The Adoption and Productivity of Modern Agricultural Technologies in the Ethiopian Highlands: A cross-sectional analysis of maize production in the West Gojam zone

March 20081

Jon Pycroft DPhil Economics Candidate University of Sussex

ABSTRACT This paper investigates the factors that determine adoption and productivity of improved seeds, one of the most important modern agricultural technologies for small farmers in Ethiopia. Accounting for almost half of GNI and some 85% of employment, agriculture is the foremost economic sector in Ethiopia. Agricultural productivity, however, has failed to keep pace with the burgeoning population. Due to the limited amount of suitable land for agricultural expansion, the need for modern agricultural technologies to boost production levels on existing farmland is paramount. Firstly, using data derived from an agricultural census of the West Gojam zone, the decision to adopt improved seeds is analysed using a censored Tobit and a probit model. The results indicate that many of the common determinants of adoption, such as having larger farms and a literate household member, apply in West Gojam. Secondly, the same data set is used to investigate the benefit of adopting improved seeds, in terms of gains in productivity, using two methods: an endogenous switching regression model and a propensity score matching model. Both models predict that were traditional seed users to adopt improved seeds, on average, they would benefit but they would not be as productive as those currently using improved seeds.

1

Prepared for presentation at the Centre for the Study of African Economies (CSAE) Conference, Economic Development in Africa, 15th-17th March 2008. An earlier version of this paper was presented at the 2nd International Symposium on Economic Theory, Policy and Applications, 6-7th August 2007, Athens, Greece, mimeo.

1

1. Introduction In 1964, Schultz observed, “The man who farms as his forefathers did, cannot produce much food, no matter how rich the land or how hard he works.” (p.3) Sadly, this quote applies to the majority of Ethiopian agriculture today. Average grain yield per hectare remains unchanged from 25 years ago at around 1 tonne per hectare. (Berhanu, 2003) Development of the sector has been modest in recent decades, despite repeated attempts, and large portions of the country remain food insecure. Schultz continues, “The farmer who has access to and knows how use what science knows about soils, plants, animals and machines can produce an abundance of food though the land be poor. Nor need he work nearly so hard and long.” This paper addresses how scientific knowledge is used in Ethiopian agriculture and how it could contribute to the development of the sector. The focus of the paper is on improved (or hybrid) seed varieties. Using such seeds represents a distinct agricultural technology that has great potential to raise agricultural output as has been demonstrated most famously in the so-called Green Revolution of South Asia. Adoption of improved seeds has been slow and generally less successful in Ethiopia, and indeed across Africa.

Rural West Gojam This study focuses on the West Gojam zone, which is located to the south of Lake Tana in the Amhara region of north western Ethiopia. (See Figures 1 and 2.) The zone is in the highlands (all districts within the zone are situated above 1500m) and has mostly reliable rainfall (9 out of the 10 districts are classified as such). Note that the majority of the Ethiopian population live in highland areas. The study is limited to rural agriculture, as urban agriculture production is relatively small and can follow markedly different patterns. Maize is the major crop in the area, where nearly all crop production is by smallholders. Some of the administration of government agricultural services is carried out at the zonal level, with further administration is carried out at the district (wereda) level.

2. Data The data set is derived from the Ethiopian Agricultural Sample Enumeration 2001/02, commonly referred to as the Agricultural Census. Through transformation of the data, observations are obtained for 6,395 households that are involved in maize production.

Productivity and Adoption of Improved Seeds As shown in Figure 3, maize productivity is clearly bi-modal and this is due to the distinction between improved seeds users and non-users (Figure 4). This clear illustration shows that, though there is considerable overlap, the improved seed users tend to be more productive by around one half of a kilogram of production per square

2

metre cultivated; the difference in means is 0.544 kg/m2. 2 This result offers a prima facie case for detailed investigation into improved seed use in rural West Gojam. Figure 5 illustrates the variation of improved seed adoption by district, which ranges from 10% is Yilmena Densa to 55% in Dembecha. A summary of all variables used is given in Tables 1a, 1b and 1c.

3. Determinants of Improved Seed Adoption The data set gives information about how much of a household’s maize production is planted with improved seeds. The dependent variable, the proportion of maize production that uses improved seeds, is censored (i.e. no household can plant less than zero percent improved seeds or more than 100 percent), so a simple OLS regression will give biased and inconsistent results. Two estimation methods to deal with this are carried out and the results of each are compared. Firstly, a double-censored Tobit model seeks to discover the parameters that effect how much of a household’s maize area uses improved seeds. Predictions outside possible values (i.e. less than zero percent or more than 100 percent) can be interpreted as how far the household is from starting to use improved seeds. So a predicted value that is negative but close to zero (in this model, say, -0.1) would suggest that a small change in the household’s circumstances would be expected to make the household start using improved seeds. A predicted value that is strongly negative (say, -3) would suggest that even a large change in household circumstances would be insufficient to stimulate improved seed use. Secondly, a probit model is estimated as to whether the household uses improved seeds or not. The marginal effects are expected to be smaller than those in the Tobit model: this is due not only to the different functional forms of the models, but also to the fact that a small amount of improved seed usage is sufficient to qualify a farmer as an improved seed user (the histogram in Figure 6 helps to understand the distribution of proportion of improved seed use among users).

Results For both models, only the preferred specification has been reported (see Table 2). The Tobit model estimates can be interpreted as marginal effects. The dprobit estimates can be read as the effect of an infinitesimal change in the variable, and the interpretations below all assume mean values for all other variables. Interpretations for both models assume ceteris paribus. Reassuringly, the two models support each other. In both cases, the same variables are significant (with the same sign) and the magnitudes are comparable. The marginal effects in the probit model are consistently smaller, reflecting the fact that the probit is estimate a propensity over a limited range (between 0 and 1), as opposed to the slope coefficients in the Tobit, which predicts well outside the possible range. The Tobit marginal effects range from 1.5 times greater (for the coefficient on ln(tropical livestock units)) to 3.2 times greater (for the dummy on Achefer District). Addressing some of the important results of the models, by variable, one can observe that a male household head is 8.6 percentage points more likely to use improved seeds (from the probit model) and expected to plant an additional 22.9 percentage points of their maize production with improved seeds (from the Tobit 2

The values of the means are 2.729 kg/m2 (standard error: 0.015 kg/m2) for users and 2.185 kg/m2 (standard error: 0.013 kg/m2) for nonusers.

3

model). This may be indicative of males having better contacts and access to markets, especially the credit market. One anticipates that education would improve access to information and make households more likely to embrace new technology. Having a literate household member makes the household 11.5 percentage points more likely to use improved seeds at all (probit model), likely to plant an extra 27.3 percentage points of maize production with improved seeds (Tobit model). Households with more land overall, are more likely to use improved seeds. Comparing a household with one hectare to another with two hectares, ceteris paribus, the larger farm makes the household 18.9 percentage points more likely to use improved seeds (from probit) and is expected to plant 29.2 percentage points more land with improved seeds (from Tobit, author’s calculations). This can be explained as land proxies for wealth, and possibly collateral for loans. As land has often been allocated by the local government, the variable could even be associated with the favour the household has with local officials, who often provide the improved seeds and cofactors. Those who specialise more in maize production i.e. have a higher proportion of their total land devoted to maize, are more likely to use improved seeds. This is intuitively acceptable result, as the more important maize is to a household, the more efforts one makes to raise its productivity. Comparing a farm with 20% of their crop in maize to one with 40%, the latter is 14.3 percentage points more likely to use improved seeds (probit model) and likely to plant 27.2 percentage points more of their maize production with improved seeds (Tobit model, author’s calculations). Other factors that make a household more likely to adopt are having a young household head – suggesting the young are more likely to embrace new ideas and have less aversion to risk – having a low proportion of the crop damaged – suggesting better quality land (fewer pests, etc) – rented land is more likely to use improved seeds – which could result from the landlord having better access to resources, or equally, the renters in of such land can be better-off households – and there is some impact from having more livestock – perhaps used for collateral on loans or more generally as a proxy for wealth. The same six district dummies are significant in the Tobit and probit models, which must all be interpreted as difference between each district and the omitted district, Merawi. One can speculate the major differences observed between districts reflect differences in geography and in administration, as much of the distribution of seeds and cofactors in carried out by local government. Access to markets, including the road infrastructure, would also differ both within and between districts.

4. Productivity Effect of Improved Seeds The Endogeneity of the Decision to Use Improved Seeds and the Endogenous Switching Regression Model When attempting to evaluate the benefit from using improved seeds, a possible model would resemble the following3: Yi = Xi B + α(improved seeds use) i + ε i

3

The following explanation of the endogenous switching regression model draws on Maddala (1983: 260-7).

4

Where Yi is maize productivity, Xi B are other regressors and coefficients and ε i is the error term. Were it reasonable to treat the decision to use improved seeds as exogenous, and the B estimates as the same across improved seed users and non-users, the estimate of the coefficient on improved seeds, α, would reveal the productivity gain from using of improved seeds. However, there are good reasons to why this ought not to be assumed. Figure 7 shows one way in which selection bias may effect the decision to use improved seeds. The solid black circles represent the observed productivity level for improved seeds users and non-users. It is tempting to then presume that this difference represents the benefit of using improved seeds. However, this may be subject to ability bias, which is why it has been labelled the ‘“naïve” measured benefit of improved seeds’. Suppose it were the case that those who use improved seeds tend to be high ability farmers, and so even without using improved seeds they would still be highly productive (represented by point A); suppose also that those who do not use improved seeds tend to be low ability farmers, so even with improved seeds they would only be able to produce a little more (represented by point B). Were this the case, then the shallower slopes would represent the actual benefit of improved seeds. In the absence of variables that adequately proxy for a farmer’s ability, one must suspect the possibility of ability bias,4 and this would result in endogeneity of the improved seed use variable. More generally, the problem is that farmers in each of the two regimes are not the same with respect to variables that are relegated to the error term. To place this argument is an algebraic formulation: YIi = XIi BI + e1i YNi = XNi BN + e2i

(productivity of improved seed user) (productivity of non-improved seed user)

Ii = Z iγ + ε i Ii = 1 if Z iγ + εi > 0 Ii = 0 if Z iγ + εi ≤ 0

(participation decision) (improved seeds used) (non-improved seeds used)

Where Zi represents a vector of selection variables; ei is normally distributed Of course, YI and YN cannot both be observed for any single maize plot. This is unfortunate, as the difference between the two is exactly what we wish to discover, that is, we are interested in (a) the difference between the productivity of the user compared with their expected productivity without the improved seeds, and (b) the difference between what non-users produced and their expected productivity had they used improved seeds. Algebraically, these are: YIi – E(YNi | Ii = 1) YNi – E(YIi | Ii = 0)

the benefit to each improved seed user the benefit forgone by each non-improved seed user

Furthermore, as argued, the error terms u1i and u2i may be subject to selection bias. The values can be estimated using an endogenous switching regression model, which estimates the following three equations: Ii YIi YNi

= Z iγ + ε i = XIi βI + σ1ε * ø(Z iγ)/Φ(Z iγ) = XNi βN + σ2ε * ø(Z iγ)/(1 – Φ(Z iγ))

+ u2i + u1i

4

Note that as well as ability bias, similar results could be found if the improved seed users tended to have superior plots of land or had other unobserved factors that led them to be more productive.

5

Where: σIε = cov(u1, ε)5 σNε = cov(u2, ε) ø(.) = density function of the standard normal Φ(.) = cumulative distribution function of the standard normal If σ1ε and/or σ2ε are non-zero, selectivity bias exists in the estimation procedure and OLS estimates would be inconsistent. The selection equation could be estimated by the first stage of a two-stage least squares method or as part of a maximum likelihood estimation, however this paper uses a full-information maximum likelihood (FIML) method to simultaneously fit binary and continuous parts of the models, so as to yield consistent standard errors (Lokshin & Sajaia, 2004:282-3). It should be emphasised that the endogenous switching model allows not only an intercept change between the two agricultural technologies, but also changes in the slope coefficients, which is relevant in this case.6 Note also that the identification criteria require at least some variables are in the binary (selection) model, but not in the continuous model (i.e. some variables in Z are not included in X).

Switching Model Results I: Determinants of Productivity in Each Regime The switching model is designed to give consistent coefficients on all variables, which have been adjusted to remove selection bias. Only the preferred specification of the model has been reported (Table 3). All variables are significant across the three equations. Specification values are reported beneath Table 3.7 A number of the coefficients are significantly different in each regime, which is why it is necessary to split the sample. The likelihood ratio test of independent equations estimates whether the selection bias adjustment is significant (which it is, pvalue 2.35%). Note that the selection bias term is only significant for the equation 2 (as ρu1,ε is not significant), which suggests that merely splitting the sample would have been sufficient to estimate equation 1. The coefficients of a switching model can be interpreted as marginal effects, as in standard least squares regressions. The dependent variable is maize productivity, measured in kilograms of production per square metre. For non-users, there is support for the well-known result that smaller farms are more productive (e.g. a two hectare farm is predicted to be 0.040 kg/m2 less productive than 5

The covariance, σ1ε, is mathematically equivalent to the correlation coefficient multiplied by both the standard deviation of u1 and ε (i.e. σ1ε = ρ1ε * σ1 * σε). In the Stata results (using the “movestay” command), ρ ε 1 and σ1 are reported, which multiplied together give the estimate for σ1ε, as σε is normalised to 1. (see Lokshin & Sajaia, 2004: 283-4) 6 In the absence of selection bias, this could be achieved with an OLS model that includes extra variables, which repeat the values for all the variables for the household is an improved seed user and are zero otherwise. When this is done, many of these extra variables are individually significant (and together are clearly jointly significant), which suggests that many coefficient values in each regime are not the same. 7 The specification values can be used to recover the original σu1 / σu2 and ρu1,ε / ρu2,ε. Multiplying these together gives the covariance. Recall that the standard deviation of the error of the selection model is normalised to one – see footnote above. Multiplying the covariance by the inverse Mills ratio, or the compliment of the Mills ratio (both of which vary by observation) gives the magnitude of the selection bias term, the summary statistics of which are shown (algebra given above).

6

a one hectare farm). This is believed to be due to the additional care taken of a smaller plot. However, this result becomes insignificant for those using improved seeds. There is a large impact from the degree of specialisation in maize production (pcmaizearea). An extra 20 percentage points of farm area devoted to maize is associated with a rise in productivity of 0.140 kg/m2 for those using improved seeds, and a smaller (but still significant) rise of 0.047 kg/m2 for non-users. This is an intuitive result as when maize is a more important crop, more effort is made to ensure its productivity. More of the crop being subject to damage (such as from pests) lowers productivity, as one would expect. Other results show that having more farm implements benefit users, but is insignificant for non-users. Lastly, having an older household head helps non-users, presumably benefiting from the extra years of experience; however, for improved seed users the age of the household head is insignificant, probably because few households have many years of experience of the new technology. Also note that the district dummies record substantial variation in productivity levels that have not been incorporated elsewhere in the model, which may be due to differences in geography, administration and access to markets for both inputs and produce.

Switching Model Results II: The Benefit of Improved Seeds The impact of improved seeds on productivity levels is recoverable by inserting the variable values for each household into the corresponding equation to evaluate the predicted productivity. Interestingly, it is now also possible to use the parameters for the improved seed users equation to predict the values for non-users were they to adopt improved seeds, and vice versa. This results in four sets of predicted values for productivity that are summarised in Table 4. The hypothetical predictions assume that the coefficients obtained in the switching regression for improved seeds users would apply to non-users, were they to adopt them, and also the coefficients obtained for non-improved seeds users would apply to users, were they to revert. The variations in predicted values can be seen in the kernel density plots (Figures 8 and 9). Considering those who currently do not use improved seeds (Figure 8 and Table 4), it is clear that were they to adopt their predicted productivity would improve, on average. The mean predicted value for current non-users, continuing to not use improved seeds is 2.185 kg/m2, compared with 2.478 kg/m2 were they to adopt. Note that this higher value is still lower than the mean value for those who currently use improved seeds, which is 2.729 kg/m2. In turn, this value for current users is considerably above the predicted 2.195 kg/m2 were they to revert to non-improved seeds; the density plots are compared in Figure 9. The predicted benefits of adoption for non-users vary considerably with most households expected to gain, whilst around 1 in 10 expected to lose, ranging from a reduction in productivity of 19 percent to an increase of 113 percent. The median predicted increase is 17 percent, and the mean 19 percent. This demonstrates considerable diversity in the predicted response to improved seeds, which has policy implications for targeting of government programmes and private retail. The key conclusion is that the benefits to non-users of adopting improved seeds are less than the benefits to those already using. One suspects that this is part of the explanation as to why they have not yet adopted.8 8

Were time series data available, it would be interesting to explore whether the productivity of improved seed users continues to rise after adoption, reflecting some learning curve for the

7

Propensity Score Matching Model: An Alternative Approach to Isolating the Benefit of Improved Seeds An alternative approach to dealing with endogeneity was proposed by Rosenbaum & Rubin (1983), which built on earlier work by Cochran and Chambers (1965) and Cochran (1968). The goal is to replicate the conditions of a random experiment, such as is done in medical trials, with a treatment group and a control group that can be otherwise considered identical. These conditions could be largely achieved were it possible to match those who were treated with an observation from the untreated group that is identical in all important characteristics. For example, in a medical trial, this could involve matching observations on the basis of gender, age, income level, education level, exercise regime and so on. Clearly for any large number of variables, matching on all of them becomes difficult with the rapidly expanding number of subclasses. Rosenbaum and Rubin’s innovation was to suggest that an observation’s likelihood of being treated or not could be summarised using a propensity score, which can be calculated using a probit or logit model. 9 The matching of treated and untreated observations can then take place on the basis of matching the propensity score alone. If observations from the treated and control groups that have the same propensity score are, on average, identical, (the property of unconfoundedness) then the matching technique is valid. (See Imbens, 2000, for further discussion of the concept.) In practice, various tests are conducted on the respective propensity scores to determine whether these conditions are likely to be met (outlined in Becker and Ichino, 2002) As the propensity scores are continuous variables, the scores from the treated and untreated groups will not match exactly, so an algorithm for matching observations is required. There are at least four valid methods for doing so: nearest neighbour matching, radius matching, kernel matching and stratification matching (ibid). All four methods are reported below (see Tables 5 and 6). Regarding radius matching, where the radius must be specified, of the many values experimented with, two are reported: 0.01 and 0.001. Once the matching has taken place, the outcomes (in this case, the productivity of maize production) can be compared. The average difference in outcome between all the matched pairs is termed the average treatment effect on the treated.

Matching Model Results The propensity scores obtained for all observations are shown in Figure 10, which is a density plot of these scores, split by users and non-users. The observations were then matched on the basis of these scores (by each of the four methods). This was carried out first for improved seed users (the treated group), matching these observations to non-users (the untreated group), the results of which are shown in Table 5. Taking a four-way average (i.e. counting the average of the two radius matching results as one), improved seed users appear to produce 0.373 kg more maize per square metre, ceteris paribus, than non-users. Note that this is significantly less than the simple comparison of the two groups that suggests that users are 0.544 kg/m2 more productive. This indicates that while improved seeds increase productivity, new technology. Such a learning curve may well be lurking behind these cross-section results, though it is not possible to draw it out with the available data. 9 This study uses a probit model.

8

there is endogeneity in the simple comparison that is being corrected for using this method. Asking the question this way, addressed what the impact of being an improved seed user is for those who use improved seeds. The question can be reversed to investigate the effect on non-users of being non-users (the treated group is now the non-users). The results for this are shown in Table 6. Regardless of the matching method used, the average treatment effect is reduced (though not always significantly so). The fourway average is now 0.330 kg/m2, which is around one standard error lower than the previous four-way average. Considering that in many cases the calculation being made will be identical, the somewhat lower estimate from all methods is notable. However as the difference is not significant, at best, it lends weak support to the notion that current non-users would not gain as much from adopted improved seeds as current users. More importantly, the value is significantly less than the simple comparison of the two groups, which supports the result from the switching model that adoption by non-users would bring positive benefits, but they would not be expected to achieve the same productivity as current users.

5. Conclusions and Suggestions for Future Research This paper has taken a large data set for a specific area in the Ethiopian highlands where improved seeds are popular relative to other parts of the country. The results suggest that households headed by a young male, with a literate household member, are more likely to adopt improved seeds, as are larger farms with livestock on rented land, with little crop damage, that specialise in maize. These determinants of adoption are consistent with the literature and intuition, and suggestive of other factors not directly accounted for in the data, especially the availability of credit. The results in this paper suggest that the benefit of adopting improved seeds to nonusers would be positive, though would be less than the benefits of those already using improved seeds. A simple comparison shows that improved seed users are, on average, 0.544 kg/m2 more productive than non-users, however, the switching model results suggest that non-users adopting would increase productivity by 0.293 kg/m2, whereas the matching model suggests that the benefit would be 0.330 kg/m2. The switching model highlighted diversity in the predicted response of farms to adoption, with some expected to response well to adoption, while others expected to have small or negative changes in productivity. This warrants further investigation into the characteristics under which the new technology is successful. Caution is required when drawing conclusions from these results, as the limitations of the dataset do not allow investigation into some aspects of the adoption and use processes. Indeed the fairly small average increase in productivity (and large overlap in productivity between the two groups) from the use of improved seeds makes it more likely that an excluded factor(s) could be responsible for the difference. The following paragraphs outline some of the possible considerations, which also serve as recommendations for the focus of future research. Firstly, though an attempt has been made to control for ability, the data available (such as age, education, household size) may not be sufficient to proxy for general farming ability, which is expected to be correlated with the decision to adopt. A second and related point is the list of factors for which data or proxies are unavailable. This list includes the climatic and soil conditions, the quality of the road infrastructure and the proximity to an urban centre. Such variables would be expected to influence the diffusion and uptake of improved seeds. Thirdly, it is well known that the timely use of cofactors (such as fertiliser, pesticide and fungicide, as well as labour) has a significant impact on yield. Some data is

9

available regarding labour availability (household size and whether the farm hires labour), which have been utilised in variations of the model. However, isolating the impact proved difficult due to multicollinearity with other variables. The proportion of production for which fertilizer was used is known, however for improved seed users this was often identical to the proportion of land planted by improved seeds, i.e. for many households the use of improved seeds and the use of fertiliser was a joint decision. Furthermore, details about the intensity of fertiliser use are not available. Fourthly, the success of maize cultivation can depend on how it is interplanted with other crops and can be important in maintaining long-term soil quality. Scott (1998:Chpt 8) argues that modern agricultural science struggles to incorporate an understanding of poly-cropping, partly because it does not fit conveniently into the typical methods of scientific analysis. Nevertheless, in general terms it is acknowledged that, in many cases, successful poly-cropping techniques have been developed in many regions over many generations and this would be an intriguing area for future investigation.

6. References Becker, Sascha and Andrea Ichino (2002) ‘Estimation of Average Treatment Effects Based on Propensity Scores,’ The Stata Journal, Vol.2, No.4, pp.358-77. Chamberlin, Jordan, John Pender & Bingxin Yu (Nov 2006) “Development Domains for Ethiopia: Capturing the Geographical Context of Smallholder Development Options,” Development Strategy and Governance Division Discussion Paper No.43, Environment and Production Technology Division Discussion Paper No.159, International Food Policy Research Institute (IFPRI), Washington DC Cochran W. G. & S. Paul Chambers (1965) ‘The Planning of Observational Studies of Human Populations,’ Journal of the Royal Statistical Society, Series A (General), Vol. 128, No. 2, pp. 234-266. Cochran, W. G. (1968) ‘The Effectiveness of Adjustment by Subclassification in Removing Bias in Observational Studies,’ Biometrics, Vol.24, No.2, pp. 295-313, June. CSA, IFPRI & EDRI (2006) Atlas of the Ethiopian Rural Economy, Central Statistics Authority (CSA), International Food Policy Research Institute (IFPRI) and the Ethiopian Development Research Institute (EDRI), Addis Ababa Imbens, Guido (2000) ‘The Role of the Propensity Score in Estimating DoseResponse Functions,’ Biometrika, Vol.87, No.3, pp.706-10. Lokshin, Michael & Zurab Sajaia (2004) “Maximum Likelihood Estimation of Endogenous Switching Regression Models” The Stata Journal Vol.4, No.3, pp.282289 Maddala, G. S. (1983) Limited-dependent and Qualitative Variables in Econometrics, Cambridge University Press National Bank of Ethiopia (2007) Exchange Rate Statistics downloaded from http://www.nbe.gov.et/MEFR/Exchange%20Rate%20for%20web%20Jan.%20%200 7.pdf on 3rd July 2007 Nega, Berhanu (2003) “The Ethiopian National Synthesis Report” Roles of Agriculture Project, PowerPoint Presentation prepared for the Roles of Agriculture International Conference, 20-22 October, 2003 – Rome, Italy, Agricultural and Development Economics Division (ESA), Food and Agriculture Organization of the United Nations Ramakrishna, G. and Assefa Demeke (2002) “An Empirical Analysis of Food Insecurity in Ethiopia: The Case of North Wollo,” Africa Development, vol. XXVII,

10

Nos. 1 & 2, 2002: 127-43, Council for the Development of Social Science Research in Africa Rosenbaum, Paul R. and Rubin, Donald B. (1983), "The Central Role of the Propensity Score in Observational Studies for Causal Effects", Biometrika, 70(1), 41-55. Schultz, Theodore (1964) Transforming Traditional Agriculture, University of Chicago Press Scott, James (1998) Seeing Like A State: How Certain Schemes To Improve The Human Condition Have Failed. Yale University Press

11

Table 1a Binary variables Variable Name

Description

Imuser Female Litanymem D_Achefer D_BahirDarZuria D_YilmanaDensa D_Sekela D_Quarit D_DegaDamot D_Dembecha D_JabTahinan D_BureWemberma Merawi

User of improved seeds (pcimseed > 0) Female-headed household Literate member of household

District dummies (omitted district: Merawi)

(for comparison)

Number of ones (out of 6,395) 2,460 989 2,009 692 689 708 553 591 418 713 641 694 696

Percentage

38.5% 15.5% 31.4% 10.8% 10.8% 11.1% 8.6% 9.2% 6.5% 11.1% 10.0% 10.9% 11.0%

Table 1b Continuous variables and variables truncated at zero Variable

Description

Mean

Std dev

Min

Max

Mprdvty

Maize productivity in kg per sq metre planted (20 zeros)

2.394

0.824

0

4.760

Household size Ln(household size) Age of household head in years Tropical Livestock Units (weighted sum of hhld livestock10) Total value of farm implements and storage facilities in 1000s of Birr (194 zeros)

5.054 1.521 41.89

2.073 0.476 14.00

1 0 13

14 2.639 98

5.187

5.065

0

131.4

0.101

0.101

0

2.927

Lareaha

Ln(total area of farm in hectares)

-0.047

1.045

-5.054

2.212

Pcmaizearea

Proportion of total farm area planted with maize

0.309

0.213

0.001

0.992

Hhsize Lhhsize Age Tlu Valsttool

Table 1c Proportional variables (truncated at zero and one) Variable

Description

Mean

Pcimseed

Proportion of maize area planted with improved seeds Proportion of maize area privately owned by the household Proportion of maize area suffering damage

0.2886

Pcprivown

Pcdamage

10

0.9047

0.1578

Std dev 0.3960

0.2469

0.2229

Zeros

Ones

3,935

Btw’n 0&1 1,741

(61.5%)

(27.2%)

(11.2%)

719

175

1,334

4,886

(2.7%)

(20.9%)

(76.4%)

3,434

2,940

21

(53.7%)

(46.0%)

(0.3%)

Weights are as follows: Cattle = 1, Goat = 0.15, Sheep = 0.15 (ranked as goats as is common in the literature), Horse = 1, Mule = 1.15, Donkey = 0.65, Camel = 1.45, Poultry = 0.005. (Ramakrishna & Demeke, 2002)

12

Table 2 Determinants of Adoption of Improved Seed Technology: Tobit and Probit Results (Doublecensored) Tobit

Probit (robust stnd errors)

Dependent variable

pcimseed Lower limit = 0 Upper limit = 1

imuser

Female

-0.22919*** (4.24)** 0.27341*** (7.79)** -0.29158*** (4.78)** -0.00743*** (6.09)** 0.49603*** (18.93)** 1.35803*** (13.65)** -0.76730*** (9.45)** 0.01299* (1.73) 0.17761*** (3.48)** -0.64906*** (10.43)** -0.64174*** (10.19)** -0.11131* (1.82) 0.48487*** (7.72)** 0.48201*** (9.88)** -0.14865* (1.84)

-0.24548*** (3.97)** 0.32319*** (8.21)** -0.35892*** (4.79)** -0.00765*** (5.39)** 0.69898*** (20.38)** 1.99024*** (16.54)** -0.83657*** (9.42)** 0.02374*** (2.68)** 0.14731** (2.53)* -0.69847*** (10.65)** -0.65813*** (9.80)** -0.18295*** (2.62)** 0.44196*** (5.84)** 0.47026*** (7.92)** -0.34064*** (3.60)**

Litanymem Pcprivown Age Ln(areaha) Pcmaizearea Pcdamage Ln(tlu) D_Achefer D_BahirDarZuria D_YilmanaDensa D_Sekela D_DegaDamot D_Dembecha Constant

dPr(imuser=1)/dx (robust stnd errors) imuser

-0.08552*** (3.97)** 0.11520*** (8.21)** -0.13041*** (4.79)** -0.00278*** (5.39)** 0.25397*** (20.38)** 0.72314*** (16.54)** -0.30396*** (9.42)** 0.00862*** (2.68)** 0.05474** (2.53)* -0.21492*** (10.65)** -0.20518*** (9.80)** -0.06409*** (2.62)** 0.17000*** (5.84)** 0.18029*** (7.92)**

Observations 6395 6395 Pseudo R2 0.2058 0.3510 Left-censored obs 3,935 Right-censored obs 724 Observed prob 0.3847 0.3847 Predicted prob (at x-bar) 0.3327 0.3327 Tobit model: Absolute value of t statistics in parentheses Probit model: Robust z statistics in parentheses Dprobit model: z-stats correspond to the test of the underlying coefficient being 0 † dF/dx is for discrete change of dummy variable from 0 to 1 ‡ predictions > 0.5 classified as 1; < 0.5 classified as 0 * significant at 10%; ** significant at 5%; *** significant at 1%

6395 0.3510

13

Table 3 Determinants of Maize Productivity: Switching Model Results

Ln(areaha)

Equation 1: Improved seed users (Dep var: mprdvty)

Equation 2: Non-improved seed users (Dep var: mprdvty)

0.02955 (0.75) 0.70184 (5.58)** -0.26155 (2.94)** 0.03019 (1.88) -0.00085 (0.76) 0.58882 (11.97)** -0.54836 (6.99)** -0.27614 (3.43)** -0.34113 (5.17)** -0.69144 (11.74)** 0.07039 (1.54) -0.03356 (0.64) 0.29969 (6.27)**

-0.05715 (2.42)* 0.23665 (3.01)** -0.41968 (8.25)** 0.01420 (1.30) 0.00133 (1.72) 0.37631 (8.46)** -0.16615 (3.59)** -0.15424 (3.38)** -0.82764 (18.73)** -0.08296 (1.89) 0.25797 (4.68)** 0.20126 (4.53)** 0.46230 (9.94)**

Selection model (select on imuser)

0.68333 (22.25)** Pcmaizearea 2.02416 (17.43)** Pcdamage -0.81167 (8.96)** Ln(valsttool) 0.05160 (2.54)* Age -0.00796 (5.65)** D_Achefer 0.21793 (2.88)** D_BahirDarZuria -0.64808 (7.86)** D_YilmanaDensa -0.59181 (7.12)** D_Sekela 0.05434 (0.69) D_Quarit 0.52371 (6.77)** D_Dembecha 0.03803 (0.49) D_JabTahinan 0.17775 (2.30)* D_BureWemberma -0.09055 (1.09) Female -0.22879 (3.69)** Litanymem 0.30569 (7.70)** Pcprivown -0.35010 (4.86)** Ln(tlu) 0.01804 (1.98)* D_DegaDamot 0.51330 (6.00)** Constant 2.60040 2.13623 -0.26691 (27.34)** (28.88)** (2.19)* LR test of indep. eqns. : chi2(1) = 5.13 Prob > chi2 = 0.0235

Absolute value of z statistics in parentheses; * significant at 10%; ** significant at 5%; *** significant at 1% Specification variables Cons ln(σu1) Cons ln(σu2) Cons Fisher transformation of corr(u1,ε) Cons Fisher transformation of corr(u2,ε) Observations Summary of Mills Ratios Inverse Mills Ratio Complement of the Mills Ratio

14

Value -0.40535 -0.37995 0.03622 -0.16898 6395

Mean 1.161 0.652

Std. Dev. 0.675 0.448

Min 0.0432 0.0000048

z-stat (28.11)** (28.33)** (0.35) (1.59) Max 4.955 2.477

Table 4 Summary of Predicted Values for Maize Productivity Type of user

Obs

Mean

St.dev

Min

Max

Users_using (real) Users_not using (hypothetical)

2,460 2,460

2.729 2.195

0.375 0.348

1.676 0.795

3.689 2.832

Nonusers_not using (real) Nonusers_using (hypothetical)

3,935 3,935

2.185 2.478

0.401 0.413

0.890 1.512

3.031 3.708

Where: Users_using Users_not using Nonusers_not using Nonusers_using

predicted values for improved seed users, continuing to use predicted values for improved seed users, if they used non-improved seeds predicted values for current non-improved seed users, continuing to not use predicted values for non-improved seed users, if they used improved seeds

Table 5 Results of Propensity Score Matching Model: Improved Seed Users Matching estimation method

No. of Treated (users)

No. of Controls (nonusers)

Average treatment on the treated (ATT)

Bootstrapped standard error

Nearest neighbour

2,460

1,156

0.327

0.030

11.0

Radius (0.01)

2,459

3,912

0.477

0.025

18.9

Radius (0.001)

2,264

3,405

0.463

0.031

14.9

Kernel

2,460

3,912

0.355

0.019

18.3

Stratification

2,460

3,912

0.338

0.028

12.2

4-way average

t statistic

0.373

Table 6 Results of Propensity Score Matching Model: Non-users Matching estimation method

No. of Treated (nonusers)

No. of Controls (users)

Nearest neighbour

3,935

1,177

- 0.296

0.049

-6.0

Radius (0.01)

3,935

2,458

- 0.418

0.025

-17.0

Radius (0.001)

3,428

2,264

- 0.399

0.028

-14.5

Kernel

3,935

2,459

- 0.311

0.035

-8.9

Stratification

3,935

2,459

- 0.306

0.044

-6.9

4-way average

Average treatment on the treated (ATT)

Bootstrapped standard error

t statistic

- 0.330

15

• Moisture reliable, highland • Drought prone, highland • Moisture reliable, lowland • Drought prone, lowland • Lowland pastoralist

West Gojam

Addis Ababa

Figure 1 Five Main Agro-ecological Zones of Ethiopia Source: Chamberlin et al (2006:26)

Figure 2 West Gojam Zone Source: CSA et al (2006:6) Note: The district marked on this map as Adet is now officially known as Yilmana Densa.

16

Density of Maize Productivity compared with normal density .6 .5

Density

.4 .3 .2 .1 0 0

1

2 3 Maize kg/sq-metre

4

5

Kernel density estimate Normal density

Figure 3 Kernel Density of Maize Productivity

Density of Maize Productivity Improved Seed Users versus Non-users .6 .5 .4 .3 .2 .1 0 0

1

2 3 productivity (kg per square metre)

Non-improved Seed Users

4

5

Improved Seed Users

Figure 4 Kernel Density of Maize Productivity by Use of Improved Seeds

17

Proportion of Improved Maize Seed Mean Proportion of Improved Seeds

Mean Area Planted per Household by District .6

0.55

.5 .4

0.40

0.38

0.35

0.30

.3

0.26 0.21

0.22

.2 0.12

0.10

.1

D am ot D em be ch Ja a b T ah Bu in re an W em be rm a

a

Q ua ri t D eg

Ba h

Ac he

fe ir r D ar Zu Yi lm ria an a D en sa M er aw i Se ke la

0

Figure 5 Proportion of Improved Maize Seed by District

Proportion of Improved Seed Use

>9 0% 10 0% us er s

<1 0%

10 % -2 0% 20 % -3 0% 30 % -4 0% 40 % -5 0% 50 % -6 0% 60 % -7 0% 70 % -8 0% 80 % -9 0%

0

200

Frequency 400 600

800

Improved Seed Users Only

Proportion of Maize Production Planted with Improved Seed

Figure 6 Histogram of the Proportion of Improved Seeds Planted among Users

18

productivity measured difference between users and non-users

actual benefit of improved seeds

A

“naïve” measured benefit of improved seeds B

not-using improved seeds

using improved seeds

Figure 7 Potential Source of Endogenous Selection Bias

Density Plot: Predicted Values for Maize Productivity

0

.5

1

1.5

split by current and supposed improved seed use

1

2 3 Maize Productivity (kg per sq-metre) current nonusers not using

4

current non-users adopting

Figure 8 Benefit of Improved Seeds to Current Non-users (Predicted Values)

19

Density Plot: Predicted Values for Maize Productivity

0

.5

1

1.5

split by current and supposed improved seed use

1

2 3 Maize Productivity (kg per sq-metre) current users using

4

current users not using

Figure 9 Benefit of Improved Seeds to Current Users (Predicted Values)

0

1

kdensity

2

3

Propensity Scores: Users vs Non-users

0

.2

.4

propensity score for non-users

.6

.8

1

propensity score for users

Figure 10 Propensity Scores from the Matching Model: Comparing Users and Nonusers

20

The Adoption and Productivity of Modern Agricultural Technologies in ...

Mar 17, 2008 - Firstly, using data derived from an agricultural census of the West ... represents a distinct agricultural technology that has great potential to raise.

296KB Sizes 0 Downloads 253 Views

Recommend Documents

Adoption of Information and Communications Technologies in ...
Information and communications technologies (ICT) have played a key role in .... development can be restricted as a network administrator's job is considered a ...

Land Markets, Resource Allocation, and Agricultural Productivity
Further emergence of private property during this period resulted in powerful ...... zones with the highest and lowest efficiency gain as these outliers cloud out.

Land Markets, Resource Allocation, and Agricultural Productivity
family farms are the basic unit of production and we use the detailed household-level data of ..... Notes: All variables are in log scale for the purpose of illustration. ..... fixed effect, farm productivity, and land quality, and then we obtain the

The Agricultural Productivity Gap
For example, the adjusted gaps are larger, all else equal, when the quality of a ...... To explore this possibility, we analyze individual-level data from LSMS ...

Agricultural Productivity, Comparative Advantage ... - ScienceDirect.com
many developing countries2 In an open trading system, where prices are ... link between agricultural productivity and economic growth. An economy with less ... of closed economy models of endogenous growth, uses cross-country data.

The Agricultural Productivity Gap
Nov 11, 2013 - Computing “Raw” Agricultural Productivity Gaps. Measures of VAa and VAn .... Not Artifact of Secondary Jobs. Sector of Hours Worked. Country.

The Agricultural Productivity Gap in Developing Countries
Dataset downloaded from http://cbs.gov.np/nada/index.php/catalog on 08/11/2011. Central Statistical Office, Labour Fource Survey 2005 (LFS 2005),Version 1.1 ...

Agricultural Productivity Differences Across Countries
Jan 14, 2014 - Host of measurement concerns. • Exclude home production, which is very common in poor countries? • Agricultural output data otherwise badly ...

Adoption of Technologies with Network Effects: An ...
chines, and the increased use of the Internet. In such networks .... into the relationship between network size and a bank's propensity to adopt ATMs that ... tuting the automated teller for the human one during normal business hours, and will.

Land, Occupational Choice, and Agricultural Productivity
The prevalence of untitled land in poor countries helps explain the international agricultural productivity differences. Since untitled land cannot be traded across farmers, it creates land misallocation and distorts individuals' occupational choice

Adoption of Technologies with Network Effects: An ...
ness and Economics Program at MIT for financial support. ... their accounts. .... tuting the automated teller for the human one during normal business hours, and will .... decline of the cost of adopting decreases over time, the smallest T that .....

State Rights over Water and Agricultural Productivity ...
Apr 25, 2013 - Email address: [email protected] (Shreyasee Das) ...... in-differences estimate of -0.077.19 In Column 3, I add ..... in/kwdtaward.html. [26] Iyer ...

State Rights over Water and Agricultural Productivity ...
Apr 25, 2013 - Bay of Bengal. The Krishna Basin spreads ...... An additional concern is that with a long post-decision time window (my analysis extends from ...

A New Look at Agricultural Productivity and Economic ...
May 13, 2010 - choose the optimal consumption (i.e. crop production) and savings (i.e. ..... and/or gaseous molecules interact with the soil to deposit nitrogen.

Adoption of Management Practices in the Public Sector of Bangladesh
Jan 20, 2017 - Awareness of management concepts (Plan-Do-Check-Act, Total Quality Management, ..... Table 1 reports summary statistics of all variables.

Adoption of Management Practices in the Public Sector of Bangladesh
Jan 20, 2017 - management training about concrete methods and practices may improve the adoption of ... These concepts (especially the first two) are common in introductory business and ... During the past 25 years, Bangladesh has achieved stable ...

The Role of Consumer Innovativeness in the Adoption ...
Domain-specific innovativeness was found to have the biggest impact. (b=0.61), followed by Internet ..... Industrial Management and Data Systems, 100 (7): 294-300. Fram, E. H., and Grady, ... New York: Free Press. Sheth, J. N. and Sisodia, ...

DEVELOPMENT, ADOPTION AND PERFORMANCE OF Bt COTTON ...
Page 1 of 13. 73. DEVELOPMENT, ADOPTION AND PERFORMANCE OF Bt COTTON. IN PAKISTAN: A REVIEW. Raania Ahsan* and Zafar Altaf**. INTRODUCTION. Globally, cotton and other crop plants. require an intensive use of pesticides to. inhibit insect pests popula

'Cloud' Hanging Over the Adoption of Cloud Computing in Australian ...
Dec 11, 2016 - of what the term cloud computing means and its benefits; the 23% of .... all wireless and wired systems that permit users in sharing resources.

'Cloud' Hanging Over the Adoption of Cloud Computing in Australian ...
Dec 11, 2016 - In Australia, cloud computing is increasingly becoming important especially with the new accessibility provided by the development of the ...