Hierarchical multistate models from population data: an application to parity statuses

Introduction

Multistate population models are widely used in demographic analyses, as they can reflect any number of states and any pattern of movements between those states. The first multistate analysis is apparently due to DuPasquier (1912), but multistate analyses received little attention until the 1970s (e.g., Rogers, 1975). In the 1980s, their nature and significance to demography was well established (Land & Rogers, 1982; Schoen, 1988). In general, a model with N living states has N² possible transfers: N(N − 1) between living states, and N from each state to the “dead” state.

Population data on the number of persons in each state at the beginning and end of an age/time interval can provide only N equations, N − 1 for transfers between living states and 1 for the overall rate of death. Various procedures have been advanced to estimate all of the elements of the matrix that transforms the population at the beginning of an interval into the population at the end of the interval.

The leading technique is iterative proportional fitting (IPF), also known as the Deming-Stephan procedure, bi-proportional filling, and the RAS method (Bishop, Fienberg & Holland, 1975; Willekens, 1982). IPF yields the solution that maximizes entropy, i.e., it finds the transfers that can occur in the greatest number of ways. Schoen & Jonsson (2003) advanced an alternative, more demographic, approach based on estimating changes in the attractiveness of model states. More recently, Schoen (2016) presented a Quadratic Estimation of Rates of Transfer (QERT) procedure, based on the assumption that the products of selected pairs of rates can be assumed constant. All three approaches assume a pattern that is believed to characterize the interstate movements, and thus produce estimated rather than calculated values.

This article focuses on “hierarchical” multistate models, that is on N state models with only N − 1 possible transfers connecting the living states and 1 unknown risk of death. In hierarchical models, a direct algebraic solution for the transfer matrix is possible, without the need to assume any pattern of interstate movements. We first discuss the parallels between hierarchical models and graphs, then present techniques for using sequential population data to calculate rates of transfer and transition probabilities, and proceed to apply the approach to calculate a Parity Status Life Table for US Women, 2005–10.

The nature of the hierarchical model

The N living states of a hierarchical model are linked by N − 1 occurrence/exposure rates of the form m_ij, where i is the origin state and j is the destination state. The dead state is indicated by the symbol δ. For every living state there is at least one connection to another living state, though all states are not necessarily reachable from any given state. For example, consider a model with two living states, A and B. With interstate transfer rate m_AB, it is possible to go from state A to state B, but not from state B to state A. The hierarchical model greatly restricts the range of possible interstate movements. It is a special case of the general multistate model, but one where the rates over an interval can be found from the populations at the beginning and end of the interval.

Graph theory can provide some useful nomenclature and context. The living states of a multistate model correspond to the vertices or nodes of a graph, and the rates connecting them are the edges or connecting lines of the graph. Hierarchical multistate models are connected graphs with N vertices and N − 1 edges, and hence constitute a “tree” (Chartrand & Zhang, 2012). Since the rates take persons from one specified state to another, the tree is “oriented” or, using the term coined by Rebane & Pearl (1987), a “polytree.”

Graph theorists have explored the number of distinct polytrees, or distinct multistate configurations, in models with N states (or vertices). When N = 2, there is only one form, with the single transfer rate m₁₂. When N = 3, there are 3 forms, which can be described by the rate pairs (m₁₂, m₂₃), (m₁₂, m₁₃), and (m₁₃, m₂₃). There is no general rule for the number of forms, but when N = 4 there are 8 forms, when N = 5 there are 27 forms, and when N = 6 there are 91 forms (Chartrand & Zhang, 2012). Polytrees thus encompass a substantial number of distinct hierarchical multistate models.

Polytrees have demographic significance because transfers in every multistate model that is a polytree can be fully determined from knowledge of the population, by state, at the beginning and end of every interval of observation. Such multistate models can be used to analyze of a wide range of demographic behaviors, including parity progression, educational attainment, career trajectories, and the spread of infections. A procedure for carrying out the calculations needed to produce a hierarchical multistate life table from cross-sectional data is presented in the next sections.

Finding multistate rates from beginning and ending populations

With N states, the beginning and ending populations support N equations that describe the flows (movements) between model states. Those N equations can be used to determine the N − 1 rates of transfer between those living states, and an overall rate of transfer to death. Each distinct hierarchical model (polytree) has a different set of N flow equations, but all those sets of equations are readily solvable because their N equations are independent.

For example, consider the case where N = 3 and the polytree form is (m₁₂, m₁₃). The flow equations can then be written ${\displaystyle P_1\left(x+n,n,t+n\right)=P_1\left(x,n,t\right)-P{P}_1\left(x,n,t\right)\left[m_{\delta }\left(x,n,t\right)+m_{12}\left(x,n,t\right)+m_{13}\left(x,n,t\right)\right]}$ ${\displaystyle P_2\left(x+n,n,t+n\right)=P_2\left(x,n,t\right)-P{P}_2\left(x,n,t\right)m_{\delta }\left(x,n,t\right)+P{P}_1\left(x,n,t\right)m_{12}\left(x,n,t\right)}$ (1)${\displaystyle P_3\left(x+n,n,t+n\right)=P_3\left(x,n,t\right)-P{P}_3\left(x,n,t\right)m_{\delta }\left(x,n,t\right)+P{P}_1\left(x,n,t\right)m_{13}\left(x,n,t\right)}$ where P_j(x, n, t) denotes the number of persons in state j between the ages of x and x + n at time t, and PP_j(x, n, t) is the number of person-years lived in state j between times t and t + n by persons aged x to x + n at time t. The person-year values can be found from the beginning and ending population values in a number of ways (Schoen, 1988, Chap. 4). Here we use the linear relationship (2)${\displaystyle P{P}_j\left(x,n,t\right)=\left(n∕2\right)\left[P_j\left(x,n,t\right)+P_j\left(x+n,n,t+n\right)\right]}$ that assumes that the P(x, n, t) function is linear between ages x and x + n. Even with 5-year age intervals, that linear assumption has been found to be satisfactory for most demographic work.

From Eqs. (1), straightforward algebra yields the solutions ${\displaystyle m_{\delta }\left(x,n,t\right)=\frac{P_1\left(x,n,t\right)+P_2\left(x,n,t\right)+P_3\left(x,n,t\right)-P_1\left(x+n,n,t+n\right)-P_2\left(x+n,n,t+n\right)-P_3\left(x+n,n,t+n\right)}{P{P}_1\left(x,n,t\right)+P{P}_2\left(x,n,t\right)+P{P}_3\left(x,n,t\right)}}$ ${\displaystyle m_{12}\left(x,n,t\right)=\frac{\left[P{P}_1+P{P}_3\right]\left[P_2\left(n\right)-P_2\left(0\right)\right]-P{P}_2\left[P_1\left(n\right)+P_3\left(n\right)-P_1\left(0\right)-P_3\left(0\right)\right]}{P{P}_1\left(x,n,t\right)\left[P{P}_1\left(x,n,t\right)+P{P}_2\left(x,n,t\right)+P{P}_3\left(x,n,t\right)\right]}}$ (3)${\displaystyle m_{13}\left(x,n,t\right)=\frac{\left[P{P}_1+P{P}_2\right]\left[P_3\left(n\right)-P_3\left(0\right)\right]-P{P}_3\left[P_1\left(n\right)+P_2\left(n\right)-P_1\left(0\right)-P_2\left(0\right)\right]}{P{P}_1\left(x,n,t\right)\left[P{P}_1\left(x,n,t\right)+P{P}_2\left(x,n,t\right)+P{P}_3\left(x,n,t\right)\right]}}$

where the age and time identifiers have been dropped from the PP_j functions in the numerators and the age and time identifiers of the beginning and ending populations are simply denoted by (0) and (n), respectively, to simplify the presentation. The overall death rate is just the beginning number of persons minus the ending number, divided by the total number of person-years lived in the interval. The expressions for the transfer rates are a bit more complicated, but represent a weighted difference in state sizes over the interval, divided by the origin state number of person-years. Note that there is no assumption restricting a person to only one transfer during an interval (cf. Schoen, 1988, Ch. 4).

When the transfer rates take persons only from state 1 to state 2, state 2 to state 3, etc., the model can be termed “strictly” hierarchical. The flow equations of an N state strictly hierarchical model are ${\displaystyle P_1\left(x+n,n,t+n\right)=P_1\left(x,n,t\right)-P{P}_1\left(x,n,t\right)\left[m_{\delta }\left(x,n,t\right)+m_{12}\left(x,n,t\right)\right]}$ ${\displaystyle P_j\left(x+n,n,t+n\right)=P_j\left(x,n,t\right)-P{P}_j\left(x,n,t\right)\left[m_{\delta }\left(x,n,t\right)+m_{j,j+1}\left(x,n,t\right)\right]+P{P}_{j-1}\left(x,n,t\right)m_{j-1,j}\left(x,n,t\right)}$ (4)${\displaystyle P_N\left(x+n,n,t+n\right)=P_N\left(x,n,t\right)-P{P}_N\left(x,n,t\right)m_{\delta }\left(x,n,t\right)+P{P}_{N-1}\left(x,n,t\right)m_{N-1,N}\left(x,n,t\right)}$ where 1 < j < N. Those flow equations yield the death and transfer rates (5a)${\displaystyle m_{\delta }\left(x,n,t\right)=\frac{{\mathrm{\Sigma }}_i\left[P_i\left(x,n,t\right)-P_i\left(x+n,n,t+n\right)\right]}{{\mathrm{\Sigma }}_i\left[P{P}_i\left(x,n,t\right)\right]}}$ (5b)${\displaystyle m_{j,j+1}\left(x,n,t\right)=\frac{{\mathrm{\Sigma }}_i\left[P_i\left(x,n,t\right)-P_i\left(x+n,n,t+n\right)-P{P}_i\left(x,n,t\right)m_{\delta }\left(x,n,t\right)\right]}{P{P}_j\left(x,n,t\right)}}$ where the sum in Eq. (5a) goes from 1 to N, and the sum in Eq. (5b) goes from 1 to j. The overall death rate, m_δ(x, n, t), is again the difference between the beginning and ending populations divided by the total number of person-years lived in the interval. The transfer rate from state j to state j + 1 is the difference between the beginning and ending populations through state j, less the losses to death through state j, divided by the number of person-years lived in state j during the interval.

In general, the flow equations that describe the specific polytree form (or the structure of the desired multistate model), plus the initial and final population numbers in each state, can yield all of the transfer rates of a hierarchical multistate model. With the rates known, all model functions can readily be found (cf. Schoen, 1988). The only assumptions involved are that the population data are accurate, the population is closed, there is the same rate of death (or attrition) from every state, and the linear calculation assumption is appropriate. All of those assumptions can be relaxed with additional information. For example, if mortality is known, the interstate transfer rates can be found from cross-sectional surveys of the beginning and ending populations.

Finding probabilities of transfer

The basic problem in life table construction is how to go from rates of transfer to probabilities of transfer. In multistate models, including hierarchical models, the problem is best approached by arraying the transfer rates in a matrix. With N living states, the N × N rate matrix can be written (6)${\displaystyle \mathbf{M}\left(\mathbf{x},\mathbf{n},\mathbf{t}\right)=\left[\begin{array}{ccccc}{\mathrm{\Sigma }}_j{m}_{1j}\hfill & -m_{12}\hfill & -m_{13}\hfill & \cdots \hfill & -m_{1N}\hfill \\ -m_{21}\hfill & {\mathrm{\Sigma }}_j{m}_{2j}\hfill & -m_{23}\hfill & \cdots \hfill & -m_{2N}\hfill \\ \cdots \hfill & \cdots \hfill & \cdots \hfill & \hfill & \cdots \hfill \\ -m_{N1}\hfill & -m_{N2}\hfill & -m_{N3}\hfill & \cdots \hfill & {\mathrm{\Sigma }}_j{m}_{Nj}\hfill \end{array}\right]}$ where the interval identifiers have been dropped from the m’s and the sums over j in the diagonal elements exclude m_jj and include the rate to the dead state. In hierarchical models, only N transfer rates, including m_δ, are nonzero.

There are a number of ways to calculate the array of transition probabilities from M. One is to assume that the underlying risks of transition are constant over the interval and find the probabilities by exponentiating M (cf. Schoen, 2006b, Chap.1). Here, we will continue to use the linear assumption underlying Eq. (2), and employ the general linear solution of Rogers & Ledent (1976, Eq. (2)), as written in Schoen (2006b, Chap. 1). We then have (7)${\displaystyle \mathbf{\Pi }\left(\mathbf{x},\mathbf{n},\mathbf{t}\right)=\left[\mathbf{I}-\left(n∕2\right)\mathbf{M}\left(\mathbf{x},\mathbf{n},\mathbf{t}\right)\right]{\left[\mathbf{I}+\left(n∕2\right)\mathbf{M}\left(\mathbf{x},\mathbf{n},\mathbf{t}\right)\right]}^{-1}}$ where I is the N × N identity matrix. Transition probability matrix Π(x, n, t) is an N × N matrix whose ijth element is π_ij(x, n, t), the probability that a person in state i at age x and time t will be in state j exactly n years later. In hierarchical models, Π can be written as an upper triangular matrix, that is a matrix where all of the elements below the main diagonal are zero. That can be accomplished by numbering the states so that every nonzero m_ij has i < j.

The age-specific transition probability matrices for a period allow the construction of a multistate life table for that period. That life table provides, among other functions, the number of persons in each state at every age. Those life table survivorship values follow from the projection relationship (8)${\displaystyle {\mathbf{\ell }}^{\mathbf{T}}\left(\mathbf{x}+\mathbf{n}\right)={\mathbf{\ell }}^{\mathbf{T}}\left(\mathbf{x}\right)\mathbf{\Pi }\left(\mathbf{x},\mathbf{n},\mathbf{t}\right)}$ where the jth element of the N element survivorship vector ℓ(x) is ℓ_j(x), the number of persons in the life table cohort who are in state j at exact age x. The superscript T indicates that column vector ℓ(x) is transposed into a row vector. The initial (radix) vector, ℓ(0), is assumed to be known. The remaining multistate life table functions follow in the usual manner (cf. Schoen, 1988, Chap. 4).

The calculation that provides the transition probabilities from observed adjacent population values proceeds from the observed populations to the rates of transfer and then to the transition probabilities. It is thus rate centered, not probability based. Rates are preferable because they directly reflect the underlying behavior, i.e., the risk of movement from state i to state j, while probabilities only reflect that behavior in the context of other prevailing risks. In the hierarchical model, there are only N − 1 nonzero rates of transfer between living states while, because of multiple moves within an interval, there can be N(N − 1)∕2 nonzero probabilities of transfer. That larger (for N > 2) number of probabilities cannot be determined from adjacent population data alone, unless the underlying rate structure is hierarchical. The proposed rate based approach accommodates multiple moves with an interval via the underlying Markov assumption that the risk of transfer depends only on current age and state.

An application to parity status life tables

The procedure described in the previous two sections can find the rates and probabilities of any hierarchical multistate model from initial and final population data alone. To illustrate the procedure numerically, let us consider an important demographic application: the parity status model.

Summary and conclusions

Many demographic phenomena can be represented by hierarchical models, that is by models with N living states that have N − 1 rates of movement between them. In such models, the rates of transfer between living states and to death can be found from data on the number of persons in each state at the beginning and end of a given interval. With the rates of transfer provided by Eq. (5), the probabilities of transfer can be found from Eq. (7) and used to produce a multistate life table. Exploiting the ability of population counts to infer rates of transfer in hierarchical contexts broadens the scope of demographic analysis.

The potential of the approach is illustrated by an application to parity status. A parity status life table for US Women, 2005–2010, is constructed from survey data on the fraction of women, by age, in each parity status in 2004, 2006, and 2010. The results show that the US proportion childless is continuing to increase, with nearly a quarter of women aged 42.5 remaining at parity zero. The 2–4 child pattern of the past has been replaced by a 0–3 child pattern.

Supplemental Information