Data were collected from Gwanda and Mangwe districts of the Matabeleland South province of Zimbabwe in 2019. The districts are semi-arid, in the same agro-ecological zone IV (<450 mm annual rainfall). A multistage sampling procedure was used to select households, who were mainly representatives of communal farming areas of the study districts (Figure 1). First, 10 wards (municipalities)–four from Gwanda and six from Mangwe–were selected using a probability proportional to size (PPS) sampling approach. Second, three villages were randomly selected to represent sampled wards of communal farming communities, resulting in 30 study villages. Following this, a sampling frame was developed in selected villages with the help of extension agents and local administrative offices. The overall sampling frame covered 6,780 households, with a sampling frame for specific villages ranging from about 150 households in the smallest village to 300 households in the largest village. In the final stage, study households were selected proportional to population using systematic random sampling. With an equal probability of households in the sampling frame being selected, a random number was generated by dividing the total number of households in the sampling frame of a village by the target sample size of the village. This random number was then used a denominator of counting to select participants until the target sample size was reached in each village. This process produced a proportionate representation in the sample where samples from larger villages were more frequent than those from smaller villages. Following this procedure, 645 households were interviewed using a semi-structured questionnaire. However, this paper used data from 625 households due to missing market participation and outcome data. ¹

Sixty percent of the households were male-headed, with an average age of about 57 years and 7 years of completed schooling (Table 1). The average household had six family members, two hectares of land and 11 tropical livestock units (TLU). ² Farming was the main occupation for 78% of households. The average per capita income from all sources was about 3,137 Zimbabwean dollars (Z$) (US$1 = Z$15.60 on average at the end of 2019), while the average household asset value was about Z$21,938. About 46% are members of farmers’ groups. Moreover, 76 and 29% reported, respectively, accessing extension and credit services. Grazing rangelands was the dominant livestock feed, with 32% of the farmers receiving feed support during drought seasons.

The livestock module elicited households’ livestock ownership, most important livestock keeping purpose, livestock input use and marketing related to cattle and goats. Livestock market participation was defined as whether a household sold live animals and livestock products over the 12 months before the survey. About 49% participated in livestock markets. Supplementary Tables A1 and A2 in the Supplementary material present information used to construct the outcome variables. On average, 63 and 93% of the households owned cattle and goats, respectively. All households reported owning at least one type of livestock. Households’ livestock production-orientation is elicited using the most important purpose for each type of livestock: (1) sale of live animals, (2) sale of animal products, (3) food consumption, (4) crop farming (draft power and manure), (5) precautionary savings and insurance, and (6) signaling social status.

Supplementary Table A1 describes the distribution of main purposes of livestock ownership and production. The sales of live animals and livestock products were considered as market-oriented purposes (Moll, 2005). As shown in Supplementary Table A1, the principal motivation for keeping livestock is income from the sale of live animals (72%), while a few households also sell livestock products (3%). About 15% reported keeping livestock for food consumption, and 9% kept livestock for precautionary savings and self-insurance against shocks. But livestock services for crop farming were rare. Overall, about 76% of livestock ownership is market-oriented (Table 1). As hypothesized, livestock market participants are more likely to engage in market-oriented production.

The remaining rows of Supplementary Table A1 provide information on improved livestock inputs. About 53% vaccinated their livestock, but only about 24% fed their livestock supplementary commercial concentrates. Most households own indigenous breeds, with only about 15% exotic or hybrid breeds. In this study, improved livestock input use is defined as adopting at least one of the modern livestock inputs: vaccination of livestock, feeding commercial concentrates, or keeping exotic or hybrid breeds. Overall, 63% of the farmers used at least one of the improved livestock inputs (Table 1). Market participants are more likely to use modern livestock inputs than non-participants.

As outlined, household dietary diversity (HDD) and per capita food expenditure are jointly used to measure food security. The HDD is a qualitative measure for nutrient adequacy and quality of the household diet (Swindale and Bilinsky, 2006). HDD is based on households’ food consumption in the past 7 days. These food items were classified into 12 food groups. Supplementary Table A2 in the Supplementary material shows patterns in individual food groups. Table 1 shows the average household consumes from about six food groups. Market participants consume a more diverse diet than non-participants. Household food expenditure is computed as the total value of food consumed from own production, market purchase and gifts and other sources. On average, households spent Z$1,272 on food per capita (Table 1). However, it does not differ between market participants and non-participants.

While suggestive, simple mean comparisons can be misleading as market participants and non-participants might differ in observed and unobserved characteristics. This may be true as several covariates significantly differ between market participants and non-participants (Table 1). Below, we use empirical strategies that address this concern.

Our goal is to estimate the effects of livestock market participation on livestock production-orientation, modern livestock input use, HDD, and food expenditure. The empirical specification is:

where Y i is a vector of our outcome variables for individual i, M i is livestock market participation status for individual i, X i is a vector of covariates, and ε i is a random error term. β s are vectors of parameters to be estimated. We control for a wide range of individual, household and contextual variables. District dummies and village fixed effects are included throughout the regressions to capture observed and unobserved characteristics due to differences in local governance, market opportunities and political contexts beyond the controlled covariates. We run ordinary least squares (OLS) regressions for HDD and per capita food expenditure. But we estimate a linear probability model (LPM) for market-orientation and improved livestock input use, as LPM provides tests for the appropriateness of the IV approach (see below). We transform per capita food expenditure to its logarithmic value to address distributional issues.

As indicated, livestock market participation may be endogenous in the outcome models for different reasons. First, there may be potential reverse causality from the outcomes to market participation. Second, unobservable characteristics may drive both the outcomes and market participation simultaneously beyond the controlled covariates. For example, farmers may decide to participate in markets due to their innate entrepreneurial skills and technical abilities in understanding and profitably tapping market opportunities. Third, measurement errors in market participation may lead to a lower bound estimate (Theil, 1971), as typical in recall household surveys. Thus, estimating Equation (1) requires addressing these concerns, where corr( M i , ε i ) ≠ 0. We use an instrumental variables (IV) approach to address threats for identification. We estimate the following two-stage model:

M i = ∝ 0 + ∝ 1 Z i + ∝ 2 X i + u i and(2)

where Z i is a vector of valid instruments, u i is the error term of the first-stage regression, and other variables are as explained in Equation (1).

As often, the challenge is to find credible instruments, Z i . Theoretically, valid instruments are those correlated with the endogenous variable but not directly correlated with outcome variables. Using insights from the literature, we identify two relevant instruments for livestock market participation: distance to the nearest permanent market and farmers’ perception about livestock market opportunities. Distance to markets is a key indicator of smallholder market access, as it influences transaction costs and ease of doing business (Renkow et al., 2004; Melesse and Cecchi, 2017). Most smallholders in remote rural areas have poor access to transport and market infrastructure, and in most cases, they must transport their livestock products themselves to the market. As a result, shorter distances to markets are likely linked with lower transaction costs and increased market participation (Key et al., 2000). In addition to access to markets, the propensity of smallholders’ market participation is likely to be governed by their capacities and attitudes toward available market opportunities and benefits of participation (FAO, 2014). Smallholders would be encouraged to invest in their livestock and practice market-oriented management if they could access improved market opportunities and expect better returns from their investments. Thus, we use smallholders’ perceptions about livestock market opportunities over the last 10 years to capture their attitudes toward market participation. Table 1 presents data on the two instruments. On average, farmers are located about 25 km away from the nearest major market. About 52% of the farmers believed that livestock market opportunities had improved over time, which might signal incentives to participate in livestock markets.

The validity of an IV strategy rests on two criteria: the relevance and exclusion restriction criteria. The relevance criterion is that instruments should be good predictors of the endogenous regressor. To formally examine the relevance criterion, Table 2 reports the short-hand form of the first-stage regression results. The instruments are significantly correlated with market participation, confirming that they are relevant predictors of market participation. In addition to the instruments, market participation is significantly correlated with several other control variables, including male, household size, livestock herd, household asset, per capita income, group membership and access to extension, with expected signs of association. This implies that farmers with larger herds, larger household sizes, more asset and income, group membership and accessing extension services, are more likely to participate in livestock markets.

The second criterion for valid instruments is the exclusion restriction. It requires that instruments should not be directly correlated with outcome variables, except via the endogenous variable. This is difficult to compellingly satisfy and more so to prove. Households have limited scope to influence distance to markets. One concern is that households concerned about market access may migrate to areas characterized by better market access. While this could violate the exclusion restriction, we believe this does not pose a threat to our data because of the land ownership structure. Zimbabwe’s dominant land tenure systems, including our study areas, are communal lands managed under customary tenure (Tatsvarei et al., 2018). In most cases, households are not allowed to transfer land freely to others. Land can only be transferred generationally through marriages and in the event of the death of the landholder. The absence of active land markets restricts rural households’ migration, while individual labor migration is high.

Regarding the livestock market perception instrument, it is less likely to directly affect livestock production decisions other than through market prospects. However, a potential concern with the livestock market perception instrument, which is not necessarily connected with the exclusion restriction, is that it may be endogenous in the livestock market participation model, i.e., the first-stage regression. This may be a valid concern, as their participation in markets might partly shape individual’s perceptions about market opportunities. While our dataset does not include valid instruments to address this potential threat, we maintain this instrument because it helps us to capture non-infrastructure related drivers of market participation. A placebo test where the instruments are regressors of outcomes of non-market participants confirms the absence of a direct relationship between the instruments and outcome variables. Further, Supplementary Table A3 in the Supplementary material presents further standard statistical tests that support the validity of our instruments (Supplementary Table A3). The Anderson canon and Stock and Yogo (2004) tests reject that the endogenous regressor is weakly identified (p < 0.01). The Sargan and Hansen tests (i.e., over-identification restriction tests) fail to reject the null of zero correlation between instruments and the error term of the models, implying that our instruments are exogenous and valid, requiring at least one instrument to be exogenous (Murray, 2006). In our case, distance to the nearest market can largely be considered as exogenous. While these tests can help us minimize the concern that our results might be affected by this contaminated instrument, we realize that some reservations may remain about the exogeneity of the livestock market perception in the first-stage regressions. Despite this, we believe that our results remain informative, even in the presence of such contaminated causality in the first-stage regressions.

We proceed in steps to establish the effect of market participation on the outcome variables. Table 3 presents results from simple regressions, suggesting significant correlations between livestock market participation and all outcome variables. The coefficients are strong with expected signs, ranging from 0.113 for improved livestock production to 0.444 for HDD.

Table 4 also reveals other correlates of the outcome variables. Farmers with farming as their main occupation are more likely to engage in market-oriented livestock production. Those who received feed support during drought seasons are less likely to keep livestock for market purposes. Land, livestock and asset ownership are significantly associated with improved livestock production. Similarly, male farmers are more likely to keep livestock for market purposes and use modern livestock inputs than female farmers. Households with more livestock and asset tend to consume more diversified diets and spend more on food. Household dietary diversity increases with education level and access to extension services, highlighting the importance of knowledge and information for nutrition behavior and outcome (Melesse and Van den Berg, 2021). Conversely, households led by an older head, larger in size, members of a group and primarily depending on farming a livelihood strategy were less food secure.

However, as discussed earlier, market participation may be endogenous. Indeed, the Wu–Hausman and Durbin–Wu–Hausman tests (Supplementary Table A3) confirm that livestock market participation is endogenous as exogeneity is rejected for all outcomes (p < 0.05). We employ an IV strategy to attenuate this concern, using distance to the nearest permanent market and household livestock market perception as instruments. Table 4 presents the IV estimation results for all outcome variables. The coefficients for market participation remain significant with the expected sign but are generally higher for the IV estimates than those from the simple regressions (Table 3). Such differences between simple regression and IV estimates are consistent with measurement errors that lead to an attenuation bias toward zero in simple regressions (Theil, 1971). IV approaches often mitigate measurement errors (Gujarati, 2003), because the identification in the IV comes from households responding to changes triggered by the instruments. Thus, the instrumented coefficients can be interpreted as Local Average Treatment Effects (LATE).

Finally, we control for district dummy and village fixed effects throughout all regressions, partially controlling for observed and unobserved community-level characteristics beyond controlled covariates, such as differences in local governance, market opportunities and political contexts. Thus, our results are less likely to be driven by omitted variables. Overall, the results indicate that market participation has a positive effect on the four outcome variables.

While other competing hypotheses cannot be ruled out, we look at smallholder livestock keepers’ marketing behavior and decision-making process to get some insight into potential mechanisms for the differences between market participants and non-participants. We considered whether smallholders make marketing decisions consistent with a commercial approach, use market information for decision-making, choose formal market channels, and capture value through improved practices. For this purpose, farmers were asked to state their agreement on several statements using a five-point scale, ranging from “completely disagree” to “completely agree.” Table 5 summarizes the percentage of respondents that “agree” or “completely agree” with each statement.

The results suggest that market participants are more likely than non-participants to actively look for livestock market information, claim to negotiate market prices, keep records of their costs and returns, and sell livestock through formal market channels, which may require more commercial practices. Market participants are also more likely to capture livestock market value through formulating prices and valuing livestock according to market conditions, participating in commodity groups to get better market terms and selling livestock during high-price months (seasonal arbitrage). However, only 8% of the respondents sell livestock when market prices are high, implying that non-price factors drive most livestock sales. This suggests that livestock marketing decision-making is significantly influenced by other competing needs and interests, and socio-cultural considerations, which in many cases even overshadow the purely economic ones. The decision to sell livestock is often needs-driven; livestock sales peak during a period (November to April), when farmers are in need of cash to balance food deficits, schools reopening and hence the need to pay for children’s school fees and buying inputs for the next agricultural season (Homann-Kee Tui et al., 2022). For farmers with smaller herds and less cropland in a farming context of variable climate, pests and diseases, economic upheavals, and mortality losses from previous seasons, holding back livestock to time sales with better market terms may overshadow purely economic goals.

Particularly, the sale price offered to the farmer is not the most important consideration when deciding when and to whom to sell. This is manifested in the low off-take percentage and lacking participation in commercial production. This can also reflect market imperfections, as livestock markets in rural Zimbabwe differ in effectiveness, limiting the transfer of information and incentives to smallholder farmers (Homann-Kee Tui et al., 2022). In Districts, like Gwanda, livestock markets are more commercialized and target high-income markets, as owners of large cattle and goat herds exist that regularly supply livestock, along with many smallholder farmers owning smaller livestock and selling occasionally. Efforts to improve market participation must distinguish between contexts with more and less developed livestock markets, and within those contexts, the behaviors of market-oriented farmers and those that are needs-driven.

We assess the robustness of our findings in several ways. First, we consider different functional forms for some outcomes. Livestock market orientation and modern input use are binary, and the HDD is a count variable. While applying the linear model provides tests for the IV strategy, using linear models for these outcomes may not be appropriate. As such, we estimate the IV probit model for market-orientation and farming intensity and IV Poisson for HDD, which fits the exponential conditional mean and implements the generalized method of moments (GMM) estimator. The results are significant and robust (Table 6), suggesting that the non-linearity of these variables does not drive the results.

Second, because livestock market participation is a binary variable, (i) there may be a possibility for non-linear relationship between market participation and outcome variables, and (ii) it does not fit the continuous endogenous variable assumption of estimated models (e.g., the IV probit model). We apply a control function estimation, an appropriate equivalent to instrumental variables estimation in the potential presence of thresholds and non-linearity in effects. It also accounts for the binary nature of the treatment variable and addresses endogeneity by implementing the control function estimators (Wooldridge, 2010). The results presented in the Supplementary material confirm strong and significant effects of market participation on outcome variables (Supplementary Table A4). The coefficients are interpreted as improvements from potential outcomes that would otherwise be attained when farmers do not participate in markets.

Finally, we carry out multiple hypothesis testing (MHT) to assess whether reported results are true effects and can jointly be identified. We follow the procedure by Barsbai et al. (2020), ³ which is an improvement on the previous procedure (mhtexp) of List et al. (2019). The Barsbai et al. (2020) procedure is a regression-based approach that permits multiple outcomes or treatments with other control variables. For completeness, we first perform the MHT by excluding other covariates and later by including them. Results for value of ps adjusted for multiple hypotheses based on different approaches are presented in the Supplementary material (Supplementary Table A5). While there are some differences in coefficient magnitudes in the specifications with and without covariates, the effects of market participation largely remain significant and consistent. A minor difference is a slight change in the significance levels for the Bonferroni correction, which is the most conservative correction. Overall, our results are robust to the possibility of false discoveries.