IFPRI Discussion Paper 02918 
September 2023 
The (Perceived) Quality of Agricultural Technology and Its Adoption 
Experimental Evidence from Uganda 
Caroline Miehe 
Robert Sparrow 
David Spielman 
Bjorn Van Campenhout 
Innovation Policy and Scaling Unit
INTERNATIONAL FOOD POLICY RESEARCH INSTITUTE 
The International Food Policy Research Institute (IFPRI), a CGIAR Research Center established in 1975, 
provides research-based policy solutions to sustainably reduce poverty and end hunger and malnutrition. 
IFPRI’s strategic research aims to foster a climate-resilient and sustainable food supply; promote healthy 
diets and nutrition for all; build inclusive and efficient markets, trade systems, and food industries; 
transform agricultural and rural economies; and strengthen institutions and governance. Gender is 
integrated in all the Institute’s work. Partnerships, communications, capacity strengthening, and data and 
knowledge management are essential components to translate IFPRI’s research from action to impact. 
The Institute’s regional and country programs play a critical role in responding to demand for food policy 
research and in delivering holistic support for country-led development. IFPRI collaborates with partners 
around the world.  
AUTHORS 
Caroline Miehe (caroline.miehe@kuleuven.be) is a Researcher at LICOS, KU Leuven, Leuven, Belgium. 
Robert Sparrow (robert.sparrow@wur.nl) is an Associate Professor at the Development Economics 
Group, Wageningen University and Research, Wageningen, the Netherlands and the International 
Institute of Social Studies, Erasmus University Rotterdam, the Netherlands. 
David Spielman (d.spielman@cgiar.org) is Director of the International Food Policy Research Institute’s 
Innovation Policy and Scaling Unit, Washington, DC.  
Bjorn Van Campenhout (b.vancampenhout@cgiar.org) is a Research Fellow in IFPRI’s Innovation Policy 
and Scaling Unit, Leuven, Belgium and an Associate Researcher at LICOS, KU Leuven, Belgium. 
Notices 
1 Author order is alphabetical. The experiment on which this paper is based is registered in the American Economic Association's 
registry for randomized controlled trials as AEARCTR-0006361. This project received clearance from Makerere University's School 
of Social Sciences Research Ethics Committee (08.20.436/PR1) as well as from the International Food Policy Research Institute's 
Institutional Review Board (DSGD-20-0829). The research was also registered at the Ugandan National Commission for Science 
and Technology (SS603ES). 
2 IFPRI Discussion Papers contain preliminary material and research results and are circulated in order to stimulate discussion and 
critical comment. They have not been subject to a formal external review via IFPRI’s Publications Review Committee. Any opinions 
stated herein are those of the author(s) and are not necessarily representative of or endorsed by IFPRI.  
3 The boundaries and names shown and the designations used on the map(s) herein do not imply official endorsement or 
acceptance by the International Food Policy Research Institute (IFPRI) or its partners and contributors. 
4 Copyright remains with the authors. The authors are free to proceed, without further IFPRI permission, to publish this paper, or any 
revised version of it, in outlets such as journals, books, and other publications.
Abstract
Recently, issues related to the (perceived) quality of inputs and technologies
have been proposed as an important constraint to their adoption by smallholder
farmers in low income countries. Taking maize seed embodying genetic gain as a
case, we train random agro-dealers to test whether under-adoption by farmers is
caused by low quality due to sellers' lack of knowledge about proper storage and
handling. In a second hypothesis, we randomly introduce an information clear-
inghouse similar to popular crowd-sourced review platforms such as yelp.com or
trustpilot.com to test whether information asymmetries crowd out quality seed.
We nd that the information clearinghouse treatment improves outcomes for both
agro-dealers and farmers, with agro-dealers receiving more customers and report-
ing higher revenues from maize seed sales, and farmers reporting signicantly
higher use of improved maize seed varieties obtained from agro-dealers, leading
to higher maize productivity after two seasons. The primary mechanisms behind
this impact appear to be an increased eort to signal quality by agro-dealers and
a general restoration of trust in the market for improved seed. The agro-dealer
training does not have a clear impact on agro-dealers, nor on farmers in associ-
ated catchment areas. However, we do nd that the information clearinghouse
increases agro-dealer knowledge about proper seed storage and handling. Upon
exploring interaction eects between the training and the clearinghouse treatment,
we also nd that the training becomes eective for agro-dealers that are also in
the clearinghouse treatment group. This underscores the importance of incentives
to make supply side interventions such as trainings eective.
Keywords: agricultural technology adoption, agricultural input quality, agro-
dealers, knowledge, information asymmetries, perceptions, information clearing-
house
JEL Codes: D82, D83, O13, O33, Q12, Q16, C93
iii
Acknowledgments
We would like to thank Charles Marc Wanume, Leocardia Nabwire, Richard Ariong,
and Wilberforce Walukano for their support in the eld. We would also like to ex-
press our gratitude to Johan Swinnen, Eva-Marie Meemken, Francesco Cecchi, and
Martin Paul Jr. Tabe-Ojong for their helpful comments and suggestions. We fur-
ther thank participants of the Centre for the Study of African Economies Conference
2023, the Netherlands - Consultative Group for International Agricultural Research
(NL-CGIAR) Conference, the AgEconMeeting 2022, and seminars at Wageningen Uni-
versity and KU Leuven for valuable discussions. This project received funding from
the NL-CGIAR research programme on Seed Systems Development, grant number
W08.240.105, funded by the Netherlands Organisation for Scientic Research. The
research was also supported by the Fonds Wetenschappelijk Onderzoek - Vlaanderen
and the Fonds de la Recherche Scientique under EOS project G0G4318N. Additional
support was provided by the CGIAR Research Program on Policies, Institutions, and
Markets (PIM), the CGIAR Seed Equal Research Initiative, and the CGIAR Market
Intelligence Research Initiative which are funded by contributors to the CGIAR Fund
(https://www.cgiar.org/funders/).
iv
1 Introduction
The adoption of new agricultural inputs and technologies (such as seed embodying ge-
netic gain, inorganic fertilizers, or agro-chemicals) remains tepid in areas where they can
make the largest dierence in terms of food security, poverty reduction, and biodiver-
sity preservation (Suri and Udry, 2022; Gollin, Hansen, and Wingender, 2021; Borlaug,
2007). Several explanations for the low uptake of agricultural technology among small-
scale, resource-poor farmers in low- and middle-income countries have been explored
and tested with increasing depth and rigor in recent years. These include access to in-
formation about existence, use, and benets of the technology (Ashraf, Giné, and Kar-
lan, 2009; Van Campenhout, 2021), procrastination and time-inconsistent preferences
(Duo, Kremer, and Robinson, 2011), heterogeneity in the net benets derived from
the technology (Suri, 2011), missing markets for risk and credit Karlan et al. (2014),
and challenges related to learning about new technologies (Hanna, Mullainathan, and
Schwartzstein, 2014).
More recently, issues related to the quality of inputs and technologies have been
proposed as a key constraint to their adoption by smallholder farmers. Bold et al.
(2017) build on the observation that farmers generally cannot easily assess quality
from visual inspection at the time of purchase, so information asymmetries between
sellers and buyers characterize the market for agricultural inputs, in turn crowding out
the market for quality inputs in Uganda, similar to the lessons learned from Akerlof's
seminal Market for Lemons study (1970). However, subsequent research suggests
ambiguity in whether these quality issues are attributable to agro-dealers intentionally
adulterating their products, or whether they lack the requisite knowledge and skills to
preserve quality (Barriga and Fiala, 2020). Furthermore, it is not always clear whether
these quality issues are signicant: while some studies argue that input quality is indeed
lacking (Ashour et al., 2019), others argue that farmers may mistakenly perceive quality
deciencies even when the product meets the required standards (Michelson et al., 2021;
Wossen, Abay, and Abdoulaye, 2022).
We investigate some of these issues through a eld experiment conducted with both
agro-dealers and smallholder farmers in their catchment areas in the nascent market
for improved, high-yielding maize varieties in eastern Uganda.1 We focus partly on
agro-dealers because they are an essential marketing channel for inputs and technolo-
gies in countries with large smallholder farmer populations living in remote areas with
poor infrastructure. A reasonably dense network of semi-formal agro-dealers provides
access to seed, fertilizer, agro-chemicals, and tools, and may be combined to provide
agricultural advisory services and credit facilities, both formal or informal.
Yet small-scale agro-dealers may also be weak links in the supply chain for quality
inputs. For instance, the semi-formal nature and scale of agro-dealers may mean they
lack knowledge on proper handling and storage of agro-inputs. Barriga and Fiala (2020)
1In the context of this study, high-yielding maize varieties refers to both open-pollinated varieties
and hybrids. While the distinction may be lost on some, it is a nuance that is recognizable to many
colleagues working in the elds of plant breeding and seed sector development.
1
document various issues related to handling and storage that may reduce input quality
in the Ugandan seed supply chain. For example, agro-dealers often repack seed from
larger bags packed by seed companies into smaller bags in order to oer quantities
that are convenient and aordable to smallholders. As a result, important information
including variety name, expiry date, or planting instructions are lost, while the material
used for re-packagingair-tight polyethylene bagsaect aeration, moisture, and seed
viability. Other agro-dealers sell from open bags, which can similarly lower seed viability
(Bold et al., 2017). In a rst hypothesis, we thus posit that simply providing information
to agro-dealers will increase seed quality and subsequently improve farmers' product
experience, ultimately encouraging them to adopt improved maize varieties.
However, we recognize that providing information only to agro-dealers may not
necessarily remedy the fact that seed quality cannot easily be observed by farmers.
Furthermore, we expect that small scale agro-dealers are subjected to less regulation
and oversight than their larger counterparts in manufacturing, import, or wholesale
operations that are situated further up the supply chain. This means that agro-dealers
may still be incentivizedeven with trainingto underinvest in quality management
and preservation. In a context similar to ours, Homann et al. (2021) examine maize
(grain, not seed) in rural Kenyan markets and nd an absence of incentives for sellers to
address food safety problems because they are not observable to their buyers. The same
issue carries into seed markets, and in a worst-case scenario, agro-dealers may inten-
tionally sacrice quality to reduce costs and increase prots, e.g., by mixing improved
or fresh seed with local or old seed, or even with grain. There is some evidence that may
be consistent with this kind of adulteration and counterfeiting in the Ugandan agricul-
tural input supply chain. Bold et al. (2017) nd that hybrid maize seed contains less
than 50% authentic seeds and that 30% of nutrient is missing in fertilizer. Ashour et al.
(2019) nd that the average bottle of herbicide is missing 15% of the active ingredient
and nearly one in three bottles contains less than 75% of the ingredient advertised.
But even in cases where agro-dealers provide quality inputs, the fact that quality
can not be easily assessed by farmers at the time of purchase may be problematic
if farmers hold negatively skewed cognitive beliefs about seed quality sold by agro-
dealers.2 Michelson et al. (2021) show that the nutrient content of inorganic fertilizers
in Tanzania meets industry standards, but farmers nonetheless persist in their belief
that it is adulterated. Wossen, Abay, and Abdoulaye (2022) show that farmers in
Nigeria routinely misperceive the cassava variety they are cultivating, and that eorts
to address misperceptions could potentially improve farmers' investment choices and
productivity outcomes.
In a second hypothesis, we conjecture that making quality observable to both buyers
and sellers will increase adoption through various mechanisms. First, buyers may shift
from sellers that sell low quality inputs to sellers that provide high quality inputs.
Second, sellers may start to compete on quality, either by increasing quality if there is
2There are many reasons why these beliefs are likely to be biased downwards, including negativity
bias and loss aversion (Rozin and Royzman, 2001; Kahneman and Tversky, 1979).
2
still room for improvement, or by making the quality attribute of their products more
salient to buyers. Finally, in light of the new information, farmers may adjust their
perceptions of the quality of technologies.
The two hypotheses are tested in a randomized control trial (RCT) among 350 agro-
dealers and an associated 3,500 smallholder maize farmers in their catchment areas in
eastern Uganda over the course of two agricultural seasons. The rst hypothesis involves
a fairly standard intervention where we provide a one day training for agro-dealership
owners and managers on proper storage and handling of seed. For the second hypoth-
esis, we build on Hasanain, Khan, and Rezaee (2023) and implement a decentralized
information clearinghouse that is based on crowd-sourced information on quality pro-
vided by agro-dealers which is then aggregated and made public, much like yelp.com or
tripadvisor.com. In particular, we ask farmers to provide star ratings to agro-dealers
in their vicinity, and use these ratings to construct scores and rank agro-dealers. We
then disseminate these scores to farmers. We also provide the score, together with their
relative position in the rankings, to the agro-dealers.
We nd that the information clearinghouse improves outcomes for both agro-dealers
and farmers. Clearinghouse treated agro-dealers receive more customers and have
higher revenues from maize seed than control agro-dealers. Clearinghouse treated farm-
ers are signicantly more likely to use improved maize varieties from agro-dealers, and
have higher yields than control farmers after two seasons. Impact also seems to stem
from treated agro-dealers who increase their eorts and expand the services that they
provide to farmers. Treated agro-dealers are also more likely to be registered with the
Uganda National Agro-input Dealers Association (UNADA), perhaps to signal quality
now that it has been made salient to farmers. Finally, we nd that farmers in the treat-
ment group rate maize seed of agro-dealers in their neighborhood better, suggesting
that the clearinghouse treatment is also eective in changing perceptions.
The agro-dealer training does not have a clear impact on dealers, nor on farmers
in associated catchment areas. Interestingly, we do nd that the information clearing-
house increases agro-dealer knowledge about proper seed storage and handling. Upon
exploring interaction eects between the training and the clearinghouse treatment, we
also nd that the training becomes eective for agro-dealers that are also in the clear-
inghouse treatment group. This is consistent with Bold et al. (2022) who point out
the importance of simultaneously addressing demand-side constraints to make training
more eective.
Our study contributes to a large literature on the eectiveness of providing training
to small businesses in developing countries. Helping entrepreneurs to grow small rms
by teaching them business skills has yielded mixed results when subjected to rigorous
impact evaluation methods (eg. Karlan and Valdivia, 2011; Drexler, Fischer, and Schoar,
2014; Giné and Mansuri, 2021)}. While these studies often suer from methodological
issues such as low statistical power, it has also been argued that simply providing
knowledge may be insucient to move the needle (McKenzie and Woodru, 2013).
More promising results have emerged recently when the focus shifts from traditional
3
trainings to trainings designed to instill personal initiative(Campos et al., 2017).3 Our
study similarly shows the importance of (external) motivation in making trainings reach
their objective.
Our study also contributes to the literature that shows how providing product in-
formation to consumers can solve the lemons problem through a variety of economic
mechanisms. First, by enabling consumers to screen on quality, they can now shift to
better quality products. For instance, Lane, Schonholzer, and Kelley (2022) show that
commuters in Nairobi choose for the safe busses after information on safety records
of dierent busses was made publicly available. Second, public disclosure of product
information makes it possible for sellers to dierentiate on quality, and indeed, start
competing on it, in turn lifting the market out of the low quality equilibrium. For exam-
ple, Bennett and Yin (2019) show that entry of a chain store (with a solid reputation for
quality) leads to higher overall drug quality and lower prices in India. For the market
for antimalarial drugs in Uganda, Björkman Nyqvist, Svensson, and Yanagizawa-Drott
(2022) show that the presence of a non-governmental organization providing a superior
product led to a stark reduction in the share of rms selling fake drugs.
We also contribute to a growing literature on the importance of social comparison,
self-image, and social norms in determining behavior. Allcott and Rogers (2014) nd
that a social comparison-based intervention consisting of mailing reports of home energy
use to households reduced their energy consumption dramatically. Gosnell, List, and
Metcalfe (2020) report on an experiment with airline pilots where dierent strategies
to increase fuel eciency (including performance feedback and prosocial incentives) are
tested. We suspect that, in addition to the threat of farmers shifting to better rated
agro-dealers, psychological factors such as professional identity and a sense of social
obligation may even be more important drivers for agro-dealers to improve.
The article further ts into an emerging literature that tests how crowd-sourced
information can be used to to reduce information asymmetries. Even though advances
in Information and Communications Technology and the rise of e-commerce has led to
numerous platforms that allow for consumer feedback and a variety of websites that
aggregate crowd-sourced reviews, there is surprisingly little evidence on the eects of
these developments. The few rigorous studies that are available report impressive im-
pact. Reimers and Waldfogel (2021) compare the eects of professional critics and
Amazon star ratings of books on consumer welfare and nd the eect of star ratings
on consumer surplus to be more than ten times the eect of traditional (expert) review
outlets. In the context of smallholder agriculture, Hasanain, Khan, and Rezaee (2023)
implement a crowd-sourced information clearinghouse in the market for articial insem-
ination of livestock in Punjab, Pakistan, where individual signals of quality are noisy.
They nd that farmers who receive information enjoy 25% higher insemination success.
3Personal initiative is dened as a self-starting, future-oriented, and persistent proactive mindset.
4
2 Experimental design
We designed an experiment with two interventions (detailed in the next section) that
aim to induce quality improvements (or perceptions thereof) in the seed market, and
evaluate their impact on a set of outcomes related to market performance, technol-
ogy adoption, and productivity. The interventions are randomized at the agro-dealer
catchment area level. These catchment areas are clusters of towns, villages, markets,
trading centers, and other key market sheds where agricultural market activity tends
to operate, and are typically host to several agro-dealers. Clustering agro-dealers into
catchment areas is done on the basis of geographical location.4
We randomize at the level of the catchment area (instead of opting for randomiza-
tion at the less aggregate agro-dealer level) for three reasons. Firstly, randomizing at
the level of the individual agro-dealer prompted ethical concerns and was thus ruled out
a priori. Specically, in cases where two or more agro-dealers operate in very close prox-
imity to each other, treating only one of them may lead to a competitive (dis)advantage.
Randomizing at the catchment level substantially reduces the risk of (dis)advantaging
agro-dealers in this way. Secondly, catchment-level randomization reduces the likelihood
of spillovers from treated to control agro-dealers. Thirdly, catchment-level randomiza-
tion allows us to extend the evaluation to the measurement of treatment eects on
farmers (and not just agro-dealers) because all farmers in the catchment area are now
exposed to agro-dealers who all received the same treatment.
We used simulations to determine the sample sizes required to detect eects of the
treatment on selected outcomes at both farmer and agro-dealer levels.5 The simula-
tions show that if the number of catchment areas is larger than 112, our experiments
will return statistically signicant results 80% of the time on a selection of primary
outcomes. This corresponds to approximately 318 agro-dealers. Based on further sim-
ulations to study impacts at the farm-household level, we decide to collect information
on 10 farmers per agro-dealer, leading to a sample size of 3,180 households.6 The two
interventions are combined in a eld experiment that takes the form of a 22 factorial
design. The power simulations focused on the individual treatments, implying that we
are likely to be under-powered to estimate interaction eects between the interventions
(Lin, 2013; Muralidharan, Romero, and Wüthrich, 2019)
4We use a haversine function to construct an adjacency matrix based on GPS coordinates, and
agro-dealers that are less than 5 kilometer apart are assigned to the same catchment area. The 5
kilometer threshold was selected based on a visual inspection of a map, the size of an average village in
our sample and the reported distance between farmers and agro-dealers in survey data from a previous
study of the maize value chain that can be found here.
5Simulation provides a exible and intuitive way to analyze statistical power. Furthermore, instead
of relying on theoretical distributions for the outcome variables that make assumptions and return
analytic solutions, we run simulations that (re-)sample from real data that was collected in previous
surveys. In particular, we use data from 78 agro-dealers and 1,529 smallholder farmers in the catchment
areas of these agro-dealers that were collected in three districts in eastern Uganda in July 2019. The
data are publicly accessible here.
6More detailed information including the expected treatment eect sizes can be found in the pre-
analysis plan which was pre-registered at the AEA RCT registry under RCT ID 0006361.
5
3 Interventions
3.1 Agro-dealer training
Training content and material
To determine the content of the training and to ensure it is suciently anchored in
the study site and context, we consulted experts from several Ugandan organizations
using semi-structured interviews and a workshop to identify problems in the seed sec-
tor and in agro-dealer retailing practice, and to discuss eective and realistic solutions
and best practices to address seed storage and handling issues. We then developed a
training manual to ensure standardization and a simple but visually appealing poster
illustrating the most important practices. Participants in the interviews and workshops
included individuals from the Ministry of Agriculture, Animal Industries and Fisheries
(MAAIF), the National Agricultural Research Organization (NARO) the Uganda Seed
Trade Association (USTA), and the Uganda National Agro-Dealers Association (UN-
ADA). The roll-out of the experiment began in 2021 (Figure 1).
Training
In each treated catchment area, all agro-dealers were selected for a training, and for
each treated agro-dealer, both the owner and shop manager were invited. The owner
was invited because several recommended techniques and practices required that new
investments were made in the agro-dealership, while the shop manager was invited
because many of the recommendations are hands-on practices that would ultimately
fall under the manager's purview for day-to-day operations. Of 166 agro-dealers that
were invited, 140 sent at least one person, leading to a compliance of 84%.7
The trainings took place in May 2021, a time when agro-dealers are not too busy,
and early enough to ensure they could apply newly learned practices in the second
agricultural season. Trainings were held in locations that were easily reachable by the
agro-dealers. Trainings were organized in small groups, with an average of 10-15 agro-
dealers participating. To deal with COVID in a responsible manner, participants and
trainers were required to keep the proper distance, wear face masks, and frequently
disinfect their hands. All attendants were compensated for transport, and both lunch
and refreshments were provided. Participants were further incentivized to engage in
the training and pay close attention with an oer of a free portable seed moisture
meter, subject to passing a short content knowledge tests comprised of a short series of
multiple-choice questions at the end of the training.8 The agro-dealers were also given
a copy of an informational poster used in the training to remind them of best practices.
7For 80 agro-dealers, both owner and manager attended; for 50 agro-dealers, only the owner at-
tended; for 10 agro-dealers, only the manager attended.
8This was just to encourage agro-dealers to pay attention. Every agro-dealer that attended got a
moisture meter, regardless of how they scored on the test.
6
maize planting maize harvesting maize planting maize harvesting maize planting
Mar/Apr Jun/Jul Aug/Sep Nov/Dec Mar/Apr
2021 2022
Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr
CH rating collection 1 dealer training CH rating dissemination 1 CH rating collection 2 CH rating dissemination 2
April 2021 May 2021 SMS to farmers: July 2021 January 2022 visited farmers: Jan/Feb 2022
visited dealers: July 2021 visited dealers: Jan/Feb 2022
visited farmers: August 2021 SMS to farmers: March 2022
Figure 1: Timeline
In each training, the trainers explained correct handling and storage practices for
improved maize seed and used the poster and an example seed bag for illustration.
Participants then rehearsed the more challenging practices like measuring moisture
using a moisture meter. The trainings were organized and conducted in collaboration
with UNADA.
3.2 Information clearinghouse
Rating collection and computation
At the time of baseline data collection (April 2021), we asked sampled farmers to
rate all agro-dealers that operated in the catchment area on multiple characteristics.
Enumerators were guided by a tablet-based application that iterated through all agro-
dealers in the catchment area. For each agro-dealer, we provided the common names
that are used to refer to the agro-dealer, a description of where the store is located, and
a picture of the store front (obtained during the agro-dealer censussee Subsection
5.1). If farmers knew the agro-dealer, they were asked to provide ratings using the
questions listed in Table 1. For example, we asked farmers to rate the maize seed that
an agro-dealer sells on a scale of one to ve stars on seed germination. As can be seen
in Figure 1 , we implemented the clearinghouse in two consecutive seasons, so ratings
were collected again in January 2022. Ratings were always collected after harvest, when
smallholders were able to assess seed quality based on observing germination and yield,
the resistance against droughts, pests and diseases, and how quickly the crop matured
7
Table 1: Questions for farmers to rate agro-dealers
min max
Do you know this shop name or dealer name, no yes
sometimes called nickname, located in market name?
The place can be described as description.
Please rate this agro-dealer on:
Quality and authenticity of seed 1 star 5 stars
Please rate the maize seed that this agro-dealer sells on:
General quality 1 star 5 stars
Yield as advertised 1 star 5 stars
Drought tolerance as advertised 1 star 5 stars
Pest/disease tolerance as advertised 1 star 5 stars
Speed of maturing as advertised 1 star 5 stars
Germination 1 star 5 stars
(i.e., duration).
A potential concern arises from asking treated farmers to rate agro-dealers because
it may increase awareness among farmers of the existence of all agro-dealers in the
area, so that this awareness eect could confound the clearinghouse eect. To address
this concern, we also iterated through the agro-dealers in the catchment areas with
farmers in the the control group to make them similarly aware of the existence of agro-
dealers in their vicinity. However, control farmers were not asked to rate agro-dealers
as the process of rating an agro-dealer's seed could make quality more salient, which
we consider to be a key aspect of the treatment.
Based on the answers of all farmers about all agro-dealers in a catchment area,
we computed an average rating for each agro-dealer. These ratings were translated
into both words and star ratings to ensure that they were easily comprehensible to
farmers and agro-dealers when disseminated. See Appendix A.2 for details on the
rating computations..
Rating dissemination to farmers
Our ability to test the eectiveness of the clearinghouse treatment requires precise
timing for the dissemination of these agro-dealer ratings. Dissemination occurred before
farmers started buying seed for the next agricultural season, allowing treated farmers
to use the new rating information when choosing whether and where to purchase inputs
(Figure 1). Ratings were disseminated to farmers through short message service (SMS)
and in person, as detailed below.
Text messages Farmers were sent one text message per agro-dealer in their proximity
by SMS in one of three local languages---Lusoga, Lugwere, or Samia---chosen at the
8
sub-county level to increase specicity. For farmers in the treatment area, the message
read:
Hello from AgroAdvisor! Did you know that customers from [name of the
agro-dealer ] rate the quality of maize seed sold there as [okay/good/very
good/excellent ]?
To isolate the eect of the ratings from more general eects that may arise from sending
SMS messages, we also use a placebo for the control group that consisted of an "empty"
SMS that only pointed out the existence of the agro-dealers in the control farmer's
catchment area. This also makes it more dicult for farmers to identify if they are
being treated or not, thus reducing the likelihood of reactivity eects and experimenter
bias.
In person The enumerators also re-visited the farmers in our sample. For this pur-
pose, we designed a visually appealing tablet-based application that cycles through
all agro-dealers in the catchment area of each farmer and provides their ratings. The
application generates the following statement:
We wanted to let you know that customers from [name of the agro-dealer ]
rate the quality of maize seed sold there as [okay/good/very good/excellent ]!
The quality of the maize seed that this agro-dealer sells received a score of
[score] out of 5!
The application also displayed the stars associated with the score. Again, for control
group farmers, the application cycled through the agro-dealers in the control areas
without providing ratings to control for any eect that may arise from simply being
reminded of the existence of agro-dealers.
Rating dissemination to agro-dealers
Agro-dealers received their ratings by means of a report on laminated paper that was
delivered to their agro-dealerships. The front of the report is a visually appealing
certicate with a logo and the agro-dealer's general rating (Figure 2). We encouraged
agro-dealers to prominently display the ratings in the agro-dealership, similar to a
certicate of excellence from TripAdvisor or similar rating apps.
The back of the report provides additional information, including the individual
ratings that the seed sold by the agro-dealer received for overall quality, yield, drought
and disease resistance, speed of maturing, and germination, and the average (combined)
ratings of other agro-dealers in the same catchment area in a table, visualized by stars.
This shows agro-dealers their relative position in the area, potentially incentivizing the
agro-dealer to improve their ratings.
The entire process of collecting and disseminating ratings was done twice, the rst
time targeting the second agricultural season of 2021 and the second time targeting the
9
Kafuko Farm Supply
scores
4.0 out of 5
This was based on 18 reviews.
This score means that farmers in the area think that the
quality of the maize seed this shop sells is:
Excellent!
SeedAdvisor certificate 2021
Figure 2: SeedAdvisor certicate
rst agricultural season of 2022 (see Figure 1). Repeating the treatment was important
to capture dierent dynamics of particular impact channels on certain outcomes. For
instance, if seed quality is good but farmers hold pessimistic beliefs about the quality,
disseminating information may already result in increased adoption and yield eects
after a single season. However, if agro-dealers engage in counterfeiting, the threat of
farmers switching to more honest agro-dealers may lead them to improve quality, which
will only be reected in subsequent ratings. This in turn could increase adoption but the
eect of this on yields will only become apparent during harvest in the second season.
Repetition may also be important for the eectiveness of certain impact pathways. For
instance, agro-dealers may be more likely to change their behavior if they know that
they will be scored again in the near future.
4 Empirical strategy
We estimate intention-to-treat eects on outcomes at both the agro-dealer level and the
farmer level. To increase power, we condition the estimates on (mean-centered) baseline
values of the outcome variables. We estimate the following specication using Ordinary
Least-Squares to obtain the average treatment eect for agro-dealer level outcomes:
Yij = α + βTj + γ′Xij + δY0ij + εij (1)
where Yij is the outcome for agro-dealer i in catchment area j at mid- or endline,
Y0ij is the corresponding outcome at baseline, Tj is a dummy for the treatment status
10
of catchment area j, Xij is a vector of controls for the orthogonal treatments in the
factorial design (demeaned and interacted with the main treatment eect, see Lin, 2013;
Muralidharan, Romero, and Wüthrich, 2019), and εij an error term that is potentially
correlated within catchment areas. The coecient β is the estimated average treatment
eect. For farmer-level outcomes, a similar equation is estimated, where Yij is now the
outcome variable for farmer i in catchment area j at midline or endline, Y0ij is the
corresponding outcome at baseline, and all other terms are dened as in the agro-dealer
regression above.
Because we randomize at the catchment-area level, we use cluster-robust variance-
covariance matrices that cluster standard errors at this level. For outcomes at the
farmer level where we have almost 3,500 observations in 130 clusters, the original form
of the sandwich estimator that does not make any small-sample correction, is used.
For outcomes at the agro-dealer level where we have almost 350 observations in 130
clusters, we use the BellMcCarey adjustment (Imbens and Kolesár, 2016).
We also follow several pre-registered principles for variable construction. For con-
tinuous variables, trimmed values are used to reduce the inuence of outliers. In par-
ticular, we trim 1% of each side of the distribution for agro-dealer level outcomes and
2.5% of each side of the distribution for farmer level outcomes. Inverse hyperbolic sine
transformations are used if variables are skewed, with skewness being dened as the
adjusted Fisher-Pearson coecient of skewness exceeding 1.96. Outcomes for which
95% of observations have the same value within the relevant sample are omitted from
the analysis.
We account for multiple hypothesis testing by aggregating dierent outcomes within
a family into summary indices, following Anderson (2008).9 While these indices are use-
ful to answer the question of overall impact of the intervention on a family of outcomes,
it is not straightforward to interpret the eect size. Zooming in on individual outcomes
within each family, eect sizes become more meaningful and show which variables drive
the results. That is why we also report the treatment eects on individual variables,
though we advise care in interpretation.
5 Data
5.1 Sample
The agro-dealer sample was obtained by listing all agro-dealers in 11 districts in south-
eastern Uganda. After the census, which resulted in a sample of 348 eligible agro-
dealers, these agro-dealers were assigned to 130 catchment areas (for details, see Section
2, Footnote 4 in particular). This procedure led to an average of three agro-dealers per
catchment area, ranging from a minimum of 1 to a maximum of 18.
9Each index is computed as a weighted mean of the standardized values of the outcome variables.
The weights of this ecient generalized least squares estimator are calculated to maximize the amount
of information captured in the index by giving less weight to outcomes that are highly correlated with
each other.
11
Table 2: Factorial design
agro-dealer training
1 0
33 areas 32 areas
1 96 agro-dealers 97 agro-dealers
960 farmers 970 farmers
clearinghouse
33 areas 32 areas
0 70 agro-dealers 85 agro-dealers
700 farmers 850 farmers
To connect agro-dealers to customers, we asked agro-dealers for the names of the
villages where most of their customers come from. Then enumerators were instructed to
randomly sample ten households that grow maize in these villages. Consequently, about
3,500 smallholder maize farmers were sampled. Allocation of farmers, agro-dealers and
catchment areas to the dierent treatment cells of the factorial design is summarized
in Table 2.
Baseline data was collected from agro-dealers in September and October 2020 and
from farmers in April 2021. Midline data from both farmers and agro-dealers was
collected in January and February 2022, and endline data from farmers and agro-dealers
was collected in July and August 2022.
At the level of the agro-dealer, enumerators were instructed to interview the person
who is most knowledgeable about the day-to-day operations, which was usually the
shop manager. As part of this initial quality assessment process, a bag of maize seed
was also purchased at each agro-dealer, although only 232 of the 348 sampled agro-
dealers had seed in stock at the time of the baseline interview. Enumerators were also
instructed to note down a series of objectively veriable quality indicators related to
storage. Often, this also meant that enumerators inspected stores at dierent locations
that were separate from the sales outlet.
At the farm household level, enumerators were instructed to interview the person
most knowledgeable about maize farming. However, a set of questions deals with the
household head, who could be or could not be the respondent. In addition to general
questions about farming and input use, the farmer was asked to enumerate all maize
plots, from which a random plot was chosen, and detailed data was collected on that
plot.10
10This was mainly done to reduce on data collection costs, time, and burden. As plots were chosen
randomly, averages should be representative at the household level.
12
5.2 Descriptive statistics
This subsection describes the baseline sample. Information about the average agro-
dealer can be found in Table 3. The average respondent is 32 years old; 60% are male
and more than 90% nished primary education. In 55% of the cases, the respondent is
also the owner of the agro-dealership.
We see substantial heterogeneity among agro-dealers. Some are small informal stores
that are located in rural areas and sell maize seed and other agricultural inputs in
addition to consumer items to a small customer base and only during the planting
season. Others have many customers, are located in towns and only sell inputs and
equipment used in agricultural production. The average agro-dealership was established
5 years prior to the baseline survey date, is located 7 km from the nearest tarmac road,
and services an average of 41 customers per day. Among the sampled agro-dealers, 74%
only sell farm inputs and equipment. 60% reported that they provide credit and 46%
that they oer advisory services.
Information was also collected to provide an initial assessment of the quality of maize
seed sold at the sampled agro-dealers. This included specic questions on seed stor-
age and handling. Furthermore, with the shop manager's permission, enumerators
drawing on training provided to them at the outset of the studyinspected the area
where seed was stored and noted the conditions. We nd that 65% of agro-dealers had
problems with pests such as rats or insects, while 16% store maize seed in open con-
tainers, thus exposing the seed to a range of pests and contaminants. Not surprisingly
two thirds of the agro-dealers sampled reported that they had received at least one
complaint about seed they sold from a customer during the prior season.
Turning to the seed samples obtained from the agro-dealers, our measurements
of moisture content in the bag indicated an average of 13.6%, with a minimum of
10.3% and a maximum of 17.4%. On average, these moisture rates were above the
recommended rate of 13%, suggesting potential for the growth of molds and pests that
can negatively aect seed quality and performance. In terms of labeling for quality,
68% of the purchased seed bags contained a printed packaging date, only 18% had an
expiry date, and only 8% displayed a quality indication label issued by the National
Seed Certication Services (NSCS).11
Table 4 reports means in the farmer sample. The average household head in our
sample is 49 years old; 78% are male and 51% have nished primary education. The
average household size is 9 people, and the typical homestead is located 4 kilometers
from the nearest agro-dealer and 9 kilometers from the nearest tarmac road. The
average farmer has 23 years of experience with maize cultivation and cultivates 3 acres
of land for all crop including maize.
Half of the farmers in our sample planted improved maize seed on at least one of
their plots during last season, with 1 out of 3 farmers purchasing this seed at an agro-
11Typically, maize seed is certied by NSCS, a division of the Department of Crop Inspection and
Certication (DCIC) in MAAIF. Certied seed is indicated as such with a blue tag or sticker axed
to the package, and quality declared seed is indicated by a green tag or sticker.
13
dealer. Only 25% applied inorganic fertilizers such as di-ammonium phosphate (DAP)
or nitrogen, phosphorus, and potassium (NPK) on the randomly selected plot. Yields
on these plots is about 440 kilograms per acre.
5.3 Orthogonality tests of randomization balance
We include standard orthogonality tables with pre-registered variables for both agro-
dealers and farmers to test if treatment and control groups are comparable in terms of a
set of baseline characteristics (Tables 3 and 4 respectively). Some of these characteristics
are unlikely to be aected by the intervention, while others are drawn from the set of
outcome variables that will be used to measure the intervention's impact and explore
the underlying mechanisms at play in the next sections.
For outcomes at the agro-dealer level reported in Table 3, we nd that from a
total of 32 comparisons, only one is signicant at the 5% signicance level and two are
signicant at the 10% level. For outcomes at the farmer level reported in (Table 4, out
of 32 comparisons, one is signicant at the 10% level. In all, we conclude that these
results show reasonable balance at baseline.
5.4 Attrition
Table 5 reports attrition levels in the treatment and comparison groups. We failed to
collect data from 12% of agro-dealers and 2% of farmers at midline, and from 14% of
agro-dealers and 1% of farmers at endline. To test if non-response is related to one
of the treatments, we regress the likelihood of leaving the sample on the treatment
indicators. We nd that clearinghouse treated agro-dealers are signicantly less likely
to leave the sample.
The dierential attrition may be due to a larger share of control agro-dealers going
out of business, while the clearinghouse treatment cushioned some agro-dealers in the
treatment catchment areas against bankruptcy at a time when COVID-19 hit. As in
biomedical RCTs where dierential attrition rates may be due to excess mortality in
the control group, the attritors are likely the ones that would have beneted most from
the treatment. As such, the unadjusted selection-contaminated estimates provide lower
bounds for the true treatment eect (Angrist, Bettinger, and Kremer, 2006; Duo,
Glennerster, and Kremer, 2007).
6 Results
We now present impact of the two interventions on both agro-dealer level outcomes and
farmer level outcomes. We separately report eects one agricultural season after the
intervention (referred to as impact at midline) and two seasons after the intervention
(referred to as impact at endline).
14
Table 3: Descriptive statistics and orthogonality tests - Agro-dealer
mean training CH
Respondent's age in years 32.43 0.56 -2.24+
(11.49) (1.19) (1.21)
Respondent is male 0.59 0.02 -0.01
(0.49) (0.06) (0.06)
Respondent nished primary education 0.92 0.01 -0.01
(0.27) (0.03) (0.03)
Respondent owns shop 0.55 0.03 0.02
(0.50) (0.06) (0.06)
Respondent received training on maize seed handling 0.53 0.05 0.12+
(0.50) (0.07) (0.07)
Respondent knows how to store seed after repackaging 0.27 0.07 0.08
(0.44) (0.06) (0.06)
Agro-dealer's distance to nearest tarmac road in km 6.56 -0.92 -1.58
(10.39) (2.21) (2.24)
Agro-dealer only sells farm inputs 0.74 -0.09 0.03
(0.44) (0.07) (0.06)
Years since Agro-dealer establishment 5.34 -0.09 0.21
(6.30) (0.77) (0.78)
Number of customers per day 41.49 11.35 6.43
(46.49) (7.16) (6.72)
Quantity of maize seed sold in kg 695.50 201.06 176.31
(1497.18) (252.97) (235.92)
Amount of maize seed lost/wasted last season in kg 3.50 1.99 2.40
(18.65) (2.47) (2.30)
Agro-dealer has problem with pests 0.65 -0.01 -0.03
(0.48) (0.06) (0.06)
Agro-dealer stores maize seed in open containers 0.16 0.00 0.08
(0.36) (0.05) (0.05)
Agro-dealer received seed related complaint from customer 0.64 -0.11∗ 0.07
(0.48) (0.05) (0.05)
Moisture in bag of maize seed in % 13.56 0.25 -0.18
(1.44) (0.25) (0.26)
Note: Column (1) reports sample means at baseline and standard deviations below; columns (2)-(3) report dierences
between treatment and control groups and standard errors below; they are clustered at the level of randomization; **,
*, and + denote signicance at the 1, 5 and 10% levels.
15
Table 4: Descriptive statistics and orthogonality tests - Farmer
mean training CH
Household head's age in years 48.62 -0.08 -0.24
(13.38) (0.56) (0.56)
Household head is male 0.78 -0.02 0.03
(0.42) (0.03) (0.03)
Household head nished primary education 0.51 0.00 0.04
(0.50) (0.03) (0.03)
Homestead's distance to nearest tarmac road in km 9.39 0.33 -1.23
(10.81) (1.69) (1.71)
Homestead's distance to nearest agro-dealer in km 3.78 -0.11 0.11
(4.79) (0.37) (0.37)
Number of people in household (incl. respondent) 8.70 -0.16 -0.09
(3.98) (0.18) (0.18)
Number of rooms in house 3.49 -0.01 0.02
(1.45) (0.09) (0.09)
Farmer's land for crop production in acres 3.35 0.07 0.00
(4.32) (0.21) (0.22)
Years since farmer started growing maize 23.09 0.61 -0.55
(13.14) (0.55) (0.58)
Yield in kg/acre 443.01 27.15+ -6.14
(304.99) (13.71) (13.52)
Farmer used quality maize seed on any plot 0.49 0.02 0.01
(0.50) (0.02) (0.02)
Farmer bought this seed at agro-dealer 0.32 -0.01 0.01
(0.47) (0.02) (0.02)
Amount of this seed farmer bought at agro-dealer in kg 9.52 0.16 -0.34
(6.92) (0.53) (0.53)
Farmer thinks maize seed at agro-dealer is adulterated 0.68 0.01 0.00
(0.46) (0.03) (0.03)
Farmer used DAP/NPK 0.25 0.04 0.02
(0.43) (0.03) (0.04)
Farmer used organic manure 0.07 -0.01 0.01
(0.26) (0.01) (0.01)
Note: Column (1) reports sample means at baseline and standard deviations below; columns (2)-(3) report dierences
between treatment and control groups and standard errors below; they are clustered at the level of randomization; **,
*, and + denote signicance at the 1, 5 and 10% levels.
16
Table 5: Attrition
mean training CH
midline
Agro-dealer left the sample 0.121 -0.007 -0.108∗∗
(0.326) (0.034) (0.035)
Farmer left the sample 0.018 -0.005 0.001
(0.134) (0.005) (0.005)
endline
agro-dealer left the sample 0.144 0.017 -0.079+
(0.351) (0.040) (0.042)
Farmer left the sample 0.008 -0.003 -0.001
(0.091) (0.003) (0.003)
Note: Column (1) reports sample means at mid- or endline and standard deviations below; columns (2)-(3) report dier-
ences between treatment and control groups and standard errors below; they are clustered at the level of randomization;
**, *, and + denote signicance at the 1, 5 and 10% levels.
For reasons of transparency and replicability, all outcome variables are described in
the registered pre-analysis plan which can be found in the American Economic Associa-
tion Randomized Controlled Trial Registry. In addition, before midline data collection,
the entire econometric analysis was run on simulated data and also registered in a mock
report (Humphreys, De la Sierra, and Van der Windt, 2013).12 All documents, code,
and data were under revision control and are publicly accessible in a GitHub reposi-
tory which provides time-stamped records of all changes made over the course of the
project.13
Tables 6 to 15 present results in a common layout. Column (1) provides baseline
sample means with standard deviations in parentheses to help gauge eect sizes.14
In column (2), we provide the average treatment eect of the agro-dealer training
at midline, while column (3) reports the average treatment eect of the information
clearinghouse treatment at midline. Standard errors are reported in parentheses below
the coecient estimates. Column (4) reports the number of observations used in the
12A mock report is a dynamic report that integrates all code, such that when midline and endline
data becomes available, one simply has to replace the simulated data with the real data. We use the
knitr engine to integrate R code in LATEX (Xie, 2017).
13The presentation of results in this paper diers somewhat from the way it was pre-registered
and presented in the mock report (and midline report, endline report, and previous versions of this
manuscript). In particular, we reorganized the presentation of the results to match a structure where we
rst look at impact on outcomes at the end of the causal chain and then look at impact on intermediate
outcomes to explore potential mechanisms. While this change does aect the construction of some
of the indices, overall conclusions remain the same. Several pre-registered tables can be found in
Appendix A.3. The entire analysis that follows the pre-registered structure can be found through the
project history in GitHub, for instance here.
14Note that these sample means are reported in levels to allow for this interpretation of eect sizes,
even though we may report the dierence between treatment and control group after using inverse
hyperbolic sine transformations.
17
estimations at midline. Columns (5) and (6) report average treatment eects of the
training and the clearinghouse treatment respectively at endline, that is, after two
seasons. Column (7) reports the number of observations that was used in the endline
estimations. As noted in Section 4, we account for multiple hypothesis testing by
aggregating dierent outcomes within families into overall summary indices, following
Anderson (2008). Results for these indices are reported at the bottom of the tables.15
6.1 Impact on agro-dealers
We start by testing if the interventions aected general business operations of agro-
dealers (Table 6). Sales volume and price, revenue, and number of customers and
maize varieties in stock are the key outcomes of interest. A measure of sales volume was
constructed by asking agro-dealers how much of a specic maize variety they sold in the
previous season. We restrict attention to the four most popular improved varieties, two
of which are hybrids (Longe 7H and Longe 10H) and two of which are open pollinated
varieties (OPVs) (Longe 4 and Longe 5). Total quantity sold is the sum of quantities
sold of these four varieties. We also asked agro-dealers about the sales price of the
four varieties at the start of the season and then calculated the simple average at the
agro-dealer level. We then calculate the revenue (expressed in million UGX) by rst
multiplying prices with quantities sold and then summing over the four varieties.16 We
also include the number of customers that bought maize seed on an average day at the
start of the season, as well as the number of maize varieties that the agro-dealer had in
stock.
Table 6 shows that we do not nd an impact of training agro-dealers on their business
operations. At both midline and endline, the impact on the index is not signicantly
dierent from zero. No particular pattern emerges to explain these insignicant results,
and there is little signicance among any of the outcome variables when estimated
separately. At midline, we nd a negative impact of the training on the average sales
price. At endline, the training seems to have reduced amounts sold, which is also
reected in a lower revenue.
We do nd a positive impact of the information clearinghouse intervention on agro-
dealer operations. At midline, the overall index is signicantly higher among agro-
dealers in the clearinghouse treatment group. Among individual outcomes, treated
agro-dealers sold more maize seed at a higher price, albeit not signicantly so. However,
in combination, this lead to revenues that are almost 20% higher (and this dierence
is signicant at the 10% level).17 At endline, the positive eect of the clearinghouse
15In the regressions with these overall indices, we do not control for the baseline values because this
would imply having the result only for dealers and farmers who have no missing values for any of the
variables constituting these indices at mid-/endline and at baseline, severely reducing statistical power.
16One dollar was about 3600 UGX at the time of the study.
17For reasonably large values, coecients of regressions that involve a dependent variable that has
been transformed using the inverse hyperbolic sine can be interpreted as elasticities (Bellemare and
Wichman, 2020).
18
intervention seems to become stronger, with the overall index now signicant at the 1%
level. The eect is driven by a 31% increase in the number of customers that a treated
agro-dealer attracts, which translates into 6 additional customers per day.
The next set of results focuses on the eect of the interventions on operations related
to one particular variety, specically, the most recently released hybrid (Longe 10H) in
Table 7 and the most recently released OPV (Longe 5) in Table 8. In addition to the
business operation outcomes reported in Table 6, we also include outcomes related to
stock management, given that seed quality decreases with shelf-life. We asked the agro-
dealers how much of the particular seed was carried over from the previous season. Many
agro-dealers reported that they did not carry over any seed, leading to low baseline
means. Furthermore, we asked the agro-dealers to estimate how much they bought
from any provider during the same season. For both varieties, this is slightly more than
what agro-dealers reported to have sold. We expect our treatments to decrease the
amount of seed carried forward and increase the amount of fresh seed procured from
providers. We also asked the agro-dealers to estimate how much of the seed stock was
lost or wasted during the season, and how often they ran out of stock. We expect the
interventions to reduce both losses and stock-outs.
For both varieties, we do not nd signicant eects of the training nor the clear-
inghouse treatment at midline. At endline however, all individual coecient estimates
move in the expected direction for the information clearinghouse, and when outcomes
are combined in an index, the eect is positive and statistically signicant.
Moving one step further up the impact chain, we explore whether reported increases
in the number of customers, sales, and revenues are likely driven by an increase in the
quality of maize seed sold by these agro-dealers. To do so, we instructed enumerators
to buy a random bag of seed from each agro-dealer. This bag was then inspected on a
range of attributes (bag integrity, lot number, packaging date, and shelf-life, etc) and
moisture was measured. Our quality tests of the seed purchased and collected during
the survey did not indicate any eects, although due to the fact that not all agro-dealers
had seed in stock, we may be facing statistical power issues. More information can be
found in Appendix A.1.
6.2 Impact on smallholder farmers
We rst examine harvest-related outcomes for farmers and report the results in Table
9. We start by examining production, plot size, and production scaled by plot size
(i.e., yield) on a randomly selected maize plot. We also look at market participation
(amount sold, sales price, and revenue from maize sales) and how much grain farmers
save to use as seed in the next season. While we expect positive eects on harvest and
sales, the amount kept as seed enters the index negatively.
The coecient estimates for the overall index show no eect of the agro-dealer
training, and a positive eect of the information clearinghouse, albeit only after two
seasons of implementation. Farmers that live in areas where the clearinghouse was
implemented report higher production and productivity at endline than control farmers
19
20
Table 6: Eects on agro-dealer outcomes: Operations
baseline midline endline
mean training CH obs. training CH obs.
Quantity of maize seed sold in kg§† 695.503 -0.092 0.284 292 -0.499+ 0.239 286
(1497.183) (0.220) (0.227) (0.250) (0.253)
Sales price of maize seed in UGX/kg† 4273.897 -192.784+ 99.272 275 -33.867 145.861 264
(955.073) (114.934) (113.292) (143.152) (138.816)
Revenue from maize seed in mln UGX§† 2.890 -0.069 0.185+ 292 -0.227+ 0.143 286
(6.286) (0.104) (0.108) (0.118) (0.118)
Number of maize seed customers per day§† 19.764 -0.056 0.127 294 -0.190 0.310∗∗ 288
(20.689) (0.098) (0.101) (0.116) (0.112)
Number of maize varieties in stock† 2.834 0.042 0.245 295 -0.216 0.221 292
(1.589) (0.266) (0.245) (0.234) (0.220)
Overall index 0.031 -0.130 0.197∗ 274 -0.131 0.238∗∗ 270
(0.610) (0.095) (0.092) (0.086) (0.082)
Max. number of obs. 306 297
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
§Due to the skewness of this variable, the regression was run after an Inverse Hyperbolic Sine transformation. Coecient estimates can therefore be interpreted as
percentage changes. The baseline mean column shows the untransformed variable.
21
Table 7: Eects on agro-dealer outcomes: Operations - Longe 10H
baseline midline endline
mean training CH obs. training CH obs.
Quantity sold in kg§† 288.384 0.050 0.236 256 -0.205 0.352 242
(727.049) (0.206) (0.204) (0.231) (0.239)
Sales price in UGX/kg§† 9.417 -0.025 -0.013 194 -0.019 0.010 187
(0.145) (0.026) (0.026) (0.030) (0.029)
Revenue in mln UGX§† 1.625 0.008 0.130 255 -0.106 0.173 241
(3.839) (0.119) (0.123) (0.130) (0.136)
Amount carried over in kg§† 2.679 -0.186 0.090 262 -0.012 -0.034 250
(12.137) (0.212) (0.215) (0.138) (0.134)
Amount shop bought from provider in kg§† 294.672 0.118 0.206 257 -0.022 0.283 243
(741.810) (0.218) (0.213) (0.250) (0.253)
Amount lost/wasted in kg§† 0.036 -0.001 0.019 257 -0.058 -0.038 243
(0.405) (0.093) (0.097) (0.037) (0.041)
Number of times per month shop ran out§† 1.039 -0.236+ -0.045 192 -0.180 -0.205 185
(1.575) (0.129) (0.133) (0.128) (0.136)
Overall index 0.080 0.030 0.029 244 0.021 0.217∗∗ 233
(0.437) (0.067) (0.070) (0.052) (0.057)
Max. number of obs.1 268 254
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
§Due to the skewness of this variable, the regression was run after an Inverse Hyperbolic Sine transformation. Coecient estimates can therefore be interpreted as
percentage changes. The baseline mean column shows the untransformed variable.
1The comparisons were only made for agro-dealers which had Longe 10H in stock at mid- or endline.
22
Table 8: Eects on agro-dealer outcomes: Operations - Longe 5
baseline midline endline
mean training CH obs. training CH obs.
Quantity sold in kg§† 389.492 -0.040 0.304 261 -0.215 0.316 259
(716.556) (0.222) (0.216) (0.234) (0.230)
Sales price in UGX/kg§† 8.730 0.017 -0.015 249 -0.002 0.013 241
(0.110) (0.016) (0.016) (0.022) (0.022)
Revenue in mln UGX§† 1.193 0.019 0.111 261 -0.080 0.114 258
(2.175) (0.099) (0.096) (0.100) (0.105)
Amount carried over in kg§† 4.312 0.247 -0.092 270 -0.095 -0.004 263
(19.088) (0.324) (0.306) (0.148) (0.155)
Amount shop bought from provider in kg§† 431.451 -0.005 0.253 262 -0.179 0.289 260
(803.696) (0.221) (0.215) (0.232) (0.235)
Amount lost/wasted in kg§† 1.756 -0.150 0.031 266 -0.055 -0.033 261
(10.173) (0.128) (0.128) (0.055) (0.058)
Number of times per month shop ran out§† 0.839 0.053 0.086 248 0.094 -0.054 237
(1.509) (0.100) (0.101) (0.120) (0.126)
Overall index 0.039 0.037 0.012 256 -0.038 0.152∗ 252
(0.401) (0.068) (0.062) (0.058) (0.058)
Max. number of obs.1 275 269
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
§Due to the skewness of this variable, the regression was run after an Inverse Hyperbolic Sine transformation. Coecient estimates can therefore be interpreted as
percentage changes. The baseline mean column shows the untransformed variable.
1The comparisons were only made for agro-dealers which had Longe 5 in stock at mid- or endline.
that live areas where the clearinghouse was not implemented. Yield dierences are
signicant at the 1% level and amount to 10% of the baseline mean. Finally, we look at
the amount of maize that farmers retain for seed in the next season. At midline, we see
that, in line with expectations, clearinghouse treated farmers save less grain for seed.
To further explore the large and signicant eect of the clearinghouse on yields,
we investigate whether specic subgroups of farmers experienced yield gains more than
others. Rerunning the regression only for farmers who did not adopt at baseline led to
a coecient of 56.44 with a standard error of 17.38 (hence, signicance at the 1% level).
For farmers that did adopt at baseline, we nd a coecient of 30.79 with a standard
error of 20.38 (hence, no signicance). This indicates that the eect is plausibly driven
by farmers who did not use improved maize seed at baseline, started using improved
maize seed due to the clearinghouse and, in turn, realized higher yields.
Moving up on the causal chain, we test if the interventions aect the use of agri-
cultural technology by farmers. In particular, we examine the use of purchased maize
seed as a second important family of outcomes at the smallholder level. For the agro-
dealer training, we do not nd any eect at midline, nor at endline. The eect of the
information clearinghouse treatment on overall use (or adoption for convenience) as
measured by the index is positive and signicant at the 5% level at mid- and endline.
Zooming in on individual outcomes, we start with a subjective assessment of seed
used by asking farmers if they "...used any quality maize seed (like OPV or hybrid in
seed) on any of their plots". We see that at midline, farmers that were subjected to the
clearinghouse treatment were 3.5 percentage points more likely to answer this question
with yes than control farmers. After two agricultural seasons, the dierence between
treatment and control farmers increases to 4.2 percentage points. Related, we ask if
farmers bought high-yielding maize varieties at an agro-dealer for any plot. At midline,
we nd a dierence between the clearinghouse treatment and control groups of about 6
percentage points, and this amounts to an almost 20% increase relative to the baseline
mean. At endline, the dierence is about 3 percentage points, but not signicant. We
do not nd an impact of the clearinghouse on the amount of seed that farmers bought
at agro-dealers. However, note that estimates are based on a small sample size (n=599
at midline and 621 at endline) that was conditional on having purchased seed from an
agro-dealer.
Next, we turn our attention to the use of purchased maize seed on a randomly
selected plot. For the use of seed from either hybrid or open-pollinated maize varieties,
we nd positive treatment eects of the clearinghouse, although the coecients are
insignicant.18 As for the more general questions above, we also ask if the seed that
was used on the random plot was obtained from an agro-dealer. We nd an almost
5 percentage point treatment eect for the clearinghouse at midline and an almost 4
percentage point eect at endline. Conversely, we estimate the clearinghouse eect
18Here, we asked farmers which variety they planted in the previous season. If a farmer used Longe
10H, Longe 7H, Longe 7R/Kayongo-go, Bazooka, Longe 6H, Longe 5/Nalongo, Longe 4, Panner, Wema,
KH series, or other hybrid/OPV, and this seed was not recycled or farmer-saved but newly purchased,
it counted as hybrid/open-pollinated maize seed.
23
24
Table 9: Eects on farmer outcomes: Harvest on specic maize plot
baseline midline endline
mean training CH obs. training CH obs.
Production in kg† 463.203 -0.806 -20.372 2884 16.959 43.937∗ 2898
(399.595) (14.050) (14.529) (17.957) (17.922)
Area in acres 1.094 -0.013 -0.003 3004 0.000 0.006 3066
(0.655) (0.029) (0.029) (0.032) (0.038)
Yield in kg/acre† 443.222 -12.216 -23.006 2878 5.118 56.436∗∗ 2889
(304.964) (16.234) (16.964) (15.596) (17.382)
Amount sold in kg§† 195.295 -0.046 -0.201 3063 -0.147 0.173 3137
(297.545) (0.126) (0.124) (0.159) (0.173)
Sales price in UGX/kg 506.954 -7.787 33.027∗ 610 -47.215 12.614 639
(139.389) (14.395) (14.244) (30.547) (41.238)
Revenue in UGX§† 97.783 -0.141 -0.393 3058 -0.354 0.355 3109
(156.538) (0.260) (0.257) (0.341) (0.363)
Amount kept as seed in kg§ 14.506 -0.098 -0.188∗ 2931 -0.043 0.036 2861
(18.530) (0.092) (0.092) (0.108) (0.104)
Overall index -0.020 -0.015 -0.061 2932 0.018 0.097∗ 2900
(0.784) (0.039) (0.039) (0.041) (0.041)
Max. number of obs. 3407 3441
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
§Due to the skewness of this variable, the regression was run after an Inverse Hyperbolic Sine transformation. Coecient estimates can therefore be interpreted as
percentage changes. The baseline mean column shows the untransformed variable.
on the use of farmer-saved seed on the randomly selected plot. Againand in line
with our expectationswe nd that farmers that were exposed to the clearinghouse
treatment reduced their use of saved seed, albeit only signicantly so at midline. Finally,
we examine the product of the amount and the price of maize seed, i.e., the total
expenditure on seed for the plot. We see that in areas where the clearinghouse was
implemented, farmers invest signicantly more in seed.
7 Causal chain and mechanisms
For the agro-dealer training, the underlying mechanism is fairly straightforward: it
potentially changes agro-dealer knowledge and behavior through learning, which re-
sults from exposure to, and uptake of, new and salient information. The information
clearinghouse is a multifarious intervention in that it solves a variety of potentially in-
terlinked information problems simultaneously. If the quality of maize seed is sucient
but some farmers believe that agro-dealers provide sub-standard quality, a clearing-
house may correct their perceptions. If the quality of seed diers between agro-dealers,
the clearinghouse provides farmers with information that may help them to switch to
agro-dealers selling better quality products. Furthermore, the rating system directly
incentivizes agro-dealers to stay ahead of immediate competitors. They can do this by
improving the quality of the products they oer, or by signaling that the quality of
products they oer is good. In this section, we investigate the relative importance of
these dierent impact pathways.
7.1 Agro-dealer knowledge
The primary mechanism underlying the agro-dealer training is learning, which is in turn
expected to increase knowledge of treated agro-dealers. To test if the interventions aect
agro-dealer knowledge, we construct two indices that summarize dierent measures of
knowledge. The rst index aims to measure knowledge about seed storage and handling
and tests if agro-dealers retained the information that was provided during the training.
The test is a short multiple choice quiz of ve questions related to seed carryover between
agricultural seasons, how seed should be stored after repackaging, how seed should be
stored in the storeroom, and whether seed should be repackaged. The exact questions,
the options presented to farmers, and the correct answers are outlined in Appendix A.4.
The second knowledge index aims to capture knowledge about seed more broadly,
and does not focus on seed handling recommendations covered in the training. We
again use multiple choice questions to test if agro-dealers know which seed variety to
recommend if a farmer complains about poor soil or lack of rain, if a farmer is late
for planting, and whether they know what to tell clients who inquire about the yield
benets of hybrid or OPV seed. Again, the questions and (correct) answer options are
explained in Appendix A.4.
Table 11 suggests a positive impact of the agro-dealer training on knowledge at
25
26
Table 10: Eects on farmer outcomes: Adoption
baseline midline endline
mean training CH obs. training CH obs.
Farmer used quality maize seed on any plot† 0.492 -0.021 0.035+ 3206 -0.009 0.042∗ 3282
(0.500) (0.020) (0.020) (0.020) (0.020)
Farmer bought maize seed at agro-input shop for any plot† 0.325 -0.014 0.059∗∗ 3145 0.004 0.031 3225
(0.468) (0.021) (0.021) (0.019) (0.020)
Amount of this maize seed farmer bought at agro-input shop in kg 9.519 0.512 -0.105 599 0.457 0.378 621
(6.920) (0.348) (0.358) (0.419) (0.431)
Farmer used hybrid/open-pollinated maize seed on specic plot1† 0.432 -0.019 0.035 2954 0.009 0.030 3047
(0.495) (0.023) (0.023) (0.023) (0.023)
Farmer bought maize seed at agro-input shop for specic plot† 0.330 -0.010 0.047∗ 3153 0.012 0.036+ 3240
(0.470) (0.022) (0.022) (0.019) (0.019)
Farmer used farmer-saved maize seed on specic plot 0.579 0.020 -0.042+ 3153 -0.009 -0.016 3240
(0.494) (0.022) (0.022) (0.020) (0.020)
Cost of maize seed used on specic plot in UGX§† 14078.272 -0.181 0.499∗ 2848 0.283 0.350+ 2942
(24654.685) (0.235) (0.235) (0.208) (0.209)
Overall index -0.013 -0.030 0.087∗ 2854 0.015 0.086∗ 2978
(0.899) (0.043) (0.042) (0.039) (0.039)
Max. number of obs. 3407 3441
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
1For this variable, only non-recycled (newly purchased, not farmer-saved) seed counted hybrid/open-pollinated seed.
§Due to the skewness of this variable, the regression was run after an Inverse Hyperbolic Sine transformation. Coecient estimates can therefore be interpreted as
percentage changes. The baseline mean column shows the untransformed variable.
midline, but the coecient is just shy of signicance at the 10% level. The (insignicant)
eect of the training is strongest at midline, which seems reasonable as the training was
organized only once at the start of the study (Figure 1). Interestingly, we nd knowledge
eects from the clearinghouse treatment, particularly for agro-dealer knowledge related
to seed storage seed and handling. This eect becomes stronger over time, which
again seems reasonable as this treatment was repeated. This result suggests that the
clearinghouse treatment, with its focus on seed quality, prompts agro-dealers to actively
search for information on better ways to store and handle seed.
The above suggests that providing only knowledge through training is unlikely to
improve outcomes when demand-side constraints are binding, echoing Bold et al. (2022).
In our setting, extra knowledge at the agro-dealer level is only useful if farmers are able
to appreciate the eect of it (see also Homann et al. (2021)). To examine this further,
we exploit the factorial design of the experiment, and focus on the subset of agro-
dealers that were assigned to both the training and the clearinghouse treatment. We
indeed nd a signicant positive interaction eect on key outcomes at the agro-dealer
level at endline (most notably on the overall operations index, as in Table 6). The
positive interaction eect seems to be driven by signicant improvements in eort and
practices (as in Table 12). We also nd a positive interaction eect on the index of
agro-dealer knowledge about seed at midline, but the eect is not signicant, possibly
due to insucient statistical power to test interactions. Results on these interaction
eects are available from the authors upon request.
7.2 Agro-dealer eorts, services, and practices
The information clearinghouse provides agro-dealers with an incentive to become better
than their direct competitors to attract more farmers (Lane, Schonholzer, and Kelley,
2022). Furthermore, the performance feedback provided through the treatment may
motivate agro-dealers to become better and increase quality through behavioral chan-
nels such as social comparison and self-image eects (Gosnell, List, and Metcalfe, 2020).
Agro-dealers can attempt to increase quality by changing the way they store and
handle seed. Furthermore, and particularly if they already use appropriate storage and
handling methods, they may increase eort and start providing more or better services
in an attempt to (indirectly) aect ratings.19
In Table 12, we provide evidence that agro-dealers who are exposed to the clear-
inghouse indeed invest more eort than agro-dealers in the control group. The table
shows results for one overall index and four individual indices that each capture dif-
ferent dimensions of eort, services, and practices. The rst index focuses on eort
and service provision as reported by agro-dealers themselves, and is composed of seven
19For example, if agro-dealers already provide quality inputs but farmers complain about low yields,
agro-dealers may provide training/advice and recommend complementary inputs to get optimal results.
This could increase yields, in turn increasing farmer perceptions of seed quality. Or, agro-dealers may
provide insurance or cash back guarantees, which may then be interpreted by farmers as a signal that
agro-dealers sell high quality.
27
28
Table 11: Eects on agro-dealer outcomes: Knowledge
baseline midline endline
mean training CH obs. training CH obs.
Index of dealer knowledge about seed storage1† 0.000 0.091 0.115 306 0.030 0.124∗ 297
(0.482) (0.076) (0.075) (0.053) (0.055)
Index of dealer knowledge about seed2† 0.000 0.102 0.065 306 -0.009 -0.007 297
(0.533) (0.072) (0.070) (0.080) (0.078)
Overall index 0.000 0.208 0.211+ 306 0.038 0.142 297
(0.729) (0.125) (0.119) (0.107) (0.102)
Max. number of obs. 306 297
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
1The index of dealer knowledge about seed storage contains 5 variables: whether dealer knows how long seed can be carried over, how seed should be stored after
repackaging, what the min. distance between oor and seed is, how seed should be stored in storeroom, whether seed should be repackaged.
2The index of dealer knowledge about seed contains 4 variables: whether dealer knows which seed variety to recommend if farmer complains about poor soil, if farmer
complains about little rain, if farmer is late for planting, what to tell clients about yield benets of hybrid seed.
variables: whether the agro-dealer (1) oers explanations on how to use the seed they
sell to farmers, (2) recommends complementary inputs to get optimal results from
improved varieties, (3) provides advisory services or training, (4) oers discounts for
large-quantity purchases, (5) oers credit, (6) received a seed-related customer com-
plaint since last season, and (7) accepts mobile money payments. A second index
summarizes the perceptions of farmers who purchase from these agro-dealers. This in-
dex is also constructed from seven variables: whether a agro-dealer (1) oers refunds or
insurance, (2) provides credit, (3) oers training or advice to customers, (4) delivers to
the farmgate, (5) provides after-sales service, (6) accepts dierent payment methods,
and (7) sells small quantities. Farmers' answers to these questions are aggregated at
the agro-dealer level before the index is computed.
To handle and store seed correctly, a combination of investments and labor-intensive
practices are necessary. Also during the agro-dealer training, we recommended a mix
of practices that are in reach of dierent types of agro-dealers, some of which may have
excess labor while others may have access to money to invest. A third index groups
a set of labor-intensive seed handling and storage practices. It contains six variables:
whether seed is stored (1) in a dedicated area, (2) in correct lighting, (3) on appropriate
surfaces, and (4) not in open containers; whether the agro-dealer has a pest problem;
and whether the agro-dealership looks clean and professional. Data for these variables
were collected by enumerators through visual inspection, and none are self-reported
by agro-dealers. A fourth index summarizes capital-intensive seed-handling practices,
based on six variables: whether the roof is (1) leak-proof, (2) insulated; (3) whether
the walls are insulated; whether the agro-dealership is (4) ventilated, and (5) displays
any ocial certicate; and (6) whether expired seed is handled correctly. Most of these
variables were collected or at least conrmed by enumerators through visual inspection,
only one of them (whether expired seed is handled correctly) is self-reported.20
We nd that the clearinghouse intervention increases agro-dealer eort and services,
especially at midline, where the coecient of the overall index is signicant at the 1%
level. This eect is driven by treated agro-dealers who signicantly raised their eort
and services, according to farmers. We see that the impact persists until endline, where
the signicant eect on the overall index seems to be driven by the self-reported measure
of eort. We do not nd that the agro-dealer training improves services or practices.
In markets characterized by asymmetric information, signaling is often used to solve
the ineciency problem (Spence, 1973). In our case, as it is dicult to assess seed qual-
ity via visual inspection, agro-dealers may use various strategies to signal to customers
that their products are of good quality. Becoming a member of professional organiza-
tions is one way to do so, as these memberships signal professionalism. Agro-dealers who
try to signal quality will also not shy away from inspections. On the contrary, they may
actively seek inspection so that they can advertise the result in their agro-dealerships.
20To test whether social desirability aects this result, we exclude the last variable from the index
and rerun the analysis, since agro-dealers might report that they handle expired seed correctly but, in
reality, choose not to do so for strategic reasons such as cost management. Doing this does not change
the coecients for the index of capital-intensive seed-handling practices in any notable way.
29
30
Table 12: Eects on agro-dealer outcomes: Eorts and practices
baseline midline endline
mean training CH obs. training CH obs.
Index of dealer eorts and services, self-reported1† 0.000 -0.063 0.066 243 -0.031 0.086+ 297
(0.454) (0.062) (0.060) (0.051) (0.048)
Index of dealer eorts and services, according to farmers2† -0.027 -0.151∗ 0.301∗∗ 259 0.006 0.086 271
(0.583) (0.074) (0.069) (0.092) (0.084)
Index of labor-intensive seed handling practices3† 0.010 0.058 0.099 285 0.083 0.074 274
(0.484) (0.070) (0.065) (0.067) (0.068)
Index of capital-intensive seed handling practices4† 0.000 -0.019 0.000 270 -0.087 0.070 265
(0.508) (0.063) (0.072) (0.092) (0.081)
Overall index 0.032 -0.029 0.359∗∗ 189 0.006 0.165+ 234
(0.540) (0.121) (0.113) (0.099) (0.091)
Max. number of obs. 306 297
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
1The index of dealer eorts and services, self-reported contains 7 variables: whether shop oers explanations, complementary input recommendations, extension/training,
discounts for larger quantities, credit, did not receive seed related customer complaint, accepts mobile money.
2The index of dealer eorts and services, according to farmers contains 7 variables: whether shop oers refund/insurance, credit, training/advice, delivery, after-sales
service, accepts dierent payment methods, sells small quantities. The answers are aggregated at dealer level, then the index is computed.
3The index of labor-intensive seed handling practices contains 6 variables: whether seed is stored in dedicated area, in correct lighting, on correct surface, not in open
containers, whether shop has no pest problem, cleanness and professionality rating by enumerator.
4The index of capital-intensive seed handling practices contains 6 variables: whether roof is leak-proof, roof is insulated, walls are insulated, shop is ventilated, shop
displays ocial certicate, expired seed is handled correctly.
Table 13 collects a set of variables related to signaling quality, including member-
ships in UNADA and other professional associations, trading licenses, the number of
inspections in the last season, and warnings or conscations of seed after inspection.
We nd that at endline, judging by the overall index, the clearinghouse treatment led
to a signicant increase in compliance with or participation in quality assurance mea-
sures. Looking at the individual outcomes, the overall eect seems to be driven by
an increase in registrations with UNADA. We also see that treated agro-dealers were
inspected signicantly more often.
7.3 Switching
An important potential mechanism underlying the eect of the information clearing-
house at the farmer level is the possibility that farmers switch from lower rated agro-
dealers to those with better ratings. To explore this mechanism, we asked farmers if
they switched agro-dealers since the previous season. Results (Table 14) indicate that
only 17% of farmers reported switching at baseline. However, at midline, a signicantly
higher share of farmers in the clearinghouse treatment group reported switching agro-
dealers. Also at endline, we nd a higher propensity for switching among clearinghouse
treated farmers.
Increased switching in itself does not necessarily mean that farmers move from
lower-rated agro-dealers to higher-rated ones. To investigate this, we calculate the
dierence between the rating of agro-dealer the farmer is switching to and that of the
agro-dealer the farmer is switching from. If farmers move to better-rated agro-dealers,
this dierence is positive. We nd that this is indeed the case, and more so during
the second season, although the dierence is not signicantly dierent from zero at
conventional levels (p-value = 0.166).
At the agro-dealer level, we nd that the clearinghouse treatment led to a signicant
increase in the number of customers (Table 6). This can not be explained by farmers
switching, since switching involves an increase in customers for some agro-dealers at
the expense of others, leaving the average number of customers of agro-dealers in the
treatment group unaected. When we look at changes in customers over time, we
actually do not see that the increase in customers is largest for agro-dealers with high
ratings. This suggests that the impact of the clearinghouse treatment on perceptions
to which we turn nextmay be particularly important for agro-dealers with low ratings,
osetting any potential loss that is due to the switching mechanism.
7.4 Perceptions
Finally, the clearinghouse may change farmer perceptions of the quality of seed sold
by agro-dealers. Table 15 provides impact on two measures of farmer perceptions of
quality. First, we asked farmers if they think that maize seed that can be bought at
agro-dealers is counterfeit or adulterated. At baseline, two in three farmers responded
armatively to this question, indicating substantial pessimism about quality. Columns
31
32
Table 13: Eects on agro-dealer outcomes: Memberships, licenses, inspections
baseline midline endline
mean training CH obs. training CH obs.
Shop is registered with UNADA† 0.442 0.040 0.066 252 -0.050 0.118+ 258
(0.497) (0.072) (0.068) (0.072) (0.070)
Shop is member of other professional association† 0.345 -0.035 0.058 268 0.001 0.069 267
(0.476) (0.051) (0.052) (0.073) (0.066)
Shop has trading license issued by local government† 0.749 -0.042 0.021 288 -0.033 0.008 285
(0.435) (0.053) (0.054) (0.056) (0.057)
Number of shop inspections§† 1.532 0.037 -0.097 293 0.038 0.292∗ 273
(1.859) (0.247) (0.259) (0.109) (0.111)
Shop received warning after inspection† 0.317 0.045 0.005 291 0.013 -0.009 284
(0.466) (0.072) (0.073) (0.062) (0.063)
Shop's products were conscated after inspection† 0.145 0.021 -0.027 293 0.014 -0.025 285
(0.353) (0.046) (0.046) (0.033) (0.036)
Overall index -0.004 -0.005 0.047 266 -0.006 0.203∗∗ 253
(0.433) (0.056) (0.055) (0.078) (0.074)
Max. number of obs. 306 297
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
§Due to the skewness of this variable, the regression was run after an Inverse Hyperbolic Sine transformation. Coecient estimates can therefore be interpreted as
percentage changes. The baseline mean column shows the untransformed variable.
33
Table 14: Eects on farmer outcomes: Switching behavior
midline midline endline
mean training CH obs. training CH obs.
Farmer switched to dierent agro-input shop1 0.168 -0.013 0.042∗∗ 3407 -0.024 0.026+ 3441
(0.374) (0.014) (0.014) (0.015) (0.015)
Max. number of obs. 3407 3441
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
1We report the mean and standard deviation at midline because this variable was not collected at baseline.
(2) to (5) show the impact of the clearinghouse for the full sample. The treatment
does not signicantly aect farmer perceptions as measured by this variable at midline
or endline. However, we expect the eect of the clearinghouse on perceptions to be
strongest for farmers who did not use improved maize varieties at baseline. Therefore,
we repeat the analysis for this subgroup of farmers in columns (6) to (9). At midline,
farmers that did not adopt at baseline and live in areas exposed to the clearinghouse
are 12.5 percentage points less likely to think that agro-dealers sell adulterated seed
than similar farmers in areas not assigned to the treatment.
Second, we look at rating data to assess farmer perceptions of product quality. In
particular, we look at the index of maize seed ratings that combine ratings for general
quality, yield, drought tolerance, pest and disease tolerance, time of maturity, and
germination of seed obtained from various agro-dealers.21 These ratings are aggregated
at the farmer level (since one farmer generally rates multiple agro-dealers) and then
the index is calculated.22 We see that the index is positively and signicantly aected
by the clearinghouse treatment, even though the eect is only signicant at the 10%
level. If we restrict the sample to farmers that did not adopt improved maize varieties
at baseline, the treatment eect on the ratings is signicant at the 5% level. The
impact on the overall index that combines the two perception related indicators is also
signicant for this sub-sample.
The improvement in farmer opinions and ratings may reect a real increase in the
quality of seed, because the clearinghouse improved agro-dealers' seed handling eorts
and practices, thus improving seed quality. However, in Appendix A.1 we show that we
do not nd clear evidence that the clearinghouse treatment aected a set of (imperfect)
quality proxies of the seed the agro-dealer sell. Furthermore, looking at Table 12, we
nd no evidence that agro-dealers change seed handling practices in response to the
clearinghouse treatment. As such, it does not seem that the change in perceptions
reects a real change in quality.
An alternative explanation is that the quality of maize seed at most agro-dealers in
our sample is sucient but (non-adopting) farmers misperceive it. This is in line with
Michelson et al. (2021) and Wossen, Abay, and Abdoulaye (2022) who establish that
input quality is generally good but that farmers' beliefs are often incorrect, so that one
simply needs to rectify this misperception to increase adoption. Consistent with this
explanation, we nd that two in three farmers thought that maize seed at agro-dealers
is counterfeit or adulterated at baseline. At the same time, note that the average agro-
dealer was rated 3.4 out of 5 at baseline which indicates that perceived seed quality
was reasonable. The fact that ratings are relatively high is probably due to the fact
that farmers who rate have experience with seed from agro-dealers. Because of the
clearinghouse treatment, mostly non-adopting pessimistic farmers notice that adopting
21As the act of rating agro-dealers was an essential part of the clearinghouse treatment, we only
collect ratings in control areas at endline and so can only test this hypothesis at endline.
22To compute this index at the farmer level, we restrict the sample to observations where a farmer
needs to have rated at least one agro-dealer in the catchment area on all components of the index.
This procedure leads to a sample size reduction, which in turn may aect statistical power.
34
35
Table 15: Eects of the clearinghouse on farmer outcomes: Perceptions
full sample sub-sample
baseline midline endline midline endline
mean CH obs. CH obs. CH obs. CH obs.
Farmer thinks maize seed at agro-dealer is adulterated† 0.685 -0.041 2113 0.020 2167 -0.125∗∗ 903 0.010 944
(0.465) (0.027) (0.028) (0.036) (0.035)
Index of farmer's maize seed ratings of agro-dealers within area1† 0.000 0.092+ 1664 0.141∗ 693
(0.637) (0.054) (0.063)
Overall index 0.019 0.104 1462 0.160∗ 596
(0.770) (0.071) (0.074)
Max. number of obs. 3407 3441 1719 1741
Note: Column (1) reports baseline means and standard deviations below; columns (2) and (4) report dierences between treatment and control groups and standard
errors below; they are clustered at the level of randomization; columns (3) and (5) report number of observations; columns (6) to (9) mirror this structure for the
sub-sample of farmers that did not adopt at baseline, **, *, and + denote signicance at the 1, 5 and 10% levels; † indicates that the variable is included in the overall
index; larger indices indicate more desirable outcomes.
1The index of farmer's maize seed ratings contains 6 ratings: general quality, yield, drought tolerance, pest/disease tolerance, time of maturity, germination. The ratings
are aggregated at farmer level (one farmer rates multiple agro-dealers), then the index is computed. Note that treatment and control groups can only be compared
at endline. At base- and midline, only clearinghouse treated farmers rated dealers in their proximity because being confronted with these questions is part of the
treatment. Hence control dealers were not rated and this line is left blank at midline. At endline, all farmers rated all agro-dealers, so that this variable can be investigated.
peers do not share their pessimism, and they adjust their perceptions. Furthermore,
the clearinghouse aects several measures of adoption already at midline. If we assume
that changing agro-dealer behavior and farmers noticing this change takes some time,
rectifying incorrect perceptions of smallholders must have played an important role in
increasing their adoption.
8 Conclusion
In this study, we hypothesize that seed quality deteriorates because agro-dealers lack
knowledge and/or because asymmetric information results in excessive search costs
for farmers and reduced incentives for agro-dealers. Our hypothesized solutions to
these problems are a training to inform agro-dealers about correct seed handling and
storage practices, and an information clearinghouse based on crowd-sourced ratings of
the quality of seed that agro-dealers sell to farmers. We explore the impact of these
solutions on a range of agro-dealer and farmer outcomes with an experimental design
that allows for causal attribution and investigation of several underlying mechanisms.
Results show that training agro-dealers is generally ineective: our training does
not change agro-dealer knowledge about proper storage and handling practices, their
day-to-day operations, or the seed they sell to farmers. Nor does the training aect
farmers' perceptions of seed quality purchased from these agro-dealers or subsequent use
of purchased seed. The punchline from these results is that training seems insucient
without additional incentives.
This is where the clearinghouse results come into play. They suggest that if agro-
dealers have the right incentives, they may actively seek out knowledge necessary to
improve their operations and stay ahead of their competitors. If agro-dealers are ex-
posed to both incentives and information, they handle and store seed better and attract
more business.
Recently, Dar et al. (2023) showed that an information intervention targeted at
private input suppliers is eective in changing their behavior and in increasing farmer-
level adoption, which seems to contradict our null result for the training at rst sight.
However, their information treatment is very dierent from ours: instead of training
these suppliers in seed handling and storage at the agro-input shop, they give them
access to a new seed variety for their own learning, hoping that they will spread this
information to their clients who will use it on the farm. Moreover, Dar et al. (2023) nd
that business incentives and reputational concerns play a key role when dealers share
information and give advice, which is in line with our nding that incentives matter.
The information clearinghouse clearly aected the market for maize seed in our
study area: sellers and buyers started behaving in line with our theory of change. Agro-
dealers reported more business and smallholders reported increased use of purchased
maize varieties, in turn increasing production outcomes. This eect seems to partly
originate from agro-dealers who expanded service provision and signal quality to farmers
to outperform their competitors. We nd some evidence that the clearinghouse induced
36
farmers to switch between agro-agro-dealers, but most of the impact on farmer outcomes
seems to be driven by the fact that the clearinghouse improved the opinions that farmers
held about agro-dealers and their products.
The signicant impact of the clearinghouse indicates that farmers cannot judge the
quality of maize seed at the time of purchase. If the quality of a product can not be
easily assessed at the time of purchase, one solution is to make sure consumers do not
have to, through regulation and quality assurance. Most LMICs regulate seed quality
in the formal market with codied standards, inspections, and certication systems.
However, and particularly in countries that lack institutional capacity to implement
and enforce the regulatory framework, the reach of these quality assurance systems
and the seed market in generalis limited. As a result, seed certication provides
farmers with a relatively weak and unreliable indication of quality. A decentralized
approach that relies on crowd-sourced quality signals such as the clearinghouse may be
more eective.
Furthermore, peer ratings are likely to measure the dimensions of seed quality that
matter most to smallholders. While it is possible to objectively measure seed qual-
ity (e.g., by sending mystery shoppers, followed by DNA ngerprinting) or agro-dealer
practices (e.g., by sending objective inspectors incognito), it is plausible that farm-
ers are not concerned about genetic purity (indicating whether the seed embodies the
genetic characteristics of a specic variety) but mainly care about seed performance
(e.g., germination rate, vigor, and yield). The opinion of peers who are familiar with
the heterogeneous conditions farmers face, may be more useful and trustworthy for
smallholders than the judgment of an inspector or DNA test.
A key assumption underlying the clearinghouse mechanisms is that, while individual
farmers can not assess the quality of seed at the time of purchase, collective experience
does provide useful information. This is because farmers can learn from experience,
and use this information when making decisions in subsequent seasons. Empirically,
however, farmers may have diculties drawing a causal link between seed quality, on
the one hand, and plant emergence, growth, and harvest, on the other hand given the
extensive set of confounding variables at play (rainfall, soil quality, pest and disease
pressure, inputs, management, and plant genetics), the complexity of genotype-by-
environment-by management (GÖEÖM) interactions, or the stochastic nature of many
of these variables. In addition, Bayesian learning takes time, and opportunities to
learn are few (even tough in Uganda there are two agricultural seasons). This partly
explains the rich body of research on the role of peer eects in technology adoption: the
ability to combine own experience with the experience of farmers in a similar location
is therefore likely to provide a good signal about the quality of seed (Conley and Udry,
2010; Bandiera and Rasul, 2006; Foster and Rosenzweig, 1995). As farmers are thus
unlikely to discover quality of inputs themselves, an information clearinghouse that
relies on peer ratings is expected to increase data points that farmers can use when
making adoption decisions.
While the crowd-sourced information clearinghouse tested in this study may also be
feasible at scale, the clearinghouse idea cannot not be approached naively. There is con-
37
siderable controversy surrounding the credibility of reviewing platforms and the ability
of sellers on these platforms to improve their ratings with payments to consumers to
provide favorable ratings, automated bot reviews, and other strategic practices. There
are also more practical considerations, such as how long ratings should remain valid,
or how to deal with sellers who receive high ratings on one set of attributesseed
quality, in this casebut perform poorly in other categories, for example, by engaging
in unfair labor practices, discriminating against certain types of customers, engaging
in anti-competitive behavior, or promoting environmentally hazardous products. Rat-
ing platforms could potentially amplify these practices and biases rather than address
them. While this paper oers both proof of concept and an empirical test of theoret-
ical predictions about the relationship between observable product quality and buyer
and seller behavior in a market characterized by asymmetric information, further pi-
loting and testing is needed at increasingly larger scales to test additional hypotheses
about market performance, supply and demand responses, and productivity, welfare,
and resilience outcomes.
38
References
Akerlof, G. A. 1970. The Market for "Lemons": Quality Uncertainty and the Market
Mechanism. The Quarterly Journal of Economics 84 (3): 488500.
Allcott, H. and T. Rogers. 2014. The Short-Run and Long-Run Eects of Behav-
ioral Interventions: Experimental Evidence from Energy Conservation. American
Economic Review 104 (10): 300337.
Anderson, M. L. 2008. Multiple inference and gender dierences in the eects of early
intervention: A reevaluation of the Abecedarian, Perry Preschool, and Early Training
Projects. Journal of the American Statistical Association 103 (484): 14811495.
Angrist, J., E. Bettinger, and M. Kremer. 2006. Long-Term Educational Consequences
of Secondary School Vouchers: Evidence from Administrative Records in Colombia.
American Economic Review 96 (3): 847862.
Ashour, M., D. O. Gilligan, J. B. Hoel, and N. I. Karachiwalla. 2019. Do Beliefs About
Herbicide Quality Correspond with Actual Quality in Local Markets? Evidence from
Uganda. The Journal of Development Studies 55 (6): 12851306.
Ashraf, N., X. Giné, and D. Karlan. 2009. Finding missing markets (and a disturbing
epilogue): Evidence from an export crop adoption and marketing intervention in
Kenya. American Journal of Agricultural Economics 91 (4): 973990.
Bandiera, O. and I. Rasul. 2006. Social Networks and Technology Adoption in Northern
Mozambique*. The Economic Journal 116 (514): 869902.
Barriga, A. and N. Fiala. 2020. The supply chain for seed in Uganda: Where does it
go wrong? World Development 130: 104928.
Bellemare, M. F. and C. J. Wichman. 2020. Elasticities and the Inverse Hyperbolic
Sine Transformation. Oxford Bulletin of Economics and Statistics 82 (1): 5061.
Bennett, D. and W. Yin. 2019. The Market for High-Quality Medicine: Retail Chain
Entry and Drug Quality in India. The Review of Economics and Statistics 101 (1):
7690.
Björkman Nyqvist, M., J. Svensson, and D. Yanagizawa-Drott. 2022. Can good prod-
ucts drive out bad? A randomized intervention in the antimalarial medicine market
in Uganda. Journal of the European Economic Association 20 (3): 9571000.
Bold, T., K. C. Kaizzi, J. Svensson, and D. Yanagizawa-Drott. 2017. Lemon technolo-
gies and adoption: Measurement, theory and evidence from agricultural markets in
Uganda. The Quarterly Journal of Economics 132 (3): 10551100.
39
Bold, T., S. Ghisol, F. Nsonzi, and J. Svensson. 2022. Market Access and Quality
Upgrading: Evidence from Four Field Experiments. American Economic Review
112 (8): 251852.
Borlaug, N. 2007. Feeding a Hungry World. Science 318 (5849): 359359.
Campos, F., M. Frese, M. Goldstein, L. Iacovone, H. C. Johnson, D. McKenzie, and
M. Mensmann. 2017. Teaching personal initiative beats traditional training in boost-
ing small business in West Africa. Science 357 (6357): 12871290.
Conley, T. G. and C. R. Udry. 2010. Learning about a New Technology: Pineapple in
Ghana. American Economic Review 100 (1): 3569.
Dar, M., A. De Janvry, E. Sadoulet, K. Emerick, and E. Wiseman. 2023. Private input
suppliers as information agents for technology adoption in agriculture. American
Economic Review: Applied Economics forthcoming.
De, A., C. Miehe, and B. Van Campenhout. 2022. Gender bias in consumer perceptions:
The case of agro-input dealers in Uganda. IFPRI Discussion Paper .
Drexler, A., G. Fischer, and A. Schoar. 2014. Keeping It Simple: Financial Literacy
and Rules of Thumb. American Economic Journal: Applied Economics 6 (2): 131.
Duo, E., R. Glennerster, and M. Kremer. 2007. Using randomization in development
economics research: A toolkit. Handbook of Development Economics 4: 38953962.
Duo, E., M. Kremer, and J. Robinson. 2011. Nudging farmers to use fertilizer: The-
ory and experimental evidence from Kenya. American Economic Review 101 (6):
235090.
Foster, A. D. and M. R. Rosenzweig. 1995. Learning by doing and learning from others:
Human capital and technical change in agriculture. Journal of Political Economy
103 (6): 11761209.
Giné, X. and G. Mansuri. 2021. Money or Management? A Field Experiment on
Constraints to Entrepreneurship in Rural Pakistan. Economic Development and
Cultural Change 70 (1): 4186.
Gollin, D., C. W. Hansen, and A. M. Wingender. 2021. Two Blades of Grass: The
Impact of the Green Revolution. Journal of Political Economy 129 (8): 23442384.
Gosnell, G. K., J. A. List, and R. D. Metcalfe. 2020. The Impact of Management
Practices on Employee Productivity: A Field Experiment with Airline Captains.
Journal of Political Economy 128 (4): 11951233.
Hanna, R., S. Mullainathan, and J. Schwartzstein. 2014. Learning through noticing:
Theory and evidence from a eld experiment. The Quarterly Journal of Economics
129 (3): 13111353.
40
Hasanain, S. A., M. Y. Khan, and A. Rezaee. 2023. No bulls: Experimental evidence on
the impact of veterinarian ratings in Pakistan. Journal of Development Economics
161: 102999.
Homann, V., S. K. Mutiga, J. W. Harvey, R. J. Nelson, and M. G. Milgroom. 2021.
Observability of food safety losses in maize: Evidence from Kenya. Food Policy 98:
101895.
Humphreys, M., R. S. De la Sierra, and P. Van der Windt. 2013. Fishing, commitment,
and communication: A proposal for comprehensive nonbinding research registration.
Political Analysis 120.
Imbens, G. W. and M. Kolesár. 2016. Robust Standard Errors in Small Samples: Some
Practical Advice. The Review of Economics and Statistics 98 (4): 701712.
Kahneman, D. and A. Tversky. 1979. Prospect Theory: An Analysis of Decision under
Risk. Econometrica 47 (2): 263291.
Karlan, D. and M. Valdivia. 2011. Teaching Entrepreneurship: Impact of Business
Training on Micronance Clients and Institutions. The Review of Economics and
Statistics 93 (2): 510527.
Karlan, D., R. Osei, I. Osei-Akoto, and C. Udry. 2014. Agricultural decisions after
relaxing credit and risk constraints. The Quarterly Journal of Economics 129 (2):
597652.
Lane, G., D. Schonholzer, and E. Kelley. 2022. Information and Strategy in Lemon
Markets: Improving Safety in Informal Transit. PEDL Research Paper .
Lin, W. 2013. Agnostic notes on regression adjustments to experimental data: Reex-
amining Freedman's critique. Annals of Applied Statistics 7 (1): 295318.
McKenzie, D. and C. Woodru. 2013. What Are We Learning from Business Training
and Entrepreneurship Evaluations around the Developing World? The World Bank
Research Observer 29 (1): 4882.
Michelson, H., A. Fairbairn, B. Ellison, A. Maertens, and V. Manyong. 2021. Mis-
perceived quality: Fertilizer in Tanzania. Journal of Development Economics 148:
102579.
Muralidharan, K., M. Romero, and K. Wüthrich. 2019. Factorial designs, model se-
lection, and (incorrect) inference in randomized experiments. Tech. rep., National
Bureau of Economic Research.
Reimers, I. and J. Waldfogel. 2021. Digitization and pre-purchase information: The
causal and welfare impacts of reviews and crowd ratings. American Economic Review
111 (6): 194471.
41
Rozin, P. and E. B. Royzman. 2001. Negativity Bias, Negativity Dominance, and
Contagion. Personality and Social Psychology Review 5 (4): 296320.
Spence, M. 1973. Job Market Signaling. The Quarterly Journal of Economics 87 (3):
355374.
Suri, T. 2011. Selection and comparative advantage in technology adoption. Econo-
metrica 79 (1): 159209.
Suri, T. and C. Udry. 2022. Agricultural Technology in Africa. Journal of Economic
Perspectives 36 (1): 3356.
Van Campenhout, B. 2021. The Role of Information in Agricultural Technology Adop-
tion: Experimental Evidence from Rice Farmers in Uganda. Economic Development
and Cultural Change 69 (3): 12391272.
Wossen, T., K. A. Abay, and T. Abdoulaye. 2022. Misperceiving and misreporting
input quality: Implications for input use and productivity. Journal of Development
Economics 157: 102869.
Xie, Y. 2017. Dynamic Documents with R and knitr. CRC Press.
42
A Appendix
A.1 Eects on seed quality at agro-dealer
To assess the eect of the training and clearinghouse intervention on the quality of
seed sold by agro-dealers, we bought a bag of seed at each agro-dealer and inspected
it on a number of attributes. First, we measured the moisture content of the seed.
In Table A1, we see that the clearinghouse treatment reduced moisture as expected,
but the parameter is estimated imprecisely, perhaps due to the smaller data set as we
were not able to source seed from all agro-dealers and the comparisons were only made
for agro-dealers from which the enumerator was able to buy a bag of maize seed at
mid- or endline. We further look at the integrity of the package and whether it shows
important information such as the packaging date, shelf-life and the lot number. We do
not nd that the clearinghouse treatment nor the agro-dealer training aected quality
proxies of the seed that agro-dealers sell. However, because our proxies of seed quality
are far from perfect and we rely on a smaller sample, we cannot safely conclude that
the treatments did not aect seed quality.23
23Assessing seed quality is not only challenging for farmers, but also for researchers. Even though
investigating moisture is an attempt to test the quality of seeds in a quantitative and objective way,
this variable is one-dimensional and only a weak indication of seed quality. Additional ways to capture
seed quality include lab testing to measure purity as the rate of extraneous, non-seed material in the
bag, germination using a germination chamber, grow-out tests for genetic purity using morphological
analysis, and DNA ngerprinting to test for genetic purity using single nucleotide polymorphisms tests.
43
44
Table A1: Eects on agro-dealer outcomes: Bag of maize seed
baseline midline endline
mean training CH obs. training CH obs.
Moisture in %† 13.564 0.017 -0.122 175 -0.041 -0.220 261
(1.482) (0.142) (0.144) (0.198) (0.197)
Bag shows packaging date† 0.689 0.053 0.050 179 -0.091 0.035 265
(0.464) (0.069) (0.072) (0.063) (0.064)
Shelf-life in days1† 60.951 -18.930 -8.272 164 13.091 6.352 240
(40.960) (22.091) (20.869) (8.243) (8.289)
Seed is in original undamaged bag† 0.940 0.025 0.002 179 0.006 0.051 265
(0.238) (0.044) (0.046) (0.053) (0.055)
Bag shows lot number† 0.508 0.025 -0.001 179 -0.138∗ 0.027 265
(0.501) (0.106) (0.107) (0.062) (0.064)
Overall index 0.065 0.083 0.108 160 -0.067 0.108 236
(0.364) (0.103) (0.103) (0.094) (0.090)
Max. number of obs.2 179 265
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
1Days since the packaging date or, if the bag does not show the packaging date, days since the expiry date minus 6 months.
2The comparisons were only made for agro-dealers in which the enumerator was able to buy a bag of maize seed at mid- or endline. Also, we do not control for the
baseline values of the outcome variables in the entire table because only 144 of the 179 dealers who had seed at midline also had seed at baseline and only 183 of the
265 dealers who had seed at endline also had seed at baseline, so that controlling for baseline values would reduce the sample sizes drastically.
A.2 Rating computation: details
What to do if a treated dealer does not receive a single rating? If a shop in
a treated catchment area is not rated by a single farmer, e.g., because no farmer in our
sample knows him or her, we could ll in the catchment area mean as his or her rating.
However, this is not as innocent as it seems because it is likely that the lack of ratings
is not random. Poor quality dealers have less customers, so their likelihood to get rated
is lower. Giving them average catchment area ratings inates the ratings of these low
quality dealers. Instead, we simply told farmers that we do not have information about
this shop (implicitly informing the farmer that it exists). 16 of 193 treated dealers were
not rated by a single farmer in the rst round.
Should more ratings lead to better ratings? Some agro-dealers were not rated by
any smallholder in the rst round, while others were rated by up to 22 smallholders. If
dealer A is rated by 10 farmers and gets rating 3,5 and dealer B is rated by 1 farmer and
gets rating 3,6, we treat dealer B as the better dealer. Even though receiving many (few)
ratings can be related to good (poor) quality (the lack of ratings could be nonrandom,
see previous paragraph), there could be other reasons why dealers are rated by many
(few) farmers. Furthermore, giving higher ratings to better-known dealers could harm
new dealers entering the market and dealers who are discriminated, e.g., due to their
gender. Also on TripAdvisor, having more reviews than a rival hotel does not lead to
a better rating.
Should ratings depend on catchment area dealer performance? The following
examples show that ratings should not depend on catchment area averages. In an area
with poor quality dealers in which one dealer is a bit better than the rest but still
poor, we do not want this dealer to be rated well (i.e., expose farmers to poor quality
dealers). Similarly, in an area with good dealers in which one dealer is a bit worse than
the rest but still good, we do not want this dealer to be rated poorly (which would be
unfair towards him or her). On the other hand, less than 9% of agro-dealers received
a rating below 3 out of 5, so we would throw away valuable data if we would only
disseminate good scores without any variation. Therefore, we take the distribution of
ratings into account by using quintiles. Consequently, less dealers receive rating 4 or 5,
more dealers receive rating 1 or 2. This could strengthen the eect of the treatment on
dealer eort. If dealers get ratings 1 or 2 instead of 4 or 5, they could feel more inclined
to improve their scores. Consequently, also the eect on seed quality itself could be
larger. However, the clearinghouse should also have a signaling eect, which might be
weaker if more dealers are rated 1 or 2 instead of 4 or 5 (dealers would seem to be of
worse quality to farmers). Therefore, we chose words with a positive connotation as
the quintile names for rating dissemination. As most dealers received a good or very
good rating before taking the distribution into account, we ensure that even a 2 is still
communicated as good to farmers to not weaken the signaling eect. That is why
the rst quintile is translated to okay and gets one star, the second one is named
45
good and receives two stars, the third quintile is very good and gets three stars, the
fourth and fth one are excellent and awarded with four and ve stars. This way of
considering the distribution of the original ratings when choosing the names also helps
us to disseminate ratings as truthfully, purely and as closely to reality as possible.
Are female dealers rated worse than male dealers? Because we found signi-
cant dierences between the ratings of female (41% of dealers) and male agro-dealers
(59% of dealers) after controlling for some potentially confounding variables like edu-
cation and for several indications of quality, we have no reason to believe that these
dierences in perception can be explained by dierences in real quality. Instead, it is
likely that women are perceived to be worse due to discrimination (De, Miehe, and
Van Campenhout, 2022), so that we adjusted the ratings of female dealers accordingly
to prevent that they are harmed by our intervention. We regressed all seed quality at-
tributes on the gender dummy and added the resulting coecients to the initial ratings
of female dealers.
46
A.3 Outcome variables and results as they were pre-registered
47
48
Table A2: Eects on primary dealer outcomes
baseline midline endline
mean training CH obs. training CH obs.
Quantity of maize seed sold in kg§† 695.503 -0.092 0.284 292 -0.499+ 0.239 286
(1497.183) (0.220) (0.227) (0.250) (0.253)
Sales price of maize seed in UGX/kg 4273.897 -192.784+ 99.272 275 -33.867 145.861 264
(955.073) (114.934) (113.292) (143.152) (138.816)
Revenue from maize seed in mln UGX§† 2.890 -0.069 0.185+ 292 -0.227+ 0.143 286
(6.286) (0.104) (0.108) (0.118) (0.118)
Number of maize seed customers per day§† 19.764 -0.056 0.127 294 -0.190 0.310∗∗ 288
(20.689) (0.098) (0.101) (0.116) (0.112)
Moisture in randomly selected seed bag in % 13.563 0.017 -0.122 175 -0.041 -0.220 261
(1.442) (0.142) (0.144) (0.198) (0.197)
Index of capital-intensive seed handling practices1† 0.000 -0.019 0.000 270 -0.087 0.070 265
(0.508) (0.063) (0.072) (0.092) (0.081)
Index of labor-intensive seed handling practices2† 0.010 0.058 0.099 285 0.083 0.074 274
(0.484) (0.070) (0.065) (0.067) (0.068)
Index of all seed handling practices3 0.009 0.042 0.052 251 0.021 0.083 248
(0.382) (0.051) (0.053) (0.063) (0.059)
Index of dealer's eorts and services4† 0.000 -0.063 0.066 243 -0.031 0.086+ 297
(0.454) (0.062) (0.060) (0.051) (0.048)
Index of shop's maize seed ratings by farmers5 -0.018 0.020 0.122 327
(0.595) (0.102) (0.101)
Overall index 0.007 -0.004 0.214+ 215 -0.058 0.239∗ 258
(0.591) (0.130) (0.121) (0.128) (0.117)
Max. number of obs. for dealer survey outcomes 306 297
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
§Due to the skewness of this variable, the regression was run after an Inverse Hyperbolic Sine transformation. Coecient estimates can therefore be interpreted as
percentage changes. The baseline mean column shows the untransformed variable.
1The index of capital-intensive seed handling and storage practices contains 6 variables: whether roof is leak-proof, whether roof is insulated, whether walls are insulated,
whether shop is ventilated, whether any ocial certicate is displayed, whether expired seed is handled correctly.
2The index of labor-intensive seed handling and storage practices contains 6 variables: whether seed is stored in dedicated area, whether shop has no pest problem,
whether seed is stored in correct lighting, whether seed is stored on correct surface, whether seed is not stored in open containers, cleanness and professionality rating
by enumerator.
3The index of all seed handling and storage practices contains 12 variables: the ones included in the index of capital-intensive practices and the ones included in the
index of labor-intensive practices.
4The index of dealer's eorts and services contains 7 variables: whether shop oers explanations, complementary input recommendations, extension/training, discounts
for larger quantities, credit, did not receive seed related customer complaint, accepts mobile money.
5The index of shop's maize seed ratings by farmers contains 6 ratings: general quality, yield, drought tolerance, pest/disease tolerance, time of maturity, germination.
Ratings are aggregated at shop level (one shop is rated by multiple farmers), then the index is computed.
49
Table A3: Eects on secondary dealer outcomes: Indices
baseline midline endline
mean training CH obs. training CH obs.
Index of dealer's motivation and satisfaction1 0.000 0.033 0.000 306 -0.109 -0.076 286
(0.674) (0.082) (0.085) (0.082) (0.086)
Index of dealer's self-ratings2 0.000 -0.068 -0.002 306 -0.132 0.080 297
(0.651) (0.084) (0.079) (0.086) (0.079)
Index of dealer's eorts and services according to farmers3 -0.027 -0.151∗ 0.301∗∗ 259 0.006 0.086 271
(0.583) (0.074) (0.069) (0.092) (0.084)
Index of dealer's knowledge about seed storage4 0.000 0.091 0.115 306 0.030 0.124∗ 297
(0.482) (0.076) (0.075) (0.053) (0.055)
Index of dealer's knowledge about seed5 0.000 0.102 0.065 306 -0.009 -0.007 297
(0.533) (0.072) (0.070) (0.080) (0.078)
Max. number of obs. 306 297
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
1The index of dealer's motivation and satisfaction contains 3 variables: whether dealers see themselves working as agro-dealers in future, would recommend working as
dealers, how happy dealers feel when they come to work. We report the mean and standard deviation at midline because these variables were not collected at baseline.
2The index of dealer's self-ratings contains 5 ratings: location, price, product quality, stock, reputation.
3The index of dealer's eorts and services according to farmers contains 7 variables: whether shop oers refund/insurance, credit, training/advice, delivery, after-sales
service, accepts dierent payment methods, sells small quantities. The answers are aggregated at dealer level, then the index is computed.
4The index of dealer's knowledge about seed storage contains 5 variables: whether dealer knows how long seed can be carried over, how seed should be stored after
repackaging, what the min. distance between oor and seed is, how seed should be stored in storeroom, whether seed should be repackaged.
5The index of dealer's knowledge about seed contains 4 variables: whether dealer knows which seed variety to recommend if farmer complains about poor soil, if farmer
complains about little rain, if farmer is late for planting, what to tell clients about yield benets of hybrid seed.
50
Table A4: Eects on primary farmer outcomes
baseline midline endline
mean training CH obs. training CH obs.
Farmer planted improved maize seed on any plot† 0.492 -0.021 0.035+ 3206 -0.009 0.042∗ 3282
(0.500) (0.020) (0.020) (0.020) (0.020)
Farmer bought maize seed at agro-input shop for any plot† 0.325 -0.014 0.059∗∗ 3145 0.004 0.031 3225
(0.468) (0.021) (0.021) (0.019) (0.020)
Amount of this seed farmer bought at agro-input shop in kg 9.519 0.512 -0.105 599 0.457 0.378 621
(6.920) (0.348) (0.358) (0.419) (0.431)
Index of farmer's maize seed ratings of agro-dealers within catchment area1 0.000 0.021 0.092+ 1664
(0.637) (0.054) (0.054)
Index of farmer's general ratings of agro-dealers within catchment area2 0.000 -0.026 -0.005 1706
(0.657) (0.043) (0.042)
Index of services of agro-dealers within catchment area according to farmers3 -0.037 -0.138+ 0.161∗ 312 0.034 0.131+ 320
(0.609) (0.073) (0.067) (0.081) (0.077)
Farmer switched to dierent agro-input shop4† 0.168 -0.013 0.042∗∗ 3407 -0.024 0.026+ 3441
(0.374) (0.014) (0.014) (0.015) (0.015)
Index of farmer's practices on randomly selected plot5† 0.008 0.011 -0.026 2929 0.001 0.016 3053
(0.400) (0.019) (0.019) (0.021) (0.021)
Farmer thinks maize seed at agro-dealer is adulterated 0.685 -0.033 -0.041 2113 -0.041 0.020 2167
(0.465) (0.027) (0.027) (0.028) (0.028)
Farmer planted land race maize seed on randomly selected plot† 0.448 0.015 -0.013 2954 0.009 -0.024 3047
(0.497) (0.021) (0.020) (0.022) (0.022)
6
Overall index 0.009 0.008 0.017 2933 -0.023 0.063+ 3083
(0.698) (0.033) (0.034) (0.034) (0.034)
Max. number of obs. 3407 3441
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
1The index of farmer's maize seed ratings contains 6 ratings: general quality, yield, drought tolerance, pest/disease tolerance, time of maturity, germination. The ratings
are aggregated at farmer level (one farmer rates multiple agro-dealers), then this index is computed.
2The index of farmer's general ratings contains 6 ratings: general quality, location, price, product quality, stock, reputation. The ratings are aggregated at farmer level
(one farmer rates multiple agro-dealers), then this index is computed.
3The index of services of agro-dealers within catchment area contains 7 variables: whether shop oers refund/insurance, credit, training/advice, delivery, after-sales
service, accepts dierent payment methods, sells small quantities. The answers are aggregated at shop level, then the index is computed at farmer level. Only 320
farmers answered all 7 questions for at least one shop within the catchment area at baseline and at endline.
4We report the mean and standard deviation at midline because this variable was not collected at baseline.
5The index of farmer's practices contains 10 variables: whether farmer spaced seed correctly, sowed correct number of seeds/hill, applied organic manure, DAP/NPK,
Urea, pesticides/herbicides/fungicides, weeded suciently, weeded at correct time, planted at correct time, re-sowed.
6We report the mean and standard deviation at midline because not all variables in this index were collected at baseline.
51
Table A5: Eects on secondary farmer outcomes: Adoption on randomly selected maize plot
baseline midline endline
mean training CH obs. training CH obs.
Farmer planted hybrid seed† 0.264 0.002 0.009 2654 -0.023 0.032 2700
(0.441) (0.022) (0.022) (0.023) (0.023)
Farmer planted open-pollinated seed† 0.260 -0.017 0.002 2654 0.010 -0.007 2700
(0.439) (0.022) (0.022) (0.020) (0.021)
Farmer planted farmer-saved seed† 0.579 0.020 -0.042+ 3153 -0.009 -0.016 3240
(0.494) (0.022) (0.022) (0.020) (0.020)
Farmer planted seed bought at agro-input shop† 0.330 -0.010 0.047∗ 3153 0.012 0.036+ 3240
(0.470) (0.022) (0.022) (0.019) (0.019)
Farmer planted hybrid or open-pollinated seed1 0.432 -0.019 0.035 2954 0.009 0.030 3047
(0.495) (0.023) (0.023) (0.023) (0.023)
Overall index -0.003 0.000 0.002 2867 -0.010 0.026 2963
(0.553) (0.024) (0.024) (0.025) (0.025)
Max. number of obs. 3407 3441
Note: Column (1) reports baseline means and standard deviations below; columns (2), (3), (5), and (6) report dierences between treatment and control groups and
standard errors below; they are clustered at the level of randomization; columns (4) and (7) report number of observations; **, *, and + denote signicance at the 1, 5
and 10% levels; † indicates that the variable is included in the overall index; larger indices indicate more desirable outcomes.
1For this variable, only seed which was not farmer-saved counted as hybrid seed and only seed which was not recycled too often counted as open-pollinated seed.
A.4 Multiple choice questions to measure dealer knowledge
Dealer knowledge about seed storage
1. How long can seed be carried over before losing viability?
(a) Seed can be carried over into the next seasons as you can store seed for 12
months.
(b) Seed cannot be carried over into the next seasons as 6 months is the longest
seed can be stored.
(c) This depends on the seed: hybrids cannot be carried over, OPVs can be
carried over for 5 seasons.
(d) I don't know.
2. How should seed best be stored after repackaging?
(a) Airtight in polyethylene bags.
(b) In paper bags or perforated polyethylene bags.
(c) In a sealed tin/plastic container.
(d) I don't know.
3. What is the minimum recommended distance between the oor and where seed
is stored?
(a) 0 inches, seed should be stored directly on the oor for maximum stability.
(b) Minimum 2 inches from the oor.
(c) Minimum 6 inches from the oor.
(d) I don't know.
4. How should seed ideally be stored in your store room?
(a) In sealed cardboard boxes.
(b) Stacked on pallets.
(c) Arranged on shelves with sucient space between packets.
(d) I don't know.
5. Which statement do you agree most with?
(a) You should repackage all your seed to visually verify that you are selling
good quality seed.
(b) You should repackages all your seed so you can sell more to small farmers.
(c) You should avoid repackaging your seed as much as possible.
(d) I don't know.
52
Dealer knowledge about seed
1. If a farmer complains about poor soil, which maize variety do you recommend?
(a) Longe 5.
(b) Bazooka.
(c) Longe 10H.
(d) I don't know.
2. What do you tell clients who inquire about the yield benets of hybrid seeds?
(a) Hybrid seeds double maize yields (increasing yield from about 4 to 8
bags/acre).
(b) Hybrid seeds triple maize yields (increasing yield from about 4 to 12
bags/acre).
(c) Hybrid seeds increase yields tenfold (increasing yield from about 4 to 40
bags/acre).
(d) I don't know.
3. If a farmer misses the rains or lives in an area that receives little rain, which maize
variety do you recommend?
(a) Longe 10H.
(b) Longe 7H.
(c) Wema.
(d) I don't know.
4. If a farmer is late for planting in the short season and needs a fast maturing
variety, which maize variety do you recommend?
(a) Bazooka.
(b) Longe 10H.
(c) Myezi mitatu (mm3).
(d) I don't know.
53
ALL IFPRI DISCUSSION PAPERS 
All discussion papers are available here 
They can be downloaded free of charge 
INTERNATIONAL FOOD POLICY RESEARCH INSTITUTE 
www.ifpri.org 
IFPRI HEADQUARTERS 
1201 Eye Street, NW 
Washington, DC 20005 USA 
Tel.: +1-202-862-5600 
Fax: +1-202-862-5606 
Email: ifpri@cgiar.org