Asian Journal of Economics and Empirical Research ISSN(E) : 2409-2622 ISSN(P) : 2518-010X Vol. 3, No. 2, 145-155, 2016 http://asianonlinejournals.com/index.php/AJEER 145 Impact of Milk Marketing Channel Choice Decision on Income, Employment and Breeding Technologies among Dairy Farmer Households in Kericho County, Kenya Elijah K. Ng’eno1 1 University of Kabianga, Department of Agricultural Biosystems and Economics, Kericho, Kenya and PhD Student, Moi University, Eldoret, Kenya Abstract The study examined the impact of milk marketing channel choice decisions on dairy farmer household income, employment and breed technologies among dairy farmer households in Kericho County. Data was collected from 432 dairy farmer households using multistage cluster sampling technique. Both primary and secondary data were used in the analysis. Processing and analysis of survey data was carried out using STATA version 12. Multivariate probit and propensity score matching was used in data analyse. Propensity score matching was also used to account for selection bias. Matching results show that the average effect of the farmer household that sold milk to commercial buyers had higher probability of obtaining Kenya shillings 16.00 per day as compared to households that did not sell milk to commercial buyer. While selling through commercial milk buyers had significant positive effect on farmer welfare, majority of dairy farmers were hesitant to engage with them. Milk buyers value security in supply which comes from trusted relationships and contracts. Establishing such relationships is in the long-term interest of the dairy farmer. Therefore, to improve farmer welfare, group formation and partnership development should be strengthened, milk cooperative societies needs to be bolstered and an increased financial investment in livestock markets by national and county governments. Keywords: Propensity score matching, Milk marketing channels, Dairy farmer households. Contents 1. Introduction ....................................................................................................................................................................... 146 2. Objectives ........................................................................................................................................................................... 146 3. Research Methodology ...................................................................................................................................................... 146 4. Empirical Results and Discussion ..................................................................................................................................... 151 References .............................................................................................................................................................................. 155 Citation | Elijah K. Ng’eno (2016). Impact of Milk Marketing Channel Choice Decision on Income, Employment and Breeding Technologies among Dairy Farmer Households in Kericho County, Kenya. Asian Journal of Economics and Empirical Research, 3(2): 145-155. DOI: 10.20448/journal.501/2016.3.2/501.2.145.155 ISSN(E) : 2409-2622 ISSN(P) : 2518-010X Licensed: This work is licensed under a Creative Commons Attribution 3.0 License Funding: This study received no specific financial support. Competing Interests: The author declares that there are no conflicts of interests regarding the publication of this paper. Transparency: The author confirms that the manuscript is an honest, accurate, and transparent account of the study was reported; that no vital features of the study have been omitted; and that any discrepancies from the study as planned have been explained. History: Received: 28 September 2016/ Revised: 14 November 2016/ Accepted: 17 November 2016/ Published: 19 November 2016 Ethical: This study follows all ethical practices during writing. Publisher: Asian Online Journal Publishing Group http://creativecommons.org/licenses/by/3.0/ http://crossmark.crossref.org/dialog/?doi=10.20448/journal.501/2016.3.2/501.2.145.155 https://orcid.org/orcid-search/quick-search?searchQuery=Elijah K. Ng%E2%80%99eno http://search.crossref.org/?q=10.20448/journal.501/2016.3.2/501.2.145.155 http://crossmark.crossref.org/dialog/?doi=10.20448/journal.501/2016.3.2/501.2.145.155 https://orcid.org/orcid-search/quick-search?searchQuery=Elijah K. Ng%E2%80%99eno http://search.crossref.org/?q=10.20448/journal.501/2016.3.2/501.2.145.155 http://crossmark.crossref.org/dialog/?doi=10.20448/journal.501/2016.3.2/501.2.145.155 https://orcid.org/orcid-search/quick-search?searchQuery=Elijah K. Ng%E2%80%99eno http://search.crossref.org/?q=10.20448/journal.501/2016.3.2/501.2.145.155 Asian Journal of Economics and Empirical Research, 2016, 3(2): 145-155 146 1. Introduction Output prices received by farmers significantly determine their welfare especially in rural areas where there is weak non-farm income which limits diversification of agricultural production amongst producers. While there is some debate about the actual and potential impacts of having a wide array of commercial milk buyers on broader welfare of the rural poor, case study evidence suggests that farmers are worst placed when faced with a privately owned or government-controlled monopsony (Sadler, 2006; Gorton 2007). The choice of a milk marketing channel depends on a number of aspects. These include availability of the markets, milk prices offered in the markets, distance to the milk market and the potential of the market to absorb the volume on sale (Paterson, 1997). However, there is relatively little evidence linking milk marketing channel choices to dairy farmer household income and employment outcomes in the study area. While the theoretical arguments in favor of marketing cooperatives are well known, in practice their performance in developing countries has been patchy (Glover, 1987). Erratic rent seeking government intervention may reinforce these problems. While case studies (Striewe, 1999; Cocks et al., 2005; Gorton et al., 2006) and aggregate market analysis identify these difficulties, there is an absence of cross-sectional data analysis on the impact of milk marketing channel choice decisions on income, employment and technology among dairy farmer households in Kericho County, Kenya. This study therefore, attempts to fill this gap by analyzing the impacts of the available commercial milk buying channels on dairy farmer households in Kericho County, Kenya. However, certain factors are beyond the scope of the dairy farmers in the study area. For example, weak rural infrastructure, land fragmentation and weak capital base. Therefore, most dairy farmers are constrained by high transaction costs due to the great distances to milk markets, lack of adequate access to finance, and in some cases social factors (Nkosi and Kirsten, 1994). Also, information about markets and market prices guide the farmer in making informed decisions (Nkosi and Kirsten, 1994). Making uninformed decisions may result in the farmer accessing the market when it is not profitable to do so. This is a common situation where livestock farmers approach saturated markets with the wrong price signals (Nkosi and Kirsten, 1994). The study contributes to the literature by empirically examining the impacts of the milk marketing channel choice decisions on dairy farmer household income, employment and technology. The study used propensity score matching model to control for self-selection since the choice decision for a particular milk buying channel is not random, with the group of dairy farmers being systematically different. By considering the causal relationship between participation in selling milk to commercial milk buyers and dairy farmer household welfare, this paper seek to address counterfactual queries that could be important in forecasting the impacts of policy changes. The study analyzed independently both farmers selling under commercial channels and for those that sell to final consumers in order to assess their impacts on the choice decisions made by the dairy farmer. 2. Objectives 2.1. General Objective The general objective of the study was to evaluate the impact of milk marketing channel choice decision on dairy farmer household’s income, employment and technology adoption in Kericho County, Kenya. 2.2. Specific Objectives The specific objective of the study was to estimate the impact of milk marketing channel choice on dairy farmer household’s income, labor hours and breed technology in Kericho County. 2.3. Hypothesis The study tested the following hypothesis:- HO1: The milk marketing channel choice has no impact on dairy farmer household income, labour hours and breed technology in Kericho County. 3. Research Methodology 3.1. Research Design This study used cross-sectional and correlational research designs. A cross-sectional survey of dairy farmer households was carried out in early 2015. The study estimated income, employment and technology functions, endogenously stratified for each of the marketing channel used. 3.2. Study Area Primary data was collect from smallholder dairy farmer households in six sub counties of Kipkelion East, Kipkelion West, Kericho West, Kericho East, Sigowet/Soin and Bureti of Kericho County as shown in table 2.1. Table-1. Smallholder Livestock Milk Producers and Cooperative Societies Sub – County Number of households Number of Dairy Farmers Dairy Cattle Population Average Number of Dairy Cows /Farmer Dairy Cooperative societies /companies Dairy Self- help groups Kipkelion East 27791 13,996 20666 5 2 1 Soin/Sigowet 20940 15,141 12808 2 1 2 Kericho west 31394 17,111 26007 2 1 0 Bureti 30977 28,304 11400 2 11 24 Kericho East 27700 8,150 10498.8 3 2 3 Kipkelion West 14615 11,725 18667 4 4 3 Kericho County 153417 94,427 100047 3 21 33 Source: Kericho County Development Profile, 2013 Asian Journal of Economics and Empirical Research, 2016, 3(2): 145-155 147 3.3. Target Population The primary unit of analysis was the household with dairy cows with milk for marketing. The target population was restricted to the 94,427 smallholder livestock milk producers and marketers, divided proportionately amongst the six sub-counties of Kericho County as shown in Table 1. Given the objective of the study, the population of interest was defined as the primary milk producer household who sell cows’ milk to another supply chain actor. Therefore, farmers without dairy cows, those who did not sell any of the milk produced or those who processed all the milk themselves were excluded from the study. With the given focus, these restrictions were justified, and it meant that the sample could not be directly compared to official data on the structure of milk production. 3.4. Sampling Procedure A multistage cluster sampling procedure was used to get the total population and sample size of interest.To achieve the study objectives, the county was clustered into six sub-counties, namely, Kipkelion East, Kipkelion West, Kericho West, Kericho East, Sigowet/Soin and Bureti as shown in Table 2. These sub-counties formed the sample sites for the study and the mean of the results from all these sites constituted the results for the whole county. Table-2. Sub-County Sampling Areas Constituency Sub-counties Divisions Area (Km 2 ) Number of Locations Number of Sub locations Ainamoi Kericho East Ainamoi 239.9 11 24 Belgut Kericho West Kabianga and Belgut 337.4 12 27 Sigoewet/Soin Sigowet Soin and Sigowet 473.2 13 38 Kipkelion West Kipkelion West Kunyak, Chilchila, Kamasian and Kipkelion 333 16 35 Kipkelion East Kipkelion East Londiani, Sorget and Chepseon 774.4 14 32 Bureti Bureti Bureti, Roret and Cheborge 321.1 19 53 Source: County Commissioner’s Annual Report, Kericho, 2013 To achieve representative sample size, the six sub-counties formed the first-stage cluster that had the target population. These six clusters were selected based on the fact that small scale dairy farming was dominant and practiced throughout the county. Furthermore, it reflected significant differences in structure of the dairy milk marketing industry in the county. Further, within the six sub-counties, second-stage cluster sample of wards and villages with high concentration of small scale dairy farmers was then selected for the study. Sample selection of dairy farmer households from the clustered wards was done using random sampling and effort was made to include statistically significant sub‐samples of dairy milk producers representing different milk marketing channels and sizes for each of the sub counties. The sampled milk producing nth smallholder farmer household was determined by the proportionate size sampling methodology (Anderson et al., 2007). 2 2 0 e pqZ N  (1) Where 0N was the sample size, Z is the standard normal value of 1.96 significant at 5 percent confidence level, e is themargin of error (the sampling error or desired level of precision), p is the estimated population proportion of smallholder dairy farmers with characteristics of interest assumed at 70 percent (Table 3.3). Thus taking p at 70 percent gave a representative size with minimal error making q = 1-p, i.e. 1-0.7 = 0.3, Z = 1.96, and e = 0.04 for a good precision respectively. 504 04.0 3.0*7.0*96.1 2 2       N dairy farmer households. Finally, based on the above calculation, the sample units (number of dairy farmers households) were calculated proportionately based on the number of dairy farmer households in each sub county and as a proportion of the total dairy farmers in the county against the desired sample size of 504 as shown in Table 3. Table-3. Proportionate Distributions of Dairy Farmer Households Sub – County Number of Households Number of Dairy Farmers Percent Dairy farmer Households Total proportion Kipkelion East 27,791 13,996 15 75 Soin/Sigowet 20,940 15,141 16 81 Kericho west 31,394 17,111 18 91 Bureti 30,977 28,304 30 150 Kericho East 27,700 8,150 8 44 Kipkelion West 14,615 11,725 12 63 Total Kericho County 153,417 94,427 100 504 Source: Author’s Computation from County Data, 2016 Therefore, a random sample of 504 dairy farmer households was set for the whole county with the intention of sampling representative cross-section of small scale dairy farmer households selling raw milk to different marketing channels available at farm gate. After data entry and cleaning, a total of 432 households were finally used for data analysis (Table 4). Within the county, sampling was weighted to the six sub-counties that had significant dairy cow production. The results that were obtained were assumed valid for the whole County. Asian Journal of Economics and Empirical Research, 2016, 3(2): 145-155 148 Table-4. Distribution of Sample Smallholder Dairy Households Sub-county Respondents Percentage Ainamoi 54 12.50 Bureti 122 28.24 Kericho West 70 16.20 Kipkelion East 50 11.57 Kipkelion West 61 14.12 Soin/Sigowet 75 17.36 Total 432 100.00 Source: Author’s Computation from Survey Data, 2016. 3.5. Data Types and Sources Both primary and secondary data were used in this study. Primary data were collected through a survey. A structured pre-tested questionnaire was used and administered by trained enumerators through direct interviews amongst selected dairy farmer households. Primary data was also collected through discussion and observations of the farmers’ dairy farming activities. Seasonal observations were also used to correlate those dairy farmer households left out from the network with the secondary milk market in order to estimate their supply and price variations. This involved observing the natural behavior (nonverbal expression of feelings, determine who interacts with whom, how dairy farmers communicate with each other, and to check for how much time is spent on various activities) of the dairy farmer households in order to describe existing situations and to obtain information that was relevant to the goals of the study. Secondary data was obtained mainly from various sources including economic surveys, economic journals, statistical abstracts, conference reviews, books, magazines, official government of Kenya reports and documents such as statistical abstracts and bulletins, national and district development plans, national and county development and strategic plans, and Kenya Dairy Board records and annual reports. Different documents of livestock production and marketing, regional level reports and consultants’ reports as well as National Bureau of Statistics publications were reviewed to gather more relevant information. Desktop literature and internet were also used to access credible information from available and accessible documents, published and unpublished reports, books and agricultural journals. Farm records from a few dairy farmer households were also used to supplement secondary data sources. Given the importance accorded to the involvement of smallholder milk producers in the various milk marketing channels in this study, data types encompassed representative sample of households representing various categories of households, types of marketing channels (commercial and non-commercial), and changing structure of dairy sector was adopted. In order to analyze the response of the smallholder milk producers, the study focused mainly on whether the dairy farmer household sold milk at farm gate to commercial milk marketing channels (Y1) and if farmer household chose to sell also to final consumers (non-commercial channel) (Y0) or otherwise. Commercial milk marketing channels in this study were taken to mean three major marketing channels: organized cooperative societies, organized private sector milk buyers, and traditional/unorganized milk buyers. For a given village, there were four types of farmers: (i) farmers who chose to supply milk to the organized cooperative societies, (ii) farmers who chose to sell milk to the organized private sector milk buyers, (iii) farmers who chose to supply milk to the traditional or unorganized milk buyers such as milk vendors, restaurants, or directly to consumers and contractors and (iv) farmers who supplied milk to multiple channels like milk cooperative societies self-help groups, traditional and private milk buyers. The data collected included dairy farmers’ socio-economic characteristics, actual milk production, milk market competitiveness and other related obligations with the milk buyers. The socio-economic data collected comprised the farmer’s age, education level, household size, gender, and farm ownership, off farm income, access to credit, access to extension service, membership to milk cooperative society and access to other milk marketing channels. The farm production data collected comprised the size of land under dairy production, average volume of milk produced per year, amount of livestock inputs such as feeds, breeding methods, types of labor used, capital used, cost of inputs, and farm gate prices of livestock outputs. Respondents were also expected to provide information regarding market competitiveness and an estimated total number of potential commercial buyers for their milk. This would capture the degree of switching power from one commercial buyer to the other that farmers have in marketing their milk. The study also included data on whether the farmer sold total milk output on contract or on signing agreements or on spot cash sale as an independent variable. Farmers may sell their milk on signing agreements with milk buyers rather than via spot cash sales. Agreements with buyers provide a greater degree of certainty for buyers regarding the availability of supply, for which a buyer may pay a premium (Gow and Swinnen, 2001). To capture the trustworthiness of commercial milk buyers, a measure of trust on the commercial milk buyer by the dairy milk farmer was included. This attribute was analyzed by a proxy that identified the perception that the dairy milk farmer had in relation to their trust in the commercial milk buyer. Finally regarding milk marketing characteristics, a dummy variable was introduced that captured whether the farmer sells via milk cooling/chilling plants, milk sheds or milk bars or not. Time series data on farm gate milk prices received by farmers over a period of three years (2013, 2014 and 2015) was also collected from the farmers. This entailed the use of pairwise comparison of the six sub county mean milk prices (means that were significantly different from each other) for the three years using Tukey's HSD (honest significant difference) test. 3.6. Instruments of Data Collection A structured questionnaire was used as an instrument for data collection. The questionnaire was designed to address the objectives of the study. The questionnaires were administered by trained enumerators. The enumerators were identified from among the people who were conversant with the sub county wards and villages in the county to Asian Journal of Economics and Empirical Research, 2016, 3(2): 145-155 149 aid in data collection. The enumerators were trained for two days and the training culminated in the pre-testing of the questionnaire on the third week of December, 2015. Pre-test of data collection tool on the four dairy farmer households was done in Kericho East and Bureti sub counties respectively. Observations were also used to correlate those left out from the sampled population with the secondary milk marketing channels in order to provide estimates of their milk supply and milk prices received. 3.7. Data Analysis and Diagnostics STATA version 12 was used for data analysis. The collected primary data was collated, cleaned, coded and stored in excel worksheets and IBM- SPSS version 21 before they were transferred to STATA for analysis. Empirical analysis in this study consisted of two stages. In the first stage, multivariate probit model was used to estimate the factors which determined the milk marketing channel choice decision equation, specifically whether farmers sell only to a commercial milk buyer or sell also to final consumers of milk. Secondly, propensity score matching model was used to analyze the impact of farmers’ marketing choices on gross dairy income, milk yield, labor hours, and on breed technology). Diagnostic tests were also conducted from the regression results of STATA output. To check on multicollinearity, the study used variance inflation factor (VIF) and contingency coefficient (CC) among discrete and continuous variables, respectively. All assumptions were tested and corrected accordingly using STATA. 3.8. Analytical Frameworks 3.8.1. Theoretical Framework The farmer or producers behave like neoclassical firms who control the transformation of inputs into valuable outputs in order to maximize profits (Varian, 2000). The decision on whether or not to adopt a new technology is considered under the general framework of utility or profit maximization (Norris and Sandra, 1987; Pryanishnikov and Katarina, 2003). It is assumed that economic agents, including smallholder subsistence farmers, use certain livestock milk marketing systems only when the perceived utility or net benefit from using such a method is significantly greater than is the case without it. Again smallholder dairy farmers are assumed to be rational and they want to derive the highest utility from the choices they make; either to market their produce independently or under a certain milk marketing channel. They make their choices with respect to random utility theory, which states that a decision maker is guided by unobservable, observable and random characteristics when making a decision. Although utility is not directly observed, the actions of economic agents are observed through the choices they make. Suppose that Yj and Yk represent a household’s utility for two milk marketing choices, which are denoted by Uj and Uk , respectively. The linear random utility model could then be specified as: jij XU   and kikk XU   (2) Where; Uj and Uk are perceived utilities of using a certain milk marketing channel j and k, respectively. Xi is the vector of explanatory variables that determines and or influences the perceived desirability of the choice of the milk marketing channel, Bj and Bk are parameters to be estimated, and εj and εk are error terms assumed to be independently and identically distributed (Greene, 2003). Therefore, for the case of choice of a livestock milk marketing channel, if a household (dairy farmer) decides to use option j marketing channel, it follows that the perceived utility or benefit from option j marketing channel is greater than the utility from other options (say k) marketing channel depicted as follows: ),(()( 11 kikikjijij XUXU   k ≠ j (3) The probability that a dairy farmer will choose milk marketing channel j among the set of livestock milk marketing channels to market his milk instead of the k marketing channel could then be defined as )()|1( ijij UUPXYP  (4) Therefore, )|0( 11 XXXP kikjij   Hence )|0( 11 XXXP kjikij   )(|0( *** ii XFXXXP   (5) Where; P is a probability function, Uij, Uik,, and Xi are as defined above, ε* = εj –εkis a random disturbance term, )( 11* kjj   is a vector of unknown parameters that can be interpreted as a net influence of the vector of independent variables influencing the decision to sell a commercial milk marketing channel, and )( * iXBF is a cumulative distribution function of the error terms (ε*) evaluated at iXB* . The exact distribution of F depends on the distribution of the random disturbance term, ε*. Depending on the assumed distribution that the random disturbance term follows, several qualitative choice models can be estimated (Greene, 2003). Propensity score matching which is used in this study’s analysis requires no assumption about the functional form in specifying the relationship between outcomes and predictors of outcome, unlike the parametric methods mentioned above. However, the drawback of the approach is the Conditional Independence Assumption (CIA), which states that for a given set of covariates, participation is independent of potential outcomes (Smith and Todd, 2005). Further, Smith and Todd (2005) note that there may be systematic differences between the outcomes of participants and non-participants, even after conditioning on observables. Such differences may arise because of selection into treatment based on unmeasured characteristics. To address the selectivity bias problems associated with choice decision of a milk marketing channel, this study employed the matching techniques in assessing the impact of selling to a commercial milk marketing channel on average dairy farmer household’s gross income, employment and dairy breeding technology uptake. The propensity score matching approach addresses the problem of the limited distributional assumption of the errors, and more importantly allows for a decomposition of the treatment effect on outcomes (Heckman, 1999). Also the counterfactual framework could detect “two important sources of bias in the estimation of treatment effects. Asian Journal of Economics and Empirical Research, 2016, 3(2): 145-155 150 These include the initial differences between the group selling to commercial milk market channels and those selling to final consumers in the absence of treatment, and the difference between the two groups in the potential effect of the treatment. 3.8.2. Empirical Modeling of Effects of Market Channel Choices on Income, Labor Hours and Technology Farmer’s milk marketing channel choice decisions in this study were hypothesized to have not significant impact on various technological and economic parameters, such as income, productivity; employment and technology (breed composition). Here, the study estimated income, employment and technology functions, again endogenously stratifying for each of the marketing channel used. Since the separation of producers by market channel introduces a bias derived from an endogenous stratification of market channels, this bias needed to be corrected. The regression equations were estimated for the group selling to commercial milk market channels and those who sell to final consumers. Therefore, the structural model that was adopted for the analysis was the propensity score matching model (PSM) as shown in equation 2.7 and as adopted from Heckman (1999). iiii uKXY   (6) Where; Yi is household income, employment or technology; Xi is a vector of the explanatory variables, representing personal and household characteristics and assets, and distance; Ki is the dummy variable representing one, if a dairy farmer sells to a commercial milk buyer and 0 for those selling also to final consumers;  are the coefficients and iu is the error term. The specification above in equation treats milk market choice decision by the dairy farmer household as an exogenous variable on the premise that households opts for a milk buyer to increase their income, employment of resources or to improve on their technological status. Nonetheless, this need not be the case, since better-off dairy farmer households may be better predisposed to several commercial milk markets as compared to the poor dairy farmer households. Furthermore, the decisions to choose or not to choose a particular milk marketing channel choice may be dependent on the benefits from the choice of the milk marketing channel. Thus, the choice of the milk marketing channel is not random, with the group of dairy farmers being systematically different. However, selection bias occurs if unobservable factors influence both the error term ( u ) of the choice equation, and the error term (  ) of the income equation, thus resulting in correlation of the error terms. To evaluate the impact of various milk marketing channel choice decisions on smallholder dairy farmer’s income, both farmers selling under commercial channels and for those not were expected to show the same observable characteristics. The study assumed that those selling through commercial milk marketing channels were taken as treatment and those not were taken as control. The average treatment effect (ATE) of involving commercial buyers is the difference between the actual income and the income for involving the commercial buyers in milk marketing, which is expressed as; 1/( 01  iii KYYEATE (7) Where; iY1 is the income when ith farmer sells to commercial buyer, iY0 is the income when the ith farmer markets independently to final consumers and Ki is a dummy variable denoting the involvement of the commercial buyer, 1 = selling to commercial milk buyers, 0 = otherwise. The mean difference (D) between observable and control can be written as in equation 8 below.  ATEKYEKYED ii )0/()1/( 01 (8) Where;  is the bias. The estimated model used for the fourth hypothesis related to the impact on farmers’ milk marketing choices, Y1i, (a binary variable which takes the value one if the farmer sells to commercial milk buyers only and zero if the farmer decides to sell also to final consumers), and their impacts on farmers’ income, employment, and technology (Zij), is as specified in the equation below: ij ij uIMRFAMILYSIZELANDSIZE EXPHERDVETFDSPARTROADEDCAGEZ   1099 8765432,10 PR   (9) Zij is a set of variables that were hypothesized to affect the farmer’s marketing channel choices (Y1j). These were the gross dairy income, milk yield, employment, and share of crossbred animals as dependent variables. Ideally, the dependent variable was the net dairy income. Regrettably, it was quite difficult to obtain accurate data on the value of some of dairy inputs. This was mainly true of dairy inputs for which livestock markets were not well developed, such as labor, home grown feeds and fodder, home-made feed ratios and in some cases costs data were missing completely. A major reason why propensity score matching was employed was to address potential unobserved heterogeneity. As observed by Hujer et al. (2004) a possible hidden bias might occur if there are unobserved variables that tend to influence simultaneously commercial milk marketing channel choice decision and dairy farmer household income, employment and breed technology. Given that it is not possible to estimate the magnitude of selection bias with non-experimental data. Rosenbaum and Donald (1985) suggested the use of the bounding approach to examine the influence of unmeasured variables on the selection process. As a consequence, the study used gross dairy income per animal per household as the dependent variable in the second stage of the Heckman model. The Inverse Mills’ Ratio was also be used to correct the error terms in the impact equations to achieve consistent and unbiased estimates. 3.9. Diagnostic Tests for Multinomial Logit The study used variance inflation factor (VIF) and contingency coefficient (CC) among discrete and continuous variables, respectively. Potential multicollinearity among explanatory variables was tested and it was found not to Asian Journal of Economics and Empirical Research, 2016, 3(2): 145-155 151 have any potential influence on estimates from the model. The highest pair-wise correlation was 0.4 whereas multicollinearity is a serious problem if pair-wise correlation among regressors is in excess of 0.5 (Gujarati, 2004). An analysis of variance inflation factor did not show any problem since none of the VIF of a variable exceeded 8 (Greene, 2003). 4. Empirical Results and Discussion Probit model was used to estimate the propensity scores. The scores were only used to balance the observed distribution of covariates across the treated group (farmers selling to commercial buyers) and the untreated group (farmers not selling to commercial buyers). The independent variables used in the probit regression model to predict the propensity scores were based on past research on determinants of participation in nonfarm employment (Barrett et al., 2001) in Owusu et al. (2014). From Table 5 results, most of the variables included in the estimation have the expected sign. In particular, participation in milk marketing and household size were found to be positively and significantly related to milk marketing. The coefficient for dairy farmer household participation in milk marketing was positive and significant. The presence of a milk buyer enhanced the probability of participation in the milk market. According to Kousar and Abdulai (2015) endowments with valuable household assets represents household’s wealth and the presence of a development project in an area enhance the probability of participation for both male and female in non-farm earning activities. The coefficient of household size was positive and significant balancing factor for propensity score matching. This suggests that the presence of labor availability in the household tend to increase the milk output levels which will then increase the probability of dairy farmer households selling their milk to commercial milk buyers. These results are in contrast to the study of Barrett et al. (2001) and in line with the study of Kousar and Abdulai (2015) in the case of male labor supply. The distribution of propensity of propensity scores before and after matching clearly indicate that estimating the p-score appears to balance the treated and untreated groups extremely well than without the p-score, a result which underscored the significance of the propensity score matching approach for this study. Table-5. Probit Estimates of the Propensity Score for Dairy Farmer Household’s Milk Marketing Involvement Major farm gate milk buyers (Commercial or Final consumers) Coefficient Standard Error Z P>|z| Age of household head -0.0026244 0.0083386 -0.31 0.753 Education level 0.0103616 0.0597299 0.17 0.862 Distance to Milk Market -0.0257749 0.0162404 -1.59 0.012** Milk marketing participation 0.319914 0.1308435 2.45 0.014** Price risk -0.2414148 0.1867471 -1.29 0.006* Veterinary feeds 0.0512779 0.1309099 0.39 0.695 Farming experience 0.0074178 0.0091395 0.81 0.017** Total farm size 0.0105474 0.0100193 1.05 0.002* Total household size 0.0484578 0.0275248 1.76 0.008* Gross Income 0.005435 0.0022667 2.40 0.016** Net Income -0.0042365 0.0021526 -1.97 0.049** Employment hour 0.003493 0.0031162 1.12 0.002* Technology (Breeding) -0.0014355 0.0018119 -0.79 0.028** Constant -0.1393711 0.454589 -0.31 0.759 Key: Caliper = 0.001 Nearest Neighbour = 1 Number of observations = 432 LR chi 2 (13) = 25.97 Prob > chi 2 = 0.0172 Log likelihood = -284.59954 Pseudo R 2 = 0.0436 * Significant at 1 percent; ** significant at 5 percent; *** significant at 10 percent Source: Author’s Computation from Survey Data, 2016 Pseudo-R 2 from probit estimation indicated the goodness of fit of the model or how well the regressors explained the probability to sell to commercial buyers. From the table of results, after matching, pseudo- R2 was fairly low (0.0436), which showed that the matching procedure balanced the determining factors (covariates) very well for this study. 4.1. The Mean Differences in Outcome Variables in the Matching Analyses Table 6 compares the mean differences in the outcome variables and other household and farm-level variables between dairy farmer households selling milk to commercial buyers and dairy farmer group not selling milk to commercial buyers. Given that the mean difference comparisons do not account for the effect of other characteristics of farm households, they confound the impact on household gross income, employment and breed technology adoption (dairy breeding) status with the influence of other characteristics. The significance levels suggest that there are some differences between those who sell to commercial milk buyers and those not with respect to household and farm-level characteristics. With regard to the outcome variables, there were statistically significant differences in household income and employment hours between the two categories of dairy farmers. Therefore, we again rejected the null hypothesis that the milk marketing channel choice has no impact on dairy farmer household income, labour hours and breed technology in Kericho County. Asian Journal of Economics and Empirical Research, 2016, 3(2): 145-155 152 Table-6. Propensity Score (Pscore) Matching Analysis Variable Sample Treated Controls Difference S.E. T-stat Net Income Unmatched 172.252586 148.97167 23.2809153 12.3018885 1.89 ATT 172.450649 154.408731 18.0419177 17.2355653 1.05 ATU 148.849432 162.68549 13.8360576 . . ATE 16.1310585 . . Gross Income Unmatched 206.621994 190.094018 16.5279754 12.8656873 1.28 ATT 207.007684 188.762198 18.2454867 18.3727788 0.99 ATU 189.250915 193.497668 4.24675287 . . ATE 11.8854063 . . Employment hours Unmatched 10.8287367 9.75894024 1.06979644 2.48402087 0.43 ATT 10.925244 6.45855443 4.46668959 2.76165675 1.62 ATU 9.85423381 8.3365533 -1.51768051 . . ATE 1.74779545 . . Technology Unmatched 55.0968127 58.0400175 -2.94320479 3.50114827 -0.84 ATT 55.3412638 56.0343963 -0.693132423 5.2095683 -0.13 ATU 58.5095023 60.2395584 1.73005617 . . ATE 0.407801035 . . Note: Standard error does not take into account that the propensity score is estimated. Source: Author’s Computation from Survey Data, 2016 Results in Table 6 further shows that the means of the treated group (farmers selling to commercial buyers) and the control group (farmers not selling to commercial buyers) are different. The magnitudes of the coefficients of the treatment effects indicate that the average treatment effects for the treated (ATT) are higher than the average treatment effects for the entire sample (ATE) and the average treatment effects for the untreated (ATU) except for technology outcome variable. The matching estimates generally indicated that selling to commercial milk buyers exerted positive, significant and unbiased impacts on household gross income and hence net income (Table 6). These ATT effects demonstrate that dairy farmer households who sell milk through these commercial milk buyers increase their gross income thereby improving their welfare over and above those that are less motivated to commercial milk buying channels. After matching, dairy farmer households selling to commercial milk buyers raised their daily net income in the household by 18.00 Kenya shillings on average per dairy cow. According to Owusu and A. (2009) households that have a higher probability of participating in non-farm work are able to obtain higher incomes and improve their food security status over and above those that are less inclined to participate in non-farm work. This is in convergence with the current study findings. The matching estimates of the ATU effect indicates that if farmer households who did not sell their milk to commercial milk buyers had actually sold the milk (counterfactual condition), then their household income and employment hours would be on average higher than that of those who did sell their milk to commercial milk buyers.Owusu and A. (2009) further notes that the implication of such an outcome is that income and food security gains from participation in non-farm employment are slightly higher for households with a higher probability of participating than households with slightly lower chances of participating in non-farm employment. Results further revealed that the matching estimates for gross income per animal on the treatment group, while balancing for the original level of gross income before and after treatment, also increased by 18.24 Kenya shillings (the nearest neighbour estimate of the average gain) after matching. Similarly, employment hours increased on average by 4 hour 46 minutes for dairy farmer households in the treatment group (those selling to commercial buyers) after matching results of the unobserved. This confirm earlier findings by Heshmati (2007) that the underlying technologies employed defines a production function to estimate the mean output rather than the maximum output. After matching, the percentage of cows bred using modern breeding technologies for example AI or sexed semen decreased for the dairy farmer households in the treatment group (dairy farmer households selling milk to commercial buyers) by about 69 percent. This can be attributed to the high cost of dairy cow breeding in the study area that has been brought about by asymmetric information flow. Of particular concern have been breeding (A.I) prices that do not fully reflect quality because dairy farmers and breeders do not have the same information. This result is in convergence and in conformity with earlier studies by Eggertson (1990) who argues that before making a decision about how to market a product and to whom to sell it, producers must determine the price that they expect to receive. Further, Eggertson (1990) argues that transaction costs arise when market information is asymmetric as this induces activities such as information searchers, bargaining, market contracting, monitoring, enforcement and protection of property rights, which are, by nature costly. The magnitude of the coefficients of the treatment effects indicated that the average treatment effects for the treated (ATT) were higher than the average treatment effects for the entire sample (ATE) and the average treatment effects for the untreated (ATU) for the outcome variables except for technology (Table 6). These results indicated that farmers who sold milk to commercial buyers had a higher probability of obtaining higher gross income per animal and had a higher probability of improving their welfare over and above those farmers who did not. The average effect of the treatment (ATE) for a dairy farmer household drawn from the overall population at random was Kenya shillings 16.00 higher because of selling to commercial milk buyers. This was because a positive effect was estimated for the dairy farmer households not selling to commercial buyers (ATU). The ATU effect by caliper matching estimates indicated that if farmer households that did not sell milk to commercial milk buyers had actually sold the milk, a counterfactual condition, then their household income and employment hours would be on average lower than that of those who did sell their milk to commercial milk buyers. The implication here is that the net income and employment hour gains from selling to commercial milk buyers are slightly higher for dairy farmer Asian Journal of Economics and Empirical Research, 2016, 3(2): 145-155 153 households with a higher probability of selling to commercial milk buyers than to dairy farmer households with slightly lower chances of selling to commercial milk marketing channel. In performing PSM using the common option imposes a common support by dropping treatment observations whose pscore is higher than the maximum or less than the minimum pscore of the controls (Table 7). According to the results, five observations were dropped from the entire observations. Three from the treated group (farmers selling to commercial buyers) and two from the untreated group (farmers not selling to commercial buyers), respectively. Table-7. Becker and Ichino (Psmatch2) PSM Estimation psmatch2 psmatch2 Treatment assignment Common support Off support On support Total Untreated 2 194 196 Treated 3 233 236 Total 5 427 432 Source: Author’s Computation from Survey Data, 2016 4.2. Matching Success for Impact/Outcome Factors Table 8 give result of t-test on the hypothesis that the mean value of each of the outcome variables; namely age, education level, distance to the milk market, milk market participation, price risk, livestock feeds, farming experience, total farm size, household size, gross and net household income, employment hours and breed technology adoption was the same for the treatment group (farmers selling to commercial buyers) and non-treatment group (farmers not selling to commercial buyers). It was done both before and after matching. Pstest was used to check for the success of the matching for the outcome variables. Further, a bias before and after matching was calculated for each of the variables and the change in the bias stated. Table-8. Indicators of Matching Quality before and after Matching (PSM Results) Major farm gate milk buyers Commercial or Final Unmatched Mean %reduction t-test V(T)/V(C) Outcome Variables Matched Treated Control %bias |bias| T p>|t| Age of household head U 48.75 47.995 6.5 0.67 0.504 0.95 M 48.639 51.12 -21.2 -228.5 -2.25 0.025* 0.89 Education level U 3.0932 3.0561 3.5 0.36 0.717 0.91 M 3.0944 3.0215 6.9 -96.7 0.74 0.460 0.88 Distance to Milk Market U 3.0275 3.4564 -10.9 -1.13 0.259 1.08 M 3.0399 3.378 -8.6 21.2 -0.93 0.352 1.10 Milk marketing participation U 0.4661 0.3520 23.3 2.41 0.017 1.09 M 0.4592 0.4506 1.8 92.5 0.19 0.853 1.00 Price risk U 0.8517 0.8725 -6.0 -0.62 0.536 1.13 M 0.8498 0.8584 -2.5 58.6 -0.26 0.794 1.05 Veterinary feeds U 0.6568 0.6327 5.0 0.52 0.603 0.97 M 0.6567 0.7082 -10.7 -113.5 -1.19 0.233 1.09 Farming experience U 19.127 18.041 10.3 1.06 0.290 1.09 M 19.03 21.485 -23.2 -126.0 -2.45 0.015* 0.99 Total farm size U 5.4667 4.4128 15.8 1.61 0.109 1.90* M 4.885 4.4895 5.9 62.5 0.90 0.369 2.64* Total household size U 6.4025 5.9745 18.8 1.95 0.052 0.88 M 6.382 6.4034 -0.9 95.0 -0.10 0.923 0.74* Gross Income U 172.25 148.97 18.5 1.89 0.059 1.66* M 172.45 148.45 19.1 -3.1 2.09 0.037* 1.87* Net Income U 206.62 190.09 12.5 1.28 0.200 1.56* M 207.01 182.98 18.2 -45.4 1.99 0.048* 1.68* Employment hours U 10.829 9.7589 4.3 0.43 0.667 3.81* M 10.925 6.7606 16.6 -289.3 1.83 0.048* 5.00* Technology (Breeding) U 55.097 58.04 -18.1 -0.84 0.401 0.89 M 55.341 53.947 3.8 52.6 0.42 0.672 0.98 Source: Author’s Computation from Survey Data, 2016 The bias before and after matching was calculated for each variable. This bias was the difference between the mean values of the treatment group and the control group, divided by the square root of the average sample variance in the treatment group and the not matched control group. Table 8 shows the differences in the values of the exogenous variables between the two groups before and after matching. For example, 46.61% and 35.2% of the treatment and control group respectively participated in milk marketing. This means that the results have significant influence on the treatment probability. The indicators of matching quality presented in Table 9 show substantial reduction in absolute bias for all the outcome variables for both the treated and non-treated. As indicated in the table (column 5), the mean bias in the covariates Z after matching lies below the 20 % level of bias reduction suggested by Rosenbaum and Donald (1985). This indicates that the covariates were significantly balanced as a result of the propensity score procedure. Asian Journal of Economics and Empirical Research, 2016, 3(2): 145-155 154 Table-9. Indicators of Matching Quality before Matching and After Matching Sample Ps R 2 LR Chi 2 p>chi 2 Mean Bias Med Bias B R % Var Unmatched 0.044 25.97 0.017 11.0 10.3 49.3* 1.01 31 Matched 0.032 20.77 0.078 10.7 8.6 40.5* 1.86 38 * If variance ratio outside [0.77; 1.29] for Unmatched and [0.77; 1.29] for Matched Ps R 2 – Pseudo R 2 , LR Chi 2 – Chi square likelihood ratio Note:* p-value of Likelihood Ratio Test (Pr > 2 ) Note: Pseudo-R 2 from probit estimation indicates the goodness of fit or how well the regressors explain the probability to participate in an employment activity. Source: Author’s Computation from Survey Data, 2016 The pseudo-R 2 from the propensity score estimation and from re-estimation of the propensity score matching on the matched samples for both the groups was 0.044 and 0.032, respectively (table 3.5). However, if p>0.05, the null hypothesis could not be rejected on the 5% significance level. Therefore, the joint significance of the regressors on the treatment status could not be rejected after matching. It was, however, not rejected before matching either. The null hypothesis that the mean values of the two groups do not differ after matching cannot be rejected for the variables except for age, farming experience, and gross income. Therefore, it is possible to generate a control group, which is similar enough to the treatment group to be used for the ATT estimation. The relatively low pseudo- R 2 after matching and the p-values of the likelihood-ratio test of the joint significance of the regressors imply that there is no systematic difference in the distribution of covariates between treatment and non-treatment groups after matching, suggesting that the overall results from the matching procedure are satisfactory in balancing the covariates between the treatment and non-treatment (Caliendo et al., 2005). 4.3. Distribution of the Propensity Scores for Treated and Untreated Groups Figure 1 gives the histograms of estimated propensity scores for the treated and non-treated groups of dairy farmer households. It shows the distribution and overlap (regions of common support) conditions of the propensity scores after matching for the two groups of dairy farmers (nearest neighbours).Visual inspection of the results reveals that the densities of the mean propensity scores are similar after matching and there is a clear overlap of the distributions. Propensity scores before and after matching as shown clearly indicate that estimating the p-score appears to balance the treated and untreated groups extremely well than without the p-score, a result which underscored the significance of the propensity score matching approach for this study. There is a clear balance between the two coordinates as they moved towards a central and a common area. Figure-1. Density Distribution (Covariate Balance) of the Estimated Propensity Scores Source: Author’s Computation from Survey Data, 2016 4.4. Summary and Conclusions This study examined the impacts of milk marketing channel choice decisions on dairy farmer household income, employment and breed technologies, using a cross-sectional sample of 432 dairy farmer households from six sub counties in Kericho County, Kenya. A propensity score matching model was employed to account for selection bias Asian Journal of Economics and Empirical Research, 2016, 3(2): 145-155 155 that normally occurs when unobservable factors influence both participation and non-participation in dairy milk marketing on the outcomes. The paper addressed dairy farmer heterogeneity by explicitly providing separate estimates for treated group (farmers selling to commercial buyers) and the untreated group (farmers not selling to commercial buyers). Results of the propensity score matching showed that the magnitudes of the coefficients of the average treatment effects for the treated (ATT) were higher than the average treatment effects for the entire sample (ATE), and the average treatment effects for the untreated (ATU) for the outcome variables except for breeding technology. The results indicated that farmers that sold milk to commercial milk buyers had a higher probability of obtaining higher gross income per animal per day and could improve their welfare over and above those farmers who did not. The average effect of the treatment (ATE) for a dairy farmer household drawn from the overall population at random was Kenya shillings 25.60 higher because of selling to commercial milk buyers. The ATU effect by caliper matching estimates revealed that if a farmer household who did not sell milk to commercial milk buyers had actually sold the milk, their household income and employment hours would be on average lower than that of those who did sell their milk to commercial milk buyers. The implication here is that the net income and employment hour gains from selling to commercial milk buyers are slightly higher for dairy farmer households with a higher probability of selling to commercial milk buyers than to dairy farmer households with slightly lower chances of selling to commercial milk marketing channel. The growing interest of policy makers in promoting dairy milk production, particularly in the rural areas of Kenya is in the right direction as revealed by the results of this study. Besides being a valuable source of income for rural dairy farmer households in developing countries, dairy production also helps in smoothing household incomes, which in turn smoothens consumption hence improving farmer household welfare over long periods of time. References Anderson, D.R., J.D. Sweeny, T.A. Williams, J. Freeman and E. Shoesmith, 2007. Statistics for business and economics. 9th Edn.: Thomson Learning. Thomson Press (India) Ltd. Barrett, C.B., T. Reardon and P. Webb, 2001. Nonfarm income diversification and household livelihood strategies in rural Africa: Concepts, dynamics, and policy implications. Food Policy, 26(4): 315–331. Caliendo, M., M. Caliendo and S. Kopeinig, 2005. Some practical guidance for the implementation of propensity score matching. (1588). Cocks, J., H.R. Gow and R. Westgren, 2005. Public facilitation of small farmer access to international food marketing channels: An empirical analysis of the USDA market assistance program in Armenia., AAEA Annual Meetings 2005. Providence, Rhode Island. Eggertson, T., 1990. Economic behavior and institutions: Principles of neoinstitutional economics. Glover, D.J., 1987. Increasing the benefits to smallholders from contract farming: Problems for farmers’ organizations and po licy makers. In World Development, 15(4): 441-448. Gorton, M., M. Dumitrashko and J. White, 2006. Overcoming supply chain failure in the agri-food sector: A case study from Moldova. Food Policy, 31(1): 90– 103. Gorton , W.J., 2007. Transformation and contracting in the supply chains of the former Soviet Union. In: Swinnen, J.F.M. (Ed.), Global Supply Chains Standards and The Poor: How the Globalisation of Food System and the Standards Affects Rural Development and Poverty. CABI, Wal. J.F.M. (Ed.). Belgium: CABI. pp: 175–187. Gow, H.R. and J.F.M. Swinnen, 2001. Private enforcement capital and contract enforcement in transitional economies. American Journal on Agricultural Economics, 83(3): 686–690. Greene, W.H., 2003. Econometric analysis. 5th Edn., Upper Saddle River, NJ: Prentice Hall, Upper Saddle River, NJ. Gujarati, 2004. Basic econometrics. 4th Edn., New York: Tata McGraw Hill Publisher. Heckman, J.J., 1999. Policies to foster human capital. Research in Economics, 54(1): 3-56. Heshmati, Y.-B.S.Y.-R.K., 2007. Commercialization and transfer of technology: Major country case studies Y.-R. K. Almas Heshmati, Young- Bock Sohn, ed., New York: Nova Science Publishers, Incl. Hujer, R., M.C. and S.L.T., 2004. New evidence on the effects of job creation schemes in Germany - A matching approach with threefold heterogeneity. Research in Economics, 58(4): 257–302. Kousar, R. and A. Abdulai, 2015. Gender based labor activity of farm households. International Journal of Business and Management Study – IJBMS, 2(2): 183–187. Nkosi, S.A. and J. Kirsten, 1994. Small holder’s market participation: Concepts and evidence from Eastern and Southern Africa. A Paper Presented at FAO Workshop on Staple Food and Market Policy Options for Promoting Development in Eastern and Southern Africa, Rome, March 1-2, 1994. Norris, P.E. and S.B. Sandra, 1987. Virginia farmers' soil conservation decisions: An application of tobit analysis. Southern Journal of Agricultural Economics, 19(1): 79-90. Owusu, V. and A. A., 2009. Non-farm work and food security among farm households in Northern Ghana. Food Policy, 36(2): 0–37. Owusu, V., F.K. Frimpong and V. Owusu, 2014. Impact of cocoa agroforests on yield and household income: Evidence from Ghana, Paris, France. 88th Annual Conference of the Agricultural Economics Society, Agro Paris Tech, Paris, France. Paterson, A.G., 1997. Pastoralists are people: Why development for pastoralists not the development of pastoralists? The Rural Extension Bulletin, No. 4. Pryanishnikov, I. and Z. Katarina, 2003. Multinomial logit models for the Austrian labor market. Austrian Journal of Statistics, 32(4): 267–282. Rosenbaum, P.R. and B.R. Donald, 1985. Constructing a control group using multivariate matched sampling methods that incorporate the propensity score. American Statistican, 39(1): 33–38. Sadler, M., 2006. Comparative analysis of cotton supply chains in Central Asia. In: Swinnen, J.F.M. (Ed.), Case Studies on Vertical Co- ordination in Agrofood Chains in Europe and Central Asia, ECSSD. Washington, DC: World Bank. Smith, J. and P. Todd, 2005. Does matching overcome Lalonde ’ s critique of nonexperimental by does matching overcome Lalonde ’ s critique of nonexperimental estimators ? 1. Journal of Econometrics, 125(1–2): 305–353. Striewe, L., 1999. Grain and oilseed marketing in Ukraine, In: Swinnen, J.F.M. (Ed.), Global Supply Chains Standards and The Poor: How the Globalisation of Food System and the Standards Affects Rural Development and Poverty. Iowa State University Ukraine Agricultural Policy Project (UAPP). Kiev: CABI, Wal. J.F.M. (Ed.). pp: 175–187. Varian, H.R., 2000. Intermediate microeconomic theory: A modern approach Eighth Edi. J. Repcheck, ed., Cheltenham, UK: Edward Elgar Publishing Limited, UK. Asian Online Journal Publishing Group is not responsible or answerable for any loss, damage or liability, etc. caused in relation to/arising out of the use of the content. Any queries should be directed to the corresponding author of the article.