Sequential Historical Cohorts and Power Priors

Why Cohort Order Needs Care

Historical control data often come from several cohorts observed at different calendar times. More recent cohorts may be more relevant to a current trial, but chronology must be represented deliberately: multiplying likelihood contributions in time order does not, by itself, make recent observations count more. A power prior provides one way to assign each cohort a prespecified amount of influence.

The central construction has two parts: identify a likelihood contribution for each distinct cohort, then attach the intended weight to that cohort. The chronological labels make the assignment interpretable. They do not change the fact that likelihood factors multiply commutatively.

Definition: Let $D_{(1)},\ldots,D_{(J)}$ denote historical control data from $J$ cohorts, indexed in chronological order from oldest to most recent. Let $L_j(\theta)=p(D_{(j)}\mid\theta)$ be the likelihood contribution from cohort $j$, and let $a_j\in[0,1]$ be its fixed power-prior weight. Given a baseline prior density $\pi_0(\theta)$, the normalized power prior is $$ \pi(\theta\mid D_{(1)},\ldots,D_{(J)},\mathbf a) = \frac{\pi_0(\theta)\prod_{j=1}^J L_j(\theta)^{a_j}} {C(\mathbf a)}, \qquad C(\mathbf a)=\int \pi_0(u)\prod_{j=1}^J L_j(u)^{a_j}\,du. $$ Here $\mathbf a=(a_1,\ldots,a_J)$ is fixed in this tutorial. A weight of zero removes that cohort’s likelihood contribution; a weight of one uses it without discounting.

The construction assumes the cohort likelihoods are appropriate for the parameterization and data structure. In the common-parameter case, cohorts are distinct sets of individuals whose outcomes are conditionally independent given the shared parameter $\theta$. Their joint historical likelihood then factors as $\prod_j L_j(\theta)$. The powered product discounts each factor separately. If a cohort has different outcome ascertainment, eligibility, or follow-up, these features must be handled in the model or the data definition rather than assumed away by choosing a weight.

How Chronology Enters the Construction

Chronology enters through the cohort labels and the rule assigning weights to them. For example, if a prespecified policy gives more influence to more recent cohorts, one might require $a_1\leq a_2\leq\cdots\leq a_J$. That monotonicity is a modeling choice, not a mathematical requirement of the power prior. When elapsed time matters, the weights should be associated with actual cohort dates or time gaps, rather than treating equally spaced cohort ranks as if they necessarily represented equal calendar intervals.

The likelihood product itself has no order:

L_1(\theta)^{a_1}L_2(\theta)^{a_2} = L_2(\theta)^{a_2}L_1(\theta)^{a_1}.

Thus “sequential” does not mean that multiplying the oldest cohort first and the newest last creates a time trend. It means the data have been organized into chronological, distinct contributions, and each weight is attached to the intended cohort. If the weights are reversed while the data are left in place, the model changes. If the factors are merely written in a different multiplication order while each weight remains attached to its cohort, the model does not change.

A critical data requirement is that $D_{(1)},\ldots,D_{(J)}$ represent non-overlapping individuals or valid conditional increments. Separate cohorts observed in successive periods can often be modeled as distinct contributions. In contrast, two cumulative reports that contain many of the same individuals are not two independent cohorts. Their likelihoods cannot simply be multiplied as though they described separate patients.

Binary Outcomes: The Weighted Beta-Binomial Form

Suppose cohort $j$ contains $n_j$ distinct control patients, of whom $y_j$ experience a specified binary outcome, and assume a shared event probability $p$. Conditional on $p$, take $Y_j\sim\mathrm{Binomial}(n_j,p)$, independently across cohorts. The cohort likelihood is

L_j(p) = \binom{n_j}{y_j}p^{y_j}(1-p)^{n_j-y_j}.

With a Beta baseline prior $p\sim\mathrm{Beta}(\alpha_0,\beta_0)$, the powered historical likelihood contributes $a_jy_j$ to the event exponent and $a_j(n_j-y_j)$ to the non-event exponent. The binomial coefficient raised to $a_j$ does not depend on $p$, so it is absorbed into the normalizing constant.

Result: Under the shared-$p$ binomial model and a $\mathrm{Beta}(\alpha_0,\beta_0)$ baseline prior, the historical power prior is $$ p\mid D_{(1)},\ldots,D_{(J)},\mathbf a \sim \mathrm{Beta}\left( \alpha_0+\sum_{j=1}^J a_jy_j,\; \beta_0+\sum_{j=1}^J a_j(n_j-y_j) \right). $$ If a separate current-trial control group has $y_c$ events among $n_c$ patients and its likelihood is not discounted, add $y_c$ to the first shape parameter and $n_c-y_c$ to the second.

This result follows directly by multiplying the prior kernel by the powered cohort likelihoods. The Beta prior kernel is $p^{\alpha_0-1}(1-p)^{\beta_0-1}$, while cohort $j$’s powered likelihood kernel is $p^{a_jy_j}(1-p)^{a_j(n_j-y_j)}$. Collecting exponents gives the displayed Beta parameters. The resulting weighted event and non-event totals can be fractional; they are contributions to a prior distribution, not literal counts of patients.

Worked Example: Three Ordered Historical Cohorts

Consider a hypothetical binary control outcome measured consistently in three distinct cohorts. The oldest cohort has 4 events among 40 patients, the middle cohort has 9 among 30, and the most recent has 10 among 20. Assign fixed weights $a_1=0.2$, $a_2=0.5$, and $a_3=0.8$, respectively. Use a $\mathrm{Beta}(1,1)$ baseline prior, and suppose the current trial’s control group has 12 events among 30 patients. Assume independent outcomes conditional on a shared event probability $p$, and assume the cohorts do not overlap.

The historical weighted event contribution is $$ 0.2(4)+0.5(9)+0.8(10)=0.8+4.5+8=13.3. $$ The historical weighted non-event contribution is $$ 0.2(40-4)+0.5(30-9)+0.8(20-10) =7.2+10.5+8=25.7. $$

Add the current-trial outcomes without discounting. The posterior is $$ p\mid \text{all data} \sim \mathrm{Beta}(1+13.3+12,\;1+25.7+18) = \mathrm{Beta}(26.3,44.7). $$ Its posterior mean is $26.3/(26.3+44.7)=26.3/71=0.3704$. Under the stated model, this is the posterior mean event probability for the current control population. It is not a claim that 37.04% of observed patients had events; the observed current-trial proportion is $12/30=0.4000$.

Why the Weight Must Stay With Its Cohort

Chronological order matters when weights encode a recency policy. To see this, retain the hypothetical data above but incorrectly assign weights $0.8,0.5,0.2$ to the oldest, middle, and most recent cohorts. The weighted event contribution becomes $$ 0.8(4)+0.5(9)+0.2(10)=3.2+4.5+2=9.7, $$ and the weighted non-event contribution becomes $$ 0.8(36)+0.5(21)+0.2(10)=28.8+10.5+2=41.3. $$

With the same current data and baseline prior, this reversed assignment gives $\mathrm{Beta}(22.7,60.3)$, with mean $22.7/83=0.2735$. It differs from the preceding posterior because the cohorts have different outcome rates and the weights have been attached differently. This calculation illustrates sensitivity to assignment; it does not establish which weighting policy is scientifically appropriate. That policy must be justified before examining the resulting posterior, rather than selected to produce a preferred answer.

Worked Example: Deriving the Posterior Kernel

For a smaller hypothetical illustration, suppose two ordered, non-overlapping cohorts have $(y_1,n_1)=(2,10)$ and $(y_2,n_2)=(6,15)$, with weights $a_1=0.3$ and $a_2=0.7$. Let the baseline prior be $\mathrm{Beta}(2,2)$. Show the historical power-prior distribution.

The weighted event total is $0.3(2)+0.7(6)=0.6+4.2=4.8$. The weighted non-event total is $0.3(8)+0.7(9)=2.4+6.3=8.7$. Therefore the unnormalized prior kernel is $$ p^{2-1}(1-p)^{2-1}\,p^{4.8}(1-p)^{8.7} = p^{5.8}(1-p)^{9.7}. $$

Matching this kernel to $p^{\alpha-1}(1-p)^{\beta-1}$ gives $\alpha=6.8$ and $\beta=10.7$. Thus the normalized historical power prior is $\mathrm{Beta}(6.8,10.7)$, with mean $6.8/(6.8+10.7)=0.3886$. The fractional shape contributions arise from the specified powers; they are not integer-valued patient counts.

Disjoint Cohorts Versus Cumulative Reports

A likelihood factor represents information from the observations it describes. If a report at time 2 includes the time 1 patients again, treating both reports as independent factors counts those patients twice. A valid sequential construction instead partitions the records into disjoint increments, or uses a correctly specified conditional likelihood for the new information given what was already observed.

Worked Example: Converting Cumulative Snapshots to Increments

Suppose a hypothetical report at the first cutoff records 5 events among 20 patients. A later cumulative report records 8 events among 35 patients, including those original 20 patients. Assume the later report adds 15 distinct patients and that 3 of them have events. Let the first increment have weight $0.4$, the second increment weight $0.8$, and let $p$ be a shared event probability. Use a $\mathrm{Beta}(2,3)$ baseline prior.

The distinct increments are $(5,20)$ and $(8-5,35-20)=(3,15)$, not $(5,20)$ and $(8,35)$. The weighted event contribution is $0.4(5)+0.8(3)=2+2.4=4.4$; the weighted non-event contribution is $0.4(15)+0.8(12)=6+9.6=15.6$. The historical power prior is therefore $$ \mathrm{Beta}(2+4.4,\;3+15.6)=\mathrm{Beta}(6.4,18.6). $$ Its mean is $6.4/25=0.2560$. Multiplying likelihoods for both cumulative reports would reuse the first 20 patients and would not correspond to the stated independent-increment model.

Assumptions, Diagnostics, and Alternatives

A cohort-specific power prior is most defensible when the endpoint and its definition are comparable across cohorts, the cohorts are correctly separated, and the shared-parameter likelihood is a reasonable approximation after any necessary modeling of design differences. Check the data lineage: verify which patients appear in each dataset, how each cohort’s time window is defined, and whether reported counts are cumulative or incremental. Check the model inputs as well: denominators must match the outcome counts, weights must be attached to the intended chronological cohorts, and the direction of the event must be consistent.

A poor fit or major differences in observed outcome rates across cohorts can signal that a shared $p$ is inadequate. Such differences may reflect calendar-time changes, changed eligibility or care, different follow-up, or chance variation. Discounting can limit a cohort’s influence, but it does not repair a misspecified outcome definition or make non-comparable groups exchangeable. If cohort-specific probabilities are scientifically needed, the likelihood can be written in terms of $p_j$, but borrowing across cohorts then requires a model linking those parameters; simply raising unrelated likelihoods to powers does not supply that link.

The alternatives answer different questions. A common fixed power applies the same discount to every historical likelihood contribution; cohort-specific fixed powers allow different prespecified influence by cohort. Full pooling corresponds to weight one for all cohorts under the shared-parameter model, while using zero for all historical weights leaves the current analysis without historical likelihood contributions. An analysis that excludes historical data entirely avoids the shared-parameter borrowing assumption but gives up that information. The weights describe the chosen borrowing construction; they do not establish that the historical and current populations are clinically interchangeable.

Wrong reasoning: “The most recent likelihood is multiplied last, so it automatically has the greatest influence.” Multiplication order has no effect on the product.
Correction: Chronological labels identify cohorts; any recency preference must be encoded by assigning weights to those labels or by an explicit time model.
Wrong reasoning: “Two reports are two cohorts, so both likelihoods should be included.” Reports may contain overlapping patients.
Correction: Define disjoint cohorts or derive valid conditional increments before constructing the product.
Wrong reasoning: “A fractional weighted count is a literal fractional number of patients.”
Correction: It is a likelihood exponent and a contribution to the prior shape parameters, not a data count.

Key takeaway: Order the historical cohorts by calendar time, define a valid likelihood contribution for each distinct increment, and attach each fixed power to the cohort it is intended to discount. The ordering informs the weight assignment; it does not make the likelihood product itself sequentially order-dependent.

Check Your Understanding

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

Two independent historical cohorts have likelihoods $L_1(\theta)$ and $L_2(\theta)$, with weights $a_1$ and $a_2$. Does writing the second factor first alter the power prior? Why?
For a binary outcome, a cohort has 7 events among 25 patients and weight $0.4$. What are its weighted event and non-event contributions?
A historical report has 6 events among 24 patients at one cutoff, then 9 among 40 cumulatively. How many events and patients are in the later increment?
Under a shared-$p$ model with a Beta prior, what does setting every historical cohort weight to zero do to the posterior when there is also a current control group?

Solutions. 1. No. The product $L_1(\theta)^{a_1}L_2(\theta)^{a_2}$ is unchanged by reordering its factors, provided each weight remains attached to its cohort. 2. The weighted event contribution is $0.4(7)=2.8$; the weighted non-event contribution is $0.4(25-7)=7.2$. 3. The later increment has $9-6=3$ events among $40-24=16$ patients. 4. The historical likelihood factors become $L_j(\theta)^0=1$, so the posterior is determined by the baseline prior and the current control likelihood.