Experimental analysis of adoption of domestic mopane worm farming technology in Zimbabwe

Abstract

Seasonal outbreaks of mopane worms, caterpillars of the moth Imbrasia belina, provide an important source of income and food for rural people in the semi-arid woodlands of southern Africa. Outbreaks are erratic and periodically fail to produce caterpillars of harvestable size, which has generated interest in a new technology for domestic farming of mopane worms at the household level. Using a choice experiment, the authors explore the preferences of harvesters across alternative farm management scenarios in four villages located in the mopane woodlands of rural Zimbabwe. The results highlight preference heterogeneity across investment cost, labour effort, harvest price and harvest yield attributes depending on age, location and latent class decomposition. They conclude that design specifications need to respond to socio-ecological variability and significant household investment constraints in order for the technology to be adopted by rural households living under extreme economic hardship.

1. INTRODUCTION

Caterpillars of the emperor moth Imbrasia belina are an important natural resource for rural people living in the mopane woodlands of Botswana, Namibia, northern South Africa and southern Zimbabwe (Stack et al., 2003). The moth is widespread but is particularly common in areas dominated by the mopane tree (Colophospermum mopane). The larvae (mopane worms) have two main seasons of emergence in December–January and April–May. During these times mopane worms can be extremely abundant in the wild and are widely harvested for food or income (Gardiner, 2003). Styles (1994) estimated a population of 9500 million mopane worms in South Africa's mopane woodlands, of which 40 per cent were harvested by rural people, who are often poor women. The market value of this harvest was estimated to be US$84 million.1 The areas dominated by mopane generally have low agricultural potential. Cultivation is risky, with regular failure of staple grain crops producing a high degree of livelihood vulnerability. Mopane worms thus represent an important source of food and income at a time when many households have limited livelihood alternatives (Frost, 2005). Nevertheless, outbreaks in the wild are unpredictable and the caterpillars may fail to grow to a harvestable size. This has generated interest in options for small-scale farming of mopane worms at the household level.

In Zimbabwe, a contracting economy and high unemployment in recent years has led to increased dependency on mopane worm harvesting. Spiralling inflation has increased the price of agricultural inputs and food imports, so that mopane worms have grown in popularity in urban areas as an affordable substitute source of protein (Stack et al., 2003). In the southern arc of mopane woodlands, most households will harvest mopane worms when outbreaks occur, with up to three-quarters of households selling some of their harvest (Frost, 2005). Rural–urban value-chain analysis indicates that consumers can pay four to five times the price received by a rural harvester. These trade linkages stretch across neighbouring borders into Botswana, the Democratic Republic of Congo, South Africa and Zambia. It is estimated that income from mopane worm harvesting can contribute up to one-quarter of total annual cash income for rural households (Stack et al., 2003). These factors make mopane worms an important natural resource for rural people and contribute to livelihood security in various ways:

• Offsetting seasonal shortages of income or food.

• Buffering households against drought, illness or other shocks.

• Supplementing household expenditure on education, health, food, agriculture or other productive or consumptive activities.

Recent research has attempted to identify how markets and marketing for mopane worms might be improved (Stack et al., 2003), how the resource can be more sustainably managed and used (Frost, 2005), and what opportunities there might be for domestic farming of mopane worms by households (Gardiner, 2003; Gardiner & Gardiner, 2005). Findings indicate that important biotic factors, principally disease and parasite pressure, limit mopane worm populations in the wild (Ghazoul, 2006).2 Wild outbreaks are unpredictable, being influenced by a complex range of biotic, climatic and other factors, which makes harvesting undependable. Research indicates that some of the biotic factors can be controlled or their negative impacts reduced in semi-domesticated farming systems. This has led to trials being carried out to test the feasibility of a household-scale semi-domestication system for breeding and processing mopane worms for income and food benefits.

Exploring social adoption of domestic mopane worm farm management provides information to decision-makers in the process of designing feasible and desired interventions. This is useful where adoption of technology may be constrained by factors that are unanticipated or simply unknown. Analysis of experimental adoption scenarios allows a provisional quantification of livelihood gains with an understanding of how adoption preferences may vary by age, location or other latent factors. Specifically, we explore the following questions:

• Which domestic mopane worm farm-management technology options can meet the needs and offset the constraints of potential users?

• How will trade-offs between users' investments and returns influence choice?

• Which management attributes will be likely to influence adoption by different social groups?

In the remainder of the paper, Section 2 sets out the constraints to adoption of rural technology, and explains how choice experiments offer a method to model exploratory scenarios, Section 3 describes the study villages and context, Section 4 presents the descriptive results and the modelling analysis, and Section 5 concludes.

2. ODDS AGAINST ADOPTION

Perret and Stevens (2006) argue that the ‘odds against adoption’ of technology by rural users are often high because there is frequently a disconnection between the developers of a technology and its potential users. While the development literature has recently converged on placing people rather than technology at the ‘centre of development’, this often seems to be rhetoric rather than a guiding principle for designing interventions and policies that meet the priorities of the intended beneficiaries (Chambers, 1988; Crewe & Harrison, 1998; Hope, 2006). Adoption is influenced by attitudes, cultural norms, assets and capacities, and wider economic and environmental conditions. Being able to provide defensible evidence of a likely social response to an adoption opportunity requires understanding of these influences. A leap of faith is often required by potential adopters, as they not only have to accept the risks inherent in most new technologies but must also weigh up the uncertainties of successful adoption under conditions of environmental, economic, political and social variability, and shocks. This is likely to require adapting rather than adopting the technology – see Perret and Stevens (2006) for a discussion. This is a subtle distinction but one with significant implications for potential users and for the likelihood of sustained adoption.

Under conditions of scarcity and risk, adoption may be slower than anticipated because of a failure by those promoting a technology to understand the constraints beyond how place, culture and existing practices influence user adoption. Dynamic and temporal uncertainties are often beyond the competence of technological design but need to be estimated when exploring the likelihood of adoption by potential users. Exploring users' preferences when faced with plausible adoption scenarios can provide decision-makers and technology designers with evidence of the probable adoption responses of diverse groups of intended users to different design scenarios. Choice experiments offer one approach to explore and evaluate such uncertainties in an effort to improve the match between technology design and user adoption.

2.1 Exploring choices under uncertainty

Choice experiments offer a method to explore social preferences for proposed or predicted future scenarios that cannot be assessed using existing knowledge (Louviere et al., 2000). The results allow decision-makers to evaluate various design scenarios before committing funds and capacity to a proposed intervention. A key feature of a choice experiment is that it prompts respondents to make trade-offs between alternative scenarios that vary across attributes and attribute levels. For example, trade-offs in labour investment can be explored across a range of labour days to estimate how important and sensitive this attribute is compared with other identified design attributes. Choice experiments provide a flexible exploratory tool that has been commonly used for many years in environmental management, marketing, transport economics, medicine and psychology, with the methodological basis, design criteria and econometric models well established.

A limitation of the technique is that choices are shaped by the way the options are framed. Scoping analysis and identification of key attributes are critical to structuring the experimental design. An advantage of choice experiments is that they combine qualitative and quantitative social research methods. Canvassing stakeholder perceptions beforehand can enrich the experimental design, which can then be iteratively tested and modified before being implemented more widely. The results from the analysis can be referred again to stakeholders for validation, thus combining qualitative methods of stakeholder engagement to increase the internal validity of selecting attributes and attribute levels for an experimental design that can be used to elicit data for quantitative analysis. Moreover, experimental designs can be illustrated pictorially, which increases the participation of less literate respondents, who may often be an intended beneficiary group.

3. STUDY LOCATION AND METHODS

Four villages were selected in the mopane woodland zone of Matobo district, southwestern Zimbabwe, where biological and socio-economic studies had been conducted since 2001 as part of the mopane worm project (Ghazoul, 2006) (Figure 1). These villages are not necessarily representative of the wider rural population in Zimbabwe or beyond. Nevertheless, they reflect the general conditions of rural development in southern Africa, with high levels of income poverty and diversified livelihoods involving cultivation, livestock husbandry, migration, dependency on remittances and reliance on harvesting environmental goods and services (Frost, 2003). All the villages are located in the arc of low-lying (less than 1000 metres above sea level), semi-arid mopane woodland across southern Zimbabwe, a region characterised by low rainfall (300–700 mm per year) with high inter-annual and intra-annual variability, high temperatures (a mean annual temperature of 28°C) and high potential evaporation rates (1600 mm per year).

Figure 1: Location of study sites

Sasane and Kapeni villages are located in Beula Ward in Mambali Communal Area, approximately 190 km south of Bulawayo, in an area of relatively high mopane woodland density. Ndiweni and Mahetshe villages are located further north in Madwaleni Ward in the Tshatshani Communal Area, approximately 100 km from Bulawayo by road. Mopane woodland density in this area is lower due to woodland conversion to arable land. Thirty households were randomly sampled in each village using transect walks with systematic sampling of every nth household depending on the size of the community. Questionnaires were administered in Ndebele or Shona based on an English-language version that had been designed and pre-tested in the scoping phase in September 2005. Prior to the pilot phase, the applicability of the methodology and questionnaire design was discussed in a workshop with local researchers and government forestry officials who had extensive knowledge of the relevant biological, social, economic, marketing and institutional issues. This pilot phase led to revisions and amendments that determined the final experimental design. All of the respondents confirmed that they and their households collected mopane worms. More than four in five respondents were women, which required weights to be applied in the data analysis. A skewed profile of female respondents is expected for social and economic reasons. Nevertheless, the sample ratio of 24 men to 100 women is significantly lower than expected from both earlier project studies conducted by the authors, which gave a 63:100 male:female ratio for mopane worm collection, and the results of the 1992 national census, in which the gender ratio for Matobo district in the 15-years-and-older age group was 77:100. To correct for potential bias, a 75:100 male:female ratio is calculated by the standard approach of weighting data by the inverse probability of selection.

3.1 Questionnaire design

A household questionnaire was used to elicit information on current harvesting practices and the social impacts of collecting mopane worms. The questionnaire comprised four sections: household selection and measures of data quality; mopane worm harvesting; the choice experiment; and household characteristics (Hope et al., 2006). The questionnaire took, on average, 46 minutes to complete (range of 25–75 minutes). The average age of the respondents was 45 years (range of 15–85 years).

3.2 Choice experiment design

Following a pilot phase, attributes and attribute levels were collectively agreed on in a multi-partner workshop (Table 1). Five attributes were identified: investment cost (US$), labour effort (days), harvest price (US$), domestic harvest yield (20-litre bucket of dried mopane worms), and mopane worm harvesting scenarios. The choice-card design presents five trade-off scenarios on each card. The respondents were required to choose or vote for their preferred option across these attributes (see Table 1 and Figure 2).

Table 1 : Choice attributes and levels

Attribute	Level
Investment cost (US$)	1	2	3	4
Labour effort (days)	4	8	12	16
Domestic harvest yield (bucket)	0.5	1	1.5	2
Harvest price (US$)	2	4	6	8
Harvest scenario	Good domestic, good wild	Good domestic, poor wild	Poor domestic, good wild	Poor domestic, poor wild

Figure 2: Dummy card for choice experiment

The choice of investment cost was set at one-quarter, one-half, three-quarters and the full market price of a 20-litre bucket of dried mopane worms. This was done for two reasons. First, devaluation and multiple parallel currency exchanges make it difficult for many people to follow price changes over time. Longitudinal research on prices carried out by the project team indicates that the price of a 20-litre bucket of dried mopane worms has remained constant in real terms at US$4 at the point of consumption. This provides a useful and simplified comparable metric.

Attribute levels for the domestic mopane worm harvest are difficult to determine because the variability of harvest is influenced by the size of tree, number of trees and a natural variation in outbreaks. There is also insufficient empirical trial data to estimate domestic harvest levels. Two buckets per season was considered to be an upper limit to harvested yields, but the only accurate method to estimate this will be through field trials. Harvest price was depicted as a number of grain sacks with a full sack equivalent to the value of a 20-litre bucket of dried mopane worms. Harvest price levels were set at US$2, US$4, US$6 and US$8 per bucket.

Labour effort in farming was illustrated by a 16-pin die and an adjacent image of a woman. We have no empirical measure of the additional effort required to farm mopane worms but the 4, 8, 12 or 16 labour-days scenario reflects a probable range based on field experience with differing management approaches. Harvesting scenarios are depicted by a combination of three figures: a hut, a leafy tree, and/or a bare tree. The hut represents the homestead. The tree closest to the homestead represents the domestic harvest, and the distant tree the wild one. Good harvests are shown by trees with leaves, poor harvests by leafless ones. All four permutations were included in the design. Finally, a status quo choice was included in all the choice cards as a fifth option (equivalent to a ‘no vote’ clause). The status quo option is important, as respondents must always be given the opportunity to opt out or reject the scenarios presented.

The attribute levels result in a 4⁵ factorial design. Running a mains-effects orthogonal design function in SPPS (version 11.5) resulted in a 16-card main-effects design. We decided that respondents would answer up to eight choice cards each, to limit fatigue. The two sets of eight choice cards each were offered alternately to successive respondents. Because mopane woodland density was also considered likely to influence the adoption of domestic farming, the enumerators classified each household according to one of four mopane woodland types by using a standard pictorial guide: low, low/medium, medium, and high density (see Figure 3).

Figure 3: Mopane woodland zones

4. RESULTS

4.1 Descriptive analysis

4.1.1 Socio-demographic conditions

Two language groups are dominant in the study villages: Ndebele and Kalanga (a branch of the Shona language group). Languages tend to dominate by village. For example, Ndiweni and Mahetshe are mainly populated by Ndebele speakers (83 per cent and 80 per cent, respectively) while Sasane and Kapeni have a higher proportion of Kalanga speakers (80 per cent and 60 per cent, respectively). Household composition reveals little variation between villages, with average household size of around six people per dwelling, four of whom are adults (here, older than 16 years of age). Two in five households report a woman as head of household, with a higher percentage in Kapeni (62 per cent) than in Ndiweni (28 per cent). Approximately one in six household heads are illiterate (16 per cent), with almost all Kapeni household heads being literate (Table 2).

Table 2 : Socio-demographic profiles of the four study villages

Characteristic	Sample village				Full sample (n=120)
	Ndiweni (n=30)	Mahetshe (n=30)	Sasane (n=30)	Kapeni (n=30)
Household members^a	6.67 (2.74)	6.87 (3.18)	6.47 (2.52)	6.50 (2.14)	6.63 (1.31)
Household members aged 5–16 years^a	2.30 (1.59)	2.27 (1.53)	2.90 (1.42)	2.97 (1.53)	2.61 (0.77)
Head illiterate (%)	17	20	22	2	16
Female-headed household (%)	28	42	47	62	44
Main language (%)
Ndebele	83	80	17	40	55
Kalanga	13	10	80	60	41
Other/missing	3	10	3	0	4
Dwelling condition (%)
Good	43	33	50	63	48
Average	40	43	27	23	33
Poor	17	23	23	13	19
Sanitation access (%)
Blair toilet	83	60	63	57	66
Open field	10	40	37	37	31
Other	7	0	0	7	3
Drinking water access^b (%)
River	37 (33)	27 (33)	0 (3)	43 (57)	27 (32)
Spring	43 (43)	7 (3)	0 (0)	0 (0)	13 (12)
Community borehole	10 (13)	40 (37)	100 (93)	50 (37)	50 (45)
Private borehole	10 (10)	17 (7)	0 (0)	3 (0)	8 (4)
>500 metres	67 (67)	77 (70)	73 (70)	73 (70)	73 (68)
Electricity (%)	0	0	0	0	0
Radio (%)	47	23	23	63	39
Bicycle (%)	40	47	63	80	57
Cart (%)	33	33	47	57	43
Cattle^a	2.57 (4.87)	2.83 (4.70)	1.07 (1.94)	2.67 (3.98)	2.29 (2.03)
Goats^a	5.97 (4.65)	6.93 (9.15)	7.48 (5.34)	7.33 (6.12)	6.92 (3.29)
^aData presented as the mean (standard deviation).
^bData presented as percentage for dry season (percentage for wet season).

Classifying dwellings by their relative condition indicates that almost one in two respondents' homes rank as ‘good’ (48 per cent), in the sense of having a brick house with a tin roof. One-third of homes are ‘average’, with either a mud structure or a tin roof. The remaining 19 per cent are classified as ‘poor’, characterised by inferior structure and conditions, such as a mud hut and thatched roof. Village-level analysis shows that Sasane and Kapeni have more ‘good’ homes (50 per cent and 63 per cent, respectively) while Ndiweni and Mahetshe have more ‘average’ homes (40 per cent and 43 per cent, respectively).

Access to improved sanitation is represented by the presence of Blair toilets,3 which are used by 66 per cent of households, in contrast to 31 per cent of households that use ‘open field’ sanitation (see Table 2). Ndiweni has the highest level of access to Blair toilets (83 per cent), whereas two in five households in Mahetshe, Sasane and Kapeni report using ‘open field’ sanitation. Community boreholes, river water and springs are the main sources of drinking water in both the dry and wet season. In the dry season, 50 per cent of households use community boreholes as the main drinking water source, 27 per cent use river water, 13 per cent use spring water, and only 8 per cent use private boreholes. In the wet season, the use of river water increases to 32 per cent of households, while the percentage using community boreholes and springs falls to 45 per cent and 12 per cent, respectively. In Ndiweni, and to a lesser extent Kapeni, markedly fewer households than expected have access to borehole water, whereas the reverse is true in Sasane. Over one in seven households are further than 500 metres from their main drinking water supply in the dry season, with a small reduction in the wet season. None of the four villages has electricity. Material assets common across villages include radios (39 per cent), bicycles (57 per cent) and carts (43 per cent). Livestock are owned in all villages, with an average of two head of cattle and seven goats per household.

4.1.2 Woodland density and harvesting activities

The villages are divided between those with higher densities of mopane woodlands (Ndiweni and Mahetshe) and those with lower densities (Sasane and Kapeni) (see Table 3). Mopane worm harvesting is a common activity for most household members over the age of 5 years. Most people harvest in the area around the homestead (54 per cent), although this is associated with woodland density. For example, over one-half of the Ndiweni and Mahetshe households will harvest away from the homestead in areas such as state forests, crop fields or grazing areas. Accordingly, fewer of these households than those from Sasane and Kapeni rate the homestead as the best harvesting site. Ndiweni and Mahetshe households travel an average of 6 km to their main mopane worm collection areas, compared with less than 4 km for households from Sasane and Kapeni. Overall, households report spending over 6 hours harvesting mopane worms on an average collection day.

Table 3 : Harvesting activities and use of harvested mopane worms

	Ndiweni	Mahetshe	Sasane	Kapeni	Full sample
Mopane woodland zone (%)
Low/medium	100	73	30	33	60
Medium/high	0	26	70	67	40
Household members older than 5 years who usually harvest mopane worms (%)	65	67	68	73	68
Usually harvest at homestead (%)	37	47	70	63	54
Homestead is best harvest site (%)	3	17	37	37	23
Distance to main mopane worm collection area (km)^a	6.00 (5.40)	5.11 (4.90)	3.32 (4.01)	3.51 (5.48)	4.48 (5.05)
Collection time on an average harvesting day (minutes)	493 (136)	377 (135)	333 (145)	382 (156)	396 (154)
Harvest of dried mopane worms (20-litre bucket)
Good year^b	8.98 (6.51)	7.51 (4.86)	7.78 (5.06)	7.98 (4.03)	8.07 (5.03)
Bad year^c	2.03 (1.80)	2.34 (1.80)	2.20 (1.47)	2.04 (1.10)	2.15 (1.55)
Mopane worm use (%)
Good year^d
Sell immediately	45	31	30	52	40
Food source	17	23	25	14	19
Store for off-season sale	29	32	36	25	31
Other	9	14	9	9	10
Bad year^e
Sell immediately	44	24	28	39	33
Food source	33	55	53	46	47
Store for off-season sale	23	21	18	14	19
Other	0	0	0	0	0
Note: Data presented as the percentage or the mean (standard deviation).
^aOutliers >30 km are excluded (n=117).
^bOutliers >25 buckets are excluded (n=119).
^cOutliers >6 buckets are excluded (n=116).
^dOutliers >25 buckets are excluded (n=119).
^eOutliers >6 buckets are excluded (n=116).

Households estimate that in a good year a wild harvest yields eight buckets of mopane worms, with little variation between median and modal values across villages. Conversely, in a bad year households only harvest just over two buckets. Harvest allocation across good and bad years is classified under four general headings: sell the harvest immediately, use as a household food source, store for off-season sale, and other. In a good year, households sell, on average, 40 per cent of the harvest immediately. Households in Kapeni and Ndiweni sell slightly more (52 per cent and 45 per cent, respectively) while households in Mahetshe and Sasane sell slightly less (31 per cent and 30 per cent, respectively). These latter villages use around one-quarter of the harvest as a food source, while Ndiweni and Kapeni report using just 17 per cent and 14 per cent, respectively, for household food. Households across all four villages report storing at least one-quarter of the harvest for later sale. Other uses (gifts, barter) represent generally less than 10 per cent of the harvest.

In a bad year, household decisions about allocating the harvest change. In general, a higher proportion of the harvest is used as a food source, which results in proportional reductions in the quantity sold, stored or used in other ways. In Mahetshe and Sasane more than one-half of the smaller harvest is used for household food in a bad year, compared with one-quarter in a good year. Only households in Ndiweni sell a similar proportion of the wild harvest in both good and bad years. This is reflected in the smaller increase in the amount of the harvest used for food within households (from 17 per cent to 33 per cent) in bad years. Across the sample, the proportion of the wild harvest allocated to household use increases by 28 per cent (from 19 per cent to 47 per cent) in a bad year. This results in a proportional fall of 7 per cent in immediate mopane worm sales, a 12 per cent fall in storage for off-season sale and a 10 per cent fall in allocation to other uses compared with a good year. Across the sample, the median difference in harvest between a good year and a bad year results in little aggregate change in the allocation to household food consumption.4 The impacts of this are mainly lost income from immediate sales and reduced future income from higher prices in the off-season because less is stored for later sale.

4.2 Choice experiment analysis

4.2.1 Multinomial logit models

Each of the 120 respondents completed eight choice tasks. The analysis of choice responses is based on random utility theory, which models the probability that individual n will choose option i over any other option j belonging to the complete choice set C as:

where V is an observable systematic component of alternative I, and e is a random component.

If it is assumed that the stochastic elements of the utilities follow a Gumbel distribution, the multinomial logit model can be specified as:

Following estimation of these parameters, the marginal rate of substitution (MRS) between any pair of attributes a and b can be calculated from the following equation:

The economic valuation of various scenarios can be estimated from the following equation:

where the implicit value represents an economic surplus derived from a monetary coefficient (β_m) and the change in utility associated with shifting from one state to another, V₀ – V₁ (Colombo et al., 2006). Of particular interest in this study is the implicit value of a good domestic harvest relative to a poor domestic harvest when the wild harvest is held constant.

Estimates from the multinomial logit model are presented in Table 4. Model I provides an aggregate specification. All attributes are significant at the 5 per cent level or lower, and their signs are consistent with expectations. For example, a 1 per cent increase in investment cost is associated with a negative value (–0.59), whereas an increase in harvest price has a positive value (0.38). The coefficient for committing additional labour to domestic farming is negative (–0.09), with the coefficient for domestic harvest yield being positive (0.13). Mopane worm harvest scenario attributes are associated with positive utility and the values are consistent with farmers preferring ‘good’ to ‘poor’ harvests. Model I has a significant model fit,5 supported by a high pseudo R ² value of 0.45.

Table 4 : Choice models

	Model I^a	Model II		Model III		Model IV^b
		Younger than 40 years	Older than 40 years	Low density	High density	Class I	Class II
Investment in mopane worm farming (US$)	−0.59*	−0.61*	−0.58*	−0.57*	−0.96*	−1.78**	−0.61*
Labour effort in mopane worm farming (days)	−0.09*	−0.11*	−0.08**	−0.09*	−0.11*	0.24	−0.00
Domestic mopane worm harvest (20-litre buckets of dried mopane worms)	0.13**	0.08	0.17**	0.13**	0.15	−1.94	0.44*
MW price (US$ per 20-litre bucket)	0.38**	0.55***	0.26	0.33***	0.88	−2.34	0.53**
Mopane worm harvest scenarios
Good domestic, good wild	4.67*	4.46*	4.98*	4.46*	34.41	51.81	1.24**
Good domestic, poor wild	3.93*	3.72*	4.21*	3.75*	33.71	53.39	0.16
Poor domestic, good wild	2.23*	1.61**	2.76**	2.17*	30.55	34.36	0.06
Poor domestic, poor wild	1.76*	1.18***	2.29*	1.61*	30.70	38.67	−0.03
Constant I	0.17	0.17	0.15	0.19	0.16	9.41***	−0.67
Constant II	0.19	0.38	0.04	0.27	−0.70	8.36***	−1.09*
Constant III	0.69*	0.77**	0.61***	0.67**	0.82	6.33***	0.09
Model specification
Observations	1200	530	670	1000	200	1200
Log-likelihood at constant	−239.83	−102.45	−134.08	−209.76	−25.71	−115.92
Likelihood ratio test	373.86*	200.62*	179.25*	299.00*	83.41*	642.53*
Pseudo R ^2	0.45	0.48	0.44	0.43	0.62	0.73
^aModel I fails the Independence from Irrelevant Alternatives test (χ^2 = 96.59; degrees of freedom = 8; p < 0.05; option 1 excluded), suggesting use of less restricted models. ^bProbability of class membership is 69 per cent (Class I) and 31 per cent (Class II), significant at the 1 per cent level. *Significant at the 1 per cent level, **significant at the 5 per cent level, ***significant at the 10 per cent level.

The model violates the Independence from Irrelevant Alternatives test (Hausman & McFadden, 1984), which suggests exploring other model specifications. First, we decompose the aggregate results by simple age and woodland density thresholds, as reported under Models II and III (Table 4). Second, we model a latent class specification in Model IV to estimate heterogeneity in preferences across the sample (Boxall & Adamowicz, 2002).

Model II differentiates the responses of respondents younger than 40 years of age and those 40 years of age and older.6 The results indicate similar values for investment cost and labour effort. The differences are that older respondents have a higher positive value for domestic harvest yield (0.17) compared with Model I (0.13), whereas younger respondents are influenced more by domestic harvest price (0.55). Older respondents also have higher positive values across all four harvest scenarios compared with younger respondents and the aggregate (Model I) results.

In Model III the sample is broken down by location into high-density and low-density mopane woodland zones.7 The results show that households in high-density zones appear more constrained by investment costs and labour effort than households in low-density zones, whose coefficients are similar to those of Model I. All other attributes are either similar to Model I results (for low-density households) or not significant (for high-density households). The higher pseudo R ² value reported for the high-density model is notable (0.62, compared with 0.45 for Model I), although some caution is needed in interpreting this because of the relatively low number of observations.

4.2.2 Latent class specification

Model IV provides a latent class specification. This introduces the complexity of preference heterogeneity by assuming the variance in the error term (e _i ) is not constant but depends on individual observations (heteroscedasticity). A latent class model assumes a discrete number of support points (say, S) is sufficient to describe the joint density function. Latent classes correspond to underlying market segments or groups, each of which is characterised by unique tastes β _s , where s = 1, …, S. Membership of class s can be characterised by variance differences across preferences, β _s , and scale, λ _s (Swait, 1994). If the indirect utility function for members of class s is:

and the ϵ _iq/s values are conditionally independently and identically distributed extreme values of type I within class, the choice probability for members of class s is estimated by:

Swait (1994) provides details of a classification mechanism to predict an individual's membership in a class. The unconditional probability of choosing alternative i is given by:

where W _qs is the probability of being in class s.

An advantage of using a latent class specification, in this context, is that an understanding of preference heterogeneity can be estimated without arbitrary decomposition by age or location, thus potentially revealing preferences of classes (or groups) hidden within aggregate models. Where such classes can be identified, an associated probability of class membership can also be estimated. The results indicate a large improvement in model fit from a pseudo R ² value of 0.45 in Model I to 0.73 in Model IV. Two classes that are significant at the 1 per cent level are revealed. The estimated probability of class membership is 69 per cent for Class 1 and 31 per cent for Class 2. Of particular note is the high and significant negative value associated with the investment cost coefficient for Class I (–1.78), which is three times greater than the coefficient in Model I and almost twice as much as that estimated for respondents from high-density woodland zones. All other attributes are not significant for Class I other than the constants, which are significant and positive. This increases the analyst's confidence in respondents' understanding of and commitment to the choice tasks, as it is common in the literature for constants to be negative (Adamowicz et al., 1998). At least three reasons may be suggested when constants are negative: the choice task is too complex, there is uncertainty about the trade-off the respondent would be willing to make, and respondents do not trust effective implementation of the changes. Here, the results support the intuitive conclusion that the principal constraint to adoption is investment cost. Class II attributes show significant coefficients for investment cost, which is of similar scale to the Model I estimate, and harvest yield and harvest price, which are both positive and higher than the Model I estimates. Finally, the harvest scenario for ‘good domestic, good wild’ is positive and significant but is lower (1.24) than the Model I estimate (4.67).

4.2.3 Implicit values

Implicit values are estimated from Equation (4) and are presented in Table 5. The hypothetical scenario of particular policy interest is where a domestic harvest moves from ‘poor’ to ‘good’, with wild harvest held constant at ‘poor’ (i.e. good domestic, poor wild to poor domestic, poor wild). This allows implicit valuation of adopting domestic mopane worm farming. This can be specified in terms of US$ (using investment cost as a proxy for marginal utility of income) or estimated in labour days by substituting investment cost or days in the coefficient (β _m ) in Equation (4). The results show a range of values from US$3.31 and 24 days (Model II, older than 40 years) to US$4.16 and 23 days (Model II, younger than 40 years). This implies a rural wage rate of US$0.14–0.18 per day. Estimates for willingness-to-pay (WTP) for one domestically farmed bucket may be inferred from Equation (3), with attribute estimates for harvest yield divided by investment cost (or labour effort). Similarly, an estimate of willingness-to-accept (WTA) for one domestically farmed bucket may be calculated by dividing harvest price by investment cost (or labour effort). Estimates from Model I indicate a WTP of US$0.22 or 1.44 days and a WTA of US$0.64 or 4.22 days. This suggests respondents expect a threefold return on investment. It also indicates a value share of 16 per cent of the estimated urban point-of-purchase price of one bucket of dried mopane worms (US$4). These estimates compare with a higher WTP by older respondents (US$0.29) and households in areas of low-density woodland (US$0.23) and a lower WTA by the same households (US$0.58) but a higher WTA by younger respondents (US$0.90). Comparable estimates are provided by Class II coefficients, which reflect the highest WTP (US$0.72) with a relatively small gain represented by a WTA of US$0.87.

Table 5 : Implicit values in US$ (days)

	Model I	Model II		Model III		Model IV
		Younger than 40 years	Older than 40 years	Low density	High density	Class I	Class II
Good domestic harvest with poor wild harvest (days)	US$3.68 (24.11)	US$4.16 (23.09)	US$3.31 (24.00)	US$3.75 (23.78)	n/a	n/a	n/a
Harvested bucket (days)	US$0.22 (1.44)	n/a	US$0.29 (2.13)	US$0.23 (1.44)	n/a	n/a	US$0.72 (n/a)
Sold bucket (days)	US$0.64 (4.22)	US$0.90 (5.00)	n/a	US$0.58 (3.67)	n/a	n/a	US$0.87 (n/a)
n/a = not available.

5. CONCLUSION

Exploring experimental scenarios of social adoption of domestic mopane worm farm management provides information to decision-makers in the process of designing feasible and desired rural development interventions. This is useful in conditions of scarcity or risk where adoption may be constrained by factors unanticipated by promoters of the technology. A choice experiment investigates identified attributes of uncertainty in an effort to improve the connection between technology design and user adoption. The method tests adoption scenarios using pictorial choice cards, which increases the participation of less literate respondents. This permits an inclusive although provisional quantification of livelihood gains from an innovative rural diversification intervention, and an understanding of how adoption preferences may vary by age, location and other latent factors.

While estimated gains of domestic farming appear modest (US$3.31–4.16), this has to be set in the context of its role as a coping mechanism when wild harvests fail, and of the significant developmental implications associated with the disintegration of the economy in Zimbabwe. For poorer and more vulnerable people in income-constrained rural settings, income-generating activities can assume a disproportionately higher value due to poorer people simply having less money than wealthier people, along with the impacts of ‘cash hunger’ (Scott, 1978), which may prejudice livelihood opportunities moderated by market-based exchange (e.g. farm inputs, basic household needs, school fees). A condition of cash hunger is illustrated here by a low, estimated rural wage rate of US$0.14–0.18 per day. This dampens the level of investment in domestic farming per bucket to a rate of US$0.22 (or 1.44 days), which is roughly one-third of the price for which respondents are estimated to be willing to sell one bucket (US$0.64 or 2.44 days). This implies a harvester–seller ratio of 16 per cent (e.g. US$0.64/4.00), which is lower than wider analysis of rural Zimbabwean harvesters receiving ratios of 20–25 per cent at the final point-of-sale price (Frost, 2005). Data disaggregation indicates that older people and households located in lower woodland density zones are prepared to invest more per bucket. This is consistent with wider evidence across sub-Saharan Africa of more vulnerable rural people being willing to invest in low-value labour activities, given a lack of alternatives.

But this tells only part of the story, as the latent class analysis reveals cash investment as the overriding constraint for a majority class of respondents with a coefficient three times as large as the aggregate model (–1.78 compared with –0.59, respectively). Further weight is added to this finding as all constants are significant, which increases confidence in the level of respondents' understanding and commitment to these choice tasks. In contrast to this large ‘cash hungry’ class, a smaller class of respondents is identified, who appear willing to invest more per bucket (US$0.72) than any other model specification, although they require a higher rate of return per bucket (US$0.87). These findings underline the heterogeneous and complex nature of social decision-making and the associated difficulty of identifying uniform policy responses under conditions of risk, scarcity and uncertainty.

Given the exploratory nature of this research, a number of issues are identified for further consideration. First, the social impacts of domestic mopane worm farming are thought likely to be most significant when outbreaks in the wild fail. This remains a working hypothesis until empirical evidence can be gathered to accept or reject this position. Second, it can be expected that social adoption of farming innovations will respond to market conditions. Existing research calls for greater knowledge of market-chain analysis, storage and value-addition (Stack et al., 2003), with which this study would concur. Third, there is no empirical evidence of the relationship between domestic yield and various domestic farm management regimes. Fourth, domestic price sensitivity to wild outbreak failure would provide firmer insights into the value of domestic production. Fifth, monitoring and evaluation of early adopters of this technology could provide important insights into social impacts and adaptive management practices for wider promotion. While it is unlikely that adoption of this innovative technology alone will lift rural people out of poverty, it may provide a new mechanism that offers a timely and flexible opportunity to buffer poor and vulnerable people for whom few other livelihood options exist.

The authors thank Jayne Stack, Eva Gardiner, Tendayi Gondo and Member Mushongahande, who contributed in the design phase; the field team of Tendayi Gondo, Witness Kozanayi and Benias Madzividza; the Plant Production Systems Group, University of Wageningen; and two anonymous referees. This publication is an output from a research project funded by the UK Department for International Development (DFID) for the benefit of developing countries. The views expressed are not necessarily those of DFID (R7822 – Forestry Research Programme).

Notes

¹At an exchange rate of US$1=approximately 3.6 ZAR in June 1994, this would be about R302.4 million (www.oanda.com/convert/fxhistory).

²This research forms part of a larger study, the ‘Mopane woodlands and mopane worms’ project (R7822, DFID Forestry Research Programme).

³Characterised by a chimney ventilation shaft.

⁴Median harvest values for a good year (six buckets) and a bad year (two buckets) are multiplied by proportional allocation to food in a good year (0.19) and in a bad year (0.47), resulting in a net loss of mopane worm food allocation of 0.20 buckets in a bad year.

⁵Likelihood ratio test = 373.86, p < 0.001

⁶From qualitative inquiry, age of respondent appeared to be a determinant influencing people's willingness to adopt domestic mopane worm farming.

⁷High incorporates high and medium categories; low incorporates low and medium/low categories.