 Research
 Open access
 Published:
Power and sample sizes estimation in clinical trials with treatment switching in intentiontotreat analysis: a simulation study
BMC Medical Research Methodology volume 23, Article number: 49 (2023)
Abstract
Background
Treatment switching, also called crossover, is common in clinical trials because of ethical concerns or other reasons. When it occurs and the primary objective is to identify treatment effects, the most widely used intentiontotreat analysis may lead to underpowered trials. Here, we presented an approach to preview power reductions and to estimate sample sizes required to achieve the desired power when treatment switching occurs in the intentiontotreat analysis.
Methods
We proposed a simulationbased approach and developed an R package to perform power and sample sizes estimation in clinical trials with treatment switching.
Results
We simulated a number of randomized trials incorporating treatment switching and investigated the impact of the relative effectiveness of the experimental treatment to the control, the switching probability, the switching time, and the deviation between the assumed and the real distributions for the survival time on power reductions and sample sizes estimation. The switching probability and the switching time are key determinants for significant power decreasing and thus sample sizes surging to maintain the desired power. The sample sizes required in randomized trials absence of treatment switching vary from around fourfifths to oneseventh of the sample sizes required in randomized trials allowing treatment switching as the switching probability increases. The power reductions and sample sizes increase with the decrease of switching time.
Conclusions
The simulationbased approach not only provides a preview for power declining but also calculates the required sample size to achieve an expected power in the intentiontotreat analysis when treatment switching occurs. It will provide researchers and clinicians with useful information before randomized controlled trials are conducted.
Background
Clinical trials are often used to assess the effectiveness of new treatments. The most reliable clinical trial is the randomized controlled trials (RCT). In an RCT, patients are randomly assigned to the control group, where they receive a placebo or an existing standard treatment, or the experimental group, where they receive the experimental treatment. For treatments against diseases like cancer, the effectiveness of the new treatment is determined by comparing the observed survival time of patients in the experimental group to those in the control group. The power of a statistical test, which is the probability to correctly tell if the experimental treatment is better when it is truly better, is extremely important. Being able to correctly determine the effectiveness of an experimental treatment gives justification for its production, allows for better utilization of resources, and helps those who will benefit from the experimental treatment.
A common method used to interpret the data from a RCT is the intentiontotreat analysis (ITT). The ITT analysis includes all patients with randomization in statistical analysis and compares their responses to determine the effectiveness of the experimental treatment according to the treatment group that was initially assigned to them, regardless of what treatment they actually received. Patients, however, do not always comply with the treatment they were assigned to. Because of ethical concerns or other reasons, they may switch treatments from the control group to the experimental group. For example, it may occur when a disease progresses or when the healthcare provider believes the patient’s prognosis will improve with the experimental treatment. When treatment switching, also called crossover, is permitted, the ITT analysis is confounded, which decreases the power of a statistical test. A simple and alternative approach, such as the perprotocol analysis, excludes the participants who switch treatments from the statistical analysis. Nevertheless, this may heavily bias the analysis results if the participants included and excluded in the analysis significantly differ in prognosis, i.e., treatment switching is associated with prognostic variables [1].
Various statistical methods dealing with treatment switching have been proposed. Law and Kaldor [2] proposed the adjusted Cox model by splitting the study population into four groups according to their initial and final treatment groups and by assuming different hazard functions with time dependent covariates defined by the time of switching treatment for the four group. Loeys and Goetghebeur [3] proposed the causal proportional hazards estimator by assuming that patients in one arm complete their treatment while the patients in the other arm either completely fulfills their treatment or does not fulfill their treatment at all. Robins and Tsiatis [4] proposed correcting for noncompliance by using the rank preserving structural accelerated failure time models (RPSFT). Based on the RPSFT, Branson and Whitehead [5] and Zhang and Chen [6] provided iterative parametric estimation and modified iterative parametric estimation methods, respectively, for fast and reliably estimating the treatment effect. Morden et al. [1] and Latimer et al. [7, 8] performed simulations studies to compare several adjustment methods mentioned above.
However, sample sizes estimation in most RCTs is still based on the assumption of no treatment switching and predetermined statistical tests, such as the logrank test in ITT analysis. Given the sample sizes estimated from noswitching designs, the ITT analysis may underestimate the positive treatment effect when treatment switching occurs. In this study, we proposed a simulationbased approach, PowerSwitchingTrial, to preview power reductions and calculate sample sizes required in RCTs allowing treatment switching. We investigated the impact of the relative effectiveness of the experimental treatment to the control, the switching probability, the switching time, and the deviation of the assumed survival distribution from the real one on power reductions and sample sizes estimation. We found the switching probability and the switching time are key determinants for power decreasing and thus sample sizes increasing to maintain the desired power in ITT analysis. PowerSwitchingTrial is freely available at https://github.com/darwinhub/PowerSwitchingTrial.
Methods
PowerSwtichingTrial overview
PowerSwitchingTrial is a simulationbased R package, developed for exploring statistical powers and sample sizes required in RCT trials allowing treatment switching when the logrank test is used in ITT analysis. The flowchart of PowerSwitchingTrial is shown in Fig. 1.
PowerSwtichingTrial simulates a twoarm randomized clinical trial study with an accrual time of T_{a} and an additional followup time of T_{e} − T_{a}, where T_{e} is the time from the start to the end of the trial. Participants are assumed to enter the study uniformly during the accrual time period, i.e., the entry time follows a uniform distribution U(0, T_{a}), and soon be randomly assigned to one of the two treatment groups, the control group or the experimental group, with an allocation ratio of r (r is the ratio of the numbers of participants in the experimental and the control groups). T_{a} = 0 indicates all participants enter the study at the beginning of the study. The survival time of participants in the control and the experimental groups, denoted by T_{1} and T_{2}, is assumed to follow exponential distributions with median values of m_{1} and m_{2}, respectively. The censoring time of participants in the two groups, denoted by C_{1} and C_{2}, is defined as the time from randomization to dropout or to the end of the trial if participants do not experience the event of interest. The censoring consists of both dropout censoring and administrative censoring. The distributions of the censoring time for C_{1} and C_{2} are assumed to be the same and can be formulated as follows [9]:
where v is the entry time, and d(c) and \(\overline{D}(c)\) are the density function and survival function of the dropout censoring, respectively. I(∙) is the indicator function. For simplicity, we assume a uniform distribution for the dropout censoring, i.e., d(c) = h^{−1} I(0 < c < h), and h is determined by the formula of P(C_{1} < T_{1}) with a given censoring rate of the control group under no treatment switching. P(C_{1} < T_{1}) is a function of h and can be explicitly expressed (See Supplementary Materials for details).
PowerSwtichingTrial allows participants in the control group have a chance to switch to the experimental group if predetermined conditions are satisfied. For example, if patients with cancer have a disease progression before death (assume death is the event of interest), they may switch from the standard treatment to the new treatment after disease progression and being evaluated by the investigator [10]. The switching time, denoted by s, is the time from randomization to the moment when a participant may switch with a probability p. p, defined as the switching probability, is the probability that a participant who qualifies for treatment switching will switch from the control group to the experimental group after being evaluated by physicians or health care professionals. Five options were provided for the switching time. The first three options assume s is correlated with T_{1} and s = X × T_{1}, where X is independent of the survival time T_{1} and follows one of the three distributions, beta, gamma, or uniform U(0, 1). Since most often patients are allowed to switch to the new therapy at the point of disease progression, the assumption that switching time is correlated with the survival time is appropriate. The factor X controls the distribution of the switching time and the correlation between the switching time and the survival time T_{1} by the three different distributions and the associated parameters. If X follows beta or uniform distributions, s < T_{1}, denoting the switching time is always less than the survival time. In contrast, if X follows gamma distributions, s has a chance to be greater than T_{1}. The parameters in the beta and the gamma distributions are determined by given parameters pt and rho (Supplementary Material). pt = E(s)/E(T_{1}) denotes the ratio of the average switching time to the average survival time of the control group and rho denotes the correlation between s and T_{1}. If the switching time is the time when disease progresses, pt and rho can be estimated by the time of progression and the time of events from previous studies. If X ~ U(0, 1), then pt = 0.5 and rho is a constant of 0.775 when T_{1} follows an exponential distribution (Supplementary Material). The other two options assume that s is independent of T_{1}, where s either follows an exponential distribution with its mean determined by a given pt, or s is a constant. The survival time for participants starting from switching is assumed to increase by a constant multiplier of A, based on the rank preserving structural failure time model [4]. That is, the survival time of the participants with treatment switching will be \({T}_1^{\ast }=s+\left({T}_1s\right)\times A\), A = m_{2}/m_{1}. Therefore, the observable survival time Y_{1} for the participants who are originally assigned to the control group will be equal to min(T_{1}, C_{1}) without treatment switching and \(\mathit{\min}\left({T}_1^{\ast },{C}_1\right)\) with treatment switching. The observable survival time Y_{2} for the participants in the experimental group will be equal to min (T_{2}, C_{2}). Finally, the logrank test is used to compare Y_{1} and Y_{2}.
Two main functions of PowerSwtichingTrial
PowerSwtichingTrial provides two main functions, LogRankTestMix2PowerMedian and LogRankTestMix2Nmedian. LogRankTestMix2PowerMedian calculates the power, given sample sizes, the difference in the survival time between two groups, and the other parameters. LogRankTestMix2NMedian estimates sample sizes, given an expected power to detect the difference in the survival time between two groups, and the other required parameters.
LogRankTestMix2PowerMedian, a function to calculate the power, includes 13 input parameters to simulate a number of scenarios, n, m_{1}, m_{2}, r, T_{a}, T_{e}, p, alpha, s.dist, pt, rho, censor.rate and reps. n is the number of participants in the control group, m_{1}, m_{2}, r, T_{a}, T_{e}, and the switching probability (p) were defined in the Section 2.1, and alpha is the significance level. s.dist is the option for the switching time, which should be one of the five options: “beta”, “gamma”, “unif”, “indepExp”, or a constant value. Those parameters in the beta, gamma, or independent exponential distributions (“indepExp”) are determined by the given pt and rho (Section 2.1 and Supplementary Material). censor.rate is a predetermined censoring rate of the control group under no treatment switching. When censor.rate = “AC.only”, it assumes no dropout censoring but administrative censoring only. Reps is the number of computer simulations. The function will return the power, that is, the average number of rejecting the null hypothesis which suggests no difference in survival time between the two treatment groups across reps simulations, and the expected number of events for the two groups.
LogRankTestMix2Nmedian is a function to calculate the required sample sizes to meet a given power and other settings. LogRankTestMix2NMedian uses the same parameters as LogRankTestMix2PowerMedian, except the parameter of n being replaced by power, lower, and upper. The lower and upper is the minimum and maximum sample sizes that users set to explore. The function uses the bisection method to find the sample size (n) required between the lower and upper bounds.
Results
A scenario from a real clinical trial
We first demonstrated the application of the two R functions using a real clinical trial setting. The study was an openlabel phase III trial to compare the benefit of survivals in patients with chemotherapyrefractory metastatic colorectal cancer who were randomly assigned to either panitumumab + best supportive care (BSC) or BSC alone [10]. After 1040 patients were screened, 231 patients were randomly assigned to panitumumab + BSC and 232 to BSC alone. The maximum followup time was 113 weeks (~ 26 months). Thus, we assumed T_{a} = 0 and T_{e} = 26. Among BSC alone patients, 85% had disease progression, and 76% crossed over to panitumumab + BSC after being evaluated by the investigator. That is, the switch probability p was 0.89 (=0.76/0.85). According to their ITT results reported on ClinicalTrial.gov (NCT00113763), the median of overall survival (OS) in the panitumumab + BSC group was 6.4 months (m_{2} = 6.4) and the median of OS in the BSC alone group was 6.2 months (m_{1} = 6.2). There was no significant difference between the two treatment groups.
When there was no treatment switching, we assumed the actual OS median in the BSC alone would be shorter than 6.2 months since the panitumumab + BSC group had better progressionfree survival than the BSC alone (logrank test, pvalue < 0.0001). If the actual OS median in the BSC alone was 4.43 months (m_{1} = 4.43), the power to detect the difference between two treatment groups could attain 0.9 at the significance level of 0.01 with the sample size n= 232 and r = 1.
When the treatment switching was allowed, with the sample size n= 232 and r = 1, the power to detect the difference in OS were around 0.02–0.10 and 0.07 assuming s.dist = “gamma” and “indepExp”, respectively. That is, it was hard to detect the difference in OS between two treatment groups, which was consistent with the original result. To achieve a power of 0.9 at the significance level of 0.01, the sample size needs to be more than 2000 (Table 1). Here, we chose s.dist = “gamma” (i.e., s = XT_{1} and X follows a gamma distribution) since some patients didn’t experience disease progression prior to death. The parameters of shape and rate parameters in the gamma distribution were determined by pt and rho, where pt = 0.3 (=1.96/6.4). pt was estimated by the ratio of E(s) = 1.96 months (from the reported mean PFS time, 8.5 weeks) and E(T_{1}) = 6.4 months (= m_{1} /log (2)). We considered rho= 0.1, 0.3, 0.5, 0.7, and 0.9 to model low to high correlations between PFS and OS since rho was unknown. In addition, we chose s.dist = “indepExp” (i.e., s is independent of T_{1} and follows an exponential distribution), where E(s) = 0.3E(T_{1}). The censoring rate was set to be 0.02 (=4/232) since 4 censoring cases were reported in the 232 BSC alone patients.
Simulation scenarios
A number of simulation scenarios were generated to explore the effect of treatment switching on powers and sample sizes estimation in ITT analysis, which included the relative different effectiveness of the experimental treatment to the control, the different switching probability, the different types of switching time, and the deviation of the assumed distribution from the real distribution. In each scenario, the accrual time T_{a} was set to 3, the study time T_{e} was set to 5, and m_{1} was set to 1. Under no treatment switching, the expected power to detect the survival difference between two treatment groups was set as 0.8 and the significance level was set as 0.05. We also set r = 1, pt = 0.5, rho = 0.775, and censor.rate = 0.2. The number of computer simulations was set at 5000.
Effect of the relative effectiveness on power and sample sizes estimation
We first investigated the impact of the relative effectiveness of the experimental treatment to the control (m_{2}/m_{1}) on power and sample sizes estimation, which ranged at 1.5 and 2. The power decreased slightly and the ratio (n_{s}/n_{no − s}) of sample sizes with treatment switching (n_{s}) to those absence of switching (n_{no − s}) increased with the rising relative effectiveness (Table 2). For example, when m_{2}/m_{1}= 1.5 and p = 0.4, n_{s}/n_{no − s} = 1.65 (n_{no − s} = 130 and n_{s} = 215) if s.dist = “beta”. When the relative effectiveness grew from 1.5 to 2, the ratio n_{s}/n_{no − s} increased from 1.65 to 1.81, and the power decreased from 0.57 to 0.56. The changes in the ratios and powers in the four types of random switching time had similar trends. The change with s.dist = “indepExp” was the smallest, but larger sample sizes were required.
Effect of the switching probability on power and sample sizes estimation
We further evaluated the effect the switching probability (p) on power and sample sizes estimation, which was set at 0.2, 0.4, 0.6, 0.8, and 1.0. The power decreased and the ratio of sample sizes (n_{s}/n_{no − s}) increased significantly with the increased switching probability (Table 2). At m_{2}/m_{1}= 1.5, when the switching probability increased from 0.2 to 1.0, the ratio of sample sizes n_{s}/n_{no − s} increased from 1.28 to 4.82 if s.dist = “beta”, 1.27 to 6.02 if s.dist = “gamma”, 1.29 to 5.05 if s.dist = “unif”, and 1.30 to 7.21 if s.dist = “indepExp”. The power decreased from 0.70 to 0.24 if s.dist = “beta”, 0.70 to 0.23 if s.dist = “gamma”, 0.70 to 0.24 if s.dist = “unif”, and 0.69 to 0.18 if s.dist = “indepExp”. The changes of power and sample sizes were similar at m_{2}/m_{1}= 2. The results indicate that the switching probability is critical for power and sample sizes estimation. A small change of the switching probability would require much more sample sizes to maintain the expected power.
Effect of the switching time on power and sample sizes estimation
In addition, we explored the effect of the switching time on power and sample sizes estimation. We used the constant switching time (s) to make it easy to understand the impact. The switching time was set at 0.5/log (2) and 1/log (2), which were equivalent to pt = 0.5 and 1. The ratio of sample sizes (n_{s}/n_{no − s}) increased and power decreased significantly when the switching time advanced, especially when the switching probability p was greater than 0.6. For example, the ratio of sample sizes decreased from 1.91 to 4.83 when the switching time advanced from 1/log (2) to 0.5/log (2) under p = 1 and m_{2}/m_{1} = 1.5 (Table 3). The results indicate that the switching time is also a key factor for power reductions and sample sizes increase in clinical trials with treatment switching. More simulation results with different m_{2}/m_{1} and r were included in Supplementary Table s1 and s2. The change of the allocation ratio r had very small effect on sample sizes and power.
Power estimation when true distributions deviate from the assumed distributions
Finally, we performed simulation studies to investigate the robustness of power if the true distributions for the survival time deviate from the assumed exponential distributions. We first calculated the required sample sizes based on the survival time following the exponential distributions, given power = 0.8, m_{1} = 1, m_{2} = 1.5, r = 1, T_{a} = 3, T_{e} = 5, alpha = 0.05, pt = 0.5, rho = 0.775, and censor.rate = 0.2. Then, we assumed the true distributions of the control and the experimental groups follow Weibull distributions with the same median survival time as those in the assumed exponential distributions (m_{1} = 1 and m_{2} = 1.5). Thus, the shape and scale parameters in Weibull distributions had the relationship of scale = m (log(2))^{−1/shape}, where m was the median. When shape = 1, the Weibull distribution is reduced to the assumed exponential distribution. In contrast, the Weibull distribution deviates from the exponential distribution if shape ≠ 1 (Supplementary Figure s1). Table 4 listed power and the expected numbers of events under a low (p = 0.2) or high switching probability (p = 0.6). The shape of 0.5 and 0.75 led to power reductions (< 0.8), while power was improved when shape = 1.25 (> 0.8). To be noted, power was altered in a similar magnitude in both low and high switching probabilities. When the shape changed from 1 to 0.75, for example, power was reduced from 0.8 to 0.53 at a switching probability of 0.2, and from 0.8 to 0.54 at a switching probability of 0.6. In summary, there might be power reductions when the true distributions deviate from the assumed distribution, but the treatment switching seems not to cause additional power reductions. That is, there is no interaction effect between switching probabilities and distribution deviation.
Discussion
Our simulation study showed a standard ITT analysis failing to consider treatment switching resulted in a significant reduction of statistical powers especially when the switching probability is high and the switching time is early, which indicated that a much larger sample size was required to maintain the expected power. The sample sizes required in RCTs absence of treatment switching vary from around fourfifths to oneseventh of the sample sizes required in RCTs allowing treatment switching as the switching probability increases. The growth of the relative effectiveness and the switching probability increased the ratio of samples and decreased the power, while the later switching time decreased the ratio of sample sizes and increased the power. Moreover, the event rate in the control group was below the expected event rate (i.e., 1 – censor.rate), and decreased with the increased switching probability in trials with treatment switching. When m_{2}/m_{1} = 1.5 and censor.rate = 0.2, for example, the event rate E1/ n = 0.80 (104/130) under no switching, where E1 was the expected number of events in the control group (Table 2). In contrast, E1/ n = 130.8/166 = 0.79 if p = 0.2, E1/ n = 225/294 = 0.77 if p = 0.6, and E1/ n = 464.9/627 = 0.74 if p = 1.0 when s.dist = “beta” (Table 2).
PowerSwitchingTrial assumed the effects of the experimental treatment were the same for the participants switched from the control group to the experimental group and those initially assigned to the experimental group, which is the same as the “common treatment effect” assumption made by RPSFTM [8]. In some cases, the assumption may be problematic. The participants switched from the control group to the experimental group may be associated with worse survival. Properly adjusting the accelerated factor A may be useful to fit the scenario. Multiplying a constant that is less than 1 to A may be a solution, but it is difficult to determine the constant value before clinical trials even if we can borrow the information from previous similar studies.
The assumption of exponential distributions for survival time also limits the flexibility and practice of PowerSwitchingTrial. Survival times do not always follow exponential distributions well. Weibull or gamma distribution may be alternatives to exponential distributions. However, it is not easy for clinicians to determine a nonexponential distribution with at least two parameters. Median survival or mean survival is still commonly used, which leads to favoring exponential distributions.
The current version of PowerSwitchingTrial is built for twoarm RCTs. It can be extended to multiarm RCTs, straightforwardly but not simply. More parameters need to be determined in multiarm clinical trial designs. We plan to extend PowerSwitchingTrial for multiarm RCTs in the near future.
Conclusions
In this study, we proposed a simulationbased approach, PowerSwitchingTrial, to evaluate statistical powers and sample sizes required in RCTs allowing switching in ITT analysis when the logrank test is used. The approach not only provides a preview for power declining but also calculates the required sample size to achieve an expected power when treatment switching occurs. It will provide researchers and clinicians with useful information before RCTs are conducted.
Availability of data and materials
Additional file 1: Supplementary Material.
Abbreviations
 ITT:

Intentiontotreat
 RCT:

Randomized controlled trials
 RPSFT:

Rank preserving structural accelerated failure time models
References
Morden JP, Lambert PC, Latimer N, Abrams KR, Wailoo AJ. Assessing methods for dealing with treatment switching in randomised controlled trials: a simulation study. BMC Med Res Methodol. 2011;11:4.
Law MG, Kaldor JM. Survival analyses of randomized clinical trials adjusted for patients who switch treatments. Stat Med. 1996;15:2069–76.
Loeys T, Goetghebeur E. A causal proportional hazards estimator for the effect of treatment actually received in a randomized trial with allornothing compliance. Biometrics. 2003;59:100–5.
Robins JM, Tsiatis AA. Correcting for noncompliance in randomized trials using rank preserving structural failure time models. Communications in Statistics  Theory and Methods. 1991;20:2609–31.
Branson M, Whitehead J. Estimating a treatment effect in survival studies in which patients switch treatment. Stat Med. 2002;21:2449–63.
Zhang J, Chen C. Correcting treatment effect for treatment switching in randomized oncology trials with a modified iterative parametric estimation method. Stat Med. 2016;35:3690–703.
Latimer NR, Abrams KR, Lambert PC, Crowther MJ, Wailoo AJ, Morden JP, et al. Adjusting for treatment switching in randomised controlled trials – a simulation study and a simplified twostage method. Stat Methods Med Res. 2017;26:724–51.
Latimer NR, Abrams KR, Lambert PC, Morden JP, Crowther MJ. Assessing methods for dealing with treatment switching in clinical trials: a followup simulation study. Stat Methods Med Res. 2018;27:765–84.
Hsu CY, Chen CH, Hsu KN, Lu YH. A useful design utilizing the information fraction in a group sequential clinical trial with censored survival data. Biometrics. 2019;75:133–43.
Van Cutsem E, Peeters M, Siena S, Humblet Y, Hendlisz A, Neyns B, et al. Openlabel phase III trial of panitumumab plus best supportive care compared with best supportive care alone in patients with chemotherapyrefractory metastatic colorectal cancer. J Clin Oncol. 2007;25:1658–64.
Acknowledgements
Not applicable.
Funding
This work was supported by the National Institutes of Health [P30CA068485, U24CA163056, U24CA213274, P50CA236733, P50CA098131, U54CA163072].
Author information
Authors and Affiliations
Contributions
L.D. and CY.H. were major contributors in writing the manuscript and developing the R package. Y.S. was the major contributor in the conception and design of the work. All authors reviewed the manuscript. The author(s) read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
Deng, L., Hsu, CY. & Shyr, Y. Power and sample sizes estimation in clinical trials with treatment switching in intentiontotreat analysis: a simulation study. BMC Med Res Methodol 23, 49 (2023). https://doi.org/10.1186/s12874023018641
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s12874023018641