1Introduction
A passenger who presses an elevator hall-call button that is already lit is, from the controller’s point of view, doing nothing. Under collective control the first press creates a queue entry for that floor and direction, and later presses on the same button are discarded (Sorensen and Mbeki, 2021). The practice is nevertheless common: 61% of 2,114 surveyed commuters reported pressing a lit button at least occasionally (Aldous and Ferreira-Nkosi, 2018), and nothing limits the number of presses that one person may apply. We could not find a study that relates the number of presses either to a benefit or to the reaction of the people waiting nearby.
These are the two questions of a phase I dose-finding study: how large an exposure is tolerated, and whether the exposure does anything. The standard designs address the first. The 3+3 design escalates in cohorts of three according to fixed rules (Storer, 1989), the continual reassessment method (CRM) updates a one-parameter dose-toxicity model after each cohort (O'Quigley et al., 1990), and the Bayesian optimal interval design (BOIN) compares the observed toxicity rate at the current dose with two fixed boundaries (Liu and Yuan, 2015). Designs that use efficacy as well as toxicity exist (Thall and Cook, 2004), but they generally assume that efficacy is present at some dose and ask where the best trade-off lies.
If the exposure has no efficacy at any dose, a toxicity-only design still places most participants near the maximum tolerated dose (MTD), since nothing in the design asks whether the exposure achieves anything. We add to a CRM a futility rule on a continuous efficacy endpoint. The rule is a one-sided posterior probability on the slope of the endpoint against dose, and it has a closed-form boundary (Section 3.2). We evaluate the design by simulation against the 3+3 design and BOIN (Section 4) and report a trial in which the dose was the number of redundant presses (Section 5).
2Setting and endpoints
2.1Dose
An episode begins when a trained research assistant (the presser) stands within 1 m of a hall-call panel at which the call for the presser’s direction is latched and at least one other passenger is waiting within 4 m. The presser presses the lit button once, then waits 5 s, then applies further presses at 1 Hz, and then waits for the car. The dose is the number of redundant presses. There are dose levels, . Episodes in the same lobby were at least 10 minutes apart and no lobby hosted more than one episode on a given morning.
2.2Dose-limiting event
Bystander reactions were graded on a four-point scale adapted from Vasquez and Linde (2020). Grade 1 is a non-verbal reaction such as a look or an audible exhalation. Grade 2 is a remark to a companion. Grade 3 is a remark addressed to the presser or to the button. Grade 4 is a bystander touching or covering the button, or leaving the lobby by the stairs while remarking on the presses. A dose-limiting toxicity (DLT) is a reaction of grade 3 or higher by any bystander during the episode. Two raters coded audio transcripts from which the press sounds had been removed. On 120 transcripts coded by both, agreement on the DLT indicator was 94.2% (Cohen’s ).
2.3Efficacy endpoint
The efficacy hypothesis is that redundant presses bring the car sooner. The endpoint is the ETA error, the time at which a car door opened at the lobby minus the arrival time that the controller displayed when the presser pressed for the first time. Negative values mean that the car came earlier than predicted. The displayed estimate is computed from the dispatch queue and so does not respond to discarded presses. The planning value for the standard deviation of the ETA error, from the previous month’s controller logs, was s.
2.4Conduct and ethics
The trial ran on 11 weekday mornings in January and February 2026 in the lobbies of 11 office towers, with one cohort of three episodes per morning in a different tower. The posterior was updated and the next dose assigned after each cohort. The study was approved by the Oakmere University Research Ethics Committee (OU-REC-2026-0147) as observation and staged behaviour in a public place. Building managers consented, and signs at the entrances stated that research on lobby behaviour was in progress. Bystanders were not identifiable from the audio transcripts, and the presser could end any episode at any time. No episode was ended early.
3Design
3.1Toxicity model and allocation
Let be the probability of a DLT at level and the target. The MTD is the highest level with . We use the one-parameter power model
with skeleton . The prior variance of 1.34 follows Cheung (2011). After cohorts, let and be the number of episodes and DLTs at level and let be the posterior mean. The next cohort is treated at the level that minimises , subject to not exceeding the current level by more than one (Goodman et al., 1995). Cohorts have three episodes and the maximum sample size is 60, or 20 cohorts. All posterior calculations use a grid of 201 values of on .
3.2Futility rule
Let and be the dose and ETA error of episode , . We assume
with known, a flat prior on and with s per press. Efficacy corresponds to . Let , , and let be the least-squares slope. The posterior is with
Let s per press be the smallest slope of interest. After each cohort from the eighth onward, enrolment stops for futility if , with . Before the eighth cohort the rule is not applied, because the dose spread is small and the posterior probability can fall below by chance.
Proposition 1 (Futility boundary). Let . The rule stops enrolment if and only if
As the right-hand side tends to , and if the true slope is and , the probability of stopping at a single look tends to .
Proof. By Eq. (3), , which is below if and only if , that is, . Substituting and noting that gives Eq. (4). As , and the factor in front tends to one, so the boundary is . If then .∎
The boundary depends on the doses only through , so that the rule needs a spread of doses before it can stop. With and we have , and enrolment stops if s per press. With , which is the value at the end of the trial in Section 5, and the boundary is s per press. The probability of stopping when the true slope is exactly exceeds in practice because the rule is applied at up to thirteen looks (Table 3).
3.3Final selection
When enrolment ends, by futility or at 60 episodes, the MTD estimate is the level that minimises . If enrolment stopped for futility, no dose is recommended for further study whatever the MTD estimate.
4Operating characteristics
4.1Scenarios and comparators
We used five true dose-toxicity curves (Table 1). In each, the true MTD is the highest level with . The comparators were the 3+3 design in its standard form, in which a dose with 2 or more DLTs among 3 to 6 episodes ends escalation and the MTD is the next lower dose, and BOIN with target 0.25, boundaries and , cohorts of three, a maximum of 60 episodes, and the dose chosen by isotonic regression of the observed rates (Liu and Yuan, 2015). The CRM column uses the design of Section 3 without the futility rule. Each entry is based on 2,000 replicates. The 3+3 design has no fixed sample size, and its mean is reported.
| Scenario | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | MTD |
|---|---|---|---|---|---|---|---|---|---|
| S1 | 0.02 | 0.05 | 0.10 | 0.25 | 0.40 | 0.55 | 0.70 | 0.80 | 4 |
| S2 | 0.05 | 0.12 | 0.25 | 0.38 | 0.50 | 0.62 | 0.74 | 0.85 | 3 |
| S3 | 0.01 | 0.02 | 0.04 | 0.08 | 0.15 | 0.25 | 0.38 | 0.52 | 6 |
| S4 | 0.25 | 0.40 | 0.52 | 0.64 | 0.74 | 0.82 | 0.88 | 0.93 | 1 |
| S5 | 0.02 | 0.03 | 0.05 | 0.07 | 0.10 | 0.14 | 0.20 | 0.25 | 8 |
4.2Toxicity-only results
The CRM and BOIN selected the correct dose in similar proportions in S1 to S4, and both were well above the 3+3 design (Table 2). The CRM was ahead in S3 (66.7% against 52.5%) and BOIN was ahead in S5, where the MTD is the top level (38.9% against 27.8%). BOIN treated fewer episodes above the MTD in S1 to S3, and the CRM treated more at the MTD. The 3+3 design used fewer episodes, a mean of 7.7 to 24.7, but at the price of selecting the correct dose in 30% to 42% of replicates. In S4 the 3+3 design treated 35.0% of its episodes above the MTD.
| Scenario | Design | Correct (%) | At MTD (%) | Above MTD (%) | DLT (%) | Mean episodes |
|---|---|---|---|---|---|---|
| S1 | 3+3 | 35.8 | 21.3 | 14.9 | 17.1 | 16.9 |
| BOIN | 73.5 | 42.8 | 19.0 | 21.7 | 60.0 | |
| CRM | 73.8 | 49.0 | 23.3 | 24.1 | 60.0 | |
| S2 | 3+3 | 34.3 | 25.1 | 18.1 | 21.2 | 13.8 |
| BOIN | 67.3 | 43.0 | 21.6 | 22.9 | 60.0 | |
| CRM | 69.5 | 46.2 | 25.6 | 24.9 | 60.0 | |
| S3 | 3+3 | 29.9 | 13.0 | 9.2 | 13.1 | 22.3 |
| BOIN | 52.5 | 29.3 | 15.4 | 18.5 | 60.0 | |
| CRM | 66.7 | 40.7 | 15.8 | 20.4 | 60.0 | |
| S4 | 3+3 | 42.0 | 65.0 | 35.0 | 37.8 | 7.7 |
| BOIN | 90.6 | 75.5 | 24.5 | 29.1 | 60.0 | |
| CRM | 89.8 | 79.8 | 20.2 | 28.7 | 60.0 | |
| S5 | 3+3 | 31.4 | 7.4 | 0.0 | 10.4 | 24.7 |
| BOIN | 38.9 | 21.6 | 0.0 | 14.4 | 60.0 | |
| CRM | 27.8 | 14.7 | 0.0 | 14.8 | 60.0 |
4.3Futility rule
Table 3 gives the performance of the CRM with the futility rule for three of the scenarios and three true slopes. At the rule stopped 45.7% of trials in S1, 85.0% in S3 and 94.8% in S5. The proportion depends on the doses reached, as the boundary in Eq. (4) does through : when the MTD is low the trial explores a narrow range of doses and has less information about the slope in 60 episodes. At s per press the rule stopped at most 0.8% of trials, and the proportion of correct selections was within 0.2 percentage points of the CRM without the rule in S1 (73.6% against 73.8%). At the stopping probability was 10.7% to 15.1%, above the single-look value of 5% because of repeated looks.
| Scenario | True slope (s/press) | Stopped (%) | Mean episodes | Correct (%) |
|---|---|---|---|---|
| S1 | 0 | 45.7 | 46.1 | 62.5 |
| −0.5 | 10.7 | 56.8 | 73.5 | |
| −1.0 | 0.8 | 59.7 | 73.6 | |
| S3 | 0 | 85.0 | 32.6 | 47.6 |
| −0.5 | 10.9 | 56.9 | 64.0 | |
| −1.0 | 0.1 | 60.0 | 67.7 | |
| S5 | 0 | 94.8 | 29.5 | 20.5 |
| −0.5 | 15.1 | 56.0 | 28.0 | |
| −1.0 | 0.1 | 60.0 | 28.9 |
5Application
Enrolment began at level 1 (one redundant press). The first four cohorts, at levels 1 to 4, had no DLTs (Table 4 and Figure 1). The fifth cohort, at level 5 (6 presses), had one. The model then moved to level 6 (8 presses), where it stayed for the remaining six cohorts. Among the 18 episodes at 8 presses there were 5 DLTs, 27.8%. There were 6 DLTs in 33 episodes overall (18.2%), and none was of grade 4. The mean number of bystanders per episode was 2.6.
The futility rule was first applied after the eighth cohort. The posterior probability that the slope was steeper than s per press was 0.170 at cohort 8, 0.090 at cohort 9, 0.059 at cohort 10 and 0.018 at cohort 11, and enrolment stopped (Figure 1). At that point , the least-squares slope was 0.05 s per press, and a 95% interval using the residual standard deviation of 4.55 s was to . The boundary in Eq. (4) at this information was s per press, and the observed slope exceeded it.
The posterior mean DLT probabilities at the stop were 0.016, 0.040, 0.071, 0.115, 0.182, 0.270, 0.393 and 0.535 for levels 1 to 8. The estimated MTD was therefore level 6, 8 presses, with a posterior mean DLT probability of 0.27. Because the trial stopped for futility, no dose was recommended for further study.
| Level | Presses | Episodes | DLTs | Posterior DLT prob. | ETA error (s) |
|---|---|---|---|---|---|
| 1 | 1 | 3 | 0 | 0.016 | −1.4 (2.2) |
| 2 | 2 | 3 | 0 | 0.040 | 4.9 (7.2) |
| 3 | 3 | 3 | 0 | 0.071 | −0.5 (4.6) |
| 4 | 4 | 3 | 0 | 0.115 | 1.8 (5.0) |
| 5 | 6 | 3 | 1 | 0.182 | 4.7 (3.2) |
| 6 | 8 | 18 | 5 | 0.270 | 1.3 (4.3) |
| 7 | 12 | 0 | 0 | 0.393 | |
| 8 | 16 | 0 | 0 | 0.535 |
6Discussion
The futility rule did what it was designed to do. Of the 33 episodes, 15 were spent on escalation through levels 1 to 5 and 18 at 8 presses, the level preferred by the toxicity model. The slope of 0.05 s per press was far enough from that the posterior probability of an efficacy of at least fell below the threshold at cohort 11. The 95% interval for the slope, to , still contains , so the data do not exclude an efficacy as small as 0.5 s per press. The rule stopped because such an efficacy is unlikely given the prior and the data, and not because it was ruled out. The probability depends modestly on the prior and on . A flat prior on gives 0.019 at cohort 11 instead of 0.018, and using the residual standard deviation of 4.55 s in place of 4.0 gives 0.032. The rule would have stopped in both cases.
The MTD of 8 presses rests on 18 episodes at that level and on the skeleton for the other levels. The posterior DLT probabilities at levels 6 and 7 differ by 0.12, and the observed rate of 5 DLTs in 18 episodes has an exact 95% interval of 9.7% to 53.5%. The simulations show that the correct dose is selected in about two thirds of trials of this size when the MTD is in the middle of the range, and in about 28% of trials when it is the top level.
There are several limitations. The pressers were trained research assistants who applied presses at a fixed rate, and untrained passengers press in irregular bursts. The DLT definition depends on the grading of remarks, and a lexicon that is accepted in one city may not be in another. The ETA error is a sensitive efficacy endpoint only if the controller’s estimate is accurate. A design with two efficacy endpoints, or with the perceived wait rather than the measured one, would need a different rule. The trial was conducted in one cohort per morning, so that tower and day are confounded with dose assignment. We did not model this. We also did not study whether the toxicity depends on the number of bystanders, the time of day or the presence of a companion.
The proposed rule stops a trial when the data cannot support a slope of at least and gives no information about benefits at other levels of the endpoint. It could be used for any exposure whose tolerability is dose-limited and whose benefit is doubtful. In the trial reported here the toxicity model identified a dose and the efficacy model found no reason to use it.
Data availability
Episode-level doses, DLT indicators and ETA errors and the simulation code are available from the corresponding author. Audio recordings were not retained.
ACohort log
Table 5 lists the allocation and the futility statistics after each cohort. The least-squares slope and posterior probability are not defined at the first cohort because all three episodes were at one dose.
| Cohort | Presses | DLTs | Slope | Probability |
|---|---|---|---|---|
| 1 | 1 | 0 | ||
| 2 | 2 | 0 | 6.28 | 0.097 |
| 3 | 3 | 0 | 0.44 | 0.273 |
| 4 | 4 | 0 | 0.40 | 0.187 |
| 5 | 6 | 1 | 0.79 | 0.016 |
| 6 | 8 | 1 | −0.16 | 0.184 |
| 7 | 8 | 1 | −0.25 | 0.215 |
| 8 | 8 | 1 | −0.22 | 0.170 |
| 9 | 8 | 0 | −0.13 | 0.090 |
| 10 | 8 | 1 | −0.08 | 0.059 |
| 11 | 8 | 1 | 0.05 | 0.018 |