The impact of SMS-based agricultural information on farmers: Insights from a randomized control trial
The impact of SMS-basedagricultural information onfarmers: Insights from arandomized control trialMarcel Fafchamps and Bart Minten
I. Introduction Mobile phones are rapidly spreading: global penetration rate at about 76 per 100 inhabitants in 2010 Poor developing countries increasingly part of this widespread use of mobile phones Rapid spread offers new possibilities for poor rural and agricultural households in developing countries: can overcome important barriers of physical distance and improve access to information
I. Introduction Mobile phones increasingly used in rural areas in Africa and Asia to disseminate daily prices of commodities, possibly leading to improved welfare:1. Better information would lead to better allocation of production factors;2. Information can improve bargaining position of small farmers and competition between traders;3. Farmers can use information to switch between end markets;4. Famers can use information to make choices around timing of marketing; Mobile phones should thus reduce erratic price variations as better arbitrage over time and space
I. Introduction Purpose of this study is to ascertain whether the distribution of agricultural information through mobile phones generates important economic effects in rural/agricultural areas Randomized control trial of a commercial service (Reuters Market Light - RML), the largest and best established private provider of agricultural price information via SMS Offered a one-year free subscription to RML to a random sample and analyze impacts on different indicators
I. Introduction RML one of many initiatives in this area; In India: e.g. IFFCO Kisan Sanhar, Qualcomm/Tata Teleservices; In Africa: e.g. Manobi, Eseko.
II. Description of the intervention Thomson-Reuters: transnational corporation specialized in information services, its core business Thomson-Reuters started up an innovative service to distribute agricultural information to farmers in India, called RML; Subscribers get SMS messages, in total 75 or 100 messages a month, at a price of $1.5/month; At the time of survey, they had 25,000 RML subscribers
II. Description of the intervention Content offering of RML:1. Market information: subscriber gets regular price updates for three markets of his choice for a chosen crop (crops covered at time of survey: soybean, cotton, pigeon pea, wheat, green gram, chick pea, maize, onion, pomegranates, grapes, oranges);2. Weather forecast: temperature, humidity, rainfall, and probability of rainfall at tehsil (district) level3. Crop advisory tips: given for the chosen crop4. Commodity news: news about agriculture and price trends of commodities at national/international level
III. Experimental design Randomized control trial, in close collaboration with Thomson Reuters Five crops selected in the state of Maharashtra (India): tomato (Pune); pomegranate (Nashik); onions (Ahmadnagar); wheat (Dhule); soybean (Latur). Assumption that information might have different value for different crops (e.g. perishable versus non-perishable) 100 villages selected: 20 villages in each district not previously targeted by RML; 10 farmers per village
III. Experimental design Treatment 1: All 10 farmers are offered the RML service Treatment 2: 3 farmers randomly selected (to test if treatment of some farmers benefit others as well) Control: No farmers treated Stratification improves efficiency in RCT (Bruhn/McKenzie): triplets of similar villages selected based on longitude, latitude, size population, distance to road, road availability index, distance markets, distance cell phone tower, number of extension visits
III. Experimental design Extension agents: randomly half of treatment 1 and 2 villages (to test if information given to them makes individual information unnecessary) No point of estimating effect on people who are unlikely to sign up or benefit from RML; Sample of selected farmers limited to:1. Farmers that grow selected crop for sale;2. Farmers with a cellphone (all 100 villages had cellphone coverage/electricity; half of farmers in state had cellphone).
III. Experimental design Each farmer interviewed twice:- Baseline survey: June-July 2009;- Impact survey: June-July 2010 Survey done after main harvest; survey covers one complete agricultural season and two cropping cycles (rabi/karif); Same survey instrument used in the two rounds; Questionnaire focus: agricultural information; use of sms information; price and time at which farmer sold crop; total revenue from sales of crop; crop yields; input use.
III. Experimental design Names and phone numbers collected for everybody in first round; those ones selected for treatment were sent to RML which then tried to hook up selected farmers; Farmers were free to refuse service and some did; To ascertain extent of contamination and non-compliance: additional phone interview after impact survey (to see if difference between target and actual user).
IV. Conceptual framework Two models in paper Channel 1: Timely access to market price information at the time of harvest helps farmers to arbitrage across markets; treated farmers should receive a better price than uninformed farmers; Channel 2: Farmers that are better informed about prices can obtain higher prices from traders because of better negotiation power (especially when selling at the farm-gate)
IV. Conceptual framework Other effects not modeled:1. Better knowledge of quality price differentials would induce farmers to upgrade quality of output;2. Weather information helps with farm operations (e.g. timing of use of pesticides);3. Crop advisory would lead to adoption of more appropriate technologies.
V. Testing strategy Intent-to-treat estimates (W=1 if offered free subscription) because of self-selection issues (Y is outcome variable): Yi = θ + βWi + ei IV-LATE (Instrumental variables – Local Average Treatment Effect) where S (=1 if farmer signed up for RML) is instrumented by W: Yi = θ + αSi + ei Heterogeneous effects (X is characteristics of farmer that are thought to affect the effect of treatment on the outcome variable, e.g. bigger farmers or less experienced ones):
V. Testing strategy Caveats:1. Value of information changes with circumstances; for information to be useful, it must be provided in a timely matter; information can be useful in one year but not in another;2. Information circulates through other channels than RML; assumption that circulation not so rapid that control villages benefit as well; however, cannot completely rule out spillovers across villages;3. Price information possibly leads to increased bargaining power; as traders and commission agents can not distinguish RML and non-RML users and if they change behavior, all farmers might benefit; in this case, we would underestimate benefits from RML
VI. The context and data Take-up of RML is one possible measure to measure benefits:Table 1. Number of agricultural holdings and RMLsubscribers in the five districts studied in MaharashtraDistrict: Crop followed Number of Number of RML subscribers** in survey agricultural holdings* 2007 2008 2009 2010Ahmadnagar onion 916,724 711 1,377 3,763 1,637Dhule wheat 230,216 108 1,296 1,028 840Latur soya 305,706 163 914 1,048 826Nashik pomegranate 591,763 2,176 1,561 3,934 6,514Pune tomato 667,365 392 653 3,495 781Total 2,711,774 3,550 5,801 13,268 10,598*: Government of India, Agricultural Census, 2000/01**: Thompson-Reuters
VI. The context and data Attrition:- 1000 farmers in baseline;- 933 in impact survey;- To test if there is anything specific about attrition, we run regression of attrition dummy on household characteristics;- Only onion dummy is significant dummy (more likely to drop out); is controlled for by triplet dummies.
VI. The context and data Compliance: Only 59% of farmers that were offered RML actually used it; Reasons:- Refusal (suspicion that they had to pay later; illiterate households);- Never activated RML (had to chose crops/markets they wished information on);- Changes in phone numbers/migration;- Chinese phones could not display the Marathi script. Reasons given an indication of some lack of interest in price information.
VI. The context and data Contamination: Low; 10 farmers or about 3.7% of the interviewed farmers in control group villages; Balancedness: Good balance on all variables included, area planted to target crop, marketing, transaction costs, past weather, and technological innovation; Follow Deaton (2009) to not include unnecessary control variables as to not artificially inflate t-values.
VII. Empirical analysis Intent-to-treat analysis. Average Treatment effect on the Treated (ATT) is calculated using nearest neighbor matching methodology (matching performed by triplet dummy); IV-LATE regression with instrumented actual RML use with random assignment to treatment; Standard errors clustered by village triplet; In presentation: results of whole sample (in paper also treatment 1 versus control only; heterogeneous effects).
Sources of information (baselinesurvey; % of farmers) Own observations/experimentation main source of information
RML Usage Treated farmers are significantly more likely to mention RML as a source of information (but much below 100%); IV-LATE effects > ATT.
Knowledge/information sharing Treated farmers appear more knowledgeable about crop prices Treated farmers share more information with other farmers Know price before sale: Share one one one at informationWhole sample day week month planting farmingNumber of observations 722 723 722 722 922Nearest neighbor matching (a)ATT Coeff 0.084 0.095 0.097 0.078 0.063 z-value 3.100 2.830 3.090 2.540 4.050Regression results (b)1. IV-LATETreatment Coeff 0.130 0.158 0.181 0.146 0.119 t-value 2.180 1.980 3.280 1.960 4.080Intercept Coeff 0.717 0.304 -0.034 0.010 0.676 t-value 64.780 20.640 -3.280 0.720 116.330
Realized prices Different crops different variation (CV wheat=0.07; CV soya=014; CV tomato=0.22; CV onions=0.44; CV pomegranate=0.45); RML expected to be more beneficial for perishable crops; Dependent variable: log(unit price in sales transactions over last 12 months of the target crop) Different methods: (a) ATT comparing treated/control; (b) ATT using dif-in-dif; (c) IV- LATE; (d) OLS long model to control for omitted variables; (e) Heterogeneous effects
Realized prices ATT (a) ATT (b) IV-LATE OLS Heterogeneous effects(d) long-model(c) OLS IVFor whole sample(e) No obs. 1480 688 1480 1425 1464 1457Treatment Coeff -0.031 -0.043 -0.062 -0.028 -0.034 -0.026 t-value -2.000 -0.520 -1.670 -1.510 -1.860 -1.430Intercept Coeff 2.260 2.159 2.248 2.249 t-value 309.620 21.250 93.64 99.080Dummy young head of hh Coeff 0.021 -0.013 -0.013 t-value 0.990 -0.500 -0.530Total crop area cultivated Coeff 0.005 0.001 0.001 t-value 5.720 1.900 1.550Dummy if sold to a trader Coeff -0.011 -0.006 -0.008 t-value -0.250 -0.190 -0.290Treatment extension agent Coeff -0.013 t-value -0.500Interaction with treatmentDummy young head of hh Coeff 0.057 0.059 t-value 1.750 1.850Total crop area cultivated Coeff -0.001 -0.000 t-value -0.590 -0.240Dummy if sold to a trader Coeff 0.085 0.091 t-value 1.750 1.830
Realized prices Overall no significant effect on prices received for treated households; Farmers that grow more get better prices; Younger treated farmers obtain a significantly higher price (6%); Treated farmers that sell to traders get significantly higher prices (9% higher); Results consistent when we do regressions at the crop level; also no effect of treatment on prices; Treatment does not decrease price variance in treated villages (in contrast with e.g. Aker, 2008).
Profitability Four measures:1/ Transactions costs: transportation; loading/unloading; rates paid to commission agents2/ Net price: Gross price minus transaction costs3/ Total sales revenues of the crop: quantity sold times prices4/ Value added: total sales minus monetary cost of inputs
Profitability No effect of RML on profitability measures Transaction Net Sale Value cost price revenues addedFor whole sampleNumber of observations 713 713 713 713Nearest neighbor matching (a)ATT Coeff 0.078 -0.760 48,247 46,352 z-value 1.420 -1.480 0.580 0.580Regression results (b)IV-LATETreatment Coeff 0.146 -1.450 87,074 84,530 t-value 1.050 -1.730 0.880 0.910Intercept Coeff 1.576 8.906 66,545 59,235 t-value 59.060 55.350 3.500 3.320
Spatial arbitrage and marketchanges Most sales take place at the market (except for pomegranate, farm-sales account only for 2% of sales) Market diversification varies from crop to crop; sales of perishable crops geographically concentrated (98% and 81% of tomato and pomegranate sold at one market) Overlap index of markets:1=same location for the two years (100% overlap);0=nothing sold at the same location between the two years (0% overlap).
Spatial arbitrage and market changes Treated households do more spatial arbitrage and change markets Number of markets Overlap added dropped index (c)For whole sampleNumber of observations 691 691 691Nearest neighbor matching (a)ATT Coeff 0.099 0.087 -0.095 z-value 2.980 2.680 -3.030Regression results (b)IV-LATETreatment Coeff 0.208 0.194 -0.197 t-value 2.120 2.080 -2.090Intercept Coeff 0.575 0.463 0.493 t-value 30.430 25.850 27.290
Other marketing characteristics Small effects on other marketing indicators Sold if whole- Sold Sold Crop was in whole- sale through to graded/ sale market, a trader sorted market chosen commis- before because sion sale closest agentFor whole sampleNumber of observations 1477 1352 1482 1470 1478Nearest neighbor matchingATT Coeff 0.030 -0.078 0.006 0.046 0.033 z-value 2.540 -3.220 0.230 1.740 2.260IV-LATETreatment Coeff 0.063 -0.131 0.539 0.084 0.055 t-value 1.750 -0.940 0.844 1.050 1.120Intercept Coeff 0.923 0.199 0.933 0.450 0.925 t-value 132.800 6.940 89.080 28.350 98.140
Effect on marketing RML helped farmers realize that they could obtain a higher price at a more distant market and some tried it out; However, treated farmers are not emboldened to sell more at farmgate or at local markets; Treated farmers pay more attention to grading and sorting.
Weather information No effect of weather information on treated farmers Did not Because of storm, heavy rainfall, incur output output output storm/ loss loss increaseFor whole sample heavy rain before harvest at harvestNumber of observations 915 529 529 529Nearest neighbor matchingATT Coeff -0.031 0.029 -0.042 0.046 z-value -1.000 0.720 -1.240 1.620IV-LATETreatment Coeff -0.108 -0.034 -0.025 0.095 t-value -2.080 -0.450 -0.310 1.390Intercept Coeff 0.288 0.690 0.369 0.024 t-value 27.740 39.630 19.860 1.540
Crop varieties grown andcultivation practices 31% of the farmers changed varieties between the two survey years; Of those who changed variety, 65% stated that they did so to improve profitability; 22% of farmers stated that they changed cultivation practices.
Crop varieties grown and cultivation practices Influence of RML on changes, but no differences in practices for treated and control Change If yes, If profita- Change If of crop reason bility, in change, variety is because cultivation because since profita- of practices ofFor whole sample last year bility RML last year RMLNumber of observations 895 240 156 911 203Nearest neighbor matchingATT Coeff 0.029 0.020 0.155 -0.027 0.211 z-value 1.100 0.460 2.830 -1.110 3.990Regression resultsIV-LATETreatment Coeff 0.043 0.006 0.200 -0.045 0.410 t-value 0.970 0.090 2.060 -1.240 5.170Intercept Coeff 0.525 0.374 -0.033 0.476 -0.016 t-value 59.350 30.950 -2.060 65.780 -0.950
VIII. ConclusionsRCT on effect of better information. Threepossible effects:1. The price information is expected to improvefarmers ability to negotiate with buyers and toenable them to arbitrage better across space;2. Weather information is expected to helpfarmers reduce crop losses due to extremeweather events such as storms;3. Crop advisory information is expected toinduce farmers to adopt new crop varieties andimprove their cultivation practices.
VIII. Conclusions-Treated farmers make use of theinformation that RML provides. However,no statistically significant average effect oftreatment on:1. the price received by farmers,2. crop losses resulting from rainstorms,3. the likelihood of changing crop varietiesand cultivation practices.- Disappointing results but in line with thetake-up rate of RML services
VIII. Conclusions1. Some effect of treatment:- Less likely to sell at the farmgate and more likely to sell at different, more distant markets;- Small increase in the likelihood of grading/sorting;2. Arbitrage does not results in higher prices, possibly because of two reasons:- Very few farmers sold at farmgate- Limited range of alternative market destinations (spatial concentrations of crop sales)
VIII. Conclusions Effect of food inflation during the study period:- Would expect information to be valuable in high-inflationary environments- Conceivable that magnitude of the change blunts our capacity to identify a significant price difference, given how variable prices are; One year experiment; maybe more time needed.
VIII. Conclusions Possible external validity considerations:1. Price information could help if spatial arbitrage across agricultural markets does not hold (segmented; thin; disorganized);2. Price information could help for those farmers that sell at the farmgate (little the case in our survey);3. Effect on crop quality obtained if price information detailed by variety and grade and farmers are provided with complementary information on how to produce this quality products. Need for similar experiments in other settings