moremata

. ssc install moremata, replace

. net install moremata, replace from(https://raw.githubusercontent.com/benjann/moremata/master/)

01dec2020
- mm_quantile():
  o definitions 6-9 with weighted data and fw=0: the adjustments in the 
    denominator are now in terms of the sample size, not the sum of weights;
    the adjustments in the numerator are now relative to the weights, not
    absolute; the changes imply that results no longer depend on
    the scaling of the weights
  o definitions 3 with weighted data and fw=0: the rule for picking the lower
    or upper value in case of equal distance is now defined in terms of the
    indices of the observations, not the running sum of weights

23oct2020
- mm_mloc() and mm_mscale() added (robust M estimation of location and scale)
- mm_quantile() now also supports the computation of "high" quantile (def=0)
- mm_srswor() now has argument -alt- to select an alternative algorithm that
  is typically much faster than the default algorithm.
- mm_sample() now has an additional -alt- argument that is passed through to 
  mm_srswor()
- mm_sample() now has an additional -nowarn- argument that is passed through to 
  mm_upswor()

21oct2020
- mm_hl(), mm_qn(), and mm_mc() added (robust pairwise-based measures of
  location, scale, and skewness)

19oct2020
- mm_ranks() implicilty assumed weights to be nonnegative and produced meaningless
  results if mid!=0 was specified in presence of negative weights; this is fixed

03sep2020
- mm_median() had argument fw that did not do anything; the argument has now been
  removed

24aug2020
- mm_density() now returns error if bandwith cannot be determined 
  (e.g. if data is constant); function D.h() returns missing in this case

18aug2020
- function mm_kderiv_triweight() returned incorrect results; this is fixed

17aug2020
- mm_density():
  o new public functions D.K() and D.Kd() for observation-level evaluation of
    kernel function or derivative of kernel function using current settings 
    (including boundary correction)
  o some internal changes in organization of approximation estimator to avoid
    redundant computations in some situations

12aug2020
- mm_linbin() and mm_exactbin() are now implemented in terms of loops over
  grid points (instead of loops over observations) and are faster (and more 
  accurate) in large datasets
- new _mm_linbin() and _mm_exactbin() functions for use with sorted data
- new _mm_fastexactbin() for use with regular grid
- _mm_fastlinbin() is now slightly faster
- mm_ddens() and ISJ bandwidth selector in mm_density() now make use of 
  _mm_exactbin() and mm_fastexactbin()
- mm_density() now makes use _mm_linbin()
- D.bw() in mm_density() now allows argument adjust also in case of 
  user-provided bandwidth

11aug2020
- mm_density()
  o applied some renaming: D.bwmethod() is now D.bw() (furthermore, D.bw()
    now returns the user bandwith, if set, instead of the bandwidth method)
    D.bcmethod() is now D.bc(); D.bwadjust() is now D.adjust()
  o D.support() without argument now returns (lb(), ub())
- mm_ddens() as well as ISJ bandwidth selector in mm_density(): now using exact
  binning as in code by Botev; exact binning leads to inaccurate results if the
  grid size is small, but the error vanishes with increasing grid size; linear
  binning is more precise for small grid sizes, but it leads to non-vanishing
  error at the boundaries (doubling the first and last grid count does not
  seem to help); mm_ddens() now uses default grid size of 2^14 (as in code by
  Botev); ISJ in mm_density() enforces a grid of at least 2^10

10aug2020
- new mm_ddens() function for diffusion density estimation
- mm_density():
  o ISJ bandwidth selector wrongly used grid size instead of number of obs when
    rescaling the bandwidth; this is fixed
  o increased padding of approximation grid to +/- 10% of data range (instead of
    +/- 5% percent)
  o D.n() now has an additional argument to set the padding proportion

07aug2020
- mm_density():
  o now using DPI if SJPI/ISJ fails
  o ISJ now uses same root-finding algorithm as SJPI
  o SJPI and DPI now compute the scale from the binned data
  o SJPI now uses min of sd and iqr as scale measure when computing the
    oversmoothed bandwidth; this is at odds with h_os(), but may add some
    robustness; furthermore, root finder now uses full precision
  o extension of automatic grid is now limited to 5% of range on either side
  o D.kernel() always selected gaussian; this is fixed
  o D.support(.,"",1) returned error; this is fixed

06aug2020
- new mm_density() funtion for (univariate) kernel density estimation
- new mm_minim() funtion for univariate minimization without derivatives

04aug2020
- new mm_prod()/mm_rowprod()/mm_colprod() funtions to compute products of 
  elements in a matrix
- new mm_seq() function to generate regular sequences

17jul2020
- improved quantile functions; underscore functions no longer assume weights
  to be nonzero and now allow multiple columns in P
- new mm_issorted() function

14jul2020
- mm_quantile() has been rewritten; it now supports all 9 quantile definitions
  from from Hyndman and Fan (1996); weights are supported for all 
  definitions; new argument -fw- requests treating the weights as frequency
  weights
  ***
  IMPORTANT CHANGE:
      argument -altdef- in mm_quantile() and mm_iqrange() has been replaced
      by argument -def- that can take on values 1 to 9; altdef!=0 in the 
      previous version is equivalent to def=6 in the new version
  ***
- new functions _mm_quantile(), _mm_median(), and _mm_iqrange() that assume
  sorted data
- new functions mm_unique(), mm_unique_tag(), mm_uniqrows(), 
  mm_uniqrows_tag() to obtain or tag unique values in a vector or unique
  rows in a matrix; mm_uniqrows() differs from official uniqrows() in that
  it has an option to determin the order in which the result is returned
- new functions _mm_nunique(), _mm_unique(), _mm_unique_tag(), 
  _mm_nuniqrows(), _mm_uniqrows(), and _mm_uniqrows_tag() to count, obtain,
  or tag unique values/rows without sorting the data
- function mm_ipolate() is now faster, especially if there are ties
- new _mm_ipolate() function that assumes sorted data
- new mm_fastipolate() function that assumes sorted and unique data 
- new mm_group() function for creating a group index
- new mm_sort()/mm_order() functions for stable sorting
- new mm_diff() function for lagged differences
- new mm_clip() function to clip/limit values in a matrix
- new mm_kderiv() function for kernel derivatives
- new mm_ecdf2()/_mm_ecdf2() functions that return the CDF at unique values
  of X
- argument -mid- in mm_ranks() did not make sense with ties=0 or ties=4; 
  this is fixed
- function mm_relrank() has been reqritten; it now has additional 
  arguments support breaking ties and to compute nonnormalized ranks
- new _mm_ecdf() function that assumes sorted data
- new _mm_ranks() function that assumes sorted data
- new _mm_relrank() function that assumes sorted data
- mm_ranks() now uses quad precision in Stata 10 or newer
- mm_ecdf(), mm_ranks(), and mm_relrank() now have separate help files
- mm_colrunsum() now has argument -missing- to treat missing values as missing
  (instead of zero) and argument -quad- to request quad precision in Sata 10 
  or newer
- mm_isconstant() now uses allof() instead of all() and is thus faster

17apr2020
- installation files added to GitHub distribution

21aug2019
- mm_ebal(): handling of collinearity/redundant constraints improved

04may2019
- strange problem caused by mm_ebal(): it left junk behind in memory; this
  had something to do with keeping an optimization object within a structure, 
  but then passing the structure as an argument to the optimization object; this
  fixed
- mm_greedy() added

26apr2019
- mm_ebal() added

30may2017
- mm_sqrt() added

01feb2017
- mm_regexr() added

01jun2015
- mm_pieces() now supports unicode (Stata 14) 

16may2014
- mm_finvert() now has optional argument to pass on to &f()

29jan2014
- mm_integrate_sr() and mm_integrate_sr38() added

19feb2009:
- mm_collapse() added

10feb2009:
- mm_rbinomial(): note added that Stata 10.1 provides -rbinomial()-
- mm_invtokens(): note added that Stata 10 provides -invtokens()-
- new mm_pieces() functiom using genuine Mata code instead of extended macro 
  funtion -: piece-

26mar2008:
- mm_gini() updated so that it correctly handles ties. (Results depended on 
  sort order in case of ties)

29feb2008:
- mm_cond() added

11jan2008
- redirection of colrunsum in Stata 10 improved; _mm_colrunsum10() now faster
  if only 1 column
- mm_invtokens() now also works with column vectors and has a -noclean- option
- the default algorithm in mm_quantile() had precision problems if
  noninteger weights were specified
- mm_quantile() now properly handles zero weights
- mm_mgof() now displays progress dots
- mm_mgof():
  - error message in cases where noninteger f is not allowed
  - mc method now rounds sum(f) to the nearest integer to prevent
    sampling (n-1) obs in case of imprecision

29aug2007
- mm_matlist() added
- mm_colrunsum() now redirects itself to runningsum() if used in Stata 10

07aug2007
- mm_cauchy() functions added
- mm_colrunsum() now no longer uses the mean update formula; the mean update formula
  is problematic with integers
- mm_ranks() now has a normalize option (so that max(ecdf/relrank) is exactly 1)
  mm_gini(), mm_ecdf(), mm_relrank() updated
- linbin/fastlinbin/exactbin now support data outside of grid
- mm_ranks() has new syntax: new -mid- option for half-step (midpoint) method
  (replaces method==5); mm_relrank() now also has the mid option

27jun2007
- mm_benford() added
- mm_upswor() now has a -nowarn- option
- mm_ranks() changed (method=5 introduced; adjust removed; __mm_ranks() 
  for sorted data)
- mm_relrank() now based on mm_ranks()
- mm_nunique did not work with 'string rowvector' (because of transposeonly()); 
  this is fixed
- mm_freq2() and _mm_freq2() added; _mm_freq() added
- mm_freq() now allows matrix as input
- mm_nuniqrows() is computed slightly differently now (faster if x has many 
  columns)
- _mm_panels() is faster now
- mm_isconstant() added
- _mm_strexpand() added
- bug with single quotes in strings with mm_pieces() fixed
- mm_subset() etc. added
- mm_mgof() added
- mm_colrunsum(x) now works again if rows(x)==0 (the bug has been introduced
  on 12apr2007)
- mm_which() now works if nothing is selected from a scalar

12apr2007
- mm_colrunsum() now uses the mean-update formula

05apr2007
- mm_pieces(), mm_npieces(), and _mm_pieces() added
- mm_kern.mata: makes use of new capability of findexternal() to find
  functions; default kernel now epan2

03aug2006
- plot() added

13jul2006
- polint() added
- kernel integrals for xK(x) and x^2K(x) added
- slight changes to ipolate()
- mm_nrroot, mm_finvert added
- mm_locate, mm_hunt added
- mm_root() added
- fixed bug in mm_quantile
- mm_kern: kernel functions added
- default for m in makegrid() now 512 (previous: 401)
- fixed bug with missings in variance0, mse, sse
- w optional in quantile, iqrange, median, ecdf, ranks, freq, gini,
  histogram
- P optional in quantile
- g optional in histogram
- mm_bs() and mm_jk() added
- callf() added
- slight change to mm_panels: info1 will be filled even if X1 is
  absent; info1 will contain two columns if Y==. or void
- sse(), colsse() added
- mse(), colmse() added
- expand(), repeat() added
- fw option deleted from linbin(), fastlinbin(), and exactbin()
- nobs() added
- nobs() now used in histogram()
- quantile():
  * weighted version for altdef (only frequency weights)
  * speed improvements for unweighted algorithms
- quantile() now has an altdef option (interpolation)
- rank() now hat ties==4 option (order ties by w)
- p in quantile(x,w,p) may now be matrix
- q in relrank(x,w,q) may now be matrix

23may2006
- quantile, median, iqrange, ecdf, relrank, ranks now work with
  matrix X (statistics are computed for each column of X)
- gini now works with matrix X (gini of each column of X)
- fixed bug with with mm_sample() if stratified and n==0
- mm_gini() added
- mm_(mean)variance0(), mm_(mean)colvar() added
- mm_rank() now has adjust option
- mm_panels() etc: input now transmorphic vector
- mm_nunique, mm_nuniqrows added
- mm_ranks() added, mm_ecdf() now in terms of mm_ranks()
- mm_npanels() added, mm_panels() can now be used with void strata
   and void cluster
- mm_ipolate has new syntax (and is faster in most applications)
  (extrapolation not supported anymore; now using closest extremes)
- mm_fastlinbin() added

14apr2006
- bug fixed in mm_sample() (nn[i] rather than n)
- bug fixed in mm_sample() (strata[i,2] rather than cluster[i])
- declarations fixed in rbinomial, cebinomial, outsheet
- stable sort order in -mm_linbin()- and -mm_exactbin()-

01apr2006
- mm_unorder2(), mm_jumble2(), mm__jumble2() added
- mm_sample() (etc.) added
- mm_outsheet(): append/replace option
- mm_panels() added
- relrank(), ecdf(): range(1,I,1) changed to (1::I)
- freq() added
- cut() added
- rbinomial() and cebinomial() added
- posof() function added
- insheet and infile:
  * much faster now (code based on cat() version 2)
  * now support reading specific range of file (line1-line2)

15sep2005
- released on SSC