For classical statistics, I'd recommend GLM.jl over MultivariateStatsjl. The former is written by a statistician (and R Core member) and the latter is written by a machine learner who is very focused on speed. The main reason for this recommendation is not speed, but that GLM.jl is quite similar to the lm/glm functions in R when used with the DataFrames.jl package.

The difference in precision is the difference between using a Cholesky and a QR based solution for the least squares problem. Personally, I'm not too worried about the loss of precision in the Cholesky solution because the statistical error is so much larger than the numerical error.

On Sat, Jan 16, 2016 at 12:38 PM, Riccardo (Jack) Lucchetti <> wrote:
On Sat, 16 Jan 2016, Sven Schreiber wrote:

Am 15.01.2016 um 20:39 schrieb Allin Cottrell:
Following up Jack's comment at

in current git there's a basic "preview" of Julia support in gretl.


# NIST's certified coefficient values
matrix nist_b = {-3482258.63459582, 15.0618722713733,
    -0.358191792925910E-01, -2.02022980381683,
    -1.03322686717359, -0.511041056535807E-01,

Since I don't have it installed yet, could you comment on whether the
results match (between gretl/Julia/NIST)?

These are the results I get

Log-relative errors, Longley coefficients:

       gretl       julia
      12.228      8.0224
      10.920      7.5300
      11.797      7.5697
      12.528      8.1421
      13.169      8.3801
      11.770      7.2368
      12.235      8.0333

Column means
      12.092      7.8449


So it would seem that the MultivariateStats julia module leaves a bit to be desired for the moment, at lest in terms of precision.

  Riccardo (Jack) Lucchetti
  Dipartimento di Scienze Economiche e Sociali (DiSES)

  Università Politecnica delle Marche
  (formerly known as Università di Ancona)

Gretl-devel mailing list