Challenges in Evaluating Technological Interventions for Affect Regulation

We evaluated whether a vibrotactile breathing pacer would influence two measures of affect during a cognitive stressor. In particular, we examined whether changes in breathing would be evident, and if so, whether these would mediate the effects of breathing pacer on self-report anxiety and skin conductance. Our results were surprising: although we observed the expected effects on breathing, we were unable to demonstrate that changes in breathing parameters were responsible for the observed changes in self-report anxiety. In addition, we failed to show that the pacer had an impact on skin conductance. In this paper, we investigate why we did not observe the expected effects. We believe our negative results have implications for evaluating technological interventions for affect regulation.

Because of these challenges, we have focused our work around the affect regulation strategy of slow-paced breathing, which has been robustly shown to reduce skin conductance and self-reported anxiety in various contexts either with or without a stressor [20], [21], [22], [23], [24], [25], [26], [27].In particular, we have built a vibrotactile-based breathing pacer and demonstrated a link between the vibration patterns that guides a user through slow-paced breathing and a decrease in self-reported measure of anxiety [17].
In this paper, we extended our prior work to determine (1) whether the pacer also reduced skin conductance and (2) whether breathing modulation, induced by the pacer, would mediate the relationship between the vibrotactile patterns and the two affective measures of skin conductance and selfreported anxiety.The context for this extension was an experimental paradigm in which a group of 100 college students engaged in a set of cognitively challenging stressor tasks during the first block, and then half of them were randomly assigned to an intervention condition with slow-paced breathing and half them were not.During this entire sequence we assessed their breathing and skin conductance and repeatedly asked about their self-reported anxiety.We expected a two-step casual relationship between the pacer and skin conductance, as well as a three-step causal relationship in which changes in breathing behavior, guided by the pacer, would explain reduced self-reported anxiety and skin conductance.
In the sections that follow, we describe the analyses we performed.We then discuss possible explanations for why the hypothesized effects were not evident.We conclude with recommendations regarding experimental design for measuring the effectiveness of affect regulation assisted by technology.To the best of our knowledge, this is the first paper that presents an in-depth examination of the mechanism that links a vibrotactile breathing pacer to affective measures of skin conductance, breathing, and self-reported anxiety.It is also the first paper that discusses the difficulties in designing experiments that measure the effectiveness of technology-based affect regulation interventions.
The specific contributions of this paper are as follows: We report the results of a mixed-design pre-registered experiment to evaluate the effect of a breathing pacer on psychophysiology measures.We interrogate the role of breathing behavior in changing affect via a set of repetitive patterns of vibrations produced by a breathing pacer.

HYPOTHESES
The confirmatory analyses are as follows: 1. Efficacy of pacer in reducing breathing rate.We hypothesized an interaction effect between group (treatment and control) and condition (Stressor 1 and Stressor 2) for the dependent variable of estimated breathing rate.We expected to observe no change from Stressor 1 to Stressor 2 in the control group, but a lower breathing rate in the treatment group during Stressor 2 when the pacer was active, which would indicate that participants were successful in following the pacer despite being cognitively engaged in performing stressful tasks.
2. Efficacy of pacer in reducing skin conductance.We hypothesized an interaction effect between group (treatment and control) and condition (Stressor 1 and Stressor 2) for the dependent variable of skin conductance.We expected to observe no change in the control group, but a lower skin conductance level in the treatment group, from Stressor 1 to Stressor 2. This is because we expected that slow-paced breathing deployed by the treatment group in the Stressor 2 condition when the pacer was active would be more effective in reducing skin conductance compared to other forms of affect regulation they may have deployed in Stressor 1.
Conditional upon observing the effect of the pacer on reducing breathing rate (confirmation of hypothesis 1), reducing skin conductance (confirmation of hypothesis 2), and previously observed reduction in self-reported anxiety, we had the following confirmatory mediational hypotheses.
3. Indirect effect of pacer in reducing skin conductance through breathing rate.We hypothesized that the breathing rate should mediate the impact of the pacer on changes in skin conductance measure from Stressor 1 to Stressor 2.
4. Indirect effect of pacer in reducing anxiety through breathing rate.We hypothesized that breathing rate should mediate the impact of the pacer on the change (in this case drop) in self-reported measure of anxiety (STAI-6) from Stressor 1 to Stressor 2.
The exploratory hypothesis was envisioned to test whether all participants in the treatment group equality benefited from the breathing manipulation.If not, we were interested whether the effects of interests (hypotheses 1-4) were evident in a subset of participants who were most effective in their breathing manipulation.

Short-and Long-Term Benefits of Slow-Paced Breathing
Slow-paced breathing within the range of 4 to 8.5 breaths per minute has been applied as a non-pharmacological method in stress management and biofeedback training.In general, during slow-paced breathing, inhalation temporarily suppresses parasympathetic activity and causes an immediate increase in heart rate, while exhalation decreases the heart rate and restores parasympathetic activity.This results in greater heart rate variability than the resting state.When an individual breathes at their resonance frequency [20] (which is unique for each individual and typically falls in the range from 4.5 to 7 breaths per minute), this effect is magnified significantly.Additionally, prolonged exhalations at this rate reduces physiological arousal quickly (within 60 seconds).In addition to immediate benefits, there have been long term benefits on cardiovascular and respiratory functions.For example, practicing slow-paced breathing for 20 minutes every day for a few months has been shown to decrease blood pressure and help with hyperventilation.In addition, as we age, our lung capacity and functioning decrease, and practicing slow deep breathing regularly can slow these gradual decreases.[20], [21], [22], [23], [24], [25], [26] Although it is recommended for an individual to practice slow-paced breathing at their particular resonance frequency for optimal benefits, both Leher et.al and Khazan's protocols [28], [29] require multiple sessions over a course of multiple weeks which render them impractical for an inlab study.For that reason, in PIV [30], we introduced an approximation protocol, individualized to each participant, to find a breathing rate close to their resonance frequency that reduced skin conductance and increased heart rate variability.In this study, we use this protocol to tune our pacer to guide the treatment group through slow-paced breathing.

Vibrotactile Technologies for Affect Regulation
Compared to other studies evaluating vibrotactile technologies for affect regulation, this study measured a greater range of psychophysiological indicators, which allowed us to assess more comprehensively the mechanism of the breathing pacer on affect change.Besides our pacer, several other devices have been studied that use vibrotactile technology for affect regulation.These include 2 : Doppel [12], BoostMeUp [15], and ambienBeat [18], all three of which provide heartbeat-like stimulation on the wrist; Spire Stone [14], which monitors users' respiration and delivers interventions through an app; and Just Breathe [10], Breath Booster [31], Calm Commute [32] a series of in-car haptic and audio interventions for use while driving.Of the studies used to evaluate the other vibrotactile technologies mentioned, only two analyzed skin conductance as a measure of anxiety (Doppel, Just Breathe), and one (Just Breathe) failed to show a reliable change in skin conductance.In addition, only two analyzed respiratory data in any form (Just Breathe and Calm Commute) and one (Calm Commute) used measures of Heart Rate Variability to justify a calming effect . 3 Furthermore, none of these studies analyzed breathing modulation in terms of breathing regularity, percentage of breathing modulation, or chest-to-abdomen displacement ratio, all of which are known to have a link to change in affect [29], [33].Lastly, none of these studies explored a three-step causal relationship (vibration pattern affect regulation strategy affect).Our study used the measured skin conductance and breathing data, and also derived measures of breathing behavior (e.g., breathing rate, irregularity, and the extent to which treatment group followed the pacer).We were able to analyze the effect of using the breathing pacer on each measure and also develop and analyze an explanatory hypothesis about the affect regulation mechanism of the breathing pacer by testing for breathing mediation effect to explain the pacer's calming effect.

METHODS
In this section, we briefly describe the study design.An extensive explanation can be found in [17].

Participants
Participants were recruited via a university pool of Stanford undergraduate students with the incentive of either course credit or hourly payment for their time.Out of the total of 100 volunteers, 3 had to be excluded due to procedural errors.44 participants (29 female, 15 male) were randomly assigned to the treatment group and 53 (32 female, 21 male) to the control group.The IRB review and approval prior to and consents from all the participants during the study were obtained.

Protocol
Before starting the study, participants, with the help of research assistants, placed a few devices on different locations of their body.Based on the recommendations from [30], each participant had two tactors taped to their abdomen.They were also fitted with breathing gauges around the chest and abdomen and electrodermal activity sensors placed on the index and ring fingers of their non-dominant hand.Participants were given noise-cancelling headphones to aid in blocking out any potential vibration sounds. 4 The study consisted of the following stages: Baseline, Meditation, Personalization, Vibrotactile Breathing Practice, Stressor Task Practice, Block 1, and Block 2 (see Fig. 1).All participants followed the same procedure until the end of Block 1, at which point the treatment group received the breathing pacer intervention while the control group did not.Below, we summarize the stages of the protocol and describe the stressor task used in Blocks 1 and 2. For more detailed information about the protocol and the stressor task, see the Methods section of Miri et al. [17].
In the Baseline stage, participants watched a 5-minute video of a pipe changing colors and were asked to count how many times the pipe turned green; the objective of this stage was to obtain baseline physiology measurements.In the Meditation stage, participants listened to a 5-minute audio guiding them through slow-paced abdominal breathing techniques.The objective of this stage was twofold: to train participants to prepare them for synchronous breathing with the vibrations, and to estimate their slow-paced respiratory rate in breaths per minute.This is the personalized respiratory rate that was used to guide the users through vibrotacticle slow-paced breathing for the rest of the study.
In the Personalization stage, we established personalized frequencies and amplitude of vibrations for each participant using a staircase method.We determined two frequencies of vibrations to allow participants to distinguish between the two phases of inhalation and exhalation.Typically, these two frequencies were approximately 30 to 40 Hz apart to allow them to be easily distinguished from each other.Research assistants worked with participants to arrive at these personalized frequencies and amplitudes.The details of the personalization protocol can be found in [30].
In the Vibrotactile Breathing (V-Breathing) Practice stage, participants practiced synchronizing their breathing with the personalized vibrations for 90 seconds.This stage allowed the research assistants to determine whether the choices for pace, amplitude, and frequencies allowed the 2. For a comprehensive review of vibrotactile devices for affect regulation, see Background sections in [17], [18].In this paper we are only focused on those devices whose effect on psychophysical measures have been reported.
3. Heart Rate Variability (HRV) and breathing rate are markers for two different organ systems, namely the cardiovascular system and the respiratory system.But, the markers are linked: short-term increase in HRV occurs while breathing near the resonant frequency.Hence, we believe that HRV should be used as a manipulation check on slowpaced breathing rather than as a measure of affective outcome.
4. C2 tactors used in this study can make audible noise in a quiet lab environment when vibrating at frequencies over 165Hz [34] without noise canceling headphones.In the personalization procedure, C2 tactors were not driven beyond 180 Hz.Beyond this frequency, C2 tactors vibrations become audible even when noise canceling headphones are in use.
participant to breathe slowly while effortlessly in sync with the vibrations.If not, then the research assistant would adjust these values so that breathing was effortless.
In the Stressor Task Practice stage, participants were introduced to the compound remote associate (CRA) task to be used as a stressor.Compound remote associate questions consist of three English words whose answer is a fourth word that is associated with all of them.For example, the answer to the CRA question "river/note/account" would be "bank."The specific CRA questions were taken from Bowden & Jung-Beeman [35].Block 1 consisted of the first stressor and two waiting periods occurring before and after the stressor.Waiting periods, called Pre-stressor 1 and Poststressor 1, allowed two minutes to pass between stressors so that any anxiety from the stressor would not affect measurements taken at other stages.The stressor task was 4 minutes long and consisted of 27 CRA questions with a time limit of 9 seconds each.Running out of time on a question was equivalent to an incorrect answer.If a question was answered correctly, participants would still have to wait the full 9 seconds before seeing the next question; this was done to ensure that all participants spent an equal amount of time in the stressor task.During Block 2, which was otherwise identical to Block 1, the treatment group received ongoing personalized vibrations and were instructed to conform their breathing with the sensation of the pacer.

MEASURES
We used Qualtrics to collect self-reported measures and FlexComp System with BioGraph Infiniti Software T7555 M by Thought Technology to acquire all physiology measures in this study.

Respiration Measures
Our primary respiratory measure was breathing rate (number of breaths per minute).In addition to this primary outcome, to determine compliance with the pacer, we calculated active breathing percentage which reflects the extent to which the treatment group followed the pacer in a response window.For exploratory purposes, we also computed additional breathing indices including Breathing irregularity (measured through coefficient of variance) and chest to abdomen displacement ratio (slope of a regression line fitted over a Konno-Mead plot [36], [37]).

Skin Conductance
Skin conductance (SC) is a measure of the degree to which the skin transmits a small electric current between two electrodes, and it reflects sympathetic arousal (SA: i.e., activation of the sympathetic nervous system), which is associated with the so-called fight-or-flight response, as the body anticipates and prepares for action, either mental or physical.

Self-Reported Anxiety
To measure anxiety, we collected responses to the State-Trait Anxiety Inventory (STAI-6) after Baseline, V-Breathing Practice, Pre-stressor 1, Post-stressor 1, and Post-stressor 2 (indicated by a color-coded pen-and-paper icon in Fig. 1).We chose STAI-6 -the brief six-item version of the State-Trait Anxiety Inventory -over the more commonly used STAI-20 because it reduced the burden on the participant.In addition, STAI-6 scores correlate highly with scores on the full-length STAI-20 (r ¼ 0:95) [38].The results of this measure were previously reported in [17].

Self-Reported Regulation
At the end of the study, we asked open-ended questions about the type of affect regulation participants used during block 1 and block 2.

DATA PROCESSING 6.1 Respiratory Data
Before extracting respiratory measurements of interest (mentioned above), we used an empirical mode decomposition (EMD) based signal processing technique, inspired by Liu et al. [39], to remove tissue artifacts (e.g., any muscle movement), posture change, and background noise from respiratory signals.In addition, we applied a smoothing technique using a rolling window of 2  3 Â sampling frequency (in our case 256 Hz).Below, we describe in more detail our approach to artifact removal.At that point, participants who were randomly assigned to the control group went through Block 2 while receiving no vibrations (indicated by a solid black line), while participants who were randomly assigned to the treatment group go through Block 2 while receiving vibrations.Subjective measures of anxiety (i.e., STAI-6) were collected at five conditions of Baseline, V-Breathing Practice, Pre-stressor 1, Post-stressor 1, and Post-stressor 2 (indicated by a color-coded penand-paper icon).Physiological (i.e., skin conductance and respiration) measures were collected during the entire procedure.[17].

Artifact Removal of Respiration Signal
Empirical Mode Decomposition (EMD) [40] decomposes a time-varying signal into a sum of finite components, referred to as intrinsic mode functions (IMFs), each of which represents an oscillatory mode (see Fig. 2).A critical step in achieving an artifact-free signal is selecting the appropriate IMFs that contribute to the signal and excluding those IMFs that mainly contribute to the contamination of the signal.In the context of a respiration signal derived from ribcage or abdomen displacement, tissue artifacts (e.g., any muscle movement unrelated to breathing), and background noise are usually present in higher frequency bands (i.e., several first IMFs) while lower frequency bands (i.e., several last IMFs) often correspond to artifacts resulting from posture changes.By removing these high and low frequency bands corresponding to such artifacts, an artifact-free respiratory signal can be reconstructed by summing up the remaining IMFs.
We investigated different applications of EMD for removing artifacts ( [39], [41], [42]), each applied to differing ways the breathing signal was captured.We found none of them worked well with our signal.Instead, we found that removing the first two IMFs (which always corresponded to tissue artifacts) and the last IMF (which always corresponded to the participant's posture change).This resulted in less artifact removal, but enough to estimate breathing rate using both peak finding and Fast Fourier Transform (FFT) approaches.We found that the removal of the lowest computed IMF was more important than the first two highest computed IMFs in measuring breathing rates, as described next.

Breathing Rate
For the purposes of confirmatory analyses, we chose a simple and robust approach for measuring breathing rates.
After removing artifacs using the generic EMD approach described above, we identified all peaks and troughs 5 We then computed a set of breathing rate averages using a sliding window of 5 breaths, and returned the median of this set as the dominant breathing frequency.The window size of 5 was chosen because we did not find a significant difference between what this algorithm computes versus what an FFT-based approach, described in Appendix A, which can be found on the Computer Society Digital Library at http:// doi.ieeecomputersociety.org/10.1109/TAFFC.2022.3175687,computed for segments over which it returned a valid result.This appendix section also explains why we did not use the FFT-based approach.

Degree to Which the Treatment Group
Followed the Pacer For the purposes of exploratory analyses which were designed to assess whether participants in the treatment group varied in benefits they enjoyed from the pacer, we computed a measure of how well that participant followed the pacer.We computed the set of breathing rate averages using a sliding window of 5 breaths, and partitioned this set into two, using the cutoff value of 14 breaths per minute to distinguish passive from active breathing rates [29].Then, we computed the ratio of active breathing rate averages to the total number of averages in the original set.Thus, those who followed the pacer fully -and were thus actively breathing -were assigned a measure of 1, those who 5. To extract the peak-related information, we used Python find_peaks with the following settings: find_peaks(signal, distance = 2 Â sampling_frequency, width =. 8 Â sampling_frequency, prominence =. 3).
ignored the pacer were assigned 0, and those remaining were assigned a value between 0 and 1.Then, based on visual inspection of Fig. 4, we further classified a participant with an active breathing ratio > 0.85 as an active participant, less than 0.4 as passive, and the remaining as inbetween participants.This resulted in 17 participants in active, 16 in passive, and 11 in in-between groups.The inbetween group members could not fully shift their breathing to match the pacer rate at all times and so their attempts led to a greater breathing inconsistency compared to the passive and active groups.In contrast to the slow, deep, and diaphragmal breaths associated with proper slow paced breathing chemistry, inconsistent breathing can be linked to negative affective states such as anxiety [29].Because of this, in some of our exploratory analysis, we excluded the in-between group participants to focus on testing the distinction between passive, active, and control groups who all exhibit approximately consistent breathing patterns.Fig. 3 shows examples of participants belonging to each of these categories.Note that in addition to the difference in breathing rate, active breathing also resulted in a larger breathing wave amplitude as compared to passive breathing.
We believe that whether participants followed the pacer or not was not as important as whether they managed to slow down their breathing sufficiently enough to have a beneficial effect.When we did directly compare the pacer phase with the breathing wave, we observed misalignments during the stressor phases.For example, we would frequently see a participant take a deep slow breath when the pacer was more than half way through the inhalation phase.We think it was important to count this slow breath as an attempt to follow the pacer and so a model of how much participants aligned their breathing with the pacer was not contextually appropriate to use.

Efficacy of the Pacer in Reducing
Breathing Irregularity For the purposes of exploratory analysis we measured the irregularity in breathing depth, which is the coefficient of variance (CV) of the respiration signals.This measure is known as the relevant standard deviation expressed as a percentage of a signal mean (i.e., SD Mean Â 100).A higher CV corresponds to more irregularity in the depth of breathing.

Efficacy of the Pacer in Reducing Chest to Abdomen Displacement Ratio
For the purposes of exploratory analysis, we examined ribcage-abdominal motion using the method of Konno and Mead [36], [37], where the extensions of the abdomen and rib cage are displayed on an XY plot.There is no axis corresponding to time on a Konno-Mead plot; instead, the breathing waves together generate a trace on the Cartesian plane.
When the chest and abdomen move relatively in synchronization, the trace is a figure that slopes upwards and to the right.We then fit in an ordinary least squares (OLS) regression line which is a measure of the ratio of abdomen to chest motion: slopes less than 1 result from more abdominal compared to chest breathing.Appendix C, available in the online supplemental material explores this approach in more detail.

Skin Conductance Data
For the purposes of confirmatory analyses, we used two general approaches for skin conductance analysis: a modelbased approach (called CDA, for continuous decomposition analysis) which uses mathematical models to map between observed SC and SA, and a model-free approach, such as the average of an observed SC within a certain time frame.Ledalab [43] is a widely used package that infers SA from observed SC.We used two model-based measures produced by Ledalab: CDA.nSCR, which is the total number of significant peaks in the phasic (fast-changing) drive, and CDA.SCR, which is the average phasic driver value in a response window. 6We also used the model-free measure SC.mean, which is the averaged observed skin conductance over a response window.See the Appendix B, available in the online supplemental material for details on how the LedaLab package computes CDA.nSCR and CDA.SCR.

DATA ANALYSES
To test the confirmatory analyses of (1) efficacy of pacer in reducing breathing rate and (2) efficacy of pacer in reducing skin conductance, we ran linear mixed models to test the interaction effect between group (treatment and control) and condition (Stressor 1 and Stressor 2) on the dependent measures of interest while controlling for stable individual differences.Models were implemented in R with the lme4 package according to the following specifications: lmer(DV $ condition * group + (1 | id)).The effect confidence intervals were bootstrapped.
To test the confirmatory analyses of (3) indirect effect of pacer in reducing skin conductance through breathing rate and (4) indirect effect of pacer in reducing anxiety through breathing rate, we used the mediate function of the mediation package in R (see AppendixG, available in the online supplemental material for sample code).To assess the impact of compliance with our instructions in the treatment group, we quantified the degree to which the treatment group adhered to the pacer and classified the treatment participants into three groups of active (largely followed the pacer), passive (largely ignored the pacer), and in between (which occasionally followed the pacer: we excluded this group from further analysis).We then tested for the interaction effect between group (passive, active, and control) and condition (Stressor 1 and Stressor 2) for the two affect measures of skin conductance and self-reported anxiety.
In addition to these four confirmatory analyses and one exploratory analysis, we also conducted some additional analyses to better understand the patterns of findings that emerged from these analyses.

RESULTS
During the personalization procedure, we tailored the vibrotactile breathing patterns to the needs of each individual participant.The slow breathing pace was in the range of 5.5 to 9 with median of 7 and mean of 7.15.Below we discuss the details of our findings.

Efficacy of the Pacer in Reducing
Breathing Rate In the first confirmatory analysis, we ran linear mixed models on the measure of breathing rate and found an interaction effect between the condition (Stressor 1 and Stressor 2) and group (treatment and control).Consistent with our hypothesis, the treatment group's breathing rate dropped from Stressor 1 to Stressor 2 as a result of the pacer intervention (see Fig. 4, left, and Fig. 5).The observable large effect (intercept = 17.02, b = 3.27, p < .0001, CI = ½2:18; 4:36) suggests that on average, participants lowered their breathing rate.Some participants were more successful than others in implementing slow-paced breathing throughout the Stressor 2 window.The subsection below describes follow up exploratory analyses on this in more details.

Exploratory Analyses: Efficacy of the Pacer in Increasing Ratio of Active Breathing
To quantify how successful treatment group participants were in following the pacer, we derived a measure of active breathing percentage, as explained in Section 6.1.3.Fig. 4, right, shows the ratio of active breathing in two conditions of Stressor 1 and Stressor 2. During Stressor 1, the participants (treatment and control groups combined) performed an average of 19% of active breathing. 7During Stressor 2, the proportion of active breathing percentage increased to approximately 57% for the treatment group when the pacer was active, but did not significantly change for the control group.While the average breathing rate in the treatment group dropped significantly during Stressor 2, the majority of the participants were not able to breathe consistently at a rate slow enough to be classified as following the pacer.Instead, we observed a significant variability in the breathing rates, suggesting that participants were more likely to switch back and forth between passive breathing when solving a problem and active breathing when they were able to attend to the pacer.This pattern of breathing resulted in more irregularity in breathing data, and every time a switch between passive to active breathing occurred, more chest breathing became involved. 8

Exploratory Analyses: Were Participants'
Performance Compromised Due to How Much They Engaged With the Pacer?
Stressors and stressful situations come in various types, ranging from low to high cognitive demand, as well as anticipatory, in-the-moment, or ruminating stress.In the case of the CRA task, it is highly cognitively demanding, inthe-moment stressor.We chose the CRA task because of its ecological validity.People often become anxious while engaging in cognitively demanding activities.In addition, the CRA task was well suited to our mixed design experiment.Our mixed design required two stressor blocks, and so the stressor must be repeatable.Lastly, the CRA tasks don't have a learning effect: one doesn't get better at answering the CRA-generated questions.
6.For replicability purposes, here is a sample invocation of the Ledalab script with a skin conductance sampling rate of 256 Hz: Ledalab('path_to_data', 'open', 'text', 'downsample', +2, 'analyze', 'CDA', 'optimize', 2, 'export_era', [6 24005 1]).The response window is 6 to 240 seconds, and the threshold for SCR peaks is.05 microsiemens.7.That is, of the breaths taken in this condition, 19% were at a rate slow enough to be considered to be active breathing.Because of the rolling average used to compute the breathing rates, the first 4 breaths were not included in this percentage.
8. See Appendix C, available in the online supplemental material for analysis of pacer effect in modulating breathing behavior in terms of breathing irregularity (coefficient of variance), and ratio of abdomen to chest displacement (slope of a regression line fitted over Konno-Mead plot).
Our findings suggest that the pacer was usable with the CRA stressor.In particular, the degree of following the pacer did not affect CRA performance.For instance, the CRA scores of the passive group did not differ significantly from the average score (intercept = 10.13,b = -0.73,p =. 50, CI = ½À2:85; À1:48).Likewise, the active group didn't have a score any higher or lower than the average score (intercept = 10.47 , b = -1.18,p =. 24, CI = ½À3:17; :80).Therefore, following or ignoring the pacer did not result in boosted or reduced CRA task performance.

Efficacy of the Pacer in Reducing Skin Conductance
In the second confirmatory analysis, we ran linear mixed models on three measures of skin conductance: CDA.SCR, CDA.nSCR, and CS.mean.Although we only pre-registered our hypothesis for the measure CDA.SCR, when we did not observe the expected interaction effect, we used the other two measures to ensure that none of the three measures showed the hypothesized effect.We did not find an observable interaction effect between group (treatment and control  6, center and right).That is, we did not find a significant drop in skin conductance (reflecting a drop in sympathetic arousal) as a result of the breathing pacer intervention using either model-based or model-free measures of skin conductance.The treatment and the control groups are indistinguishable in terms of skin conductance measures in Fig. 6).Contrary to our hypothesis, the results suggest that although the treatment group participants attempted to use slow-paced breathing, doing so was not any more or less effective than whatever affect regulation method they may have deployed during the Stressor 1 condition during which there were no pacer vibrations (within-subject comparison) or whatever the control group participants deployed (between-subject comparison).

Indirect Effect of Pacer in Reducing Skin
Conductance Through Breathing Rate In the third confirmatory analysis, because the two-step causal link between the pacer and skin conductance was not observed (see subsection above), we were unable to show that there exists a three-step link between these two measures mediated through breathing rate.

Indirect Effect of Pacer in Reducing Anxiety
Through Breathing Rate In the fourth confirmatory analysis, we failed to find the mediation effect of the breathing rate to explain the link  between the pacer and self-reported anxiety.We previously found an interaction effect between group (treatment and control) and condition (Stressor 1 and Stressor 2) for STAI-6, which suggested a calming effect of the pacer in reducing stress (see [17], Results and Discussion section).We tested whether this effect can be explained by the breathing rate, and did not find the average causal mediation effects (ACME) =. 03, CI = ½À1:69; 1:71, p =. 97; ADE = À2:45, CI = ½À5:56; :5, p =. 09; total effect = À2:41, CI = ½À4:83; À:10, p =. 04).

Did Other Measures of Breathing Modulation
Cause a Drop in Self-Reported Anxiety?
We have shown that the pacer modulates breathing behavior and that the breathing behavior was causally linked to self-reported anxiety.But we have failed to show that the pacer produced a reliable change in self-reported anxiety through breathing modulation.This has led us to further test whether the other measures of breathing we collected (see Appendix C, available in the online supplemental material, i.e., breathing regularity, chest to abdomen breathing ratio, and active breathing ratio) could explain the pacer's effect as a mediator.Unfortunately, none of them showed a mediation effect to explain the drop in selfreported anxiety (STAI-6).This suggests that identifying more measures of breathing behavior may not be a fruitful direction to pursue.Instead, we may benefit more by focusing on improving methods for collecting self-reported anxiety measures.

Efficacy of the Pacer in Reducing Anxiety and Skin Conductance in Subset of Treatment Group Participants
In the exploratory analysis, after observing that not all participants in the treatment group equality benefited from the breathing manipulation (see section 8.1.1),we tested whether the effects of the pacer was evident in a subset of participants who were most effective in their breathing manipulation.Fig. 7 (Left) shows that participants in both active and passive groups both reported benefits from the pacer as compared to the control group.However, the control, active, and passive groups were indistinguishable in terms of skin conductance during Stressor 2 Fig. 7 (Right).

Summary of Results
In this study, we found that the pacer had the intended impact on breathing but contrary to expectations, there were no effects of the paced breathing on skin conductance and breathing modulation did not mediate the effect of the pacer on either skin conductance and self-reported anxiety (reported separately).In addition, we failed to show that the effects of the pacer was evident in a subset of participants who were most effective in their breathing manipulation.In the following sections, we will consider why we might have failed to observe the expected effects and attempt to draw some conclusions and provide suggestions about future works in this area.

Why Did We Not Observe the Effect of Pacer on Skin Conductance?
In an attempt to understand why the control, active, and passive groups were indistinguishable in terms of skin conductance, we first checked that the skin conductance measurement and the experimental design were meaningful (see Appendix E, available in the online supplemental material for more details).We then investigated several explanations based on the data we collected.Below, we describe two explanations that we believe are the most and one that we think is the least plausible.In Appendix F, available in the online supplemental material we listed other explanations that could be possible but are less data-driven as compared to the explanations presented below.1.The control group was not passively experiencing stress: They were spontaneously regulating.
One possible explanation of our failure to find reductions in skin conductance during Stressor 2 is that the affect regulation spontaneously deployed by the control group was as effective as the slow-paced breathing deployed by the treatment group, thus concealing the hypothesized effect of slow-paced breathing on skin conductance.
The possibility that the control group was spontaneously engaging in affect regulation was supported by their selfreport of the affect regulation they deployed during Prestressor 2 (a stress anticipation window prior to stressor tasks) and Stressor 2 (when they were performing the stressor task).We grouped the control participants into those who didn't report using an affect regulation strategy in Pre-Stressor 2 and Stressor 2, and those who did use some affect regulation strategy.For some participants, it was unclear what affect regulation strategy they deployed.For example, participants s076 reported "thinking of words that I know".We included such participants in the group that used some affect regulation strategy.We identified five participants (10%) in Pre-stressor 2 and 9 in Stressor 2 (18%) who reported using no affect regulation strategy.Therefore, most the control group did not report passively experiencing stress.
2. The active group followed the pacer at a cost: Cognitive load induced by synchronizing breathing with the pacer The second possible explanation as to why the treatment group may not have shown the reduction to skin conductance is because skin conductance is sensitive not only to affective responses but also to cognitive load.The reason this is relevant is that the participants in the treatment group had two tasks, whereas the control participants had only one.More specifically, the treatment group both had to perform the CRA tasks and had to follow to the pacer, while the control group only had to perform the CRA tasts.On this account, even though some of the participants showed reduced anxiety and reduced breathing rate, they may have shown equivalent skin conductance responses to the control group because of the greater effort that was expended trying to conform to the demands of the CRA tasks as well as to the pacer.We grouped the treatment participants into two: those who mentioned something about needed to context switch between breathing and stressor tasks and those who did not.Fourteen participants (33%) were in the first group, illustrating the difficulty of following the pacer while engaged in the CRA task.
3.Not observing the impact of slow-paced breathing on skin conductance does not imply lack of its existence We believe that not observing a difference between active and passive groups in terms of skin conductance does not contradict the well-known fact that slow-paced breathing has an effect on skin conductance.We have found evidence that the pacer modulated the average breathing rate, and we observed that breathing behavior was unidirectionally linked to temporal changes in skin conductance.Specifically, for each participant, regardless of group membership (Treatment or Control) or condition (pre-stressors or stressors), we observed a unidirectional relationship between the time series of their breathing signal (chest or abdominal displacement) and their correspondent skin conductance.We made this observation using Convergent Cross Mapping (CCM) [44], which is a method that distinguishes causality from correlation in nonseparable weakly connected dynamic systems.We tested this causal relationship and have shown that, indeed, breathing caused a change in skin conductance, and not the other way around.See Appendix D, available in the online supplemental material for more details on CCM analysis.

Why Did We Not Observe the Indirect Effect of
Pacer in Reducing Anxiety Through Breathing Rate?One plausible explanation for not observing the mediation effect is because the self-reported anxiety measures were not collected immediately after the stressors sections.Instead, they were collected at the end of the 2-minute recovery time (Post-stressor 1 and Post-stressor 2).Observe in Fig. 6, right, that during the recovery time, the level of sympathetic arousal had significantly dropped for both the treatment and control groups, but there was no significant difference between the two in the level of this drop.Therefore, it is possible that the retrospective measure of STAI didn't capture the nature of anxiety that could have been captured had the questionnaires been administered immediately after the stressor conditions.Perhaps that is why we didn't observe a larger interaction effect and we failed to observe the mediation effect through breathing rate.Another plausible explanation is that the self-reported anxiety measures were not collected as a time-series measure.In future studies, inspired by Xiao Yang et.al study [45], participants could be asked to rewatch video recordings of their Stressor 1 and 2 segments (each about 4 minutes long) and use a joystick to rate their anxiety level in a continuous manner.A casual relationship between the two contentious measures of breathing and the self-reported measure of anxiety during Stressor 2 could then be computed using Convergent Cross Mapping technique [44].Should a casual relationship be found, the anxiety measure could be collapsed to a single data point and test for the mediation effect.Independent of the issue of the type of stressor, is practice using the pacer with the stressor important?In general, practice using a new technology is often important in being able to effectively use that technology.How much is needed?How persistent is the effect of practice'for example, would a third stressor block have shown a reduction in skin conductance by the treatment group?Sufficient practice may allow participants to master the act of matching their breathing with the pacer without using up cognitive resources required by the stressor task.These questions suggest a 2 Â 3 Â 2 table with stressor type (low versus high cognitive load), situation (in the moment versus, anticipatory, or after) as one axis and the experience level (novice versus expert) as the other axis.This study addresses the combination of novice users of the pacer experiencing an in-the moment, cognitively demanding stressor.In the future, it would be useful to sample from other combinations as well.

CONCLUSION AND FUTURE WORK
Our findings indicate that the breathing pacer led to expected change in breathing behavior.However, despite the documented efficacy of this pacer in decreasing self reported anxiety [17], and its success in modulating breathing, it did not produce reliable changes in skin conductance.We also failed to show that changes in self-reported anxiety were mediated by changes in breathing parameters.
We expected to find the pacer led to reductions in skin conductance, especially since the link between slow-paced breathing and reduced skin conductance is well established.Not finding the expected effect was surprising: we separately showed that the pacer modulates breathing behavior and the breathing behavior unidirectionally, on a causal account, was linked to skin conductance.The experimental method was well designed and similar to previous studies [10], [12], [15] that used affective measures of skin conductance and self-reported anxiety to determine efficacy of a technological intervention.In addition, the experiment was adequately powered [46] to detect a medium size effect 9 .Instead, we observed that no matter how effective the pacer was in the absence of a stressor, or whether its usage began prior to engagement with stressor tasks, or whether we focused only on the active group, the psychophysiological calming effect of the pacer was indistinguishable from any other forms of affect regulation used by the control group.
These findings point to a puzzling situation: one measure of affect (skin conductance) suggested no differences between the group who fully modulated their breathing with the pacer, the group who didn't modulate their breathing with the pacer, and the group who neither received nor modulated their breathing with the pacer.The other affective measure (self-reported anxiety), on the other hand, showed differences between those who experienced the pacer (regardless of whether they modulated their breathing or not) and those who didn't.See Fig. 7. Aggregating these two observations strongly suggests a placebo effect of the pacer for both active and passive groups.
However, such a conclusion strongly contradicts the following two findings: the pacer did modulate breathing and the breathing behavior, using a time-series analysis, was unidirectionally linked to skin conductance.The qualitative data on use of affect regulation during various stages of the study shed light into this paradoxical situation to suggest that presence of the spontaneous regulation in the control group and cognitive load to comply with the pacer in the treatment group could have contributed to not observing the effects of the interest.
Confronting our perplexing results, we have identified a set of considerations for future research in evaluating technology that assists in affect regulation.One consideration is to include the stressor anticipation phase in addition to the stressor phases.The difference in cognitive load between these two allow us to detect any masking effect the stressor could have on the affective measure of interest.Second, to include a collective of multiple measures of affect, in case one measure is masked by the cognitive load of the stressor choice, others could facilitate understanding the regulatory mechanism of the intervention.Another consideration is to collect qualitative data on the type of deployed affect regulation and collect measures that can verify them.Lastly, one should aim to understand the causal links between the technological interventions and affect in the context of the interaction among stressor types (e.g., low versus high cognitive load), situation (in the moment versus, anticipatory, or after), and user expertise (novice versus expert).Understanding how other technologies perform in this interaction space is important as well.More generally, further systematic research is needed to better understand the conditions under which a technological intervention will have impacts on an affective measure.

ACKNOWLEDGMENTS
We would like to thank our dedicated research assistants Alyssa Jackson, Anna Speder, Eman Magzoub, Elaheh Salehi, Satvir Basran, Hannah Finegold, Jonas Sch€ one, and Emily Jusuf for their time and commitment to this project.The first author wishes to express her gratitude to Drs.Amanda Lazar, Eun Kyoung Choe, and Hernisa Kacorri who were abundantly helpful and offered invaluable assistance, support, and guidance.9.The minimal sample size was 39 to detect a medium effect size, with 80 percent power, as large as d = 0.64 which was observed in [12].In our study both treatment and control groups had a sample size larger than 39.

Fig. 1 .
Fig.1.Procedure flowchart of the study.All participants went through the same procedure until end of Post-stressor 1.At that point, participants who were randomly assigned to the control group went through Block 2 while receiving no vibrations (indicated by a solid black line), while participants who were randomly assigned to the treatment group go through Block 2 while receiving vibrations.Subjective measures of anxiety (i.e., STAI-6) were collected at five conditions of Baseline, V-Breathing Practice, Pre-stressor 1, Post-stressor 1, and Post-stressor 2 (indicated by a color-coded penand-paper icon).Physiological (i.e., skin conductance and respiration) measures were collected during the entire procedure.[17].

Fig. 2 .
Fig.2.An example of empirical mode decomposition (EMD) approach to decompose a respiratory signal into nine intrinsic mode functions (IMFs) components.We removed the first two and the last IMF to filter tissue artifacts (e.g., any muscle movement), posture change, and background noise from the respiratory signal.

Fig. 3 .
Fig. 3. Examples of breathing signals.The red dots are rolling average breathing rates with window of five breaths, and the red line shows the linear regression for the breathing rates.Top left: Passive breathing.Top right: Active breathing.Bottom: Combination of both passive and active breathing.All three examples were from participants in the treatment group who received pacer intervention during the Stressor 2 condition.The cutoff to categorize a breathing rate as active was 14 breaths per minute or slower.

Fig. 5 .
Fig. 5. Linear mixed model fit to breathing rate estimates.Note that the model fit lines with low negative slopes for the control group and lines with large negative slopes for the treatment group (shown in red), in both cases considering individualized intercepts.The model predicts a significant drop in breathing rate for the treatment group going from the Stressor 1 to Stressor 2 condition.

Fig. 4 .
Fig.4.Left: Average breathing rate estimate during conditions of the study for both treatment and control groups.An interaction effect was observed between group (treatment and control) and condition (Stressor 1 and Stressor 2): the treatment group receiving vibrotactile patterns during Stressor 2 experienced a drop in the breathing rate compared to the control group.The solid line indicates the treatment group; the dotted line indicates the control group.Note that the treatment group breathing was significantly lower compared to the control group when the pacer was active (highlighted by the red circle).Right: Average active breathing rate during Stressor 1 and 2 conditions as reported on a scale of 1 to 100 by both treatment and control groups.An interaction effect was observed between group (treatment and control) and condition (Stressor 1 and Stressor 2): we did find that receiving vibrotactile patterns had an influence on portion of active breathing.

Fig. 6 .
Fig. 6.Three measures of skin conductance all indicating no difference between the treatment and control group.Right: The large inferred number of peaks in the phasic drive (CDA.nSCR)during Stressor 1 and Stressor 2 demonstrated that the participants experienced significantly higher levels of sympathetic arousal during the stressor tasks.

Fig. 7 .
Fig. 7. Left: Average anxiety self-reported (STAI-6) rate estimate during conditions of the study for both treatment and control groups.Two two-way interaction effects was observed between group (passive, active, and control) and condition (Stressor 1 and Stressor 2): both passive and active groups receiving vibrotactile patterns during Stressor 2 experienced a drop in their self-reported anxiety compared to the control group.The solid line indicates the active group; the dotted lines indicate the passive (in pink) and control (in gray) groups.Note that both passive and active groups anxiety were significantly lower compared to the control group when the pacer was active (highlighted by the red circle) which may suggest an existence of a placebo effect.Right: Average skin conductance (specifically CDA.nSCR) during Stressor 1 and Stressor 2 conditions as reported on a scale of 1 to 100 by passive, active, and control groups.No interaction effect was not observed between group (passive, active, and control) and condition (Stressor 1 and Stressor 2): Passive, active, and control groups were indistinguishable.

9. 4
What is the Relationship Between the Type of Stressor and the Skill of the User in Breathing With the Pacer?