Targeting Incremental Treatment Interventions

Why Tilt Treatment Odds?

A deterministic regime asks what would happen if treatment were set to a specified value, while a stochastic intervention specifies a probability distribution for treatment. An incremental intervention offers a particular stochastic alternative: it changes the odds of treatment relative to the observed treatment mechanism without requiring treatment to be assigned with probability one. The incremental factor describes the size and direction of the shift.

This construction is especially useful when deterministic regimes would demand treatment levels rarely observed for some histories. The intervention preserves the support of the observed treatment mechanism, while shifting its probabilities. Its targeting covariate is a cumulative product of intervention-to-observed treatment density ratios. Because the intervention itself is defined using the propensity score, however, that covariate alone does not describe every contribution to the efficient influence function.

Definition and Treatment Density

Let \(H_t\) denote the measured history immediately before the binary treatment \(A_t\in\{0,1\}\) at time \(t\), and let \(g_t(a\mid h)=P(A_t=a\mid H_t=h)\) be the observed treatment mechanism. For a fixed incremental factor \(\delta>0\), define the intervention distribution by multiplying the odds of treatment by \(\delta\).

The intervention odds equal \(\delta\) times the observed odds whenever \(0<g_t(h)<1\): \[ \frac{q_{\delta,t}(1\mid h)}{q_{\delta,t}(0\mid h)} =\delta\frac{g_t(h)}{1-g_t(h)}. \] Thus \(\delta=1\) leaves the mechanism unchanged, \(\delta>1\) increases treatment odds, and \(0<\delta<1\) decreases them. For finite positive \(\delta\), the intervention probabilities sum to one because their numerators sum to \(d_{\delta,t}(h)\).

Longitudinal Density Ratios and Targeting Covariates

For decisions at times \(t=0,\ldots,K\), the intervention distribution for the treatment sequence, conditional on the evolving histories, is the product of its conditional probabilities. Relative to the observed treatment sequence, its density ratio is therefore the product of the corresponding one-time ratios.

This covariate is the analogue of the clever covariate for a stochastic intervention. In a sequential targeting step, it weights the residual at that time according to how much more or less likely the observed treatment path is under the intervention than under the observed mechanism. It is not the probability of the observed treatment path under the intervention: it is the ratio of that probability to its observed-mechanism counterpart.

Proof. By the chain rule for conditional probabilities, the observed conditional probability of a treatment sequence \(a_0,\ldots,a_K\), given the histories generated along that sequence, is \(\prod_{t=0}^{K}g_t(a_t\mid h_t)\). Under the intervention it is \(\prod_{t=0}^{K}q_{\delta,t}(a_t\mid h_t)\). Where the observed probability is positive, dividing these expressions gives \[ \frac{\prod_{t=0}^{K}q_{\delta,t}(a_t\mid h_t)} {\prod_{t=0}^{K}g_t(a_t\mid h_t)} =\prod_{t=0}^{K}\frac{q_{\delta,t}(a_t\mid h_t)}{g_t(a_t\mid h_t)}. \] The partial product through \(t\) is the corresponding density ratio for the sequence up to time \(t\). \(\square\)

Why the Propensity-Score Dependence Matters

For a fixed stochastic intervention, the cumulative ratio weights the outcome-regression residuals in the usual longitudinal targeting construction. An incremental intervention differs in one important respect: \(q_{\delta,t}\) is defined using \(g_t\). Changing the treatment mechanism therefore changes both the observed-data distribution and the intervention being evaluated. Consequently, the cumulative ratio is the outcome-residual targeting covariate, but it is not by itself the entire influence-function contribution for this parameter.

Derivation. Fix a history \(h\) and write \(g=g(h)\). The intervention mean conditional on \(h\) is \(V(h)=q_\delta(1\mid h)Q(1,h)+q_\delta(0\mid h)Q(0,h)\). Differentiating the intervention probabilities with respect to \(g\) gives \[ \frac{\partial q_\delta(1\mid h)}{\partial g}=\frac{\delta}{d(h)^2}, \qquad \frac{\partial q_\delta(0\mid h)}{\partial g}=-\frac{\delta}{d(h)^2}. \] Thus the derivative of \(V(h)\) from changing \(g\) is \(\{\delta/d(h)^2\}\{Q(1,h)-Q(0,h)\}\dot g(h)\). A score for the Bernoulli treatment mechanism along this change is \(\{A-g(h)\}\dot g(h)/\{g(h)(1-g(h))\}\). The conditional covariance of \(\{A-g(h)\}\) with that score is \(\dot g(h)\), because \(E[(A-g)^2\mid H=h]=g(h)\{1-g(h)\}\). Multiplying \(A-g(h)\) by \(\{\delta/d(h)^2\}\{Q(1,h)-Q(0,h)\}\) therefore reproduces the required derivative. The outcome-residual term follows from the conditional mean derivative: its multiplier is \(q_\delta(A\mid H)/g(A\mid H)\). The centered baseline term accounts for variation in the distribution of \(H\). \(\square\)

In a longitudinal problem, the same distinction applies at each decision: the outcome-residual covariate is the cumulative ratio through that decision, while dependence of the intervention on each \(g_t\) also contributes to the derivative of the target. The relevant contrast at time \(t\) is between the downstream intervention values after setting \(A_t=1\) and \(A_t=0\); histories at time \(t\) are distributed under the earlier intervention. A targeting procedure must account for the intervention’s dependence on the treatment mechanisms rather than treating the estimated \(q_{\delta,t}\) as fixed without justification.

Assumptions, Diagnostics, and Limitations

For a causal interpretation, the usual longitudinal requirements remain relevant: consistency of observed outcomes with the treatment actually received, sequential exchangeability given the measured histories, and adequate measurement of confounders at each decision. The treatment and outcome histories must be ordered so that \(H_t\) is available before \(A_t\). For standard regularity arguments, one commonly assumes a fixed finite number of decisions, suitable integrability of the outcome, and treatment probabilities bounded away from zero and one. The incremental intervention is defined for any finite \(\delta>0\).

Incremental interventions preserve the observed treatment support: if \(g_t(1\mid h)=0\), then \(q_{\delta,t}(1\mid h)=0\); if \(g_t(1\mid h)=1\), then \(q_{\delta,t}(0\mid h)=0\). This can avoid the unsupported treatment assignments required by some deterministic regimes. But it does not create information about treatment effects for histories at which one treatment is never observed. Nor does boundedness of the density-ratio factors guarantee precise estimation: long products, noisy nuisance estimates, and rare histories can still cause unstable finite-sample behavior.

Useful diagnostics include inspecting the fitted \(g_t(H_t)\) distributions at every decision, checking that computed \(q_{\delta,t}(1\mid H_t)\) lies in \([0,1]\), and checking the mean and upper tail of each cumulative ratio. Verify that the ratios are computed as \(q/g\), not \(g/q\), and that the product includes only decisions up to the time of the residual being targeted. Also check whether the fitted mechanism and the intervention used in targeting are updated consistently.

A common mistake is to interpret the incremental factor as an addition to the treatment probability, or to use \(\delta g\) as the intervention probability. Neither operation preserves the intended odds tilt or necessarily gives a probability. The correction is to normalize the tilted probabilities by \(1+(\delta-1)g\). Another mistake is to claim that the cumulative ratio alone gives the full efficient influence function. It gives the outcome-residual multiplier; because \(q_\delta\) depends on \(g\), there is also a treatment-mechanism derivative contribution.

Compared with deterministic treatment regimes, incremental interventions make smaller, graded shifts and retain the observed mechanism’s support. Compared with a stochastic intervention specified independently of \(g\), they adapt to the observed propensity score, which is their defining feature and also adds a derivative term to inference and targeting. They do not automatically solve confounding, model misspecification, or sparse-data problems; the factor \(\delta\) defines a different causal target, rather than a correction for bias.