{smcl}
{* *! version 1.6 03 Jan 2019}{...}
{vieweralsosee "" "--"}{...}
{vieweralsosee "Install invtmvnormal" "ssc install MVTNORM"}{...}
{vieweralsosee "Help invtmvnormal (if installed)" "help invtmvnormal"}{...}
{viewerjumpto "Syntax" "invtmvnormal##syntax"}{...}
{viewerjumpto "Description" "invtmvnormal##description"}{...}
{viewerjumpto "Options" "invtmvnormal##options"}{...}
{viewerjumpto "Remarks" "invtmvnormal##remarks"}{...}
{viewerjumpto "Examples" "invtmvnormal##examples"}{...}
{title:Title}
{phang}
{bf:invtmvnormal} {hline 2} Quantiles of the Truncated Multivariate Normal Distribution
{marker syntax}{...}
{title:Syntax}
{p 8 17 2}
{cmdab:invtmvnormal}{cmd:,} [{it:options}]
{synoptset 20 tabbed}{...}
{synopthdr}
{synoptline}
{syntab:Required}
{synopt:{opt p(#)}} Probability.{p_end}
{synopt:{opt lowert:runcation(numlist miss)}} Vector of lower truncation points.{p_end}
{synopt:{opt uppert:runcation(numlist miss)}} Vector of upper truncation points.{p_end}
{break}
{syntab:Optional}
{synopt:{opt me:an(numlist)}} Mean vector.{p_end}
{synopt:{opt s:igma(string)}} Covariance matrix.{p_end}
{synopt:{opt t:ail(string)}} Which type of quantile should be computed.{p_end}
{synopt:{opt max:_iter(string)}} Maximum number of allowed iterations in the root-finding algorithm.{p_end}
{synopt:{opt tol:erance(string)}} Desired tolerance in the root-finding algorithm.{p_end}
{synopt:{opt int:egrator(string)}} Method for evaluating required multivariate normal distribution functions.{p_end}
{synopt:{opt shi:fts(#)}} Number of shifts of the Quasi-Monte Carlo integration algorithm.{p_end}
{synopt:{opt sam:ples(#)}} Number of samples in each shift of the Quasi-Monte Carlo integration algorithm.{p_end}
{synoptline}
{p2colreset}{...}
{p 4 6 2}
{marker description}{...}
{title:Description}
{pstd}
{cmd:invtmvnormal} computes equicoordinate quantiles of a specified truncated multivariate normal distribution through Brent's root finding algorithm (Brent, 1973).
Arbitrary means, covariance matrices, and truncation points are supported.
{p_end}
{pstd}
For one- and two-dimensional multivariate normal distributions it makes use of the functionality provided by {help invnormal} and {help binormal} respectively. For multivariate normal distributions of dimension three or more, support is available to evaluate requisite multivariate normal distribution functions using the approaches of both {help pmvnormal} and {help mvnormalcv} (see option {opt int:egrator}).
{marker options}{...}
{title:Options}
{dlgtab:Required}
{phang}
{opt p(#)} A real giving the desired tail probability. It must be between zero and one inclusive.
{p_end}{break}{phang}
{opt lowert:runcation(numlist miss)} A numlist (vector) giving the lower truncation points for the truncated multivariate normal distribution under consideration.
Its length must be equal to the length of option {opt low:er}.
Missing values are allowed, as . can be used to indicated a value is -Infinity.
{p_end}{break}{phang}
{opt uppert:runcation(numlist miss)} A numlist (vector) giving the upper truncation points for the truncated multivariate normal distribution under consideration.
Its length must be equal to the length of option {opt low:er}, and the elements of {opt uppert:runcation} must be strictly greater than the corresponding elements in {opt lowert:runcation}.
Missing values are allowed, as . can be used to indicated a value is +Infinity.
{dlgtab:Optional}
{phang}
{opt me:an(numlist)} A numlist (vector) giving the mean of the non-truncated multivariate normal distribution from which the truncated multivariate normal distribution under consideration is constructed.
If specified, its length must be equal to the length of option {opt lowert:runcation}.
If left unspecified, it will default internally to the zero vector of the correct length.
{p_end}{break}{phang}
{opt s:igma(string)} A string (matrix) giving the covariance matrix of the non-truncated multivariate normal distribution from which the truncated multivariate normal distribution under consideration is constructed.
If specified, It must be symmetric positive-definite, and its number of rows (equivalently columns) must be equal to the length of option {opt lowert:runcation}.
If left unspecified, it will default internally to the identity matrix of the correct dimension.
{p_end}{break}{phang}
{opt t:ail(string)} A string specifying which type of quantile should be computed.
It must be either lower, upper, or both.
lower gives the x such that P(X <= x) = p, upper such that P(X >= x) = p, and both such that P(-x <= X <= x) = p.
It left unspecified, it defaults internally to lower.
{p_end}{break}{phang}
{opt max:_iter(#)} An integer giving the maximum allowed number of iterations in the root-finding algorithm.
It must be greater than or equal to one.
It defaults to 1000000.
{p_end}{break}{phang}
{opt tol:erance(#)} A real giving the desired tolerance in the root-finding algorithm.
It must be strictly positive.
It defaults to 0.000001.
{p_end}{break}{phang}
{opt int:egrator(string)} A string giving the method for evaluating the required multivariate normal distribution functions.
It must be either pmvnormal (to use the method from {help pmvnormal}) or mvnormal (to use {help mvnormalcv} as released in Stata 15).
If left unspecified, it will default internally to pmvnormal.
{p_end}{break}{phang}
{opt shi:fts(#)} Only used when {opt int:egrator} is set to pmvnormal.
It is then an integer giving the number of shifts to use in the Quasi-Monte Carlo integration algorithm.
It must be strictly positive.
Theoretically, increasing {opt shi:fts} will increase run-time but reduce the error in the returned value of the quantile.
It defaults to 12.
{p_end}{break}{phang}
{opt sam:ples(#)} Only used when {opt int:egrator} is set to pmvnormal.
It is then an integer giving the number of samples to use in each shift of the Quasi-Monte Carlo integration algorithm.
It must be strictly positive.
Theoretically, increasing {opt sam:ples} will increase run-time but reduce the error in the returned value of the quantile.
It defaults to 1000.
{p_end}
{marker examples}{...}
{title:Examples}
/// Example 1: Without specifying {opt me:an} or {opt s:igma}
{phang}
{stata invtmvnormal, p(0.95) lowertruncation(-2, -2, -2) uppertruncation(2, 2, 2)}
/// Example 2: Specifying {opt me:an}, {opt s:igma}, and {opt t:ail}
{phang}
{stata mat Sigma = (1, 0.5, 0.5 \ 0.5, 1, 0.5 \ 0.5, 0.5, 1)}
{phang}
{stata invtmvnormal, p(0.95) lowertruncation(-2, -2, -2) uppertruncation(2, 2, 2) mean(1, 1, 1) sigma(Sigma) tail(upper)}
{marker results}{...}
{title:Stored results}
{pstd}
{cmd:invtmvnormal} stores the following in {cmd:r()}:
{synoptset 20 tabbed}{...}
{p2col 5 20 24 2: scalars:}{p_end}
{synopt:{cmd:r(quantile)}} Computed quantile.{p_end}
{synopt:{cmd:r(error)}} Estimated error in the computed quantile.{p_end}
{synopt:{cmd:r(flag)}} Flag indicating the progress of the root-finding algorithm. Non-zero values indicate a potential problem.{p_end}
{synopt:{cmd:r(fquantile)}} Value of the objective function at the computed quantile.{p_end}
{synopt:{cmd:r(iterations)}} Number of iterations used by the root-finding algorithm.{p_end}
{title:Authors}
{p}
Dr Michael J Grayling
Population Health Sciences Institute, Newcastle University, UK
Email: {browse "michael.grayling@newcastle.ac.uk":michael.grayling@newcastle.ac.uk}
Prof Adrian P Mander
Centre for Trials Research, Cardiff University, Cardiff, UK
{title:See also}
{bf:References:}
{phang}
Brent R (1973) {it:Algorithms for minimization without derivatives}. Prentice-Hall: New Jersey, US.
{phang}
Genz A, Bretz F (2009) {it:Computation of multivariate normal and t probabilities}. Lecture Notes in Statistics, Vol 195. Springer-Verlag: Heidelberg, Germany.
{phang}
Gibson GJ, Glasbey CA, Elston DA (1994) Monte Carlo evaluation of multivariate normal integrals and sensitivity to variate ordering.
In {it:Advances in numerical methods and applications}, ed Dimov IT, Sendov B, Vassilevski PS, 120-6. River Edge: World Scientific Publishing.
{phang}
Grayling MJ, Mander AP (2018) {browse "https://www.stata-journal.com/article.html?article=st0542":Calculations involving the multivariate normal and multivariate t distributions with and without truncation}. {it:Stata J} {bf:18}(4){bf::}826-43.
{phang}
Tong YL (2012) {it:The multivariate normal distribution}. Springer-Verlag: New York, US.
{bf:Related commands:}
{help mvtnorm} (for an overview of the functionality provided by {bf:mvtnorm})
Multivariate normal distribution
{help drawnorm} (an official Stata command for multivariate normal random deviates)
{help invmvnormal} ({bf:mvtnorm} command for multivariate normal quantiles)
{help lnmvnormalden} (an official Stata command for multivariate normal densities)
{help mvnormalcv} (an official Stata command for the multivariate normal distribution function)
{help mvnormalden} ({bf:mvtnorm} command for multivariate normal densities)
{help pmvnormal} ({bf:mvtnorm} command for the multivariate normal distribution function)
{help rmvnormal} ({bf:mvtnorm} command for multivariate normal random deviates)
Multivariate t distribution
{help invmvt} ({bf:mvtnorm} command for multivariate t quantiles)
{help mvtden} ({bf:mvtnorm} command for multivariate t densities)
{help mvt} ({bf:mvtnorm} command for the multivariate t distribution function)
{help rmvt} ({bf:mvtnorm} command for multivariate t random deviates)
Truncated multivariate normal distribution
{help invtmvnormal} ({bf:mvtnorm} command for truncated multivariate normal quantiles)
{help rtmvnormal} ({bf:mvtnorm} command for truncated multivariate normal random deviates)
{help tmvnormal} ({bf:mvtnorm} command for the truncated multivariate normal distribution function)
{help tmvnormalden} ({bf:mvtnorm} command for truncated multivariate normal densities)
Truncated multivariate t distribution
{help invtmvt} ({bf:mvtnorm} command for truncated multivariate t quantiles)
{help rtmvt} ({bf:mvtnorm} command for truncated multivariate t random deviates)
{help tmvt} ({bf:mvtnorm} command for the truncated multivariate t distribution function)
{help tmvtden} ({bf:mvtnorm} command for truncated multivariate t densities)