Robust and efficient multiple-unit switchback experimentation

Christian Delb\'e; Guido Imbens; Lorenzo Masoero; Paul Missault; Thomas Richardson

REVIEW 3 major objections 5 minor 13 references

Robust and efficient multiple-unit switchback experimentation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Regular Balanced Switchback Designs (RBSDs) randomize treatment across items and time under double balance, and the paper argues they estimate item-level average treatment effects with far lower variance than item-randomization or…

desk verdict RBSD is a genuinely useful design for point estimation, but the variance estimator is unproven and likely inconsistent, so the paper's inference claims overreach. read the letter →

arxiv 2506.12654 v1 pith:4UURBVW5 submitted 2025-06-14 stat.ME stat.AP

Paul Missault , Lorenzo Masoero , Christian Delbé , Thomas Richardson , Guido Imbens This is my paper · ORCID

classification stat.MEstat.AP MSC 62K0562K1062D05

keywords RegularBalancedSwitchbackDesignsexperimentsitemrandomizationcarryovereffectsHorvitz-Thompsonestimationonlinecausalinference

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

User-randomized A/B tests are often infeasible, illegal, or practically risky (for pricing changes, for example), so experimenters randomize at the item level, but item-randomization typically suffers from low power. This paper proposes Regular Balanced Switchback Designs (RBSDs), which re-randomize treatment over both items and time while keeping two margins fixed: at every timestep a fixed fraction of items are treated, and every item is treated the same number of times. Under no cross-item spillover and carryover effects bounded by m periods, a lag-l Horvitz-Thompson estimator is unbiased, and the balancedness sharply reduces estimator variance, especially when item-level metrics are skewed. In simulations on e-commerce sales data and in two online experiments, RBSDs yield tighter confidence intervals than item-randomized and regular switchback designs, so smaller effects can be detected with the same data. If the claim holds, item-level experiments become a much more attractive fallback when user-level randomization cannot be used.

What carries the argument

The load-bearing object is the RBSD assignment matrix: each row (item) is a random permutation of a fixed binary vector with $pS$ ones, and rows are stacked in complementary pairs (each row with its bitwise complement), which enforces both row and column balance. The estimator that carries the inference is the lag-$\ell$ Horvitz-Thompson estimator of eq. (5), which uses the probability of observing a run of $\ell+1$ identical assignments to reweight each item-time observation; Lemma 4 gives the closed-form run probability for the paired-permutation design. Regularity provides robustness to carryover, balancedness provides variance reduction, and the combination is what the paper claims dominates both simpler designs.

What would settle it

Run a placebo (A/A) simulation under an RBSD with paired complement rows for $N=10{,}000$ items and $S=14$ days, draw 100 assignment matrices, and compare the Monte Carlo standard deviation of $\hat{\tau}_\ell$ to the average of the Theorem 2 standard-error estimates. If the average estimated standard error is systematically below the empirical standard deviation, or if 95% z-intervals cover the true value less than 95% of the time, the variance-estimator consistency used for the reported tests is refuted.

Watch

Extended reading notes

Core claim

The central claim is that combining regular switchback randomization (treatment is held constant between breakpoints, so temporal carryover can be modeled) with double balancedness (fixed row and column sums) makes item-level randomized experiments both robust to carryover effects and substantially more efficient. The paper defines the lag-$\ell$ average treatment effect $\tau_\ell$, which discards the first $\ell$ timesteps so that bounded carryover effects do not contaminate the estimand, and shows that the Horvitz-Thompson estimator in eq. (5) is unbiased for $\tau_\ell$ under Assumptions 1, 2, and 4 for any design with positive assignment probabilities. Theorems 1 and 2 provide a simple variance estimator based on the dispersion of per-item lagged individual treatment effects. Empirically, the paper argues that RBSD estimates have the smallest standard errors among item-randomization, multi-unit regular switchback, and RBSD, with no additional bias, and that the improvement carries over to real e-commerce interventions.

Load-bearing premise

The confidence intervals rely on treating each item's estimated effect as nearly independent, yet the recommended sampling scheme deliberately pairs every item with its mirror-image treatment sequence, so those estimates are negatively correlated; if that correlation is not negligible, the intervals would be too narrow.

Editorial extensions

If this is right

With the same number of items and days, an RBSD experiment can detect smaller treatment effects than item-randomization, without inflating false positives.
When carryover effects of known duration are present, using the lag-$\ell$ estimator with an RBSD removes the bias that would otherwise hit naive switchback estimates.
RBSD designs make item-level randomization practical in settings where user-randomization is legally or ethically barred, such as price experiments.
The variance gains are largest in skewed item-level metrics, where a few items dominate sales.
Clustering items into families, as in the second online experiment, extends the same efficiency benefits to settings where cross-item interference is a concern.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's standard-error formulas treat per-item estimates as nearly independent, but the recommended paired-complement construction couples each item with its mirror image; if that negative dependence is non-negligible, the reported confidence intervals could be optimistic until a design-aware variance estimator is used.
The minimax optimality of the half-half allocation is a worst-case guarantee; on real skewed data, a different treatment fraction might do better, so choosing the treatment proportion adaptively is a natural extension the paper does not explore.
The efficiency comparison assumes no cross-item spillover; with spillovers, clustering becomes necessary, and the relative gain of RBSD over clustered item-randomization remains to be quantified.
The paired-complement construction works directly only for an even number of items, so odd-sized inventories would require a separate balancing device and may not inherit the same variance guarantees.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First, the useful part: RBSDs are a sensible new design class in a practical area. Combining regular switchbacks (for carryover robustness) with double balancedness (for variance reduction) is not in the prior work, the sampling construction in Lemmas 2–3 is concrete, and the closed-form probabilities in Lemma 4 are useful. The unbiasedness of the Horvitz–Thompson estimators under the stated assumptions is proven cleanly in Appendices B.1–B.2. The simulation and the two online experiments show the design can be dramatically more precise than item-randomization on skewed e-commerce data. That part is credible and worth building on.

The load-bearing problem is the standard-error estimator. Theorems 1 and 2 call sigma-hat and sigma-hat_l consistent, but the appendices prove unbiasedness only; no consistency proof appears. The estimator is the sample standard deviation of the item-level dITEs divided by sqrt(N), which is consistent only when those item-level terms are mutually uncorrelated. The designs do not ensure that. Under Assumption 1, a common time component is allowed: if Y_{n,s}=a_s for all n, every dITE_n is identical, sigma-hat=0, while tau-hat has positive design variance. The RBSD construction in Lemma 3 pairs each row with its complement, producing negative dependence between item-level estimates, and that dependence is not accounted for. This is not a pedantic corner case; it means the false-positive rate, false-negative rate, and confidence intervals reported in Sections 5–6 rest on an unproven and in general false premise. The point-estimate variance improvements may well be real, but the paper's claim to 'systematically outperform' on inference is not supported.

A few smaller things: Definition 6 gives rejection sampling while Lemmas 3–4 use paired complements; these are different sampling distributions and the paper should pick one and stick to it. The minimax result in Lemma 1 is imported from single-unit results; that is fine, but it should be flagged as an import rather than a new derivation. No code or data is released, which will slow replication.

Who this is for: experimenters in online marketplaces who want a concrete multi-unit switchback design and are willing to check the variance question themselves. It deserves a serious referee, but the referee should demand a real consistency proof or a revised statement that only gives conservative variance under stronger assumptions. In current form, the inference claims are not defensible.

Referee Report

3 major / 5 minor

Summary. The paper proposes a class of experimental designs, Regular Balanced Switchback Designs (RBSDs), that randomize treatment jointly over items and time while enforcing both row and column balance. The authors define Horvitz-Thompson estimators for the average treatment effect and for a lag-ℓ version that is designed to remain unbiased in the presence of m-period carryover effects. They state theorems on unbiasedness and on consistency of standard-error estimators, provide a practical sampling scheme based on complement pairing of treatment rows, and support the method with Monte Carlo simulations on e-commerce sales data and with two field experiments.

Significance. The proposed design addresses a real practical need: item-randomized experiments often have low power, and standard switchback designs may be biased under carryover. If the paper's claims were fully established, RBSD would be a useful addition to the experimenter's toolkit. The paper is transparent about its assumptions, gives explicit algorithms, and provides closed-form assignment probabilities. A clear strength is that the unbiasedness of the point estimators is proved correctly in Appendices B.1 and B.2 (provided the lag ℓ is matched to the carryover order m). However, the paper's central efficiency and inference claims rest on standard-error estimators whose consistency is neither proved nor generally true, so the empirical conclusions in Sections 5 and 6 are not yet convincingly supported.

major comments (3)

[Section 3.3, Theorem 1] The theorem claims that sigma-hat(W) is a consistent estimator of the standard error of tau-hat under any randomization design, but Appendix B.1 proves only unbiasedness of tau-hat. The claim is false in general. Under the switchback design of Definition 2, all items share the same assignment sequence (W_{n,s}=W_{1,s} for all n). Suppose the potential outcomes satisfy Assumption 1 and are identical across items up to an item-invariant time series, e.g., Y_{n,s}(W)=a_s for all n. Then dITE_n(W) is identical for all n for every W, so sigma-hat(W)=0, while tau-hat = dITE_1(W) has positive design variance under any nondegenerate assignment distribution. Thus sigma-hat can severely underestimate the true standard error. This directly affects the z-tests and confidence intervals in Sections 5.2 and 6 and the variance comparisons in Section 5.1, where sigma-hat is reported as the standard error of the ATE.
[Section 3.3, Theorem 2] The same problem affects sigma-hat_l. Under the common-time example above, the l-lagged item-level estimates are identical across items under switchback, so sigma-hat_l=0 while tau-hat_l has positive variance. In addition, Theorem 2 states unbiasedness under Assumption 4 without requiring that the lag ℓ be at least the carryover order m. If m>ℓ, the estimator in eq. (5) weights only the last ℓ+1 assignments and does not control for the older assignments W_{n,s-m:s-ℓ-1}; the expectation is then not equal to Y^T-Y^C in general. Appendix B.2 hides this issue by using the same symbol p for both the lag and the carryover order. The theorem should state ℓ ≥ m, and the proof should be rewritten with distinct notation for the lag and the carryover horizon.
[Section 4.2.3, Lemma 3] The recommended RBSD sampling scheme stacks each treatment row with its complement, so item-level dITEs are dependent across items: the two rows in each complement pair have opposite assignments, and any common time component affects both rows. Theorems 1 and 2 do not state conditions under which the sample-variance estimators remain consistent under this dependence, and no proof is supplied. Because Section 5.1 reports distributions of sigma-hat as evidence of efficiency, and Section 5.2 and Section 6 use z-tests built on sigma-hat for false-positive/false-negative rates and confidence intervals, the empirical claim that RBSDs 'systematically outperform' existing designs is not yet established. The authors should either prove consistency under complement pairing or provide a corrected variance estimator (for example, based on independent pairs or on the known assignment distribution) and re-run the affected empirical analyses.

minor comments (5)

[Appendix B.2] The proof uses the symbol p for both the lag ℓ and the carryover order m, which is confusing and obscures the required condition ℓ ≥ m; please use distinct notation.
[Lemma 4] The notation is inconsistent: the event is written as '1_{s-m:s}' in the statement and as '1_{s-j:s}' in the display; it should be W_{n,s-j:s}=1^{j+1}.
[Definition 3] Definition 3 says 'A probability distribution ... is a switchback design,' which is too vague and makes the uniform-distribution example appear tautological; please state the intended class explicitly.
[Section 5.2] The statement that observed false-positive rates below 0.05 'highlight that the variance calculated per Theorem 1 or Theorem 2 is conservative' is not logically implied; under the variance-estimator concerns above, the FPR could be below nominal for other reasons. Please qualify the claim.
[Section 4.2.3] Definition 6 allows arbitrary treatment probability p, but the practical sampling scheme in Lemma 3 realizes only p=1/2; please clarify whether the theoretical and empirical results are restricted to balanced p=1/2 designs.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the estimators and variance claims are derived from stated assumptions or imported from external results, and the efficiency claims rest on independent simulations and field experiments rather than on fitted inputs or self-citation.

full rationale

The derivation chain is self-contained in the relevant sense. Assumptions 1-4 plus the Horvitz-Thompson structure directly give unbiasedness of \hat\tau and \hat\tau_\ell, and the proofs in Appendices B.1 and B.2 show the expectation calculations without assuming the conclusion. Lemma 4 derives the treatment-window probabilities from the permutation construction rather than imposing them. Lemma 1 imports the minimax result of Bojinov et al. (2023) as an external theorem; it is not a self-citation and is not used to define the paper's own designs. The paper's central efficiency claim about RBSDs is supported by Monte Carlo simulations with fixed, non-fitted effect sizes and by two online experiments, not by a parameter fitted to the outcome. The citation to Masoero et al. (2023) is a self-citation, but it is used only as background motivation for balanced designs and is independently established prior work; it does not carry the empirical conclusions of the present paper. The most serious concern, the claimed consistency of \hat\sigma and \hat\sigma_\ell under arbitrary designs, is asserted without proof and may fail under shared time shocks or under the complement-pair dependence induced by Lemma 3, but that is a statistical validity gap rather than a circular reduction: no equation is defined in terms of its conclusion, and no prediction is equivalent to a fitted input. Hence the circularity score is 0.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The method itself has no fitted parameters; p and l are design choices. The load-bearing assumptions are the stable-unit-treatment-value Assumption 1 (no cross-item spillover), non-anticipation, and bounded known carryover order m. The variance estimator additionally imports an unstated independence assumption across items.

assumptions (5)

domain assumption Assumption 1 (SUTVA, no cross-item spillovers): potential outcomes for one item depend only on its own treatment assignments.
All unbiasedness results and efficiency claims rest on this, stated in Section 3.2; violations via competing items are acknowledged in Remark 1.
domain assumption Assumption 2 (Non-anticipating potential outcomes): outcomes at time s do not depend on future assignments.
Used in proofs of Theorems 1 and 2 to reduce Y_n,s(W) to a function of current and past assignments only.
domain assumption Assumption 4 (m-Carryover with known m): the outcome at time s depends on at most m previous assignments.
The lag-l estimator is unbiased only if carryover is bounded by a known m and l is set to m; if the true carryover horizon exceeds l, bias enters, a limitation acknowledged in Section 7.
standard math Assumption 5 (Bounded potential outcomes): |Y_n,s(W)| <= B for all n, s, W.
Technical condition for minimax Lemma 1, implicitly used to apply Bojinov et al. 2023's variance bound.
domain assumption Independence of item-level ITE estimators: the variance formulas in Theorems 1 and 2 assume the dITE_n terms are independent or uncorrelated across items.
Not stated or verified under Lemma 3's paired complement design, where rows are dependent; this is a gap in the inference claims.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Robust and efficient multiple-unit switchback experimentation." pith.science (2026). https://pith.science/paper/4UURBVW5

@misc{pith2026250612654,
  author       = {Pith},
  title        = {Pith review of: Robust and efficient multiple-unit switchback experimentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4UURBVW5}},
  note         = {Machine review of arXiv:2506.12654}
}

read the original abstract

User-randomized A/B testing has emerged as the gold standard for online experimentation. However, when this kind of approach is not feasible due to legal, ethical or practical considerations, experimenters have to consider alternatives like item-randomization. Item-randomization is often met with skepticism due to its poor empirical performance. To fill this gap, in this paper we introduce a novel and rich class of experimental designs, "Regular Balanced Switchback Designs" (RBSDs). At their core, RBSDs work by randomly changing treatment assignments over both time and items. After establishing the properties of our designs in a potential outcomes framework, characterizing assumptions and conditions under which corresponding estimators are resilient to the presence of carryover effects, we show empirically via both realistic simulations and real e-commerce data that RBSDs systematically outperform standard item-randomized and non-balanced switchback approaches by yielding much more accurate estimates of the causal effects of interest without incurring any additional bias.

Figures

Figures reproduced from arXiv: 2506.12654 by the authors.

**Figure 2.** Results of a Monte Carlo simulation, over 100 re-draws of the assignment matrix [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

**Figure 3.** Empirical distribution of ˆσ (left) and ˆσ1 (right) under the same setting of fig. 2 0 100 200 300 400 500 10−5 10−4 10−3 10−2 Units sold Frequency (Log Scale) Log-Normal with µ = 2.4507 and σ = 1.4764 (a) Distribution of item sales after removing items for which no sales were measured, and winsorizing at the 99th percentile. Skew coefficient of 4.3. 1 2 3 4 5 6 0 0.5 1 Units purchased Frequency (Log Scale) Power la… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Distributions of ASIN sales and user purchases in the A/A data described in section 5.1. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 8 canonical work pages

[1]

Aronow and Cyrus Samii

Peter M. Aronow and Cyrus Samii. Estimating average causal effects under general interference, with application to a social network experiment. The Annals of Applied Statistics , 11(4):1912 – 1947,

work page 1912
[7]

Garrett Johnson, Randall A Lewis, and Elmar Nubbemeyer

URL https://arxiv.org/abs/2312.15574. Garrett Johnson, Randall A Lewis, and Elmar Nubbemeyer. The online display ad effectiveness funnel & carryover: Lessons from 432 field experiments. Available at SSRN 2701578 ,

arXiv
[9]

Lorenzo Masoero, Suhas Vijaykumar, Thomas Richardson, James McQueen, Ido Rosen, Brian Burdick, Pat Bajari, and Guido Imbens

URL https://www.amazon.science/publications/ efficient-switchback-experiments-via-multiple-randomization-designs . Lorenzo Masoero, Suhas Vijaykumar, Thomas Richardson, James McQueen, Ido Rosen, Brian Burdick, Pat Bajari, and Guido Imbens. Multiple randomization designs: Estimation and inference with interference. arXiv preprint arXiv:2401.01264 ,

work page arXiv
[10]

Using ego-clusters to measure network effects at LinkedIn

20 Guillaume Saint-Jacques, Maneesh Varshney, Jeremy Simpson, and Ya Xu. Using ego-clusters to measure network effects at LinkedIn. arXiv preprint arXiv:1903.08755 ,

arXiv 1903
[13]

Moreover, Assumption 2 and Assumption 4 allow us to write each potential outcome Yn,s as a function of only the assignments 23 [Wn,s−p,

Assumption 1 allows us to express potential outcomes only as a function of the individual unit’s assignments. Moreover, Assumption 2 and Assumption 4 allow us to write each potential outcome Yn,s as a function of only the assignments 23 [Wn,s−p, . . . , Wn,s]. In turn, we can write E [ˆτℓ] = NX n=1 SX s=p+1 E 1(Wn,s−p:s = 1p+1)Yn,s(Wn,s−p:s = 1p+1) N (S −...

work page 2023
[1864]

Interference, bias, and variance in two- sided marketplace experimentation: Guidance for platforms

Hannah Li, Geng Zhao, Ramesh Johari, and Gabriel Y Weintraub. Interference, bias, and variance in two- sided marketplace experimentation: Guidance for platforms. In Proceedings of the ACM Web Conference 2022, pages 182–192,

work page 2022
[1941]

Science of price experimentation at Amazon

Joe Cooprider and Shima Nassiri. Science of price experimentation at Amazon. In AEA 2023, NABE 2023 ,

work page 2023
[1949]

Optimal experimental design for staggered rollouts

Ruoxuan Xiong, Susan Athey, Mohsen Bayati, and Guido Imbens. Optimal experimental design for staggered rollouts. Management Science, 70(8):5317–5336, 2024a. Ruoxuan Xiong, Alex Chin, and Sean J Taylor. Data-driven switchback experiments: Theoretical tradeoffs and empirical Bayes designs. arXiv preprint arXiv:2406.06768 , 2024b. F Yates. The gain in effici...

arXiv

Show all 13 references

[1952]

Switchback experiments under geometric mixing

Yuchen Hu and Stefan Wager. Switchback experiments under geometric mixing. arXiv preprint arXiv:2209.00197,
[2017]

URL https://doi.org/10.1214/16-AOAS1005

doi: 10.1214/16-AOAS1005. URL https://doi.org/10.1214/16-AOAS1005. Patrick Bajari, Brian Burdick, Guido W Imbens, Lorenzo Masoero, James McQueen, Thomas S Richardson, and Ido M Rosen. Experimental design in marketplaces. Statistical Science, 38(3):458–476,

doi
[2019]

EJ Williams

Accessed: 2024- 12-19. EJ Williams. Experimental designs balanced for pairs of residual effect. Australian Journal of Chemistry , 3 (3):351–363,

2024
[2023]

The use of triangulation in qualitative research

Nancy Carter. The use of triangulation in qualitative research. Number 5/September 2014 , 41(5):545–547,

2014
[2025]

Accessed: 2025-01-27

URL https://developers.google.com/ search/docs/crawling-indexing/website-testing. Accessed: 2025-01-27. Spence Green, Jeffrey Heer, and Christopher D Manning. The efficacy of human post-editing for language translation. In Proceedings of the SIGCHI conference on human factors ...

2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Aronow and Cyrus Samii

Peter M. Aronow and Cyrus Samii. Estimating average causal effects under general interference, with application to a social network experiment. The Annals of Applied Statistics , 11(4):1912 – 1947,

work page 1912

[2] [7]

Garrett Johnson, Randall A Lewis, and Elmar Nubbemeyer

URL https://arxiv.org/abs/2312.15574. Garrett Johnson, Randall A Lewis, and Elmar Nubbemeyer. The online display ad effectiveness funnel & carryover: Lessons from 432 field experiments. Available at SSRN 2701578 ,

arXiv

[3] [9]

Lorenzo Masoero, Suhas Vijaykumar, Thomas Richardson, James McQueen, Ido Rosen, Brian Burdick, Pat Bajari, and Guido Imbens

URL https://www.amazon.science/publications/ efficient-switchback-experiments-via-multiple-randomization-designs . Lorenzo Masoero, Suhas Vijaykumar, Thomas Richardson, James McQueen, Ido Rosen, Brian Burdick, Pat Bajari, and Guido Imbens. Multiple randomization designs: Estimation and inference with interference. arXiv preprint arXiv:2401.01264 ,

work page arXiv

[4] [10]

Using ego-clusters to measure network effects at LinkedIn

20 Guillaume Saint-Jacques, Maneesh Varshney, Jeremy Simpson, and Ya Xu. Using ego-clusters to measure network effects at LinkedIn. arXiv preprint arXiv:1903.08755 ,

arXiv 1903

[5] [13]

Moreover, Assumption 2 and Assumption 4 allow us to write each potential outcome Yn,s as a function of only the assignments 23 [Wn,s−p,

Assumption 1 allows us to express potential outcomes only as a function of the individual unit’s assignments. Moreover, Assumption 2 and Assumption 4 allow us to write each potential outcome Yn,s as a function of only the assignments 23 [Wn,s−p, . . . , Wn,s]. In turn, we can write E [ˆτℓ] = NX n=1 SX s=p+1 E 1(Wn,s−p:s = 1p+1)Yn,s(Wn,s−p:s = 1p+1) N (S −...

work page 2023

[6] [1864]

Interference, bias, and variance in two- sided marketplace experimentation: Guidance for platforms

Hannah Li, Geng Zhao, Ramesh Johari, and Gabriel Y Weintraub. Interference, bias, and variance in two- sided marketplace experimentation: Guidance for platforms. In Proceedings of the ACM Web Conference 2022, pages 182–192,

work page 2022

[7] [1941]

Science of price experimentation at Amazon

Joe Cooprider and Shima Nassiri. Science of price experimentation at Amazon. In AEA 2023, NABE 2023 ,

work page 2023

[8] [1949]

Optimal experimental design for staggered rollouts

Ruoxuan Xiong, Susan Athey, Mohsen Bayati, and Guido Imbens. Optimal experimental design for staggered rollouts. Management Science, 70(8):5317–5336, 2024a. Ruoxuan Xiong, Alex Chin, and Sean J Taylor. Data-driven switchback experiments: Theoretical tradeoffs and empirical Bayes designs. arXiv preprint arXiv:2406.06768 , 2024b. F Yates. The gain in effici...

arXiv