Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Markovchain Package

Download as pdf or txt
Download as pdf or txt
You are on page 1of 70

The markovchain Package: A Package for Easily

Handling Discrete Markov Chains in R

Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav

Abstract
The markovchain package aims to fill a gap within the R framework providing S4
classes and methods for easily handling discrete time Markov chains, homogeneous and
simple inhomogeneous ones as well as continuous time Markov chains. The S4 classes
for handling and analysing discrete and continuous time Markov chains are presented, as
well as functions and method for performing probabilistic and statistical analysis. Finally,
some examples in which the package’s functions are applied to Economics, Finance and
Natural Sciences topics are shown.

Keywords: discrete time Markov chains, continuous time Markov chains, transition matrices,
communicating classes, periodicity, first passage time, stationary distributions.

1. Introduction
Markov chains represent a class of stochastic processes of great interest for the wide spectrum
of practical applications. In particular, discrete time Markov chains (DTMC) permit to model
the transition probabilities between discrete states by the aid of matrices.Various R packages
deal with models that are based on Markov chains:

❼ msm (Jackson 2011) handles Multi-State Models for panel data.


❼ mcmcR (Geyer and Johnson 2013) implements Monte Carlo Markov Chain approach.
❼ hmm (Himmelmann and www.linhi.com 2010) fits hidden Markov models with covari-
ates.
❼ mstate fits ‘Multi-State Models based on Markov chains for survival analysis (de Wreede,
Fiocco, and Putter 2011).

Nevertheless, the R statistical environment (R Core Team 2013) seems to lack a simple package
that coherently defines S4 classes for discrete Markov chains and allows to perform probabilis-
tic analysis, statistical inference and applications. For the sake of completeness, markovchain
is the second package specifically dedicated to DTMC analysis, being DTMCPack (Nichol-
son 2013) the first one. Notwithstanding, markovchain package (Spedicato 2017) aims to
offer more flexibility in handling DTMC than other existing solutions, providing S4 classes
for both homogeneous and non-homogeneous Markov chains as well as methods suited to
perform statistical and probabilistic analysis.
The markovchain package depends on the following R packages: expm (Goulet, Dutang, Mae-
chler, Firth, Shapira, Stadelmann, and expm-developers@lists.R-forge.R-project.org 2013) to
2 markovchain package: discrete Markov chains in R

perform efficient matrices powers; igraph (Csardi and Nepusz 2006) to perform pretty plotting
of markovchain objects and matlab (Roebuck 2011), that contains functions for matrix ma-
nagement and calculations that emulate those within MATLAB environment. Moreover, other
scientific softwares provide functions specifically designed to analyze DTMC, as Mathematica
9 (Wolfram Research 2013b).
The paper is structured as follows: Section 2 briefly reviews mathematics and definitions re-
garding DTMC, Section 3 discusses how to handle and manage Markov chain objects within
the package, Section 4 and Section 5 show how to perform probabilistic and statistical model-
ling, while Section 6 presents some applied examples from various fields analyzed by means
of the markovchain package.

2. Review of core mathematical concepts

2.1. General Definitions


A DTMC is a sequence of random variables X1 , X2 , . . . , Xn , . . . characterized by the Markov
property (also known as memoryless property, see Equation 1). The Markov property states
that the distribution of the forthcoming state Xn+1 depends only on the current state Xn
and doesn’t depend on the previous ones Xn−1 , Xn−2 , . . . , X1 .

P r (Xn+1 = xn+1 |X1 = x1 , X2 = x2, ..., Xn = xn ) = P r (Xn+1 = xn+1 |Xn = xn ) . (1)

The set of possible states S = {s1 , s2 , ..., sr } of Xn can be finite or countable and it is named
the state space of the chain.
The chain moves from one state to another (this change is named either ‘transition’ or ‘step’)
and the probability pij to move from state si to state sj in one step is named transition
probability:

pij = P r (X1 = sj |X0 = si ) . (2)


(n)
The probability of moving from state i to j in n steps is denoted by pij = P r (Xn = sj |X0 = si ).
A DTMC is called time-homogeneous if the property shown in Equation 3 holds. Time
homogeneity implies no change in the underlying transition probabilities as time goes on.
P r (Xn+1 = sj |Xn = si ) = P r (Xn = sj |Xn−1 = si ) . (3)

If the Markov chain is time-homogeneous, then pij = P r (Xk+1 = sj |Xk = si ) and


(n)
pij = P r (Xn+k = sj |Xk = si ), where k > 0.
The probability distribution of transitions from one state to another can be represented into
a transition matrix P = (pij )i,j , where each element of position (i, j) represents the transition
probability pij . E.g., if r = 3 the transition matrix P is shown in Equation 4
 
p11 p12 p13
P =  p21 p22 p23  . (4)
p31 p32 p33
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 3

The distribution over the states


P can be written in the form of a stochastic row vector x (the
term stochastic means that i xi = 1, xi ≥ 0): e.g., if the current state of x is s2 , x = (0 1 0).
As a consequence, the relation between x(1) and x(0) is x(1) = x(0) P and, recursively, we get
x(2) = x(0) P 2 and x(n) = x(0) P n , n > 0.
DTMC are explained in most theory books on stochastic processes, see Brémaud (1999) and
Dobrow (2016) for example. Valuable references online available are: Konstantopoulos (2009),
Snell (1999) and Bard (2000).

2.2. Properties and classification of states


A state sj is said accessible from state si (written si → sj ) if a system starting in state si has
a positive probability to reach the state sj at a certain point, i.e., ∃n > 0 : pnij > 0. If both
si → sj and sj → si , then si and sj are said to communicate.
A communicating class is defined to be a set of states that communicate. A DTMC can be
composed by one or more communicating classes. If the DTMC is composed by only one
communicating class (i.e., if all states in the chain communicate), then it is said irreducible.
A communicating class is said to be closed if no states outside of the class can be reached
from any state inside it.
If pii = 1, si is defined as absorbing state: an absorbing state corresponds to a closed com-
municating class composed by one state only.
The canonic form of a DTMC transition matrix is a matrix having a block form, where the
closed communicating classes are shown at the beginning of the diagonal matrix.
A state si has period ki if any return to state si must occur in multiplies of ki steps, that is
ki = gcd {n : P r (Xn = si |X0 = si ) > 0}, where gcd is the greatest common divisor. If ki = 1
the state si is said to be aperiodic, else if ki > 1 the state si is periodic with period ki . Loosely
speaking, si is periodic if it can only return to itself after a fixed number of transitions ki > 1
(or multiple of ki ), else it is aperiodic.
If states si and sj belong to the same communicating class, then they have the same period ki .
As a consequence, each of the states of an irreducible DTMC share the same periodicity. This
periodicity is also considered the DTMC periodicity. It is possible to classify states according
to their periodicity. Let T x→x is the number of periods to go back to state x knowing that
the chain starts in x.

❼ A state x is recurrent if P (T x→x < +∞) = 1 (equivalently P (T x→x = +∞) = 0). In


addition:

1. A state x is null recurrent if in addition E(T x→x ) = +∞.


2. A state x is positive recurrent if in addition E(T x→x ) < +∞.
3. A state x is absorbing if in addition P (T x→x = 1) = 1.

❼ A state x is transient if P (T x→x < +∞) < 1 (equivalently P (T x→x = +∞) > 0).

It is possible to analyze the timing to reach a certain state. The first passage time (or hitting
time) from state si to state sj is the number Tij of steps taken by the chain until it arrives
for the first time to state sj , given that X0 = si . The probability distribution of Tij is defined
4 markovchain package: discrete Markov chains in R

by Equation 5

hij (n) = P r (Tij = n) = P r (Xn = sj , Xn−1 6= sj , . . . , X1 6= sj |X0 = si ) (5)

and can be found recursively using Equation 6, given that hij (n) = pij .

X
hij (n) = pik hkj (n−1) . (6)
k∈S−{sj }

A commonly used quantity related to h is its average value, i.e. the mean first passage time
P (n)
(also expected hitting time), namely h̄ij = n=1...∞ n hij .
If in the definition of the first passage time we let si = sj , we obtain the first return time
Ti = inf{n ≥ 1 : Xn = si |X0 = si }. A state si is said to be recurrent if it is visited infinitely
often, i.e., P r(Ti < +∞|X0 = si ) = 1. On the opposite, si is called transient if there is a
positive probability that the chain will never return to si , i.e., P r(Ti = +∞|X0 = si ) > 0.
Given a time homogeneous Markov chain with transition matrix P, a stationary
P distribution
z is a stochastic row vector such that z = z · P , where 0 ≤ zj ≤ 1 ∀j and j zj = 1.
If a DTMC {Xn } is irreducible and aperiodic, then it has a limit distribution and this distri-
bution is stationary. As a consequence, if P is the Pk k × k transition matrix of the chain and
z = (z1 , ..., zk ) is the eigenvector of P such that i=1 zi = 1, then we get

lim P n = Z, (7)
n→∞

where Z is the matrix having all rows equal to z. The stationary distribution of {Xn } is
represented by z.
(AB)
Given two absorbing states sA (source) and sB (sink), the committor probability qj is the
probability that a process starting in state si is absorbed in state sB (rather than sA ) (Noé,
Schütte, Vanden-Eijnden, Reich, and Weikl 2009). It can be computed via

(AB) (AB) (AB) (AB)


X
qj = Pjk qk with qA =0 and qB =1 (8)
k∋A,B

2.3. A short example


Consider the following numerical example. Suppose we have a DTMC with a set of 3 possible
states S = {s1 , s2 , s3 }. Let the transition matrix be:
 
0.5 0.2 0.3
P =  0.15 0.45 0.4  . (9)
0.25 0.35 0.4

In P , p11 = 0.5 is the probability that X1 = s1 given that we observed X0 = s1 is 0.5, and so
on.It is easy to see that the chain is irreducible since all the states communicate (it is made
by one communicating class only).
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 5

Suppose that the current state of the chain is X0 = s2 , i.e., x(0) = (0 1 0), then the probability
distribution of states after 1 and 2 steps can be computed as shown in Equations (10) and
(11).
 
0.5 0.2 0.3
x(1) = (0 1 0)  0.15 0.45 0.4  = (0.15 0.45 0.4) . (10)
0.25 0.35 0.4
 
0.5 0.2 0.3
x(n) = x(n−1) P → (0.15 0.45 0.4)  0.15 0.45 0.4  = (0.2425 0.3725 0.385) . (11)
0.25 0.35 0.4

If, f.e., we are interested in the probability of being in the state s3 in the second step, then
P r (X2 = s3 |X0 = s2 ) = 0.385.
6 markovchain package: discrete Markov chains in R

3. The structure of the package

3.1. Creating markovchain objects


The package is loaded within the R command line as follows:

R> library("markovchain")

The markovchain and markovchainList S4 classes (Chambers 2008) are defined within the
markovchain package as displayed:

Class "markovchain" [package "markovchain"]

Slots:

Name: states byrow transitionMatrix name


Class: character logical matrix character

Class "markovchainList" [package "markovchain"]

Slots:

Name: markovchains name


Class: list character

The first class has been designed to handle homogeneous Markov chain processes, while
the latter (which is itself a list of markovchain objects) has been designed to handle non-
homogeneous Markov chains processes.
Any element of markovchain class is comprised by following slots:

1. states: a character vector, listing the states for which transition probabilities are
defined.
2. byrow: a logical element, indicating whether transition probabilities are shown by row
or by column.
3. transitionMatrix: the probabilities of the transition matrix.
4. name: optional character element to name the DTMC.

The markovchainList objects are defined by following slots:

1. markovchains: a list of markovchain objects.


2. name: optional character element to name the DTMC.

The markovchain objects can be created either in a long way, as the following code shows
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 7

R> weatherStates <- c("sunny", "cloudy", "rain")


R> byRow <- TRUE
R> weatherMatrix <- matrix(data = c(0.70, 0.2, 0.1,
R+ 0.3, 0.4, 0.3,
R+ 0.2, 0.45, 0.35), byrow = byRow, nrow = 3,
R+ dimnames = list(weatherStates, weatherStates))
R> mcWeather <- new("markovchain", states = weatherStates, byrow = byRow,
R+ transitionMatrix = weatherMatrix, name = "Weather")

or in a shorter way, displayed below

R> mcWeather <- new("markovchain", states = c("sunny", "cloudy", "rain"),


R+ transitionMatrix = matrix(data = c(0.70, 0.2, 0.1,
R+ 0.3, 0.4, 0.3,
R+ 0.2, 0.45, 0.35), byrow = byRow, nrow = 3),
R+ name = "Weather")

When new("markovchain") is called alone, a default Markov chain is created.

R> defaultMc <- new("markovchain")

The quicker way to create markovchain objects is made possible thanks to the implemented
initialize S4 method that checks that:

❼ the transitionMatrix to be a transition matrix, i.e., all entries to be probabilities and


either all rows or all columns to sum up to one.
❼ the columns and rows names of transitionMatrix to be defined and to coincide with
states vector slot.

The markovchain objects can be collected in a list within markovchainList S4 objects as


following example shows.

R> mcList <- new("markovchainList", markovchains = list(mcWeather, defaultMc),


R+ name = "A list of Markov chains")

3.2. Handling markovchain objects


Table 1 lists which methods handle and manipulate markovchain objects.
The examples that follow shows how operations on markovchain objects can be easily per-
formed. For example, using the previously defined matrix we can find what is the probability
distribution of expected weather states in two and seven days, given the actual state to be
cloudy.

R> initialState <- c(0, 1, 0)


R> after2Days <- initialState * (mcWeather * mcWeather)
R> after7Days <- initialState * (mcWeather ^ 7)
R> after2Days
8 markovchain package: discrete Markov chains in R

Method Purpose
* Direct multiplication for transition matrices.
[ Direct access to the elements of the transition matrix.
== Equality operator between two transition matrices.
as Operator to convert markovchain objects into data.frame and
table object.
dim Dimension of the transition matrix.
names Equal to states.
names<- Change the states name.
name Get the name of markovchain object.
name<- Change the name of markovchain object.
plot plot method for markovchain objects.
print print method for markovchain objects.
show show method for markovchain objects.
sort sort method for markovchain objects.
states Name of the transition states.
t Transposition operator (which switches byrow slot value and modifies
the transition matrix coherently).

Table 1: markovchain methods for handling markovchain objects.

sunny cloudy rain


[1,] 0.39 0.355 0.255

R> round(after7Days, 3)

sunny cloudy rain


[1,] 0.462 0.319 0.219

A similar answer could have been obtained defining the vector of probabilities as a column
vector. A column - defined probability matrix could be set up either creating a new matrix
or transposing an existing markovchain object thanks to the t method.

R> initialState <- c(0, 1, 0)


R> after2Days <- (t(mcWeather) * t(mcWeather)) * initialState
R> after7Days <- (t(mcWeather) ^ 7) * initialState
R> after2Days

[,1]
sunny 0.390
cloudy 0.355
rain 0.255

R> round(after7Days, 3)

[,1]
sunny 0.462
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 9

cloudy 0.319
rain 0.219

The initial state vector previously shown can not necessarily be a probability vector, as the
code that follows shows:

R> fvals<-function(mchain,initialstate,n) {
R+ out<-data.frame()
R+ names(initialstate)<-names(mchain)
R+ for (i in 0:n)
R+ {
R+ iteration<-initialstate*mchain^(i)
R+ out<-rbind(out,iteration)
R+ }
R+ out<-cbind(out, i=seq(0,n))
R+ out<-out[,c(4,1:3)]
R+ return(out)
R+ }
R> fvals(mchain=mcWeather,initialstate=c(90,5,5),n=4)

i sunny cloudy rain


1 0 90.00000 5.00000 5.00000
2 1 65.50000 22.25000 12.25000
3 2 54.97500 27.51250 17.51250
4 3 50.23875 29.88063 19.88062
5 4 48.10744 30.94628 20.94628

Basic methods have been defined for markovchain objects to quickly get states and transition
matrix dimension.

R> states(mcWeather)

[1] "sunny" "cloudy" "rain"

R> names(mcWeather)

[1] "sunny" "cloudy" "rain"

R> dim(mcWeather)

[1] 3

Methods are available to set and get the name of markovchain object.

R> name(mcWeather)
10 markovchain package: discrete Markov chains in R

[1] "Weather"

R> name(mcWeather) <- "New Name"


R> name(mcWeather)

[1] "New Name"

Also it is possible to alphabetically sort the transition matrix:

R> markovchain:::sort(mcWeather)

New Name
A 3 - dimensional discrete Markov Chain defined by the following states:
cloudy, rain, sunny
The transition matrix (by rows) is defined as follows:
cloudy rain sunny
cloudy 0.40 0.30 0.3
rain 0.45 0.35 0.2
sunny 0.20 0.10 0.7

A direct access to transition probabilities is provided both by transitionProbability met-


hod and "[" method.

R> transitionProbability(mcWeather, "cloudy", "rain")

[1] 0.3

R> mcWeather[2,3]

[1] 0.3

The transition matrix of a markovchain object can be displayed using print or show methods
(the latter being less verbose). Similarly, the underlying transition probability diagram can
be plotted by the use of plot method (as shown in Figure 1) which is based on igraph
package (Csardi and Nepusz 2006). plot method for markovchain objects is a wrapper of
plot.igraph for igraph S4 objects defined within the igraph package. Additional parameters
can be passed to plot function to control the network graph layout. There are also diagram
and DiagrammeR ways available for plotting as shown in Figure 2. The plot function also uses
communicatingClasses function to separate out states of different communicating classes.
All states that belong to one class have same colour.

R> print(mcWeather)

sunny cloudy rain


sunny 0.7 0.20 0.10
cloudy 0.3 0.40 0.30
rain 0.2 0.45 0.35
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 11

Weather transition matrix

rain 0.35

0.45
0.2

0.3

0.1

cloudy 0.4
0.3
0.2
sunny 0.7

Figure 1: Weather example. Markov chain plot

R> show(mcWeather)

New Name
A 3 - dimensional discrete Markov Chain defined by the following states:
sunny, cloudy, rain
The transition matrix (by rows) is defined as follows:
sunny cloudy rain
sunny 0.7 0.20 0.10
cloudy 0.3 0.40 0.30
rain 0.2 0.45 0.35

Attaching package: ✬igraph✬

The following objects are masked from ✬package:stats✬:

decompose, spectrum

The following object is masked from ✬package:base✬:

union

If one would like to use the MmgraphR package (Adamopoulou 2018) to plot the transition
matric, the following code shows how to do:

R> library("MmgraphR")

Loading required package: TraMineR


12 markovchain package: discrete Markov chains in R

0.35
rain

0.45 0.1
0.2
0.3

0.4 0.2 0.7


cloudy sunny
0.3

Figure 2: Weather example. Markov chain plot with diagram

TraMineR stable version 2.0-10 (Built: 2019-01-15)

Website: http://traminer.unige.ch

Please type ✬citation("TraMineR")✬ for citation information.

Loading required package: colorspace

R> stochastic_matrix_to_plot <- as(mcWeather,"matrix")


R> trmatplot(stochastic_matrix_to_plot,main="Weather MC plot using MmgraphR")

Weather MC plot using MmgraphR

3
States

t t+1

Time
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 13

Import and export from some specific classes is possible, as shown in Figure 3 and in the
following code.

R> mcDf <- as(mcWeather, "data.frame")


R> mcNew <- as(mcDf, "markovchain")
R> mcDf

t0 t1 prob
1 sunny sunny 0.70
2 sunny cloudy 0.20
3 sunny rain 0.10
4 cloudy sunny 0.30
5 cloudy cloudy 0.40
6 cloudy rain 0.30
7 rain sunny 0.20
8 rain cloudy 0.45
9 rain rain 0.35

R> mcIgraph <- as(mcWeather, "igraph")

R> require(msm)

Loading required package: msm

R> Q <- rbind ( c(0, 0.25, 0, 0.25),


R+ c(0.166, 0, 0.166, 0.166),
R+ c(0, 0.25, 0, 0.25),
R+ c(0, 0, 0, 0) )
R> cavmsm <- msm(state ~ years, subject = PTNUM, data = cav, qmatrix = Q, death = 4)
R> msmMc <- as(cavmsm, "markovchain")
R> msmMc

Unnamed Markov chain


A 4 - dimensional discrete Markov Chain defined by the following states:
State 1, State 2, State 3, State 4
The transition matrix (by rows) is defined as follows:
State 1 State 2 State 3 State 4
State 1 0.853958721 0.08836953 0.01475543 0.04291632
State 2 0.155576908 0.56663284 0.20599563 0.07179462
State 3 0.009903994 0.07853691 0.65965727 0.25190183
State 4 0.000000000 0.00000000 0.00000000 1.00000000

R> library(etm)
R> data(sir.cont)
R> sir.cont <- sir.cont[order(sir.cont$id, sir.cont$time), ]
R> for (i in 2:nrow(sir.cont)) {
14 markovchain package: discrete Markov chains in R

R+ if (sir.cont$id[i]==sir.cont$id[i-1]) {
R+ if (sir.cont$time[i]==sir.cont$time[i-1]) {
R+ sir.cont$time[i-1] <- sir.cont$time[i-1] - 0.5
R+ }
R+ }
R+ }
R> tra <- matrix(ncol=3,nrow=3,FALSE)
R> tra[1, 2:3] <- TRUE
R> tra[2, c(1, 3)] <- TRUE
R> tr.prob <- etm(sir.cont, c("0", "1", "2"), tra, "cens", 1)
R> tr.prob

Multistate model with 2 transient state(s)


and 1 absorbing state(s)

Possible transitions:
from to
0 1
0 2
1 0
1 2

Estimate of P(1, 183)


0 1 2
0 0 0 1
1 0 0 1
2 0 0 1

R> etm2mc<-as(tr.prob, "markovchain")


R> etm2mc

Unnamed Markov chain


A 3 - dimensional discrete Markov Chain defined by the following states:
0, 1, 2
The transition matrix (by rows) is defined as follows:
0 1 2
0 0.0000000 0.5000000 0.5000000
1 0.5000000 0.0000000 0.5000000
2 0.3333333 0.3333333 0.3333333

Coerce from matrix method, as the code below shows, represents another approach to create
a markovchain method starting from a given squared probability matrix.

R> myMatr<-matrix(c(.1,.8,.1,.2,.6,.2,.3,.4,.3), byrow=TRUE, ncol=3)


R> myMc<-as(myMatr, "markovchain")
R> myMc
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 15

Import − Export from and to markovchain objects

etm matrix msm

markovchain

dataframe
igraph

sparseMatrix

table

Figure 3: The markovchain methods for import and export

Unnamed Markov chain


A 3 - dimensional discrete Markov Chain defined by the following states:
s1, s2, s3
The transition matrix (by rows) is defined as follows:
s1 s2 s3
s1 0.1 0.8 0.1
s2 0.2 0.6 0.2
s3 0.3 0.4 0.3

Non-homogeneous Markov chains can be created with the aid of markovchainList object.
The example that follows arises from health insurance, where the costs associated to patients
in a Continuous Care Health Community (CCHC) are modelled by a non-homogeneous Mar-
kov Chain, since the transition probabilities change by year. Methods explicitly written for
markovchainList objects are: print, show, dim and [.

R> stateNames = c("H", "I", "D")


R> Q0 <- new("markovchain", states = stateNames,
R+ transitionMatrix =matrix(c(0.7, 0.2, 0.1,0.1, 0.6, 0.3,0, 0, 1),
R+ byrow = TRUE, nrow = 3), name = "state t0")
R> Q1 <- new("markovchain", states = stateNames,
R+ transitionMatrix = matrix(c(0.5, 0.3, 0.2,0, 0.4, 0.6,0, 0, 1),
R+ byrow = TRUE, nrow = 3), name = "state t1")
R> Q2 <- new("markovchain", states = stateNames,
R+ transitionMatrix = matrix(c(0.3, 0.2, 0.5,0, 0.2, 0.8,0, 0, 1),
R+ byrow = TRUE,nrow = 3), name = "state t2")
R> Q3 <- new("markovchain", states = stateNames,
R+ transitionMatrix = matrix(c(0, 0, 1, 0, 0, 1, 0, 0, 1),
16 markovchain package: discrete Markov chains in R

R+ byrow = TRUE, nrow = 3), name = "state t3")


R> mcCCRC <- new("markovchainList",markovchains = list(Q0,Q1,Q2,Q3),
R+ name = "Continuous Care Health Community")
R> print(mcCCRC)

Continuous Care Health Community list of Markov chain(s)


Markovchain 1
state t0
A 3 - dimensional discrete Markov Chain defined by the following states:
H, I, D
The transition matrix (by rows) is defined as follows:
H I D
H 0.7 0.2 0.1
I 0.1 0.6 0.3
D 0.0 0.0 1.0

Markovchain 2
state t1
A 3 - dimensional discrete Markov Chain defined by the following states:
H, I, D
The transition matrix (by rows) is defined as follows:
H I D
H 0.5 0.3 0.2
I 0.0 0.4 0.6
D 0.0 0.0 1.0

Markovchain 3
state t2
A 3 - dimensional discrete Markov Chain defined by the following states:
H, I, D
The transition matrix (by rows) is defined as follows:
H I D
H 0.3 0.2 0.5
I 0.0 0.2 0.8
D 0.0 0.0 1.0

Markovchain 4
state t3
A 3 - dimensional discrete Markov Chain defined by the following states:
H, I, D
The transition matrix (by rows) is defined as follows:
H I D
H 0 0 1
I 0 0 1
D 0 0 1

It is possible to perform direct access to markovchainList elements, as well as to determine


Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 17

the number of markovchain objects by which a markovchainList object is composed.

R> mcCCRC[[1]]

state t0
A 3 - dimensional discrete Markov Chain defined by the following states:
H, I, D
The transition matrix (by rows) is defined as follows:
H I D
H 0.7 0.2 0.1
I 0.1 0.6 0.3
D 0.0 0.0 1.0

R> dim(mcCCRC)

[1] 4

The markovchain package contains some data found in the literature related to DTMC models
(see Section 6. Table 2 lists datasets and tables included within the current release of the
package.

Dataset Description
blanden Mobility across income quartiles, Jo Blanden and Machin (2005).
craigsendi CD4 cells, B. A. Craig and A. A. Sendi (2002).
kullback raw transition matrices for testing homogeneity, Kullback et al. (1962).
preproglucacon Preproglucacon DNA basis, P. J. Avery and D. A. Henderson (1999).
rain Alofi Island rains, P. J. Avery and D. A. Henderson (1999).
holson Individual states trajectiories.
sales Sales of six beverages in Hong Kong Ching et al. (2008).

Table 2: The markovchain data.frame and table.

Finally, Table 3 lists the demos included in the demo directory of the package.

R Code Filee Description


bard.R Structural analysis of Markov chains from Bard PPT.
examples.R Notable Markov chains, e.g., The Gambler Ruin chain.
quickStart.R Generic examples.
extractMatrices.R Generic examples.

Table 3: The markovchain demos.

4. Probability with markovchain objects


The markovchain package contains functions to analyse DTMC from a probabilistic per-
spective. For example, the package provides methods to find stationary distributions and
18 markovchain package: discrete Markov chains in R

identifying absorbing and transient states. Many of these methods come from MATLAB
listings that have been ported into R. For a full description of the underlying theory and
algorithm the interested reader can overview the original MATLAB listings, Feres (2007) and
Montgomery (2009).
Table 4 shows methods that can be applied on markovchain objects to perform probabilistic
analysis.

Method Returns
absorbingStates the absorbing states of the transition matrix, if any.
steadyStates the vector(s) of steady state(s) in matrix form.
meanFirstPassageTime matrix or vector of mean first passage times
committorAB committor probabilities
communicatingClasses list of communicating classes.
sj , given actual state si .
canonicForm the transition matrix into canonic form.
is.accessible checks whether a state j is reachable from state i.
is.irreducible checks whether a DTMC is irreducible.
period the period of an irreducible DTMC.
recurrentClasses list of recurrent classes.
summary DTMC summary.
transientStates the transient states of the transition matrix, if any.

Table 4: markovchain methods: statistical operations.

4.1. Conditional distributions


The conditional distribution of weather states, given that current day’s weather is sunny, is
given by following code.

R> conditionalDistribution(mcWeather, "sunny")

sunny cloudy rain


0.7 0.2 0.1

4.2. Stationary states


A stationary (steady state, or equilibrium) vector is a probability vector such that Equation
12 holds

0 ≤ πj ≤ 1
P
j∈S πj = 1 (12)
π∗P =π
Steady states are associated to P eigenvalues equal to one. Therefore the steady states vector
can be identified by the following:

1. decompose the transition matrix in eigenvalues and eigenvectors;


Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 19

2. consider only eigenvectors corresponding to eigenvalues equal to one;


3. normalize such eigenvalues so that the sum of their components is one.

Numeric issue (negative values) can arise when the Markov Chain contains more closed classes.
If negative values are found in the initial solution, the above described algorithm is performed
on the submatrix corresponding to recurrent P classes. Another vignette in the package
focuses on this issue.
The result is returned in matrix form.

R> steadyStates(mcWeather)

sunny cloudy rain


[1,] 0.4636364 0.3181818 0.2181818

It is possible for a Markov chain to have more than one stationary distribution, as the gambler
ruin example shows.

R> gamblerRuinMarkovChain <- function(moneyMax, prob = 0.5) {


R+ require(matlab)
R+ matr <- zeros(moneyMax + 1)
R+ states <- as.character(seq(from = 0, to = moneyMax, by = 1))
R+ rownames(matr) = states; colnames(matr) = states
R+ matr[1,1] = 1; matr[moneyMax + 1,moneyMax + 1] = 1
R+ for(i in 2:moneyMax)
R+ { matr[i,i-1] = 1 - prob; matr[i, i + 1] = prob }
R+ out <- new("markovchain",
R+ transitionMatrix = matr,
R+ name = paste("Gambler ruin", moneyMax, "dim", sep = " ")
R+ )
R+ return(out)
R+ }
R>
R> mcGR4 <- gamblerRuinMarkovChain(moneyMax = 4, prob = 0.5)

Loading required package: matlab

Attaching package: ✬matlab✬

The following object is masked from ✬package:stats✬:

reshape

The following objects are masked from ✬package:utils✬:

find, fix
20 markovchain package: discrete Markov chains in R

The following object is masked from ✬package:base✬:

sum

R> steadyStates(mcGR4)

0 1 2 3 4
[1,] 0 0 0 0 1
[2,] 1 0 0 0 0

4.3. Classification of states


Absorbing states are determined by means of absorbingStates method.

R> absorbingStates(mcGR4)

[1] "0" "4"

R> absorbingStates(mcWeather)

character(0)

The key function used within Feres (2007) (and markovchain’s derived functions) is .commclassKernel,
that is called below.

R> .commclassesKernel <- function(P){


R+ m <- ncol(P)
R+ stateNames <- rownames(P)
R+ T <- zeros(m)
R+ i <- 1
R+ while (i <= m) {
R+ a <- i
R+ b <- zeros(1,m)
R+ b[1,i] <- 1
R+ old <- 1
R+ new <- 0
R+ while (old != new) {
R+ old <- sum(find(b > 0))
R+ n <- size(a)[2]
R+ matr <- matrix(as.numeric(P[a,]), ncol = m,
R+ nrow = n)
R+ c <- colSums(matr)
R+ d <- find(c)
R+ n <- size(d)[2]
R+ b[1,d] <- ones(1,n)
R+ new <- sum(find(b>0))
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 21

R+ a <- d
R+ }
R+ T[i,] <- b
R+ i <- i+1 }
R+ F <- t(T)
R+ C <- (T > 0)&(F > 0)
R+ v <- (apply(t(C) == t(T), 2, sum) == m)
R+ colnames(C) <- stateNames
R+ rownames(C) <- stateNames
R+ names(v) <- stateNames
R+ out <- list(C = C, v = v)
R+ return(out)
R+ }

The .commclassKernel function gets a transition matrix of dimension n and return a list of
two items:

1. C, an adjacency matrix showing for each state sj (in the row) which states lie in the
same communicating class of sj (flagged with 1).
2. v, a binary vector indicating whether the state sj is transient (0) or not (1).

These functions are used by two other internal functions on which the summary method for
markovchain objects works.
The example matrix used in Feres (2007) well exemplifies the purpose of the function.

R> P <- matlab::zeros(10)


R> P[1, c(1, 3)] <- 1/2;
R> P[2, 2] <- 1/3; P[2,7] <- 2/3;
R> P[3, 1] <- 1;
R> P[4, 5] <- 1;
R> P[5, c(4, 5, 9)] <- 1/3;
R> P[6, 6] <- 1;
R> P[7, 7] <- 1/4; P[7,9] <- 3/4;
R> P[8, c(3, 4, 8, 10)] <- 1/4;
R> P[9, 2] <- 1;
R> P[10, c(2, 5, 10)] <- 1/3;
R> rownames(P) <- letters[1:10]
R> colnames(P) <- letters[1:10]
R> probMc <- new("markovchain", transitionMatrix = P,
R+ name = "Probability MC")
R> .commclassesKernel(P)

$C
a b c d e f g h i j
a TRUE FALSE TRUE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
b FALSE TRUE FALSE FALSE FALSE FALSE TRUE FALSE TRUE FALSE
22 markovchain package: discrete Markov chains in R

c TRUE FALSE TRUE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
d FALSE FALSE FALSE TRUE TRUE FALSE FALSE FALSE FALSE FALSE
e FALSE FALSE FALSE TRUE TRUE FALSE FALSE FALSE FALSE FALSE
f FALSE FALSE FALSE FALSE FALSE TRUE FALSE FALSE FALSE FALSE
g FALSE TRUE FALSE FALSE FALSE FALSE TRUE FALSE TRUE FALSE
h FALSE FALSE FALSE FALSE FALSE FALSE FALSE TRUE FALSE FALSE
i FALSE TRUE FALSE FALSE FALSE FALSE TRUE FALSE TRUE FALSE
j FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE TRUE

$v
a b c d e f g h i j
TRUE TRUE TRUE FALSE FALSE TRUE TRUE FALSE TRUE FALSE

R> summary(probMc)

Probability MC Markov chain that is composed by:


Closed classes:
a c
b g i
f
Recurrent classes:
{a,c},{b,g,i},{f}
Transient classes:
{d,e},{h},{j}
The Markov chain is not irreducible
The absorbing states are: f

All states that pertain to a transient class are named “transient” and a specific method has
been written to elicit them.

R> transientStates(probMc)

[1] "d" "e" "h" "j"

Listings from Feres (2007) have been adapted into canonicForm method that turns a Markov
chain into canonic form.

R> probMcCanonic <- canonicForm(probMc)


R> probMc

Probability MC
A 10 - dimensional discrete Markov Chain defined by the following states:
a, b, c, d, e, f, g, h, i, j
The transition matrix (by rows) is defined as follows:
a b c d e f g h i
a 0.5 0.0000000 0.50 0.0000000 0.0000000 0 0.0000000 0.00 0.0000000
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 23

b 0.0 0.3333333 0.00 0.0000000 0.0000000 0 0.6666667 0.00 0.0000000


c 1.0 0.0000000 0.00 0.0000000 0.0000000 0 0.0000000 0.00 0.0000000
d 0.0 0.0000000 0.00 0.0000000 1.0000000 0 0.0000000 0.00 0.0000000
e 0.0 0.0000000 0.00 0.3333333 0.3333333 0 0.0000000 0.00 0.3333333
f 0.0 0.0000000 0.00 0.0000000 0.0000000 1 0.0000000 0.00 0.0000000
g 0.0 0.0000000 0.00 0.0000000 0.0000000 0 0.2500000 0.00 0.7500000
h 0.0 0.0000000 0.25 0.2500000 0.0000000 0 0.0000000 0.25 0.0000000
i 0.0 1.0000000 0.00 0.0000000 0.0000000 0 0.0000000 0.00 0.0000000
j 0.0 0.3333333 0.00 0.0000000 0.3333333 0 0.0000000 0.00 0.0000000
j
a 0.0000000
b 0.0000000
c 0.0000000
d 0.0000000
e 0.0000000
f 0.0000000
g 0.0000000
h 0.2500000
i 0.0000000
j 0.3333333

R> probMcCanonic

Probability MC
A 10 - dimensional discrete Markov Chain defined by the following states:
a, c, b, g, i, f, d, e, h, j
The transition matrix (by rows) is defined as follows:
a c b g i f d e h
a 0.5 0.50 0.0000000 0.0000000 0.0000000 0 0.0000000 0.0000000 0.00
c 1.0 0.00 0.0000000 0.0000000 0.0000000 0 0.0000000 0.0000000 0.00
b 0.0 0.00 0.3333333 0.6666667 0.0000000 0 0.0000000 0.0000000 0.00
g 0.0 0.00 0.0000000 0.2500000 0.7500000 0 0.0000000 0.0000000 0.00
i 0.0 0.00 1.0000000 0.0000000 0.0000000 0 0.0000000 0.0000000 0.00
f 0.0 0.00 0.0000000 0.0000000 0.0000000 1 0.0000000 0.0000000 0.00
d 0.0 0.00 0.0000000 0.0000000 0.0000000 0 0.0000000 1.0000000 0.00
e 0.0 0.00 0.0000000 0.0000000 0.3333333 0 0.3333333 0.3333333 0.00
h 0.0 0.25 0.0000000 0.0000000 0.0000000 0 0.2500000 0.0000000 0.25
j 0.0 0.00 0.3333333 0.0000000 0.0000000 0 0.0000000 0.3333333 0.00
j
a 0.0000000
c 0.0000000
b 0.0000000
g 0.0000000
i 0.0000000
f 0.0000000
d 0.0000000
e 0.0000000
24 markovchain package: discrete Markov chains in R

h 0.2500000
j 0.3333333

The function is.accessible permits to investigate whether a state sj is accessible from state
si , that is whether the probability to eventually reach sj starting from si is greater than zero.

R> is.accessible(object = probMc, from = "a", to = "c")

[1] TRUE

R> is.accessible(object = probMc, from = "g", to = "c")

[1] FALSE

In Section 2.2 we observed that, if a DTMC is irreducible, all its states share the same
periodicity. Then, the period function returns the periodicity of the DTMC, provided that
it is irreducible. The example that follows shows how to find if a DTMC is reducible or
irreducible by means of the function is.irreducible and, in the latter case, the method
period is used to compute the periodicity of the chain.

R> E <- matrix(0, nrow = 4, ncol = 4)


R> E[1, 2] <- 1
R> E[2, 1] <- 1/3; E[2, 3] <- 2/3
R> E[3,2] <- 1/4; E[3, 4] <- 3/4
R> E[4, 3] <- 1
R>
R> mcE <- new("markovchain", states = c("a", "b", "c", "d"),
R+ transitionMatrix = E,
R+ name = "E")
R> is.irreducible(mcE)

[1] TRUE

R> period(mcE)

[1] 2

The example Markov chain found in Mathematica web site (Wolfram Research 2013a) has
been used, and is plotted in Figure 4.

R> require(matlab)
R> mathematicaMatr <- zeros(5)
R> mathematicaMatr[1,] <- c(0, 1/3, 0, 2/3, 0)
R> mathematicaMatr[2,] <- c(1/2, 0, 0, 0, 1/2)
R> mathematicaMatr[3,] <- c(0, 0, 1/2, 1/2, 0)
R> mathematicaMatr[4,] <- c(0, 0, 1/2, 1/2, 0)
R> mathematicaMatr[5,] <- c(0, 0, 0, 0, 1)
R> statesNames <- letters[1:5]
R> mathematicaMc <- new("markovchain", transitionMatrix = mathematicaMatr,
R+ name = "Mathematica MC", states = statesNames)
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 25

c 0.5

0.5
0.5

d 0.5

0.67

0.33
0.5

0.5

e 1

Figure 4: Mathematica 9 example. Markov chain plot.

Mathematica MC Markov chain that is composed by:


Closed classes:
c d
e
Recurrent classes:
{c,d},{e}
Transient classes:
{a,b}
The Markov chain is not irreducible
The absorbing states are: e

4.4. First passage time distributions and means


Feres (2007) provides code to compute first passage time (within 1, 2, . . . , n steps) given the
initial state to be i. The MATLAB listings translated into R on which the firstPassage
function is based are:

R> .firstpassageKernel <- function(P, i, n){


R> G <- P
R> H <- P[i,]
R> E <- 1 - diag(size(P)[2])
R> for (m in 2:n) {
R> G <- P %*% (G * E)
R> H <- rbind(H, G[i,])
R> }
R> return(H)
R> }
26 markovchain package: discrete Markov chains in R

We conclude that the probability for the first rainy day to be the third one, given that the
current state is sunny, is given by:

R> firstPassagePdF <- firstPassage(object = mcWeather, state = "sunny",


R+ n = 10)
R> firstPassagePdF[3, 3]

[1] 0.121

To compute the mean first passage times, i.e. the expected number of days before it rains
given that today is sunny, we can use the meanFirstPassageTime function:

R> meanFirstPassageTime(mcWeather)

sunny cloudy rain


sunny 0.000000 4.285714 6.666667
cloudy 3.725490 0.000000 5.000000
rain 4.117647 2.857143 0.000000

indicating e.g. that the average numer of days of sun or cloud before rain is 6.67 if we start
counting from a sunny day, and 5 if we start from a cloudy day. Note that we can also specify
one or more destination states:

R> meanFirstPassageTime(mcWeather,"rain")

sunny cloudy
6.666667 5.000000

The implementation follows the matrix solutions by (Grinstead and Snell 2006). We can check
the result by averaging the first passage probability density function:

R> firstPassagePdF.long <- firstPassage(object = mcWeather, state = "sunny", n = 100)


R> sum(firstPassagePdF.long[,"rain"] * 1:100)

[1] 6.666664

4.5. Committor probability


The committor probability tells us the probability to reach a given state before another given.
Suppose that we start in a cloudy day, the probabilities of experiencing a rainy day before a
sunny one is 0.5:

R> committorAB(mcWeather,3,1)

sunny cloudy rain


0.0 0.5 1.0
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 27

5. Statistical analysis
Table 5 lists the functions and methods implemented within the package which help to fit,
simulate and predict DTMC.

Function Purpose
markovchainFit Function to return fitted Markov chain for a given sequence.
predict Method to calculate predictions from markovchain or
markovchainList objects.
rmarkovchain Function to sample from markovchain or markovchainList objects.

Table 5: The markovchain statistical functions.

5.1. Simulation
Simulating a random sequence from an underlying DTMC is quite easy thanks to the function
rmarkovchain. The following code generates a year of weather states according to mcWeather
underlying stochastic process.

R> weathersOfDays <- rmarkovchain(n = 365, object = mcWeather, t0 = "sunny")


R> weathersOfDays[1:30]

[1] "cloudy" "rain" "cloudy" "sunny" "sunny" "cloudy" "cloudy"


[8] "rain" "cloudy" "rain" "cloudy" "rain" "cloudy" "rain"
[15] "cloudy" "cloudy" "rain" "rain" "cloudy" "sunny" "sunny"
[22] "sunny" "sunny" "sunny" "sunny" "sunny" "cloudy" "rain"
[29] "cloudy" "rain"

Similarly, it is possible to simulate one or more sequences from a non-homogeneous Markov


chain, as the following code (applied on CCHC example) exemplifies.

R> patientStates <- rmarkovchain(n = 5, object = mcCCRC, t0 = "H",


R+ include.t0 = TRUE)
R> patientStates[1:10,]

iteration values
1 1 H
2 1 I
3 1 D
4 1 D
5 1 D
6 2 H
7 2 H
8 2 H
9 2 H
10 2 D
28 markovchain package: discrete Markov chains in R

Two advance parameters are availabe to the rmarkovchain method which helps you decide
which implementation to use. There are four options available : R, R in parallel, C++
and C++ in parallel. Two boolean parameters useRcpp and parallel will decide which
implementation will be used. Default is useRcpp = TRUE and parallel = FALSE i.e. C++
implementation. The C++ implementation is generally faster than the R implementation. If
you have multicore processors then you can take advantage of parallel parameter by setting
it to TRUE. When both Rcpp=TRUE and parallel=TRUE the parallelization has been carried
out using RcppParallel package (Allaire, Francois, Ushey, Vandenbrouck, Geelnard, and Intel
2016).

5.2. Estimation
A time homogeneous Markov chain can be fit from given data. Four methods have been
implemented within current version of markovchain package: maximum likelihood, maximum
likelihood with Laplace smoothing, Bootstrap approach, maximum a posteriori.
Equation 13 shows the maximum likelihood estimator (MLE) of the pij entry, where the nij
element consists in the number sequences (Xt = si , Xt+1 = sj ) found in the sample, that is

nij
p̂M
ij
LE
= k
. (13)
P
niu
u=1

Equation (14) shows the standardError of the MLE (Skuriat-Olechnowska 2005).

p̂M
ij
LE
SEij = √ (14)
nij

R> weatherFittedMLE <- markovchainFit(data = weathersOfDays, method = "mle",name = "Weathe


R> weatherFittedMLE$estimate

Weather MLE
A 3 - dimensional discrete Markov Chain defined by the following states:
cloudy, rain, sunny
The transition matrix (by rows) is defined as follows:
cloudy rain sunny
cloudy 0.3478261 0.3304348 0.3217391
rain 0.5125000 0.3125000 0.1750000
sunny 0.1952663 0.1065089 0.6982249

R> weatherFittedMLE$standardError

cloudy rain sunny


cloudy 0.05499613 0.05360360 0.05289359
rain 0.08003905 0.06250000 0.04677072
sunny 0.03399149 0.02510438 0.06427681
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 29

The Laplace smoothing approach is a variation of the MLE, where the nij is substituted by
nij + α (see Equation 15), being α an arbitrary positive stabilizing parameter.

nij + α
p̂LS
ij = k
(15)
P
(niu + α)
u=1

R> weatherFittedLAPLACE <- markovchainFit(data = weathersOfDays,


R+ method = "laplace", laplacian = 0.01,
R+ name = "Weather LAPLACE")
R> weatherFittedLAPLACE$estimate

Weather LAPLACE
A 3 - dimensional discrete Markov Chain defined by the following states:
cloudy, rain, sunny
The transition matrix (by rows) is defined as follows:
cloudy rain sunny
cloudy 0.3478223 0.3304355 0.3217422
rain 0.5124328 0.3125078 0.1750594
sunny 0.1952908 0.1065491 0.6981601

(NOTE: The Confidence Interval option is enabled by default. Remove this option to fasten
computations.) Both MLE and Laplace approach are based on the createSequenceMatrix
functions that returns the raw counts transition matrix.

R> createSequenceMatrix(stringchar = weathersOfDays)

cloudy rain sunny


cloudy 40 38 37
rain 41 25 14
sunny 33 18 118

stringchar could contain NA values, and the transitions containing NA would be ignored.
An issue occurs when the sample contains only one realization of a state (say Xβ ) which is
located at the end of the data sequence, since it yields to a row of zero (no sample to estimate
the conditional distribution of the transition). In this case the estimated transition matrix is
corrected assuming pβ,j = 1/k, being k the possible states.
Create sequence matrix can also be used to obtain raw count transition matrices from a given
n ∗ 2 matrix as the following example shows:

R> myMatr<-matrix(c("a","b","b","a","a","b","b","b","b","a","a","a","b","a"),ncol=2)
R> createSequenceMatrix(stringchar = myMatr,toRowProbs = TRUE)

a b
a 0.6666667 0.3333333
b 0.5000000 0.5000000
30 markovchain package: discrete Markov chains in R

A bootstrap estimation approach has been developed within the package in order to provide
an indication of the variability of p̂ij estimates. The bootstrap approach implemented within
the markovchain package follows these steps:

1. bootstrap the data sequences following the conditional distributions of states estima-
ted from the original one. The default bootstrap samples is 10, as specified in nboot
parameter of markovchainFit function.
2. apply MLE estimation on bootstrapped data sequences that are saved in bootStrapSamples
slot of the returned list.
3. the pBOOT ST RAP ij is the average of all pM LE ij across the bootStrapSamples list, nor-
malized by row. A standardError of pM LE ˆ estimate is provided as well.
ij

R> weatherFittedBOOT <- markovchainFit(data = weathersOfDays,


R+ method = "bootstrap", nboot = 20)
R> weatherFittedBOOT$estimate

BootStrap Estimate
A 3 - dimensional discrete Markov Chain defined by the following states:
cloudy, rain, sunny
The transition matrix (by rows) is defined as follows:
cloudy rain sunny
cloudy 0.3473521 0.3351076 0.3175402
rain 0.5174037 0.3084494 0.1741469
sunny 0.1891193 0.1095270 0.7013537

R> weatherFittedBOOT$standardError

cloudy rain sunny


cloudy 0.011135229 0.012318630 0.011048099
rain 0.011755452 0.013216004 0.010068409
sunny 0.005930542 0.004387205 0.006975656

The bootstrapping process can be done in parallel thanks to RcppParallel package (Allaire
et al. 2016). Parallelized implementation is definitively suggested when the data sample size
or the required number of bootstrap runs is high.

R> weatherFittedBOOTParallel <- markovchainFit(data = weathersOfDays,


R> method = "bootstrap", nboot = 200,
R> parallel = TRUE)
R> weatherFittedBOOTParallel$estimate
R> weatherFittedBOOTParallel$standardError

The parallel bootstrapping uses all the available cores on a machine by default. However, it is
also possible to tune the number of threads used. Note that this should be done in R before
calling the markovchainFit function. For example, the following code will set the number of
threads to 4.
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 31

R> RcppParallel::setNumThreads(2)

For more details, please refer to RcppParallel web site.


For all the fitting methods, the logLikelihood (Skuriat-Olechnowska 2005) denoted in Equa-
tion 16 is provided.
X
LLH = nij ∗ log(pij ) (16)
i,j

where nij is the entry of the frequency matrix and pij is the entry of the transition probability
matrix.

R> weatherFittedMLE$logLikelihood

[1] -343.7674

R> weatherFittedBOOT$logLikelihood

[1] -343.8035

Confidence matrices of estimated parameters (parametric for MLE, non - parametric for
BootStrap) are available as well. The confidenceInterval is provided with the two matrices:
lowerEndpointMatrix and upperEndpointMatrix. The confidence level (CL) is 0.95 by
default and can be given as an argument of the function markovchainFit. This is used to
obtain the standard score (z-score). Equations 17 and 18 (Skuriat-Olechnowska 2005) show
the confidenceInterval of a fitting. Note that each entry of the matrices is bounded between
0 and 1.

LowerEndpointij = pij − zscore(CL) ∗ SEij (17)


U pperEndpointij = pij + zscore(CL) ∗ SEij (18)

R> weatherFittedMLE$confidenceInterval

NULL

R> weatherFittedBOOT$confidenceInterval

$confidenceLevel
[1] 0.95

$lowerEndpointMatrix
cloudy rain sunny
cloudy 0.3290363 0.3148453 0.2993677
rain 0.4980677 0.2867110 0.1575858
sunny 0.1793644 0.1023107 0.6898798
32 markovchain package: discrete Markov chains in R

$upperEndpointMatrix
cloudy rain sunny
cloudy 0.3656680 0.3553700 0.3357127
rain 0.5367397 0.3301878 0.1907080
sunny 0.1988742 0.1167433 0.7128276

A special function, multinomialConfidenceIntervals, has been written in order to obtain


multinomial wise confidence intervals. The code has been based on and Rcpp translation of
package’s MultinomialCI functions Villacorta (2012) that were themselves based on the Sison
and Glaz (1995) paper.

R> multinomialConfidenceIntervals(transitionMatrix =
R+ weatherFittedMLE$estimate@transitionMatrix,
R+ countsTransitionMatrix = createSequenceMatrix(weathersOfDays))

$confidenceLevel
[1] 0.95

$lowerEndpointMatrix
cloudy rain sunny
cloudy 0.2521739 0.23478261 0.2260870
rain 0.4125000 0.21250000 0.0750000
sunny 0.1301775 0.04142012 0.6331361

$upperEndpointMatrix
cloudy rain sunny
cloudy 0.4491348 0.4317435 0.4230479
rain 0.6361202 0.4361202 0.2986202
sunny 0.2652502 0.1764928 0.7682088

The functions for fitting DTMC have mostly been rewritten in C++ using Rcpp Eddelbuettel
(2013) since version 0.2.
It is also possible to fit a DTMC object from matrix or data.frame objects as shown in
following code.

R> data(holson)
R> singleMc<-markovchainFit(data=holson[,2:12],name="holson")

The same applies for markovchainList.

R> mcListFit<-markovchainListFit(data=holson[,2:6],name="holson")
R> mcListFit$estimate

holson list of Markov chain(s)


Markovchain 1
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 33

Unnamed Markov chain


A 3 - dimensional discrete Markov Chain defined by the following states:
1, 2, 3
The transition matrix (by rows) is defined as follows:
1 2 3
1 0.94609164 0.05390836 0.0000000
2 0.26356589 0.62790698 0.1085271
3 0.02325581 0.18604651 0.7906977

Markovchain 2
Unnamed Markov chain
A 3 - dimensional discrete Markov Chain defined by the following states:
1, 2, 3
The transition matrix (by rows) is defined as follows:
1 2 3
1 0.9323410 0.0676590 0.0000000
2 0.2551724 0.5103448 0.2344828
3 0.0000000 0.0862069 0.9137931

Markovchain 3
Unnamed Markov chain
A 3 - dimensional discrete Markov Chain defined by the following states:
1, 2, 3
The transition matrix (by rows) is defined as follows:
1 2 3
1 0.94765840 0.04820937 0.004132231
2 0.26119403 0.66417910 0.074626866
3 0.01428571 0.13571429 0.850000000

Markovchain 4
Unnamed Markov chain
A 3 - dimensional discrete Markov Chain defined by the following states:
1, 2, 3
The transition matrix (by rows) is defined as follows:
1 2 3
1 0.9172414 0.07724138 0.005517241
2 0.1678322 0.60839161 0.223776224
3 0.0000000 0.03030303 0.969696970

Finally, given a list object, it is possible to fit a markovchain object or to obtain the raw
transition matrix.

R> c1<-c("a","b","a","a","c","c","a")
R> c2<-c("b")
R> c3<-c("c","a","a","c")
R> c4<-c("b","a","b","a","a","c","b")
R> c5<-c("a","a","c",NA)
34 markovchain package: discrete Markov chains in R

R> c6<-c("b","c","b","c","a")
R> mylist<-list(c1,c2,c3,c4,c5,c6)
R> mylistMc<-markovchainFit(data=mylist)
R> mylistMc

$estimate
MLE Fit
A 3 - dimensional discrete Markov Chain defined by the following states:
a, b, c
The transition matrix (by rows) is defined as follows:
a b c
a 0.4 0.2000000 0.4000000
b 0.6 0.0000000 0.4000000
c 0.5 0.3333333 0.1666667

$standardError
a b c
a 0.2000000 0.1414214 0.2000000
b 0.3464102 0.0000000 0.2828427
c 0.2886751 0.2357023 0.1666667

$confidenceLevel
[1] 0.95

$lowerEndpointMatrix
a b c
a 0.07102927 0 0.07102927
b 0.03020599 0 0.00000000
c 0.02517166 0 0.00000000

$upperEndpointMatrix
a b c
a 0.7289707 0.4326174 0.7289707
b 1.0000000 0.0000000 0.8652349
c 0.9748283 0.7210291 0.4408089

The same works for markovchainFitList.

R> markovchainListFit(data=mylist)

$estimate
list of Markov chain(s)
Markovchain 1
Unnamed Markov chain
A 3 - dimensional discrete Markov Chain defined by the following states:
a, b, c
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 35

The transition matrix (by rows) is defined as follows:


a b c
a 0.5 0.5 0.0
b 0.5 0.0 0.5
c 1.0 0.0 0.0

Markovchain 2
Unnamed Markov chain
A 3 - dimensional discrete Markov Chain defined by the following states:
a, b, c
The transition matrix (by rows) is defined as follows:
a b c
a 0.3333333 0.3333333 0.3333333
b 1.0000000 0.0000000 0.0000000
c 0.0000000 1.0000000 0.0000000

Markovchain 3
Unnamed Markov chain
A 3 - dimensional discrete Markov Chain defined by the following states:
a, b, c
The transition matrix (by rows) is defined as follows:
a b c
a 0.5000000 0.0000000 0.5000000
b 0.5000000 0.0000000 0.5000000
c 0.3333333 0.3333333 0.3333333

Markovchain 4
Unnamed Markov chain
A 2 - dimensional discrete Markov Chain defined by the following states:
a, c
The transition matrix (by rows) is defined as follows:
a c
a 0.5 0.5
c 1.0 0.0

Markovchain 5
Unnamed Markov chain
A 2 - dimensional discrete Markov Chain defined by the following states:
a, c
The transition matrix (by rows) is defined as follows:
a c
a 0 1
c 0 1

Markovchain 6
Unnamed Markov chain
A 3 - dimensional discrete Markov Chain defined by the following states:
36 markovchain package: discrete Markov chains in R

a, b, c
The transition matrix (by rows) is defined as follows:
a b c
a 0.3333333 0.3333333 0.3333333
b 0.3333333 0.3333333 0.3333333
c 0.5000000 0.5000000 0.0000000

If any transition contains NA, it will be ignored in the results as the above example showed.

5.3. Prediction
The n-step forward predictions can be obtained using the predict methods explicitly written
for markovchain and markovchainList objects. The prediction is the mode of the conditional
distribution of Xt+1 given Xt = sj , being sj the last realization of the DTMC (homogeneous
or non-homogeneous).

Predicting from a markovchain object


The 3-days forward predictions from markovchain object can be generated as follows, assu-
ming that the last two days were respectively “cloudy” and “sunny”.

R> predict(object = weatherFittedMLE$estimate, newdata = c("cloudy", "sunny"),


R+ n.ahead = 3)

[1] "sunny" "sunny" "sunny"

Predicting from a markovchainList object


Given an initial two year Healty status, the 5-year ahead prediction of any CCRC guest is

R> predict(mcCCRC, newdata = c("H", "H"), n.ahead = 5)

[1] "H" "D" "D"

The prediction has stopped at time sequence since the underlying non-homogeneous Markov
chain has a length of four. In order to continue five years ahead, the continue=TRUE parameter
setting makes the predict method keeping to use the last markovchain in the sequence list.

R> predict(mcCCRC, newdata = c("H", "H"), n.ahead = 5, continue = TRUE)

[1] "H" "D" "D" "D" "D"

5.4. Statistical Tests


In this section, we describe the statistical tests: assessing the Markov property (verifyMarkovProperty),
the order (assessOrder), the statinarity (assessStationarity) of a Markov chain sequence,
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 37

and the divergence test for empirically estimated transition matrices (divergenceTest). Most
of such tests are based on the χ2 statistics. Relevand references are Kullback et al. (1962)
and Anderson and Goodman (1957).
All such tests have been designed for small samples, since it is easy to detect departures
from Markov property as long as the sample size increases. In addition, the accuracy of
the statistical inference functions has been questioned and will be thoroughly investigated in
future versions of the package.

Assessing the Markov property of a Markov chain sequence


The verifyMarkovProperty function verifies whether the Markov property holds for the given
chain. The test implemented in the package looks at triplets of successive observations. If
x1 , x2 , . . . , xN is a set of observations and nijk is the number of times t (1 ≤ t ≤ N − 2) such
that xt = i, xt+1 = j, xx+2 = k, then if the Markov property holds nijk follows a Binomial
distribution with parameters nij and pjk . A classical χ2 test can check this distributional
P P P n −nij pˆjk
∼ χ2 |S|3 where |S| is the cardinality of the state

assumption, since i j k ijknij pˆjk
space.

R> sample_sequence<-c("a", "b", "a", "a", "a", "a", "b", "a", "b", "a",
R+ "b", "a", "a", "b", "b", "b", "a")
R> verifyMarkovProperty(sample_sequence)

Warning in verifyMarkovProperty(sample_sequence): The accuracy of the


statistical inference functions has been questioned. It will be thoroughly
investigated in future versions of the package.

Testing markovianity property on given data sequence


Chi - square statistic is: 0.28
Degrees of freedom are: 8
And corresponding p-value is: 0.9999857

Assessing the order of a Markov chain sequence


The assessOrder function checks whether the given chain is of first order or of second order.
For each possible present state, we construct a contingency table of the frequency of the future
state for each past to present state transition as shown in Table 6.

past present future future


a b
a a 2 2
b a 2 2

Table 6: Contingency table to assess the order for the present state a.

Using the table, the function performs the χ2 test by calling the chisq.test function. This
test returns a list of the chi-squared value and the p-value. If the p-value is greater than the
given significance level, we cannot reject the hypothesis that the sequence is of first order.
38 markovchain package: discrete Markov chains in R

R> data(rain)
R> assessOrder(rain$rain)

Warning in assessOrder(rain$rain): The accuracy of the statistical


inference functions has been questioned. It will be thoroughly investigated
in future versions of the package.

The assessOrder test statistic is: 26.09575


The Chi-Square d.f. are: 12
The p-value is: 0.01040395

Assessing the stationarity of a Markov chain sequence


The assessStationarity function assesses if the transition probabilities of the given chain
change over time. To be more specific, the chain is stationary if the following condition meets.

pij (t) = pij for all t (19)

For each possible state, we construct a contingency table of the estimated transition proba-
bilities over time as shown in Table 7.

time (t) probability of transition to a probability of transition to b


1 0 1
2 0 1
. . .
. . .
. . .
16 0.44 0.56

Table 7: Contingency table to assess the stationarity of the state a.

Using the table, the function performs the χ2 test by calling the chisq.test function. This
test returns a list of the chi-squared value and the p-value. If the p-value is greater than the
given significance level, we cannot reject the hypothesis that the sequence is stationary.

R> assessStationarity(rain$rain, 10)

Warning in assessStationarity(rain$rain, 10): The accuracy of the


statistical inference functions has been questioned. It will be thoroughly
investigated in future versions of the package.

Warning in chisq.test(mat): Chi-squared approximation may be incorrect

Warning in chisq.test(mat): Chi-squared approximation may be incorrect

Warning in chisq.test(mat): Chi-squared approximation may be incorrect


Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 39

The assessStationarity test statistic is: 4.181815


The Chi-Square d.f. are: 54
The p-value is: 1

Divergence tests for empirically estimated transition matrices


This section discusses tests developed to verify whether:

1. An empirical transition matrix is consistent with a theoretical one.

2. Two or more empirical transition matrices belongs to the same DTMC.

The first test is implemented by the verifyEmpiricalToTheoretical function. Bein fij


f
the raw transition count, Kullback et al. (1962) shows that 2 ∗ ri=1 rj=1 fij ln fi. P (Eijj |Ei ) ∼
P P

χ2 (r ∗ (r − 1)). The following example is taken from Kullback et al. (1962):

R> sequence<-c(0,1,2,2,1,0,0,0,0,0,0,1,2,2,2,1,0,0,1,0,0,0,0,0,0,1,1,
R+ 2,0,0,2,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,0,0,0,0,2,1,0,
R+ 0,2,1,0,0,0,0,0,0,1,1,1,2,2,0,0,2,1,1,1,1,2,1,1,1,1,1,1,1,1,1,0,2,
R+ 0,1,1,0,0,0,1,2,2,0,0,0,0,0,0,2,2,2,1,1,1,1,0,1,1,1,1,0,0,2,1,1,
R+ 0,0,0,0,0,2,2,1,1,1,1,1,2,1,2,0,0,0,1,2,2,2,0,0,0,1,1)
R> mc=matrix(c(5/8,1/4,1/8,1/4,1/2,1/4,1/4,3/8,3/8),byrow=TRUE, nrow=3)
R> rownames(mc)<-colnames(mc)<-0:2; theoreticalMc<-as(mc, "markovchain")
R> verifyEmpiricalToTheoretical(data=sequence,object=theoreticalMc)

Warning in verifyEmpiricalToTheoretical(data = sequence, object =


theoreticalMc): The accuracy of the statistical inference functions has
been questioned. It will be thoroughly investigated in future versions of
the package.

Testing whether the


0 1 2
0 51 11 8
1 12 31 9
2 6 11 10
transition matrix is compatible with
0 1 2
0 0.625 0.250 0.125
1 0.250 0.500 0.250
2 0.250 0.375 0.375
[1] "theoretical transition matrix"
ChiSq statistic is 6.551795 d.o.f are 6 corresponding p-value is 0.3642899

$statistic
0
6.551795
40 markovchain package: discrete Markov chains in R

$dof
[1] 6

$pvalue
0
0.3642899

The second one is implemented by the verifyHomogeneity function, inspired by (Kullback


et al. 1962, section 9). Assuming that i = 1, 2, . . . , s DTMC samples are available and that
the cardinality of the state space is r it verifies whether the s chains belongs to the same
unknown one. Kullback et al. (1962) shows that its test statistics follows a chi-square law,
n∗f
2 ∗ si=1 rj=1 rk=1 fijk ln fi.. fijk ∼ χ2 (r ∗ (r − 1)). Also the following example is taken from
P P P
.jk
Kullback et al. (1962):

R> data(kullback)
R> verifyHomogeneity(inputList=kullback,verbose=TRUE)

Warning in verifyHomogeneity(inputList = kullback, verbose = TRUE): The


accuracy of the statistical inference functions has been questioned. It
will be thoroughly investigated in future versions of the package.

Testing homogeneity of DTMC underlying input list


ChiSq statistic is 275.9963 d.o.f are 35 corresponding p-value is 0

$statistic
[1] 275.9963

$dof
[1] 35

$pvalue
[1] 0

5.5. Continuous Times Markov Chains

Intro
The markovchain package provides functionality for continuous time Markov chains (CTMCs).
CTMCs are a generalisation of discrete time Markov chains (DTMCs) in that we allow time
to be continuous. We assume a finite state space S (for an infinite state space wouldn’t fit in
memory). We can think of CTMCs as Markov chains in which state transitions can happen
at any time.
More formally, we would like our CTMCs to satisfy the following two properties:
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 41

❼ The Markov property - let FX(s) denote the information about X upto time s. Let
j ∈ S and s ≤ t. Then, P (X(t) = j|FX(s) ) = P (X(t) = j|X(s)).
❼ Time homogenity - P (X(t) = j|X(s) = k) = P (X(t − s) = j|X(0) = k).

If both the above properties are satisfied, it is referred to as a time-homogeneous CTMC. If


a transition occurs at time t, then X(t) denotes the new state and X(t) 6= X(t−).
Now, let X(0) = x and let Tx be the time a transition occurs from this state. We are interested
in the distribution of Tx . For s, t ≥ 0, it can be shown that $ P(T x > s+t | T x > s) =
P(T x > t) $
This is the memory less property that only the exponential random variable exhibits. There-
fore, this is the sought distribution, and each state s ∈ S has an exponential holding parameter
1
λ(s). Since ETx = λ(x) , higher the rate λ(x), smaller the expected time of transitioning out
of the state x.
However, specifying this parameter alone for each state would only paint an incomplete picture
of our CTMC. To see why, consider a state x that may transition to either state y or z. The
holding parameter enables us to predict when a transition may occur if we start off in state
x, but tells us nothing about which state will be next.
To this end, we also need transition probabilities associated with P the process, defined as
follows (for y 6= x) - pxy = P (X(Ts ) = y|X(0) = x). Note that y6=x pxy = 1. Let Q denote
this transition matrix (Qij = pij ). What is key here is that Tx and the state y are independent
random variables. Let’s define λ(x, y) = λ(x)pxy
We now look at Kolmogorov’s backward equation. Let’s define Pij (t) = P (X(t) = j|X(0) = i)
for i, j ∈ S. The backward equation is given by (it can be proved) Pij (t) = δij e−λ(i)t +
Rt −λ(i)t
P
0 λ(i)e k6=i Qik Pkj (t − s)ds. Basically, the first term is non-zero if and only if i = j
and represents the probability that the first transition from state i occurs after time t. This
would mean that at t, the state is still i. The second term accounts for any transitions that
may occur before time t and denotes the probability that at time t, when the smoke clears,
we are in state j.
This equation can be represented compactly as follows P ′ (t) = AP (t) where A is the generator
matrix. (
λ(i, j) if i 6= j
A(i, j) =
−λ(i) else.
Observe that the sum of each row is 0. A CTMC can be completely specified by the generator
matrix.

Stationary Distributions
The following theorem guarantees the existence of a unique stationary distribution for CT-
MCs. Note that X(t) being irreducible and recurrent is the same as Xn (t) being irreducible
and recurrent.
Suppose that X(t) is irreducible and recurrent. Then X(t) has an invariant measure η, which
is unique up to multiplicative factors. Moreover, for each k ∈ S, we have

πk
ηk =
λ(k)
42 markovchain package: discrete Markov chains in R

where π is the unique invariant measure of the embedded discrete time Markov chain Xn.
Finally, η satisfies

0 < ηj < ∞, ∀j ∈ S
P
and if i ηi < ∞ then η can be normalised to get a stationary distribution.

Estimation
Let the data set be D = {(s0 , t0 ), (s1 , t1 ), ..., (sN −1 , tN −1 )} where N = |D|. Each si is a state
from the state space S and during the time [ti , ti+1 ] the chain is in state si . Let the parameters
be represented by θ = {λ, P } where λ is the vector of holding parameters for each state and
P the transition matrix of the embedded discrete time Markov chain.
Then the probability is given by

P r(D|θ) ∝ λ(s0 )e−λ(s0 )(t1 −t0 ) P r(s1 |s0 ) · . . . · λ(sN −2 )e−λ(sN −2 )(tN −1 −tN −2 ) P r(sN −1 |sN −2 )

Let n(j|i) denote the number of i->j transitions in D, and n(i) the number of times si occurs
in D. Let t(si ) denote the total time the chain spends in state si .
Then the MLEs are given by

ˆ = n(s) , P r(j|i)
λ(s) ˆ =
n(j|i)
t(s) n(i)

Expected Hitting Time


The package provides a function ExpectedTime to calculate average hitting time from one
state to another. Let the final state be j, then for every state i ∈ I, where I is the set of
all states and holding time qi > 0 for every i 6= j. Assuming the conditions to be true,
expected hitting time is equal to minimal non-negative solution vector p to the system of
linear equations: (
pk = 0 k=j
P (20)
− l∈I qkl pk = 1 k 6= j

Probability at time t
The package provides a function probabilityatT to calculate probability of every state ac-
cording to given ctmc object. Here we use Kolmogorov’s backward equation P (t) = P (0)etQ
for t ≥ 0 and P (0) = I. Here P (t) is the transition function at time t. The value P (t)[i][j]
at time P (t) describes the probability of the state at time t to be eqaul to j if it was equal
to i at time t = 0. It takes care of the case when ctmc object has a generator represented
by columns. If inital state is not provided, the function returns the whole transition matrix
P (t).

Examples
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 43

To create a CTMC object, you need to provide a valid generator matrix, say Q. The CTMC
object has the following slots - states, generator, byrow, name (look at the documentation
object for further details). Consider the following example in which we aim to model the
transition of a molecule from the σ state to the σ ∗ state. When in the former state, if it
absorbs sufficient energy, it can make the jump to the latter state and remains there for some
time before transitioning back to the original state. Let us model this by a CTMC:

R> energyStates <- c("sigma", "sigma_star")


R> byRow <- TRUE
R> gen <- matrix(data = c(-3, 3,
R+ 1, -1), nrow = 2,
R+ byrow = byRow, dimnames = list(energyStates, energyStates))
R> molecularCTMC <- new("ctmc", states = energyStates,
R+ byrow = byRow, generator = gen,
R+ name = "Molecular Transition Model")

To generate random CTMC transitions, we provide an initial distribution of the states. This
must be in the same order as the dimnames of the generator. The output can be returned
either as a list or a data frame.

R> statesDist <- c(0.8, 0.2)


R> rctmc(n = 3, ctmc = molecularCTMC, initDist = statesDist, out.type = "df", include.T0 =

states time
1 sigma_star 0.319699317028095
2 sigma 1.08757360349614
3 sigma_star 1.22777987434733

n represents the number of samples to generate. There is an optional argument T for rctmc.
It represents the time of termination of the simulation. To use this feature, set n to a very
high value, say Inf (since we do not know the number of transitions before hand) and set T
accordingly.

R> statesDist <- c(0.8, 0.2)


R> rctmc(n = Inf, ctmc = molecularCTMC, initDist = statesDist, T = 2)

[[1]]
[1] "sigma" "sigma_star" "sigma" "sigma_star" "sigma"
[6] "sigma_star" "sigma" "sigma_star"

[[2]]
[1] 0.0000000 0.3245419 0.6881124 1.0515503 1.4466669 1.4591308 1.5751001
[8] 1.8251498

To obtain the stationary distribution simply invoke the steadyStates function

R> steadyStates(molecularCTMC)
44 markovchain package: discrete Markov chains in R

sigma sigma_star
[1,] 0.25 0.75

For fitting, use the ctmcFit function. It returns the MLE values for the parameters along
with the confidence intervals.

R> data <- list(c("a", "b", "c", "a", "b", "a", "c", "b", "c"),
R+ c(0, 0.8, 2.1, 2.4, 4, 5, 5.9, 8.2, 9))
R> ctmcFit(data)

$estimate
An object of class "ctmc"
Slot "states":
[1] "a" "b" "c"

Slot "byrow":
[1] TRUE

Slot "generator":
a b c
a -0.9090909 0.6060606 0.3030303
b 0.3225806 -0.9677419 0.6451613
c 0.3846154 0.3846154 -0.7692308

Slot "name":
[1] ""

$errors
$errors$dtmcConfidenceInterval
$errors$dtmcConfidenceInterval$confidenceLevel
[1] 0.95

$errors$dtmcConfidenceInterval$lowerEndpointMatrix
a b c
a 0 0 0
b 0 0 0
c 0 0 0

$errors$dtmcConfidenceInterval$upperEndpointMatrix
a b c
a 0.0000000 1 0.8816179
b 0.8816179 0 1.0000000
c 1.0000000 1 0.0000000

$errors$lambdaConfidenceInterval
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 45

$errors$lambdaConfidenceInterval$lowerEndpointVector
[1] 0.04576665 0.04871934 0.00000000

$errors$lambdaConfidenceInterval$upperEndpointVector
[1] 1 1 1

One approach to obtain the generator matrix is to apply the logm function from the expm
package on a transition matrix. Numeric issues arise, see Israel, Rosenthal, and Wei (2001).
For example, applying the standard method (‘Higham08’) on mcWeather raises an error, whilst
the alternative method (eigenvalue decomposition) is ok. The following code estimates the
generator matrix of the mcWeather transition matrix.

R> mcWeatherQ <- expm::logm(mcWeather@transitionMatrix,method=✬Eigen✬)


R> mcWeatherQ

sunny cloudy rain


sunny -0.863221 2.428723 -1.565502
cloudy 4.284592 -20.116312 15.831720
rain -4.414019 24.175251 -19.761232

Therefore, the “half - day” transition probability for mcWeather DTMC is

R> mcWeatherHalfDayTM <- expm::expm(mcWeatherQ*.5)


R> mcWeatherHalfDay <- new("markovchain",transitionMatrix=mcWeatherHalfDayTM,name="Half Da
R> mcWeatherHalfDay

Half Day Weather Transition Matrix


A 3 - dimensional discrete Markov Chain defined by the following states:
sunny, cloudy, rain
The transition matrix (by rows) is defined as follows:
sunny cloudy rain
sunny 0.81598647 0.1420068 0.04200677
cloudy 0.21970167 0.4401492 0.34014916
rain 0.07063048 0.5146848 0.41468476

The ctmcd package (Pfeuffer 2017) provides various functions to estimate the generator matrix
(GM) of a CTMC process using different methods. The following code provides a way to join
markovchain and ctmcd computations.

R> require(ctmcd)

Loading required package: ctmcd

R> require(expm)

Loading required package: expm


46 markovchain package: discrete Markov chains in R

Loading required package: Matrix

Attaching package: ✬expm✬

The following object is masked from ✬package:Matrix✬:

expm

R> #defines a function to transform a GM into a TM


R> gm_to_markovchain<-function(object, t=1) {
R+ if(!(class(object) %in% c("gm","matrix","Matrix")))
R+ stop("Error! Expecting either a matrix or a gm object")
R+ if ( class(object) %in% c("matrix","Matrix")) generator_matrix<-object else generator
R+ #must add importClassesFrom("markovchain",markovchain) in the NAMESPACE
R+ #must add importFrom(expm, "expm")
R+ transitionMatrix<-expm(generator_matrix*t)
R+ out<-as(transitionMatrix,"markovchain")
R+ return(out)
R+ }
R> #loading ctmcd dataset
R> data(tm_abs)
R> gm0=matrix(1,8,8) #initializing
R> diag(gm0)=0
R> diag(gm0)=-rowSums(gm0)
R> gm0[8,]=0
R> gmem=gm(tm_abs,te=1,method="EM",gmguess=gm0) #estimating GM
R> mc_at_2=gm_to_markovchain(object=gmem, t=2) #converting to TM at time 2

5.6. Pseudo - Bayesian Estimation


Hu, Kiesel, and Perraudin (2002) shows an empirical quasi-bayesian method to estimate
transition matrices, given an empirical P̂ transition matrix (estimated using the classical
approach) and an a - priori estimate Q. In particular, each row of the matrix is estimated
using the linear combination α ∗ Q + (1 − 1alpha) ∗ P , where α is defined for each row as
Equation 21 shows

K̂i
 α̂i =
v(i)+ P K̂i
v(i)2 − j Yij2 (21)
K̂i = P 2
j (Yij −v(i)∗qij )

The following code returns the pseudo bayesian estimate of the transition matrix:

R> pseudoBayesEstimator <- function(raw, apriori){


R+ v_i <- rowSums(raw)
R+ K_i <- numeric(nrow(raw))
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 47

R+ sumSquaredY <- rowSums(raw^2)


R+ #get numerator
R+ K_i_num <- v_i^2-sumSquaredY
R+ #get denominator
R+ VQ <- matrix(0,nrow= nrow(apriori),ncol=ncol(apriori))
R+ for (i in 1:nrow(VQ)) {
R+ VQ[i,]<-v_i[i]*apriori[i,]
R+ }
R+
R+ K_i_den<-rowSums((raw - VQ)^2)
R+
R+ K_i <- K_i_num/K_i_den
R+
R+ #get the alpha vector
R+ alpha <- K_i / (v_i+K_i)
R+
R+ #empirical transition matrix
R+ Emp<-raw/rowSums(raw)
R+
R+ #get the estimate
R+ out<-matrix(0, nrow= nrow(raw),ncol=ncol(raw))
R+ for (i in 1:nrow(out)) {
R+ out[i,]<-alpha[i]*apriori[i,]+(1-alpha[i])*Emp[i,]
R+ }
R+ return(out)
R+ }

We then apply it to the weather example:

R> trueMc<-as(matrix(c(0.1, .9,.7,.3),nrow = 2, byrow = 2),"markovchain")


R> aprioriMc<-as(matrix(c(0.5, .5,.5,.5),nrow = 2, byrow = 2),"markovchain")
R>
R> smallSample<-rmarkovchain(n=20,object = trueMc)
R> smallSampleRawTransitions<-createSequenceMatrix(stringchar = smallSample)
R> pseudoBayesEstimator(
R+ raw = smallSampleRawTransitions,
R+ apriori = aprioriMc@transitionMatrix
R+ ) - trueMc@transitionMatrix

s1 s2
s1 0.17758007 -0.17758007
s2 -0.06885246 0.06885246

R> biggerSample<-rmarkovchain(n=100,object = trueMc)


R> biggerSampleRawTransitions<-createSequenceMatrix(stringchar = biggerSample)
R> pseudoBayesEstimator(
R+ raw = biggerSampleRawTransitions,
48 markovchain package: discrete Markov chains in R

R+ apriori = aprioriMc@transitionMatrix
R+ ) - trueMc@transitionMatrix

s1 s2
s1 -0.02298889 0.02298889
s2 -0.04711818 0.04711818

R> bigSample<-rmarkovchain(n=1000,object = trueMc)


R> bigSampleRawTransitions<-createSequenceMatrix(stringchar = bigSample)
R> pseudoBayesEstimator(
R+ raw = bigSampleRawTransitions,
R+ apriori = aprioriMc@transitionMatrix
R+ ) - trueMc@transitionMatrix

s1 s2
s1 -0.01138052 0.01138052
s2 -0.01003458 0.01003458

5.7. Bayesian Estimation


The markovchain package provides functionality for maximum a posteriori (MAP) estimation
of the chain parameters (at the time of writing this document, only first order models are
supported) by Bayesian inference. It also computes the probability of observing a new data
set, given a (different) data set. This vignette provides the mathematical description for the
methods employed by the package.

Notation and set-up


The data is denoted by D, the model parameters (transition matrix) by θ. The object of
interest is P (θ|D) (posterior density). A represents an alphabet class, each of whose members
represent a state of the chain. Therefore

D = s0 s1 ...sN −1 , st ∈ A

where N is the length of the data set. Also,

θ = {p(s|u), s ∈ A, u ∈ A}
P
where s∈A p(s|u) = 1 for each u ∈ A.
Our objective is to find θ which maximises the posterior. That is, if our solution is denoted
by θ̂, then

θ̂ = argmaxP (θ|D)
θ

where the search space is the set of right stochastic matrices of dimension |A|x|A|.
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 49

P
n(u, s) denotes the number of times the word us occurs in D and n(u) = s∈A n(u, s). The
hyperparameters are similarly denoted by α(u, s) and α(u) respectively.

Methods
Given D, its likelihood conditioned on the observed initial state in D is given by
Y Y
P (D|θ) = p(s|u)n(u,s)
s∈A u∈A

Conjugate priors are used to model the prior P (θ). The reasons are two fold:

1. Exact expressions can be derived for the MAP estimates, expectations and even vari-
ances
2. Model order selection/comparison can be implemented easily (available in a future re-
lease of the package)

The hyperparameters determine the form of the prior distribution, which is a product of
Dirichlet distributions

Yn Γ(α(u)) Y o
P (θ) = Q p(s|u)α(u,s))−1
u∈A s∈A Γ(α(u, s)) s∈A

where Γ(.) is the Gamma function. The hyperparameters are specified using the hyperparam
argument in the markovchainFit function. If this argument is not specified, then a default
value of 1 is assigned to each hyperparameter resulting in the prior distribution of each chain
parameter to be uniform over [0, 1].
Given the likelihood and the prior as described above, the evidence P (D) is simply given by
Z
P (D) = P (D|θ)P (θ)dθ

which simplifies to
Q
Γ(α(u)) s∈A Γ(n(u, s)
+ α(u, s)) o
Yn
P (D) = Q
u∈A s∈A Γ(α(u, s)) Γ(α(u) + n(u))

Using Bayes’ theorem, the posterior now becomes (thanks to the choice of conjugate priors)
Yn Γ(n(u) + α(u)) Y o
P (θ|D) = Q p(s|u)n(u,s)+α(u,s))−1
u∈A s∈A Γ(n(u, s) + α(u, s)) s∈A

Since this is again a product of Dirichlet distributions, the marginalised distribution of a


particular parameter P (s|u) of our chain is given by
P (s|u) ∼ Beta(n(u, s) + α(u, s), n(u) + α(u) − n(u, s) − α(u, s))

Thus, the MAP estimate θ̂ is given by


n n(u, s) + α(u, s) − 1 o
θ̂ = , s ∈ A, u ∈ A
n(u) + α(u) − |A|
50 markovchain package: discrete Markov chains in R

The function also returns the expected value, given by


n n(u, s) + α(u, s) o
Epost p(s|u) = , s ∈ A, u ∈ A
n(u) + α(u)

The variance is given by


n(u, s) + α(u, s) n(u) + α(u) − n(u, s) − α(u, s)
Varpost p(s|u) =
(n(u) + α(u))2 n(u) + α(u) + 1

The square root of this quantity is the standard error, which is returned by the function.
The confidence intervals are constructed by computing the inverse of the beta integral.

Predictive distribution
Given the old data set, the probability of observing new data is P (D′ |D) where D′ is the new
data set. Let m(u, s), m(u) denote the corresponding counts for the new data. Then,
Z
P (D |D) = P (D′ |θ)P (θ|D)dθ

We already know the expressions for both quantities in the integral and it turns out to be
similar to evaluating the evidence
Q
Γ(α(u)) s∈A Γ(n(u, s) + m(u, s) + α(u, s))
Yn o

P (D |D) = Q
u∈A s∈A Γ(α(u, s)) Γ(α(u) + n(u) + m(u))

Choosing the hyperparameters


The hyperparameters model the shape of the parameters’ prior distribution. These must
be provided by the user. The package offers functionality to translate a given prior belief
transition matrix into the hyperparameter matrix. It is assumed that this belief matrix
corresponds to the mean value of the parameters. Since the relation
α(u, s)
Eprior p(s|u) =
α(u)

holds, the function accepts as input the belief matrix as well as a scaling vector (serves as a
proxy for α(.)) and proceeds to compute α(., .).
Alternatively, the function accepts a data sample and infers the hyperparameters from it.
Since the mode of a parameter (with respect to the prior distribution) is proportional to one
less than the corresponding hyperparameter, we set

α(u, s) − 1 = m(u, s)

where m(u, s) is the u → s transition count in the data sample. This is regarded as a ‘fake
count’ which helps α(u, s) to reflect knowledge of the data sample.

Usage and examples


Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 51

R> weatherStates <- c("sunny", "cloudy", "rain")


R> byRow <- TRUE
R> weatherMatrix <- matrix(data = c(0.7, 0.2, 0.1,
R+ 0.3, 0.4, 0.3,
R+ 0.2, 0.4, 0.4),
R+ byrow = byRow, nrow = 3,
R+ dimnames = list(weatherStates, weatherStates))
R> mcWeather <- new("markovchain", states = weatherStates,
R+ byrow = byRow, transitionMatrix = weatherMatrix,
R+ name = "Weather")
R> weathersOfDays <- rmarkovchain(n = 365, object = mcWeather, t0 = "sunny")

For the purpose of this section, we shall continue to use the weather of days example intro-
duced in the main vignette of the package (reproduced above for convenience).
Let us invoke the fit function to estimate the MAP parameters with 92% confidence bounds
and hyperparameters as shown below, based on the first 200 days of the weather data. Ad-
ditionally, let us find out what the probability is of observing the weather data for the next
165 days. The usage would be as follows

R> hyperMatrix<-matrix(c(1, 1, 2,
R+ 3, 2, 1,
R+ 2, 2, 3),
R+ nrow = 3, byrow = TRUE,
R+ dimnames = list(weatherStates,weatherStates))
R> markovchainFit(weathersOfDays[1:200], method = "map",
R+ confidencelevel = 0.92, hyperparam = hyperMatrix)

$estimate
Bayesian Fit
A 3 - dimensional discrete Markov Chain defined by the following states:
cloudy, rain, sunny
The transition matrix (by rows) is defined as follows:
cloudy rain sunny
cloudy 0.4126984 0.2698413 0.3174603
rain 0.4634146 0.3902439 0.1463415
sunny 0.1553398 0.0776699 0.7669903

$expectedValue
cloudy rain sunny
cloudy 0.4090909 0.27272727 0.3181818
rain 0.4545455 0.38636364 0.1590909
sunny 0.1603774 0.08490566 0.7547170

$standardError
[,1] [,2] [,3]
52 markovchain package: discrete Markov chains in R

[1,] 0.06006657 0.05440960 0.05690292


[2,] 0.07422696 0.07258509 0.05452441
[3,] 0.03547494 0.02694693 0.04159431

$confidenceInterval
$confidenceInterval$confidenceLevel
[1] 0.92

$confidenceInterval$lowerEndpointMatrix
[,1] [,2] [,3]
[1,] 0.32117517 0.1801276 0.2266730
[2,] 0.35036388 0.2810536 0.0000000
[3,] 0.08492395 0.0000000 0.6946682

$confidenceInterval$upperEndpointMatrix
[,1] [,2] [,3]
[1,] 0.5508869 0.3696812 0.4276070
[2,] 1.0000000 0.5612194 0.2404126
[3,] 0.2143174 0.1249744 1.0000000

$logLikelihood
[1] -170.3136

R> predictiveDistribution(weathersOfDays[1:200],
R+ weathersOfDays[201:365],hyperparam = hyperMatrix)

[1] -163.8482

The results should not change after permuting the dimensions of the matrix.

R> hyperMatrix2<- hyperMatrix[c(2,3,1), c(2,3,1)]


R> markovchainFit(weathersOfDays[1:200], method = "map",
R+ confidencelevel = 0.92, hyperparam = hyperMatrix2)

$estimate
Bayesian Fit
A 3 - dimensional discrete Markov Chain defined by the following states:
cloudy, rain, sunny
The transition matrix (by rows) is defined as follows:
cloudy rain sunny
cloudy 0.4126984 0.2698413 0.3174603
rain 0.4634146 0.3902439 0.1463415
sunny 0.1553398 0.0776699 0.7669903
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 53

$expectedValue
cloudy rain sunny
cloudy 0.4090909 0.27272727 0.3181818
rain 0.4545455 0.38636364 0.1590909
sunny 0.1603774 0.08490566 0.7547170

$standardError
[,1] [,2] [,3]
[1,] 0.06006657 0.05440960 0.05690292
[2,] 0.07422696 0.07258509 0.05452441
[3,] 0.03547494 0.02694693 0.04159431

$confidenceInterval
$confidenceInterval$confidenceLevel
[1] 0.92

$confidenceInterval$lowerEndpointMatrix
[,1] [,2] [,3]
[1,] 0.32117517 0.1801276 0.2266730
[2,] 0.35036388 0.2810536 0.0000000
[3,] 0.08492395 0.0000000 0.6946682

$confidenceInterval$upperEndpointMatrix
[,1] [,2] [,3]
[1,] 0.5508869 0.3696812 0.4276070
[2,] 1.0000000 0.5612194 0.2404126
[3,] 0.2143174 0.1249744 1.0000000

$logLikelihood
[1] -170.3136

R> predictiveDistribution(weathersOfDays[1:200],
R+ weathersOfDays[201:365],hyperparam = hyperMatrix2)

[1] -163.8482

Note that the predictive probability is very small. However, this can be useful when comparing
model orders. Suppose we have an idea of the (prior) transition matrix corresponding to the
expected value of the parameters, and have a data set from which we want to deduce the
MAP estimates. We can infer the hyperparameters from this known transition matrix itself,
and use this to obtain our MAP estimates.

R> inferHyperparam(transMatr = weatherMatrix, scale = c(10, 10, 10))

$scaledInference
cloudy rain sunny
54 markovchain package: discrete Markov chains in R

cloudy 4 3 3
rain 4 4 2
sunny 2 1 7

Alternatively, we can use a data sample to infer the hyperparameters.

R> inferHyperparam(data = weathersOfDays[1:15])

$dataInference
cloudy sunny
cloudy 3 4
sunny 3 8

In order to use the inferred hyperparameter matrices, we do

R> hyperMatrix3 <- inferHyperparam(transMatr = weatherMatrix,


R+ scale = c(10, 10, 10))
R> hyperMatrix3 <- hyperMatrix3$scaledInference
R> hyperMatrix4 <- inferHyperparam(data = weathersOfDays[1:15])
R> hyperMatrix4 <- hyperMatrix4$dataInference

Now we can safely use hyperMatrix3 and hyperMatrix4 with markovchainFit (in the
hyperparam argument).
Supposing we don’t provide any hyperparameters, then the prior is uniform. This is the same
as maximum likelihood.

R> data(preproglucacon)
R> preproglucacon <- preproglucacon[[2]]
R> MLEest <- markovchainFit(preproglucacon, method = "mle")
R> MAPest <- markovchainFit(preproglucacon, method = "map")
R> MLEest$estimate

MLE Fit
A 4 - dimensional discrete Markov Chain defined by the following states:
A, C, G, T
The transition matrix (by rows) is defined as follows:
A C G T
A 0.3585271 0.1434109 0.16666667 0.3313953
C 0.3840304 0.1558935 0.02281369 0.4372624
G 0.3053097 0.1991150 0.15044248 0.3451327
T 0.2844523 0.1819788 0.17667845 0.3568905

R> MAPest$estimate

Bayesian Fit
A 4 - dimensional discrete Markov Chain defined by the following states:
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 55

A, C, G, T
The transition matrix (by rows) is defined as follows:
A C G T
A 0.3585271 0.1434109 0.16666667 0.3313953
C 0.3840304 0.1558935 0.02281369 0.4372624
G 0.3053097 0.1991150 0.15044248 0.3451327
T 0.2844523 0.1819788 0.17667845 0.3568905

6. Applications
This section shows applications of DTMC in various fields.

6.1. Weather forecasting


Markov chains provide a simple model to predict the next day’s weather given the current
meteorological condition. The first application herewith shown is the “Land of Oz example”
from J. G. Kemeny, J. L.Snell, and G. L. Thompson (1974), the second is the “Alofi Island
Rainfall” from P. J. Avery and D. A. Henderson (1999).

Land of Oz
The Land of Oz is acknowledged not to have ideal weather conditions at all: the weather is
snowy or rainy very often and, once more, there are never two nice days in a row. Consider
three weather states: rainy, nice and snowy. Let the transition matrix be as in the following:

R> mcWP <- new("markovchain", states = c("rainy", "nice", "snowy"),


R+ transitionMatrix = matrix(c(0.5, 0.25, 0.25,
R+ 0.5, 0, 0.5,
R+ 0.25,0.25,0.5), byrow = T, nrow = 3))

Given that today it is a nice day, the corresponding stochastic row vector is w0 = (0 , 1 , 0)
and the forecast after 1, 2 and 3 days are given by

R> W0 <- t(as.matrix(c(0, 1, 0)))


R> W1 <- W0 * mcWP; W1

rainy nice snowy


[1,] 0.5 0 0.5

R> W2 <- W0 * (mcWP ^ 2); W2

rainy nice snowy


[1,] 0.375 0.25 0.375

R> W3 <- W0 * (mcWP ^ 3); W3


56 markovchain package: discrete Markov chains in R

rainy nice snowy


[1,] 0.40625 0.1875 0.40625

As can be seen from w1 , if in the Land of Oz today is a nice day, tomorrow it will rain or
snow with probability 1. One week later, the prediction can be computed as

R> W7 <- W0 * (mcWP ^ 7)


R> W7

rainy nice snowy


[1,] 0.4000244 0.1999512 0.4000244

The steady state of the chain can be computed by means of the steadyStates method.

R> q <- steadyStates(mcWP)


R> q

rainy nice snowy


[1,] 0.4 0.2 0.4

Note that, from the seventh day on, the predicted probabilities are substantially equal to the
steady state of the chain and they don’t depend from the starting point, as the following code
shows.

R> R0 <- t(as.matrix(c(1, 0, 0)))


R> R7 <- R0 * (mcWP ^ 7); R7

rainy nice snowy


[1,] 0.4000244 0.2000122 0.3999634

R> S0 <- t(as.matrix(c(0, 0, 1)))


R> S7 <- S0 * (mcWP ^ 7); S7

rainy nice snowy


[1,] 0.3999634 0.2000122 0.4000244

Alofi Island Rainfall


Alofi Island daily rainfall data were recorded from January 1st, 1987 until December 31st,
1989 and classified into three states: “0” (no rain), “1-5” (from non zero until 5 mm) and “6+”
(more than 5mm). The corresponding dataset is provided within the markovchain package.

R> data("rain", package = "markovchain")


R> table(rain$rain)
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 57

0 1-5 6+
548 295 253

The underlying transition matrix is estimated as follows.

R> mcAlofi <- markovchainFit(data = rain$rain, name = "Alofi MC")$estimate


R> mcAlofi

Alofi MC
A 3 - dimensional discrete Markov Chain defined by the following states:
0, 1-5, 6+
The transition matrix (by rows) is defined as follows:
0 1-5 6+
0 0.6605839 0.2299270 0.1094891
1-5 0.4625850 0.3061224 0.2312925
6+ 0.1976285 0.3122530 0.4901186

The long term daily rainfall distribution is obtained by means of the steadyStates method.

R> steadyStates(mcAlofi)

0 1-5 6+
[1,] 0.5008871 0.2693656 0.2297473

6.2. Finance and Economics


Other relevant applications of DTMC can be found in Finance and Economics.

Finance
Credit ratings transitions have been successfully modelled with discrete time Markov chains.
Some rating agencies publish transition matrices that show the empirical transition probabi-
lities across credit ratings. The example that follows comes from CreditMetrics R package
(Wittmann 2007), carrying Standard & Poor’s published data.

R> rc <- c("AAA", "AA", "A", "BBB", "BB", "B", "CCC", "D")
R> creditMatrix <- matrix(
R+ c(90.81, 8.33, 0.68, 0.06, 0.08, 0.02, 0.01, 0.01,
R+ 0.70, 90.65, 7.79, 0.64, 0.06, 0.13, 0.02, 0.01,
R+ 0.09, 2.27, 91.05, 5.52, 0.74, 0.26, 0.01, 0.06,
R+ 0.02, 0.33, 5.95, 85.93, 5.30, 1.17, 1.12, 0.18,
R+ 0.03, 0.14, 0.67, 7.73, 80.53, 8.84, 1.00, 1.06,
R+ 0.01, 0.11, 0.24, 0.43, 6.48, 83.46, 4.07, 5.20,
R+ 0.21, 0, 0.22, 1.30, 2.38, 11.24, 64.86, 19.79,
R+ 0, 0, 0, 0, 0, 0, 0, 100
R+ )/100, 8, 8, dimnames = list(rc, rc), byrow = TRUE)
58 markovchain package: discrete Markov chains in R

It is easy to convert such matrices into markovchain objects and to perform some analyses

R> creditMc <- new("markovchain", transitionMatrix = creditMatrix,


R+ name = "S&P Matrix")
R> absorbingStates(creditMc)

[1] "D"

Economics
For a recent application of markovchain in Economic, see Jacob (2014).
A dynamic system generates two kinds of economic effects (Bard 2000):

1. those incurred when the system is in a specified state, and


2. those incurred when the system makes a transition from one state to another.

Let the monetary amount of being in a particular state be represented as a m-dimensional


column vector cS , while let the monetary amount of a transition be embodied in a C R matrix
in which each component specifies the monetary amount of going from state i to state j in a
single step. Henceforth, Equation (22) represents the monetary of being in state i.

m
X
ci = cSi + R
Cij pij . (22)
j=1

Let c̄ = [ci ] and let ei be the vector valued 1 in the initial state and 0 in all other, then, if fn is
the random variable representing the economic return associated with the stochastic process
at time n, Equation (23) holds:

E [fn (Xn ) |X0 = i] = ei P n c̄. (23)

The following example assumes that a telephone company models the transition probabilities
between customer/non-customer status by matrix P and the cost associated to states by
matrix M .

R> statesNames <- c("customer", "non customer")


R> P <- zeros(2); P[1, 1] <- .9; P[1, 2] <- .1; P[2, 2] <- .95; P[2, 1] <- .05;
R> rownames(P) <- statesNames; colnames(P) <- statesNames
R> mcP <- new("markovchain", transitionMatrix = P, name = "Telephone company")
R> M <- zeros(2); M[1, 1] <- -20; M[1, 2] <- -30; M[2, 1] <- -40; M[2, 2] <- 0

If the average revenue for existing customer is +100, the cost per state is computed as follows.

R> c1 <- 100 + conditionalDistribution(mcP, state = "customer") %*% M[1,]


R> c2 <- 0 + conditionalDistribution(mcP, state = "non customer") %*% M[2,]
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 59

For an existing customer, the expected gain (loss) at the fifth year is given by the following
code.

R> as.numeric((c(1, 0)* mcP ^ 5) %*% (as.vector(c(c1, c2))))

[1] 48.96009

6.3. Actuarial science


Markov chains are widely applied in the field of actuarial science. Two classical applications
are policyholders’ distribution across Bonus Malus classes in Motor Third Party Liability
(MTPL) insurance (Section 6.3.1) and health insurance pricing and reserving (Section 6.3.2).

MPTL Bonus Malus


Bonus Malus (BM) contracts grant the policyholder a discount (enworsen) as a function of
the number of claims in the experience period. The discount (enworsen) is applied on a pre-
mium that already allows for known (a priori) policyholder characteristics (Denuit, Maréchal,
Pitrebois, and Walhin 2007) and it usually depends on vehicle, territory, the demographic
profile of the policyholder, and policy coverages deep (deductible and policy limits).\ Since
the proposed BM level depends on the claim on the previous period, it can be modelled by a
discrete Markov chain. A very simplified example follows. Assume a BM scale from 1 to 5,
where 4 is the starting level. The evolution rules are shown in Equation 24:
     
bmt+1 = max (1, bmt − 1) ∗ Ñ = 0 + min 5, bmt + 2 ∗ Ñ ∗ Ñ ≥ 1 . (24)

The number of claim Ñ is a random variable that is assumed to be Poisson distributed.

R> getBonusMalusMarkovChain <- function(lambda) {


R+ bmMatr <- zeros(5)
R+ bmMatr[1, 1] <- dpois(x = 0, lambda)
R+ bmMatr[1, 3] <- dpois(x = 1, lambda)
R+ bmMatr[1, 5] <- 1 - ppois(q = 1, lambda)
R+
R+ bmMatr[2, 1] <- dpois(x = 0, lambda)
R+ bmMatr[2, 4] <- dpois(x = 1, lambda)
R+ bmMatr[2, 5] <- 1 - ppois(q = 1, lambda)
R+
R+ bmMatr[3, 2] <- dpois(x = 0, lambda)
R+ bmMatr[3, 5] <- 1 - dpois(x=0, lambda)
R+
R+ bmMatr[4, 3] <- dpois(x = 0, lambda)
R+ bmMatr[4, 5] <- 1 - dpois(x = 0, lambda)
R+
R+ bmMatr[5, 4] <- dpois(x = 0, lambda)
R+ bmMatr[5, 5] <- 1 - dpois(x = 0, lambda)
60 markovchain package: discrete Markov chains in R

R+ stateNames <- as.character(1:5)


R+ out <- new("markovchain", transitionMatrix = bmMatr,
R+ states = stateNames, name = "BM Matrix")
R+ return(out)
R+ }

Assuming that the a-priori claim frequency per car-year is 0.05 in the class (being the class
the group of policyholders that share the same common characteristics), the underlying BM
transition matrix and its underlying steady state are as follows.

R> bmMc <- getBonusMalusMarkovChain(0.05)


R> as.numeric(steadyStates(bmMc))

[1] 0.895836079 0.045930498 0.048285405 0.005969247 0.003978772

If the underlying BM coefficients of the class are 0.5, 0.7, 0.9, 1.0, 1.25, this means that the
average BM coefficient applied on the long run to the class is given by

R> sum(as.numeric(steadyStates(bmMc)) * c(0.5, 0.7, 0.9, 1, 1.25))

[1] 0.534469

This means that the average premium paid by policyholders in the portfolio almost halves in
the long run.

Health insurance example


Actuaries quantify the risk inherent in insurance contracts evaluating the premium of insu-
rance contract to be sold (therefore covering future risk) and evaluating the actuarial reserves
of existing portfolios (the liabilities in terms of benefits or claims payments due to policyholder
arising from previously sold contracts), see Deshmukh (2012) for details.
An applied example can be performed using the data from De Angelis, Paolo and Di Falco,
L. (2016) that has been saved in the exdata folder.

R> ltcDemoPath<-system.file("extdata", "ltdItaData.txt",


R+ package = "markovchain")
R> ltcDemo<-read.table(file = ltcDemoPath, header=TRUE,
R+ sep = ";", dec = ".")
R> head(ltcDemo)

age pAD pID pAI pAA


1 20 0.0004616002 0.01083364 0.0001762467 0.9993622
2 21 0.0004824888 0.01079719 0.0001710577 0.9993465
3 22 0.0004949938 0.01177076 0.0001592333 0.9993458
4 23 0.0005042935 0.01159394 0.0001605731 0.9993351
5 24 0.0005074193 0.01260574 0.0001606504 0.9993319
6 25 0.0005154267 0.01526364 0.0001643603 0.9993202
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 61

The data shows the probability of transition between the state of (A)ctive, to (I)ll and Dead.
It is easy to complete the transition matrix.

R> ltcDemo<-transform(ltcDemo,
R+ pIA=0,
R+ pII=1-pID,
R+ pDD=1,
R+ pDA=0,
R+ pDI=0)

Now we build a function that returns the transition during the t + 1 th year, assuming that
the subject has attained year t.

R> possibleStates<-c("A","I","D")
R> getMc4Age<-function(age) {
R+ transitionsAtAge<-ltcDemo[ltcDemo$age==age,]
R+
R+ myTransMatr<-matrix(0, nrow=3,ncol = 3,
R+ dimnames = list(possibleStates, possibleStates))
R+ myTransMatr[1,1]<-transitionsAtAge$pAA[1]
R+ myTransMatr[1,2]<-transitionsAtAge$pAI[1]
R+ myTransMatr[1,3]<-transitionsAtAge$pAD[1]
R+ myTransMatr[2,2]<-transitionsAtAge$pII[1]
R+ myTransMatr[2,3]<-transitionsAtAge$pID[1]
R+ myTransMatr[3,3]<-1
R+
R+ myMc<-new("markovchain", transitionMatrix = myTransMatr,
R+ states = possibleStates,
R+ name = paste("Age",age,"transition matrix"))
R+
R+ return(myMc)
R+
R+ }

Cause transitions are not homogeneous across ages, we use a markovchainList object to
describe the transition probabilities for a guy starting at age 100.

R> getFullTransitionTable<-function(age){
R+ ageSequence<-seq(from=age, to=120)
R+ k=1
R+ myList=list()
R+ for ( i in ageSequence) {
R+ mc_age_i<-getMc4Age(age = i)
R+ myList[[k]]<-mc_age_i
R+ k=k+1
R+ }
62 markovchain package: discrete Markov chains in R

R+ myMarkovChainList<-new("markovchainList", markovchains = myList,


R+ name = paste("TransitionsSinceAge", age, sep = ""))
R+ return(myMarkovChainList)
R+ }
R> transitionsSince100<-getFullTransitionTable(age=100)

We can use such transition for simulating ten life trajectories for a guy that begins “active”
(A) aged 100:

R> rmarkovchain(n = 10, object = transitionsSince100,


R+ what = "matrix", t0 = "A", include.t0 = TRUE)

[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10] [,11] [,12] [,13]
[1,] "A" "A" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D"
[2,] "A" "A" "A" "A" "A" "A" "A" "A" "D" "D" "D" "D" "D"
[3,] "A" "A" "A" "A" "D" "D" "D" "D" "D" "D" "D" "D" "D"
[4,] "A" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D"
[5,] "A" "A" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D"
[6,] "A" "A" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D"
[7,] "A" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D"
[8,] "A" "A" "A" "D" "D" "D" "D" "D" "D" "D" "D" "D" "D"
[9,] "A" "A" "A" "A" "A" "D" "D" "D" "D" "D" "D" "D" "D"
[10,] "A" "A" "A" "A" "A" "D" "D" "D" "D" "D" "D" "D" "D"
[,14] [,15] [,16] [,17] [,18] [,19] [,20] [,21] [,22]
[1,] "D" "D" "D" "D" "D" "D" "D" "D" "D"
[2,] "D" "D" "D" "D" "D" "D" "D" "D" "D"
[3,] "D" "D" "D" "D" "D" "D" "D" "D" "D"
[4,] "D" "D" "D" "D" "D" "D" "D" "D" "D"
[5,] "D" "D" "D" "D" "D" "D" "D" "D" "D"
[6,] "D" "D" "D" "D" "D" "D" "D" "D" "D"
[7,] "D" "D" "D" "D" "D" "D" "D" "D" "D"
[8,] "D" "D" "D" "D" "D" "D" "D" "D" "D"
[9,] "D" "D" "D" "D" "D" "D" "D" "D" "D"
[10,] "D" "D" "D" "D" "D" "D" "D" "D" "D"

Lets consider 1000 simulated live trajectories, for a healty guy aged 80. We can compute the
expected time a guy will be disabled starting active at age 80.

R> transitionsSince80<-getFullTransitionTable(age=80)
R> lifeTrajectories<-rmarkovchain(n=1e3, object=transitionsSince80,
R+ what="matrix",t0="A",include.t0=TRUE)
R> temp<-matrix(0,nrow=nrow(lifeTrajectories),ncol = ncol(lifeTrajectories))
R> temp[lifeTrajectories=="I"]<-1
R> expected_period_disabled<-mean(rowSums((temp)))
R> expected_period_disabled

[1] 1.156
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 63

Assuming that the health insurance will pay a benefit of 12000 per year disabled and that
the real interest rate is 0.02, we can compute the lump sum premium at 80.

R> mean(rowMeans(12000*temp%*%( matrix((1+0.02)^-seq(from=0, to=ncol(temp)-1)))))

[1] 11426.91

6.4. Sociology
Markov chains have been actively used to model progressions and regressions between social
classes. The first study was performed by Glass and Hall (1954), while a more recent applica-
tion can be found in Jo Blanden and Machin (2005). The table that follows shows the income
quartile of the father when the son was 16 (in 1984) and the income quartile of the son when
aged 30 (in 2000) for the 1970 cohort.

R> data("blanden")
R> mobilityMc <- as(blanden, "markovchain")
R> mobilityMc

Unnamed Markov chain


A 4 - dimensional discrete Markov Chain defined by the following states:
Bottom, 2nd, 3rd, Top
The transition matrix (by rows) is defined as follows:
2nd 3rd Bottom Top
Bottom 0.2900000 0.2200000 0.3800000 0.1100000
2nd 0.2772277 0.2574257 0.2475248 0.2178218
3rd 0.2626263 0.2828283 0.2121212 0.2424242
Top 0.1700000 0.2500000 0.1600000 0.4200000

The underlying transition graph is plotted in Figure 5.


The steady state distribution is computed as follows. Since transition across quartiles are
shown, the probability function is evenly 0.25.

R> round(steadyStates(mobilityMc), 2)

Bottom 2nd 3rd Top


[1,] 0.25 0.25 0.25 0.25

6.5. Genetics and Medicine


This section contains two examples: the first shows the use of Markov chain models in genetics,
the second shows an application of Markov chains in modelling diseases’ dynamics.

Genetics
P. J. Avery and D. A. Henderson (1999) discusses the use of Markov chains in model Prepro-
gucacon gene protein bases sequence. The preproglucacon dataset in markovchain contains
the dataset shown in the package.
64 markovchain package: discrete Markov chains in R

1970 mobility

2nd 0.29

0.11

0.17
0.38

0.22
Top 0.42

0.26 0.16

0.24 0.25

0.28
Bottom 0.21
0.22
0.28

0.25

3rd 0.26

Figure 5: 1970 UK cohort mobility data.

R> data("preproglucacon", package = "markovchain")

It is possible to model the transition probabilities between bases as shown in the following
code.

R> mcProtein <- markovchainFit(preproglucacon$preproglucacon,


R+ name = "Preproglucacon MC")$estimate
R> mcProtein

Preproglucacon MC
A 4 - dimensional discrete Markov Chain defined by the following states:
A, C, G, T
The transition matrix (by rows) is defined as follows:
A C G T
A 0.3585271 0.1434109 0.16666667 0.3313953
C 0.3840304 0.1558935 0.02281369 0.4372624
G 0.3053097 0.1991150 0.15044248 0.3451327
T 0.2844523 0.1819788 0.17667845 0.3568905

Medicine
Discrete-time Markov chains are also employed to study the progression of chronic diseases.
The following example is taken from B. A. Craig and A. A. Sendi (2002). Starting from six
month follow-up data, the maximum likelihood estimation of the monthly transition matrix is
obtained. This transition matrix aims to describe the monthly progression of CD4-cell counts
of HIV infected subjects.
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 65

R> craigSendiMatr <- matrix(c(682, 33, 25,


R+ 154, 64, 47,
R+ 19, 19, 43), byrow = T, nrow = 3)
R> hivStates <- c("0-49", "50-74", "75-UP")
R> rownames(craigSendiMatr) <- hivStates
R> colnames(craigSendiMatr) <- hivStates
R> craigSendiTable <- as.table(craigSendiMatr)
R> mcM6 <- as(craigSendiTable, "markovchain")
R> mcM6@name <- "Zero-Six month CD4 cells transition"
R> mcM6

Zero-Six month CD4 cells transition


A 3 - dimensional discrete Markov Chain defined by the following states:
0-49, 50-74, 75-UP
The transition matrix (by rows) is defined as follows:
0-49 50-74 75-UP
0-49 0.9216216 0.04459459 0.03378378
50-74 0.5811321 0.24150943 0.17735849
75-UP 0.2345679 0.23456790 0.53086420

As shown in the paper, the second passage consists in the decomposition of M6 = V · D · V −1


in order to obtain M1 as M1 = V · D1/6 · V −1 .

R> eig <- eigen(mcM6@transitionMatrix)


R> D <- diag(eig$values)

R> V <- eig$vectors


R> V %*% D %*% solve(V)

[,1] [,2] [,3]


[1,] 0.9216216 0.04459459 0.03378378
[2,] 0.5811321 0.24150943 0.17735849
[3,] 0.2345679 0.23456790 0.53086420

R> d <- D ^ (1/6)


R> M <- V %*% d %*% solve(V)
R> mcM1 <- new("markovchain", transitionMatrix = M, states = hivStates)

7. Discussion, issues and future plans


The markovchain package has been designed in order to provide easily handling of DTMC
and communication with alternative packages.
Some numerical issues have been found when working with matrix algebra using R internal
linear algebra kernel (the same calculations performed with MATLAB gave a more accurate
result). Some temporary workarounds have been implemented. For example, the condition
66 markovchain package: discrete Markov chains in R

for row/column sums to be equal to one is valid up to fifth decimal. Similarly, when extracting
the eigenvectors only the real part is taken.
Such limitations are expected to be overcome in future releases. Similarly, future versions of
the package are expected to improve the code in terms of numerical accuracy and rapidity.
An intitial rewriting of internal function in C++ by means of Rcpp package (Eddelbuettel
2013) has been started.

8. Aknowledgments
The package was selected for Google Summer of Code 2015 support. The authors wish to
thank Michael Cole, Tobi Gutman and Mildenberger Thoralf for their suggestions and bug
checks. A final thanks also to Dr. Simona C. Minotti and Dr. Mirko Signorelli for their
support in drafting this version of the vignettes.
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 67

References

Adamopoulou P (2018). MmgraphR: Graphing for Markov, Hidden Markov, and Mixture
Transition Distribution Models. R package version 0.3-1, URL https://CRAN.R-project.
org/package=MmgraphR.

Allaire J, Francois R, Ushey K, Vandenbrouck G, Geelnard M, Intel (2016). RcppParallel:


Parallel Programming Tools for ’Rcpp’. R package version 4.3.19, URL http://rcppcore.
github.io/RcppParallel/.

Anderson TW, Goodman LA (1957). “Statistical inference about Markov chains.” The Annals
of Mathematical Statistics, pp. 89–110.

B A Craig, A A Sendi (2002). “Estimation of the Transition Matrix of a Discrete-Time Markov


Chain.” Health Economics, 11, 33–42.

Bard JF (2000). “Lecture 12.5 - Additional Issues Concerning Discrete-Time


Markov Chains.” URL http://www.me.utexas.edu/~jensen%20/ORMM/instruction/
powerpoint/or_models_09/12.5_dtmc2.ppt.

Brémaud P (1999). “Discrete-Time Markov Models.” In Markov Chains, pp. 53–93. Springer.

Chambers J (2008). Software for Data Analysis: Programming with R. Statistics and compu-
ting. Springer-Verlag. ISBN 9780387759357.

Ching WK, Ng MK, Fung ES (2008). “Higher-order multivariate Markov chains and their
applications.” Linear Algebra and its Applications, 428(2), 492–507.

Csardi G, Nepusz T (2006). “The igraph Software Package for Complex Network Research.”
InterJournal, Complex Systems, 1695. URL http://igraph.sf.net.

De Angelis, Paolo, Di Falco, L (2016). Assicurazioni sulla salute: caratteristiche, modelli


attuariali e basi tecniche. Il Mulino. Il Mulino. ISBN 9788815260840. URL https://
books.google.it/books?id=D56bjgEACAAJ.

de Wreede LC, Fiocco M, Putter H (2011). “mstate: An R Package for the Analysis of
Competing Risks and Multi-State Models.” Journal of Statistical Software, 38(7), 1–30.
URL http://www.jstatsoft.org/v38/i07/.

Denuit M, Maréchal X, Pitrebois S, Walhin JF (2007). Actuarial modelling of claim counts:


Risk classification, credibility and bonus-malus systems. Wiley.

Deshmukh S (2012). Multiple Decrement Models in Insurance: An Introduction Using R.


SpringerLink : Bücher. Springer-Verlag. ISBN 9788132206590.

Dobrow RP (2016). Introduction to Stochastic Processes with R. John Wiley & Sons.

Eddelbuettel D (2013). Seamless R and C++ Integration with Rcpp. Springer-Verlag, New
York. ISBN 978-1-4614-6867-7.

Feres R (2007). “Notes for Math 450 MATLAB Listings for Markov Chains.” URL http:
//www.math.wustl.edu/~feres/Math450Lect04.pdf.
68 markovchain package: discrete Markov chains in R

Geyer CJ, Johnson LT (2013). mcmc: Markov Chain Monte Carlo. R package version 0.9-2,
URL http://CRAN.R-project.org/package=mcmc.

Glass D, Hall JR (1954). “Social Mobility in Great Britain: A Study in Intergenerational


Change in Status.” In Social Mobility in Great Britain. Routledge and Kegan Paul.

Goulet V, Dutang C, Maechler M, Firth D, Shapira M, Stadelmann M, expm-


developers@listsR-forgeR-projectorg (2013). expm: Matrix Exponential. R package version
0.99-1, URL http://CRAN.R-project.org/package=expm.

Grinstead CM, Snell LJ (2006). Grinstead and Snell’s Introduction to Probability. Version
dated 4 july 2006 edition. American Mathematical Society. URL http://math.dartmouth.
edu/~{}prob/prob/prob.pdf.

Himmelmann SSDDL, wwwlinhicom (2010). HMM: HMM - Hidden Markov Models. R


package version 1.0, URL http://CRAN.R-project.org/package=HMM.

Hu YT, Kiesel R, Perraudin W (2002). “The estimation of transition matrices for sovereign
credit ratings.” Journal of Banking and Finance, 26(7), 1383–1406. ISSN 03784266. doi:
10.1016/S0378-4266(02)00268-6.

Israel RB, Rosenthal JS, Wei JZ (2001). “Finding generators for Markov chains via empirical
transition matrices, with applications to credit ratings.” Mathematical finance, 11(2), 245–
265.

J G Kemeny, J LSnell, G L Thompson (1974). Introduction to Finite Mathematics. Prentice


Hall.

Jackson CH (2011). “Multi-State Models for Panel Data: The msm Package for R.” Journal
of Statistical Software, 38(8), 1–29. URL http://www.jstatsoft.org/v38/i08/.

Jacob I (2014). “Is R Cost Effective?” Electronic. Presented on Manchester R Meet-


ing, URL http://www.rmanchester.org/Presentations/Ian%20Jacob%20-%20Is%20R%
20Cost%20Effective.pdf.

Jo Blanden PG, Machin S (2005). “Intergenerational Mobility in Europe and North America.”
Technical report, Center for Economic Performances. URL http://cep.lse.ac.uk/about/
news/IntergenerationalMobility.pdf.

Konstantopoulos T (2009). “Markov Chains and Random Walks.” Lecture notes.

Kullback S, Kupperman M, Ku H (1962). “Tests for Contingency Tables and Marltov Chains.”
Technometrics, 4(4), 573–608.

Montgomery J (2009). “Communication Classes.” URL http://www.ssc.wisc.edu/


~jmontgom/commclasses.pdf.

Nicholson W (2013). DTMCPack: Suite of Functions Related to Discrete-Time Discrete-


State Markov Chains. R package version 0.1-2, URL http://CRAN.R-project.org/
package=DTMCPack.
Giorgio Alfredo Spedicato, Tae Seung Kang, Sai Bhargav Yalamanchi, Deepak Yadav 69

Noé F, Schütte C, Vanden-Eijnden E, Reich L, Weikl TR (2009). “Constructing the equili-


brium ensemble of folding pathways from short off-equilibrium simulations.” Proceedings
of the National Academy of Sciences, 106(45), 19011–19016. ISSN 0027-8424, 1091-6490.
doi:10.1073/pnas.0905466106. URL https://www.pnas.org/content/106/45/19011.

P J Avery, D A Henderson (1999). “Fitting Markov Chain Models to Discrete State Series.”
Applied Statistics, 48(1), 53–61.

Pfeuffer M (2017). ctmcd: Estimating the Parameters of a Continuous-Time Markov Chain


from Discrete-Time Data. R package version 1.1, URL https://CRAN.R-project.org/
package=ctmcd.

R Core Team (2013). R: A Language and Environment for Statistical Computing. R Founda-
tion for Statistical Computing, Vienna, Austria. URL http://www.R-project.org/.

Roebuck P (2011). matlab: MATLAB emulation package. R package version 0.8.9, URL
http://CRAN.R-project.org/package=matlab.

Sison CP, Glaz J (1995). “Simultaneous confidence intervals and sample size determination
for multinomial proportions.” Journal of the American Statistical Association, 90(429),
366–369.

Skuriat-Olechnowska M (2005). Statistical inference and hypothesis testing for Markov chains
with Interval Censoring. diploma thesis, Delft University of Technology.

Snell L (1999). “Probability Book: chapter 11.” URL http://www.dartmouth.edu/~chance/


teaching_aids/books_articles/probability_book/Chapter11.pdf.

Spedicato GA (2017). “Discrete Time Markov Chains with R.” The R Journal. URL https:
//journal.r-project.org/archive/2017/RJ-2017-036/index.html.

Villacorta PJ (2012). MultinomialCI: Simultaneous confidence intervals for multinomial


proportions according to the method by Sison and Glaz. R package version 1.0, URL
http://CRAN.R-project.org/package=MultinomialCI.

Wittmann A (2007). CreditMetrics: Functions for Calculating the CreditMetrics Risk Model.
R package version 0.0-2.

Wolfram Research I (2013a). URL http://www.wolfram.com/mathematica/new-in-9/


markov-chains-and-queues/structural-properties-of-finite-markov-processes.
html.

Wolfram Research I (2013b). Mathematica. Wolfram Research, Inc., ninth edition.


70 markovchain package: discrete Markov chains in R

Affiliation:
Giorgio Alfredo Spedicato
Ph.D C.Stat ACAS, UnipolSai R&D
Via Firenze 11 Paderno Dugnano 20037 Italy
E-mail: spedygiorgio@gmail.com
URL: www.statisticaladvisor.com

Tae Seung Kang


Ph.D student, Computer & Information Science & Engineering
University of Florida Gainesville, FL, USA
E-mail: tskang3@gmail.com

Sai Bhargav Yalamanchi


B-Tech student, Electrical Engineering
Indian Institute of Technology, Bombay Mumbai - 400 076, India
E-mail: bhargavcoolboy@gmail.com

Deepak Yadav
B-Tech student, Computer Science and Engineering
Indian Institute of Technology, Varanasi Uttar Pradesh - 221 005, India
E-mail: deepakyadav.iitbhu@gmail.com

You might also like