Conditional cash transfers, uptake of maternal and child health services, and health outcomes in western rural China: A cluster controlled design

Background: Empirical evidence suggests that the uptake of maternal and child health (MCH) services is still low in poor rural areas of China. There is concern that low uptake of MCH services may detrimentally affect child health outcomes. Previous studies have not yet identified the exact nature of the impact that conditional cash transfer (CCT) have on the uptake of MCH services and ultimately, on child health outcomes. The objective of this study is to examine the relationship between CCT, uptake of MCH services and health outcomes among children in poor rural areas of western China. Methods: We designated two different sets of comparison villages and households that were used as comparison against which outcomes of the treated households could be assessed. We conducted a large-scale survey of 1,522 households at 75 villages (including 25 treatment and 50 comparison) from nine nationally-designated poverty counties in two provinces of China in 2014. In each village, 21 households were selected based on their eligibility status for the CCT program. Propensity score matching analyses were used to assess the impact of CCT on outcomes in terms of both Intention-to-treat (ITT) and average-treatment-effects-on-the-treated (ATT). Results: Overall, the uptake of MCH services in the sample households were low, especially in terms of post-partum care visit, early breast feeding, exclusive breast feeding, and physical examination of the baby. The uptake of the seven types of MCH services in the CCT treatment villages was significantly higher than that in the comparison villages. Results from both the ITT and ATT analyses showed the CCT program had a positive, although small, impact on the uptake of MCH services and the knowledge of mothers about MCH health issues. Nonetheless, the CCT program had no noticeable effect on child health outcomes. Conclusions: The CCT program generated modest improvements in the uptake of MCH services and mothers’ knowledge of MCH services in poor rural areas of Western China. These improvements, however, did not translate into substantial improvements in child health outcomes for two possible reasons: poor CCT implementation and the low quality of the rural health facilities. logistic

Development Goals [1]. The Chinese government has made great progress in maternal and child health (MCH) across large parts of the country in recent years by aggressively expanding the coverage of rural health insurance and promoting maternal delivery in hospitals [2][3][4]. Nevertheless, there is concern about MCH in western China's poor rural areas, where up to 40% of women do not receive prenatal physical examinations, and the rate of maternal delivery in hospitalswas only around 70% in 2008 [5][6][7]. Despite the launch of China's Basic Public Health Service programs in 2009, through which all basic maternal health care is supposed to be free, this low uptake of MCH services persists.
Although one reason for the low uptake of MCH services in these areas may be supply-side challenges, such as poor quality of health services and the attitudes of doctors [8,9], a less-studied aspect of the continuing challenge of promoting higher uptake of MCH services is demand-side factors. The limited research that has been published on this topic shows that the level of a woman's education, household annual income, and cost of both hospital delivery and transportation to the hospital are important factors that correlate with the uptake of prenatal physical examinations and hospital deliveries in rural areas [10][11][12]. As such, demand-side factors seem to be an important source of the differences in the uptake of MCH services among subpopulations, possibly even more important than supply-side factors [13].
Conditional cash transfer (CCT) programs are an increasingly popular method of improving participation in education and health services in developing and developed countries [14][15][16]. In their most basic form, CCT programs seek to overcome demand-side constraints by providing cash payments to poor households for behaving in a certain pre-defined and socially-responsible way [17].
A meta-study on the uptake and impact of CCTs on health-related outcomes indicated that CCT programs in Africa and Latin America were generally highly effective in promoting increased access to health services and better health outcomes [18,19]. There are, however, cases in the literature in which CCT programs did not work [20,21]. Given that CCT programs have both succeeded and failed in other developing countries, we are interested if a CCT will succeed or fail in the context of poor, rural China.
Previous research conducted within China suggests that CCT programs may be effective in improving MCH uptake in rural areas in China. As stated above, CCT provides cash payments as a way to incentivize certain socially-responsible behaviors. From our previous studies, it seems that this cash incentive should be highly effective-when women's representatives were asked about the biggest obstacle for increasing MCH uptake, the most common answer was that costs were too high.
Additionally, all surveyed village women's representatives believed that paying women to uptake MCH services-most likely through a CCT program-would be an effective way to increase uptake [13]. A CCT program could therefore be an effective way to address these demand-side issues as these cash transfers would not only resolve the main obstacle preventing women from using MCH services, but also may provide a strong, positive incentive for women to utilize MCH services. We therefore suspect that a CCT program has the potential to be very successful in rural China.
With the goal of improving the uptake of MCH services in poor rural areas, China's government, with support from UNICEF, launched a CCT pilot program in 2013 in which pregnant women and mothers with infants were incentivized with cash payments to utilize government-provided MCH services. To the best of our knowledge, this program is the first CCT initiative in the health field in China.
Therefore, there has not yet been a rigorous impact evaluation of the effectiveness of a CCT program in China's health field.
Our overall goal is to measure the effects of China's first CCT program on three sets of outcomes: MCH service utilization, mother's knowledge of MCH issues, and child health outcomes. Ultimately, the goal is to assess whether CCT programs can improve MCH service utilization, knowledge of MCH issues, and health outcomes among the poor western rural population in China.

Methods
The MCH CCT program The premise of the CCT intervention is simple. In a series of group meetings between eligible participants in a selected number of project villages in designated project townships, pregnant women and mothers-to-be (both henceforth collectively referred to as mothers) were advised that, if they undertook any one of a set of seven MCH services-of which, five are free-then they would not only have the benefits of the service that was provided, they would also receive a cash payment of a certain amount. The program was designed to make payments to the participants shortly after completion of each visit to the MCH service provider.
The list of services and the payment schedule were displayed prominently in all treatment townships.
Eligible women were told that they would receive a separate payment each time they (a) underwent a prenatal examination; (b) delivered their baby in the hospital; (c) underwent a maternal postpartum examination; (d) engaged in early breastfeeding, that is, began breastfeeding after delivery within one hour; (e) breastfed exclusively for six months; (f) gave their child all required vaccinations; or (g) took their child to a child health examination. If a mother in a CCT project township completed all CCT activities, she would receive about 1,000 Chinese yuan (equivalent to 154 US dollars). This is a relatively large sum of money in the study area, given that the average annual income in 2013 was approximately 1,500 yuan per capita [7].

Study sample
For the purpose of the evaluation, we received access to 25 CCT townships, including 9 townships from four counties in Gansu and 16 townships from five counties in Sichuan, an average of about three townships per county. Fig.1 depicts the location of the sample townships, counties, and provinces.
We designated two different sets of comparison towns, villages, and households that were used as comparison groups against which outcomes of the treated households could be assessed. To assess the comparability of the treatment and comparison groups, we assembled a list of variables. These include: ethnicity, as measured by the share of Han; village population; nature of the local township road, measured by the presence of a paved road from the town to the village; share of families who were receiving income support or welfare; and time that it would take to travel from the village office to the township health center.
Our sampling frame worked as follows. First, we went to the nine CCT project counties from the Sichuan and Gansu provinces. From the five counties in Sichuan, we went to 16 towns that were offering the CCT program (treatment towns), and 32 that were not (comparison towns). From the four counties in Gansu, we went to nine CCT towns and 18 non-CCT towns.
To choose the sample villages and households, we followed a pre-specified protocol that consisted of four steps. The first step was to take a between-village matching strategy. To do so, we randomly chose a set of treatment villages from among the 25 treatment towns that would contain households eligible to participate in the CCT program. After the treatment villages were randomly chosen, the second step involved choosing a set of comparison villages from among the 50 comparison towns. To improve the probability of having a good match, we chose two comparison villages for each treatment village. The assumption of our sampling strategy was that the two comparison villages, by nature of their proximity to the treatment villages, were likely to be close matches.
To select the comparison villages, we used secondary township-level and village-level statistics.
Utilizing all available variables noted above for each village in each town, we identified one village from within each of the two comparison towns that were similar to our treatment villages. Specifically, the comparison villages were statistically matched to the treatment villages so that each of the two comparison villages was statistically similar to the treatment village, to as great a degree as possible, on the different relevant township-level and village-level characteristics.
The sample included 25 treatment villages and 50 comparison villages, for a total of 75 study villages.
The sample villages were selected, and the overall sampling protocol was implemented in the fall of 2014. The timing of the sample selection was carried out 18 months after the launch of the CCT program, which was spring 2013.
For the third step of the sample selection protocol, we chose study households in the treatment and comparison villages. The goal was to choose three different types of households. The first type was termed Fully Eligible households (FE households), meaning that the mother became pregnant after the implementation of the CCT program in the treatment villages. This means that she would have been able to take advantage of the full services of the program. The second type was termed Partially Eligible households (PE households), wherein the mother became pregnant prior to the launch of the CCT program but did not deliver her child until after its launch in the treatment villages. This means she would have been able to access some, but not all, of the services of the program. The third type was termed Ineligible households (IE households), meaning that the mother both became pregnant and delivered her child prior to the launch of the CCT program. This would have barred her from being able to access its services even though she was in a CCT treatment village.
To select these households, we went to each village and consulted the roster of all babies from the village doctor or the women's representative. We grouped the babies into the three types based on their dates of birth. Within each type of household, we randomly selected seven babies and their mothers to become our sample households. As such, we selected 21  In total, the sample included 25 treatment villages and 50 comparison villages, for a total of 75 study villages. In addition, 21 households in each target village were selected based on their eligibility status for the CCT program. The assessment profile of this CCT program is depicted in Fig.2.

Data collection
The research team conducted the survey in October 2014 in Sichuan Province and November 2014 in Gansu Province. The survey comprised five modules that were designed to meet our objectives of measuring the impact of the CCT program on the outcomes of interest: uptake of MCH services; knowledge of the mother about MCH issues; health outcomes of the child; participation in and receipt of the CCT payments in the CCT eligible households in the treatment villages; and information on individual and family characteristics that we used as control variables in the analysis.
The first module involved the collection of information needed to assess the uptake of MCH services.
This included asking whether the caregiver of the child had participated in antenatal check-ups, inhospital delivery, post-partum checkups, child physical checkups, or child vaccinations. Using our data, we constructed seven measures of the uptake of MCH services: made any antenatal care visit (1 = yes, 0 = no), baby delivered in hospital (1 = yes, 0 = no), made any postpartum care visit (1 = yes, 0 = no), early breastfeeding (1 = yes, 0 = no), exclusive breastfeeding (1 = yes, 0 = no), compliance rate of child physical examinations (%), and compliance rate of child vaccinations (%). The compliance rates of child physical examinations and vaccinations were calculated based on the requirements of the national standards of basic public health services, which has a schedule of when to take a child for physical examinations and vaccinations by child age from birth to 6 years old. Therefore, the compliance rate is the actual frequency of physical examination and vaccination divided by the frequency of the national standards.
The second module concerned the assessment of the mother's knowledge of the MCH services that she was asked to take her child to, as well as knowledge of infant nutrition. In this knowledge scale, there was a total of 22 items, scored by giving the respondent one point for each correct answer; a mother having complete knowledge would score 22 points. Appendix A1 provides an English translation of the knowledge test [see Additional file]. Using our data, we constructed five measures of mother's knowledge about MCH services: total score on the 22-item knowledge test (full = 22 points), at least 60 percent of the 22-item knowledge test correct (1 = yes, 0 = no), score on items related to maternal care (full = 8 points), score on items related to child nutrition (full = 6 points), and thinking child physical exam necessary (1 = yes, 0 = no).
In the third module, each sample mother and child in both the treatment and comparison groups received physical examinations. Trained nurses, as part of the survey team, collected data on three indicators of health outcomes for each child and three indicators for each mother. The three measures included hemoglobin concentrations, height, and weight. Hemoglobin levels were measured using HemoCue Hb 201+ systems (HemoCue Inc., Angelholm, Sweden). Following international standards for our sample age group, we defined anemia as a hemoglobin count of less than 110 grams per liter [22]. Height and weight measurements were obtained following the World Health Organization (WHO) standard protocol. The children were measured in light clothing without shoes, hats, or accessories.
Height was measured using a standard tape measure. Weight was measured with a calibrated electronic scale (Tanita, HD-388, Japan). The nursing team was trained to set up the weighing station on level ground to ensure accuracy of the equipment. The anthropometric data were used to develop standard indicators of child development, such as length-for-age Z-scores (LAZ) and weight-for-length Z-scores (WLZ), based on international standards [23].Using our data, we constructed four measures of child health outcomes, following WHO guidelines: low birth weight (1 = less than 2,500 g, 0 = 2,500 g or more), anemia (1 = hemoglobin less than 11.0 g/dl, 0 = 11.0 g/dl or more), stunted growth (1 = LAZ less than -2 standard deviations, 0 = -2 or more), and wasting (1 = WLZ less than -2 standard deviations, 0 = -2 or more).
Although all other modules were administered to all sample households regardless of their eligibility status, the fourth was not. For households in the treatment groups, namely, FE and PE treatment households, we had one extra module in which enumerators asked detailed information about their participation in the CCT program. Specifically, enumerators asked whether the mother registered for the program. Mothers were also asked to report the amount of cash that they had received for participating in the CCT MCH activities.
The fifth module of the survey was designed for the collection of information on various factors, statistically, controls, that might directly or indirectly affect the uptake of MCH services or health outcomes. The survey contained items for mothers about their child's age, gender, ethnicity, gestational age, and pregnancy order. Enumerators also quizzed mothers about their own characteristics, including age, education, ethnicity, and occupation. A final set of items concerned overall household characteristics, including distance from home to the township health center, in terms of kilometers and travel time, and the nature of each household's durable assets.

Statistical Analysis
A power analysis was conducted based on the one of the main outcomes: the rates of hospital delivery. With the power of 0.8 to detect a difference in hospital delivery rates between the treatment and control groups in a cluster controlled trial, a suitable sample size depends on number of children per village, number of villages, probability of hospital delivery in treatment villages and controlled villages and 95% plausible interval. According to our study design with 1 CCT village and 2 control villages, the total number of villages was 75. Based on previous studies [24], we assumed the hospital delivery rates were 50% in control villages and 60% in treatment villages. We then assumed a 95% plausible interval of 0.4 to 0.75. On the basis of these parameters, we calculated that we required 18 women per village. Considering the possible sample loss and assumed impacts, we added 3 women to each village to overpower the study when the budget allowed.
All statistical analyses were performed using STATA 12.0 (StataCorp, College Station, Texas, USA); pvalues below 0.05 were considered statistically significant. We reported coefficients and 95% confidence intervals (CIs) for all main variables of interest. Comparisons between the treatment and comparison groups for all outcomes by subgroup populations were assessed using t-tests or chisquare tests.
To examine the effect of the CCT program on the uptake and knowledge of MCH services as well as on child health status, the evaluation used two dimensions of variation, i.e., between-village analysis (Evaluation Strategy 1) and within-village analysis (Evaluation Strategy 2). The first is cross-sectional and comes from comparing households with the same eligibility but from villages with different treatment status, namely, CCT treatment villages versus non-CCT comparison villages, utilizing Evaluation Strategy 1. Under Strategy 1, we estimated the impact of the CCT program, using the following least squares regressions model: Y i is the outcome of interest for household i, including uptake of MCH services, knowledge of MCH services, and health status of children; CCT i is a dummy variable that indicates whether a household comes from a CCT village, which makes β the parameter of interest; and X i is a vector of covariates that are included to capture the characteristics of children, mothers, and households. In all cases, we adjusted standard errors for clustering at the township level, using a cluster-corrected estimator.
The second dimension of variation is temporal and comes from comparing households that are fully or partially eligible (FE/PE households) against those households that are ineligible (IE households) for this CCT program under Evaluation Strategy 2. In this evaluation strategy, we estimated the impact of the CCT program, using the following least squares regressions model: Note that the only difference between equation (2) and equation (1) is that we replaced CCT i with the dummy variable Eligibility i, indicating whether a household is fully or partially eligible (FE/PE households). The rest of the variables are the same as described in equation (1). Together, Strategies 1 and 2 consist of comparing households whose children were born at different times-before, during, or after the launch of the CCT program-and by CCT status. The CCT program can be considered the treatment, and our sample households were divided into two treatment groups and a comparison/control group. The treatment groups include (a) the FE households in the CCT villages and (b) the PE households in the CCT villages. The comparison group includes all of the households in the non-CCT comparison villages (FE, PE, and IE households) as well as the IE households in the CCT villages.
We supplemented our intention-to-treat (ITT) multivariable analysis described above by examining the average-treatment-effects-on-the-treated (ATT analysis) to measure the impact on outcomes among the subpopulation of households who had heard about the CCT program. This allowed us to control for any confounding due to non-compliance, which we define as usage of MCH services without receiving a monetary transfer. For the ATT analysis, we utilized an instrumental variable (IV) approach [25], in which the treatment assignment (receiving CCT information or not) was used to account for observed compliance, or receiving a monetary transfer for using MCH services. This analysis is based on the assumption that the only reason for a woman in a CCT village to not receive a monetary transfer for using an MCH service is because she was unaware of the CCT program. The IV approach allows us to measure the average effect of treatment on the use of MCH services, mother's knowledge, and child health outcomes among the subpopulation of households that knew about the program and, thus, control for confounding due to non-compliance. The ATT analyses for the continuous outcome measures were performed using STATA's ivreg model. The ATT analyses for the binary outcome measures were performed using STATA's ivprobit model. In estimating both models, we clustered the standard errors at the village level.

Participants' characteristics
A total of 1,522 households (mother-baby pairs) were enrolled in our study. In total, there were 503 households in the 25 treatment villages. Of these, 349 were treatment households (174 fully eligible households, 175 partially eligible households) and 154 were comparison households (ineligible households). Likewise, there were 1,019 households in the 50 comparison villages (non-CCT villages).
Of these, 353 were FE (fully eligible) households, 339 were PE (partially eligible) households, and 327 were IE (ineligible) households. Table 1 presents the results for the observable characteristics of the children, mothers, and households across different treatment and comparison groups. As seen in Table 1, the mean age of all of the children in the study was 12.3 months (SD = 5.9), and 47.7% of the children were girls. The mean age of the mothers was 27.7 years. Nearly 70% of the mothers were of non-Han ethnicity, and only 25% of the mothers received a junior high school or above level of schooling. Out of the total number of households, 11.8% needed more than one hour to travel to the township health centers.
For the integrity of the evaluation, the observable characteristics across treatment and comparison households are balanced (Columns 2 to 10). In the case of all of the individual variables, the p-values were above 0.05.
In the rest of this section, we first report results from the descriptive analysis that did not adjust for any covariates (Table 2). Next, we report results from the ITT analysis, adjusted for covariates, including characteristics of the children, mothers, and households (Table 3). Both these analysis results are part of Evaluation Strategy 1. To show the robustness of the analysis, Tables 4 and 5 report the descriptive and then the ITT analysis for Evaluation Strategy 2. We also report results from the ATT analysis that examined the impact of the intervention after accounting for compliance in Table 6.
Comparisons between CCT and non-CCT villages Uptake of MCH services FE households. When we compared FE households in the CCT and non-CCT villages, two out of the seven uptake measures were significant ( Table 2; Rows 1 and 6; Columns 1, 4, and 7: both p-values were less than 0.05). Specifically, more women in the FE households in the CCT villages made at least one prenatal care visit (85%) when compared to those in the comparison villages (78%). Although the differences were significant (p = 0.04), the magnitude of the difference was modest (7%). Mothers in the FE households in the CCT villages also took their children for post-natal physical examinations at higher rates (33%) than those in the comparison villages (23%). As in the case of prenatal visits, although the difference was significant (p = 0.01), the size of the difference was relatively small (10%). In the case of the other program health care activities-delivery in hospital, postpartum care visits, early breastfeeding, exclusive breastfeeding, and compliance with child vaccinations-there were no statistically significant differences in the rates of participation between the treatment and comparison villages.
A less secure impact was found when the ITT model was applied to Evaluation Strategy 1 for FE mothers (Table 3; Rows 1-7; Columns 1-3). Although all of the point estimates were positive, all of the 95% confidence intervals overlapped zero. The magnitudes of the point estimates were mostly small.
In sum, after holding all of the control variables constant except for compliance (which we do below in the ATT model), the CCT intervention did not have a large impact on the uptake of MCH services. PE households. When we compared PE households in CCT and non-CCT villages, using both the descriptive and ITT analyses, the overall results are similar to those for FE mothers. The data suggested that the impact of the CCT program was modest for women who were only partially eligible (PE mothers). The descriptive data showed that two out of the seven uptake measures were significant ( Table 2

Impacts on Mother's Knowledge
One of the main findings derived from our descriptive statistics was that the knowledge of MCH issues is low among women in the sample areas in general, regardless of being in a treatment or comparison group ( Table 2; Rows 8 to 12). When asked questions after the program had been in operation for more than one year, women inside and outside of the program answered fewer than half the questions on MCH issues correctly. The respondents got only slightly over half of the answers correct on health issues for the mother and less than half correct on issues of child nutrition. Fewer than 20% of the women believed that child health examinations are necessary for their children's health.
Despite the overall low knowledge of MCH issues in our sample, our results indicate that the knowledge of women improved in CCT treatment villages. When comparing either FE or PE women in CCT treatment villages with their counterparts in the comparison villages, using descriptive statistics, there were statistically significant differences between a number of measures (Table 2; Rows 8-12; Columns 7 and 8).However, using either the descriptive statistics or the ITT approach, the team found that, when they estimated the impact of the CCT program on mothers' knowledge about MCH services, there was only a small effect ( Table 2 and Table 3; Rows 8-12). The impact of the CCT program on the knowledge of the mothers is therefore similar to the impact of the CCT program on MCH service uptake, in that there was a statistically significant, but small impact.
Two observations moderate any attempts to claim that the CCT program did any more than modestly improve the knowledge of the women participants. First, in the descriptive analysis, the magnitudes of the differences between the FE and PE respondents in CCT treatment and non-CCT comparison villages were small. At most, the CCT treatment village women improved their knowledge by 1 point (out of 22 for the overall knowledge scale). The measured differences in magnitudes for all of the other analyses were even smaller. Second, for the ITT analysis, which held equal the characteristics of the child, mother, and family, most of the measured impacts of the CCT intervention on either FE or PE women in the treatment villages disappeared. In other words, a number of the differences that were statistically significant in the descriptive analysis became insignificant in the ITT analysis.

Impacts on Health Outcomes
The effect of the CCT program on the health outcomes of the children in the treatment groups compared with those in the comparison groups based on our impact evaluation analysis is easy to summarize. There was almost no effect on incidences of low birth weight, anemia, stunted growth, nor wasting. The descriptive analysis, using Evaluation Strategy 1, found no improvement on any of the outcome measures between either the FE children or PE children in the treatment and comparison villages ( Table 2; Rows 13-16; Columns 7 and 8). Likewise, the ITT analysis did not detect any improvements for any of the outcomes when considering the effect of the CCT intervention on FE or PE children in the treatment villages compared to those in the comparison villages.

Robustness of Results to Alternative Evaluation Strategy
Using Evaluation Strategy 2, we examined the impact of the CCT program within CCT villages by comparing FE/PE and IE households. With both the descriptive and ITT analyses, we found that there was only a very modest effect on MCH service utilization, no differences in maternal knowledge, and almost no positive effect of the CCT program on health outcomes of children (Tables 4 and 5).

Treatment compliance and the results of the ATT analysis
In CCT villages, out of 174 sample FE households that were included in the descriptive and ITT analyses (reported on above), only 55.8% reported that they had heard about the CCT program. Out of the 175 sample PE households in CCT villages, only 62.3% reported that they had heard about the program. In total, 59.0% of women (both FE and PE women) in CCT villages said that they had been informed about the CCT program.Of the women who knew about the CCT program, the proportion who used MCH services and received a monetary transfer is even lower. The overall compliance rate of the CCT program was only 49.9%. Specifically, only 51.1% of women from FE households got at least one type of MCH services and also received a cash transfer. In the PE households, only 48.6% of women complied.
The results of the ATT analysis (using either evaluation strategy) mirrored the results of the ITT analyses. The CCT intervention had only a small impact on the uptake of MCH services and on improving mother's knowledge. There was even a smaller (indeed, most likely zero) impact on improving child health outcomes (Table 6).

Discussion
In this paper, we utilized rigorous impact evaluation approaches on a large sample of women who participated (and did not participate) in a pilot CCT program that sought to incentivize new mothers in two poor regions in Western China to improve their utilization of MCH services, upgrade their knowledge, and, ultimately, raise health outcomes. In this area, the utilization of MCH services and mother's knowledge are poor overall. Health outcomes of children are also fairly poor. The purpose of the program was to increase the benefit of utilizing MCH services, and this was expected to lead to better knowledge about child health and better child health outcomes.
The data, however, showed limited impact of this pilot program. Regardless of the evaluation strategy or the nature of the treatment and control groups, the rate of participation in MCH services did not rise systematically. Mother's knowledge improved only marginally. Further, health outcomes did not show any sign of improvement. Indeed, throughout the results-both the descriptive statistics and the multivariate analysis-there were instances in which there were significant differences between the treatment mothers/children in CCT villages and the control mothers/children in non-CCT villages.
Nevertheless, in a large proportion of the statistically significant results, the magnitudes of the shifts in outcomes (MCH service utilization, mother's knowledge, or child health outcomes) were small. This is to say, the CCT program technically worked, but the effects were so small that it could hardly be considered effective.
In this respect, this pilot project's relatively minor outcomes differ from those of similar programs implemented in other countries. In other developing countries in which CCT programs have been used to help poor women overcome barriers to MCH services, many programs have been shown to improve health. For example, CCT programs have been shown to improve birth weight [14], decrease anemic rates [26], and aid child growth in Mexico [17]. Successful programs in Brazil have reduced overall infant mortality rates [27]. Programs in Nicaragua [28], Ecuador [29], and Colombia [30] have succeeded in increasing child height.
Consistent with the results of this study, however, there have also been CCT programs that did not work. An assessment of a CCT program in Honduras found no effect of such programs on child outcomes [20]. In the Honduran study, it appeared that the poor quality of the implementation was largely to blame. Another CCT program in Kenya also found no improvement in care. In this case, the poor quality of the health care system was at least part of the reason that no improvements in health outcomes were found [21].
In reflecting on the literature, we questioned why the experience with the CCT program in China more resembles the CCT programs of the Honduras and Kenya (which did not have an impact) rather than those of successful programs elsewhere-in other words, why was this CCT program so ineffective?
Although identifying the precise reason is unfortunately beyond the scope of this study, we suspect a number of potential explanations.
One possibility is that the CCT cash transfers might not have been large enough, which would lead to low uptake. However, given the way this intervention was designed, we doubt this is the case. If a mother had taken advantage of all of the incentivized activities, she would have earned more than 1,000 yuan. In these very poor communities, this is a sizeable amount, given that average annual income in 2013 was approximately 1,500 yuan per capita [7]. Thus, this is likely not the only reason for the program's underperformance.
Another possibility is that travel distances from villages to township health centers may have been prohibitively far. However, this does not seem likely, as only a small share-less than 15%-of target households were more than one hour away from the township health centers. Thus, at least for most mothers, this was unlikely to be a constraint. Likewise, our heterogeneous analysis identified no impacts by travel time from their household to the township health center [see Appendix Table A2 in the Additional file]. This result is consistent with the previous research on geographic accessibility in ethnic minority areas of Western China [31]. We also found that mother's ethnicity and education were associated with the uptake of MCH services and knowledge [see Appendix Table A3 in the Additional file]. In China, ethnic minority status is associated with economic and educational disadvantages, rural residence, mountainous topography, and poor infrastructure. In this study, 66% women were of Yi or Tibetan minority status. Both the Yi and Tibetan groups are mostly subsistence farmers who live in remote mountainous areas of Sichuan Province, with low levels of education and little access to formal health care [31,35]. Previous research among Yi and Tibetan women in China has suggested that these women may choose to give birth at home due to the high cost of care, the poor quality of township hospitals, and the cultural inappropriateness of birthing practices that cause women discomfort and embarrassment [31,36].
Low levels of education and health care knowledge may also contribute to the low uptake of MCH care among ethnic minorities [37][38][39]. Our research similarly suggested that ethnic minority status and education were still important demand-side factors that were barriers to the use of MCH services.
However, these demand-side factors were the exact barriers to MCH service uptake that the CCT program was supposed to overcome. Thus, there must be other factors that are influencing the effectiveness of the CCT program and explain low levels of improvement we found. We suspect that problems with the CCT program's implementation, similar to those encountered in the Honduras study [20], may provide an explanation for the low observed effectiveness of the CCT program.
Although the program was supposed to be aggressively promoted, only 60% of women in the fully eligible (FE) or partially eligible (PE) households in CCT villages knew about it. Further, although women were offered CCTs to get MCH services, they might not have believed that they would get paid. In the FE and PE households in the CCT villages, compliance in the CCT program was low: only about 20% of women received at least two types of MCH services with monetary transfers, and about 50% of women received at least one type of MCH service with monetary transfers. Our further ATT analysis showed that uptake of MCH services and knowledge would be significantly improved if women had actually heard about the CCT program. In other words, the use of MCH services and mother's MCH knowledge would be improved if the CCT program were implemented more effectively.
Another potential reason for this lack of CCT participation may be due to low medical quality and consequent absence of trust in the Chinese medical system. There is a growing body of literature that documents the low quality of health care in rural China [31][32][33][34]. Because these studies were conducted in relatively better-off areas of rural China, it is entirely plausible that the quality of the doctors and general health care in these poorer and more remote areas of this study are even worse.
Hence, like the study in Kenya [21], it may be that child health outcomes did not improve due to low quality of health care. This low quality of care undermines mother's trust, creating a demand-side issue for MCH uptake and undermines the effectiveness of the CCT program.
This study has several limitations. First, because of the implementation schedule of the CCT program, we were unable to collect baseline data on either the treatment or control individuals/villages. Second, in part due to the absence of baseline data, our evaluation design was not a randomized control trial, and CCT villages were not randomly assigned. To do our best to decrease evaluation bias and increase evaluation power, we used a two-pronged evaluation strategy: a between-village matching strategy and a within-village difference-in-difference strategy. These strategies let us to compare the changes not only after the CCT program was launched between CCT treatment villages and control villages but also before and after the CCT program was launched within villages.
Another fundamental limitation of this paper is that we are not able to provide a full, empiricallybased prescription on how policymakers should focus their efforts to try to improve access to MCH services and improve health outcomes. In this paper, we can only show that uptake and knowledge of MCH services are poor and that the pilot CCT program really did not solve the problem. We did spend a lot of time addressing potential reasons for this, including, poor implementation of the program, an absence of trust in the system and poor quality health care in general (which both may reduce impact on health outcomes, given access, and may also undermine interest in the program). However, the current study does not provide an empirical basis for conclusive findings on what to do in the future.
Future research is needed to address a number of these other potential constraints.

Conclusions
In conclusion, although the CCT program had a limited number of positive impacts on MCH service utilization and mother's knowledge, the effect sizes were small. There was also almost no effect on child health outcomes. Although we were not able to empirically establish the reasons for this low performance, possible reasons for the small effects might be the low trust of participants in and the poor quality of the rural health system in rural China. Additionally ineffective CCT implementation may have further undercut CCT progress. More research needs to be done to create an effective CCT system and more emphasis on the quality of the health system should be considered in future piloting efforts.
Endnote: 1 In order to ensure the accuracy of the information, enumerators also checked children's birth certificate (where place of delivery, birth height, and weight were recorded) as well as their Availability of data and materials: The datasets generated during and/or analyzed during the current study are not publicly available, but are available from the corresponding author on reasonable request.

Competing interests:
The authors declare that they have no competing interests.  : : The   The number of observations are presented in the case of binary variables or mean in the case of continuous variables; percentage or standard deviations are presented in parentheses. FE = fully eligible for the CCT program; PE = partially eligible; IE = ineligible. The compliance rates for physical examinations and child vaccinations are calculated by the requirements of the national standards of basic public health services by child's age. a : Stunted growth: length-for-age Z-scores are less than -2 standard deviations. b : Wasting: weight-for-height Z-scores are less than -2 standard deviations. Linear and logistic regressions are used to analyze CCT's impact on uptake of health services, mother's knowledge, and child health outcomes. Covariates include child's age, gender, low birth weight, premature birth, and birth order; mother's ethnicity, education, and occupation; and number of children, whether the family received social security support, distance from household to township heath center, travel time from household to township health center, and household fixed assets.
Standard errors are clustered at the town level. a : Compliance rates of physical examinations and child vaccinations are calculated by the requirements of the national standards of basic public health services by child's age. b : Stunted growth: length-for-age Z-scores are less than -2 standard deviations. c : Wasting: weight-for-height Z-scores are less than -2 standard deviations. *p < 0.05 Table 6 Average-treatment-effect-on-the-treated (ATT) analysis for the effects of CCT treatment (N =  Figure 1 Map of Evaluation areas in poor rural areas of western China. Note: The designations employed and the presentation of the material on this map do not imply the expression of any opinion whatsoever on the part of Research Square concerning the legal status of any country, territory, city or area or of its authorities, or concerning the delimitation of its frontiers or boundaries. This map has been provided by the authors.