Subgroup analysis are only exploratory unless such analyses were pre-specified in a study protocol at the design stage
For Bayesian analysis, selecting a prior or priors is challenging and there is no automatic method that will handle all situations.
In the absence of prior (empirical) information, we recommend low-information priors, ones that provide support for a broad range of values. Even here care is needed, especially to respect the scale of a parameter. For example, a N(0, 10,000) can be either low information or high information depending on the scale of the regressor and consequently the slope. Similarly, a Uniform(0,10) prior for a sampling SD can be either low or high information depending on the measurement scale. We recommend careful attention to these issues.
If empirical information is available, we recommend summarizing it by a probability distribution, for example the normalized likelihood or if on the estimate and standard error are given, a N(estimate, c x SE), with c>1, a ``power prior" penalizing to reflect that the external data are not completely fungible with that in the current study. If the analyst also determined that the external data were biased relative to the current study, a N(estimate+delta, c x SE) prior can be used, however potential bias can also be captured by selecting an appropriate value for c.
As for all important analyses, we recommend sensitivity analyses relative to these, very explicit analytic choices.
More details are provided in Henderson et al. (2016), Bayesian analysis of heterogeneous treatment effects for patient-centered outcomes research, Health Serv Outcomes Res Method, 16: 213-233.