2011-09-19 17:01:24 +02:00
function [ dLIK , dlik , a , Pstar ] = missing_observations_kalman_filter_d ( data_index , number_of_observations , no_more_missing_observations , ...
2017-05-16 15:10:20 +02:00
Y , start , last , ...
a , Pinf , Pstar , ...
kalman_tol , diffuse_kalman_tol , riccati_tol , presample , ...
T , R , Q , H , Z , mm , pp , rr )
2011-09-19 17:01:24 +02:00
% Computes the diffuse likelihood of a state space model when some observations are missing.
%
2017-05-16 15:10:20 +02:00
% INPUTS
2011-09-19 17:01:24 +02:00
% data_index [cell] 1*smpl cell of column vectors of indices.
% number_of_observations [integer] scalar.
% no_more_missing_observations [integer] scalar.
% Y [double] pp*smpl matrix of (detrended) data, where pp is the number of observed variables.
% start [integer] scalar, first observation.
% last [integer] scalar, last observation.
2021-01-18 17:33:21 +01:00
% a [double] mm*1 vector, levels of the predicted initial state variables (E_{0}(alpha_1)).
2011-09-19 17:01:24 +02:00
% Pinf [double] mm*mm matrix used to initialize the covariance matrix of the state vector.
% Pstar [double] mm*mm matrix used to initialize the covariance matrix of the state vector.
% kalman_tol [double] scalar, tolerance parameter (rcond).
% riccati_tol [double] scalar, tolerance parameter (riccati iteration).
2017-05-16 15:10:20 +02:00
% presample [integer] scalar, presampling if strictly positive.
2011-09-19 17:01:24 +02:00
% T [double] mm*mm matrix, transition matrix in the state equations.
% R [double] mm*rr matrix relating the structural innovations to the state vector.
% Q [double] rr*rr covariance matrix of the structural innovations.
2017-05-16 15:10:20 +02:00
% H [double] pp*pp covariance matrix of the measurement errors (if H is equal to zero (scalar) there is no measurement error).
2011-09-19 17:01:24 +02:00
% Z [double] pp*mm matrix, selection matrix or pp linear independant combinations of the state vector.
% mm [integer] scalar, number of state variables.
% pp [integer] scalar, number of observed variables.
2017-05-16 15:10:20 +02:00
% rr [integer] scalar, number of structural innovations.
%
% OUTPUTS
2011-09-19 17:01:24 +02:00
% dLIK [double] scalar, MINUS loglikelihood
% dlik [double] vector, density of observations in each period.
2021-01-18 17:33:21 +01:00
% a [double] mm*1 vector, current estimate of the state vector tomorrow (E_{T}(alpha_{T+1})).
2017-05-16 15:10:20 +02:00
% Pstar [double] mm*mm matrix, covariance matrix of the states.
%
2011-09-19 17:01:24 +02:00
% REFERENCES
% See "Filtering and Smoothing of State Vector for Diffuse State Space
2017-05-16 15:10:20 +02:00
% Models", S.J. Koopman and J. Durbin (2003), in Journal of Time Series
% Analysis, vol. 24(1), pp. 85-98.
2016-06-14 17:29:31 +02:00
% and
% Durbin/Koopman (2012): "Time Series Analysis by State Space Methods", Oxford University Press,
% Second Edition, Ch. 5 and 7.2
2011-09-19 17:01:24 +02:00
2017-05-16 15:10:20 +02:00
%
2022-04-13 13:15:19 +02:00
% Copyright © 2004-2021 Dynare Team
2011-09-19 17:01:24 +02:00
%
% This file is part of Dynare.
%
% Dynare is free software: you can redistribute it and/or modify
% it under the terms of the GNU General Public License as published by
% the Free Software Foundation, either version 3 of the License, or
% (at your option) any later version.
%
% Dynare is distributed in the hope that it will be useful,
% but WITHOUT ANY WARRANTY; without even the implied warranty of
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
% GNU General Public License for more details.
%
% You should have received a copy of the GNU General Public License
2021-06-09 17:33:48 +02:00
% along with Dynare. If not, see <https://www.gnu.org/licenses/>.
2011-09-19 17:01:24 +02:00
% Get sample size.
smpl = last - start + 1 ;
% Initialize some variables.
dF = 1 ;
2021-05-02 17:07:12 +02:00
isqvec = false ;
if ndim ( Q ) > 2
Qvec = Q ;
Q = Q ( : , : , 1 ) ;
isqvec = true ;
end
2011-09-19 17:01:24 +02:00
QQ = R * Q * transpose ( R ) ; % Variance of R times the vector of structural innovations.
t = start ; % Initialization of the time index.
dlik = zeros ( smpl , 1 ) ; % Initialization of the vector gathering the densities.
dLIK = Inf ; % Default value of the log likelihood.
oldK = Inf ;
2012-06-19 16:35:44 +02:00
if isequal ( H , 0 )
H = zeros ( pp , pp ) ;
end
2012-07-01 15:19:36 +02:00
s = 0 ;
2012-06-19 16:35:44 +02:00
2015-04-08 15:49:12 +02:00
while rank ( Pinf , diffuse_kalman_tol ) && ( t < = last )
2011-09-19 17:01:24 +02:00
s = t - start + 1 ;
d_index = data_index { t } ;
2021-05-02 17:07:12 +02:00
if isqvec
QQ = R * Qvec ( : , : , t + 1 ) * transpose ( R ) ;
end
2011-09-19 17:01:24 +02:00
if isempty ( d_index )
2016-06-14 17:29:31 +02:00
%no observations, propagate forward without updating based on
%observations
2011-09-19 17:01:24 +02:00
a = T * a ;
Pstar = T * Pstar * transpose ( T ) + QQ ;
Pinf = T * Pinf * transpose ( T ) ;
else
2016-06-14 17:29:31 +02:00
ZZ = Z ( d_index , : ) ; %span selector matrix
2017-05-16 15:10:20 +02:00
v = Y ( d_index , t ) - ZZ * a ; %get prediction error v^(0) in (5.13) DK (2012)
2016-06-14 17:29:31 +02:00
Finf = ZZ * Pinf * ZZ ' ; % (5.7) in DK (2012)
2017-05-16 15:10:20 +02:00
if rcond ( Finf ) < diffuse_kalman_tol %F_{\infty,t} = 0
2016-06-14 17:29:31 +02:00
if ~ all ( abs ( Finf ( : ) ) < diffuse_kalman_tol ) %rank-deficient but not rank 0
2017-05-16 15:10:20 +02:00
% The univariate diffuse kalman filter should be used.
2011-09-19 17:01:24 +02:00
return
2016-06-14 17:29:31 +02:00
else %rank of F_{\infty,t} is 0
Fstar = ZZ * Pstar * ZZ ' + H ( d_index , d_index ) ; % (5.7) in DK (2012)
if rcond ( Fstar ) < kalman_tol %F_{*} is singular
2011-09-19 17:01:24 +02:00
if ~ all ( abs ( Fstar ( : ) ) < kalman_tol )
% The univariate diffuse kalman filter should be used.
return
2016-06-14 17:29:31 +02:00
else %rank 0
2017-05-16 15:10:20 +02:00
%pathological case, discard draw
2016-06-19 17:13:04 +02:00
return
2011-09-19 17:01:24 +02:00
end
else
iFstar = inv ( Fstar ) ;
dFstar = det ( Fstar ) ;
2016-10-28 11:40:35 +02:00
Kstar = Pstar * ZZ ' * iFstar ; %(5.15) of DK (2012) with Kstar=T^{-1}*K^(0)
2016-06-14 17:29:31 +02:00
dlik ( s ) = log ( dFstar ) + v ' * iFstar * v + length ( d_index ) * log ( 2 * pi ) ; %set w_t to bottom case in bottom equation page 172, DK (2012)
Pinf = T * Pinf * transpose ( T ) ; % (5.16) DK (2012)
2016-10-28 11:40:35 +02:00
Pstar = T * ( Pstar - Pstar * ZZ ' * Kstar ' ) * T ' + QQ ; % (5.17) DK (2012) with L_0 plugged in
2016-06-14 17:29:31 +02:00
a = T * ( a + Kstar * v ) ; % (5.13) DK (2012)
2011-09-19 17:01:24 +02:00
end
end
else
2016-06-16 18:49:34 +02:00
dlik ( s ) = log ( det ( Finf ) ) + length ( d_index ) * log ( 2 * pi ) ; %set w_t to top case in bottom equation page 172, DK (2012)
2011-09-19 17:01:24 +02:00
iFinf = inv ( Finf ) ;
Kinf = Pinf * ZZ ' * iFinf ;
2016-06-14 17:29:31 +02:00
%see notes in kalman_filter_d.m for details of computations
Fstar = ZZ * Pstar * ZZ ' + H ( d_index , d_index ) ; %(5.7) DK(2012)
2016-10-29 15:00:37 +02:00
Kstar = ( Pstar * ZZ ' - Kinf * Fstar ) * iFinf ; %(5.12) DK(2012); note that there is a typo in DK (2003) with "+ Kinf" instead of "- Kinf", but it is correct in their appendix
2016-06-14 17:29:31 +02:00
Pstar = T * ( Pstar - Pstar * ZZ ' * Kinf ' - Pinf * ZZ ' * Kstar ' ) * T ' + QQ ; %(5.14) DK(2012)
Pinf = T * ( Pinf - Pinf * ZZ ' * Kinf ' ) * T ' ; %(5.14) DK(2012)
a = T * ( a + Kinf * v ) ; %(5.13) DK(2012)
2011-09-19 17:01:24 +02:00
end
end
t = t + 1 ;
end
if t == ( last + 1 )
2017-05-16 15:10:20 +02:00
warning ( [ ' kalman_filter_d: There isn' ' t enough information to estimate the initial conditions of the nonstationary variables. The diffuse Kalman filter never left the diffuse stage.' ] ) ;
2011-09-19 17:01:24 +02:00
dLIK = NaN ;
return
end
dlik = . 5 * dlik ( 1 : s ) ;
2012-01-22 22:40:46 +01:00
dLIK = sum ( dlik ( 1 + presample : end ) ) ;