NOSMO Methodologendag, vrijdag 6 juni 2008, Amsterdam

 

 

 

Sessie

onderzoeksgroep

Longitudinale Modellen (LOMO)

 

“Continue tijd modellering van panel data”

 

Analyse van panel data vindt bijna altijd plaats met behulp van discrete-tijd methoden. Vanwege  de doorgaans grote observatie-intervallen in panel onderzoek (een à twee keer per jaar of zelfs minder frequent) is de gebruikelijke discrete-tijd analyse echter problematisch. Ongelijke intervallen maken vergelijking van de resultaten tussen en binnen onderzoeken onmogelijk. Maar zelfs bij gelijke intervallen impliceren gelijke discrete-tijd effecten niet dat de onderliggende continue-tijd effecten gelijk zijn. De bijdragen aan de sessie laten op verschillende manieren zien, hoe de valkuilen van een discrete-tijd analyse kunnen worden vermeden.      

 

 

 

 

13.30-14.00     Christian Steglich & Tom Snijders 

                        Continuous-Time Models forComplete Social Networks

 

14.00-14.30     Marc Delsing & Han Oud

 Analyzing Reciprocal Relationships by Means of  the

Continuous-Time Autoregressive Latent Trajectory (CALT) Model

 

14.30-15.00     Zita Oravecz & Francis Tuerlinckx

Analyzing Intensive Longitudinal Data with a Latent Stochastic Model

 

15.00-15.15    Pauze

 

15.15-15.45     Han Oud & Henk Folmer

Spatial Econometric Methods in Continuous Time Modeling of Panel Data with Latent Variables

 

 

 


Abstracts

 

Continuous-Time Models for Complete Social Networks

 

Christian Steglich

c.e.g.steglich@rug.nl

Tom Snijders

Faculty of Behavioural and Social Sciences

University of Groningen

 

In many applied social science disciplines, dynamics of complete social networks currently receive a lot of attention -- be it friendship networks among school children, or collaboration networks among firms. Although continuous-time process tracing data become increasingly available due to computer-assisted data collection, the most

common data format still is the panel design. Here, for a fixed group of actors, the complete network of social ties among them is assessed in a series of discrete observations over time. Differences between observation moments typically are very complex and difficult to model. Moreover, they usually can be explained by multiple, competing dynamic mechanisms. A continuous time modelling approach is introduced that enables researchers to decompose the observed total change into a stochastic sequence of smaller changes that are easier to model. In this framework, it also is easier to formulate hypotheses that test the alternative theoretical mechanisms. The approach comes at the cost of conditional independence assumptions typical for stochastic process models. As an example, the co-evolution of friendship networks at school and smoking uptake of the pupils is analysed. The proposed methods are implemented in the SIENA software  develped at the University of Groningen (see Snijders, Steglich & Schweinberger, 2007).

 

 

 

 

Analyzing Reciprocal Relationships by Means of  the

Continuous-Time Autoregressive Latent Trajectory (CALT) Model

 

Marc Delsing

m.delsing@uu.nl

Research Center Adolescent Development

Utrecht University

Han Oud

Behavioural Science Institute

Radboud University Nijmegen

 

Over the past decades, several analytic tools have become available for the analysis of reciprocal relations in a non-experimental context using structural equation modeling (SEM). A recently proposed model is the autoregressive latent trajectory (ALT) model (Bollen and Curran, 2004, Curran and Bollen, 2001), which captures features of both the autoregressive (AR) cross-lagged model and the latent trajectory (LT) model. In this presentation, strengths and weaknesses are discussed and it is demonstrated how several of the problems can be solved by a continuous time version: continuous time autoregressive latent trajectory (CALT) model. The EDM/SEM continuous time procedure, using SEM to estimate the exact discrete model (EDM), is applied to a CALT model of reciprocal relations between antisocial behavior and depressive symptoms.

 

 

 

 

Analyzing Intensive Longitudinal Data with a Latent Stochastic Model

 

Zita Oravecz

zita.oravecz@psy.kuleuven.be

Francis Tuerlinckx

Research Group Quantitative and Personality Psychology

Leuven University

 

Social sciences often aim to explore the within subject variation of certain quantities by repeatedly measuring them. Moreover, understanding the difference among individuals with respect to these variations has become a central issue. The talk describes an approach which incorporates both problems in a dynamical model. First of all, it takes advantage of a continuous time stochastic process, namely the Ornstein-Uhlenbeck process, to model the underlying dynamics in the longitudinal data of one individual. The approach involves simultaneous modelling of longitudinal variables; hence their dependency structure can also be explored. Second, a hierarchical extension of this core model is developed, allowing a complex investigation of interindividual differences for a group of individuals. For instance, the autocorrelation of the underlying process can be turned into a random effect. Furthermore, covariates can be introduced so that interindividual differences can be studied in many respects. The model can handle unbalanced datasets and is especially fit for the analysis of relatively long repeated measurement chains of several individuals. As an illustration, an application on a dataset collected in a diary study will also be presented.

 


 

Spatial Econometric Methods in Continuous Time Modeling of Panel Data with Latent Variables

 

Han Oud

j.oud@pwo.ru.nl

Behavioural Science Institute

Radboud University Nijmegen

Henk Folmer

Department of Spatial Sciences, University of Groningen

 Department of Social Sciences, Wageningen University

 

 

There are compelling arguments for continuous time modeling of panel data  (see, e.g., special issue 62:1, 2008, of Statistica Neerlandica).  Most social phenomena evolve in continuous time and analysis in discrete time (cross-lagged panel analysis) oversimplifies and even distorts reality. Discrete time analysis especially gets in trouble, when the observation intervals are unequal both within and between studies. A not less compelling  reason for analyzing cross-effects in continuous time is that equal effects found in discrete time do not guarantee at all that the underlying continuous time effects are equal. So, even in the case of equal observation intervals, discrete time analysis proves to be useless. The cross-lagged effects found in discrete time in fact are part of an ongoing process and one should conclude on the basis of the cross-lagged effect functions across time instead of the discrete time interval only.

             Spatial econometrics has been developed to analyze data, for which the central assumption of independence is not applicable. This is clear especially for spatial data, because less distant observation units in general will show higher correlation than more distant units. This kind of dependence is in no way restricted to spatial data, however, but often also present in survey data. We introduce spatial econometric methods for solving the unit dependence problem into the latent variable framework and the continuous time modeling of panel data. The methodology will be applied to German unemployment data over 4 years (200, 2001, 2002, 2003) in 439 regions.