Alleviating time poverty among the working poor: a pre-registered longitudinal field experiment

Abstract Poverty entails more than a scarcity of material resources—it also involves a shortage of time. To examine the causal benefits of reducing time poverty, we conducted a longitudinal field experiment over six consecutive weeks in an urban slum in Kenya with a sample of working mothers, a population who is especially likely to experience severe time poverty. Participants received vouchers for services designed to reduce their burden of unpaid labor. We compared the effect of these vouchers against equivalently valued unconditional cash transfers (UCTs) and a neutral control condition. In contrast to our pre-registered hypotheses, a pre-registered Bayesian ANCOVA indicated that the time-saving, UCT, and control conditions led to similar increases in subjective well-being, reductions in perceived stress, and decreases in relationship conflict (Cohen’s d’s ranged from 0.25 to 0.85 during the treatment weeks and from 0.21 to 0.36 at the endline). Exploratory analyses revealed that the time-saving vouchers and UCTs produced these benefits through distinct psychological pathways. We conclude by discussing the implications of these results for economic development initiatives. Protocol registration The Stage 1 protocol for this Registered Report was accepted in principle on 27/06/2019. The protocol, as accepted by Nature Human Behaviour, can be found at https://doi.org/10.6084/m9.figshare.c.4368455.

Section 2: Baseline sample characteristics Table S1a. Baseline differences between Attritors (did not completed the full study, (n = 83) and Remainers (n = 1,070). Notes. Reporting means and standard deviations for respondent characteristics at baseline. Time spent on paid and unpaid labor is measured as a percentage of total time reported for the past 7 days. To adjust for multiple comparisons, we used Bonferroni correction. Using this correction, the significance level for these comparisons is p < 0.004. Therefore, we can determine that Attritors did not significantly differ from Remainers on any of these baseline characteristics. Notes. Reporting means and standard deviations for respondent characteristics at baseline. Time spent on paid and unpaid labor is measured as a percentage of total time reported for the past 7 days. Given the dichotomous nature of the outcome measure, binary logistic regression was used for this analysis. The first interaction term represents Control/UCT vs. Time-saving and the second interaction represents Control/Time-Saving vs. UCT conditions. As indicated by the overall model statistics, none of the overall regression models were significant; thus, the results of interaction models should be interpreted with caution. Note. At baseline, participants provided an open-ended response to the question: "What is your primary job?" This question focused on their occupation. How this person was paid (i.e., salary-based, task-based, or hour/daily) and whether they owned a share in their business (micro-enterprise ownership) were coded separately. Responses were coded into 5 categories: all sales jobs (includes selling produce, meals, clothing, and other consumer goods; working at a kiosk); trades (includes cooks, tailors, construction workers, carpenters, electricians, artisans, and all other skilled labor); personal services (includes hairdressers, restaurant staff, drivers, house cleaners, and washing clothes); casual laborer (includes temporary workers, wage laborer and kibarua); childcare, education, and healthcare services (includes daycare workers, teachers, school administrators, and community health workers). If participants mentioned more than one job, their job code was determined based on the first job they described. Job code was marked as missing if a participant's response could not be understood (N = 25). Note. Reporting means, standard deviations, and statistics testing for differences by condition. To adjust for multiple comparisons, we used a Bonferroni correction. With this correction, the significance level for each comparison is p < 0.004. Thus, we find no significant differences by condition at baseline, which supports our assertion that random assignment was successful.

Section 3: Supplementary methods and results for manipulation check
As a manipulation check, we tested for differences between the UCT and time-saving conditions with respect to 'change in perceived burden of unpaid labor' during the intervention period. In three consecutive weekly phone surveys during the treatment (Weeks 3-5), participants in the UCT and time-saving conditions were asked: "Over the past 7 days, to what extent did receiving [cash / prepared meals / laundry services] affect your burden of unpaid labor (-3=decreased my burden of unpaid labor a lot, 0=did not change my burden of unpaid labor, 3=increased my burden of unpaid labor a lot)? Participants in the control condition were not asked this question since they received no windfalls.
A critical assumption of this research is that participants in the time-saving condition reported experiencing a lower burden of unpaid labor as compared to participants in the UCT condition. Therefore, we conducted a Bayesian independent samples t-test (one-sided). This assumption was confirmed. We find a Bayes factor of BF10 > 1000 (error % < 0.001), which is very strong evidence in support the hypothesis that time-saving services reduce participants burden of unpaid labor (Tables S2-S3, Figure S1).  Figure S1. Prior/posterior distribution density plot: Bayesian independent samples t-test results for change in perceived burden of unpaid labor.

Section 4: Exploratory analyses
We conducted exploratory analyses to examine the effects of condition assignment over the course of the experiment, including effects during the intervention (Weeks 3-5). We explored the mechanisms underlying the observed differences in in subjective well-being, perceived stress, and relationship conflict at three time points: baseline, during the intervention, and endline. Lastly, we examined individual differences in treatment effects based on baseline characteristics including level of education, occupation, microenterprise ownership, household size, income, subjective well-being, perceived stress, relationship conflict, and risk of depression.
For all exploratory analyses, we conduct two sets of analyses. First, we conduct these analyses with three pre-registered conditions (N=1,070). Second, we conduct these analyses with data from an additional exploratory time-saving condition described below (N=365). All exploratory results in the main text are reported using data from the three pre-registered conditions only.

Pre-registered time-saving vouchers condition (N=349).
Participants received either prepared meals or laundry services once per week for three consecutive weeks. To possibly amplify the benefits, participants were asked to make a plan for how they would spend the additional time they had as a result of receiving these time-saving vouchers.
Prior to teach treatment week, participants provided an open-ended response to the following question: "Next week, you will receive a [prepared meal service / laundry service] designed to save you time. How do you plan to spend this additional free time?" We then asked participants follow-up questions to increase the specificity of their plans: "Where will you complete this activity / these activities?"; "Who will you complete activity / these activities with?" Additional exploratory time-saving vouchers condition (N=365). This condition was identical to the condition described above, except that no planning questions were asked.       Note. Reporting standardized coefficients for 3 parallel mediation models: 1) effects on subjective wellbeing, 2) perceived stress, and relationship conflict. Each of these outcomes are a weighted average of responses during the treatment weeks (weeks 3-5). For each model, the independent variable is condition, where 1=time-saving and 0=UCT. The control condition was dropped from these analyses. The mediators are modelled in parallel, and the following covariates are included in estimates of both the mediator and the dependent measure: baseline income (log) and the respective baseline dependent measure. N = 620, listwise deletion of cases with missing data. 5000 bootstrapped samples. *95% confidence interval around the indirect effect does not include 0. Note. Reporting standardized coefficients for 3 parallel mediation models: 1) effects on subjective wellbeing, 2) perceived stress, and relationship conflict. Each of these outcomes are a weighted average of responses during the treatment weeks (weeks 3-5). For each model, the independent variable is condition, where 1=time-saving and 0=UCT. Includes exploratory time-saving condition. The control condition was dropped from these analyses. The mediators are modelled in parallel, and the following covariates are included in estimates of the mediator and dependent measure: baseline income (log) and baseline dependent measure. N = 913, listwise deletion of cases with missing data. 5000 bootstrapped samples. *95% confidence interval around the indirect effect does not include 0. Note. Reporting standardized coefficients for 3 parallel mediation models: 1) effects on subjective wellbeing, 2) perceived stress, and relationship conflict. Time spent on paid work and socializing are calculated as a percentage of total time reported in each week; weighted average of the three treatment weeks (weeks 3-5). For each model, the independent variable is condition, where 1=time-saving and 0=UCT. The control condition was dropped from these analyses. The mediators are modelled in parallel, and the respective baseline dependent measure is included as a covariate in estimates of the mediator and dependent measure. N = 520, listwise deletion of cases with missing data. 5000 bootstrapped samples. *95% confidence interval around the indirect effect does not include 0. Note. Reporting standardized coefficients for 3 parallel mediation models: 1) effects on subjective wellbeing, 2) perceived stress, and relationship conflict. Time spent on paid work and socializing are calculated as a percentage of total time reported in each week; weighted average of the three treatment weeks (weeks 3-5). For each model, the independent variable is condition, where 1=time-saving and 0=UCT. Includes exploratory time-saving condition. The control condition was dropped from these analyses. The mediators are modelled in parallel, and the respective baseline dependent measure is included as a covariate in estimates of the mediator and dependent measure. N = 1033, listwise deletion of cases with missing data. 5000 bootstrapped samples. * 95% confidence interval around the indirect effect does not include 0.       Note. Reporting 27 moderation analyses, each using Preacher and Hayes Process model 1 with 5000 bootstrapped samples. All models control for the respective baseline dependent measure. Occupation was coded into 4 dummy variables: 1) trades, 2) personal services, 3) casual labor, and 4) childcare, education, and healthcare services; 'all sales jobs' coded as the reference category. To adjust for multiple comparisons, we used a Bonferroni correction. With this correction, the significance level for each comparison is p < 0.002. Thus, micro-enterprise ownership was the only individual difference that influenced that effect of time-saving vouchers versus UCTs. ***p < .001.    Note: To save time and money, the eligibility survey was designed such that if a participant was excluded on any variable, the survey would be terminated immediately. For example, if a participant did not live in Kibera, they were not asked if they were free to participate. Thus, the number of exclusions reflect participants who were excluded on each variable (after inclusion on the previous variables). a A subset of participants (N =171) did not answer the first two questions and were treated as missing in these analyses. b The survey did not automatically exclude participants after they reported their gender, explaining the identical denominator for children and gender in this table. c We decided to exclude participants only if they lived 45 minutes or farther from KTC (vs. 30 minutes as per our preregistration) based on recruitment advice from our field officers. Although many participants lived more than 30 minutes away, they passed by Kibera Town Center frequently while commuting to work or running errands, thus KTC was conveniently located for most. d A subset of participants did not answer this question (N =358). Notes. Reporting means and standard deviations for respondent characteristics at baseline. To adjust for multiple comparisons, we have used Bonferroni correction. Using this correction, the significance level for these comparisons is p < 0.004. Therefore, participants who completed the endline survey over the phone differ only in terms of age and baseline hours of unpaid labor as compared to those who completed the endline survey in-person at KTC.   Note: Individual comparisons are based on a Cauchy prior distribution with an r-scale value of 0.3 for comparisons between the UCT condition and the control condition; 0.5 for comparisons between the Time-saving condition and the control condition; and 0.4 for comparisons between the UCT and Times-saving conditions. N = 764.   Note: Individual comparisons are based on a Cauchy prior distribution with an r-scale value of 0.3 for comparisons between the UCT condition and the control condition; 0.5 for comparisons between the Time-saving condition and the control condition; and 0.4 for comparisons between the UCT and Times-saving conditions. N = 604.