 'Modified TWINSPAN')
points (dca, col = k)
for (i in c(1,2,4)) ordihull (dca, groups = k, show.group = i, col = i,
draw = 'polygon', label = TRUE)
ordiplot (dca, type = 'n', display = 'si', main = 'Original assignment\n (Ellenberg 1954)')
points (dca, col = danube$env$veg.type)
for (i in c(1:3)) ordihull (dca, groups = danube$env$veg.type,
show.group = unique (danube$env$veg.type)[i], col = i,
draw = 'polygon', label = TRUE)
Some technical details how the twinspanR package works
Some further details how it works. The executable file, twinspan.exe, is stored in \exec subdirectory of the library. There is also tw.bat file, which launches twinspan.exe and feeds it with data from R (I have shamelessly stolen this idea from the way how Lubomír Tichý executes TWINSPAN in JUICE, and it is also similar to the way how Tom August implemented another Hill’s FORTRAN library, FRESCALO (Hill 2011), into R as a function frescalo in package sparta). I take compositional data in R, transform them to required *.cc! format (luckily there is a function write.CEP in rioja package, written by Steve Juggins) and save them to \exec subdirectory (where also twinspan.exe is located). Then, I create file tw.dat with input parameters for twinspan.exe (using function create.tw.dat), and use the shell function in R to launch the tw.bat file. All calculations are done using original twinspan.exe, R just reads its output from tw.PUN file. It’s in a no way an elegant approach, but works just fine.
The twinspan.exe file used for calculation of TWINSPAN in R is taken from the distribution of JUICE program (Tichý 2002). It is the original FORTRAN code written by Mark O. Hill, compiled (cca in 2002) by Stephan M. Hennekens into twinspan.exe for use in his program MEGATAB (which was used together with Turboveg for analysis of community data). The version of twinspan.exe used here implements the changes by Petr Šmilauer, related mainly to problems with algorithm convergence, which cause the results being dependent on the order of samples in the input data table. Note that this algorithm is slightly different from TWINSPAN implemented in WinTWINS software (Hill & Šmilauer 2005), which implements also other modifications by Birks and ter Braak. In fact, there are at least four different versions of TWINSPAN recently used for analysis, and in certain circumstances they differ in the results how they classify the samples (when I have time, I will elaborate this topic further).
Unfortunately, the presence of twinspan.exe file in the twinspanR library is a problem for its portability – seems like CRAN doesn’t allow packages with executables inside (for understandable security reasons), and R-Forge allows it to be uploaded, but fails to build it. For now, the channel for distribution of this package is GitHub, and it will perhaps remain there until some other solution will show up. Simply, it’s just a quick and dirty way how to get TWINSPAN functionality in R without too much hassle, before somebody manages to write fully functional implementation of TWINSPAN in R.
- Hennekens S.M. & Schaminée J.H.J. (2001): TURBOVEG, a comprehensive data base management system for vegetation data. Journal of Vegetation Science, 12: 589-591.
- Hill M.O. (1979): TWINSPAN: A FORTRAN Program for Arranging Multivariate Data in an Ordered Two-way Table by Classification of the Individuals and Attributes. Cornell University, Ithaca, NY.
- Hill M.O. & Šmilauer P. (2005): TWINSPAN for Windows version 2.3. Centre for Ecology and Hydrology & University of South Bohemia, Huntingdon & České Budějovice.
- Hill M.O. & Gauch H.G. (1980): Detrended correspondence analysis: An improved ordination technique. Vegetatio, 42: 47-58.
- Kaufman L. & Rousseeuw P. J. (2009): Finding groups in data: an introduction to cluster analysis (Vol. 344). John Wiley & Sons.
- Macnaughton-Smith P., Williams W. T., Dale M. B. & Mockett L. G. (1964): Dissimilarity analysis: a new technique of hierarchical sub-division. Nature, 202: 1034-1035.
- Mueller-Dombois D. & Ellenberg H. (1974): Aims and Methods of Vegetation Ecology. John Wiley & Sons, New York, Chichester, Brisbane, Toronto.
- Roleček J., Tichý L., Zelený D. & Chytrý M. (2009): Modified TWINSPAN classification in which the hierarchy respects cluster heterogeneity. Journal of Vegetation Science, 20:596-602.
- Schaminée J.H.J & Hennekens S.M. (2001): TURBOVEG, MEGATAB und SYNBIOSYS: neue Entwicklungen in der Pflanzensoziologie. Ber. Reinhold-Tüxen Ges., 13: 21-34.
- Tichý L. (2002): JUICE, software for vegetation classification. Journal of Vegetation Science, 13: 451-453.