Learning About a New Technology: Pineapple in

Ghana

Tim Conley, University of Chicago

and

Chris Udry, Yale University

Pro

porti

on o

f Far

mer

s C

ultiv

atin

g P

inea

pple

'experienced' 'inexperienced'1990 1992 1994 1996 1998

.1

.2

.3

.4

.5

Figure 3. Sample proportion of farmers cultivating pineapple by year. Wedefine famers who adopted before 1994 as experienced.

Goal:

• Understand Learning About New Technology in

Agrarian Economy: Rural Ghana

Background

• Once Upon a Time: Farmers Grew Maize and

Cassava

• New Thing: Price of Pineapple Went Up→ Shift

to Producing Pineapple for Export

Question:

• Is Social Learning Important As Ghanaian Farm-ers Learn How to Farm Pineapple?

Pineapple Issues: Costs/Benefits

• Profitability Relative to Old Crops

• Barriers to Entry

— Capital, Land, Labor Mkt Imperfections

∗ Min. Efficient Size for Exporters

∗ Land Characteristics Matter: Goldstein andUdry (1999)

∗ Hired Labor Needed

∗ Inputs More Expensive

— Costs of Learning How to Use New Tech.

Initial Adoption and Learning Best Inputs Clearly NOT

Separate, But Focus on Latter Here

Data

• Demographic: Family Characteristics

• Geographic: Soil Characteristics, Plot Locations

• Actions: Which Crops to Plant, Use of Fertilizer,Herbicide, Pesticide

• Outcomes: Yields, Profits, Farmer Forecast Val-uation of Unharvested Crop at End

• Knowledge and Communication: Who FarmersKnow, Learned From, Talk To About a Problem

Emprically Identifying Social Learning is Hard

• Operationally Defining What Farmers Learn From:Neighbors and Observed Info

• Identify Learning v. Other Stuff: Correlated Shocks,Mimicry, etc.

Modeling Who Is My Neighbor and What Can I See?

• Whole Village, Collective Experimentation

• Sparse Network, Limited Observability

Knowledge Info

Individual’s Roster of Contacts

• Lots of Types of Interactions

Questions About Random Six Pineapple Plots

• Do You Know the Owner of this Plot?

• Do Whether It Is A Pineapple Plot?

• What Inputs Used on It?

• What Was Harvested?

Questions About Random Seven Farmers + ‘Big Three’

• Do You Know Farner A?

• When Did You Last Talk With Farmer A?

• Ever You Ever Gone to Farmer A for Advice?

Roster Connections and Average Pineapple Plot Coordinates Village 3

1 Km

Figure 1. The circles represent the geographic center of plots for eachpineapple farmer in Village 3. The lines connect pineapple farmerslisted on each other’s roster of contacts.

Ker

nel D

ensi

ty E

stim

ate

Distance in Meters

No Information Link Information Link

0 2000 4000 6000

0

.2

.4

.6

Figure 2. Kernel density estimates for the distribution of distances betweenall pineapple farmers and those listed on the roster of contacts. AnEpanechnikov kernel with a bandwidth of 300 meters was used.

Table 2: Logit Predicting Ask For Advice

Coefficient Standard Error

Either Party Holds Traditional Office -0.55 0.26Same Religion 0.04 0.33Same Clan 0.43 0.24Same Gender 1.73 0.78Same Soil Type -0.23 0.27Absolute Age Difference (years) -0.04 0.02Absolute Wealth Difference (million cedis) 0.15 0.03Distance Between Plot Centers (kilometers) 0.46 0.16Constant -2.14 0.84

Logit MLE Estimates, Sample Size = 490, Pseudo R-squared =.12.Dependent variable is one if either party answered yes to the question:Have you ever gone to _____ for advice about your farm?

Production Function: yi,t+1 = witf(xit) + εi,t+1

• yitt+1 = pineapples

• xi,t = fertilizer (discrete) with price pit

• wi,t is a weather correlated in space/time known

to farmers not us

• εi,t+1 IID productivity shock.

Myopic Farmer’s Problem: Choose the Best Input

Given Subjective Expectation:

Eit{witf(x∗i,t, εi,t+1)}− pitx

∗it ≥

Eit

nwitf(x, εi,t+1)

o− pitx

Farmer Needs to Learn Productivity of Input Levels

Eit

nf(xkt, εk,t+1)

o= Eit

nyk,t+1wkt

|xkto

Learn Expected “Weather-adjusted Output” Given X

git(xkt) ≡ Eit

nyk,t+1wkt

|xkto

• Learning = Updating g

— Using Observations of (yi,t+1, wit, xit)

— Own and Those of Info. Neighbors

• Farmer Chooses x∗ so:

Eit{witf(x∗i,t, εi,t+1)}− pitx

∗it ≥

Eit

nwitf(x, εi,t+1)

o− pitx

• Reasons for different farmers to choose same x∗

— Prices for fertilizer (Constant, Common)

— Weather shocks (Spatial, Time Correlation)

— Related Information (Can Measure Neighbors)

Local Learning Definition

Update Only One Conditional Distribution at a Time

(y, x,w)→ Update gi,t(x) not gi,t(x) for x 6= x

Update In Direction of Observations:

yw ≥ git(x)→ git+1(x) ≥ git(x)

Example: Sample Analog Updating:

git(x) =1P

j∈Ni,τ<t−11(xj,τ=x)

Pj∈Ni,τ<t−1

1(xjτ = x)yj,τ+1wj,τ

gi,t+1(x) = αit(x) · yk,t+1wkt+ (1− αit(x))git(x)

αit(x) = (1 +P

j∈Ni,τ<t−11(xjτ = x))−1

gi,t+1(x)− gi,t(x) = αit(x) ·hyk,t+1wkt

− git(x)i

Implication for Actions

• Good News About Input Level x:

nyk,t+1wkt

≥ git(x)oornπk,t+1(x) ≥ Ei,tπi,t+1(x)

ogi,t+1(x) ≥ gi,t(x) andEi,t+1πi,t+2(x) ≥ Ei,tπi,t+1(x)

Move Weakly To x

• Bad News About x

nyk,t+1wkt

< git(x)oornπk,t+1(x) < Ei,tπi,t+1(x)

oEi,t+1πi,t+2(x) < Ei,tπi,t+1(x)

Perhaps Switch Away If at x, Else Stay Put

• Random weather

Responsiveness is probabilistic when weather is ran-

dom. Let supp(x) = {H,L} with H > L. Farmer i

chooses xit = H if witgit(H) − pitH ≥ witgit(L) −pitL. xit = H if

wit ≥ w∗it =pit(H − L)

git(H)− git(L).

w∗it is strictly declining in git(H), so the probability

given git(·) and all past wkt that i will choose xit = H

is increasing in git(H). Thus, the conditional proba-

bility that i will choose xit = H is increasing in ykt.

similar pattern in the absence of learning if weather

is not properly taken into account. suppose git(·) isknown. Farmers will choose xi,t = H when

wit ≥p(H − L)

g(H)− g(L).

this will cause and yi,t+t to be high relative to its un-

conditional expectation. If wk.t is positively spatially

and serially correlated, then farmers choices of xi,tand hence yi,t+t will be positively correlated across

space and time as well. In particular, if weather shocks

at small lags in time are highly positively correlated

for physically proximate plots, then farmer k’s choices

xk,t−1 = H and associated likely-to-be-high output

yk,t will be tend to be followed by choices of H and

higher outputs of k0s physical neighbors, solely due tothe positive dynamic correlations in weather. To the

outside observer, a higher than long-run average real-

ization of yields and profits by farmer k using the high

fertilizer level tends to be followed in near future by

an increased use of that quantity of fertilizer by his

physical neighbor farmer j who also tends to achieve

a higher than long-run average profit.

• Caveats:

— Global Knowledge of Production Fn Might Change

Predictions

— Know w Role in Production Fn Implies Can

Learn From All Neighbors, Not Just Those

With Matching w

Social Learning and Experimentation

Simplest setting - two periods. Expected profits are

V2 = maxj∈{H,L}

E2³w2θj − q2 · j + ε3

´.

So the farmer chooses x2 = H if w2µ2H − q2 ·H >

w2µ2L − q2 · L. So value of H in 1 is

V1H = w1µ1H − q1 ·H + δE1V2

= w1µ1H − q1 ·H +

δE1

"max

j∈{H,L}E2

³w2θj − q2 · j + ε3

´#= w1µ1H − q1 ·H + δE1

max

"w2

Ãy2H

σ21Hσ21H + 1

+ µ1H1

σ21H + 1

!− q2 ·H ,

w2µ1L − q2 · L] .

y2H =σ21H + 1

σ21H

õ1L +

q2w2(H − L)− µ1H

1

σ21H + 1

!

V1H

= w1µ1H − q1 ·H +

δZ Z

max

"w2

Ãy2H

σ21Hσ21H + 1

+ µ1H1

σ21H + 1

!− q2 ·H

w2µ1L − q2 · L]× f(w2|w1)× φ1H(y2H)dy2Hdw2,

= w1µ1H − q1 ·H +

δZ Z ∞

y2H

Ãw2

Ãy2H

σ21Hσ21H + 1

+ µ1H1

σ21H + 1

!− q2 ·H

!×f(w2|w1)φ1H(y2H)dy2Hdw2+δ

Z Z y2H

−∞(w2µ1L − q2 · L)

×f(w2|w1)φ1H(y2H)dy2Hdw2∂V1H∂µ1H

= w1 + δZw2(1−Φ1H(y2H))f(w2|w1)dw2,

∂V1L∂µ1H

= δZw2Φ1L(y2L)f(w2|w1)dw2

Differentiate the value function to find

1

δ

∂V1L∂µ1H

=Z Z yi2L

−∞w2f(w2|w1)φ1L(y2L)dy2Ldw2 +Z

(w2µ1H − q2H)φ1L(y2L)∂y2L∂µ1H

f(w2|w1)dw2 −Z Ãw2

(y2L

σ21Lσ21L + 1

+ µ1L1

σ21L + 1

)− q2L

!

×φ1L(y2L)∂y2L∂µ1H

f(w2|w1)dw2where y2L is defined (analogously to y2H) as the crit-

ical realization of y2 given that x1 = L such that if

y2L ≥ y2L, then x2 = L and x2 = H otherwise.

Thus y2L is defined so that

(w2µ1H − q2H) =

Ãw2

(y2L

σ21Lσ21L + 1

+µ1L

σ21L + 1

)− q2L

!.

Given the definition of y2L, the difference of the sec-

ond and third lines of the derivative calculated above

is zero and equation (??) follows.

Can show

(1− Φ1H(y2H)) > Φ1L(y2L)−1

2.

(1−Φ1H(y2H)) is the probability that x2 = H given

that the farmer chooses x1 = H. So ∂V1H∂µ1H

− ∂V1L∂µ1H

= w1 + δZw2(1−Φ1H(y2H)−

Φ1L(y2L))f(w2|w1)dw2> w1 + δ

Z−12w2f(w2|w1)dw2

= w1 −δ

2E(w2|w1) > 0

• Point: Social learning implies that farmers will in-novate in their fertilizer use after the realization

of particularly high profits by their information

neighbors. However, this will also be observed

across geographic neighbors if we don’t condition

on weather. Use geographic proximity to condi-

tion for weather.

• Also Investigate

— Any Model with Learning Convergence: Up-

dates Decreasing in Some Measure of Experi-

ence

— Does Source of Information Influence Responses?

Input Responses Are Fn ofhyk,t+1wkt

− git(x)ior Profits

Analog:nπk,t+1(xk,t)−Ei,t

hπk,t+1(xk,t)|xk,t, wk,t

ioProblems:

• No Weather Data

— Scale ofhyk,t+1wkt

− git(x)inot identified

• Short Time Series Dimension

— Cannot Estimate Subjective git(·)

Solutions:

• Exploit Spatial, Temporal Correlation to IdentifySign

• Use Rational Expectations Analog of Subjectivegit(·)

Event We Care About:nπk,t+1(xk,t) > Ei,t

hπk,t+1(xk,t)|xk,t, wk,t

io

ApproximateEi,t

hπk,t+1(xk,t)|xk,t, wk,t

iwith estimate

of Ehπk,t+1(xk,t)|xk,t, wk,t

i

Estimate Ehπk,t+1(xk,t)|xk,t, wk,t

ifor unknown w

exploiting assumption that w is spatially, serially cor-

related.

Suppose info neighborhood of farmer i contains H

neighbors close to k with common weather and input

bE hπk,t+1(xk,t)|xk,t, wk,t

i= 1

H

PHh=1 πh,t+1(xh,t)

Use Event:nπk,t+1(xk,t) >

bE hπk,t+1(xk,t)|xk,t, wk,t

io

di,k,t ≡ 1nπk,t+9(xk,t, wk,t) >

bEi

hπk,t+9(xk,t, wk,t)

io.

G∗i,t1(x = xi,t0) ≡Xk∈Ni

Xτ∈[t0,t1)

ψ(t1 − τ)di,k,τ1{xi,t0 = xk,τ}.

ψ(τ ; t0, t1) ≡ (ψ(t1 − τ)− ψ(t0 − τ))

G∗∗i,t1(x = xi,t0) ≡Xk∈Ni

Xτ∈[t0−4,t1)

ψ(τ ; t0, t1)di,k,τ1{xi,t0 = xk,τ}.

Gi,t1(x = xi,t0) ≡1

TotalP lantsi,t1Xk∈Ni

Xτ∈[t0−4,t1)

ψ(τ ; t0, t1)di,k,τP lantsk,τ1{xi,t0 = xk,τ}.

1Ni

Pk∈Ni

1

Ãπk,t(xk,t−1) >bE h

πk,t(xk,t−1)|xk,t−1, wk,t

i ! hxk,t−1 − x∗i,t−1i

• Problem: Similarity in Weather Induces Similarityin Actions

• “Solution” Use Spatial Average of Actions to ProxyFor Alignment Due to Weather

1Gi

Pk∈Gi

hxk,t−1 − x∗i,t−1

i

• Problem: Similarity in financial conditions inducesimilarity in actions

• ”Solution” use average within financial neighbor-hoods:

1Fi

Pk∈Fi

hxk,t−1 − x∗i,t−1

i

Regression Specification:

∆xi,t = ω0i,tα+

β 1Ni

Pk∈Ni

1

Ãπk,t(xk,t−1) >bE h

πk,t(xk,t−1)|xk,t−1, wk,t

i ! hxk,t−1 − x∗i,t−1i

+γ 1Gi

Pk∈Gi

hxk,t−1 − x∗i,t−1

i+ δAi,t + εi,t

Estimate via LS Using Spatial SE Allowing General

Correlation

Operational Defnitions:

• Close

bE hπk,t(xk,t−1)|xk,t−1, wk,t

i: within 700 meters and

Lagged ≤ 2

Gi : within 700 meters

• Spatial Correlation Bartlett Product Weight 1.5km Cutoffs

• New Farmers Exp≤ 3 (27%)

• Big Farmers 60,000+ suckers (27 %) (Median

suckers 22,000, mean 41,000). .

More Issues:

• Nonstationarity in updating changes not here: can-not improve with short TS

• Own experience poorly measured: some may bepicked up by 1

Gi

Pk∈Gi

hxk,t − x∗i,t−1

i

• Information neighbors may also share common ac-cess to credit arrangements: can vary Ai,t.

• Endogenous neighborhoods due to something un-observed that also influence productivity

Table 4: Predicting the Change in Input Use

Dependent Variable: Indicator of a Change in Per Plant Fertilizer Use

Good News at Lagged Fertilizer Use -0.13 0.05[1.19] [1.16]

Good News at Alternative Fertilizer Use 0.18 0.37[0.97] [1.02]

Bad News at Lagged Fertilizer Use 12.32 14.41[3.72] [4.63]

Bad News at Alternative Fertilizer Use -2.98 -4.22[1.91] [2.07]

Average Absolute Deviation from 0.49 0.49Geographic Neighbors' Fertilizer Use [0.13] [0.14]

Inexperienced Farmer 1.14 1.28[0.92] [1.04]

Talks with Extension Agent -1.39[0.75]

Village 1 -2.34 -1.70[0.88] [1.06]

Village 2 -1.96 -2.55[0.99] [0.95]

Wealth (Million Cedis) -0.04 0.04[0.13] [0.11]

Clan 1 2.05 2.39[1.07] [1.02]

Clan 2 2.72 2.88[0.94] [0.84]

Church 1 -0.18 -0.68[0.91] [0.91]

BA

Logit MLE point estimates, spatial GMM (Conley 1999) standard errors in brackets allow for heteroskedasticity and correlation as a function of physical distance, see footnote 25 for details. Sample Size = 107. Pseudo R-squareds .34 and .36, columns A and B respectively. A full set of round dummies were included but not reported. Information neighborhoods defined using responses to: Have you ever gone to farmer ____ for advice about your farm?

Table 5: Predicting Innovations in Input Use, Differential Effects by Source of Information

Dependent Variable: Innovation in Per Plant Fertilizer UseA

Index of Inputs on Successful Experiments (M) 0.99[.16]

M * Inexperienced Farmer 1.09[0.22]

M * Experienced Farmer 0.10[0.32]

Inexperienced Farmer 4.01 4.20 4.22 4.19 4.12[2.62] [2.66] [2.65] [2.65] [2.77]

Index of Experiments by Inexperienced Farmers -0.13[0.37]

Index of Experiments by Experienced Farmers 1.02[0.17]

Index of Exper. by Farmers with Same Wealth 1.03[0.18]

Index of Exper. by Farmers with Different Wealth -0.41[0.32]

Index of Experiments on Big Farms 1.10[0.14]

Index of Experiments on Small Farms 0.89[0.18]

Index of Exper. by Farmers with Same Soil 1.04[0.16]

Index of Exper. by Farmers with Different Soil 0.91[0.19]

Avg. Dev. of Lagged Use From Geographic Nbrs 0.54 0.55 0.58 0.58 0.58 0.59[0.06] [0.08] [0.06] [0.06] [0.06] [0.06]

Avg. Dev. of Lagged Use From Financial Nbrs 0.53 0.45 0.40 0.43 0.22 0.24[0.58] [0.58] [0.59] [0.55] [0.61] [0.60]

Village 1 -7.62 -7.92 -8.09 -8.24 -7.81 -7.88[1.16] [1.43] [1.36] [1.43] [1.31] [1.31]

Village 2 -0.61 -1.82 -2.15 -2.17 -1.83 -1.78[1.56] [2.02] [2.03] [2.11] [2.02] [2.07]

Wealth (Million Cedis) 0.13 0.36 0.41 0.45 0.29 0.29[0.25] [0.17] [0.17] [0.17] [0.20] [0.20]

Clan 1 -2.62 -2.42 -2.68 -2.62 -2.53 -2.55[1.29] [1.21] [1.12] [1.09] [1.11] [1.15]

Clan 2 -0.40 -0.11 -0.11 -0.15 -0.31 -0.29[1.44] [1.32] [1.32] [1.32] [1.30] [1.30]

Church 1 0.26 0.67 0.76 -0.60 0.87 0.88[1.29] [1.12] [1.06] [1.11] [1.12] [1.12]

R-squared 0.70 0.73 0.71 0.71 0.71 0.71OLS point estimates, spatial GMM (Conley 1999) standard errors in brackets allow for heteroskedasticity and correlation as a function of physical distance, see footnote 20 for details. Sample Size = 107. A full set of round dummies included but not reported. Information neighborhoods defined using responses to: Have you ever gone to farmer ____ for advice about your farm?

FB C D E

Table 6: Alternate Definitions of the Information Network

Dependent Variable: Innovation in Per Plant Fertilizer Use

B

Information Neighborhood Metric

Roster of Contacts: Full

Set of Contacts

M * Inexperienced Farmer 1.50 1.49 6.34[0.28] [0.28] [1.14]

M * Experienced Farmer 0.19 0.15 4.52[0.21] [0.22] [1.80]

Inexperienced Farmer 4.66 4.65 4.01[2.84] [2.84] [2.77]

Average Deviation of Lagged Use From Geographic Neighbors' Use 0.49 0.49 0.33

[0.09] [0.09] [0.12]

Average Deviation of Lagged Use From Financial Neighbors' Use 0.50 0.51 0.59

[0.69] [0.70] [0.82]

Village 1 -7.59 -7.52 -9.25[1.64] [1.63] [1.75]

Village 2 -2.09 -2.08 -1.86[2.11] [2.10] [2.07]

Wealth (Million Cedis) 0.35 0.35 0.16[0.17] [0.17] [0.22]

Clan 1 -2.25 -2.23 -1.66[1.37] [1.37] [1.28]

Clan 2 -0.02 0.01 0.42[1.41] [1.40] [1.33]

Church 1 0.62 0.59 0.75[1.22] [1.23] [1.23]

R-squared 0.72 0.72 0.73

OLS point estimates, spatial GMM (Conley 1999) standard errors in brackets allow for heteroskedasticity and correlation as a function of physical distance, see footnote 20 for details. Sample Size = 107. A full set of round dummies were included but not reported. Alternative information neighborhoods are as defined in Section 3.2 and Appendix 2.

CRoster of Contacts: Farm Info

Only

Predicted Advice

A

Table 7: Robustness to Changes in SpecificationDependent Variable: Innovation in Per Plant Fertilizer Use

F

Lagged Fertilizer Use

M * Inexperienced Farmer 0.93 0.39 1.03 1.85 1.11 0.34[0.19] [0.15] [0.19] [0.20] [0.27] [0.13]

M * Experienced Farmer 0.16 -0.14 -0.41 0.04 -0.24 0.08[0.23] [0.24] [0.35] [0.23] [0.45] [0.31]

Inexperienced Farmer 4.38 5.83 4.01 2.87 5.94 4.05[2.73] [3.46] [2.71] [2.71] [2.72] [2.62]

Lagged Own Fertilizer Use -0.84[0.22]

0.51 0.95 0.58 0.10 0.50 0.09[0.08] [0.09] [0.08] [0.06] [0.12] [0.17]

0.36 -0.17 0.51 1.06 0.64 0.59[0.54] [0.40] [0.55] [1.16] [0.70] [0.61]

Village 1 -7.51 -7.96 -8.13 -7.48 -13.46 -3.19[1.35] [1.58] [1.51] [2.10] [3.03] [1.55]

Village 2 -2.02 -2.83 -1.93 -1.40 -1.68 -2.70[2.09] [2.71] [2.09] [2.08] [2.37] [2.11]

Wealth (Million Cedis) 0.36 0.50 0.41 0.24 0.73 0.21[0.18] [0.18] [0.17] [0.20] [0.21] [0.19]

Clan 1 -2.30 -2.58 -2.45 -2.71 -4.04 -1.17[1.16] [1.35] [1.35] [1.21] [1.87] [1.32]

Clan 2 -0.07 0.16 -0.004 -0.62 0.25 0.71[1.28] [1.31] [1.35] [1.47] [1.39] [1.29]

Church 1 0.56 1.33 0.62 0.43 1.58 0.36[1.14] [1.17] [1.13] [1.36] [1.58] [1.16]

Soil Organic Matter 0.14[0.67]

Soil pH 4.09[2.31]

Soil Type = Loam 1.40[1.14]

Soil Type = Sandy -5.78[2.72]

Sample size 107

R-squared 0.74 0.72 0.73 0.68 0.80 0.75

Avg. Dev. of Lagged Use From Geographic Neighbors' Use

Avg. Dev. of Lagged Use From Financial Neighbors' Use

A B

No Information from Plantings

at t-1

Soil Charac-teristics

C

No Information from Plantings

at t-1 or t-2

Fertilizer Categories: Zero, Med, High (High

> 80th percentile)

Geographic Neighborhood within 500m

D E

OLS point estimates, spatial GMM (Conley 1999) standard. errors in brackets allow for heteroskedasticity and correlation as a function of physical distance, see footnote 20. Round dummies included but not reported. Alternative specifications are as defined in Section 6.2.

107 89107 93 107

Table 8: Predicting Innovations in Labor for Pineapple and Maize-Cassava Plots

Dependent Variable: First Difference in Labor Inputs for Pineapple and Maize-Cassava

0.95 0.04[.37] [0.14]

0.25[0.14]

0.48 0.83[0.22] [0.14]

-0.27 0.02[0.23] [0.11]

Village 1 770.96 -195.98 -203.94[238.43] [100.54] [98.52]

Village 2 567.83 -417.61 -389.13[316.37] [119.34] [134.16]

Village 3 -126.29 -228.58[89.21] [132.30]

Wealth (Million Cedis) 47.59 1.09 -14.05[31.46] [33.42] [37.69]

Clan 1 -802.58 213.44 -794.38[358.62] [566.83] [450.66]

Clan 2 369.36 -45.33 -3.19[232.41] [82.65] [78.85]

Church 1 93.85 -37.86 -4.69[179.96] [77.80] [86.46]

Soil Organic Matter -259.11 -99.26 -72.00[202.53] [90.41] [40.58]

Soil pH 358.52 15.19 88.98[141.46] [131.66] [58.55]

Soil Type = Loam -641.90 -97.95 -105.23[197.89] [45.75] [78.08]

Soil Type = Sandy -909.36 128.06 -76.96[382.93] [66.14] [126.10]

Sample size 346R-squared 0.55 0.42 0.24

C Maize-Cassava

(labor cost in 1000 cedis per hectare)

A BPineapple Maize-Cassava

OLS point estimates, spatial GMM (Conley 1999) standard errors in brackets allow for heteroskedasticity and correlation as a function of physical distance, see footnote 20 for details. Round/season dummies included but not reported. Information neighborhood from: Have you ever gone to farmer ____ for advice about your farm?

Crop (labor cost in cedis per plant)

(labor cost in 1000 cedis per hectare)

89 346

Index of Experiments in the Geographic Neighborhood

Average Deviation of Lagged Use From Geographic Neighbors' Use

Average Deviation of Lagged Use From Financial Neighbors' Use

Index of Experiments: M-tilde

Table 1: Descriptive Statistics

Mean Std. Deviation

Fertilizer Use (cedis per sucker) 1.938 5.620

Change in Fertilizer Use -0.315 10.335Indicator of Change in Fertilizer ≠ 0 0.496 0.502Indicies of Successful Experiments: M Ask advice 0.114 4.312 M Talk frequently -0.041 4.259 M Know each other's plots -0.033 3.308 M Roster of contacts, farm info only -0.174 3.467 M Roster of contacts, full list -0.126 3.469 M Predicted ask for advice -0.020 0.940

Wealth (million cedis) 2.331 2.835Clan 1 Indicator 0.327 .Clan 2 Indicator 0.451 .Church 1 Indicator 0.487 .pH 5.952 0.738Soil Organic Matter (%) 2.927 1.110Loamy Soil Indicator 0.434 .Sandy Soil Indicator 0.053 .

Contact with Extension Agent Indicator 0.327 .

Inexperienced Farmer Indicator 0.230 .

Avg. Dev. of Lagged Use From Geographic Neighbors' UseAvg. Dev. of Lagged Use From Financial Neighbors' Use

0.498 7.790

-0.053 0.380

Table 3: Information Connections by Cohort of Pineapple Adoption

Proportion of pairs of individuals in each other's information neighborhood

Not Farming Pineapple

Inexperienced Pineapple

Farmer

Experienced Pineapple

Farmer Neighorhood Metric

Not Farming Pineapple 0.06 0.05 0.07Inexperienced Pineapple Farmer 0.05 0.09 0.13Experienced Pineapple Farmer 0.07 0.13 0.21

Response to "Have you ever gone to ____ for

advice about your farm?"