Willingness to pay for small‐quantity lipid‐based nutrient supplements for women and children: Evidence from Ghana and Malawi

Abstract Small‐quantity lipid‐based nutrient supplements (SQ‐LNS) are designed to enrich maternal and child diets with the objective of preventing undernutrition during the first 1,000 days. Scaling up the delivery of supplements such as SQ‐LNS hinges on understanding private demand and creatively leveraging policy‐relevant factors that might influence demand. We used longitudinal stated willingness‐to‐pay (WTP) data from contingent valuation studies that were integrated into randomized controlled nutrition trials in Ghana and Malawi to estimate private valuation of SQ‐LNS during pregnancy, postpartum, and early childhood. We found that average stated WTP for a day's supply of SQ‐LNS was more than twice as high in Ghana than Malawi, indicating that demand for SQ‐LNS (and by extension, the options for effective delivery of SQ‐LNS) may be very context specific. We also examined factors associated with WTP, including intervention group, household socioeconomic status, birth outcomes, child growth, and maternal and child morbidity. In both sites, WTP was consistently negatively associated with household food insecurity, indicating that subsidization might be needed to permit food insecure households to acquire SQ‐LNS if it is made available for purchase. In Ghana, WTP was higher among heads of household than among mothers, which may be related to control over household resources. Personal experience using SQ‐LNS was not associated with WTP in either site.

With growing consensus on undernutrition as a top global health and economic development priority (Copenhagen Consensus, 2012; Food and Agricultural Organization of the United Nations and World Health Organization, 2016), micronutrient supplementation has emerged as one component of the multipronged and multisectoral responses that will be necessary to achieve targeted reductions in maternal and early childhood undernutrition .
Despite their potential benefits, in developing countries many preventative innovations designed to improve health and/or nutritional outcomes, including products such as insecticide-treated bednets and water filtration systems, suffer from low household-level demand and high price sensitivity (Dupas, 2011). And at least in one setting, it appears that SQ-LNS may face similar challenges: A year-long market trial in Burkina Faso revealed low levels of demand relative to the recommended supplementation regimen of one sachet per child per day and very high price elasticity, especially for repeat purchases (Lybbert, Vosti, Adams, & Guissou, 2016).
This finding is significant because in contrast to ready-to-use therapeutic foods, which have historically been purchased and distributed by the international donor community for free via public channels, it is unlikely that SQ-LNS will follow a similar path (Lybbert, 2011;Lybbert et al., 2016). This stems primarily from the intensity and recommended duration of SQ-LNS consumption as well as the sheer size of the potential target populations of women and young children, making public procurement of SQ-LNS at scale very resource intensive and distribution via public channels logistically complex. Instead, in settings where SQ-LNS are likely to be effective and policymakers choose to promote them, a hybrid distribution system may emerge that reaches target consumers through both public channels and retail outlets (Lybbert, 2011). The success of hybrid distribution of SQ-LNS will hinge on the answers to several key questions. In particular, what delivery platforms might be used to distribute SQ-LNS to vulnerable populations? Can private consumers reasonably be expected to cover at least some of the SQ-LNS production and distribution costs, and if so, what ratio of private consumer to socially subsidized cost would yield a price that target consumers might pay? How might policymakers and practitioners charged with reducing undernutrition design programs-targeted subsidies, social marketing campaigns, partnerships with existing health product outlets, incentives, and so forth-to help boost low private demand?
Several research activities have been undertaken to build an understanding of SQ-LNS demand, including stated willingness to pay (WTP) in Niger (Tripp et al., 2011), stated WTP and a short-term market simulation in Ethiopia (Segrè et al., 2015), an experimental auction in Ghana (Adams, Lybbert, Vosti, & Ayifah, 2016), and an experimental auction followed by a market trial in Burkina Faso . We assessed stated WTP for SQ-LNS alongside randomized controlled trials in Ghana (DYAD-G trial) and Malawi (DYAD-M trial) undertaken by the International Lipid-Based Nutrient Supplements (iLiNS) Project (http://ilins.org/). Several features of the data and analysis used in this study offer unique insight into WTP for SQ-LNS. In particular, the randomized design and data collection activities of the two trials were, by design, very similar, making a cross-site comparison of our results straightforward. Also, stark differences in the two sites in terms of socioeconomic conditions and rates of undernutrition provide contrasting yet realistic conditions under which SQ-LNS might be distributed. This paper reports results using longitudinal stated WTP data that were collected from the same households at multiple time points during pregnancy, the first 6 months postpartum and early childhood (defined here as the period from approximately 6 to 18 months of age). During this time, some mothers and their children were provided with SQ-LNS, whereas others were not. This allows us to assess the effect of consuming LNS on WTP and how that effect varies over time. Our objective is to characterize private valuation of SQ-LNS over the course of recommended consumption and, using a diverse set of covariates, explore the factors associated with WTP.

| PARTICIPANTS AND METHODS
The DYAD-G trial took place in a busy commercial corridor stretching through the Lower Manya Krobo and Yilo Krobo districts in the Eastern Region of Ghana, approximately 70 km north of Accra, the nation's capital. Women were recruited for participation in the trial

Key messages
• In contexts where SQ-LNS are likely to be effective and policymakers choose to promote them, SQ-LNS could be delivered to nutritionally vulnerable women and children via a mix of public and private channels, with private consumers covering part of the production and distribution costs. Algorithms for allocating costs across stakeholder groups need to be identified and tested.
• Stated willingness to pay (WTP) for SQ-LNS was more than twice as high in Ghana than in Malawi, suggesting that demand for and distribution strategies to deliver SQ-LNS will be context specific.
• Personal experience using SQ-LNS, identified via randomized assignment to treatment group, did not influence WTP.
from the four main health facilities operating in the semi-urban catchment area. On average, women who participated in the trial were approximately 27 years of age, had just over 7.5 years of formal education, and lived in food secure households (Adu-Afarwuah et al., 2015). Recruitment for the DYAD-M trial took place in the Mangochi district of the Southern Region of Malawi at a public hospital in the town of Mangochi, one rural hospital, and two rural health centres.
Participants in the trial in Malawi were slightly younger than in Ghana (26.5 years of age on average) and had less formal schooling (average of approximately 4 years), and over a third lived in food insecure households (Ashorn, Alho, Ashorn, Cheung, Dewey, Harjunmaa, et al., 2015). At both sites, women who were less than 20 weeks of gestation at a routine visit to one of the health facilities described above were recruited for participation in the trial. Recruitment was rolling,

| Data
Stated WTP data were collected from a random subsample (approximately 60% in Ghana and 45% in Malawi) of all households with women enrolled in the trial. Within a household, the WTP survey respondent was randomly assigned as either the mother participating in the trial or the head of her household, although in Malawi interviewing heads of household proved difficult, resulting in a substitution of the mother as the representative household respondent in almost all cases (approximately 94% of respondents).
WTP data were collected five times, divided into three periods for the purposes of this analysis. Shortly after the beginning of maternal supplementation, we elicited WTP for SQ-LNS-P&L for maternal consumption during pregnancy. At around the 35th week of gestation, we again elicited WTP for SQ-LNS-P&L during pregnancy. These two time points comprise the pregnancy period. Approximately 3 months after the birth of the infant, we elicited WTP for SQ-LNS-P&L for maternal consumption during the first 6 months postpartum, and this time point represents the postpartum period. Finally, at approximately 6 and 18 months after the birth of the infant, we elicited WTP for SQ-LNS-Child for child consumption. A timeline of WTP data collection is available in the Supporting Information.
Stated WTP for SQ-LNS was elicited using a contingent valuation survey, described in detail in the Supporting Information. In short, after receiving brief information about undernutrition and nutrient supplements in general, respondents were asked to imagine SQ-LNS were available for sale at a nearby kiosk and, bearing in mind their budget and regular expenses, were then led through a bidding tree to determine their maximum WTP. In Ghana, respondents were asked their WTP for a day's supply (one 20-g sachet), and in Malawi, respondents were asked their WTP for a week's supply (seven 20 g sachets). For purposes of cross-site comparison, WTP for a week's supply in Malawi was converted to a daily rate for all analyses. The starting bids, which were randomized across respondents, were set at GH¢ 0.20, GH¢ 0.50, or GH¢ 1.00 (approximately US $0.13, $0.33, or $0.66) for a day's supply in Ghana, and in Malawi, they were K100, K200, or K300 (approximately US $0.30, $0.60, or $0.90) for a week's supply.
The starting bids were chosen to be comparable to the prices consumers would face when purchasing traditional or local products commonly used to improve diet quality among pregnant women and/or young children (the specific comparator product used to set starting bids in Ghana was soybean flour, commonly sold by nurses at prenatal clinics, whereas in Malawi, it was a corn-soy blend called Likuni Phala, designed for children aged 6 months and older).
Given that SQ-LNS are meant to be consumed daily for many months, after respondents reported their maximum WTP for a day's supply, they were asked follow-up questions to elicit WTP for the product throughout the relevant time period (i.e., throughout pregnancy or throughout the first 6 months postpartum for SQ-LNS-P&L and from 6 to 18 months of age for SQ-LNS-Child). In Ghana, we analysed both stated maximum WTP for a day's supply and stated long-term WTP throughout the period. An error in the printing of the WTP surveys in Malawi rendered the estimates of long-term WTP unreliable, so the Malawi analysis is limited to WTP for a day's supply.
To shed light on the factors that influence WTP for SQ-LNS over the course of a child's critical window of nutritional vulnerability, we combined the WTP data with household demographic and socioeconomic data, maternal and child morbidity data, as well as birth outcome and child growth data. The covariates used in our regression analyses are defined below and in Table A2 in the Supporting Information, and the average value of each covariate by round is in Supporting Information Tables A4 and A5. A household asset index was constructed using principal components analysis to combine data collected within a few months of enrolment on ownership of a set of assets, housing characteristics, and water and sanitation sources (Vyas & Kumaranayake, 2006). The Household Food Insecurity Access Scale Score is an indicator of a household's level of food insecurity and was based on the Household Food Insecurity Access Scale (Coates, Swindale, & Bilinsky, 2007). Each household was assigned a score between 0 and 27 on the basis of how frequently the household experienced each of the nine food insecurity conditions in the 4-week period prior to the interview; a higher score indicates higher food insecurity. Maternal morbidity data were collected at biweekly home visits during pregnancy and at weekly home visits for the first 6 months postpartum. Mothers were asked to recall the number of days in the past week (Ghana) or past 2 weeks (Malawi) in which they experienced each of a range of morbidity symptoms.
Child morbidity data were collected at weekly home visits from birth to 18 months of age. With the aid of a morbidity calendar, specific dates in the previous week in which the child experienced each of a range of morbidity symptoms were recorded. Among the range of maternal and child morbidity symptoms available in the data, we selected a subset to include in our analysis primarily based on two criteria: (a) a parent or guardian might correlate the morbidity symptom with the need for or side effects associated with SQ-LNS, and (b) there was sufficient variation in the data. Morbidity variables were defined as dichotomous indicators of whether the mother or child experienced the morbidity symptom for one or more days during the reference period (reference periods defined in Table A3 of the Supporting Information).

| Empirical methods
The random assignment of mothers and their infants to receive SQ-LNS allowed us to assess whether gaining first-hand experience using them had an impact on stated WTP for SQ-LNS. For the pregnancy and child periods, where we had up to two observations per respondent, we estimated a random effects tobit model (separately for each period) for i = 1 , 2 , … , N survey respondents and for t = 1 , 2 rounds of WTP data collection for latent variable y Ã it as (Cameron & Trivedi, 2010) The dependent variable, y it , was stated WTP in 2011 US dollars for respondent i at time t. It was observed at its true value if WTP was greater than zero and censored at zero otherwise. LNS i was an indicator variable equal to one if the mother-infant dyad in respondent i's household was randomized to receive SQ-LNS and zero otherwise.
The vector T it comprised time-varying controls (passage of time from enrolment/birth to WTP survey administration and indicators of randomized starting bid). The parameter α i was a respondent-level random effect, and ε it was an idiosyncratic error. Because the error was likely correlated over time for a given respondent, standard errors were bootstrapped to account for clustering at the level of the respondent. For the postpartum period in which there was only one observation per respondent, we used a tobit model with robust standard errors to estimate the effect of intervention group on WTP with the same set of controls as in Equation 1. In all models, heterogeneity over time, by survey respondent and by maternal parity, was assessed using interactions with intervention group.
To estimate the association between stated WTP and the predicted correlates for the pregnancy and child periods, we extended Equation 1 to include a set of time-varying covariates (e.g., maternal and child morbidity) in the vector X it , and time-invariant covariates (e. g., infant gender) in the vector Z i and estimated the following random effects tobit model The vector T it again comprised time-varying controls, in this case the passage of time from enrolment to WTP survey administration (pregnancy period) or birth to WTP survey administration (postpartum and child periods), the total number of days in the morbidity reference periods, and indicators of randomized starting bid. For the postpartum period, we estimated the correlates of WTP using a tobit model with robust standard errors.
We also assessed the sensitivity of our regression results to the way the morbidity variables were defined (dichotomous vs. continuous) and the length of the morbidity reference periods. Sensitivity analysis results for each period are available in the Supporting Information.

| Effect of experience using SQ-LNS on WTP
There were no statistically significant differences in average WTP between the LNS group and the capsules group at any time point in either Ghana or Malawi (Supporting Information Figure A3). Our tobit model estimates confirmed no difference in WTP by intervention group in any period or over time in either Ghana or Malawi (see Supporting Information Sections 5 and 6). In Malawi, for each additional month from the birth of the infant to WTP survey administration, WTP for a day's supply of SQ-LNS was approximately $0.05 lower (p = .08), and the association between the household asset index and WTP was positive and significant (p = .05). Note. At each round, observations >6 SD above the mean were omitted as outliers. Differences in average WTP between rounds were assessed by testing for a significant coefficient on an indicator variable for round using ordinary least squares (OLS) with cluster-robust standard errors. Regressions were run separately for each pair of rounds. SD = standard deviation; WTP = willingness to pay.

| Factors associated with WTP during the child period
In the child period in Ghana, WTP for SQ-LNS for the infant, both for a day's supply and in the long term, was lower among mothers compared to heads of household (   nor were the measures of child nutritional status (length-for-age zscore (LAZ) and weight-for-length z-score (WLZ)).
In Malawi, the association between months from the birth of the infant and WTP was negative and significant (p < .001). There was a positive and significant (p = .01) association between WTP for a day's supply of SQ-LNS for the child and the asset index and a negative  First, because respondents to the WTP surveys did not face an actual budget constraint, stated WTP is likely influenced by positive hypothetical bias (Whittington, 2010). In several settings, stated WTP for SQ-LNS has been shown to be roughly double compared to estimates of revealed WTP, that is, WTP elicited under incentive-compatible conditions in which money changed hands Lybbert et al., 2016;Segrè et al., 2015). We therefore treat our estimates of stated WTP as an upper bound on true WTP. Conversely, it is conceivable that although respondents were asked to imagine that they were no longer receiving SQ-LNS for free, receiving free product could have put downward pressure on WTP. This "anchoring" effect may have acted as a partial counterbalance to the inflationary effect of hypothetical bias (Fischer, Karlan, McConnell, & Raffler, 2014;Kőszegi & Rabin, 2006 In the interest of protecting the scientific validity of the randomized trials and to avoid potentially misinforming respondents (because the efficacy results of the trials were not known when the WTP surveys were administered), respondents to the WTP surveys were not given any information about potential benefits (in either the short or long term) of consuming SQ-LNS. Therefore, another limitation is that what we can claim about the role of experience in shaping WTP for SQ-LNS is limited to experience accumulated in the very immediate term and does Note. Dependent variables are stated WTP in 2011 US dollars. The variable 'Mother' = 1 if the respondent to the WTP survey was the mother participating in the randomized trial and = 0 if the respondent was the head of household. Controls for randomized starting bid are included in all models (unreported). Standard errors, in parentheses, obtained via 50 bootstrap replications. HFIAS = Household Food Insecurity Access Scale; LAZ = length-for-age z-score; LNS = lipid-based nutrient supplements; WLZ = weight-for-length z-score; WTP = willingness to pay.
not account for how potential longer term benefits or costs, learned through experience, information, or social networks, may influence WTP. Nor are we able to make claims about the potential role of experience outside the context of a randomized controlled trial.
Finally, the extent to which our findings can be generalized is limited because our sample is composed of respondents from households in which a pregnant member sought out early (before 20 weeks of gestation) formal prenatal care, and households that do not seek early formal prenatal care may differ in their characteristics and valuation of a product such as SQ-LNS.
Despite these limitations, the results presented here highlight several issues relevant to policymakers, government agencies, or aid organizations interested in forecasting potential demand for SQ-LNS and designing strategies to deliver them. First, the contrasting results from  An experimental auction in Burkina Faso revealed an average WTP for SQ-LNS-Child of $0.85 for a week's supply (or $0.12/day's supply; Lybbert et al., 2016). In another study, Segrè et al. (2015) Figure 2.
Our results and the range of estimates of WTP from other studies suggest that although a hybrid delivery strategy that includes a marketbased mechanism might be feasible in settings such as semi-urban Ghana, the proportion of households who might be expected to pay an out-of-pocket cost for SQ-LNS would be much smaller in more resource-constrained settings such as Malawi. In settings where policymakers expect SQ-LNS to be effective and choose to promote them, private demand should be assessed in the local context, preferably using an incentive-compatible mechanism that captures the persistence of demand over time (e.g., Lybbert et al., 2016). Where private demand is low, higher levels of public investment will be required to reach much of the target population of women and children.
Leveraging the randomized design of the iLiNS trials that provided SQ-LNS to some households and not others, we found that in general, We found that in Ghana, WTP for SQ-LNS for both maternal and child consumption was higher among heads of households than among mothers. In our sample, over 90% of heads of household indicated that they were primarily responsible for making purchase decisions related to food and for purchase decisions related to health care for their households. Only about a quarter of mothers indicated that they were primarily responsible for these purchase decisions in their household.
The finding that WTP was lower among mothers, then, may be related to control over household resources in this setting and points to a potential need to include both mothers, who would be directly consuming SQ-LNS or feeding them to their young children, and heads of household in, for example, social marketing efforts.
We also found that WTP for SQ-LNS both for maternal consumption postpartum and for child consumption was, on average, higher for male children than female children in Ghana. In light of the lack of evidence of gender bias in nutritional outcomes in Ghana, this finding is a bit puzzling and deserves more research to uncover its underlying mechanism. At a minimum, though, policymakers should be aware of the potential for differential demand by child gender in Ghana.
In both Ghana and Malawi, we found a consistent negative association between household food insecurity and WTP for SQ-LNS, even after controlling for household assets. This result is noteworthy because it is possible that women and young children in more foodinsecure households are also those for whom SQ-LNS may help protect against critical nutrient gaps in their diets, as was found in Bangladesh (Mridha et al., 2016). This result suggests that if a cost-effective mechanism for targeting based on household food insecurity could The challenges associated with delivering a regular-use product such as SQ-LNS to target members of nutritionally vulnerable households are many. SQ-LNS is an unfamiliar product. Many of the hopedfor benefits may be difficult for households to directly observe and are long term in nature. Households have to make trade-offs in where they choose to allocate their scarce resources. Should SQ-LNS become available (either via markets, through health clinics, or by some other delivery mechanism), demand will be low for some stakeholder groups but may be higher among others. In contexts were SQ-LNS are likely to be effective, and if policymakers choose to promote them, possible delivery mechanisms for SQ-LNS will need to be carefully considered within the specific context, and education/information dissemination and targeted subsidies will likely be part of any distribution strategy.

CONFLICTS OF INTEREST
The authors declare that they have no conflicts of interest.