2. Context

In this study, we focus on the effect of time preferences on the medication adherence of expectant mothers in South Africa. In 2014, the South African National Department of Health (NDoH) developed a text-message-based information and reminder system called MomConnect. This platform aims to assist expectant mothers through their pregnancy, and in the early period after birth. It does this by, for example, reminding them to schedule pregnancy check-ups, providing them with a description of how babies develop, informing them about the health benefits of breastfeeding, and informing them about symptoms that would indicate that they should seek medical help. When we ran this study, nearly 2 million women had joined MomConnect, making it one of the largest mHealth programs of its kind. Similar programs have been developed in other places, for example in India.

One of the aims of the program is to improve medication adherence among expectant mothers in South Africa, since pregnancy and early infancy are critical periods for the development of a child. In this paper, we focus on studying adherence to the specific recommendation that pregnant women should take daily iron supplements to prevent anemia. Anemia during pregnancy has negative health effects for both mothers and babies. During the last two trimesters of pregnancy, iron is essential for the mother as her body needs to produce more blood and grow both the placenta and the fetus. Anemia is associated with an elevated risk of maternal mortality, problems with lymphocyte stimulation (related to the immune systems), risk of pre-term delivery, and risk of low birth weight (see, e.g. Allen, Reference Allen2000; Lozoff et al., Reference Lozoff, Beard, Connor, Felt, Georgieff and Schallert2006; Kalaivani, Reference Kalaivani2009; Abu-Saad & Fraser, Reference Abu-Saad and Fraser2010; Balarajan et al., Reference Balarajan, Ramakrishnan, Özaltin, Shankar and Subramanian2011). The World Health Organization estimates that in 2016 the rate of anemia amongst women of reproductive age was around $25\%$ in South Africa and $33\%$ globally (World Health Organization, 2020). This is in spite of the fact that in South Africa the government runs a program to supply free iron supplements to all pregnant women. Health workers are instructed to provide pregnant woman with a supply of the supplements at the first antenatal visit and to follow-up by checking for signs of anemia during the second visit (for further evidence documenting the status of anemia and iron supplement intake in the South African context, see, e.g. Nojilana et al., Reference Nojilana, Norman, Dhansay, Labadarios, Van Stuijvenberg and Bradshaw2007; Phatlhane et al., Reference Phatlhane, Zemlin, Matsha, Hoffmann, Naidoo, Ichihara, Smit and Erasmus2016; Tunkyi & Moodley, Reference Tunkyi and Moodley2016; Mbhenyane & Cherane, Reference Mbhenyane and Cherane2017; Harika et al., Reference Harika, Faber, Samuel, Kimiywe, Mulugeta and Eilander2017; Symington et al., Reference Symington, Baumgartner, Malan, Wise, Ricci, Zandberg and Smuts2019).

In summary, all women from whom we sample are recommended to take iron supplements and should have access to free supplements. This makes the context ideal for testing our research question because it enables us to abstract from budget and availability constraints. At the same time, the South African setting is one in which the WHO numbers suggest that there is scope for reducing anemia and where a suitable platform for scalable interventions is well-established.

3. Theoretical framework

In this section, we provide a simple theoretical framework to illustrate the effect of time preferences on the decision to take iron supplements.Footnote ³ The theoretical framework informs our experimental design and the interpretation of our results.

A pregnant woman’s decision of whether to follow the recommendation to take iron supplements daily during pregnancy has similar characteristics to an investment decision: It requires an immediate cost and involves a future benefit. The immediate cost $c$ includes the effort cost of getting and taking supplements, the cognitive cost of remembering to take the supplements, and costs of short term side effects such as an upset stomach. In general, the cost would also involve costs of buying the supplements but in our setting pills are typically available for free at the health centers. The future benefits $b_t$ capture all health benefits to the mother and her baby, e.g. the reduced risk of maternal mortality, the reduced risk of pre-term delivery, and the reduced risk of low birth weight. If the expectant mothers do not take iron supplements, they receive the best alternative payoff which, without loss of generality, we normalize to $0$.

As with other investment decisions, the decision to take daily supplements is an inter-temporal decision problem. Allowing for the present-biased time preferences (Phelps & Pollak, Reference Phelps and Pollak1968; Laibson, Reference Laibson1997; O’Donoghue & Rabin, Reference O’Donoghue and Rabin1999) as captured by the $\beta \delta-$model (O’Donoghue & Rabin, Reference O’Donoghue and Rabin1999), the net benefit of taking the supplements in period $t$ are:

(1)\begin{equation} -c + \beta \sum_{s=t+1}^{\infty} \delta^{s-t}b_s. \end{equation}

where $\beta$ is an additional discount factor between present and future periods, applied on top of the usual exponential discount factor $\delta$. The expectant mother has time-consistent preferences if $\beta=1$ and she has present-biased preferences whenever $\beta \lt 1$. As suggested by O’Donoghue and Rabin (Reference O’Donoghue and Rabin1999) individuals may or may not correctly anticipate future present bias. Let $\hat{\beta}\in[\beta;1]$ denote the prediction of $\beta$ in future periods. The standard time-consistent exponential discounter then has $\beta=\hat{\beta}=1$, the sophisticated present-biased individual has $\hat{\beta}=\beta \lt 1$, and the fully naive present-biased individual has $\beta \lt \hat{\beta}=1$. With this notation the expectant mother anticipates that the future net benefit of doing the task in a future period is $-c + \hat{\beta} \sum_{s=t+1}^{\infty} \delta^{s-t}b_s$. Hence, naive present-biased individuals mispredict their future net benefit. The distinction between naivety and sophistication may be of importance, when individuals have opportunities to delay doing a task (e.g. to later in the day),Footnote ⁴ as sophisticates will anticipate that they may not do the task if they delay. As shown by O’Donoghue and Rabin (Reference O’Donoghue and Rabin1999), sophisticates may then do the task sooner than a time-consistent individual would have done and they may use commitment devices to ensure that they take supplements. This simple model illustrates a number of ways in which time-preference may influence adherence:

• • Adherence is increasing in patience as captured by the exponential discount factor $\delta$, i.e. women who are more impatient and therefore discount the future more are less likely to take supplements.

• • For naives, adherence is increasing in the present-bias parameter $\beta$, i.e. women who display time-inconsistent delay are less likely to take supplements.

• • Sophistication can alleviate problems related to present-bias.

4. Observing time preferences in the effort domain

We set out to design a simple incentivized time-preference experiment which can be used in low resource settings to classify participants in a meaningful way according to their behavior in the experiment. Our simple task does not allow for tight identification of time preference parameters. Nevertheless, we will demonstrate the usefulness of our categorizations for predicting adherence.

Traditionally, time preferences have been elicited in the monetary domain, where participants choose between either more money now or more money later. However, recent papers have emphasized the importance of measuring time preferences within the specific domain of interest (see, e.g., Ubfal, Reference Ubfal2016; Augenblick et al., Reference Augenblick, Niederle and Sprenger2015; Cohen et al., Reference Cohen, Ericson, Laibson and White2020). Since adhering to medication can be viewed as a type of (cognitive) effort task, we investigate time preferences in the effort domain. For this purpose, we design a simplified version of the experiment used by Augenblick et al. (Reference Augenblick, Niederle and Sprenger2015)Footnote ⁵ which is suitable for a text-message based implementation.

4.1. Experimental design

The experiment begins on day $t=0$ when participants are informed that they can earn phone credits worth 20 South African Rand (equivalent to 1.4 USD) if they complete a word task. The task involves writing a list of words backwards. Participants are given two examples and are told that they will receive text messages containing the relevant word lists in the coming days. Importantly, the longer participants wait to do the task, the more words they will need to write to earn the 20 Rand phone credits. The reason for this increasing effort cost is to make procrastination expensive. The participants can choose between writing four words on day $t=1$, six words on day $t=3$, or ten words on day $t=5$. Irrespective of when they complete the task, they will receive phone credits on day $t=7$. At time $t=0$, participants are fully informed about this procedure and are asked whether and when they expect to do the task. We provide a fixed payment of 5 Rand for their answer to this question on day $t=0$ and we do not incentivize how accurately their choice on day $t=0$ matches subsequent choices.

On day $t=1$, participants receive a new message with a list of four words: “baby, parent, bottle, mom.” A participant who replies with these words written backward has completed the task. If the participant does not complete the task, she gets two additional chances to do the task: on day $t=3$ and on day $t=5$ when new lists of words are sent. The time-line of the text-message task can be seen in Figure 1 and the full set of messages for the task can be found in Appendix.

Fig. 1

Timeline of the text-message-based real-effort task

We developed this specific word task for several reasons. First, the task had to be suitable for the text-message implementation. The majority of the effort tasks commonly used in the economics literature are either too long, or infeasible for a limited contact context such as this (e.g. implemented over text-message).Footnote ⁶ Second, in our setting, it is important that participants can complete the task at any time throughout a day. This way the task might mimic the underlying effort and cognitive constraints that drive time preferences and medication adherence. In pilot tests, we found that although the word task was considered tedious and frustrating, it did not take long to do.

4.2. Interpreting choices in the time-preference elicitation

We follow the approach in Augenblick (Reference Augenblick2018) and Augenblick and Rabin (Reference Augenblick and Rabin2019) to classify time preferences from a comparison of future and immediate effort choices. The initial plan of whether and when to do the task is a future effort choice. Behavior in the subsequent periods, implicitly give us an immediate effort choice. However, the simplicity of our task implies that we cannot identify precise estimates of the time-preference parameters in the $\beta \delta$-model; instead, we classify the women based on their future effort choice as either Early-Planners (who plan to do the task as soon as possible) or Late-Planners (who plan to delay the task). These two classifications are noisy measures of the participant’s patience as often captured by the exponential discount factor. In addition, we classify the participants based on whether their future and immediate effort choices are consistent meaning that they act according to plan. The women are therefore classified as either On-Timers (who do the task when they had planned), Late-or-Never-Doers (who do the task later than planned if at all) or Early-Doers (who do the task earlier than planned).Footnote ⁷ We discuss the details of the classifications in the following subsections.

To illustrate the classification mechanism, let $k$ denote the time period when an effort decision is made, let $t$ denote the time period when the task is performed, and let $e_t$ denote the effort level, i.e. number of written words where $e_1=4$, $e_3=6$ and $e_5=10$. The wage level, $w$, is the amount of money paid out at time $T=7$. In our experiment, $w =20$ if the participant completes the task at any point in time, and $w=0$ otherwise. For tractability we assume that the disutility of effort is captured by a function $C(e_t)=ce_t$. We allow for variation in $c$ across individuals and time to capture variation in busyness, distractions, phone access, and salience. We let $c$ have distribution $F(c)$ with mean $E[c]=\frac{1}{\phi}$ such that $\phi$ can be interpreted as the expected “exchange rate” between effort and money as in Augenblick (Reference Augenblick2018).Footnote ⁸ We assume that both effort and monetary payments are discounted by a quasi-hyperbolic discounting function that discounts costs by the discount factor 1 when $t = k$ and $\beta \delta^{t-k}$ when $t \gt k$.

In this setting, where the payment for the task does not depend on when the task is completed, utility maximization is equivalent to effort-cost minimization subject to a participation constraint. The participation constraint ensures that the discounted effort cost of doing the task at the chosen time $t^*$ does not exceed the discounted payment from doing the task. That is at time $k\geq0$ the participant should expect to do the task at time

(2)\begin{equation} t^\star=\underset{t\in\{1,3,5\},t\geq k}{\operatorname{argmin}}\{\beta^{1(k \lt t)}\delta^{t-k}\cdot c_t^{(1-1(k \lt t))}(\frac{1}{\phi})^{1(k \lt t)}e_t\} \end{equation}

subject to the participation constraint $\beta \delta^{T-k}\cdot 20\geq \beta^{1(k \lt t^\star)} \delta^{{t^\star-k}}\cdot c_{t^\star}^{(1-1(k \lt t^\star))}(\frac{1}{\phi})^{1(k \lt t^\star)}e_{t^\star}$ where $1(k \lt t)$ is an indicator function for whether the effort allocation is in the future relative to the point in time when the decision is made, and where $c_t$ is the time $t$ realization of $c$ which is only known once period $t$ arrives. In both Equation 2 and in the participation constraint, $\beta$ is replaced by $\hat{\beta}$ when the participant makes a prediction about future present bias. This allows us to capture full sophistication ( $\hat{\beta}=\beta$), partial sophistication ( $\beta \lt \hat{\beta} \lt 1$) and naivete ( $\beta \lt \hat{\beta}=1$) about future present bias. Below we describe optimal behavior under various assumptions regarding sophistication. We refer to the appendix for the technical details.

4.2.1. Future effort choices: what do women plan to do?

At time $k=0$ the participant states whether and when she plans to do the task and therefore implicitly states anticipated future effort at time $t=1$, $t=3$ and $t=5$. As mentioned above we do not incentivize the accuracy with which the future effort choice matches effort choices made at later points in time (i.e. for $k \gt 0$). This is to reduce the extent to which (sophisticated) participants use their answer at time $k=0$ as a commitment device to discipline future behavior and it is important because our categorization of individuals is based on a comparison between future and immediate effort choices, and hence to get variation in behavior, we do not specifically want to encourage (or discourage) consistent behavior.

For ease of exposition, we first focus on the case were individuals at time $k=0$ are naive about future present bias.Footnote ⁹ That is the women do not foresee that present bias will influence future behavior when they state their plan. This naivete assumption leaves room for discrepancies between planned effort) from the perspective of time $k=0$) and actual behavior (at time $k \gt 0$) even in cases where the realized costs of effort is exactly as expected (i.e. $c_t=\frac{1}{\phi}$).

Solving the problem in Equation (2) for $k=0$ involves doing a number of pairwise comparisons of the discounted cost of exerting effort at $t\in\{1,3,5\}$. Table 1 summarizes the conditions generated by the pairwise comparisons. Since subjects are making the plan at time $k=0$, all effort costs and payments are in the future. This implies that $\beta$ is irrelevant and cancels out. Furthermore, all effort costs are equal to the expected level an hence choices depend on the exchange rate between effort and money $\phi$ rather than on a specific realization of $c$.

Table 1

Planned future effort

Notes: The table gives conditions under which the alternative possible timings are optimal perspective of period $k=0$ if individuals are exponential discounters or naively present-biased at $k=0$. Further details are provided in the appendix.

Table 1 clearly shows that a participant may plan not to do the task, if the participation constraints are not satisfied, e.g. because the exchange rate $\phi$ is such that the wage paid is insufficient. If at least one of the participation constraints hold, then the anticipated optimal timing for the effort only depends on the size of her exponential discount factor, $\delta$. A participant with a low level of patience (i.e. a low $\delta$) will plan to do the task later. Therefore if we assume naivete (or standard exponential discounting at time $k=0$), the plan revealed in the first part of the experiment only depends on the exponential discount factor $\delta$ and not on the present bias parameter $\beta$ or beliefs thereof. This is also illustrated in Panel A of Figure 2.

Fig. 2

Model implied planned behavior for different levels of $\beta$ and $\delta$

Notes: The figure shows the planned and actual behavior predicted by the βδ model assuming naivite in the left panel and full sophistication in the right panel. For the figure we have set φ=1.

In contrast, if the women are partially or fully sophisticated about their present bias at time $k=0$, the analysis is somewhat more complicated as the individual then takes into account that even if she would like to complete the task on a particular day from the perspective of time $k=0$, when she arrives at that particular day, she may procrastinate due to her present bias. An individual who is fully sophisticated at time $k=0$ will realize this and will therefore estimate when she will actually do the task (if at all) and she will adjust her plan accordingly. In order, to make her plan, she will therefore solve the problem backwards and start by considering what she would expect to do if she arrives at the last period and has not yet done the task. Will it then be optimal to do the task? Or will she expect not to do the task? Given expected behavior in the last period, what is then optimal behavior in the period before that, etc.? This implies that the plan made by fully or partial sophisticates at time $k=0$ depends not only on $\delta$ but also on $\hat{\beta}$ since expected behavior at time $t\geq1$ depends on both $\delta$ and $\hat{\beta}$. Panel B of Figure 2 illustrates planned behavior in this case for fully sophisticated women (who have $\hat{\beta}=\hat{\beta}$). For partial sophisticates, the $\beta$ on the horizontal axis is simply replaced by $\hat{\beta}$ and planned behavior then depends on the belief about $\beta$ rather than the actual underlying preference parameter. The figure shows that even when women are fully sophisticated at time $k=0$, it is the case, that for high or low values of $\beta$, those women who plan to do the task at time $t=1$ have a higher exponential discount rate $\delta$ than those who plan to do the task later at $t\geq3$. For intermediate values of $\beta$ there is a less clear association between planning to do it early and $\delta$.

Nevertheless, under our assumptions, our model suggests that individuals who plan to do the task early in period $t=1$ (illustrated by the white areas in Figure 2) often have a high exponential discount factor $\delta$. We therefore classify participants based on their planned behavior as either Early-Planners if they plan to do the task in period $t=1$ or as Late-Planners otherwise (illustrated by the gray shaded areas in Figure 2)Footnote ¹⁰ and use this as a noisy classification of how much the participants discount the future. Late-Planners who plan to delay the task to period $t\geq3$ are largely less patient and discount more, hence their exponential discount factor $\delta$ is lower than that of Early Planners who plan to perform the task immediately in period $t=1$.

4.2.2. Future versus immediate effort choices: do women act according to plan?

The second part of our classification of behavior involves comparing the time $k=0$ plan of future effort to immediate effort choices made in periods $k=1$, $k=3$ and $k=5$, in order to elicit possible inconsistencies between the plan and actual behavior. Solving for the optimal timing in Equation (2) for $k=\{1,3,5\}$ again involves doing a number of pairwise comparisons in order to determine if planned behavior still is optimal as time progresses, and if it is not what the women will do instead. Within our model there are two basic components which may lead to deviations from planned behavior. First, when faced with a particular task, the individual may experience effort costs that are higher or lower than expected. That is, the realization of $c_t$ may differ from $E(c_t)$. If the effort costs are higher than expected, the woman may (regardless of their time preferences) choose to postpone the task and vice versa if the effort costs are lower than expected. In section 7, we discuss the extent to which systematic variation in effort costs may have influenced our findings. However, to ease the exposition in this subsection we will assume that $c_t=E(c_t)=\frac{1}{\phi}$ for all $t$ meaning that there is no time variation in effort cost. This is clearly a simplification but it enables us to focus on the second component of the model, which can result in deviations from planned behavior: present bias. In our analysis, we allow for naivete, as well as both full and partial sophistication, and we consider these cases in turn.

A fully sophisticated women with $\hat\beta=\beta$, will do exactly as planned when we abstract from variation in effort costs. This is because these women fully understand the extent of their present bias and therefore perfectly adjust their plan to reflect their subsequent behavior.

In contrast, a present-biased naive individual with $\beta \lt \hat{\beta}=1$ plans as if she has no present bias, but once the time to do a task arrives she will be influenced by her present bias and if the present-bias is strong enough she may want to delay the effort. Panel A of Figure 3 illustrates the theoretical predictions of planned and actual behavior for different combinations of the parameters $\delta$ and $\beta$ for navies.Footnote ¹¹. Along the vertical axis the figure mirrors Panel A of 2 and shows that for navies planned behavior is determined only by the size of $\delta$ as discussed in the previous subsection. Actual behavior in contrast depends on both $\beta$ and $\delta$. The very light gray areas indicate that the task is carried out as planned. Purple areas indicate that the task is carried out with delay which can occur if the women have intermediate level present-biased preferences. The blue shaded areas indicate situations where individuals plan to do the task but fail to do it. This can happen if the women have strongly present-biased preferences ( $\beta \lt 1$) and a relatively low exponential discount factor ( $\delta$).

Fig. 3

Model implied planned and actual behavior for different levels of $\beta$ and $\delta$

Notes: The figure shows the planned and actual behavior predicted by the βδ model assuming naivete in all periods in the left panel and partial sophistication in the right panel. For the figure we have set φ=1 and for partial sophisticates we let .

The behavior of partial sophisticated individuals is illustrated in Panel B of Figure 3 where we have set $\hat\beta=\sqrt\beta$ which ensures $\beta \lt \hat\beta \lt 1$. We use the same color coding. Clearly, compared to naifs, there are more combinations of $\beta$ and $\delta$ where women behave as they said they would (light gray areas). But unlike for the fully sophisticated, there are also cases where the women delay the task longer than planned (purple areas) or fail to do it even though they had planned to do the task (blue areas).

We group participants in our experiment along the color code as the colors represent different behaviors that can serve as noisy measures of underlying time preferences:

• • On-timers: These individuals carry out the task as planned (as in the light gray areas of the figure) and fully sophisticated women. Our model suggests that these women either have time-consistent preferences or only slightly present-biased preferences ( $\beta$ close to one) or they are sufficiently sophisticated about their present bias to be able to carry out their initial plan. In addition, the effort cost of doing the task does not deviate too much from what they had expected.

• • Late-or-never-doers: These women do the task later than planned, if at all (represented by the blue or purple areas above). Our model suggests that these are women who are present biased and to some extent naive, such that they fail to carry out their initial plan. However, this group also includes individuals who may in fact have limited self-control problems but for whom the effort cost of doing the task turns out to be much greater than expected.

• • Early-doers: These are women who are either future biased or face a lower effort cost than expected.

5. Implementation and sample selection

The data collection took place from the 19 March to the 29 June 2019. In total, a sample of 18,400 women was drawn randomly from the population of MomConnect users. We used the following criteria: (i) on the 19 March 2019, there were between 105 and 130 days until their expected due date, and (ii) the women were aged 18 or older. We sampled from the entire country and did not place a restriction on language. We have information on their preferred language, whether they live in an urban or rural area and whether they signed up with a smart phone or an older mobile phone without smart-phone capabilities.

This sample of women received a text-message invitation in English 105 days prior to their estimated due date asking whether they would like to participate in a research study on healthy pregnancy behavior for which they could earn phone credits. In line with our pre-study expectations, approximately 24% (4226) of the contacted women opted in to participate.Footnote ¹² These women were then randomized into six different treatment arms that were run in parallel. In total, 694 of the women were randomly assigned to the treatment arm relevant for the current paper.

A timeline of their participation can be seen in Figure 4. In an initial survey, participants were asked some baseline questions, including how many older children they had and whether they thought iron supplements were important for a healthy mother and child.Footnote ¹³ Fifteen days after opting in to the study, the women received the instructions for the real-effort time-preference task of writing words backwards. A total of 546 participants took part in the time-preference task. Approximately five weeks later, in the final week of data collection for this study, the women were asked to complete a short survey consisting of eight questions including questions about iron intake, iron deficiency and symptoms, perceived difficulty of remembering to take iron supplements, and experienced side effects.Footnote ¹⁴

Fig. 4

Timeline of the experiment

For our main analysis, we limit our sample to women who took part in the time-preference task and answered all eight questions of the final survey. This leaves us with a final sample of 480 women.Footnote ¹⁵ Table C.7 in Appendix C reports some basic demographic information for the individuals in our final sample. The average woman was born in 1992, making her 27 years old at the time of our data collection. Thirty-five percent of our sample has English as their preferred language. Twenty-nine percent of the women live in urban areas and on average the participants have one older child and 12 years of schooling. The table also tests whether there is biased selection on observables into our final analysis sample. Compared to the full sample of women who opted in, our final sample shows a slightly lower share of urban women. Otherwise, there are no observable differences.

7. Are the women simply busy?

One question that might come to mind is whether our results are driven by time preferences or simply by the fact that some women are busier or more distracted than others. In our theoretical model, we allow effort costs to vary over time and over individuals. It may be the case that our simple time-preference task is providing a measurement of how busy and cognitively constrained mothers are in their everyday lives, as opposed to being a “clean” measure of time preferences.

To investigate this, we approximate busyness in three ways. First, throughout their pregnancy, the mothers receive multiple messages as part of the larger experiment which they are asked to respond to. One might expect that women who are busier, in general, would take a longer time to answer these messages. Second, the timing of messages depends on the individual estimated due date of each expectant mother. Since the women have different due dates that are as good as random, they receive the same message on different days of the week. This provides us with exogenous variation in which day of the week a given message is received. We can use this variation to examine the role of being busy under the assumption that busyness varies between weekdays and the weekend. Lastly, it is plausible that women who already have kids at home might be busier than those who are pregnant with their first child. We use each of these three measures of busyness to provide evidence about whether busyness is a central mechanism driving our results.

First, we examine how quickly the expectant mothers replied to other messages, unrelated to the time-preference task. At the end of the first week after signing up for the study, participants received an incentivized text message asking them to reply with the first 4 digits of their ID number. This was nine days before the time-preference task and the purpose of the question was to see if the participant was still actively engaged and responsive to the study.Footnote ²¹ Table C.14 in Appendix C shows that there is no correlation between answering the text message in the first hour or not answering at all, and on the other hand, when they planned to do or actually did the time-preference task. We take this as evidence that our time-preference task is not driven by a general state of being busy when receiving the texts.Footnote ²²

Second, we examine whether the day-of-the week of receiving the message informing them about the time-reference task influenced their anticipated or actual behavior in the task. For some women, day 1 falls on a Friday, day 3 on a Sunday, and day 5 on a Tuesday. If a woman is less busy on weekends, then she might decide to do the task on Sunday (i.e on day 3), thus looking like a Late Planner to us. However, Table C.15 in Appendix C investigates if there is any difference in planning or completion of the task, depending on the day the task was received. Table C.15 shows no significant differences in the day-of-the-week in any of the results of the time-reference task, indicating that this is not a major concern in this setting.

Third, we replicate our regressions from Table 2, but now include “number of kids” as a control (see Tables C.16) The “number of kids” variable is not significant and does not change the estimates.

Overall, these results indicate that individual heterogeneity in general “busyness” of expectant mothers is not the primary driver of our results. We cannot control for all sources of busyness or exogenous shocks, so there might be some unobserved factors, but those women, who have no time or interest in taking part in our study should have already selected out of the sample at the start. Nevertheless, participation of those who opted in was very high and very consistent across several weeks.