Abstract
Theories are sets of causal relationships between constructs and their proxy indicator variables. Theories are tested and their numerical parameters are estimated using statistical models of latent and observed variables. A considerable amount of theoretical development in Information Systems occurs by theory extension or adaptation. Moreover, researchers are encouraged to reuse existing measurement instruments when possible. As a consequence, there are many cases when a relationship between two variables (latent and/or observed) is re-estimated in a new study with a new sample or in a new context. To aid in cumulative theory building, a re-estimation of parameters should take into account our prior knowledge about their likely values. In this paper, we show how Bayesian statistical models can provide a statistically sound way of incorporating prior knowledge into parameter estimation, allowing researchers to keep a “running tally” of the best estimates of model parameters.
Recommended Citation
Evermann, Joerg and Tate, Mary, "Bayesian Structural Equation Models for Cumulative Theory Building in Information Systems" (2012). AMCIS 2012 Proceedings. 1.
https://aisel.aisnet.org/amcis2012/proceedings/ResearchMethods/1
Bayesian Structural Equation Models for Cumulative Theory Building in Information Systems
Theories are sets of causal relationships between constructs and their proxy indicator variables. Theories are tested and their numerical parameters are estimated using statistical models of latent and observed variables. A considerable amount of theoretical development in Information Systems occurs by theory extension or adaptation. Moreover, researchers are encouraged to reuse existing measurement instruments when possible. As a consequence, there are many cases when a relationship between two variables (latent and/or observed) is re-estimated in a new study with a new sample or in a new context. To aid in cumulative theory building, a re-estimation of parameters should take into account our prior knowledge about their likely values. In this paper, we show how Bayesian statistical models can provide a statistically sound way of incorporating prior knowledge into parameter estimation, allowing researchers to keep a “running tally” of the best estimates of model parameters.