Latent class models for time series analysis (Q4512133)

A method to analyse data consisting of \(N\) time series with measurements made or sampled at \(T\) times yielding a \(T\)-vector of values \(y_{n}\) \((n=1,2,\dots, N)\) is proposed in this paper. Thus the data form a \(T\) by \(N\) matrix \(Y.\) It is supposed that each of the \(n\) series belongs to one and only one of a smaller number \(S,\) of subpopulations or latent classes. It is not known in advance to which subpopulation a particular series belongs, hence it is referred to these subpopulations as latent classes. Latent class analysis of the \(N\) time series designed to classify and compare sets of series is discussed. For a particular time series \(n\) in latent class \(S\) the data \(y_{n}\) are independently normally distributed with a vector of means \(\mu=(\mu_{1}(t),\dots,\mu_{s}(t))'\) and common variance \(\sigma^2.\) The function of time, \(\mu_{s}(t),\) can be represented by a linear combination of low-order splines. The probability density function for the data of a time series is posited to be a finite mixture of spherical multivariate normal densities. The maximum-likelihood function is optimized by means of an EM algorithm. The stability of the estimates is investigated using a bootstrap procedure. Examples of real and artificial data are presented.