Search

The study of experimental design offers tremendous benefits for answering causal questions across a wide range of applications, including agricultural experiments, clinical trials, industrial experiments, social experiments, and digital experiments. Although valuable in such applications, the costs of experiments often drive experimenters to seek more efficient designs. Recently, experimenters have started to examine such efficiency questions from an optimization perspective, as experimental design problems are fundamentally decision-making problems. This perspective offers a lot of flexibility in leveraging various existing optimization tools to study experimental design problems. This manuscript thus aims to examine the foundations of experimental design problems in the context of causal inference as viewed through an optimization lens.

In experimental design, Neyman allocation refers to the practice of allocating subjects into treated and control groups, potentially in unequal numbers proportional to their respective standard deviations, with the objective of minimizing the variance of the treatment effect estimator. This widely recognized approach increases statistical power in scenarios where the treated and control groups have different standard deviations, as is often the case in social experiments, clinical trials, marketing research, and online A/B testing. However, Neyman allocation cannot be implemented unless the standard deviations are known in advance. Fortunately, the multi-stage nature of the aforementioned applications allows the use of earlier stage observations to estimate the standard deviations, which further guide allocation decisions in later stages. In this paper, we introduce a competitive analysis framework to study this multi-stage experimental design problem. We propose a simple adaptive Neyman allocation algorithm, which almost matches the information-theoretic limit of conducting experiments. Using online A/B testing data from a social media site, we demonstrate the effectiveness of our adaptive Neyman allocation algorithm, highlighting its practicality especially when applied with only a limited number of stages.

Estimating the effects of long-term treatments in A/B testing presents a significant challenge. Such treatments -- including updates to product functions, user interface designs, and recommendation algorithms -- are intended to remain in the system for a long period after their launches. On the other hand, given the constraints of conducting long-term experiments, practitioners often rely on short-term experimental results to make product launch decisions. It remains an open question how to accurately estimate the effects of long-term treatments using short-term experimental data. To address this question, we introduce a longitudinal surrogate framework. We show that, under standard assumptions, the effects of long-term treatments can be decomposed into a series of functions, which depend on the user attributes, the short-term intermediate metrics, and the treatment assignments. We describe the identification assumptions, the estimation strategies, and the inference technique under this framework. Empirically, we show that our approach outperforms existing solutions by leveraging two real-world experiments, each involving millions of users on WeChat, one of the world's largest social networking platforms.

Practitioners and academics have long appreciated the benefits of covariate balancing when they conduct randomized experiments. For web-facing firms running online A/B tests, however, it still remains challenging in balancing covariate information when experimental subjects arrive sequentially. In this paper, we study an online experimental design problem, which we refer to as the "Online Blocking Problem." In this problem, experimental subjects with heterogeneous covariate information arrive sequentially and must be immediately assigned into either the control or the treated group. The objective is to minimize the total discrepancy, which is defined as the minimum weight perfect matching between the two groups. To solve this problem, we propose a randomized design of experiment, which we refer to as the "Pigeonhole Design." The pigeonhole design first partitions the covariate space into smaller spaces, which we refer to as pigeonholes, and then, when the experimental subjects arrive at each pigeonhole, balances the number of control and treated subjects for each pigeonhole. We analyze the theoretical performance of the pigeonhole design and show its effectiveness by comparing against two well-known benchmark designs: the match-pair design and the completely randomized design. We identify scenarios when the pigeonhole design demonstrates more benefits over the benchmark design. To conclude, we conduct extensive simulations using Yahoo! data to show a 10.2% reduction in variance if we use the pigeonhole design to estimate the average treatment effect.

This article studies experimental design in settings where the experimental units are large aggregate entities (e.g., markets), and only one or a small number of units can be exposed to the treatment. In such settings, randomization of the treatment may result in treated and control groups with very different characteristics at baseline, inducing biases. We propose a variety of experimental non-randomized synthetic control designs (Abadie, Diamond and Hainmueller, 2010, Abadie and Gardeazabal, 2003) that select the units to be treated, as well as the untreated units to be used as a control group. Average potential outcomes are estimated as weighted averages of the outcomes of treated units for potential outcomes with treatment, and weighted averages the outcomes of control units for potential outcomes without treatment. We analyze the properties of estimators based on synthetic control designs and propose new inferential techniques. We show that in experimental settings with aggregate units, synthetic control designs can substantially reduce estimation biases in comparison to randomization of the treatment.

Switchback experiments, where a firm sequentially exposes an experimental unit to random treatments, are among the most prevalent designs used in the technology sector, with applications ranging from ride-hailing platforms to online marketplaces. Although practitioners have widely adopted this technique, the derivation of the optimal design has been elusive, hindering practitioners from drawing valid causal conclusions with enough statistical power. We address this limitation by deriving the optimal design of switchback experiments under a range of different assumptions on the order of the carryover effect -- the length of time a treatment persists in impacting the outcome. We cast the optimal experimental design problem as a minimax discrete optimization problem, identify the worst-case adversarial strategy, establish structural results, and solve the reduced problem via a continuous relaxation. For switchback experiments conducted under the optimal design, we provide two approaches for performing inference. The first provides exact randomization based p-values, and the second uses a new finite population central limit theorem to conduct conservative hypothesis tests and build confidence intervals. We further provide theoretical results when the order of the carryover effect is misspecified and provide a data-driven procedure to identify the order of the carryover effect. We conduct extensive simulations to study the numerical performance and empirical properties of our results, and conclude with practical suggestions.

Our work is motivated by a common business constraint in online markets. While firms respect the advantages of dynamic pricing and price experimentation, they must limit the number of price changes (i.e., switches) to be within some budget due to various practical reasons. We study both the classical price-based network revenue management problem in the distributionally-unknown setup, and the bandits with knapsacks problem. In these problems, a decision-maker (without prior knowledge of the environment) has finite initial inventory of multiple resources to allocate over a finite time horizon. Beyond the classical resource constraints, we introduce an additional switching constraint to these problems, which restricts the total number of times that the decision-maker makes switches between actions to be within a fixed switching budget. For such problems, we show matching upper and lower bounds on the optimal regret, and propose computationally-efficient limited-switch algorithms that achieve the optimal regret. Our work reveals a surprising result: the optimal regret rate is completely characterized by a piecewise-constant function of the switching budget, which further depends on the number of resource constraints -- to the best of our knowledge, this is the first time the number of resources constraints is shown to play a fundamental role in determining the statistical complexity of online learning problems. We conduct computational experiments to examine the performance of our algorithms on a numerical setup that is widely used in the literature. Compared with benchmark algorithms from the literature, our proposed algorithms achieve promising performance with clear advantages on the number of incurred switches. Practically, firms can benefit from our study and improve their learning and decision-making performance when they simultaneously face resource and switching constraints.

We study a supply chain ordering problem faced by a wholesale supplier serving unpredictable demand. In this problem, the supplier has an initial stock, and faces a stream of orders for different amounts that are unknown a priori. Each order must be either accepted or rejected immediately, and must respect the knapsack constraint, that is, an order is only acceptable if its amount can be fully served by the remaining stock. The objective is to maximize the total stock spent serving orders. We investigate randomized threshold algorithms that accept an item as long as its size exceeds the threshold. We derive two optimal threshold distributions, the first is 0.4324-competitive relative to the optimal offline integral packing, and the second is 0.4285-competitive relative to the optimal offline fractional packing. Both results require optimizing the cumulative distribution function of the random threshold, which are challenging infinite-dimensional optimization problems. We also consider the generalization to multiple knapsacks, where an arriving item has a different size in each knapsack and must be placed in at most one knapsack. We derive a 0.2142-competitive algorithm for this problem. We also show that any randomized algorithm for this problem cannot be more than 0.4605-competitive. This is the first upper bound strictly less than 0.5, which implies the intrinsic challenge of knapsack constraint. We show how to naturally implement our optimal threshold distributions in the warehouses of a Latin American chain department store. We run simulations on their order data, which demonstrate the efficacy of our proposed algorithms.

This work is motivated by our collaboration with a large consumer packaged goods (CPG) company. We have found that while the company appreciates the advantages of dynamic pricing, they deem it operationally much easier to plan out a static price calendar in advance. We investigate the efficacy of static control policies for revenue management problems whose optimal solution is inherently dynamic. In these problems, a firm has limited inventory to sell over a finite time horizon, over which heterogeneous customers stochastically arrive. We consider both pricing and assortment controls, and derive simple static policies in the form of a price calendar or a planned sequence of assortments, respectively. In the assortment planning problem, we also differentiate between the static vs. dynamic substitution models of customer demand. We show that our policies are within 1-1/e (approximately 0.63) of the optimum under stationary (IID) demand, and 1/2 of the optimum under non-stationary demand, with both guarantees approaching 1 if the starting inventories are large. We adapt the technique of prophet inequalities from optimal stopping theory to pricing and assortment problems, and our results are relative to the linear programming relaxation. Under the special case of IID single-item pricing, our results improve the understanding of irregular and discrete demand curves, by showing that a static calendar can be (1-1/e)-approximate if the prices are sorted high-to-low. Finally, we demonstrate on both data from the CPG company and synthetic data from the literature that our simple price and assortment calendars are effective.