Does organic farming jeopardize food security of farm households in Benin?

The prevalence of organic farming and other sustainability standards is increasing around the globe. While effects of organic farming on productivity, income, and poverty alleviation have been analyzed in numerous empirical studies, its effects on food security are barely understood. Using data from smallholder cotton farmers in Benin, we aim to empirically investigate how adopting organic farming affects their food security. According to our results, organic farming is conditionally associated with a notably lower experienced food security and a slightly lower dietary diversity and consumption of vitamin A-rich foods. Evaluating pathways, we find that the negative conditional association between organic farming and food security is a result of a lower household income of organic farms due to lower income from cotton farming given a smaller land area cultivated with cotton, while a larger land area cultivated with food crops cannot fully compensate for the reduced income from cotton farming. This alarming result illustrates the need for evaluating and eventually improving programs for organic farming in developing countries to ensure that good intentions for more sustainable production practices do not jeopardize the livelihoods of vulnerable smallholder farmers.


Introduction
Although substantial interventions aimed at supporting food security in sub-Saharan Africa (SSA) have been introduced over the past decades, food insecurity is still highly prevalent in the region.Particularly, the prevalence of undernourishment rose from 18.3% in 2015 to 23.2% in 2021 in SSA (FAO et al., 2022), suggesting urgent action is needed to revert this trend.Given that many of the undernourished people in SSA and worldwide are smallholder farmers who rely on agriculture for their livelihoods, changes to agricultural production systems that enable smallholder subsistence farmers to increase the quantity or nutritional quality of the produced food products or the profitability of their agricultural activities have the potential to improve the nutrition of many poor rural households.
Conventional farming practices are frequently unsustainable due to soil degradation and other environmental problems and they, thus, jeopardize food security at least in the longer run.Hence, the adoption of organic farming, which prescribes sustainable farming practices such as inter-cropping, crop rotation and diversification, legume cultivation, and the use of organic fertilizers (Meemken et al., 2017;Jouzi et al., 2017) could secure and possibly enhance the availability and diversity of food products in developing countries.On the other hand, organic G.B.D. Aïhounton and A. Henningsen While the effects of organic farming on productivity, income, and poverty alleviation have been analyzed in numerous empirical studies (e.g., Bolwig et al., 2009;Jena et al., 2012;Uematsu and Mishra, 2012;Patil et al., 2014;Ayuya et al., 2015;Chiputwa and Qaim, 2016;Parvathi and Waibel, 2016;Jena et al., 2017;Mitiku et al., 2017;Froehlich et al., 2018), its effects on food security are barely understood.To our best knowledge, only three studies investigate the effect of organic certification on food security.Using weekly household food consumption expenditure and an index of dietary quality based on the frequency of the consumption of 13 food groups, Becchetti and Costantino (2008) find a positive effect of joint organic and fairtrade certification both on food consumption expenditure and the dietary quality of farm households in Kenya.Both Chiputwa and Qaim (2016) and Meemken et al. (2017) use data from smallholder coffee farmers in Uganda and measure food security based on a seven-day recall of the consumption of more than 100 food groups.While Chiputwa and Qaim (2016) show that a certification for organic and/or fairtrade standards improves the consumption of calories and three micronutrients, Meemken et al. (2017) report similar results for organic certification but do not find a significant effect of fairtrade certification on food security.As the studies of Chiputwa and Qaim (2016) and Meemken et al. (2017) use partly the same data from the same farmers and the studies of Becchetti and Costantino (2008) and Chiputwa and Qaim (2016) analyze only the joint effect of organic and fairtrade certification, the literature provides so far only little empirical evidence on the effects of organic farming on the food security of smallholder farmers.
Gaining these insights is important because farmers in developing countries often derive their livelihoods from their farming activities, which provide food for their own consumption as well as cash income from sales of farm products (e.g., Carletto et al., 2015;Chiputwa and Qaim, 2016;Fanzo, 2018;Fraval et al., 2019), and it would be very undesirable and ethically problematic if programs that encourage the adoption of organic farming in developing countries-often initiated by private or public actors in high-income countries-compromised the food security of vulnerable smallholder farmers in low-income countries.Hence, assessing the effects of organic farming on food security is highly relevant, not only scientifically but also for governments, donors, and NGOs that engage in organic farming in developing countries.
This study contributes to the existing literature as it sets out to empirically investigate how organic farming affects farm households' food security through various pathways.In our empirical analysis, we address the potential endogeneity of choosing between organic and conventional farming by using three different approaches.First, we use a variable as control variable that captures a household's marginal utility of adopting either organic or conventional farming practices.This approach to accounting for potential unobserved heterogeneity in a selection-on-observables design has been used in a number of studies (see, e.g., Verhofstadt and Maertens, 2014;Bellemare and Novak, 2017;Rajkhowa and Qaim, 2021;Ruml and Qaim, 2021).Second, we apply the approach suggested by Oster (2019) to check the sensitivity of our estimates regarding unobserved heterogeneity.Third, we use instrumental variable regression.We use six different outcome variables to measure different aspects of food security.We apply the FAO's household food insecurity experience scale (HFIES) as well as an extension of this scale that captures smaller variations in food insecurity.Furthermore, we use the FAO's dietary diversity score and the number of vitamin A-rich food groups consumed within 24 h and within seven days as proxies for aspects of food security related to dietary quality.Our empirical analysis is based on a sample of 1247 households that produce organic or conventional cotton in Benin.Given that we use observational cross-sectional data, we cannot be sure that our three approaches remove endogeneity completely.As a precaution, we interpret our estimates not as causal effects but as conditional associations.
According to our results, organic farming is conditionally associated with a notably lower experienced food security.Moreover, organic farming tends to be conditionally associated with a slightly lower dietary diversity and number of vitamin A-rich food groups consumed within 24 h, but no notable conditional association with these two measures was found when considering the consumption within a sevenday period.Hence, organic farming is expected to jeopardize some aspects of food security of the farm households who adopt organic farming, at least when circumstances are similar to those of the farmers in our empirical analysis.To further understand these effects, we investigate the mechanism of the effects of organic farming on food security.We find that organic farming is conditionally associated with a substantially smaller land area cultivated with cotton, a substantially lower income from cotton farming, and a substantially lower household income, while it is conditionally associated with a substantially larger land area cultivated with food crops.The lower income from cotton farming indicates that the higher price for organic cotton does not fully compensate for the lower cotton yield and the smaller land area cultivated with cotton.Furthermore, the larger land area cultivated with food crops does not fully compensate for the lower income from cotton farming and, thus, the capability to purchase food.This is an important result as organic farming is gaining importance in many developing countries and evidence is needed to ensure that this does not have undesirable side effects on the participating smallholder farmers.
The remainder of this paper is organized as follows.Section 2 describes the empirical background of our study.Section 3 describes the data collection and our empirical strategy.The results are presented in Section 4. Section 5 concludes and provides policy implications.

Empirical background
In Benin, agriculture plays a key role in the economic development and largely relies on cotton production, which accounts for 40% of foreign exchange earnings, 12 to 13% of Gross Domestic Product, and provides income for more than a third of the population (INSAE, 2020).Due to favorable climate conditions, cotton farming is most prevalent in the Northern part of the country, where almost 50% of the agricultural land is cultivated with cotton (Westerberg, 2017).Given the high importance of cotton in the economy, the government of Benin has supported conventional cotton production for decades, ranging from agronomic extension service tailored to cotton farming, to provision of chemical inputs for cotton farming on credit, and to a guaranteed price for cotton farmers.As the cotton growing areas of Benin are considerably more affected by food insecurity and poverty than the coastal region of Benin (e.g., due to low and unreliable yields of food crops, limited availability of food during the lean season given seasonal production, inadequate storage and post-harvest losses, etc.), the Beninese government supports intensification of cotton production also as a policy instrument to achieve food security and poverty alleviation (Glin et al., 2012).
Each year the Cotton Inter-professional Association of Benin that consists of representatives from the National Council of Cotton Ginning Companies, the National Council of Cotton Producers, and the National Council of Importers and Distributors of Agricultural Inputs, negotiates a price for seed cotton 2 in the subsequent growing season under the regulatory role of the government, so that cotton farmers know the price before they plant cotton.The Cotton Inter-professional Association of Benin distributes the cotton growing area to different ginning companies based on their locations and ginning capacities.The 2 Farmers harvest and sell so-called seed cotton, which consists of cotton fiber and cotton seeds.Ginneries separate the fiber from the cotton seeds.While the fiber is mostly used in the textile industry, the cotton seeds are used for planting in the following year or are processed to cotton seed oil (e.g., for cooking) and cotton seed cakes (usually used as animal feed).
G.B.D. Aïhounton and A. Henningsen ginning companies have a monopsony on seed cotton in the area that was distributed to them, where they have numerous local collection points, at which they are obliged to purchase all the conventional and organic seed cotton that farmers deliver to them.Most conventional cotton farmers are members of a cooperative.Unlike other agricultural cooperatives, these cooperatives are not directly involved in production or trading activities but they play an important role in cotton farming, e.g., by facilitating the provision of extension service to farmers, the distribution of chemical inputs to farmers, and the collection of seed cotton.
In the main cotton-growing areas of Benin, the population density is rather low and farms generally cultivate larger land areas than farms in the coastal and more fertile regions of Benin.For instance, Sodjinou et al. (2015) found that on average conventional cotton farmers had 17.4 ha land and organic cotton farmers had 16.1 ha land available for farming.Cotton is usually cultivated in rotation with other crops, particularly food crops such as maize, soybean, and sorghum.While cotton farming is the main source of cash income of most farms in these regions, many farms also sell a part of the food crops that they produce.Farmers can easily obtain the recommended (quite high) quantities of chemical inputs for their entire cotton production on credit but this is not the case for other crops so that farmers use a lot of chemical inputs in cotton farming, while they use chemical inputs to a much lesser extent in the production of other crops (Aïhounton et al., 2021).The extensive use of chemical inputs in cotton farming leads to high cotton yields but also causes severe health and environmental problems.For instance, the use of chemical inputs in conventional cotton farming leads to skin diseases, food intoxication, loss of soil microorganisms, water and air pollution, and decline of biodiversity (Agbohessi et al., 2012;Azandjeme et al., 2014;Agbohessi et al., 2015;Douny et al., 2021).
In the mid-1990s, organic cotton farming was introduced in Benin by the Beninese Organization for the Promotion of Organic Agriculture (OBEPAB) supported by transnational NGOs (PAN Trust/UK, Agro Eco, Solidaridad, PAN Africa) under the bilateral Sustainable Development Agreement (SDA) between the governments of Netherlands and Benin (Glin, 2014).This initiative to promote organic cotton farming in Benin was later joined by other organizations including Helvetas Intercooperation (a Swiss Development NGO) and the German Agency for International Cooperation (GIZ) (Hougni et al., 2012).Organic cotton farming was promoted as an alternative to conventional cotton farming that aims to protect the environment and to preserve the health of the cotton farmers and their households.Organic farmers are not allowed to use chemical inputs such as inorganic fertilizers and synthetic pesticides in any part of their farming activities, i.e., neither on their cotton plots nor on their other plots.Instead, organic farmers usually produce their own biopesticides and organic fertilizers but they can also purchase biopesticides and organic fertilizers that are approved for organic farming from local enterprises.To enhance soil fertility and food production, organic farmers are recommended to do intercropping and wide crop rotations with legumes and various other crops (Aihounton et al., 2022).
All farmers, irrespective of their farm size, have the possibility to become certified organic farmers if their plots are suitable for organic farming, i.e., the plots do not bear the risk of being contaminated with chemical inputs through runoff from conventional plots, e.g., are not located on the slope below conventional plots.While membership in a cooperative is voluntary for conventional cotton farmers, farmers who want to become certified organic cotton farmers must either join an existing cooperative of organic cotton farmers in their village or in a nearby village or establish a new cooperative with other cotton farmers who also want to become certified organic farmers.Cooperatives of organic cotton farmers are similar to cooperatives of conventional cotton farmers but their members also have the obligation to mutually check if the other members of their cooperative strictly follow organic farming regulations on all their plots.If a single farmer violates the organic farming regulations and the other farmers immediately report this violation to the certification agency, only the farmer who violated the regulations will lose her/his certification.However, if the violations are more widespread or the controls are not done properly, all farmers of the cooperative get their certification status suspended (and, thus, do not get the price premium) at least for some time.The certification does not require a minimum area cultivated with cotton but the organizations supporting organic cotton farming usually limit the land area that each single farmer is allowed to cultivate with organic cotton (e.g., 5 hectares by OBEPAP) in order to ensure that the farmers are able to properly manage their plots and have sufficient resources (e.g., labor) to follow organic farming practices.The certification takes three years, i.e., the farmers have to follow organic farming regulations for three years before they can sell their crops (from the fourth year on) as certified organic produce.Certified organic farmers receive a premium price for their seed cotton that is 20% above the price of conventional seed cotton (Moumouni et al., 2013).
Organic farming comprises only a small proportion of cotton production in Benin with 3.3% of the cotton farmers (6621 organic cotton farmers Textile Exchange, 2022;200,569 cotton farmers in total DSA, 2022) and 1,3% of the land area cultivated with cotton (8199 ha organic cotton Textile Exchange, 2022;614,297 ha cotton in total DSA, 2022) and accounted for only 0.6% of the cotton fiber that was produced in the 2020/2021 growing season (Textile Exchange, 2022).Organic cotton farmers generally obtain lower yields compared to conventional cotton farmers (Westerberg, 2017;Aihounton et al., 2022).For instance, the average seed-cotton yield in our data set is 1087 kg/ha for conventional farms and 936 kg/ha for organic farms (see Table 1 in Section 4.1).

Data collection
Our analysis uses cross-sectional data from a household survey in Benin.The survey locations cover three main cotton growing districts, namely Kandi, Péhunco and Glazoué, which were selected due to the high importance of cotton farming, agro-ecological diversity, and the presence of organic cotton farming in the areas.We selected three groups of villages for our household survey.In order to obtain data from a sufficiently large number of organic cotton farmers, we selected all 25 villages that engage in organic cotton farming in the three districts.Furthermore, we selected 25 villages without organic cotton farming but with similar village-level characteristics as the 25 villages with organic cotton farming in order to obtain data from organic and conventional households that operate under similar production and market conditions.Finally, we selected a further 25 villages without organic cotton farming, which together with the two other groups of villages, provide a representative sample of all cotton growing villages in each of the three districts.
We collected data from 1361 cotton growing households through face-to-face interviews using the KoBoCollect software from March to May 2018.Data from 1247 of these households (223 organic and 1024 conventional households) were used for the empirical analysis presented in this paper. 3The gathered data include information on, e.g., household characteristics, crop farming (particularly cotton), and livestock keeping.The household head and/or his/her spouse were also asked questions about their food security and about different food groups consumed in the household (see details in the following section).

Measurement of food security
Food security is defined as the situation, "when all people, at all times, have physical, social and economic access to sufficient, safe and nutritious food to meet their dietary needs and food preferences for an active and healthy life'' (FAO, 2009, p. 1).Acknowledging that this definition of food security is widely approved, measuring food security is a challenge because of its multiple dimensions (e.g., FAO, 2016).One of the measures of food security that has been most widely used in recent years is the household food insecurity experience scale (HFIES) developed by the FAO (see FAO, 2016;Cafiero et al., 2018).It is the first survey instrument to measure people's direct experiences of food insecurity at the individual or household level to be applied on a global scale (Smith et al., 2017).Secondly, it provides an internal statistical validity of the data set using the Rasch model assumptions (Cafiero et al., 2018).The household head and/or his/her spouse were asked whether their household had experienced the following situations during the last 12 months because of a lack of money or other resources: 1.You were worried your household would not have enough food to eat. 2. Your household was unable to eat healthy and nutritious food.3.Your household ate only a few different kinds of food.4. Your household had to skip a meal. 5.Your household ate less than you thought it should.6.Your household ran out of food.7.Your household was hungry but did not eat.8.Your household went without eating for a whole day.
Instead of asking the interviewees for a binary response (''Yes''/''No'') for each of the eight potential situations (as done, e.g., by Smith et al., 2017;Cafiero et al., 2018), we extend the HFIES by providing four ordered answers: 0 = Never, 1 = Rarely, 2 = Sometimes, and 3 = Often.This allows us to obtain the original HFIES, which ranges from 0 to 8 (i.e., the number of answers that are not 0), and a more detailed HFIES that ranges from 0 to 24 by summing up the answers to the eight questions.
We use the household dietary diversity score for the periods 24 h and 7 days in order to account for short-run and long-run diversity of food consumption.While the HDDS over 24 h is less subject to recall errors, the HDDS over 7 days is more representative for the household's habitual diet (Aihounton and Christiaensen, 2024) and facilitates comparisons with studies that exclusively use dietary diversity scores over 7 days (e.g., Chiputwa and Qaim, 2016;Meemken et al., 2017;Hirvonen et al., 2021;Usman and Haile, 2022).
Based on the 16 food groups listed above, a group of 'vegetables' was derived by combining vitamin A-rich vegetables and tubers (3), dark green leafy vegetables (4), and other vegetables (5).A group of 'fruit' was derived from combining vitamin A-rich fruits (6) and other fruits (7), and a group 'meat' was derived by combining organ meat (8) and flesh meat (9), leading to a total of 12 food groups.The household's dietary diversity score is obtained by summing up the number of affirmative responses and, thus, it ranges from 0 to 12.We also derive a second indicator for quality aspects of food security as the number of vitamin A-rich foods (see FAO, 2011), which is obtained by summing up the number of vitamin A-rich food groups consumed (i.e., groups 3, 4, 6, 8, 10, and 13) by the household in the past 24 h and in the past seven days and, thus, it has a score that ranges from 0 to 6.Both the dietary diversity score and the number of vitamin A-rich food groups consumed are used for measuring quality aspects of food security as they indicate the consumption of important nutrients for a healthy life.

Main regression model
In this section, we present the estimation strategy that we use to investigate the effects of organic farming on rural households' food security.
Our core regression equation is specified as: where subscript  indicates the household,   denotes our th outcome variable,   is a dummy variable that is equal to one if household  has adopted organic farming and zero otherwise,   is a vector of control variables,   is the error term for the regression with our th outcome variable,   and   are coefficients, and   is a vector of coefficients to be estimated.Our six outcome variables (  ;  = 1, … , 6) are the original HFIES, the extended HFIES, the dietary diversity scores over 24 h and 7 days as well as the number of vitamin A-rich food groups consumed over 24 h and 7 days.We are mainly interested in the coefficients   as they indicate the average treatment effects (ATE) of organic farming on the outcomes of interest, i.e.,   = (  |  = 1) − (  |  = 0).In order to obtain unbiased estimates of the ATEs with an Ordinary Least Squares (OLS) regression, we need to control for all variables that are correlated both with the household's decision to farm organically   and the error term   .Using a standard non-separable household model (e.g., Singh et al., 1986;Key et al., 2000;Henning and Henningsen, 2007, and Section A of the Supplementary Material), we derived that our vector of control variables   should include household characteristics (age, gender, education, literacy, household size, dependency ratio, distance to nearest health facility, and experience in agriculture), farm characteristics (total land owned), and market characteristics (type of road and distance to closest market).The vector   additionally includes the quadratic or cubic terms of some of the explanatory variables (based on  tests, F tests, and RESET tests) as well as arrondissement-fixed effects to account for location-specific effects. 4Given that we use observational cross-sectional data and we cannot be sure that we have controlled for all relevant factors, we do not interpret our estimates of   as causal effects but as conditional associations.
Although all of our six outcome variables are count variables, we use OLS regression rather than count-data regression due to various reasons.First, the distributions of our six outcome variables are rather smooth, symmetric and without extreme values (see Fig. 1 in Section 4.2), which is usually advantageous for OLS regressions (Wooldridge, 2016, p. 172).Second, while OLS regression does not depend on distributional assumptions for providing unbiased estimates, count-data regression methods are based on strict distributional assumptions that are likely not fulfilled for our outcome variables; thus, using count-data regression would expose us to a high risk of obtaining unreliable estimates.For instance, count-data models assume that the dependent variable does not have an upper bound but all our six outcome variables have an upper bound so that at least this assumption of count-data regression models is violated.Third, the method suggested by Oster (2019) that we use to check sensitivity to unobserved heterogeneity cannot be applied to count data and countdata instrumental-variable regression relies on much more problematic assumptions than linear instrumental-variable regression.Finally, most other studies that have the same or similar outcome variables as we have also use linear regression models.
We transform continuous variables with very right-skewed distributions with the log-transformation (if they only include strictly positive values) or by the inverse hyperbolic sine (IHS) function (if they also include zero and/or negative values; see Section C in the Supplementary Material for details) to obtain variables with more symmetric distributions and fewer extreme values because this often fulfills the assumptions required for OLS regressions to a higher degree and makes the results less sensitive to individual observations (see, e.g., Wooldridge, 2016, p. 172).Given the design of our sampling strategy, standard errors are clustered at the village level using clustered sandwich estimators of type ''HC1'' in the terminology of MacKinnon and White (1985) (for details see, e.g., Zeileis et al., 2020).

Willingness to pay as an additional control variable
Borrowing from the theory of adoption, farmers choose whether to adopt organic farming depending on their expectations of its advantages and disadvantages.Hence, if unobserved variables that affect farmers' expectations of the advantages and disadvantages of organic farming and, thus, their participation in organic farming   also affect their food security   through pathways that are not blocked by the control variables   , OLS estimates of   are biased due to omitted explanatory variables.To overcome this potential problem, we re-estimate Eq. ( 1) with an additional control variable that proxies farmers' expectations of the advantages and disadvantages of organic farming, i.e., we add an additional element to the vector   that captures the respondent's unobserved characteristics that influence participation in organic farming   .This approach to accounting for unobserved heterogeneity was inspired by Verhofstadt and Maertens (2014), Bellemare and Novak (2017), and Ruml and Qaim (2021), who used variables that proxy each respondent's marginal utility of participating in cooperatives or contract farming to control for potential unobserved characteristics that drive participation in cooperatives or contract farming.
We obtain this additional control variable in our survey by asking each respondent, ''which type of cotton farming do you think gives you and your household, in general, a better life?''.The respondents' responses were captured on a 7-point Likert scale (1 = organic much better, 2 = organic somewhat better, 3 = organic slightly better, 4 = about the same, 5 = conventional slightly better, 6 = conventional somewhat better, 7 = conventional much better).In order to avoid unnecessarily long phrases and sentences, we denote this variable that indicates the perceived overall net advantages and disadvantages of organic and conventional cotton farming as ''WTP'' given that it proxies willingness to pay (WTP) to adopt organic or conventional cotton farming (inspired by Bellemare and Novak, 2017).
If our WTP variable had been observed before the farmers decided to adopt organic farming or to remain conventional farmers, this variable would be an excellent variable for controlling for unobserved heterogeneity between organic and conventional households.As our WTP variable was observed after this decision, it is possible that positive or negative experiences with the chosen production method may have affected both the WTP variable and the food security, i.e., creating a potential endogeneity problem of the WTP variable.However, as the decision to adopt organic farming, to a large extent, depends on ideology, which is not easily changed, and given that we are not interested in the coefficient of the WTP variable as we just use it as control variable, we are confident that this weakness does not have a major effect on our estimates of the effect of organic farming on food security.

Sensitivity to omitted-variable bias
As we cannot be sure that our selection-on-observables identification strategy-even if we use the WTP variable as an additional explanatory variable-does indeed remove all correlation between our treatment variable   and the error terms   , we use the approach suggested by Oster (2019) to assess the sensitivity of our analysis to omitted variables.We derive bias-adjusted estimates of the treatment effect βℎ  for four different values of the (assumed) R-squared value of a hypothetical regression model that gives unbiased estimates ( ℎ  ).We define the bounding sets as   = [ βℎ  , β ] and consider an estimate to be robust to potential omitted-variable biases if the bounding set excludes zero (i.e., βℎ  and β have the same sign). 5

Instrumental variables
As an alternative to the approach suggested by Oster (2019), we use instrumental-variable (IV) regression to address potential endogeneity of our treatment variable.Specifically, we use the three-step IV approach suggested by Wooldridge (2010, p. 937-942) that takes into account the fact that the endogenous regressor, i.e., the dummy variable indicating organic farming (  ), is a binary variable.The first step comprises a probit regression of the dummy variable indicating organic farming (  ) on all exogenous explanatory variables   and on the instrument: exposure to organic farming ( D ) defined as the proportion of all cotton farming households that are organic cotton farming households in each village.Similar instruments have been used in many other studies (e.g., Mason et al., 2013;Krishnan and Patnam, 2014;Smale and Mason, 2014;Magnan et al., 2015;Wuepper et al., 2018;Sellare et al., 2020;Tabe-Ojong et al., 2022) as these types of instrumental variables indicate the exposure of farmers to new opportunities and as the adoption of new opportunities is frequently influenced by social ties.More explicitly, the higher the proportion of organic farmers in the same village, the more likely other farmers learn about organic farming and its potential advantages (see, e.g., Sellare et al., 2020).The second and third step comprise a classical 2SLS regression with the probability of participation in organic farming predicted by the first-stage probit model ( D ) as the instrument.Based on this IV regression,   denotes the local average treatment effect (LATE).
In order to obtain unbiased estimates of the local average treatment effect, the instrumental variable has to be relevant and exogenous.We assess the relevance of the instrumental variable by using an F-test, 5 Section E of the Supplementary Material describes this method in detail and presents the bounding sets for all four different values of  ℎ  .For simplicity, the figures in Section 4.2 only display the bias-adjusted estimates of the treatment effect for two values of  ℎ  .
which assesses the statistical significance of the predicted probability of participation in organic farming in the second step of our estimation strategy (i.e., the first step of a classical 2SLS regression).If the F-test rejects the irrelevance of the predicted instrumental variable at a very high significance level (usually an F-statistic of 10 is used as threshold, see Staiger and Stock, 1997), we conclude that our instrumental variable is not "weak" but relevant.
The second condition for the instrumental variable, exogeneity, is fulfilled if this variable is uncorrelated with the error term, i.e., the instrumental variable only affects the outcome variables through the treatment variable   and it is not correlated to omitted variables that affect the outcome variables through pathways that are not blocked by the control variables   .Indeed, the proportion of organic cotton farming households in a village is probably not randomly distributed between villages but may depend on local conditions (see Section F in the Supplementary Material for details).However, we argue that potential channels from the instrumental variable to the outcome variables that might violate the exogeneity assumption are probably 'blocked' by the control variables   , particularly variables related to the location such as arrondissement fixed effects, distance to nearest health facility, type of road, and distance to the closest market as well as the WTP variable.These variables probably capture pre-existing differences regarding locations and the (perceived) advantages or disadvantages of organic farming, which may affect food security and, at the same time, may be correlated with our instrumental variable.Therefore, we argue that our instrumental variable is credibly exogenous.Furthermore, we estimate all regression models only with observations from the first two groups of villages (see Section 3.1), i.e., only villages with organic cotton farming and villages that are similar to those with organic cotton farming regarding a number of criteria, including criteria that are probably related to food security.Finally, we assess the exogeneity of our instrument by applying the falsification test proposed by Di Falco et al. (2011).
Even if all specification tests provide favorable results, we cannot be sure that all assumptions that are necessary for obtaining unbiased estimates are fulfilled given that our empirical analysis is based on observational cross-sectional data.As a precaution, we interpret the estimated coefficients not as LATEs but as conditional associations.

What about null findings?
In our empirical analysis, finding that the adoption of organic farming is not conditionally associated with and, thus, does not jeopardize food security, would be a highly relevant result.However, P-values and statistical significance usually do not help to answer questions of interest like those addressed in this paper (Imbens, 2021) and obtaining statistically insignificant estimates does not imply that there is no association (e.g., Brown et al., 2018;Abadie, 2020).One way to investigate whether statistical insignificance implies that there is no association is to calculate the statistical power of the significance test.However, in many empirical applications with observational data in economics, one has too little information to conduct reasonable power calculations before the data are observed and analyzed, while ex-post power calculations are ''fundamentally flawed'' (Hoenig and Heisey, 2001).
On the other hand, in many cases, a point estimate and the degree of uncertainty associated with the point estimate may be the precursor to making a decision (Imbens, 2021).Given that the absence or presence of statistical significance should not be the basis to claim any association, we investigate both the statistical and the economic significance of our estimates with confidence intervals (see, e.g., Hoenig and Heisey, 2001).Specifically, we display the 50% and 95% confidence intervals for the estimated magnitudes of the conditional associations obtained from various OLS and IV regressions.If the confidence intervals of the conditional associations do not usually encompass economically significant magnitudes, we can conclude that organic farming does not have an economically relevant conditional association with food security, which would be an important finding.

Descriptive statistics
Table 1 reports mean characteristics of organic and conventional households in our sample along with balancing tests.The majority of the household heads is male but the share of female household heads is significantly higher in organic households (14%) than in conventional households (5%).The average educational attainment and literacy are very low with, on average, 1.38 years of education and only 19% of the households heads are able to read and write in their local language.While heads of organic households have a slightly higher level of education and literacy than heads of conventional households, these differences are not statistically significant at the 10% level.In contrast, conventional households are significantly wealthier than organic households in terms of land ownership and household assets.Furthermore, organic households are significantly less frequently located on a tarred road and are, on average, further from a health facility.Therefore, they are more remote than conventional households.
Regarding the variable denoting WTP, we find a large difference between organic and conventional farmers.While 82% of the organic farmers state that organic cotton provides, in general, a better life, only 9% of the conventional farmers state this.On the other hand, 88% of conventional farmers think that conventional cotton provides a better life, while only 3% of organic farmers report that conventional cotton is better for their life.Hence, most of the farmers seem to be satisfied with their decision about whether to produce organic or conventional cotton.
Organic households are significantly (at 10% significance level) less food secure than conventional households according to the FAO's food insecurity experience scale, but we do not observe any significant difference in terms of our extended HFIES.Moreover, organic households have a significantly (even at 0.1% significance level) lower dietary diversity score over 24 h and consume significantly (at the 10% significance level) fewer vitamin A-rich foods over 24 h than conventional households.However, there is no significant difference between organic households and conventional households with regards to dietary diversity or the number of vitamin A-rich food groups consumed over seven days.Moreover, on average, organic households produce significantly fewer food crops than conventional households.Finally, the average land area cultivated with cotton by organic households is less than half the average land area cultivated with cotton by conventional households; the average income from cotton farming of organic households is roughly half the average income from cotton farming of conventional households; and the average household revenue (as proxy for household income 7 ) of organic households is less than half of the average household revenue of conventional households.All three of these differences are highly statistically significant. 6The empirical analyses were performed in the statistical software ''R'' (R Core Team, 2023) with the add-on packages ''AER'' (Kleiber and Zeileis, 2008), ''sandwich'' (Zeileis, 2004(Zeileis, , 2006;;Zeileis et al., 2020), ''xtable'' (Dahl et al., 2019), and ''stargazer'' (Hlavac, 2022).All data and code that we used for the empirical analyses are publicly available at Zenodo (Aihounton and Henningsen, 2024).
7 Getting precise information on the income of subsistence farming households in a survey is usually very challenging because the farm households usually neither do bookkeeping nor have an income declaration, they usually have difficulties recalling all revenues and expenditures from all incomegenerating activities over the past 12 months, and it is difficult to determine the value of self-produced and self-consumed agricultural products.Therefore, and in order to avoid too lengthy and tiring survey interviews, we used a simpler method to obtain an approximate indicator of total household income: We asked in the survey ''What is the share of cotton income in total household income?''.The respondent received 10 identical stones, which together should represent total household income.The respondent was asked to divide these 10 stones into two groups: income from cotton farming and income from other (agricultural or non-agricultural) income-generating activities.The number G.B.D. Aïhounton and A. Henningsen

Estimates
Table 2 presents OLS and IV estimates of the conditional associations of organic farming with our six indicators of food security based on two alternative specifications (with and without WTP as an additional control variable) as well as results of diagnostic tests of these regression analyses. 8As the F-statistics of the tests for irrelevant or weak instruments are much larger than ten, we conclude that our instrument is highly relevant.Furthermore, the falsification test proposed by Di Falco et al. (2011) does not find a statistically significant relation between the instrument and the outcome variable in any of our regression models, which indicates that there is no evidence of a correlation between the instrument and the error term in these regression models. 9Hence, we conclude that our instrument is valid.
In four of the twelve regression analyses (original HFIES with WTP variable, extended HFIES with and without WTP variable, and of stones that indicated the income from cotton farming was multiplied by 10% to obtain the proportion of income from cotton farming in total household income.Finally, we divided the revenue from cotton farming by this proportion in order to obtain a proxy for total household income, which we call ''household revenue''.Using our proxy ''household revenue'' instead of household income is a limitation of our study. 8 Complete results of these regression analyses are presented in Tables S10 to S15 in the Supplementary Material. 9The complete regression results of the falsification tests are presented in Tables S3 and S4 in the Supplementary Material.consumption of vitamin A-rich foods over seven days without WTP variable), the Hausman test indicates that OLS estimates are biased, while this test does not provide evidence against the unbiasedness of the remaining eight OLS estimations.
None of the OLS or IV estimations without the WTP variable indicates a statistically significant conditional association at the 5% level for any of our six indicators of food security.However, if we include the WTP variable for controlling for heterogeneity between organic and conventional households, which is not accounted for in traditional control variables, both the OLS and IV estimates indicate that organic farming is statistically significantly conditionally associated with higher food insecurity (both original and extended HFIES), lower dietary diversity score over 24 h, and a smaller number of vitamin A-rich foods consumed over 24 h.No statistically significant conditional association is found regarding the dietary diversity score over seven days or the number of vitamin A-rich foods consumed over seven days.
In order to assess the economic significance of the estimated conditional associations, we visualize the estimated magnitudes of these associations presented in Table 2 along with their confidence intervals and the variation of the observed values of the respective outcome variables in Fig. 1.In addition, this figure assesses the robustness of the regression results by presenting the sensitivity to omitted variables for two different values of  ℎ as described in Section 3.5, 10 estimates excluding observations from the district Péhunco, which has 10 These bounds and bounds for two additional values of  ℎ are also presented in Table S1 in the Supplementary Material.under the null hypothesis, both OLS and IV estimates are consistent, while OLS estimates are more efficient than IV estimates; under the alternative hypothesis, IV estimates are consistent, while OLS estimates are inconsistent.Rows 'Di-Falco test'' present the estimated coefficients (and their cluster-robust standard errors in parenthesis) of the instrumental variable, i.e., the share of organic farmers in the village, in regression analyses with Eq. ( 1), where the explanatory variable   is replaced by the instrumental variable and where only observations from conventional farmers are included (both the coefficients and the standard errors are multiplied by 100 for better readability).
a very low prevalence of organic farms, 11 and estimates excluding observations from villages that are dissimilar to villages with organic farming (i.e., only including observations from villages with organic farming and villages that are similar to these villages as explained in Section 3.1). 12 As already suggested by Table 2, Fig. 1 confirms that regression models that use the WTP variable to account for heterogeneity indicate that organic farming is statistically and economically significantly conditionally associated with higher food insecurity (both original HFIES and extended HFIES).While some model specifications indicate that organic farming is neither statistically nor economically significantly conditionally associated with the dietary diversity score over 24 h or the number of vitamin A-rich foods consumed over 24 h, some other model specifications indicate that organic farming is slightly conditionally associated with lower values of these two indicators of food security.Finally, all of our regression models indicate that organic farming is neither statistically nor economically significantly conditionally associated with the dietary diversity score over seven days or the number of vitamin A-rich foods consumed over seven days.

Pathways
In order to understand the pathways of the effects of adopting organic farming on food security, we analyze the conditional association of organic farming with mediating outcomes.Based on a standard nonseparable household model (e.g., Singh et al., 1986;Key et al., 2000;Henning and Henningsen, 2007, and Section A of the Supplementary Material), we identified six relevant mediating outcomes: • number of livestock species kept: as organic farming relies on organic fertilizers, adopting organic farming might increase the number of livestock species kept in order to produce more and more diverse organic fertilizers; if a household keeps a larger number of livestock species, it might result in larger consumption of livestock-based food and a more diverse diet of the household 11 Complete results of regression analyses without observations from the district Péhunco are presented in Tables S22 to S27 in the Supplementary Material.
12 Complete results of regression analyses excluding observations from villages that are dissimilar to villages with organic farming are presented in Tables S34 to S39 in the Supplementary Material.
• number of food crops produced: as organic farms are recommended to do intercropping and have a wide crop rotation, adopting organic farming might increase the number of food crops produced by the household; if a household produces a larger variety of food crops, it might result in a larger supply of self-produced food crops to the household and a more diverse diet • land area cultivated with cotton: as organic cotton farming is much more labor-intensive than conventional cotton farming, organic farms are recommended to have a wider crop rotation, and organic certification agencies often restrict the area that a single farmer can cultivate with cotton, adoption of organic farming likely decreases the land area that the farm cultivates with cotton; as cotton farming is the predominant source of cash for the farmers in our study area (some farms also sell some parts of the food crops that they produce), a decreased area cultivated with cotton might reduce the cash income of the household and, thus, its ability to buy food • land area cultivated with food crops: as we expect that adoption of organic farming reduces the land area cultivated with cotton, these households may have a larger land area available for producing food crops; however, the slightly higher labor intensity of organic farming of food crops and the considerably higher labor intensity of organic cotton production might also reduce the land area cultivated with food crops; the land area cultivated with food crops likely affects the supply of self-produced food crops to the household • income from cotton farming: adoption of organic farming reduces the costs of fertilizers and pesticides for cotton farming and gives farmers a price premium for the cotton that they sell but it is expected to reduce the area cultivated with cotton and the cotton yield; hence, the effect of adopting organic farming on income from cotton farming might be positive or negative; as cotton is the predominant source of cash for farmers in our study area, it likely affects the household's ability to buy food • household revenue (proxy for household income, see footnote 7): adopting organic farming likely affects household income through various channels (e.g., through some of the other mediating outcome variables); household income affects the household's ability to buy food.
We analyze the conditional association of organic farming with these mediating outcomes in the same way as we analyzed the conditional associations with our primary outcome variables (i.e., the six mediating outcomes are   ;  = 7, … , 12 in our main regression model (1)).Table 3 summarizes OLS and IV estimates of the conditional associations of organic farming with the six mediating outcomes with two alternative specifications (with and without WTP as an additional control variable) along with diagnostic tests of these regression analyses. 13The F-statistics of the tests for weak instruments are much 13 Complete regression results are presented in Tables S16 to S21 in the Supplementary Material.larger than ten, which indicates that our instrument is highly relevant.Moreover, the falsification test suggested by Di Falco et al. (2011) does not find a statistically significant relation between the instrument and the mediating outcomes in any of our regression models, suggesting the absence of correlation between the instrument and the error term.14Hence, we conclude that our instrument is valid.Hausman tests indicate that OLS estimates are biased in four of the twelve regression analyses (food crop area without WTP, and number of livestock species, cotton income, and household revenue with WTP), while OLS estimates do not significantly differ (at the 5% level) from IV estimates in the remaining eight estimations.
While the IV estimation with the WTP variable indicates a statistically significant (at the 5% level) conditional association between organic farming and a higher number of livestock species, no statistically significant conditional association can be found between organic farming and the number of food crops produced in any of the OLS or IV estimations.In contrast, all OLS and IV estimations with and without WTP find a statistically significant conditional association between organic farming and a larger land area cultivated with food crops (at the 1% significance level), a smaller land area cultivated with cotton, a lower income from cotton farming, and a lower household revenue (at the 10% significance level for the IV estimation with WTP for household revenue and at the 1% significance level for the remaining eleven estimations).
Analogously to Fig. 1, Fig. 2 visualizes the confidence intervals of the magnitudes of the estimated conditional associations along with the variation of the observed values for assessing the economic significance of the estimated conditional associations as well as their sensitivity to omitted variables 15 and their robustness to using two different sub-samples. 16 Fig. 2 confirms results in Table 3 that organic farming is not or only slightly positively conditionally associated with the number of livestock species and that there is no economically significant conditional association with the number of crops produced, but that organic farming is clearly conditionally associated with a larger land area cultivated with food crops, a smaller land area cultivated with cotton, a lower income from cotton farming, and a lower household income.

Conclusion and policy implication
The prevalence of organic farming and other sustainability standards is becoming increasingly important around the globe.While 15 The bounds for two additional values of  ℎ are presented in Table S2 in the Supplementary Material. 16Complete results of regression analyses without observations from the district Péhunco are presented in Tables S28 to S33, while complete results of regression analyses excluding observations from villages that are dissimilar to villages with organic farming are presented in Tables S40 to S45 in the Supplementary Material.a plethora of studies investigates the effects of organic farming on agronomic and economic aspects of smallholder farming in developing countries, only very limited evidence exists on its implications for the food security of the farmers.This study set out to analyze the effects of organic farming on the food security of farm households in Benin.As failure to properly deal with unobserved characteristics may lead to biased estimates, we use the following three distinct approaches to addressing potential omitted variable biases: (i) using a control variable that captures a household's marginal utility of (or willingness to pay for) the adoption of either organic or conventional farming practices (see Verhofstadt and Maertens, 2014;Bellemare and Novak, 2017;Ruml and Qaim, 2021), (ii) the approach suggested by Oster (2019) for checking the sensitivity of our estimates to unobserved heterogeneity, and (iii) instrumental variable regression.As we use observational cross-sectional data, we cannot be sure that these three approaches completely address all potential sources of endogeneity.Hence, as a precaution, we do not interpret our results as causal effects but as conditional associations.
According to our findings, organic farming is conditionally associated with a lower food security of farm households in Benin by a statistically and economically significant magnitude.While we do not find a statistically or economically significant conditional association with dietary diversity or the number of vitamin A-rich foods consumed within seven days, organic farming tends to be conditionally associated with a slightly lower dietary diversity and number of vitamin A-rich foods consumed over 24 h.This indicates that organic farming tends to be conditionally associated with a lower frequency of the consumption of some food categories, including vitamin A-rich foods, from daily consumption to less frequent (e.g., weekly) consumption.In spite of the intention of sustainability standards to generate positive welfare impacts, we find that adopting organic farming practices jeopardizes some aspects of food security of the involved farm households.These results contradict those of Chiputwa and Qaim (2016) and Meemken et al. (2017) but corroborate those of Meemken and Qaim (2018), who point out that organic farming is not a panacea for sustainable agriculture and food security and emphasize the need for smart combinations of organic and conventional farming practices to achieve sustainable agriculture.
Our analysis of pathways indicates that organic farming is conditionally associated with substantially lower household income due to lower income derived from cotton farming given a substantially smaller land area cultivated with cotton.On the other hand, organic farming is conditionally associated with a larger land area cultivated with food crops, but this is not sufficient to counteract the lower income derived  from cotton farming, perhaps due to lower crop yields from organic farming compared to conventional farming.Finally, we do not find a noteworthy conditional association between organic farming and the diversity of crop or livestock production, which rejects the hypothesis that organic farming improves dietary diversity due to more diverse production.
While organic farming is expected to have several positive effects on welfare as a result of, e.g., improved environmental and water quality, increased biodiversity, reduced contamination of food by pesticides, and improved occupational safety, our results suggest that the many policies and programs that promote organic farming in developing countries may well have unintended undesirable side effects on the participating smallholder farmers, at least when circumstances are similar to those of the farmers in our empirical analysis.We recommend that governments, NGOs, donors, international corporations, etc. that promote organic farming or implement schemes for organic farming in developing countries analyze how their policies and programs affect the participating households' food security (e.g., with RCTs) so that these stakeholders as well as farmers and consumers (e.g., in rich countries) can make informed decisions.Furthermore, the results of these analyses may indicate how policies and programs can be designed to avoid unintended and undesirable effects on the participating farmers' food security.
Although each of the three strategies that we used to address omitted-variable biases has some potential weaknesses, together they indicate that our results are robust to various potential misspecifications.To obtain even more reliable estimates of the effects of adopting organic farming on food security as well as on other livelihood indicators, future studies could use panel data sets (e.g., as done by Meemken et al., 2017) that allow for estimation methods that take unobserved heterogeneity into account.Alternatively, future studies could use randomized controlled trials (RCTs).Replications of our study and the studies by Chiputwa and Qaim (2016) and Meemken et al. (2017) in other contexts will assess the generalizability of existing

Fig. 1 .
Fig. 1.Estimated conditional associations of organic farming with outcomes.Notes: The gray vertical bars indicate the distribution of the values of the respective outcome variable in our data set.The vertical black line indicates the mean value in the sample.The axis on the right-hand side indicates the estimation method, where a ''A'' indicates an estimation with all observations, a ''P'' indicates an estimation without observations from Péhunco district, an ''S'' indicates an estimation with only observations from similar organic and conventional cotton growing villages, a ''−'' indicates an estimation without WTP as a control variable, a ''+'' indicates an estimation with WTP as a control variable, an ''o'' indicates an OLS regression, a ''b'' indicates Oster bounds based on the OLS regression, and an ''i'' indicates an IV regression.For each OLS or IV regression, the thick part of the horizontal line indicates the 50% confidence interval and the thin part of the horizontal line indicates the 95% confidence interval for the outcome of an average farm that adopts organic farming.For each set of Oster bounds, the thick part of the horizontal line indicates the bounds for ℎ =  + (  −   )and the thin part of the horizontal line indicates the bounds for  ℎ = 1.Hence, horizontal lines on the left-hand side of the vertical black line indicate that adopting organic farming is conditionally associated with lower values of the outcome variable, while horizontal lines on the right-hand side of the vertical black line indicate that adopting organic farming is conditionally associated with higher values of the outcome variable.When an OLS estimator is rejected by a Hausman test at the 5% significance level, the corresponding horizontal line is gray instead of black.
G.B.D.Aïhounton and A. Henningsen
Notes: HFIES = Household Food Insecurity Experience Scale; numbers in parentheses are standard deviations; two-sample -tests for equality of mean values are applied to assess differences in continuous variables; Pearson's  2 -tests for equal proportions are applied to assess differences in binary and categorical variables; in tests for binary variables, Yates' continuity correction is applied; in case of categorical variables, column 'Difference' indicates the  2 -value; asterisks denote the following levels of statistical significance obtained by these tests: *** = P < 0.01, ** = P < 0.05, and * = P < 0.1.

Table 2
OLS and IV regression results of the conditional associations between organic farming and outcomes.p < 0.1; * * p < 0.05; * * * p < 0.01.Rows 'OLS estimate' and 'IV estimate' indicate the coefficients of organic farming (  ) on the respective outcome variable estimated by the OLS method and the IV method, respectively, where the numbers in parenthesis are cluster-robust standard errors.Rows 'Weak instrument test' present F-statistics indicating the statistical significance of the predicted probabilities of being an organic farmer in regression analyses of   on the predicted probabilities of being an organic farmer and the control variables   .Rows 'Wu-Hausman test' present test statistics of Wu-Hausman tests of the null hypothesis of exogneity of   , i.e., COV(  ,   ) = 0; *

Table 3
OLS and IV regression results of the conditional associations between organic farming and mediating outcomes.Livestock species' abbreviates 'Number of livestock species kept'; 'Food crops' abbreviates 'Number of food crops produced'; 'Food crop area' is the IHS of the food crop area measured in 10 ha; 'Cotton area' is the logarithm of the cotton area measured in ha; 'Cotton income' is the IHS of the cotton income measured in million FCFA; 'Household revenue' is the logarithm of the household revenue measured in FCFA.For log-transformed and IHS-transformed outcome variables, rows 'OLS estimate' and 'IV estimate' indicate the semi-elasticities instead of the coefficients of organic farming.For log-transformed dependent variables, we calculate the semi-elasticity by equation (S13) in Section D of the Supplementary Material, while for IHS-transformed dependent variables, we calculate the semi-elasticity with equation (S7) in Section C of the Supplementary Material.We calculate the standard errors of the semi-elasticities with the delta method as described in Section D of the Supplementary Material.See further notes below Table2.