Adaptive designs with continuous treatments

This post is part of our Q&A series.

A question from graduate students in our Spring 2021 offering of the new course “Targeted Learning in Practice” at UC Berkeley:


Hi Mark,

You’ve got me thinking about selecting optimal experiments in the context of shift interventions. For the example we talked about in class, in order to avoid positivity violations, we define shift interventions such that an individual’s value of the intervention node $A$ is shifted by a specified amount $\delta$ unless there is no support for such a shift based on the covariates $W$, in which case $A$ is shifted to the maximum value available for that $W$. This makes me think that there may be situations where there is a health systems issue that causes a positivity violation that prevents the full shift $\delta$ from being evaluated for some people. For example, maybe your intervention is based on increasing the number of visits to a health center for a given problem, but people who either live a certain distance away from the health center or whose insurance has a high copay for the visits are very unlikely to participate in the number of proposed sessions for logistical and financial reasons in your dataset. You could potentially design multiple experiments where you intervene not only on $A$ but also on the $W$ covariate that is preventing certain people from having the full shift $\delta$. This could be a practically interesting question as in what would be the mean outcome under the shifted $A$ if you also intervened to decrease peoples’ distance to a health center (i.e., by offering the intervention at more health centers). I would think then that you could shift $A$ further for more people – either by $\delta$ or the maximum of $\delta$ for people living within a short distance from the health center based on the rest of the $W$ covariates. As long as the probability of this joint intervention given the rest of the covariates is not very small, then I would think the effect of such an intervention would be identifiable. Could you then define a data-adaptive target parameter as the mean of the outcome under the intervention to both shift $A$ and intervene on whichever of the $W$ covariates leads to the largest mean shift in $A$? This selection of the $W$ intervention would not need to depend on the outcome. I wonder if this could be a means of both designing and evaluating the potential impact of interventions aimed at improving access to care.




Hi L.E.D.,

Thank you for the interesting question.

For concreteness, we observe $O = (W, A, Y)$ where $A$ is a continuous or discrete-ordered variable (number of visits to healthcare per year). In addition, there are some key variables in $W$, say $V$, that are known to be important for achieving a high level of $A$. You then talk about an intervention on $V$ through another $A_1$ variable making it easier to achieve $A$ for the person. It is like providing another incentive $A_1$ such as money, travel or bringing the service closer to them. Normally, $W$ has an arrow going into $A_1$, but this $A_2$ is pre-$V$. So, the ordering becomes $(W, A_1, V, A_2, Y)$. So, we now have a longitudinal two-timepoint intervention data structure, and we are interested in two-timepoint interventions $(g^{\star}_1, g^{\star}_2)$ that optimize $\mathbb{E} Y_{g^{\star}}$. You indicate a situation in which the ability to increase $A_2 = d(V)$ or $g^{\star}_{2 \mid V}$, to high levels is very much affected by how $A_1$ is set and thereby $V$ is realized. Just as in cancer trials where we h ave two-stage treatment regimen, first-line and second-line treatment.

If we restrict to shift interventions on $A_2$, then, depending on $A_1$, we can push the shift further. So, we could define $g^{\star}_{2,\delta}$ with $\delta$ chosen maximally in response to the value of $V$. Even though we fix $g^{\star}_2$ we still have flexibility in what variable among the $V$ components we intervene on. So, we really have a set $(A_{1k}: k)$ of treatment nodes for $A_1$, where $k$ identifies which $V_k$ we target with this intervention.

So, we have $(W, (A_1(k): k), V, A_2, Y)$. We now want to consider, say, rules $d_{1k}(W)$ that intervene on $A_1(k)$ with a rule targeting $V_k$. We could consider regimens that only intervene on a single $A_1(k)$, and we are then interested in not only optimizing $d_{1k}(W)$ for a given $k$ but also in selecting among $k$, or, maybe even among joint interventions that target subsets of $A_1(k)$. Given an observational data set, we could estimate the mean outcome among a user-supplied class of such interventions or define optimal rules $d_{1k,P_0}(W)$ for each $k$ and thus also an single optimal rule $d_0$ for $A_1$. We then want to learn that rule and, given that estimated rule, we care about the mean outcome with respect to using that rule on $A_1$ and the corresponding shift $g^{\star}_{2}$ on $A_2$. So, this falls in the category of learning joint interventions from data and then performing a TMLE for the corresponding target parameter, thus using CV-TMLE.

It is very much like making the choice of interventin on $A_2$ a deterministic function of the $A_1$ intervention and then defining the optimal rule for $A_1$. It avoids the notion of sequentially defined optimal rules, as in our work with Alex Luedtke by making the second treatment a deterministic function of the first – quite interesting and sensible. Overall, this certainly falls in the category of data-adaptive target parameters defined by searching among subsets of variables of $V$ and corresponding rules based on data, using CV-TMLE as a general method for inference.

If indeed one can already do some learning without outcome data then that learning comes “for free,” so that would not be bad either. However, I think you have in mind that we can investigate how $A_1$ affects $V(A_1)$ but that is in essence looking at the outcome, so it is informative since $V$ is just one of the time-dependent markers on the pathway to the final outcome. Maybe you are getting at the point that one could learn an optimal rule or so for targeting $V$ only using data $(W, A_1, V)$, and, after having run that and set the intervention on $A_1$, we consider the next problem of learning the rule for $A_2$ based on $(W, V), A_2, Y$. If indeed it is relatively cheap to measure $V$ and impact of $A_1$ on $V$, then that would be an effective experiment.

You also appear to talk about running actual trials. This falls then in the category of adaptive sequential designs (i.e., reinforcement learning), such as targeted adaptive design learning of the optimal rule. We have work on that, including recent work by Aurelien Bibaut. There is still a lot of progress to be made there, see also my lectures notes from the class on adaptive design (Public Health 243D), covering topics such as being able to adapt to surrogate outcomes so that one does not need to wait until outcomes are measured before being able to adapt the randomization probabilities for next enrolled subject. A lot of exciting and important research to do there.

Great questions.

Best Wishes,


P.S., remember to write in to our blog at vanderlaan (DOT) blog [AT] berkeley (DOT) edu. Interesting questions will be answered on our blog!

comments powered by Disqus