A Simulation Based Hierarchical Bayesian Framework for Discovery Delay Risk in Multidisciplinary versus Disciplinary Open Access Repositories: A Conceptual and Computational Study

Discoverability differences between multidisciplinary open access repositories, for example Zenodo, and disciplinary platforms with tighter curation, for example SSRN, are widely discussed in the scholarly communication literature, but the mechanisms by which limited search engine indexing translates into a quantifiable discovery delay risk are rarely formalized. This paper is a conceptual and computational contribution. It introduces a hierarchical Bayesian model, coupled with an exponential and Weibull temporal hazard model, that formalizes discovery delay as a probabilistic quantity and demonstrates how such a model can be fitted, diagnosed, and stress tested using variance based global sensitivity analysis with Sobol indices, Monte Carlo and Latin Hypercube uncertainty propagation, and posterior predictive checks. We deliberately do not claim empirical validation against real per record OpenAIRE, Zenodo, or SSRN indexing timestamps, since no such dataset was obtained or processed in this study. Instead, all quantitative results reported here are produced by a fully reproducible, seeded simulation pipeline using illustrative hyperparameters that are qualitatively consistent with directional statements in the cited literature, namely shorter delays for curated disciplinary platforms, and not by fitting to observed repository timestamps. Under this illustrative parameterization, the model yields a posterior effect size of approximately 0.048 for the difference in delay probability between the two groups, with a 95 percent highest density interval from negative 0.077 to 0.186, a probability that the effect is positive of approximately 0.80, and a Savage Dickey Bayes factor of approximately 0.20, indicating weak evidence rather than the strong effect implied by an unqualified point estimate. The model also yields a simulated cumulative risk reduction at fifteen days of 0.0486, with a 95 percent confidence interval from 0.0482 to 0.0491. We report these results honestly, including where the evidence is weaker than a superficial reading of the point estimates would suggest, and we specify concrete, falsifiable empirical predictions that would be required to test the framework against real data. The contribution is therefore a validated, reproducible method and an explicit research roadmap, rather than an empirical finding about the indexing performance of Zenodo or SSRN.