Statistics Seminars at Imperial College Friday 10 March, Room 408, Department of Mathematics, 4th floor Huxley Building, 180 Queen's Gate, South Kensington, London. Program: 2:00 -- 3:00 Mikis Stasinopoulos, University of North London Generalized Autoregressive Moving Average Models 3:00 -- 3:30 Coffee Break (served in the same room - 408) 3:30 -- 4:30 Thomas Richardson, University of Warwick Ancestral Graph Markov models: an Alternative to Models with Latent or Selection Variables Abstracts of the talks: Generalized Autoregressive Moving Average models M. Stasinopoulos, University of North London A class of Generalized Autoregressive Moving Average (GARMA) models is developed which extends the univariate Gaussian ARMA time series model to a flexible observation driven model for non-Gaussian time series data. The dependent variable is assumed to have a conditional Exponential Family distribution given the past history of the process. It is demonstrated that model estimation can be carried out using an iteratively reweighted least squares algorithm. Properties of the model including stationarity and marginal moments are either derived explicitly or investigated using Monte Carlo simulation. The relationship of the GARMA model to other models is shown, including the autoregressive models of Zeger and Qaqish (1988), the moving average models of Li (1994) and also the reparameterised GARCH model (providing the formula for its fourth marginal moment not previously derived). The model is demonstrated by two applications of the GARMA model. The first application is to a well known time series data set of poliomyelitis counts using a Negative Binomial conditional distribution. The second is to river flow data using a Gamma conditional distribution. Ancestral Graph Markov models: an Alternative to Models with Latent or Selection Variables Thomas S. Richardson,University of Warwick A graphical model is a set of multivariate distributions obeying the conditional independence properties encoded by a graph. Two types of graph have commonly been used: undirected graphs and directed acyclic graphs. Graphical models are often used for the purpose of gaining insight into possible mechanisms which may have generated a given set of data. Such interpretation of the model becomes inherently more difficult when unmeasured `confounding' variables may be present, or when there may be bias in the sampling procedure. Latent variable models present one possible approach in these circumstances, but lack many of the nice statistical properties associated with standard graphical models. This talk will present the class of maximal ancestral graph (MAG) models which provide an alternative approach to this problem. A MAG can be used to represent the Markov structure among the observed variables in a data generating process. In the Gaussian case MAG models can be parametrized via a recursive set of linear equations with correlated errors. MAG models also retain many of the desirable properties of models based on directed and undirected graphs. EVERYBODY WELCOME!!!!!!!!!!!!! %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%