Introduction

Episodic future thinking (EFT), the vivid imagining of personal future events, has been shown to reliably reduce delay discounting, the devaluation of future rewards [1, 2]. Higher delay discounting rates are associated with poorer glycemic control and diabetes-related self-care [3]. Importantly, a change in delay discounting over one year was predictive of increased HbA1c (glycated hemoglobin, a measure of retrospective blood sugar within the last 3 months) in adults with prediabetes, while controlling for demographic variables and body mass index (BMI; [4]). In recent years, researchers have examined the viability of EFT as an intervention to decrease delay discounting and increase engagement in health-promoting behaviors. In basic laboratory arrangements, EFT has been shown to reduce caloric consumption [5, 6], calories purchased in an online grocery shopping task [7], and behavioral economic demand for fast food [8]. EFT has been similarly efficacious in naturalistic settings, reducing caloric consumption [9] and calories mothers purchase while grocery shopping [10]. Finally, in clinical settings, EFT has been shown to reduce alcohol consumption [11] and increase medication adherence in adults with type 2 diabetes mellitus (T2DM; [12]); however, at least one other translational effort to incorporate EFT as a component of a clinical intervention has been less successful [13].

Regardless of the translation setting or phase, studies examining EFT’s effects typically include a control condition for comparison. A common control condition in basic laboratory settings is episodic-recent thinking (ERT), the vivid remembering of recent personal events (i.e., within the past few days). While ERT is an appropriate control in the laboratory or naturalistic settings, repeated (i.e., daily or multiple times per day) engagement in ERT over weeks or months is impractical in clinical settings [14]. To keep cues related to recent events, ERT cues would need to be generated often; matching EFT cue generation to this schedule would be prohibitive. Additionally, control conditions in clinical experiments are more likely to promote adherence to the intervention and minimize demand characteristics if they equate expectation of improvement between groups. Thus, participants should view the control condition as clinically relevant or potentially helpful as the active participants view the potential helpfulness of the intervention. However, participants engaging in ERT do not report the belief that ERT will increase the likelihood of engagement in health-promoting behaviors or decrease delay discounting [14], which may limit adherence in clinical trials.

To address these limitations, Rung and Epstein [14] developed Health Information Thinking (HIT), a novel control condition well suited for long-term investigations on the effects of EFT on delay discounting and other health behaviors in clinical settings. In an online experiment, Rung and Epstein [14] demonstrated that participants (a general sample of 254 Amazon Mechanical Turk workers] engaging in EFT had significantly lower delay discounting than participants engaging in ERT or HIT. In the HIT control, participants read and responded to health-information vignettes and generated HIT cues by describing their reactions. HIT participants also rated the information on several dimensions, including how much they liked learning the information, the importance of understanding the information, how exciting it was to learn the information, and how useful it was to learn the information. This approach results in personalized summaries of and reactions to health information, which mirrors common survey-guided EFT cue generation methods [15]. Additionally, the health-information vignettes may be tailored for a given clinical condition, increasing the likelihood that individuals with said condition will view HIT cues as a relevant intervention component. Thus, the HIT condition is likely to control for participants’ expectation of improvement and allows for matching cue generation and regeneration schedules, making HIT a potentially powerful control condition for long-term clinical studies examining the effect of EFT on delay discounting and health behaviors.

Although Rung and Epstein [14] showed that the HIT condition did not significantly impact delay discounting, the authors recruited participants from the general population and provided health information about a broad range of topics (e.g., effects of alcohol on sleep, recognizing symptoms of depression, understanding nutrition labels). In contrast, studies exploring the clinical effects of EFT have primarily focused on specific health behaviors in individual clinical populations (e.g., diet and physical activity in patients with prediabetes). It is unknown if a modified form of the HIT condition in which the health information is relevant to the prevention or treatment of a specific clinical target and population (e.g., weight loss and glycemic control in patients with T2DM) would yield different conclusions. Specifically, individuals with T2DM and obesity who generate HIT cues based on T2DM-related health-information vignettes (e.g., self-monitoring food, aerobic activity, nutrition labels) may react to these cues with increased motivation to engage in health-promoting behaviors, potentially spurring prospection and reducing delay discounting.

To further develop HIT as a clinical control for this population, the present study describes an online experiment in which participants with self-reported T2DM and obesity were recruited using Amazon Mechanical Turk and randomized to EFT, HIT, or a no-cue control (NCC) condition. EFT and HIT participants generated cues using a self-guided survey, while NCC participants did not generate cues. Afterward, participants completed a delay discounting task while instructed to vividly imagine or consider their EFT or HIT cues; NCC participants completed the delay discounting task without these instructions. We hypothesized that participants who were instructed to engage in EFT during the delay discounting task would have lower levels of delay discounting (i.e., more frequent selection of larger later rewards) compared to both HIT and NCC participants.

Materials and methods

Participants

We recruited 434 participants with self-reported obesity and T2DM using Amazon Mechanical Turk, from February 10^th, 2022, to November 17^th, 2022. The qualification criteria in Amazon Mechanical Turk specified that participants must be located in the US, have completed at least 100 previous human-intelligence tasks, and have an approval rating of at least 99%. To prevent individuals outside of the US from accessing the eligibility screening block or attempting to screen multiple times, we applied three sequential screening criteria to all responses. First, we used IP address screening to prevent participants using a virtual proxy network (VPN) from accessing the screening block [16]. Secondly, we prevented participants with IP addresses associated with countries outside of the US from accessing the screening block. Thirdly, we prevented participants with IP addresses that had already attempted or completed the survey from accessing the survey a second time. Participants who were not flagged for using a VPN, an IP address outside of the US, or a duplicate IP address were then shown the eligibility screening block. To be eligible, participants must have self-reported height and weight corresponding with a BMI ≥ 30 and self-reported a diagnosis of T2DM. To increase confidence in the truthfulness of self-reported T2DM, we asked participants to indicate any medical conditions (diagnosed by a medical professional) using two separate medical diagnoses questions. Each diagnosis question included the same 13 answer choices (i.e., various diseases and medical conditions; see survey file hosted on GitHub [https://github.com/jeremiahmbrown/public-remedi-hit-pilot] for complete list) presented in random order. After answering the first diagnosis question, participants answered three other unrelated questions (i.e., height, weight, and age) before answering the second diagnosis question on a separate page. To continue to the informed consent block, the selected answer options in both diagnoses questions had to match, and participants could not have selected more than 4 total diagnoses on either diagnoses questions. Additionally, we prevented participants from continuing if they selected any of the following answer choices on either diagnosis question: Type 1 diabetes, chronic obstructive pulmonary disease (COPD), colorectal cancer, breast cancer, or none of the illnesses in this list. The eligibility criteria regarding BMI, T2DM, and other medical diagnoses were assessed at the same stage in the survey; violating one or more of the criteria rendered participants ineligible and ended the survey. Fig 1 depicts a flow diagram of participant enrollment, eligibility, dropout, and completion.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Participant flow diagram.

Flow diagram depicting screening of MTurkers, assessment of eligibility, randomization, and study completion. Vol. With. = Voluntary withdrawal (i.e., did not continue responding to the survey).

https://doi.org/10.1371/journal.pone.0289478.g001

We determined the target sample size using an a priori power analysis. 396 participants (132 per group) provide 95% power to detect a medium effect size difference (Cohen’s d = 0.5) in delay discounting between individual pairwise comparisons (e.g., EFT vs. HIT; EFT vs. NCC) following an omnibus ANOVA. We assumed a medium effect size based on values observed by Rung and Epstein [14]. Higher attrition rates were observed in the EFT group than in the HIT and NCC control groups; we discontinued recruitment after obtaining 434 completed responses, resulting in 172 completed responses in the NCC control group, 142 in the HIT group, and 120 in the EFT group.

Data analysis

Data were accessed continually for research purposes from February 2022 to November 2022 to monitor study progress and for interim analyses, and from November 2022 to April 2023 to conduct final analyses reported in this manuscript. All data analyses were completed using R (version 4.2.1; [19]) and RStudio [20]. The tidyverse package was used to wrangle, clean, and shape data; the gtsummary package was used to create Table 1; the ggplot2 package was used to create all figures [21–23]. We compared EFT and HIT participant’s average ratings of cues to test for group differences in liking/enjoyment, importance, and excitement using three t-tests (i.e., one for each cue characteristic; ratings of usefulness generated by HIT participants and ratings of vividness generated by EFT ratings could not be compared). We controlled for type I error rate using Bonferroni correction (α = .0167). To quantify delay discounting, ordinal area under the curve (AUC) values were calculated using the discAUC package [24, 25]. Higher values of AUC reflect lower levels of delay discounting. To assess if delay discounting varied between EFT, HIT, and NCC groups, we performed a one-way ANOVA with Tukey HSD post hoc comparisons. As a sensitivity analysis, we performed these tests twice: once including only ordinal AUC values from participants who correctly answered three out of four attention checks during the delay discounting task, and again including all ordinal AUC values (see supporting information). To test for differences in delay discounting attention check passing rates between groups, we used a logistic regression to predict attention check passing status as a function of group. To test for differences in study attrition between groups, we used Fisher’s exact test. We used a significance level of α = 0.05 to determine statistical significance for all tests, unless otherwise stated. Income was self-reported using an ordinal scale (e.g., $10,000–19,999); in order to display income data in Table 1, income was recoded as a continuous variable by assigning participants a value in the center of their selected category (e.g., participants who reported $10,000 - $19,999 were recoded as $14,999.50). Deidentified raw data and R code used to generate results are available on the authors’ Github page (https://github.com/jeremiahmbrown/public-remedi-hit-pilot).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Demographics of participants who completed all study procedures, overall and by group assignment.

https://doi.org/10.1371/journal.pone.0289478.t001

Results

Delay discounting

Fig 2 depicts the decline in the subjective value of the larger later reward relative to the smaller sooner reward (i.e., indifference points) as a function of delay across EFT, HIT, and NCC groups. Fig 3 depicts a notched boxplot of corresponding ordinal AUC values in these groups. Data in Figs 2 and 3 includes participants who passed three or more discounting attention checks (see supporting information for recreations of Figs 2 and 3, including all participants who completed the experiment). Results of a logistic regression predicting delay discounting attention check passing rates between groups revealed no significant effect of group; a model including group as a predictor of passing rates was not significantly different than the null model, X² = 1.04, df = 2, p = .60. Results of the one-way ANOVA examining differences in ordinal AUC between groups (including participants who passed three or more discounting attention checks; n = 113, n = 129, and n = 159 for EFT, HIT, and NCC groups, respectively) indicate that at least one group generated significantly different AUC values (F(2, 398) = 8.54, p < .001). Post hoc comparisons using Tukey’s HSD indicate that the EFT group demonstrated significantly higher ordinal AUC values than the HIT group (Mean difference = 0.103, 95% CI: 0.04–0.167, p < .001, Cohen’s d = .49, 95% CI: 0.23–0.75) and the NCC group (mean difference = 0.089, 95% CI: 0.0279–0.149, p = .002, Cohen’s d = .43, 95% CI: 0.18–0.67). Additionally, no difference between the NCC and the HIT group was observed (mean difference = 0.014, 95% CI: -0.042–0.072, p = .81, Cohen’s d = .07, 95% CI: -0.16–0.30). All conclusions remained consistent when including all ordinal AUC values, regardless of successfully passing three out of four attention checks during the delay discounting task (see supporting information for model results).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Group-level mean and individual indifference points.

Mean and individual indifference points of participants who passed at least three of four attention checks during the discounting task. Panel A depicts mean indifference points as a function of delay by group assignment, plotted on an ordinal scale. Panels B, C, and D depict mean indifference points within each group (bold lines); transparent lines depict individual subject discounting curves. Error bars represent standard error.

https://doi.org/10.1371/journal.pone.0289478.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Ordinal AUC by group.

Notched boxplot of grouped ordinal AUC values of participants who passed at least three of four attention checks during the discounting task. Notches represent the median ± 1.58 * IQR/√n.

https://doi.org/10.1371/journal.pone.0289478.g003

Discussion

In this online experiment, we made two contributions to the literature: 1) the HIT control (using cues generated in response to T2DM related health-information) does not result in lower delay discounting compared to a no-cue control condition in adults with self-reported obesity and T2DM, and 2) brief engagement in EFT does result in lower delay discounting compared to the HIT control and a no-cue control in adults with self-reported obesity and T2DM. We will discuss both contributions, study limitations, and future directions.

Health Information Thinking appears to be a suitable control for clinical trials involving repeated engagement in EFT as an intervention to reduce delay discounting. In this clinical sample (adults with self-reported obesity and T2DM), engaging in HIT using cues generated from T2DM-related health-information vignettes did not result in lower delay discounting relative to participants engaging in EFT. Participants engaging in EFT exhibited significantly greater ordinal AUC (i.e., lower delay discounting) than HIT and NCC conditions. No differences in ordinal AUC were observed between HIT and NCC participants, suggesting that HIT is an inert intervention concerning delay discounting. These results are consistent with Rung and Epstein [14], who showed that the HIT and ERT control conditions produced comparable levels of delay discounting. These findings have important implications; namely, that HIT does not appear to influence delay discounting, thus making it appropriate as a control condition in clinical studies designed to reduce delay discounting and influence health behavior. Likewise, the HIT condition is structurally similar to the EFT intervention, as the number of vignettes and cues can be matched to the future events used in EFT. This provides a strong control for effort and time in clinical studies, as participants in both the active intervention and control groups would be scheduled to complete cue generation at approximately the same frequency.

These results suggest that EFT may be an efficacious intervention to reduce delay discounting in adults with obesity and T2DM. To date, only one multiple-baseline small-n study has examined the effects of EFT on medication adherence and delay discounting in adults with T2DM [12]. This prior study reported a large effect size difference in discounting between baseline and EFT conditions (ES = 1.2), although the small sample size (n = 4) prevented inferential statistics. Thus, the present study’s findings are important to the literature examining EFT as a clinical intervention for individuals with T2DM. However, as these findings are limited to acute (i.e, single exposure) EFT and delay discounting, future research should examine the effects of repeated (i.e., multiple exposures) EFT on delay discounting, health behaviors, and health outcomes.

Although the present study is the first large-scale examination of EFT in adults with self-reported T2DM, other studies have shown that acute and repeated EFT reliably reduces or results in lower delay discounting compared to control conditions in adults with prediabetes [13, 26, 27], a condition of elevated blood glucose that increases the risk of developing T2DM [28]. While EFT did not impact other health behavior outcomes in the aforementioned studies of adults with prediabetes, other work in adults with overweight or obesity has been promising [5, 8–10, 29]. Thus, more work is required to explore potential moderators of the efficacy of EFT (e.g., HbA1c, education, BMI) in regard to changing health behaviors.

This study is not without limitations that should be addressed in future research. First, while we are reasonably confident that our screening strategies limited the number of participants who completed the study but do not have a diagnosis of T2DM or obesity, it is likely some unknown number who did not meet our qualifications were able to meet the screening criteria by chance. However, this number is likely to be low, as we implemented emerging best practices for online research utilizing Amazon Mechanical Turk [16], and further developed methods to limit the ability to misrepresent a T2DM diagnosis in screening. As a result, less than 3% of participants could continue past the screening questionnaire. Additionally, previous research supports the validity of self-reported T2DM via interview methods, although it is unknown if these results generalize to self-reported T2DM via an online survey [30–32]. Furthermore, the prevalence of self-reported T2DM in MTurkers aged 40–49 (measured in 2015) was not different than the prevalence of T2DM in a representative sample of Americans (i.e., the 2013 Behavioral Risk Factor Surveillance System; [33]).

Second, we observed significantly higher attrition rates in participants randomized to the EFT group compared to participants randomized to HIT or NCC groups; this difference may be due to the increased effort required to generate EFT cues and engage in EFT during the delay discounting task. Importantly, no significant group differences in demographic variables related to delay discounting were observed (i.e., age, BMI, income, education, HbA1c), which suggests that the observed differences in attrition did not introduce bias related to any measured participant characteristic. We note, however, this differential attrition may be less likely to occur in a sample of motivated participants enrolled in a clinical trial using behavioral interventions and EFT to manage T2DM better.

Thirdly, one other potential benefit of HIT was elucidated by Rung and Epstein [14] but was not explored in the present study; namely, the ability to control for participants’ expectation of improvement. Future research should measure participants’ expectations of improvement associated with EFT and HIT conditions. Fourthly, Rung and Epstein [14] and the present study were unable to examine the effects of HIT (acute or repeated) on other health behaviors (i.e., food consumption, physical activity engagement); future research is required to examine these outcomes in clinical samples and to compare expectations of improvement in the context of a clinical trial. Indeed, our group is currently using the HIT control condition in a clinical trial examining the effects of repeated engagement in EFT (embedded within a behavioral intervention) on change in delay discounting, body weight, and HbA1c in adults with type 2 diabetes and obesity (see https://clinicaltrials.gov/ct2/show/NCT05280925). While data collection is ongoing, observing significant group differences in behavioral outcomes would suggest that the HIT control is inert with regard to delay discounting, health behaviors, and health outcomes. Similarly, observing no differences in participants’ treatment acceptability ratings and perceived helpfulness would suggest that expectations of improvement are not different between EFT and HIT.

Fifth, we observed significantly different average participant ratings of EFT and HIT cues on excitement and liking/enjoyment characteristics, although no differences in ratings of importance were observed. While differences in cue ratings between active and control groups are not ideal, EFT vs. HIT cue generation are, by design, different exercises resulting in qualitatively different written descriptions (i.e., a narrative about personal engagement in possible future events vs. a personal reaction to health information). Therefore, differences in EFT and HIT cues on ratings of excitement and liking/enjoyment may be related to the expected differences in content between EFT vs. HIT. Nonetheless, when applied in a long-term RCT involving repeated, daily engagement in EFT or HIT, these content differences may reduce adherence to HIT and, therefore, its suitability as a control condition. Future research should examine possible differences in intervention adherence between EFT and HIT in long-term RCTs.

In conclusion, the present experiment increases our confidence in using the HIT control for clinical examinations of the effect of EFT on delay discounting. The HIT control cues can be generated using health-information vignettes directly relevant to the clinical sample; engaging in HIT using these cues does not result in lower delay discounting relative to a standard delay discounting task. HIT control is a promising candidate for long-term clinical trials involving EFT as an intervention to reduce delay discounting and change health behaviors.

Supporting information