Incorporating Treatment-Related Harm Into Utility

Why Harm Belongs in the Allocation Objective

An allocation target that counts only favorable outcomes can recommend assigning more patients to a treatment even when that treatment also causes important adverse effects. The ethical objective should instead reflect the expected experience of a patient assigned to each arm. This requires putting outcome benefit and treatment-related harm on a common utility scale, then optimizing the resulting net utility subject to the trial’s information requirement.

The treatment allocation considered here is a fixed target for a two-arm trial: the target share assigned to treatment, rather than a rule for randomizing the next patient. The target is conditional on the assumptions and quantities specified by the design team. It is not a claim that the treatment is beneficial, and it does not replace clinical judgment about which outcomes and harms deserve inclusion.

A harm penalty can be an expected severity-weighted loss. For example, if a particular adverse event occurs with probability \(r\) and has disutility \(L\) utility points when it occurs, its expected disutility is \(rL\). With several harms, the expected penalties can be added when the utility scale and model make that addition appropriate. If benefit and harm are already combined in a validated patient utility measure, do not subtract the harm again: that would count it twice.

A Precision-Constrained Utility Target

Let \(N\) be the fixed total sample size, \(q=n_1/N\) the share assigned to treatment, and \(1-q=n_0/N\) the control share. Assume \(0<q<1\), independent patient outcomes across arms, finite arm-specific outcome variances \(\sigma_1^2\) and \(\sigma_0^2\), and complete outcome measurement for the planning calculation. The arm utilities \(u_1,u_0\) are treated as specified planning values. The variance quantities describe the patient-level outcome used to estimate the arm contrast; they are not the utility means themselves unless the design defines them that way.

The interval derivation below first treats \(q\) as continuous. For a fixed integer sample size \(N\), allowable shares are the grid values \(\{k/N:k=1,\ldots,N-1\}\); intersect this grid with the continuous feasible interval, then choose the largest or smallest feasible grid share according to the utility ordering and verify its standard error. If the intersection is empty, no integer allocation meets the precision constraint.

Under these assumptions, the variance of the difference in arm means is \(\sigma_1^2/n_1+\sigma_0^2/n_0\). If the largest acceptable standard error is \(\delta\), the allocation must satisfy that variance being no greater than \(\delta^2\). Equivalently, define the allocation-scaled standard error \(G(q)=\sqrt{\sigma_1^2/q+\sigma_0^2/(1-q)}\). Since the actual standard error is \(G(q)/\sqrt{N}\), the precision constraint is \(G(q)\leq\sqrt{N}\delta\).

The expected net utility per randomized patient is \(U(q)=(1-q)u_0+qu_1\), a linear function of the treatment share. In the continuous relaxation, maximizing a linear function over the feasible interval selects an endpoint. For a fixed integer sample size, choose the largest feasible grid share if treatment has greater net utility, or the smallest if control has greater net utility. If utilities are equal, every feasible allocation has the same expected utility.

Deriving the Feasible Allocation Interval

Write \(a=\sigma_1^2\) and \(b=\sigma_0^2\). Multiplying the precision inequality by \(q(1-q)>0\) preserves its direction and gives a quadratic inequality. Its roots, when real and within \((0,1)\), mark the boundaries of the feasible interval.

For \(K>0\), the quadratic opens upward. If its discriminant is positive and both roots lie between zero and one, the inequality holds between the roots, so the feasible set is \([q_-,q_+]\). A real-valued root formula alone does not guarantee a usable design: check that the roots are in \((0,1)\). Also, \(\inf_{0<q<1}G(q)^2=(\sqrt{a}+\sqrt{b})^2\). If \(a,b>0\), this minimum is attained at \(q=\sqrt{a}/(\sqrt{a}+\sqrt{b})\); if exactly one is zero, the infimum is not attained; if both are zero, \(G(q)^2=0\) for every \(0<q<1\). Thus a demand \(K<(\sqrt{a}+\sqrt{b})^2\) is infeasible; even the best allocation cannot meet it.

Worked Allocations With Treatment-Related Harm

Diagnostics, Alternatives, and Limitations

Before using a target, check the inputs and the direction of every calculation. Confirm that benefit and harm are measured on the same utility scale, that harm penalties are expected values for the relevant treatment exposure, and that an adverse event has not already been incorporated in the outcome utility. Check that the variances correspond to the outcome whose contrast is being estimated, that \(K=N\delta^2\) uses the same standard-error units, and that both candidate roots lie in \((0,1)\). Substituting each selected boundary back into \(G(q)\) should reproduce the constraint to rounding accuracy.

A common mistake is to treat the treatment’s favorable outcome benefit as its net utility and then subtract harm only in the narrative. The correction is to subtract expected harm in the utility calculation before comparing arms: for benefit \(3.0\), event probability \(0.08\), and event disutility \(10\), net utility is \(3.0-0.08(10)=2.2\), not 3.0. A different error is reporting the allocation-scaled quantity \(G(q)\) as the actual standard error. Since \(SE=G(q)/\sqrt{N}\), retain the factor \(\sqrt{N}\) when translating to the trial’s standard error.

This target differs from pure welfare maximization, which would assign everyone to the arm with greater utility and could leave no control observations. It also differs from Neyman allocation, which chooses shares to minimize the variance of a specified contrast without maximizing patient utility. The method here maximizes a linear expected-utility objective while enforcing a precision bound. It does not automatically resolve ethical questions about a minimum control share, nor does it determine whether the chosen utility scale is clinically or socially appropriate.

The calculations rely on fixed planning values and a fixed total sample size. If harm probabilities, treatment effects, or variances are badly misspecified, the target can be poorly chosen. The calculation also does not account for delayed, cumulative, or interacting harms unless these are represented in the utility model. An infeasible precision bound should be reported as infeasible rather than handled by silently choosing a boundary. These limitations make sensitivity analysis of plausible harm values important when the allocation depends on a narrow utility difference.