International Conference on Sequence Analysis and Related Methods Do different approaches in social science lead to divergent or convergent models Daniel Courgeau Ined International Conference on Sequence Analysis and Related Methods
I will present here the main ideas of my paper and will ask to those wanting more details to read its full version. After a cross-sectional approach which prevailed since Graunt, a longitudinal approach appeared at the end of World-War II We will see the questions raised by the more recent approaches and then try to answer them
First the approach by event duration permits to establish risk factors and to treat censored observations. The main questions it raises are: How to treat unobserved heterogeneity? What concept of probability use with it? How to avoid the risk of atomistic fallacy?
Second the approach by sequences permits a more complete description of life histories. The main questions it raises are: How to avoid the risk of finding regularities even if they do not exist? Could clusters be an artefact for not controlling for observed variables? How to introduce networks in sequence analysis?
Third the multilevel approach permits to analyse simultaneously individuals and broader levels resolving the antagonism between atomistic and ecological fallacies. Its main questions are: How to define more pertinent social contexts? How to introduce dependencies between members of a given group? How to identify the interactions that exist between the various levels?
Fourth multilevel network analysis permits to identify the interactions between different levels. New questions it raises: How to collect data with clearer guidelines? How to analyze the effect of individual and network characteristics How to analyze more deeply the effects of time effects in these studies.
Fifth agent-based models, based on computer simulation, are a bottom-up approach with population level behaviour emerging from individual rules. They raise new problems: How micro-level rules can find a link with aggregate-level ones? How a bottom-up appraoch can take into account a top-down effect? How to verify and validate their results?
In order to find a more synthetic view of these five approaches it is useful to consider two main concepts. The first one is the creation of a fictious statistical individual as distinct from an observed individual. The second one is the creation of a statistical network as distinct from an observed network.
The study of event duration and of event sequences are directly connected to the concept of statistical individual. The multilevel approach and the network approach are directly connected to the concept of statistical network. The different problems encountered when using one of these approaches may largely disappear when considering simultaneously the statistical individual and the statistical network.
It avoids the risks of atomistic and ecological fallacy. It avoids the problems linked to the choice between Bayesian or frequentist theory. It answers some problems posed by unobserved heterogenity. A number of problems encountered by sequence analysis may be solved by a network approach. The main problems encountered by multilevel analysis may be solved by a multilevel network analysis.
The situation is more complex for agent-based modeling. It focuses on individual behavior to explain a collective behavior. The explanation of emergence is too simplistic to detect social complexity. However it introduces a model for functional structure, but its basis is too arbitrary
We need to focus on the mechanisms generating a functional structure and not a causal structure. We may use Baconian induction consisting in discovering a system’s principle from a study of its properties. We must rely on the interactions between statistical individuals and networks and seek for their structure. The boundaries between social sciences are real, but Baconian induction is general.
Thank you very much for your attention and I will be pleased to answer to your comments or questions