Initial-State Distributions In State-Probability Estimation

From Transition Matrix to Population State Probabilities

A transition matrix describes how state membership changes over a specified interval. A population, however, may begin that interval spread across several states. To describe its state distribution at the end, the transition information must be combined with the population’s starting composition. This distinction matters in applications: two populations can have the same transition matrix but different state probabilities at follow-up because they began in different proportions of the states.

The calculation is a matrix-vector multiplication. Its interpretation depends on using a consistent convention for whether origins index rows or columns, and on aligning the initial proportions with the matrix’s state ordering. The result is a vector of estimated probabilities, one for each destination state.

The initial vector may represent everyone starting in one state, or a mixture. For example, \((1,0,0)\) represents a population entirely in state 1, while \((0.50,0.30,0.20)\) represents a population with 50% in state 1, 30% in state 2, and 20% in state 3. The entries must be nonnegative and sum to one. A matrix estimated from transition data is not itself the population’s state distribution: it describes how each origin state contributes to the destination states.

Why the Multiplication Gives the Mixture

For a fixed destination state \(k\), a person can be in that state at \(t\) after starting in any of the \(K\) states. The law of total probability adds those mutually exclusive possibilities. If the initial probability of state \(j\) is \(\pi_j(s)\), then the probability of starting in \(j\) and ending in \(k\) is \(\pi_j(s)P_{jk}(s,t)\). Summing over origins gives the destination probability:

This is exactly the \(k\)th component of the row-vector product \(\boldsymbol{\pi}(s)P(s,t)\). Replacing the population quantities and transition probabilities by their estimates gives the plug-in estimate in the definition. No new transition model is created by this step: the initial distribution supplies the weights, and the matrix supplies the origin-specific transition information.

The calculation assumes that the states are mutually exclusive and collectively exhaustive for the population at each time, that the initial vector and matrix use the same state labels and ordering, and that the matrix refers to the interval and target population of interest. Each row of a valid transition-probability matrix is nonnegative and sums to one. The initial vector must also be a probability vector. These conditions make the resulting vector a probability distribution. The interpretation of the estimated matrix itself relies on the assumptions of the method used to estimate it; multiplication does not remove those assumptions.

Changing the Initial Mix Changes the Result

The matrix entries describe how each origin state contributes to destination states; the initial distribution determines how much each row contributes. Consequently, the same estimated matrix can yield different population state probabilities when applied to populations with different starting compositions. This is not a contradiction: the populations are being assigned different weights across the origin states.

Checks, Alternatives, and Common Errors

Before interpreting a product, verify that it represents the intended calculation. The initial vector has \(K\) entries, the matrix is \(K\times K\), and both use exactly the same state ordering. The interval in the matrix must match the question, and the starting proportions must describe the population to which the result will apply. Check that the vector and matrix contain no negative probabilities, that the vector sums to one, and that each matrix row sums to one. Small discrepancies may arise from displayed rounding; large discrepancies suggest a coding, transcription, or estimation problem. A resulting vector that fails to sum to one is a warning to inspect these inputs and the multiplication.

If the target population starts entirely in one state, the calculation reduces to selecting that origin row. For a mixed population, multiplying by the initial vector is the weighted average of all rows. An unweighted average of rows is appropriate only when the starting proportions are equal. Likewise, weighting the columns instead of the rows reverses the intended roles under the row-origin convention. When available, directly estimating state proportions at follow-up is another way to obtain a marginal distribution; it answers the same kind of population question only if it refers to the same target population and time. It does not replace the transition matrix when the purpose is to apply transition information to a specified, possibly different, initial mix.

A further limitation is that a correct multiplication cannot repair an unsuitable matrix. If transition behavior differs across groups in ways not represented by the matrix, applying one common matrix to a mixed population may not describe that target adequately. Also, the calculation produces a point estimate; it does not, by itself, quantify sampling uncertainty. Uncertainty assessment requires methods beyond this multiplication.