prior sensitivity
Definition
Prior sensitivity refers to the degree to which statistical inferences change in response to different specifications of prior distributions in a Bayesian analysis. In the context of Bayesian graphical modeling for psychological network data, this concept applies to two distinct sets of priors: those placed on the network structure and those placed on the edge weight parameters. The scale of the prior distribution on partial correlations is a particularly consequential specification, because even small variations in that scale can substantially alter the inclusion Bayes factor and its capacity to distinguish between the presence and absence of an edge. These effects interact with properties of the data such as sample size, number of variables, number of ordinal categories, and network density, making a simulation-based approach necessary for characterizing how sensitive conclusions about conditional independence are to prior choice.
Sources: Sekulovski et al. (2024)
Related Terms
- network psychometrics (1 shared article)
- Bayes factor (1 shared article)
- Bayesian model averaging (1 shared article)
- Bayesian variable selection (1 shared article)
- Markov random fields (1 shared article)
Applications
Prior Sensitivity and Inclusion Bayes Factor
The inclusion Bayes factor, used in Bayesian network psychometrics to test for conditional independence between pairs of variables, is directly affected by both the choice of prior on network structure and the prior on edge weight parameters. Simulation evidence shows that the scale of the prior on partial correlations is especially consequential, with small variations producing substantial changes in the Bayes factor's value and its ability to differentiate edge presence from edge absence.
Sources: Sekulovski et al. (2024)
Prior Sensitivity and Conditional Independence Testing
Conditional independence testing in Bayesian Markov Random Field models depends on prior specifications that researchers must set before analyzing empirical data. Because the network structure determines which edge weight parameters receive a prior and which are fixed to zero, sensitivity to prior choice propagates through both the structural and parametric components of the model, affecting conclusions about which psychological variables are directly related after controlling for the remaining variables in the network.
Sources: Sekulovski et al. (2024)
