Fractional Likelihoods for Historical Data

Why Fractional Likelihoods Are Used

A historical control group can contribute information to a new clinical trial, but its patients may not represent the population, care setting, or calendar period of the current trial exactly. Using the historical likelihood at full strength treats its evidence as fully relevant to the target parameter. Omitting it discards all of its evidence. A fractional likelihood provides an intermediate option: raise the historical likelihood to a power between zero and one so its contribution to relative support for parameter values is reduced.

The power changes the historical likelihood contribution; it does not change the observed historical outcomes. Nor does it establish that the historical and current populations are comparable. That scientific judgment remains essential.

Definition: Let \(D_0\) denote historical data and let \(L_0(\theta;D_0)\) be their likelihood for parameter \(\theta\). For a specified power \(a\) with \(0\leq a\leq 1\), the fractional historical likelihood contribution is \(L_0(\theta;D_0)^a\). At \(a=1\), it is the full historical likelihood. At \(a=0\), define the contribution to be 1 for every \(\theta\), so the historical data provide no likelihood contribution. This convention avoids relying on \(0^0\).

What the Power Does to Relative Support

Likelihoods compare how well different parameter values account for observed data. If two candidate values \(\theta_1\) and \(\theta_2\) have a finite, strictly positive historical likelihood ratio \(R=L_0(\theta_1;D_0)/L_0(\theta_2;D_0)\), their ratio under the fractional contribution is \(R^a\). This follows by dividing the powered likelihoods:

\frac{L_0(\theta_1;D_0)^a}{L_0(\theta_2;D_0)^a} = \left\{\frac{L_0(\theta_1;D_0)}{L_0(\theta_2;D_0)}\right\}^{a} =R^a.

For \(0<a<1\), if \(R>1\), then \(1<R^a<R\). If \(0<R<1\), then \(R<R^a<1\). Thus the ratio moves toward one, weakening the historical data’s distinction between these two candidate values. For \(a=1\), the ratio is unchanged; for \(a=0\), it is one by the no-borrowing convention.

The strict positivity condition matters. If \(R=0\), then \(R^a=0\) for every \(a>0\); discounting does not move that ratio toward one. At \(a=0\), the historical contribution is defined as 1, including at parameter values where the unpowered likelihood is zero. The power therefore removes the historical likelihood contribution altogether rather than evaluating \(0^0\).

A Binary-Control Likelihood

Suppose the historical endpoint is binary: each control patient either experiences the event or does not. Let \(Y_i=1\) indicate an event and \(Y_i=0\) indicate no event. Assume, for this illustrative model, that the \(n_0\) historical outcomes are conditionally independent Bernoulli observations with a common event probability \(p\). If \(x_0\) patients experience the event, the likelihood is \(L_0(p;D_0)=\binom{n_0}{x_0}p^{x_0}(1-p)^{n_0-x_0}\) for \(0<p<1\).

Raising this likelihood to \(a\) gives \(\binom{n_0}{x_0}^{a}p^{a x_0}(1-p)^{a(n_0-x_0)}\). The exponents show directly how the historical evidence is downweighted. This is an algebraic description of a likelihood contribution, not a claim that the trial observed \(a n_0\) patients. In particular, the powered expression is not generally the ordinary sampling likelihood of a binomial experiment with a fractional number of patients.

The model also makes a substantive assumption: the historical observations are being represented as draws from a common event probability \(p\). For borrowing to inform a current trial’s control parameter, the historical endpoint and its definition should be relevant to the current estimand, and differences in population, setting, follow-up, or standard of care should be considered. A fractional power can reduce the influence of a likelihood; it cannot repair a mismatch in outcome definition or make incompatible populations exchangeable.

Worked Example: Discounting Historical Support in a Binary-Control Study

Suppose a hypothetical historical control group has 6 events among 10 patients. Compare candidate event probabilities \(p=0.4\) and \(p=0.2\). The likelihood ratio favoring \(p=0.4\) over \(p=0.2\) is

R = \frac{\binom{10}{6}(0.4)^6(0.6)^4} {\binom{10}{6}(0.2)^6(0.8)^4} = \left(\frac{0.4}{0.2}\right)^6 \left(\frac{0.6}{0.8}\right)^4 = 2^6(0.75)^4 = 64(0.31640625) = 20.25.

Under the full historical likelihood, these data give 20.25 times as much likelihood support to \(p=0.4\) as to \(p=0.2\). If the historical likelihood is raised to \(a=0.25\), the ratio becomes

R^{0.25}=20.25^{0.25}=\sqrt{\sqrt{20.25}}=\sqrt{4.5}\approx 2.121.

The fractional contribution still favors \(p=0.4\), but the historical likelihood ratio is much closer to one. This is a statement about relative likelihood support from the historical data alone, not a posterior probability that either candidate value is correct.

Worked Comparisons and an Important Boundary Case

Worked Example: A Smaller Power, a Smaller Likelihood Ratio

For a second hypothetical binary-control dataset, suppose 3 of 5 historical patients experience the event. Compare candidate probabilities \(p=0.6\) and \(p=0.3\). The binomial coefficient cancels in the ratio:

R = \frac{\binom{5}{3}(0.6)^3(0.4)^2} {\binom{5}{3}(0.3)^3(0.7)^2} = \left(\frac{0.6}{0.3}\right)^3 \left(\frac{0.4}{0.7}\right)^2 = 8\left(\frac{4}{7}\right)^2 = \frac{128}{49} \approx 2.612.

With power \(a=0.5\), the ratio is

R^{0.5} = \sqrt{\frac{128}{49}} \approx 1.616.

The full likelihood favors \(p=0.6\) over \(p=0.3\); the half-power contribution still favors \(p=0.6\), but less strongly. The calculation discounts the likelihood contrast without changing which candidate has the larger likelihood.

Worked Example: A Zero Likelihood Is Not Rescued by a Positive Power

Suppose, hypothetically, that all 5 historical patients experience the event. Compare \(p=0\) with \(p=0.5\). Under the binary model, the likelihood at \(p=0\) is zero, while at \(p=0.5\) it is

L_0(0.5;D_0) = \binom{5}{5}(0.5)^5(0.5)^0 = 0.03125.

The ratio \(L_0(0;D_0)/L_0(0.5;D_0)\) is zero. For any positive power, such as \(a=0.4\), its powered ratio remains

0^{0.4}=0.

So the historical likelihood still assigns no support to \(p=0\) under this model; taking a positive fractional power does not make a zero likelihood positive. If \(a=0\), the historical contribution is instead defined as 1 at both candidate values, giving a ratio of 1. That is the no-borrowing case, not an evaluation of \(0^0\).

Assumptions, Diagnostics, and Alternatives

The algebra above is simple, but its use depends on modeling and design choices. In a binary example, the Bernoulli likelihood assumes a clearly defined event, independent observations conditional on \(p\), and a common event probability within the historical group. If there is clustering, repeated observations, or important variation between patients that the model does not represent, the simple binomial likelihood may not be adequate. The power does not correct those problems.

A useful diagnostic review asks whether the historical data and current trial target the same clinical event and follow-up period; whether patient eligibility and treatment practices are sufficiently comparable; and whether any differences could make the historical event rate systematically misleading. Also check that the observed counts, likelihood, and parameter values are consistent. For example, a likelihood of zero at a candidate value may arise from a boundary value of \(p\); it must not be treated as an ordinary positive ratio when applying the toward-one argument.

The main alternatives differ in how much historical likelihood they retain. Full borrowing uses \(a=1\); no borrowing uses \(a=0\); fractional borrowing uses an intermediate power. These choices can be summarized as follows.

Choice	Historical contribution	Interpretation
Full likelihood	\(L_0(\theta;D_0)\)	Historical likelihood contrasts are retained in full.
Fractional likelihood	\(L_0(\theta;D_0)^a\), with \(0<a<1\)	Historical likelihood contrasts are reduced according to the chosen power.
No historical likelihood	1	The historical data do not contribute likelihood information.

An intermediate power is not automatically a compromise that is appropriate for every trial. Its justification should reflect the relevance of the historical evidence and the consequences of borrowing if the historical and current settings differ. A single power also discounts the historical likelihood as a whole; it does not selectively fix particular sources of bias.

Common Mistakes

Calling the power a fraction of the patients. The observations remain the observed patients; \(a\) weights their likelihood contribution. It does not create or remove literal patients.
Saying every likelihood ratio moves toward one. That statement requires a finite, strictly positive ratio and \(0<a<1\). A zero ratio stays zero for every positive power.
Interpreting a likelihood ratio as a probability. A ratio compares likelihoods at specified parameter values; it is not the probability that one value is true.
Assuming discounting guarantees comparability. A smaller contribution does not make a mismatched endpoint or population suitable for borrowing.
Using \(0^0\) for no borrowing. At \(a=0\), define the historical contribution as one for every parameter value.

Key takeaway: A fractional likelihood raises the historical likelihood to a chosen power. For strictly positive likelihood ratios, a power between zero and one moves the ratio toward one; zero likelihoods require separate treatment, and discounting does not substitute for clinical and modeling judgment about relevance.

Check Your Understanding

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

If a positive historical likelihood ratio is 9 and \(a=0.5\), what is the powered ratio?
For \(0<R<1\) and \(0<a<1\), how does \(R^a\) compare with \(R\) and 1?
Why is it incorrect to describe a power of 0.4 as observing 40% of the historical patients?
If a historical likelihood ratio is zero, what is its value after raising it to a positive power? What is the contribution at \(a=0\)?
Name one assumption or comparability concern that a fractional power cannot fix.

Solutions. (1) \(9^{0.5}=\sqrt{9}=3\). (2) \(R<R^a<1\), so the ratio is closer to one. (3) The power weights the likelihood contribution; the observed sample remains unchanged and is not a fractional sample. (4) It remains zero for every positive power; at \(a=0\), the contribution is defined as one for every parameter value. (5) Examples include dependence among observations not represented in the model, a different endpoint definition, or a substantial difference between historical and current patient populations.