Why Prior Justification Matters
A prior distribution is part of the statistical model, not a neutral formatting choice. It represents information about an unknown quantity before the trial data are analyzed, and it can affect posterior estimates and decisions. In a regulatory submission, reviewers therefore need more than the prior’s equation: they need to understand where it came from, why it is clinically appropriate for the question, how its parameters map to the estimand, and how much influence it can have on the proposed analysis.
The purpose of justification is not to claim that a prior is uniquely correct. It is to make the choice traceable and assessable. A complete account states the prior and its role in the analysis, connects its scale and location to clinical quantities, explains the supporting evidence or rationale, and shows its implications before and after observing trial data. The regulatory question and decision context determine which parameter is relevant and what prior information would be appropriate to use.
What a Reviewer Needs to See
A useful justification answers a sequence of related questions. First, what parameter does the prior describe? State the estimand and its scale, including units where applicable. For example, a prior on a response probability is different from a prior on a log odds ratio, even if both ultimately inform a clinical response comparison. Next, specify the distribution completely: family, parameters, support, and any transformations. A statement such as “a normal prior was used” is incomplete without its mean, variance or standard deviation, and the parameter to which it applies.
The source should be described precisely enough that the information can be traced. This may include a prespecified statistical model, published information, previous relevant evidence, subject-matter rationale, or a combination. State what information was used and what was excluded. When evidence is used, explain its relationship to the trial’s population, endpoint definition, treatment setting, follow-up, and measurement process. The justification should not imply that evidence is interchangeable merely because it concerns the same disease or intervention.
Clinical rationale translates parameters into quantities that can be judged. For a prior on a response probability, report its mean and spread and, where useful, probabilities of clinically interpretable ranges. For a prior on a treatment contrast, explain what values represent no effect, benefit, or harm on the chosen scale. State whether the prior is directional, concentrated, or diffuse and why that shape is plausible for the intended use. A mathematically convenient parameterization is not itself a clinical rationale.
Finally, explain the prior’s role in the proposed analysis. Is it used for the primary posterior estimate, a decision probability, or another prespecified purpose? Describe how it combines with the likelihood under the stated model. Report a measure of its effective contribution when one is meaningful, and show prior predictive implications on the data scale. These descriptions should make clear that the prior is information about a parameter, not a set of observed trial participants.
Worked Example: Documenting a Beta Prior for a Response Probability
Suppose a hypothetical single-arm trial plans to estimate the probability \(p\) of a prespecified clinical response by a fixed assessment time. The planned likelihood is binomial: conditional on \(p\), each of \(n\) participants has a binary response outcome, and outcomes are modeled as independent with the same response probability. The proposed prior is \(p\sim\mathrm{Beta}(2,8)\), supported on \(0\leq p\leq1\).
The prior mean is \(2/(2+8)=0.20\), and its variance is \(2(8)/[(10)^2(11)]=0.01455\), giving a prior standard deviation of approximately \(0.1206\). These quantities describe the distribution of the unknown response probability, not a prediction that exactly 20% of any particular trial will respond. A justification would explain why a 20% prior mean and this degree of dispersion are defensible for this endpoint, population, and treatment context, and would identify the evidence or clinical rationale supporting those choices.
For a concrete account of influence, suppose the hypothetical trial observes 18 responses among 60 participants. The binomial likelihood is proportional to \(p^{18}(1-p)^{42}\). Multiplying by the prior density, proportional to \(p^{2-1}(1-p)^{8-1}\), gives a posterior proportional to \(p^{19}(1-p)^{49}\). Therefore \(p\mid\text{data}\sim\mathrm{Beta}(20,50)\), with posterior mean \(20/70=0.2857\).
The posterior mean can also be written as \((2+18)/(10+60)\): the prior contributes 2 and 8 to the two Beta parameters, while the trial contributes 18 and 42. This algebraic representation is sometimes described as a prior effective sample size of 10 for this conjugate model. It is a convenient measure of the prior’s concentration, not evidence that 10 actual participants were observed. The observed trial response proportion is \(18/60=0.30\); the posterior mean of 0.2857 is lower because the prior mean is 0.20.
Parameterization Must Connect to the Clinical Scale
The same prior information can look different under different parameterizations. Reviewers need the parameter definition and a translation back to the clinical scale. This is especially important for transformed parameters: the mean of a prior on a log scale does not generally transform into the mean on the original scale. State the transformation explicitly and avoid treating a prior’s numerical center as automatically equivalent to a clinically meaningful center.
Worked Example: Translating a Normal Prior on a Log Odds Ratio
Consider a hypothetical randomized trial whose treatment effect is represented by the log odds ratio \(\theta=\log(\mathrm{OR})\), where \(\mathrm{OR}\) is the treatment-to-control odds ratio for response. Suppose the proposed prior is \(\theta\sim N(0,0.25)\); the second argument is the variance, so the prior standard deviation is \(0.5\). The prior is centered on \(\theta=0\), corresponding to \(\mathrm{OR}=\exp(0)=1\).
A reviewer-facing interpretation should include the scale of plausible effects. The central 95% interval under this normal prior is approximately \(0\pm1.96(0.5)=[-0.98,0.98]\) on the log odds ratio scale. Exponentiating the endpoints gives an interval of approximately \([\exp(-0.98),\exp(0.98)]=[0.375,2.66]\) for the odds ratio. Thus, the prior assigns substantial probability to effects in either direction; it is not an assumption that the treatments are equivalent.
Because the normal distribution is symmetric around zero, \(P(\theta>0)=0.5\), equivalently \(P(\mathrm{OR}>1)=0.5\). This describes prior probability under the specified model, not a probability that the null hypothesis is true. Whether the spread is clinically reasonable depends on the intended setting and the evidence supporting it. The submission should explain why odds ratios between about 0.375 and 2.66 are plausible over the central 95% of this prior, and whether this scale is the one relevant to the clinical decision.
Prior Predictive Implications and a Normal-Model Derivation
A prior predictive distribution describes data values implied by the prior together with the likelihood, before observing the trial outcomes. It provides a practical check on whether the specified prior and sampling model produce clinically credible data patterns. This is different from checking only whether the prior mean seems reasonable: a distribution can have a plausible mean but implausible spread or tail behavior.
For the Beta-binomial example, if a future group of 20 participants has response count \(X\), integrating over \(p\sim\mathrm{Beta}(2,8)\) gives a beta-binomial prior predictive distribution. Its mean is \(20(2/10)=4\) responses. Its variance is \(20(2)(8)(10+20)/[(10)^2(11)]=8.727\), with standard deviation about 2.954 responses. The probability of no responses is \[ P(X=0)=\frac{B(2,8+20)}{B(2,8)} =\frac{8\cdot9}{28\cdot29}=0.08867, \] where \(B(\cdot,\cdot)\) is the beta function. A justification should ask whether a mean of 4, this dispersion, and an approximately 8.9% chance of zero responses are compatible with the clinical setting and the planned observation process. If not, the prior, likelihood, or both require reconsideration before the analysis is used.
The influence of a prior can also be displayed directly in a conjugate normal model. Suppose independent observations satisfy \(Y_i\sim N(\theta,\sigma^2)\) with known \(\sigma^2\), and the prior is \(\theta\sim N(m_0,\tau^2)\). The likelihood for the sample mean is proportional to \(\exp[-n(\bar{x}-\theta)^2/(2\sigma^2)]\), and the prior density is proportional to \(\exp[-(\theta-m_0)^2/(2\tau^2)]\). Multiplying these kernels and collecting the quadratic terms in \(\theta\) gives posterior precision \(n/\sigma^2+1/\tau^2\). Completing the square yields the posterior variance and mean below.
The posterior mean is a precision-weighted average of the prior mean and the observed sample mean. The weight on each depends on its precision, so an explanation of a normal prior’s influence should include its variance relative to the sampling variance and sample size.
Worked Example: Quantifying Prior Influence on a Continuous Endpoint
Suppose a hypothetical trial measures time to symptom resolution in days, with lower values being better. For illustration, assume independent normally distributed outcomes with known standard deviation \(\sigma=5\) days and a common mean \(\theta\). Let the prior be \(\theta\sim N(10,4)\), meaning a prior mean of 10 days and prior standard deviation \(\tau=2\) days. Suppose \(n=25\) participants have a sample mean \(\bar{x}=8.5\) days.
The variance of the sample mean is \(\sigma^2/n=25/25=1\) day-squared. Prior precision is \(1/4=0.25\), and data precision is \(1/1=1\). The posterior mean is \[ \frac{10/4+8.5/1}{1/4+1/1} =\frac{2.5+8.5}{1.25} =8.8\text{ days}, \] and the posterior variance is \(1/1.25=0.8\) day-squared, so the posterior standard deviation is \(\sqrt{0.8}=0.8944\) days. The prior supplies \(0.25/1.25=20\%\) of the total precision and the data supply 80%.
The posterior mean, 8.8 days, lies between the prior mean of 10 days and the trial mean of 8.5 days. Under the specified model, an approximate 95% posterior interval is \(8.8\pm1.96(0.8944)=[7.047,10.553]\) days. This is a posterior probability interval for \(\theta\), conditional on the model and prior; it is not a frequentist confidence interval. A regulatory justification would state that the known-variance normal model is an illustrative assumption here, and would defend the actual model and prior using the endpoint’s units, population, and analysis plan.
Documentation, Diagnostics, and Alternatives
A clear prior justification can be organized as a short audit trail. Specify the parameter, estimand, scale, prior family, hyperparameters, support, and any transformation. Identify the source and describe how it informs the parameter in the proposed trial. State the clinical reasoning for the center and spread, including what prior predictive data patterns follow. Then show how the prior enters the planned analysis and what contribution it can make to estimates or decisions. Include enough information for the analysis to be reproduced, such as the exact parameter values and any computational implementation that affects the distribution.
Diagnostics are not only computational. Check that the prior assigns zero probability to impossible parameter values, uses the intended treatment and endpoint direction, and has a scale consistent with the units. Check prior predictive implications, including ranges and tails that matter clinically. For a conjugate prior, report interpretable summaries and the update; for a more complex model, use appropriate numerical summaries or simulation to show the induced distributions. A prior that is proper and computationally convenient can still be clinically implausible.
The main alternatives involve different sources or degrees of prior information. A data-informed prior uses relevant existing evidence and must explain its applicability and how the evidence was converted to the chosen parameterization. A weakly informative prior constrains implausible extremes while allowing a broad range of effects; it still requires a rationale for its scale. A prior based on clinical judgment may be defensible when its assumptions and basis are transparent, but a label such as “expert-based” is not a substitute for specifying the distribution and its implications. A frequentist analysis does not assign a prior distribution, but it answers inferential questions under a different framework; it is not a way to justify or validate a Bayesian prior.
Common warning signs include a prior stated without a parameter definition, a source cited without explaining relevance, a distribution on a transformed scale with no clinical-scale interpretation, and an effective sample size presented as literal patients. Another failure is to show only the prior mean while omitting its dispersion and tail behavior. If the prior is influential, that is not automatically a flaw; it makes a clear account of its rationale and role especially important. If it is not influential in the realized analysis, that does not remove the need to justify the prespecified model.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A proposed prior is \(p\sim\mathrm{Beta}(3,7)\) for a response probability. What are its mean and prior effective sample size under the Beta-binomial model?
- For a prior \(\theta\sim N(0,0.09)\) on a log odds ratio, what is the prior standard deviation, and what odds ratio corresponds to the prior mean?
- Why is a prior effective sample size not necessarily the number of previously observed participants?
- Name two prior predictive quantities that could help assess a prior for a future binomial response count.
Solutions. (1) The mean is \(3/(3+7)=0.30\), and the effective sample size is \(3+7=10\). (2) The standard deviation is \(\sqrt{0.09}=0.30\); the mean log odds ratio is zero, corresponding to an odds ratio of \(\exp(0)=1\). (3) It is a parameterization-based measure of prior concentration or information in a model, not a record of actual participants; its interpretation depends on the likelihood and model. (4) Examples include the predictive mean and variance of the count, the probability of zero responses, or the probability of exceeding a clinically relevant count. Each should be interpreted against the trial setting and observation process.