Experimental Design Details
The experiment is conducted through a longitudinal three-questionnaire online survey with a general population cohort (the "Lifelines" cohort based in the north of the Netherlands). Cohort members are invited to join the survey, which is programmed in Qualtrics, via an email invitation issued by the Lifelines organisation.
Respondents are randomly assigned to a control arm or to one of two different treatment arms - the TIP treatment or the TIB treatment. Subjects in the TIP treatment arm receive the TIP intervention described in the "Intervention" section above, and subjects in the TIB treatment arm receive the TIB intervention described above. Subjects are blinded to the experiment as far as is possible in this situation - subjects are not informed that they are taking part in an experiment but it is possible that they could discover that they are part of an experiment if they communicate with subjects from other arms.
Incentives - There are three different types of incentives. Participants are randomized to one of these incentive types to facilitate another separate experiment on incentives being run in parallel with the experiment described here. The randomization occurs as follows:
A.33% of subjects, measures are incentivised monetarily using Choice-matching (Cvitanić et al. 2019), a method for eliciting honest responses to survey questions. These participants receive the "Own Prediction" Predicted PA question (see Primary outcomes section above), which is incentivised with Choice-matching. See the attached PDF "Supplemental material" to see how this is explained to subjects.
B. 33% of subjects will have all measures incentivized with Choice-Matching, except for the Predicted PA question. These participants receive the "Prediction for a similar other" Predicted PA question (see Primary outcomes section above) and they receive monetary incentives for accurate predictions in this question.
C. 33% of subjects will receive no incentives. These participants receive the "Own Prediction" Predicted PA question (see Primary outcomes section above), which unincentivised.
After randomizing subjects to an incentive type, stratified randomization is used to randomize participants to one of the the three arms in this study (TIP, TIB, Control) using a single stratification factor - incentive type. This gives three strata - incentive types A, B, and C. Participants are randomized with equal probability to one of the the three arms (Control, TIP, TIB) within strata A and C. Within stratum B, participants are randomized with 66% probability to the TIP arm, 33% probability to the control arm, and zero probability to the TIB arm. This is because the TIB intervention is not possible to implement for individuals in stratum B, as they all receive the "Prediction for a similar other" prediction question, as described in the primary outcomes section above.
When analysing the data, adjustments are made to ensure that a balance is kept in incentive type between the two arms being compared (for instance, only subjects from strata A and C are included in the TIP arm when comparing to the TIB arm, as the TIB arm contains no subjects from stratum B).
Primary analysis
Linear regression will be used to estimate treatment effects on our primary outcomes. The outcomes analysed will be:
- Sophistication for the two weeks following questionnaire 2
- Demand for commitment summary index in each of questionnaire 2 and questionnaire 3
As there are two primary outcomes, False Discovery Rate-adjusted p-values ((Benjamini and Hochberg, 1995; Benjamini et al. 2006) will be presented alongside unadjusted p-values to account for multiple hypothesis testing.
I hypothesise that the two interventions will each have a positive treatment effect on each of the primary outcomes. To test this, the following comparisons will be made for each primary outcome:
- TIB Treatment v Control
- TIP Treatment v Control
- Pooled Treatment (TIB and TIP arms combined) vs Control.
I also test whether one intervention is more effective than the other, for which I have no prior expectation, using the following comparison:
TIP treatment v TIB treatment
The following variables will be used as controls: baseline values of the outcome variable being analysed, PA level, and TIP (calculated as (Ideal-Actual)/Actual for the two week period after questionnaire 1); the psychological variables trait self-control, optimism, and information avoidance; the sociodemographic variables age, gender, marital status, household composition, education; and I will also control for medical reasons that restricts the individual's ability to do PA, whether the person likes PA or not, and whether the person likes to be told when to do PA. In the regressions on Demand for Commitment, baseline sophistication will also be controlled for.
Available case analysis will be used. This means that in each analysis, any participant with missing data for the dependent variable in that analysis will be omitted from that analysis. Control variables will be encoded as categorical variables and each will have a “missing” category for participants who have missing data for that control variable.
Secondary analysis
Secondary analysis comprises firstly of analysis of secondary outcomes which is carried out in the same manner as that of primary outcomes outlined above.
Also, as part of secondary analysis, two subgroup analyses of primary outcomes will be carried out. The first analysis will be by time inconsistency categorisation in the 2 weeks preceding the intervention in questionnaire 2. That means that I have three subgroups in this analysis: 1. Time inconsistents (i.e. Ideal[Predicted PA] as measured in questionnaire 1 > Actual PA as measured in questionnaire 2); 2. Time consistents (i.e. Ideal[Predicted PA] = Actual PA); 3. Reverse time Inconsistents (i.e. Ideal[Predicted PA] < Actual PA). The second subgroup analysis will be by baseline PA levels, with three subgroups in the analysis - low, medium, high. The thresholds will be 33rd and 66th percentile.
Thirdly, a number of qualitative questions will be asked of treatment subjects at the end of questionnaire 3 in order to get some additional qualitative information on the mechanisms through which the treatment worked (or did not work). See attached pdf "Supplemental Material" for further details.
Finally, robustness checks on primary outcome findings will also be carried out. The outcome Sophistication will be disaggregated into two separate weekly measures (rather than one fortnightly measure) and analysed. For the outcome Demand for commitment, subjects who are deemed to be inconsistent in their commitment choices will be excluded in a robustness analysis. This is determined using the consistency check questions described in the outcome measures section above. A subgroup analysis by incentive type will also be carried out for robustness.
References
Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B (Methodological), 57(1), 289-300.
Benjamini, Y., Krieger, A. M., & Yekutieli, D. (2006). Adaptive linear step-up procedures that control the false discovery rate. Biometrika, 93(3), 491-507.
Cvitanić, J., Prelec, D., Riley, B., & Tereick, B. (2019). Honesty via choice-matching. American Economic Review: Insights, 1(2), 179-92.