# Confidence intervals for transition probabilities between known discrete states

I am trying to calculate confidence intervals for transition probabilities between a set of discrete states over time (involving multiple time steps).

Here is a toy example to get the transition probabilities:

# create a toy dataset
states<- c("A","B","C")
t1 <- sample(states,30,c(2/3,1/6,1/6),replace=T)
t2 <- sample(states,30,c(2/3,1/6,1/6),replace=T)
t3 <- sample(states,30,c(2/3,1/6,1/6),replace=T)
df.t <- data.frame(t1=t1,t2=t2,t3=t3)

transition.matrix.t1 <- table(t1,t2)
transition.matrix.t2 <- table(t2,t3)

prob.trans.t1 <- t(apply(transition.matrix.t1,1,function(x) x/sum(x)))
prob.trans.t2 <- t(apply(transition.matrix.t2,1,function(x) x/sum(x)))


Each row of df.t is a sample observation over 3 time periods. prob.trans.t1 contains the transition probabilities for the time step t1->t2.

In reality, I don't have any previous knowledge about the underlying probabilities and I only have a randomly picked sample. I would like to calculate a 95% confidence interval for the above transition probabilities transition.matrix.t1, transition.matrix.t2.

• prob.trans.t1 is a stochastic matrix, so I believe your question is how do I calculate confidence intervals for x/sum(x)? We can treat these as multinomial proprtions. This may help. Nov 26, 2013 at 17:02

prob.trans.t1 is a stochastic matrix, so I assume your question is "how do I calculate confidence intervals for x/sum(x)?". These can be treated as multinomial proportions. A discussion is provided in this link.

Also, here is an R package that implements a method for building simultaneous confidence intervals for the probabilities of a multinomial distribution given a set of observations, proposed by Sison and Glaz in their paper.

And finally, below is how it can be applied in your toy example:

require(MultinomialCI)
for (r in 1:nrow(transition.matrix.t1)){
m <- multinomialCI(transition.matrix.t1[r,], 0.05);
print(paste(colnames(transition.matrix.t1)[r], " to A: [", m[1,1], m[1,2], "]"));
print(paste(colnames(transition.matrix.t1)[r], " to B: [", m[2,1], m[2,2], "]"));
print(paste(colnames(transition.matrix.t1)[r], " to C: [", m[3,1], m[3,2], "]"));
}

 "A  to A: [ 0.5 0.90486071657815 ]"
 "A  to B: [ 0 0.27986071657815 ]"
 "A  to C: [ 0.0625 0.46736071657815 ]"
 "B  to A: [ 0.5 0.997113668612868 ]"
 "B  to B: [ 0 0.297113668612868 ]"
 "B  to C: [ 0.1 0.597113668612868 ]"
 "C  to A: [ 0.5 1 ]"
 "C  to B: [ 0 0.351000820903667 ]"
 "C  to C: [ 0 0.601000820903667 ]"

• Thanks for your answer. Using a multinomial distribution to calculate the CIs does make a lot of sense. I was pondering to sample multiple datasets using the initial distribution and the transition probabilities to obtain a bootstrapped CI. I will try it and compare both. Nov 26, 2013 at 20:04
• @means-to-meaning, yes it would be interesting to compare, pls do share if you have some benchmark results. Nov 27, 2013 at 9:57
• great answer! @ zhubarb Dec 14, 2022 at 1:22