CoChat
Sign Up
Cited by 2,180Agricultural Innovations and PracticesInnovation Diffusion and ForecastingPolicy Transfer and Learning

The DOI matches its Crossref record; 107 of the 111 references it cites were found in a registry or an index; none of the 10 sources its claims cite is checked for support yet.

Learning about a New Technology: Pineapple in Ghana

Conley, Timothy G, Udry, Christopher R·2010·American Economic Review
DOI 10.1257/aer.100.1.35
Open this paper in CoChat — free account

Abstract

This paper investigates the role of social learning in the diffusion of a new agricultural technology in Ghana. We use unique data on farmers' communication patterns to define each individual's information neighborhood. Conditional on many potentially confounding variables, we find evidence that farmers adjust their inputs to align with those of their information neighbors who were surprisingly successful in previous periods. The relationship of these input adjustments to experience further indicates the presence of social learning. In addition, applying the same method to input choices for another crop, of known technology, correctly indicates an absence of social learning effects. (JEL D83, O13, O33, Q16)

Overview

AI snapshot

Written from the paper’s full text.

TL;DR

Conley and Udry use direct communication data from Ghanaian pineapple farmers to test whether farmers learn from information neighbors’ experiments, and find that farmers adjust fertilizer (and labor) inputs toward surprisingly successful neighbors’ practices, especially inexperienced farmers, with no analogous learning for established maize-cassava. They interpret these patterns as social learning, robust to controls for geographic growing conditions, credit/financial ties, clan, religion, and alternative network definitions.

Key findings

  • Farmers are more likely to change fertilizer use after bad news at their lagged input level: logit coefficient 12.32 [3.72]; a one-SD increase in this bad-news index (.12) raises the logit index by 1.5, while the 75th–25th percentile range of estimated logit indices is 2.8. Bad news at alternative levels lowers the probability of changing (coefficient -2.98 [1.91]; one-SD increase ~.13 lowers logit index ~.15).
  • For magnitude, the good-news input index M is positive and significant (0.99 [.16]); a one-SD increase in M (~4) is associated with ~4 cedis per plant increase in fertilizer use, greater than median fertilizer use per plant among users. Inexperienced farmers drive this: M*Inexperienced 1.09 [0.22] vs M*Experienced 0.10 [0.32], statistically different at 3% level.
  • Farmers respond more to news from experienced neighbors (1.02 [0.17] vs -0.13 [0.37] for inexperienced neighbors) and same-wealth neighbors (1.03 [0.18] vs -0.41 [0.32]), differences p<2%; big-farm vs small-farm and same-soil vs different-soil differences are not statistically significant.
  • Pineapple labor: M-tilde coefficient 0.95 [.37]; one-SD increase in M-tilde (396) associated with ~376 cedis per sucker increase, one-fifth median labor use per sucker (1,845). Maize-cassava: M-tilde coefficient 0.04 [0.14], virtually zero; geographic-neighborhood index 0.25 [0.14] (one-SD increase ~590,000 cedis/ha vs mean labor input 650,000 cedis/ha), attributed to growing conditions not learning.

Methods

  • Data: two-year survey (1996–98) of ~240 households in southern Ghana; four villages near Nsawam/Aburi, three where pineapple farmed; random sample of married individuals (60 couples/triples per village), 15 rounds ~6 weeks. Final estimation sample: 107 plantings by 47 farmers.
  • Information neighborhoods from direct communication: link if either farmer answered yes to “Have you ever gone to ___ for advice about your farm?”; median 2 info neighbors among pineapple farmers (max 33). Geographic neighbors within 1 km (median 12, max 25); alternative info metrics: roster of contacts (farm info only/full list) and predicted advice from logit. Financial neighborhoods used as controls.
  • Empirical strategy: model local learning about fertilizer productivity; construct indices of good/bad news at lagged and alternative input levels and index M of input levels associated with good news; estimate logit for whether fertilizer use changes and OLS for magnitude of innovation, controlling for geographic neighbors’ inputs to absorb correlated growing conditions, plus wealth, clan, religion, village, round, experience; spatial GMM standard errors (Conley 1999). Analogous labor regressions for pineapple and maize-cassava.

Limitations

  • Stated: small sample size leads to examining partitions one variable at a time rather than jointly; short panel prohibits including outcomes of lagged own experiments (only 14 farmers have more than 2 fertilizer changes).
  • Stated: data are early in learning process, so value of learning cannot be estimated well; inexperienced farmers often make mistakes, and the paper does not find increases in realized profits for those who change inputs. Labor results are cautioned because labor and fertilizer are not likely separable and are less robust to specification changes than fertilizer results.
  • Stated/evident: current pineapple farmers may be selected on ability to learn, potentially overstating responsiveness of farmers in general; information-neighborhood definitions have drawbacks (ask-advice based on random sample yields smaller-than-actual neighborhoods; roster measures include pairs who may not discuss farming and depend on subjective understanding); endogeneity of neighborhoods from unobservables is an important concern, though predicted-neighborhood results mitigate it. Evident: Monte Carlo check shows M coefficient becomes insignificant when added measurement-error SD is 4 (fertilizer mean 4, SD 7).

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1257/aer.100.1.35

Cites

107 of 107 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 of 10 cited sources checked · 0 supported · 0 flagged

Show all 92 claims

  • It may state two separate claims as one. 59
  • Probably not a substantial claim. 30
  • The quote only partly states this claim. 1
79 of 79

  • The authors state this as an assumption.

    Farmers are trying to learn about the responsiveness of output to fertilizer.

    “The basic form of this model is that farmers are trying to learn about the responsiveness of output y i,t +1 on plot i to a discrete-valued input x i,t which we will call fertilizer:”

  • The authors state this as an assumption.

    The productivity shock is expectation zero and IID across farmers and time.

    “ε i,t +1 is an expectation zero productivity shock that is IID across farmers and time.”

  • The authors state this as an assumption.

    Farmers do not know the production function; it is the object of learning.

    “Farmers do not know the function f ; it is the object of learning.”

  • The authors state this as an assumption.

    Growing conditions are positive, exogenous, influence the marginal product of fertilizer, and are correlated across farmers and time.

    “The variable w i,t is a positive, exogenous growing conditions variable influencing the marginal product of x i,t that is correlated across farmers and time.”

  • The authors state this as an assumption.

    Growing conditions are observable to farmers but not to the econometrician.

    “We assume the w i,t are observable to farmers but not the econometrician.”

  • The authors state this as an assumption.

    Farmers' subjective expectations can be defined.

    “We will not be specific about how the farmer's beliefs evolve, but simply posit that farmer i 's subjective expectations, E i,t ( · ) , can be defined.”

  • The authors state this as an assumption.

    Learning is local: a new observation with input level x only affects g(x) and not g(˜x) for ˜x ≠ x.

    “We consider learning rules that are local in the sense that a new observation with input level x will only affect g i,t +1 ( x ) and not g i,t +1 (˜ x ) for ˜ x = x.”

  • The authors state this as an assumption.

    Farmers are learning in the sense that g increases when they observe higher shock-adjusted output than expected and decreases when they observe lower than expected output.

    “Weassume in addition that farmers are learning in the sense that g i,t +1 ( x ) -g i,t ( x ) &gt; 0 when they observe higher shock-adjusted output than expected ( y w &gt; g i,t ( x )) , and likewise reduce their subjective expectation when they observe lower than expected y w .”

  • The authors state this as an assumption.

    The baseline model assumes farmers are myopic.

    “The time t problem faced by a myopic farmer i is to choose the input, x, to maximize his subjective expectation of time t +1 profits for this input choice”

  • The authors state this as an assumption.

    Inputs are discrete.

    “Since inputs are discrete, this is nothing more than choosing input level x ∗ i,t so that:”

  • The authors state this as an assumption.

    The opportunity cost of capital is the same across farmers.

    “In our theoretical discussion, we incorporate the opportunity cost of capital in the price of fertilizer and assume it is the same across farmers.”

  • The paper relies on this without support.

    A plot planted at time t receives its crucial fertilizer inputs during periods t+1 through t+4, chemically forced flowering occurs approximately at t+5, and the pineapple harvest is generally complete by t+9.

    “A plot planted at time t receives its crucial fertilizer inputs during periods t +1 through t +4 . Chemically forced flowering occurs approximately at t +5 , at which point the eventual size of the crop begins to become apparent. The pineapple harvest is generally complete by t +9 .”

  • The paper relies on this without support.

    Plots have to be harvested on a single day for efficient export of fresh fruit by air to Europe, and it is essential that fruits mature simultaneously, requiring common treatment across plants within the plot.

    “Plots have to be harvested on a single day for efficient export of the fresh fruit by air to Europe. It is essential, therefore, that the fruits mature simultaneously, which requires common treatment across plants within the plot.”

  • The authors state this as an assumption.

    Farmer i increases or decreases his expectation of the productivity of fertilizer at x_k,t according to whether realized profit minus his expectation is positive or negative.

    “Farmer i increases (decreases) his expectation of the productivity of fertilizer at x k,t according to whether y k,t +9 w k,t -g i,t ( x k,t ) is positive (negative).”

  • The authors state this as an assumption.

    Within farmer i's information neighborhood there is a set of H farmers using input level x_k,t sufficiently recently and sufficiently geographically close to k that they have common objective growing conditions.

    “Suppose that within farmer i 0 s information neighborhood there is a set of H farmers using input level x k,t sufficiently recently (say, at t -1 ) and sufficiently geographically close to k that they have common objective growing conditions.”

  • The paper relies on this without support.

    The information conveyed by an information neighbor's planting depends on the number of plants involved.

    “The information conveyed by an information neighbor's planting depends on the number of plants involved, so we weight each observation by Plants k,τ .”

  • The paper relies on this without support.

    The change in beliefs associated with a piece of information depends on the stock of previous experiments that farmer i has observed.

    “The change in beliefs associated with a piece of information depends on the stock of previous experiments that i has observed, so we normalize by the total number of plants i has observed from the beginning of our data until t 1 .”

  • The authors state this as an assumption.

    One fifth of the relevant information about the eventual harvest of a period t planting is revealed in each of the periods t+5 through t+9.

    “ψ ( . ) reflects the idea that 1 / 5 of the relevant information about the eventual harvest of a period t planting is revealed in each of the periods t +5 through t +9 .”

  • The authors state this as an assumption.

    The weighting scheme used to reflect gradual revelation of information is ad hoc.

    “Obviously this weighting scheme is ad hoc ; we examine the robustness of our results to variations in the weighting scheme in Section 6 .”

  • The authors state this as an assumption.

    Farmers are Bayesian learners.

    “so we assume that farmers are Bayesian learners”

  • The authors state this as an assumption.

    In base specifications, information about the eventual size of the harvest of a period t planting began to be revealed after flowering in period t+5, with the outcome fully known at harvest in t+9.

    “Recall that in our base specifications we assumed that information about the eventual size of the harvest of a period t planting began to be revealed after flowering in period t +5 , with the outcome fully known at harvest in t +9 .”

  • The paper relies on this without support.

    The characteristics of the maize-cassava production function are well-known to farmers in these villages and there should be no uncertainty about the role of labor in maize-cassava production.

    “The characteristics of the maizecassava production function are well-known to farmers in these villages and there should be no uncertainty about the role of labor in the production of maize-cassava.”

  • The authors state this as an assumption.

    Good productivity in period 1 is not too bad news about likely productivity in period 2; a sufficient condition is that µ(.|a) first-order stochastically dominates µ(.|b) for a > b.

    “In addition, we need to assume that good productivity in period 1 is not too bad news about the likely productivity in period 2 : a sufficient condition is that µ ( . | a ) fi rst-order stochastically dominates µ ( . | b ) for a &gt; b .”

  • The authors state this as an assumption.

    g1(H) ≥ g1(L).

    “As in the myopic model, this requires g 1 ( H ) ≥ g 1 ( L ) .”

  • The paper relies on this without support.

    The change in growing conditions from period 1 to period 2 is not so strong that yields are doubled.

    “If we do not expect the change in growing conditions from period 1 to period 2 to be so strong that yields are doubled”

  • The paper relies on this without support.

    Each neighbor's planting using a given input level provides the farmer with some information about that level's productivity.

    “Each neighbor's planting from t 0 up to t 1 -1 using a given input level provides the farmer with some information about that level's productivity.”

  • The paper relies on this without support.

    Some information in a planting τ that is 1 to 4 periods before t0 would be revealed after t0 and be news from the perspective of time t1.

    “Some information in the τ planting would be revealed after t 0 and be news from the perspective of time t 1 .”

  • The paper relies on this without support.

    Some of the information in a planting τ that is 1 to 4 periods before t0 should be reflected in the t0 planting but not all of it would be revealed at t0.

    “Some of the information they contain should be reflected in the t 0 planting but it would not all be revealed at t 0 .”
28 of 28

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Fertilizer application window ends in two different periods

    Minor

    The production timeline says fertilizer is applied from t +1 through t +4, but the later description says the window runs from t + 1 through t + 1.

1 of 1

Citations

Cites

Flagged first, then by citations

4 of the 111 references couldn’t be matched to a known paper.

Research works printed but not linked: 1

  1. Moffitt Robert A, 2001, Social dynamics, ed. Steven N. Durlauf and H. Peyton Young, 45

Cited by

Sorted by citations

See all 2,180

Related

See all

Metadata

Authors
Conley, T. G., & Udry, C. R.
Affiliations
  • University of Chicago (US)
  • Yale University (US)
ORCID
Venue
American Economic Review
Publisher
American Economic Association
Published
2010-03-01
Volume
100
Issue
1
Pages
35-69
ISSN
0002-8282 (ISSN-L)
DOI
10.1257/aer.100.1.35
OpenAlex
W2005029869
Semantic Scholar
045f9f5baee3c1f4e6089e4b073dd84d6624c77a
Cited by
2,180
References
111
Citations by year
2026: 143 · 2025: 121 · 2024: 98 · 2023: 135 · 2022: 154 · 2021: 153 · 2020: 132 · 2019: 126 · 2018: 134 · 2017: 117 · 2016: 109 · 2015: 118 · 2014: 130 · 2013: 115 · 2012: 94
FWCI
390.45
Type
journal article
Language
en
Open access
closed
Locations
  • American Economic Review (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • VTechWorks (Virginia Tech) (submittedVersion)
Topics
  • Life Sciences › Agricultural and Biological Sciences › General Agricultural and Biological Sciences › Agricultural Innovations and Practices
  • Social Sciences › Decision Sciences › Management Science and Operations Research › Innovation Diffusion and Forecasting
  • Social Sciences › Political Science and International Relations › Policy Transfer and Learning
Keywords
social learning, pineapple, farming experience, Ghana

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review