Shared priors, shrinkage and the limits of fixed pooling

Last updated: 5 Oct 20265 min read

tutorial

A shared prior pulls sparse group estimates toward a common rate, but a fixed shared prior is only a baseline for a hierarchical model.

Identify the group comparison

In one period, the North queue has four escalations in six audited cases; South has two in 41. Raw shares are 66.7% and 4.9%, respectively. North is so small that a raw rank can swing with one case. Yet pooling all 47 cases into a single 12.8% rate hides the observed difference. The question is whether each queue needs its own staffing response and how much uncertainty that response can tolerate.

Apply a fixed shared prior as a baseline

Using Beta(3, 27) independently for each queue yields posterior means 7/36, about 19.4%, for North and 5/71, about 7.0%, for South. The small North sample moves more toward the prior mean than a large sample would. This is shrinkage under a fixed prior whose parameters came from an earlier comparable period; it is not a jointly fitted hierarchical posterior. The update rule is unchanged.

Distinguish actual hierarchical learning

In a hierarchical model, group rates share a population distribution and its parameters are estimated with uncertainty. The data help determine how much pooling is appropriate. A fixed shared prior cannot learn that groups are more dispersed than expected. If North truly operates under a distinct policy or customer mix, forcing it toward the common mean may hide an issue. Group definitions and outcome windows must be comparable before sharing information.

Check the prior against both groups

The North observation of four events in six cases may be surprising under a prior centered at 10%. Examine the prior predictive probability of such counts and posterior predictive replication. It may reflect random noise, a changed queue, or a poor prior. A posterior mean alone cannot tell which. Prior predictive checks and an operational investigation should accompany escalation.

Decide with uncertainty, not a league table

Publish each group’s event count, audited denominator, posterior distribution and prior source. If a small group crosses an action threshold only under one prior choice, mark the decision as sensitive and collect more cases. Avoid ranking groups by point estimates without accounting for the large and unequal uncertainty. Threshold probabilities can be reported per group with one predeclared rule.

Implementation

def fixed_prior_group_means(group_counts, prior_alpha=3, prior_beta=27):
    if prior_alpha <= 0 or prior_beta <= 0:
        raise ValueError("prior shapes must be positive")
    means = {}
    for queue_name, (escalated, audited) in group_counts.items():
        if not 0 <= escalated <= audited:
            raise ValueError("invalid group counts")
        means[queue_name] = (prior_alpha + escalated) / (prior_alpha + prior_beta + audited)
    return means

queue_means = fixed_prior_group_means({"North": (4, 6), "South": (2, 41)})
assert round(queue_means["North"], 3) == 0.194
assert round(queue_means["South"], 3) == 0.070

Performance and operating cost

Fixed-prior updates cost O(G) time and output space for G groups. A full hierarchical model requires fitting shared parameters, checking numerical diagnostics and simulating predictions; its cost depends on group count, parameterization and inference method.

Common Mistakes

Do not label fixed-prior updates as a fitted hierarchical model.
Do not shrink a genuinely different operating group into an unrelated population.
Do not rank tiny groups by raw event share without showing uncertainty.