Parent Training and Child Development at Low Cost? Evidence From a Randomized Field Experiment in Mexico

ABSTRACT Can at-scale, government-implemented parent training programs improve parenting practices and child development outcomes? This article presents evidence on the effects of a low-cost, group-based early childhood education program that provided parent training and direct child stimulation in rural communities, evaluated in six Mexican states. Despite limited take-up, the program had positive impacts on observed parent behaviors in its first year. An index of observed parenting behaviors increased by 0.20 standard deviations, with larger effects (0.32 standard deviations) for parents of the youngest children (ages 0–35 months). An index of impacts on child development showed no statistically significantly effects, but certain aspects of child development showed suggestive evidence of positive impacts in the first year. For both parenting practices and child development, effect sizes were smaller and not statistically significant in the second year. The fade-out of effects is consistent with existing literature on parenting programs. Impacts of the program on child development were not significantly different for girls versus boys or for younger versus older children. These results suggest that parent training can be implemented at low cost, although design changes to improve implementation and take-up would likely be needed to generate sustained impacts on parenting practices.


The early education program
The Early Education Program (Programa Educación Inicial or PEI) is an example of a non-formal education program aimed at developing competencies among children, and an additional set of skills for parents.In this program, promotoras (facilitators who receive two weeks of annual training, educational materials, and a small stipend) run up to 65 group sessions during a nine-month term, where parents and caregivers receive information and training over four main aspects: (1) childcare (health and safety), (2) personal and social development (autonomy and self-regulation), (3) language and communication, and (4) exploration and knowledge of the context. 3These sessions included (1) up to 26 sessions for caregivers and parents (men and women), (2) up to 18 sessions for caregivers and parents (men and women) focused on children, (3) up to 5 sessions for parents who are men, 4 and (4) up to 8 sessions for pregnant women.In addition, promotoras could organize up to 8 additional sessions for diagnosis, planning, and evaluation.
Promotoras would invite parents of children from 0 to 47 months old (i.e., 3 years and 11 months) via home visits, distribution of flyers, or through local networks (groups or committees).Sessions were generally held for two hours, and the frequency of these sessions is defined through an initial agreement between the promotoras and participants, often one or more times per week between the sessions of type (a) and (b).Operating out of community spaces (like schools or community centers) or people's homes, two types of sessions were organized: adult learning oriented and children learning oriented.In the first case, sessions were intended to be "experiential learning," aiming to foster reflections, a shared conceptualization of problems and opportunities, and the application of acquired knowledge.Child-oriented sessions, on the other hand, were based on game-based learning, including participation of parents in some of these activities.Children's sessions aim to develop cognition, language, communication, social skills, knowledge of the environment, and socio-emotional aspects.Sessions were usually organized around a theme (e.g., nutrition).About once a month, a supervisor would observe the session, and offer feedback to the promotora after the session.Two researchers characterized the community engagement in the program as follows: This program has strong family and community participation components. . . .Community participation includes selecting [promotoras], providing spaces for educational sessions, convening community meetings for monitoring and evaluation activities (for example, to learn about program progress and conduct self-evaluations), requesting the support of the authorities, and interacting with them as needed.(Vegas & Santibáñez, 2010) Because of its participatory orientation, PEI has relatively low training costs, low-cost educational materials, use of community resources for the organization of sessions (e.g., occasional meals, free facilities), and reduced supervision expenditures.

Data
To draw our sample, staff from the planning unit of the government agency that oversaw the program (the National Council for Education Development, called the Consejo Nacional de Fomento Educativo or CONAFE) helped to identify 300 communities located in poor or very poor rural areas across six states (Chiapas, Estado de México, Oaxaca, Puebla, Querétaro, and Veracruz), all fulfilling program selection requirements.Out of this population, we randomly drew 160 communities, and representatives from CONAFE visited these communities to advertise and held introductory meetings to confirm if parents were interested in participating in this program. 5Once we confirmed willingness to participate, we conducted a pair-wise matching process to define sets of two communities to be considered either as treatment or control groups. 6 During the validation process, we concluded that some of the selected communities were not eligible to participate, mostly due to safety issues and the possibility that they would lack adequate support from CONAFE (e.g., due to an inability to find a promotora or to carry out supervision in the most remote areas), increasing the odds of an ineffective implementation.We also excluded some other communities because early care interventions from other local public agencies had been recently implemented. 7Due to these concerns, our final sample was reduced from 160 to 130 locations, later reduced to 126 because of attrition: 64 treatment communities and 62 control communities. 8 In each of the three waves of data collection, we administered a questionnaire to caregivers measuring participation in the program (in treatment communities) and caregiver practices based on the Home Observation Measurement of the Environment (or HOME) scale (National Longitudinal Surveys, 2017), for both treatment and control communities.Interviewers also observed the interaction between children and caregivers in both type of communities.To measure child development, we relied on the administration of the Ages and Stages Questionnaire (ASQ-3) (Squires & Bricker, 2009).All of our instruments were reviewed in collaboration with staff from CONAFE, to ensure that the items being measured were in line with the skills being taught to caregivers and children enrolled in the program.In addition, we administered a questionnaire to promotoras, to capture potential heterogeneity in the implementation of the program associated to individual characteristics.

Baseline characteristics and balance across treatment groups
Participating children were evenly distributed between boys and girls, and roughly 80% of children were age 0-35 months.As expected, main caregivers were overwhelmingly female (97%), and 88% of them had completed only either primary or secondary school.A relatively small proportion of caregivers spoke only an indigenous language (3%). 9While these are rural communities, almost 80% of households had a solid or cement floor, about 91% had a bathroom or toilet, and about 91% had daily access to electricity.At the same time, 51% of caregivers lived in a household with only one bedroom, only 61% had daily access to water in the household, 52% had a cellular telephone, 5% had a computer, and only 3% had access to the Internet.Table 1 shows baseline characteristics of our sample.
At baseline, we observed balance on observed characteristics (Table 2).Across child development outcomes, as measured on the ASQ, we found no significant differences in any of the measured outcomes.Across observed caregiver outcomes, we observed nearly perfect balance: out of 21 observed characteristics across families allocated to treatment and control groups, only one is statistically significantly different at the 5% level.(One in 20 is the number of statistically significant differences at the 5% level that one would expect to see by random chance.)

Attrition
Unfortunately, not all caregivers and children remained in the study.Between 69% and 76% of children remained in the study during the first year, as well as between 68% and 73% of caregivers (Table 3).Attrition was lower in the second year (more than 75% retained on average across groups), after the research term invested significant effort in tracking all the households' members.While the level of attrition is higher than desired, the principal concern is whether attrition is correlated with treatment.However, a regression analysis confirmed that attrition was balanced across treatment and comparison groups, with no statistically significant differences (Appendix Table A1).As such, there is no reason to believe that any group would be different in the treatment group than in the control group.In our results section, we run robustness checks including only those individuals who appear in all three rounds of the study, and the results do not change significantly.Note.The percentages inside the parentheses are share of participants in each succeeding year against the baseline.For example, the 69% in the first column is from dividing 553 (in the first year) by 797 (from the baseline).

Estimation strategy
To estimate the effect of the PEI, we used a regression model with the following specification: where Y ft is the outcome variable corresponding to member of the family "f" (parent or child) in term "t"; "Post" is a variable set to one after the PEI was implemented (term one or two), "Treatment" is a variable set to one if the family "f" lives in any treatment community, Post*Treatment is an interaction aimed to capture the impact of the PEI; and ε b;t is the error term across families and time/terms.β 0 through β 3 are regression parameters to be estimated.As expected, a statistically significant β 3 parameter would imply differences between treatment and control communities in the outcome of interest explained by participation in the PEI (i.e., an intent-to-treat estimate).
A positive sign of the parameter would identify whether families in treatment communities positively changed parenting practices, or whether children improved their cognitive development according to the ASQ measures.We include community fixed effects and cluster standard errors at the community level.To control for potential false positive results as a result of testing multiple hypotheses, we include indices of our two major outcome groups: parenting practices and child development outcomes (Anderson, 2008).To construct the parenting practices indices, we re-orient the variables so that a positive change is a "good" outcome, standardize each variable (pooled across all waves), take the first principal component of a principal components analysis (Stein, 2019), and then standardize the resulting index. 10For child development, we sum the scores across the individual components of the ASQ, which is a standard form of analysis for that instrument and standardize the resulting sum.(We also include the standardized first component of a principal components analysis of the ASQ as well as the non-standardized simple sum, both as robustness checks).We complement our intent-to-treat analysis with treatment-on-the-treated analysis, in which a dummy variable for actual participation is instrumented with random assignment of the community to the program.The coefficients are somewhat larger with the treatment-on-the-treated analysis (as expected), and the pattern of significance is similar.The data and do-files required to replicate the analysis in this article are available at C'árdenas et al. (2023).

Implementation of the program
The vast majority of promotoras were female (90%), with an average age of 31 years.One-third of promotoras had some previous experience in teaching or promoting childcare activities, and just under three-quarters of promotoras actually attended the group training sessions organized by CONAFE.(Supervisors often organized individual training when promotoras were not able to attend group sessions).Among the promotoras without previous experience, 36% did not participate in the group training, and 92% of promotoras had access to at least one of the handbooks designed to support their activities.
Regarding PEI groups, on average 15 children and 15 caregivers were registered in each community.They usually met in facilities that the promotoras managed to get access to free of charge, including homes (33%), schools (10%), and other facilities.A common complaint of the promotoras in the focus groups we conducted was the inadequacy of the meeting spaces.
At each round of data collection, caregivers were asked how many meetings they had attended since the last interview (Figure 1 and Table A2).At baseline, the median caregiver in program communities had attended one meeting, consistent with the orientation meetings held in all communities to identify households' interest in the program.At the end of the first year, while caregivers had attended 11 meetings on average, the median number of meetings was much lower, at only four.A careful analysis of the data reveals that a significant number of households attended no meetings at all after the orientation meetings -the 25th percentile of caregivers attended zero meetings -but that another Figure 1.Caregiver attendance at meetings -baseline, first year, and second year.Notes.The bars present a histogram of the number of parent training sessions that caregivers attended that year.The line shows the percentile rank (i.e., at baseline, the median parent had attended one meeting; at the end of the first year, the median caregiver had attended 4 meetings (that year), and at the end of the second year, the median caregiver had attended 3 meetings (that year)).The numbers underlying this figure are in Appendix Table A2.
significant number reported very high attendance, at either 20 or 40 meetings (the 75th percentile reported attendance at 20 meetings).In the second year, the mean number of meetings fell from 11 to 8.7, and the median fell from 4 to 3. The 25th percentile remained at zero, and the 75th percentile fell from 20 to 12, suggesting that, overall, even with limited participation in the first year, it dropped off still further in the second year.
We also examine whether or not being in a community that was assigned to receive the program increased the probability that target parents would participate in the program -also known as the first-stage regression (Table 4).Households in communities assigned to the program were 74% more likely to participate in the program on average: 77% in the first year and 72% in the second year.

Parent behaviors
We begin with the intent-to-treat results for children age 0-36 months.After the first year of the program, the first principal component of an index of all nine observed practices reveals an improvement of 0.32 standard deviations, statistically significant at the 5% level (Table 5). 11At the end of the second year, the improvement in the index is roughly half the size but is no longer statistically significant.The treatment-on-the treated results are larger -with an index effect of 0.41 standard deviations, significant at the 5% level (Appendix Table A3).For older children, age 36-72 months, we observe no statistically significant differences in the index of observed caregiver behavior.
In terms of individual parenting practices for the 0-to 35-month-olds, which we mention only as suggestive areas for future research, almost all are positive (Table 5).In the first year, caregivers were statistically significantly 0.31 standard deviations more likely to respond to the child (i.e., if the child spoke to the caregiver during the interview), and in the second year, caregivers were significantly more likely to hug or kiss the child during the interview (0.27 standard deviations).These results suggest that the highest value of the program is at the earliest ages.Because some children fall in the younger group during the first follow-up and then in the older group during the second follow-up, we estimate a further specification, which uses only the seven caregiver practices that are observed across all ages and includes all children (Table 6).The observed practice index effects are smaller but positive and marginally statistically insignificant (10% level).They are slightly larger in the treatment-on-the-treated estimates, with similar significance patterns (Appendix Table A5).
There are many caregiver practices that cannot be observed directly during the course of a one-hour visit.Thus, we complement the earlier analysis with caregiver reports of behavior.For children age 0-35 months, we observe no significant change in the index of reported parenting practices (Table 7).In terms of individual practices (again mentioned as avenues for future research), three are statistically significant.Children in treatment households were 0.91 standard deviations more likely to have more than one book (1% significance) and watched significantly fewer hours of television on weekdays and on weekends.We observe similar patterns for the index and the individual elements in the treatment-on-the-treated estimates (Appendix Table A6).Reported parenting behaviors are similarly not statistically significant for older children (Appendix Table A7).

Child development outcomes
The impact of the program on child development outcomes is measured with the ASQ.A standardized total score across the components shows a positive but not statistically significant improvement (Table 8).We observe statistically significant increases in two child development areas in the first year: communication (0.15 standard deviations) and gross motor skills (0.16 standard deviations).While changes in other areas (fine motor skills, problem solving, or socio-emotional skills) are positive, they are smaller and not statistically significant.We observe a similar pattern of significance in the treatment-on-the-treated estimates (Appendix Table A8).
One might expect the impact of the program to change by intensity of treatment (i.e., parents and children who participate more in the program derive larger benefits).We do not have any exogenous variation in intensity of participation: families that participate more may be different from those who participate less in important ways -they may have more flexible schedules or they may put a higher value on children's investments more broadly.As such, this analysis is purely exploratory.Simple ordinary least squares regressions of the total child development score on whether households attended at least one session, whether they attended at least the median number of sessions that household who attended at least one session attended, or on the total number of sessions attended reveal no significant results (Appendix Table A9).
When we divide the analysis by gender, effect sizes are not statistically significantly different (Appendix Table A10); statistical significance disappears for both genders on communication, most likely due to a lack of statistical power, as the actual effect size is the same as in the joint analysis.The gross motor skill effect is larger for girls in year one than for boys.
Dividing our sample by child age, using children older and younger than 36 months, we observed that the child development effects are entirely concentrated among the younger group, the same children for whom caregiver practices improved (Appendix Table A11).If we instead divide the sample at 22 months (Appendix Table A12), as a test for sensitivity, the median age of children in the sample at baseline, we surprisingly see more significant impacts for communication among the older group, while gross motor improvements remain concentrated in the younger group.It may be that the communication skills are most affected by the program in the window between 22 and 36 months, when speaking tends to begin in earnest.

Survey effects in the child development measures
One concern with the results could be that, for those items that are reported by the caregiver, the intervention could lead to answers that are too high.For example, if a caregiver has learned in the program that children should be able to clap their hands at a certain age, then she might feel compelled to report that the child does indeed clap his hands all the time, even if that is not the case.To test for this, we first compare the informed values on the ASQ for three categories: behavior reported by parents, behavior observed directly over the course of the interview, or behavior reported by parents but then confirmed by the field researcher (Appendix Table A13).Items receive a value of 0 if the child never does the activity, 5 if the child sometimes does the activity, and 10 if the child does the activity "all the time."Our results show that parents reported significantly higher values than what interviewers observed.However, even if caregivers reported biased information, this principally biases the impact evaluation estimates if caregivers in treatment communities disproportionately report biased information.To explore this, we tested the effect of treatment on ASQ-measured areas separately for those items which are directly observed and for those items that are reported by the caregiver (Appendix Table A14).The coefficients on treatment are very similar across groups.For example, in year 1 communication, the treatment effect for items that were observed is 0.14, whereas the treatment effect for items that were reported is 0.11.The estimates are likewise similar for gross motor skills, the other area where significant treatment effects were identified in the initial analysis.
A second concern could be that observed caregiver actions were biased by the ASQ experience, since caregivers -upon seeing the administration of the ASQ -may be reminded of the good caregiver practices they learned about in the program sessions and then behave more affectionately during the interview.While that would be a program impact, it could signal a very local impact that only takes place when the caregiver is primed. 12In the baseline and the second follow-up, the ASQ was administered first in only 5% of cases.In the first follow-up, the ASQ was administered first in 21% of cases.Deviations from caregiver-first were driven by availability of the caregiver and of the child.If this priming effect indeed took place in the first follow-up, that could explain the positive significant caregiver behaviors in that round of data collection.To test for this, we regress our outcome variables on a dummy for the ASQ being administered first (Appendix Table A15).The administration of the ASQ first had no significant impact on any of the ASQ values or on parenting practices for younger children, where the significant impacts were concentrated.Surprisingly, we observe a negative impact for caregiver practices among older children: Treatment caregivers seem to behave more poorly with their older children if the ASQ was administered first.The reason for this is unclear, but if this is biasing the measured impact of caregiver practices, that means the null effects we observe for older children are biased downward; there may in fact be significant impacts that we are not capturing.
Third, we re-run our main analysis regressions including only those children who appear in all three rounds of data, and the results are comparable to those we have already discussed (Appendix Tables A16-A18).
Finally, the fact that we observe changes in the index of observed behaviors but not that of reported behaviors is further suggestive evidence that those results that are statistically significant are not merely an artifact of social desirability bias, in which respondents give those answers that they believe interviewers wish to hear.This bias could be exacerbated by an intervention that teaches caregivers how they should behave.But if that were the case, we would expect to see more significant effects among reported behaviors than observed behaviors, which is not the case here.

Indices
We use the first component of a principal components analysis for our main analysis of parent behaviors.We construct two alternative indices using an inverse covariance weighted average.One is the index proposed by Schwab et al. (2020); the other is proposed by Bouguen and Varejkova (2020).Both build on the work of Anderson (2008).We observe the same pattern of results across the different indices.While the magnitude of the effects varies, we observe a significant, positive impact in the first year on observed parenting practices for parents of children ages 0-3 (Table A19) and on observed parenting practices for parents of children ages 0-6 (Table A20).We do not observe significant impacts on reported parenting practices for children ages 0-3 (Table A21) or on the child development measure (Table A22) for any of the indices.

Discussion
This evaluation demonstrates that parenting practices can be affected -at least in the short run -by a low-cost intervention.Furthermore, despite the very limited caregiver engagement, the program had modest direct effects on child development.Unfortunately, we cannot disentangle whether effects on child behavior are the result of improved parenting practices, or because of the direct contact between children and promotoras during the meetings.However, it is notable that results for both caregivers and children are concentrated among the youngest children.
This program is relatively low cost.Our own estimates of implementation cost come to roughly US $113 per child per year (in 2016 dollars), including training, stipends and salaries, and materials.This is roughly 10% of the cost of comparable programs implemented in urban areas in Mexico (Myers et al., 2013).An analysis of other programs around Latin America found that a home visiting program in Jamaica cost US$313 per child per year, whereas community daycare in Chile cost US$984 (Myers, 2008).
Results from this evaluation are promising, particularly considering the limited take-up and the relative inexperience of the promotoras in these communities.If promotoras receive better training and more support, and if it were possible to increase parent participation, the effects could potentially be more sustained without increasing costs.One way to increase take-up would be to link this program to cash transfer programs already implemented, thus providing incentives to attend for busy parents who may underestimate the long-term returns to improvements in parent practices.
A version of the same program, when promoted within Mexico's longstanding conditional cash transfer program (called Prospera at that time), found positive impacts on child development (Fernald et al., 2017). 13In other words, the first arm of the Fernald et al. study may be seen as an independent evaluation of the same program that we evaluate (taking place at roughly the same time), whereas the second arm of the Fernald et al. study shows the potential of the same program when heavily promoted.In that arm, those responsible for managing the conditional cash transfer program also drew attention to the parent training program, encouraged mothers to participate, and scheduled meetings in such a way that mothers could attend both cash transfer meetings and parent training meetings.Although cash transfer payments were not dependent on participation in the parent training program, some local program implementers thought they were (Even in those cases, cash payments were not withheld from non-participants in the parent training program).The fact that the Fernald et al. study only identified consistent impacts when the parent training was heavily promoted by another program with more immediate impacts on household budgets suggests a potential pathway for supporting both short-term needs (cash) and longer-term investments (parenting skills).
An analysis of the pathways of impact in the intervention evaluated by Fernald et al. suggests that up to one third of the child development impacts may have been driven by improved parenting practices, such as increased play activities and book reading (Knauer et al., 2016).Even with the modest cost increase that greater promotion might entail, the program would likely still remain a cost-effective option for rural populations.

Notes
1.The COVID pandemic led to school closures across most countries in 2020 and 2021, further increasing time that children spent with parents -including children who would normally have been in early childhood education programs (Anderson et al., 2022) -and increasing the potential return to improved parenting skills.While global pandemics are infrequent to date, individual countries regularly experience shocks that result in school closures (Angrist et al., 2023).2. The Fernald et al. (2017) and Knauer et al. (2016) studies -both stemming from the same evaluation -took place in Chiapas, Puebla, and Oaxaca states, whereas this study takes place in Chiapas, Estado de México, Oaxaca, Puebla, Querétaro, and Veracruz.3. The main goal is to develop among children the following competencies: language and communication, healthy practices, hygiene and nutrition, interactions with others, executive function, personal and social skills, body control, fine and gross motor skills, and representation, among others.4. We initially intended to evaluate the impact of these men's meetings on the quality of fathers' parenting, but take-up of the men's meetings was almost zero across all treatment communities, which made such an evaluation of impact impossible.A recent review found that only three percent of studies of early childhood development programs estimate impacts on fathers (Evans et al., 2021); low take-up, as observed in our program, may be part of the explanation.5. Based on the number of sites, considering a significance level of 0.05, a correlation coefficient of 0.05 and the minimum number of families to be interviewed, we reached a power of 0.80 with a minimum detectable effect (MDE) size of 0.18 (0.20 if the correlation coefficient = 0.10).Under this estimation, all the communities included in the sample had at least 8 families with at least one child per family between 0 and 42 months old.In addition, none of these communities should have been part of any ECCE program, either municipal, state, or federal, since 2007.6.We paired communities using measures of poverty, population 0 to 5 years old, population 3 years and older, indigenous language speakers, nonliterate population, population older than 15 years who abandoned primary school, population older than 15 years who abandoned lower secondary school, working population and employed, population without access to health services, population without access to other social services, number of households, number of household members, and total population in each locality.7. By implementing this validation, we addressed the problem of self-selection, given that we requested that communities convene and approve the potential participation in the PEI without informing them of whether they would be part of the treatment or the control group (in addition, no public official from CONAFE knew at this point the result of the random assignment already conducted).8.We consequently considered power calculations and changes in the MDE based on the final number of interviewed families.For the 126 communities included in the final sample, with roughly 12 households each, assuming a statistical significance of 95% and an intra-cluster correlation of 0.05, we needed a modest effect of 0.20 in order to achieve a statistical power of 0.80.Doubling the intra-cluster correlation to 0.10 had very little impact to a minimum detectable effect of 0.22.Note that the higher number of households (i.e., 12 households) interviewed in comparison to what was anticipated for the original sample (i.e., 8 households) reduce concerns about power and MDE.All sample sizes were calculated using Optimal Design software.9. Twenty-two of the sampled communities required interpreters for six languages other than Spanish (13 in Chiapas, and 9 in Oaxaca).10.The principal component captures a different subset of variation than an inverse covariance weighted average.The latter puts more weight on components that offer more new information, whereas the first simply tries to capture the most variation in the data (Samii, 2017).We focus on the former, but we provide evidence from the latter as well.11. Figure 2 provides an overview of all the results; The index effect in Table 5 is larger than the individual components, which may -for example -come about if uncorrelated classical measurement error biases the individual component effects toward zero but those errors average closer to zero in the index.12. Interviewers were instructed, in general, to implement the caregiver questionnaire first, in order to build trust with caregivers before interacting directly with the children during administration of the ASQ.13.Prospera was canceled by the Mexican government in 2019.The first column shows the estimates for the index which is also used in Table 6.Community fixed effects included.Standard errors clustered at community level.***p < .01,**p < .05,*p < .1.PCA is an index consisting of the first principal component from a principal components analysis.The ICW index is an inverse covariance weighted average constructed according to Bouguen and Varejkova (2020), which we also standardize.The SW index is an inverse covariance weighted average constructed according to Schwab et al. (2020).The index in the first column is constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the index itself is standardized (same as used in Table 5).The ICW index is an inverse covariance weighted average constructed according to Bouguen and Varejkova (2020), which we also standardize.The SW index is an inverse covariance weighted average constructed according to Schwab et al. (2020) The first column shows the estimates for the index, which is also used in Table 8.The second column, labeled "ASQ Total," simply sums the different components of the ASQ instrument.PCA is an index consisting of the first principal component from a principal components analysis which we then standardize by subtracting from the mean and dividing by the standard deviation.The ICW index is an inverse covariance weighted average constructed according to Bouguen and Varejkova (2020), which we also standardize.The SW index is an inverse covariance weighted average constructed according to Schwab et al. (2020) and is also standardized by default when computed using Stata.The first column shows the estimates for the index which is also used in Table 7. Community fixed effects included.Standard errors clustered at community level.***p < .01,**p < .05,*p < .1.PCA is an index consisting of the first principal component from a principal components analysis.The ICW index is an inverse covariance weighted average constructed according to Bouguen and Varejkova (2020), which we also standardize.The SW index is an inverse covariance weighted average constructed according to Schwab et al. (2020).

Figure 2 .
Figure 2. Coefficients from the main outcome indices.Note.The indices shown in this forest plot are the indices reported in the first columns of Tables 5-8.Children ages 0-3 years are 0-35 months.Children ages 0-6 years are 0-72 months.

Table 1 .
Descriptive statistics of the study population (at baseline).

Home assets -percentage of houses/households with
*For age, we dropped a few observations with less likely ages (e.g., under ten years.)**Secondary includes upper secondary education.
Note.ASQ references the Ages and Stages Questionnaire (ASQ-3).All characteristics have been standardized to have a mean of 0 and standard error of 1 for comparability.All errors are clustered at the community level.Robust standard errors clustered at the community-level in parentheses.***p < .01,**p < .05,*p < .1.
Note.All individual practices are standardized.The index in the first column is constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the index itself is standardized (see section 3.4 for more details on the estimation strategy).Community fixed effects included.Standard errors clustered at community level.***p < .01,**p < .05,*p < .1.
Note.All individual practices are standardized.The index in the first column is constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the index itself is standardized (see Section 3.4 for more details on the estimation strategy).Community fixed effects included.Standard errors clustered at community level.***p < .01,**p < .05,*p < .1.
The index in the first column is constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the index itself is standardized (see Section 3.4 for more details on the estimation strategy).Community fixed effects included.Standard errors clustered at community level.

Table 8 .
Impact on child development (full sample of children, age 0-72 months).All individual practices are standardized.The index in the first column is the simple sum of the standardized parenting practices, which we then standardized.Community fixed effects included.Standard errors clustered at community level.***p < .01,**p < .05,*p < .1.

Table A3 .
Impact on observed parenting practices -9 practices + an index -age 0-35 months (treatment-on-the-treated estimates).All individual practices are standardized.The index in the first column is constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the index itself is standardized (see Section 3.4 for more details on the estimation strategy).Community fixed effects included.Standard errors clustered at community level. *

Table A4 .
Impact on observed parenting practices -age 36-72 months.The index in the first column is constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the index itself is standardized (see Section 3.4 for more details on the estimation strategy).Community fixed effects included.Standard errors clustered at community level.

Table A5 .
Impact on observed parenting practices -7 practices + an index -age 0-72 months (treatment-on-the-treated estimates).All individual practices are standardized.The index in the first column is constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the index itself is standardized (see Section 3.4 for more details on the estimation strategy).Community fixed effects included.Standard errors clustered at community level.***p < .01,**p < .05,*p < .1.

Table A6 .
(Continued).The index in the first column is constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the index itself is standardized (see Section 3.4 for more details on the estimation strategy).Community fixed effects included.Standard errors clustered at community level.

Table A7 .
(Continued).The index in the first column is constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the index itself is standardized (see Section 3.4 for more details on the estimation strategy).Community fixed effects included.Standard errors clustered at community level.***p

Table A8 .
Impact on child development (full sample of children, age 0-72 months) (treatment-on-the-treated estimates).

Table A9 .
Impact on child development (full sample of children, age 0-72 months) by intensity of participation.

Table A10 .
Heterogeneous effects on child development based on gender of child.All individual practices are standardized.The index in the first column is the simple sum of the standardized parenting practices which we then standardized.Community fixed effects included.Standard errors clustered at community level.***p < .01,**p < .05,*p < 0.

Table A11 .
Heterogeneous effects on child development based on baseline age of child (divided at 36 months).

Table A13 .
Child development statistics of different ASQ implementation modes.

Table A12 .
Heterogeneous effects on child development based on baseline age of child (divided at 22 months).All individual practices are standardized.The index in the first column is the simple sum of the standardized parenting practices which we then standardized.Community fixed effects included.Standard errors clustered at community level.***p < .01,**p < .05,*p < 0.

Table A14 .
Regression analysis of ASQ implementation modes.The indices in the first two columns are the simple sum of the standardized parenting practices, which we then standardized.Community fixed effects included.Standard errors clustered at community level.***p < .01,**p < .05,*p < .1.

Table A17 .
Impact on observed parenting practices for caregivers who participated in all three waves − 9 practices + an index -ages 0-35 months.
All individual practices are standardized.The index in the first column is constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the index itself is standardized (see Section 3.4 for more details on the estimation strategy).Community fixed effects included.Standard errors clustered at community level.***p < .01,**p < .05,*p < .1.

Table A15 .
Regression analysis of ASQ and caregiver questionnaire implementation order.The indices in the first two columns are constructed from the first principal component of a principal components analysis (PCA) of the standardized parenting practices and the indices themselves are standardized (see Section 3.4 for more details on the estimation strategy).The individual development scores are standardized.***p < .01,**p < .05,*p < .1.

Table A16 .
Impact on child development for children who participated in all three waves.

Table A18 .
Impact on observed parenting practices for caregivers who participated in all three waves − 12 practices + an index -ages 36-72 months.