DCA is correspondence analysis with the arch removed and the axes rescaled so that one unit is one standard deviation of species turnover — which is what makes the gradient length on each axis readable. Magic Stat runs it from a count table as a dialog.
Detrended correspondence analysis is correspondence analysis with two corrections applied during the iteration: the second axis is detrended by segments, which removes the arch that plain CA produces on long gradients, and each axis is rescaled so that one unit corresponds to one standard deviation of compositional turnover.
That rescaling is the reason DCA is still used: the gradient length it reports on each axis tells you whether the linear methods (PCA, RDA) or the unimodal methods (CA, CCA) suit your data. It is a diagnostic as much as an ordination.
The implementation is where software quietly diverges, and this is the part worth checking. Magic Stat's DCA is a port of the numerical engine of vegan::decorana — the reciprocal averaging with tridiagonal acceleration, segment detrending and rescaling — rather than an approximation that carries the same name.
Be aware, though: detrending is a strong and somewhat arbitrary treatment, and many ecologists now treat DCA as a legacy step. If you can justify CCA or NMDS for the analysis itself, prefer them. DCA is honest about the gradient; it is not the only way to be.
Input: a non-negative, complete count or abundance table with positive row sums. Negative values or empty rows are rejected with a clear message rather than a silent NaN.
Transformations available in the dialog: None · Z-score (standardise) · ln(1+x) · ln(x) · Hellinger · Chord · Chi-square · Square root · Range 0–1. Detrending and rescaling use the standard decorana parameters; there is no spin box for the segment count here.
DCA is CA plus segment detrending and rescaling. If all you want is a descriptive ordination of a table, CA is simpler. DCA earns its place when you need gradient lengths, or when the arch of CA is distorting the second axis.
It is debated. DCA was designed to fix a real artefact, but detrending is an aggressive treatment that can remove structure as well as the arch. Many reviewers will accept CCA or NMDS instead. The gradient length remains a useful, honest number.
No. But if you want vegan's other detrending options — polynomial detrending, different segment counts, downweighting of rare species — vegan::decorana exposes them and this dialog does not. For those, R is the better tool.
It is the range of site scores on that axis, expressed in standard deviations of turnover. As a rule of thumb, values above roughly four suggest unimodal methods (CA, CCA), while short axes are usually handled adequately by linear methods (RDA, PCA). It is a guide, not a switch.
No. Everything runs locally. The only automatic signal is an anonymous installation counter with no data in it.
Correspondence analysis → Canonical correspondence analysis → Redundancy analysis →