{smcl}
{* 22jun2006}{...}
{cmd:help mata mm_colvar()}
{hline}

{title:Title}

{pstd}
{bf:mm_colvar() -- Variance, by column}


{title:Syntax}

{p 8 12 2}
{it:real rowvector}
{cmd:mm_colvar(}{it:X} [{cmd:,} {it:w}]{cmd:)}

{p 8 12 2}
{it:real matrix}{bind:   }
{cmd:mm_meancolvar(}{it:X} [{cmd:,} {it:w}]{cmd:)}


{p 8 12 2}
{it:real matrix}{bind:   }
{cmd:mm_variance0(}{it:X} [{cmd:,} {it:w}]{cmd:)}

{p 8 12 2}
{it:real matrix}{bind:   }
{cmd:mm_meanvariance0(}{it:X} [{cmd:,} {it:w}]{cmd:)}


{p 8 12 2}
{it:real matrix}{space 4}{cmd:mm_mse(}{it:X}{cmd:,} {it:w}{cmd:,} {it:mu}{cmd:)}

{p 8 12 2}
{it:real rowvector} {cmd:mm_colmse(}{it:X}{cmd:,} {it:w}{cmd:,} {it:mu}{cmd:)}


{p 8 12 2}
{it:real matrix}{space 4}{cmd:mm_sse(}{it:X}{cmd:,} {it:w}{cmd:,} {it:mu}{cmd:)}

{p 8 12 2}
{it:real rowvector} {cmd:mm_colsse(}{it:X}{cmd:,} {it:w}{cmd:,} {it:mu}{cmd:)}


{p 4 4 2}
where

{p 12 16 2}
{it:X}:  {it:real matrix X} (rows are observations, columns variables)

{p 12 16 2}
{it:w}:  {it:real colvector w}

{p 11 16 2}
{it:mu}:  {it:real rowvector mu}


{title:Description}

{pstd}
{cmd:mm_colvar(}{it:X}{cmd:,} {it:w}{cmd:)}
returns the variance of each column of {it:X}. Essentially,

        {cmd:mm_colvar(}{it:X}{cmd:,} {it:w}{cmd:)} = {cmd:diagonal(variance(}{it:X}{cmd:,} {it:w}{cmd:))'}

{pstd}
See help for {helpb mf_mean:mean()}. However,
{cmd:mm_colvar()} does not compute the covariances and is
therefore much faster than {cmd:diagonal(variance())} if {it:X}
contains more than one column. Furthermore, note
that {cmd:mm_colvar()} omits missing values in {it:X} by
column, whereas {cmd:variance()} omits missing values casewise.

{pstd}
{cmd:mm_meancolvar(}{it:X}{cmd:,} {it: w}{cmd:)}
returns a matrix containing the mean and the variance of
each column of {it:X}. (Means in row one, variances in row two.)

{pstd}
{cmd:mm_variance0(}{it:X}{cmd:,} {it:w}{cmd:)}
returns the population variance matrix of {it:X}. {cmd:mm_variance0()}
differs from official Stata's {helpb mf_variance:variance()} (see help
for {helpb mf_mean:mean()}) in that
it divides the deviation cross products by N instead of N-1, where N
is the number of observations. Essentially,

        {cmd:mm_variance0(}{it:X}{cmd:,} {it:w}{cmd:)} = {cmd:variance(}{it:X}{cmd:,} {it:w}{cmd:)} * (N-1)/N

{pstd}
However, {cmd:mm_variance0()} also produces correct results if
N==1.

{pstd}
{cmd:mm_meanvariance0(}{it:X}{cmd:,} {it: w}{cmd:)}
returns {cmd:mean(}{it:X}{cmd:,}{it:w}{cmd:)\mm_variance0(}{it:X}{cmd:,}{it:w}{cmd:)}.

{pstd}
{cmd:mm_mse(}{it:X}{cmd:,} {it:w}{cmd:,} {it:mu}{cmd:)} computes
the mean squared errors matrix, where errors are defined as
{it:X}:-{it:mu}.

{pstd}
{cmd:mm_colmse()} computes mean squared errors by column.

{pstd}
{cmd:mm_sse()} and {cmd:mm_colsse()} compute the sum of squared errors.

{pstd}
{it:w} specifies the weights.
Specify {it:w} as 1 to obtain unweighted results.


{title:Remarks}

{pstd}
Examples for {cmd:mm_colvar()}:

        {com}: x = invnormal(uniform(10000,3))
        {res}
        {com}: mm_colvar(x, 1)
        {res}       {txt}          1             2             3
            {c TLC}{hline 43}{c TRC}
          1 {c |}  {res} 1.00018384   1.002621747   1.003480729{txt}  {c |}
            {c BLC}{hline 43}{c BRC}

        {com}: mm_meancolvar(x, 1)
        {res}       {txt}           1              2              3
            {c TLC}{hline 46}{c TRC}
          1 {c |}  {res}-.0024994158   -.0091972878   -.0035865732{txt}  {c |}
          2 {c |}  {res}  1.00018384    1.002621747    1.003480729{txt}  {c |}
            {c BLC}{hline 46}{c BRC}{txt}

{pstd}
The formula implemented in {cmd:mm_mse()} and {cmd:mm_colmse()} computes the
mean squared error as the sum of squared errors divided by N,
where N is the number of observations (or sum of weights if {it:w}!=1).


{title:Conformability}

{pstd}
{cmd:mm_colvar(}{it:X}{cmd:,} {it:w}{cmd:)}:
{p_end}
		{it:X}:  {it:n x k}
		{it:w}:  {it:n x 1} or {it:1 x 1}
	   {it:result}:  {it:1 x k}

{pstd}
{cmd:mm_meancolvar(}{it:X}{cmd:,} {it:w}{cmd:)}:
{p_end}
		{it:X}:  {it:n x k}
		{it:w}:  {it:n x} 1 or 1 {it:x} 1
	   {it:result}:  2 {it:x} k

{pstd}
{cmd:mm_variance0(}{it:X}{cmd:,} {it:w}{cmd:)}:
{p_end}
		{it:X}:  {it:n x k}
		{it:w}:  {it:n x 1} or {it:1 x 1}
	   {it:result}:  {it:k x k}

{pstd}
{cmd:mm_meanvariance0(}{it:X}{cmd:,} {it:w}{cmd:)}:
{p_end}
		{it:X}:  {it:n x k}
		{it:w}:  {it:n x} 1 or 1 {it:x} 1
	   {it:result}:  ({it:k}+1) {it:x} k

{pstd}
{cmd:mm_mse(}{it:X}{cmd:,} {it:w}{cmd:,} {it:mu}{cmd:)},
{cmd:mm_sse(}{it:X}{cmd:,} {it:w}{cmd:,} {it:mu}{cmd:)}:
{p_end}
		{it:X}:  {it:n x k}
		{it:w}:  {it:n x 1} or {it:1 x 1}
	       {it:mu}:  1 {it:x k}
	   {it:result}:  {it:k x k}

{pstd}
{cmd:mm_colmse(}{it:X}{cmd:,} {it:w}{cmd:,} {it:mu}{cmd:)},
{cmd:mm_colsse(}{it:X}{cmd:,} {it:w}{cmd:,} {it:mu}{cmd:)}:
{p_end}
		{it:X}:  {it:n x k}
		{it:w}:  {it:n x} 1 or 1 {it:x} 1
	       {it:mu}:  1 {it:x k}
	   {it:result}:  1 {it:x} k

{title:Diagnostics}

{pstd} {cmd:mm_variance0()}, {cmd:mm_meanvariance0()},
{cmd:mm_mse()}, and {cmd:mm_sse()} omit from calculation rows of
{it:X} or {it:w} that contain missing values (casewise deletion). If
all rows contain missing values, then the returned result contains
all missing values.

{pstd}
Contrarily, {cmd:mm_colvar()}, {cmd:mm_meancolvar()}, {cmd:mm_colmse()}, and
{cmd:mm_colsse()} omit missing
values by column (i.e. not casewise).


{title:Source code}

{pstd}
{help moremata_source##mm_colvar:mm_colvar.mata},
{help moremata_source##mm_variance0:mm_variance0.mata},
{help moremata_source##mm_mse:mm_mse.mata}

{title:Author}

{pstd} Ben Jann, University of Bern, jann@soz.unibe.ch


{title:Also see}

{psee}
Online:  help for
{bf:{help mf_mean:[M-5] mean()}},
{bf:{help moremata}}
{p_end}