Search

In accordance with the principle of "data minimization", many internet companies are opting to record less data. However, this is often at odds with A/B testing efficacy. For experiments with units with multiple observations, one popular data minimizing technique is to aggregate data for each unit. However, exact quantile estimation requires the full observation-level data. In this paper, we develop a method for approximate Quantile Treatment Effect (QTE) analysis using histogram aggregation. In addition, we can also achieve formal privacy guarantees using differential privacy., Comment: Accepted to 2023 CODE conference as a parallel presentation

While there exists a large amount of literature on the general challenges of and best practices for trustworthy online A/B testing, there are limited studies on sample size estimation, which plays a crucial role in trustworthy and efficient A/B testing that ensures the resulting inference has a sufficient power and type I error control. For example, when sample size is under-estimated, the statistical inference, even with the correct analysis methods, will not be able to detect the true significant improvement leading to misinformed and costly decisions. This paper addresses this fundamental gap by developing new sample size calculation methods for correlated data, as well as absolute vs. relative treatment effects, both ubiquitous in online experiments. Additionally, we address a practical question of the minimal observed difference that will be statistically significant and how it relates to average treatment effect and sample size calculation. All proposed methods are accompanied by mathematical proofs, illustrative examples, and simulations. We end by sharing some best practices on various practical topics on sample size calculation and experimental design., Comment: Accepted by CIKM'23

With the growing popularity of intelligent assistants (IAs), evaluating IA quality becomes an increasingly active field of research. This paper identifies and quantifies the feedback effect, a novel component in IA-user interactions: how the capabilities and limitations of the IA influence user behavior over time. First, we demonstrate that unhelpful responses from the IA cause users to delay or reduce subsequent interactions in the short term via an observational study. Next, we expand the time horizon to examine behavior changes and show that as users discover the limitations of the IA's understanding and functional capabilities, they learn to adjust the scope and wording of their requests to increase the likelihood of receiving a helpful response from the IA. Our findings highlight the impact of the feedback effect at both the micro and meso levels. We further discuss its macro-level consequences: unsatisfactory interactions continuously reduce the likelihood and diversity of future user engagements in a feedback loop., Comment: PAKDD 2023

Technology companies are increasingly using randomized controlled trials (RCTs) as part of their development process. Despite having fine control over engineering systems and data instrumentation, these RCTs can still be imperfectly executed. In fact, online experimentation suffers from many of the same biases seen in biomedical RCTs including opt-in and user activity bias, selection bias, non-compliance with the treatment, and more generally, challenges in the ability to test the question of interest. The result of these imperfections can lead to a bias in the estimated causal effect, a loss in statistical power, an attenuation of the effect, or even a need to reframe the question that can be answered. This paper aims to make practitioners of experimentation more aware of imperfections in technology-industry RCTs, which can be hidden throughout the engineering stack or in the design process., Comment: Presented at CODE@MIT 2021

It often happens that the same problem presents itself to different communities and the solutions proposed or adopted by those communities are different. We take the case of the variance estimation of the population average treatment effect in cluster-randomized experiments. The econometrics literature promotes the cluster-robust variance estimator (Athey and Imbens, 2017), which can be dated back to the study of linear regression with clustered residuals (Liang and Zeger, 1986). The A/B testing or online experimentation literature promotes the delta method (Kohavi et al., 2010, Deng et al., 2017, 2018), which tackles the variance estimation of the ATE estimator directly using large sample theory. The two methods are seemly different as the former begins with a regression setting at the individual unit level and the latter is semi-parametric with only i.i.d. assumptions on the clusters. Both methods are widely used in practice. It begs the question for their connection and comparison. In this paper we prove they are equivalent and in the canonical implementation they should give exactly the same result.

Online experiments are the gold standard for evaluating impact on user experience and accelerating innovation in software. However, since experiments are typically limited in duration, observed treatment effects are not always permanently stable, sometimes revealing increasing or decreasing patterns over time. There are multiple causes for a treatment effect to change over time. In this paper, we focus on a particular cause, user-learning, which is primarily associated with novelty or primacy. Novelty describes the desire to use new technology that tends to diminish over time. Primacy describes the growing engagement with technology as a result of adoption of the innovation. User-learning estimation is critical because it holds experimentation responsible for trustworthiness, empowers organizations to make better decisions by providing a long-term view of expected impact, and prevents user dissatisfaction. In this paper, we propose an observational approach, based on difference-in-differences technique to estimate user-learning at scale. We use this approach to test and estimate user-learning in many experiments at Microsoft. We compare our approach with the existing experimental method to show its benefits in terms of ease of use and higher statistical power, and to discuss its limitation in presence of other forms of treatment interaction with time., Comment: Submitted to the Technometrics Journal

When interpreting A/B tests, we typically focus only on the statistically significant results and take them by face value. This practice, termed post-selection inference in the statistical literature, may negatively affect both point estimation and uncertainty quantification, and therefore hinder trustworthy decision making in A/B testing. To address this issue, in this paper we explore two seemingly unrelated paths, one based on supervised machine learning and the other on empirical Bayes, and propose post-selection inferential approaches that combine the strengths of both. Through large-scale simulated and empirical examples, we demonstrate that our proposed methodologies stand out among other existing ones in both reducing post-selection biases and improving confidence interval coverage rates, and discuss how they can be conveniently adjusted to real-life scenarios.

For ordinal outcomes, the average treatment effect is often ill-defined and hard to interpret. Echoing Agresti and Kateri (2017), we argue that the relative treatment effect can be a useful measure especially for ordinal outcomes, which is defined as $\gamma = \mathrm{pr}\{ Y_i(1) > Y_i(0) \} - \mathrm{pr}\{ Y_i(1) < Y_i(0) \}$, with $Y_i(1)$ and $Y_i(0)$ being the potential outcomes of unit $i$ under treatment and control, respectively. Given the marginal distributions of the potential outcomes, we derive the sharp bounds on $\gamma,$ which are identifiable parameters based on the observed data. Agresti and Kateri (2017) focused on modeling strategies under the assumption of independent potential outcomes, but we allow for arbitrary dependence., Comment: Accepted by Biometrics

On-line experimentation (also known as A/B testing) has become an integral part of software development. To timely incorporate user feedback and continuously improve products, many software companies have adopted the culture of agile deployment, requiring online experiments to be conducted and concluded on limited sets of users for a short period. While conceptually efficient, the result observed during the experiment duration can deviate from what is seen after the feature deployment, which makes the A/B test result biased. In this paper, we provide theoretical analysis to show that heavy-users can contribute significantly to the bias, and propose a re-sampling estimator for bias adjustment., Comment: 5 pages, CIKM'19

During the last decade, the information technology industry has adopted a data-driven culture, relying on online metrics to measure and monitor business performance. Under the setting of big data, the majority of such metrics approximately follow normal distributions, opening up potential opportunities to model them directly without extra model assumptions and solve big data problems via closed-form formulas using distributed algorithms at a fraction of the cost of simulation-based procedures like bootstrap. However, certain attributes of the metrics, such as their corresponding data generating processes and aggregation levels, pose numerous challenges for constructing trustworthy estimation and inference procedures. Motivated by four real-life examples in metric development and analytics for large-scale A/B testing, we provide a practical guide to applying the Delta method, one of the most important tools from the classic statistics literature, to address the aforementioned challenges. We emphasize the central role of the Delta method in metric analytics by highlighting both its classic and novel applications.

Inspired by the pioneering work of Rubin (1978), we employ the potential outcomes framework to develop a finite-population Bayesian causal inference framework for randomized controlled $2^K$ factorial designs with binary outcomes, which are common in medical research. As demonstrated by simulated and empirical examples, the proposed framework corrects the well-known variance over-estimation issue of the classic "Neymanian" inference framework, under various settings., Comment: To appear in Journal of Statistical Computation and Simulation

$2^K$ factorial designs are widely adopted by statisticians and the broader scientific community. In this short note, under the potential outcomes framework (Neyman, 1923; Rubin, 1974), we adopt the partial identification approach and derive the sharp lower bound of the sampling variance of the estimated factorial effects, which leads to an "improved" Neymanian variance estimator that mitigates the over-estimation issue suffered by the classic Neymanian variance estimator by Dasgupta et al. (2015)., Comment: Accepted by Statistica Neerlandica

In medical research, a scenario often entertained is randomized controlled $2^2$ factorial design with a binary outcome. By utilizing the concept of potential outcomes, Dasgupta et al. (2015) proposed a randomization-based causal inference framework, allowing flexible and simultaneous estimations and inferences of the factorial effects. However, a fundamental challenge that Dasgupta et al. (2015)'s proposed methodology faces is that the sampling variance of the randomization-based factorial effect estimator is unidentifiable, rendering the corresponding classic "Neymanian" variance estimator suffering from over-estimation. To address this issue, for randomized controlled $2^2$ factorial designs with binary outcomes, we derive the sharp lower bound of the sampling variance of the factorial effect estimator, which leads to a new variance estimator that sharpens the finite-population Neymanian causal inference. We demonstrate the advantages of the new variance estimator through a series of simulation studies, and apply our newly proposed methodology to two real-life datasets from randomized clinical trials, where we gain new insights., Comment: Accepted by Statistical Methods in Medical Research

Under the potential outcomes framework, we introduce matched-pair factorial designs, and propose the matched-pair estimator of the factorial effects. We also calculate the randomization-based covariance matrix of the matched-pair estimator, and provide the "Neymanian" estimator of the covariance matrix., Comment: To appear in Statistics & Probability Letters

Randomized controlled experiment has long been accepted as the golden standard for establishing causal link and estimating causal effect in various scientific fields. Average treatment effect is often used to summarize the effect estimation, even though treatment effects are commonly believed to be varying among individuals. In the recent decade with the availability of "big data", more and more experiments have large sample size and increasingly rich side information that enable and require experimenters to discover and understand heterogeneous treatment effect (HTE). There are two aspects in HTE understanding, one is to predict the effect conditioned on a given set of side information or a given individual, the other is to interpret the HTE structure and summarize it in a memorable way. The former aspect can be treated as a regression problem, and the latter aspect focuses on concise summarization and interpretation. In this paper we propose a method that can achieve both at the same time. This method can be formulated as a convex optimization problem, for which we provide stable and scalable implementation.

We develop finite-population asymptotic theory for covariate adjustment in randomization-based causal inference for 2K factorial designs. In particular, we confirm that both the unadjusted and covariate-adjusted estimators of the factorial effects are asymptotically normal, and the latter is more precise than the former., Comment: To appear in Statistics and Probability Letters

A/B testing is one of the most successful applications of statistical theory in modern Internet age. One problem of Null Hypothesis Statistical Testing (NHST), the backbone of A/B testing methodology, is that experimenters are not allowed to continuously monitor the result and make decision in real time. Many people see this restriction as a setback against the trend in the technology toward real time data analytics. Recently, Bayesian Hypothesis Testing, which intuitively is more suitable for real time decision making, attracted growing interest as an alternative to NHST. While corrections of NHST for the continuous monitoring setting are well established in the existing literature and known in A/B testing community, the debate over the issue of whether continuous monitoring is a proper practice in Bayesian testing exists among both academic researchers and general practitioners. In this paper, we formally prove the validity of Bayesian testing with continuous monitoring when proper stopping rules are used, and illustrate the theoretical results with concrete simulation illustrations. We point out common bad practices where stopping rules are not proper and also compare our methodology to NHST corrections. General guidelines for researchers and practitioners are also provided.

We extend the randomization-based causal inference framework in Dasgupta et al. (2015) for general 2^K factorial designs, and demonstrate the equivalence between regression-based and randomization-based inferences. Consequently, we justify the use of regression-based methods in 2^K factorial designs from a finite-population perspective.

Practitioners are interested in not only the average causal effect of the treatment on the outcome but also the underlying causal mechanism in the presence of an intermediate variable between the treatment and outcome. However, in many cases we cannot randomize the intermediate variable, resulting in sample selection problems even in randomized experiments. Therefore, we view randomized experiments with intermediate variables as semi-observational studies. In parallel with the analysis of observational studies, we provide a theoretical foundation for conducting objective causal inference with an intermediate variable under the principal stratification framework, with principal strata defined as the joint potential values of the intermediate variable. Our strategy constructs weighted samples based on principal scores, defined as the conditional probabilities of the latent principal strata given covariates, without access to any outcome data. This principal stratification analysis yields robust causal inference without relying on any model assumptions on the outcome distributions. We also propose approaches to conducting sensitivity analysis for violations of the ignorability and monotonicity assumptions, the very crucial but untestable identification assumptions in our theory. When the assumptions required by the classical instrumental variable analysis cannot be justified by background knowledge or cannot be made because of scientific questions of interest, our strategy serves as a useful alternative tool to deal with intermediate variables. We illustrate our methodologies by using two real data examples, and find scientifically meaningful conclusions.

We extend the heuristic discussion in Senn (2008) on the bias from selective inference for the treatment selection problem (Dawid 1994), by deriving the closed-form expression for the selection bias. We illustrate the advantages of our theoretical results through numerical and simulated examples., Comment: To appear in Statistics and Probability Letters

Assessing the causal effects of interventions on ordinal outcomes is an important objective of many educational and behavioral studies. Under the potential outcomes framework, we can define causal effects as comparisons between the potential outcomes under treatment and control. However, unfortunately, the average causal effect, often the parameter of interest, is difficult to interpret for ordinal outcomes. To address this challenge, we propose to use two causal parameters, which are defined as the probabilities that the treatment is beneficial and strictly beneficial for the experimental units. However, although well-defined for any outcomes and of particular interest for ordinal outcomes, the two aforementioned parameters depend on the association between the potential outcomes, and are therefore not identifiable from the observed data without additional assumptions. Echoing recent advances in the econometrics and biostatistics literature, we present the sharp bounds of the aforementioned causal parameters for ordinal outcomes, under fixed marginal distributions of the potential outcomes. Because the causal estimands and their corresponding sharp bounds are based on the potential outcomes themselves, the proposed framework can be flexibly incorporated into any chosen models of the potential outcomes, and are directly applicable to randomized experiments, unconfounded observational studies, and randomized experiments with noncompliance. We illustrate our methodology via numerical examples and three real-life applications related to educational and behavioral research., Comment: Accepted by the Journal of Education and Behavioral Statistics

For ordinal outcomes, we construct sequences of alternative hypotheses in increasing departures from the sharp null hypothesis of zero treatment effect on each experimental unit, to help assess the powers of randomization tests in randomized treatment-control experiments.