Diagnosing When Biological Knowledge Dominates the Data

Why Compare Prior and Posterior Distributions?

A pathway effect or molecular interaction can appear compelling in a fitted model because the observed omics data support it, because biological knowledge made it plausible before the data were analyzed, or because both sources contribute. A posterior estimate alone does not show which explanation applies. Comparing the prior distribution with the posterior distribution makes the influence of the data more visible.

The purpose is not to label the use of biological knowledge as good or bad. A prior can stabilize estimates when data are sparse, and that may be desirable. The diagnostic question is whether a conclusion is being presented as an empirical finding when it is substantially retained from the prior. This matters especially when interpreting pathway effects or network edges for which the current experiment has limited information.

Definition: A prior distribution describes uncertainty about a model parameter before incorporating the current data. A likelihood describes how probable the observed data are at each parameter value. The posterior distribution is the distribution after combining prior information and the likelihood: $$ p(\theta\mid y) \propto p(y\mid\theta)p(\theta), $$ where $\theta$ is the parameter and $y$ denotes the observed data. A conclusion is prior-dominated when its posterior location, uncertainty, or practical classification is determined mainly by the prior because the likelihood contributes relatively little information.

This definition is diagnostic, not a universal numerical threshold. In a particular model, useful evidence includes how far the posterior moves from the prior, how much uncertainty contracts, and how strongly the likelihood favors different parameter values. A posterior that resembles its prior can indicate weak data, but resemblance by itself does not prove that the prior dominates: the data may be informative and agree closely with the prior. Likewise, a large movement can reflect strong data, prior-data conflict, or a modeling problem.

A Normal Model for a Pathway Effect

For a pathway-level effect $\theta$, suppose the estimate $y$ has a normal sampling distribution with known standard error $s_y$, and suppose the prior for $\theta$ is normal. The model assumptions are that the estimate is adequately represented by this sampling distribution, its standard error is appropriate, and the stated prior describes pre-data uncertainty on the same effect scale. Under these assumptions, the posterior is also normal.

Result: If $y\mid\theta\sim N(\theta,s_y^2)$ and $\theta\sim N(\mu_0,\tau_0^2)$, then $$ \theta\mid y\sim N(\mu_1,\tau_1^2), \qquad \tau_1^2=\left(\frac{1}{\tau_0^2}+\frac{1}{s_y^2}\right)^{-1}, \qquad \mu_1=\tau_1^2\left(\frac{\mu_0}{\tau_0^2}+\frac{y}{s_y^2}\right). $$ Here $\mu_0,\tau_0^2$ are the prior mean and variance, and $\mu_1,\tau_1^2$ are the posterior mean and variance.

The posterior mean is a precision-weighted average: the prior precision is $1/\tau_0^2$, and the data precision is $1/s_y^2$. In this model, the fraction of posterior precision contributed by the data is $\left(1/s_y^2\right)/\left(1/\tau_0^2+1/s_y^2\right)$. This fraction is a convenient diagnostic for this specific normal model; it is not a general measure of prior influence for arbitrary models.

Deriving the Posterior Weights

The normal result follows directly from Bayes' rule. Ignoring factors that do not depend on $\theta$, the prior density times the likelihood is proportional to

\exp\left\{-\frac{1}{2}\left[ \frac{(y-\theta)^2}{s_y^2} +\frac{(\theta-\mu_0)^2}{\tau_0^2} \right]\right\}.

Collecting the terms involving $\theta^2$ gives coefficient $1/s_y^2+1/\tau_0^2$. Collecting the linear terms gives $-2\theta(y/s_y^2+\mu_0/\tau_0^2)$. Completing the square therefore yields a normal density with variance equal to the reciprocal of the quadratic coefficient and mean equal to that variance times $y/s_y^2+\mu_0/\tau_0^2$. These are precisely $\tau_1^2$ and $\mu_1$ above. The derivation also shows why a precise prior can retain substantial influence when the data standard error is large.

Worked Example: A Pathway Effect With Informative Data

Suppose a hypothetical pathway effect, measured in log-fold-change units, has prior $\theta\sim N(0,0.16)$. A hypothetical omics analysis gives $y=0.30$ log-fold-change units with standard error $s_y=0.30$. Assume the normal likelihood and prior specified above are appropriate.

The prior precision is $1/0.16=6.25$, and the data precision is $1/0.09=11.11$. Thus the posterior variance and mean are

\tau_1^2=\frac{1}{6.25+11.11}=0.0576, \qquad \mu_1=0.0576\left(6.25(0)+11.11(0.30)\right)=0.192.

The posterior standard deviation is $\sqrt{0.0576}=0.240$ log-fold-change units. The data contribute $11.11/(6.25+11.11)=0.640$, or 64.0%, of posterior precision, while the prior contributes 36.0%. A 95% normal posterior interval is approximately $0.192\pm1.96(0.240)$, or $[-0.278,0.662]$ log-fold-change units. The posterior is pulled toward the prior mean of zero, but the estimate and most of the precision contribution come from the observed data. This example does not indicate a strong prior-dominated conclusion.

Worked Example: A Pathway Effect Retaining the Prior

Now suppose a hypothetical pathway has prior $\theta\sim N(0.60,0.04)$, with effects in log-fold-change units. The current data give $y=0.10$ with standard error $s_y=0.50$. Assume the same normal-model conditions hold.

The prior precision is $1/0.04=25$, and the data precision is $1/0.25=4$. Therefore,

\tau_1^2=\frac{1}{25+4}=0.03448, \qquad \mu_1=\frac{25(0.60)+4(0.10)}{29} =\frac{15.4}{29}=0.5310.

The posterior standard deviation is approximately $0.1857$ log-fold-change units, and the data contribute only $4/29=0.138$, or 13.8%, of posterior precision. The posterior mean remains close to the prior mean, moving from $0.60$ to $0.531$. Its approximate 95% interval is $[0.167,0.895]$ log-fold-change units. The likelihood estimate alone, $0.10$, has a 95% interval of approximately $[-0.880,1.080]$, so the data are weakly informative on this scale. A positive posterior estimate here should not be described as strong independent data confirmation of a positive pathway effect: most of the posterior precision and much of its location are inherited from the prior.

Diagnosing Prior Influence for an Interaction Edge

For a candidate network edge, let $E=1$ mean the edge is present and $E=0$ mean it is absent. A prior probability $P(E=1)$ can encode interaction knowledge. To compare prior and posterior beliefs, it is useful to work with odds. If the data evidence is $y$, Bayes' rule gives

\frac{P(E=1\mid y)}{P(E=0\mid y)} = \frac{P(E=1)}{P(E=0)} \times \frac{P(y\mid E=1)}{P(y\mid E=0)}.

The second factor is the likelihood ratio for the edge against its absence. A ratio near 1 means the evidence changes the prior odds little; a ratio far below or above 1 means the data provide evidence against or for the edge. This comparison assumes that the likelihood adequately represents how the observed evidence would arise under both edge states. It does not establish that the edge is causal.

Worked Example: A Network Edge With Weak, Conflicting Evidence

Consider a hypothetical edge with prior probability $P(E=1)=0.80$, based on biological knowledge. A hypothetical assay returns a negative result. Assume its sensitivity for an existing edge is 0.70 and its specificity is 0.80. Under this model, the probability of a negative result is $0.30$ if the edge exists and $0.80$ if it does not.

The likelihood ratio for this negative result is $0.30/0.80=0.375$. The prior odds are $0.80/0.20=4$, so the posterior odds are $4(0.375)=1.5$. Converting back to probability gives

P(E=1\mid y)=\frac{1.5}{1+1.5}=0.600.

The negative assay result moves the edge probability from 0.80 to 0.60, so it is evidence against the edge, even though the posterior probability remains above 0.50. On the log-odds scale, the prior contributes $\log(4)=1.386$, while the likelihood contributes $\log(0.375)=-0.981$; their sum is $\log(1.5)=0.405$. The prior contribution is larger in magnitude in this specific comparison, and the data do not overturn the prior preference. Reporting only that the posterior probability exceeds 0.50 would hide the direction and strength of the data's contribution.

Practical Diagnostics and Failure Modes

A useful diagnosis combines several views rather than relying on one summary. For pathway effects, plot prior and posterior densities on the same scale, and report the prior and posterior means, standard deviations, and relevant intervals. For edges, report prior and posterior probabilities and, when available, the likelihood ratio or the corresponding evidence on the log-odds scale. Relate each summary to a scientific decision, such as whether an effect exceeds a prespecified meaningful threshold or whether an edge is retained for follow-up.

Check what the data can identify.
Inspect the likelihood or data-only uncertainty and the effective information in the relevant omics measurements. A broad likelihood relative to the prior suggests that the data may not substantially update the prior.

Compare prior and posterior summaries.
Assess location, spread, and conclusions at scientifically meaningful thresholds. For the normal model above, compare prior and data precision explicitly; do not apply that formula outside its assumptions.

Check agreement and model fit.
Use prior predictive checks and inspect whether the observed data are surprising under the model. Prior-data conflict can signal a misspecified likelihood or a context-mismatched prior, rather than simply a prior-dominated posterior.

State the source of the conclusion.
Describe whether the data materially changed the prior assessment, and qualify conclusions that remain largely prior-supported. Avoid presenting posterior certainty as if it were entirely learned from the current experiment.

These checks have limitations. A small posterior shift is ambiguous when the data and prior agree: strong evidence in the same direction may leave the posterior near the prior. A large shift does not guarantee a well-calibrated result; it can arise from outliers, dependence among measurements, or a likelihood that omits important structure. Prior and posterior interval overlap is not a formal test of influence, and no universal overlap percentage defines domination. For complex hierarchical models, precision shares from a simple normal model generally do not apply; examine the relevant conditional likelihood, posterior summaries, and conclusions under clearly stated alternative assumptions.

Common Mistake

Common mistake: “The posterior probability of this edge is 0.60, so the data support the edge.” A posterior probability combines the prior and the likelihood; by itself it does not identify the data's contribution. In the worked edge example, the prior probability was 0.80 and the observed negative result reduced it to 0.60. The correction is to compare prior and posterior assessments and report the evidence supplied by the data, such as the likelihood ratio, when it is available.

The main alternative to this comparison is to report only a posterior summary, which is compact but conceals how much the data updated the prior. A data-only analysis can reveal the information in the likelihood, but may be unstable or imprecise when data are sparse and does not incorporate relevant prior knowledge. Prior-versus-posterior comparison complements both approaches: it makes information flow visible without claiming that one summary alone resolves model adequacy or scientific validity.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

In the normal pathway model, what does a small data fraction of posterior precision indicate, and what assumptions make that calculation valid?
A pathway prior has mean 0.4 and the posterior mean is 0.41. Is that alone enough to conclude that the prior dominated? Explain.
An edge has prior probability 0.25 and a data likelihood ratio of 4 in favor of its presence. Calculate its posterior probability.
Why is a posterior edge probability above 0.50 not, by itself, evidence that the current data support the edge?

Solutions. (1) It indicates that the prior supplies a larger share of posterior precision than the data in the specified conjugate normal model. The calculation assumes a normal likelihood with known standard error, a normal prior on the same parameter scale, and an appropriate likelihood and prior. (2) No. The data may be informative and agree with the prior, leaving the posterior mean close to it. Compare uncertainty and likelihood information as well as location. (3) Prior odds are $0.25/0.75=1/3$. Posterior odds are $(1/3)(4)=4/3$, so the posterior probability is $(4/3)/(1+4/3)=4/7\approx0.571$. (4) The posterior combines prior odds and the data likelihood. A prior probability above 0.50 can keep the posterior above 0.50 even when the data provide evidence against the edge; compare prior and posterior odds or examine the likelihood ratio.