Learning by (virtually) doing: experimentation and belief updating - - PowerPoint PPT Presentation

Learning by (virtually) doing: experimentation and belief updating in smallholder agriculture

Emilia Tjernström∗, Travis Lybbert∗∗, Rachel Frattarola Hernández∗∗∗, Juan Sebastian Correa∗

∗University of Wisconsin - Madison, ∗∗UC Davis ∗∗∗OMB

October 24, 2019

Motivation

We think modern inputs could boost smallholder yields For a household to adopt a new technology, they want to know that it exists (and how to use it) something about profitability Other constraints matter for adoption; our design allows us to ignore many of them

Introduction MahindiMaster Data & Experimental design Results 2 / 25

Motivation

Our setting is particularly challenging for learning Substantial heterogeneity in soil quality in Kenya (Tittonell 2008) → information diffusion by central agencies difficult and inaccurate for many → learning from others harder/less beneficial (Munshi 2004; Tjernström, 201?) We still know relatively little about how farmers form and update beliefs

Introduction MahindiMaster Data & Experimental design Results 3 / 25

Motivation

Some organizations have begun experimenting with tailored input recommendations But many issues / open questions remain, including: making test results accessible real-life experimentation is risky at what level to test and recommend?

Introduction MahindiMaster Data & Experimental design Results 4 / 25

Underlying questions

Practical questions: Can we convey soil test results to smallholders in an accessible way? Can we enable farmers to learn from this important information?

Introduction MahindiMaster Data & Experimental design Results 5 / 25

Underlying questions

Practical questions: Can we convey soil test results to smallholders in an accessible way? Can we enable farmers to learn from this important information? → We created MahindiMaster

Introduction MahindiMaster Data & Experimental design Results 5 / 25

Underlying questions

Practical questions: Can we convey soil test results to smallholders in an accessible way? Can we enable farmers to learn from this important information? → We created MahindiMaster Fundamental research question: How does MM change farmers’ understanding of production conditions & optimal inputs? How does MM change (short-term) behavior? → We find evidence of learning consistent with enhanced productivity

Introduction MahindiMaster Data & Experimental design Results 5 / 25

Research questions: details

Formation and evolution of beliefs

do farmers update their beliefs? is this effect stronger for unfamiliar inputs?

Changes in behavior

do farmers change behavior (i.e., does belief updating reflect learning?)

Experimentation within the app

do farmers experiment more with unknown inputs now that cost/risk is lower? does past experience correlate with experimentation?

Introduction MahindiMaster Data & Experimental design Results 6 / 25

Research questions: details

Formation and evolution of beliefs

do farmers update their beliefs? Yes, farmers revise beliefs about returns ↑ is this effect stronger for unfamiliar inputs? Yes (suggestive)

Changes in behavior

do farmers change behavior (i.e., does belief updating reflect learning?) Effects are concentrated among those with high ex ante returns

Experimentation within the app

do farmers experiment more with unknown inputs now that cost/risk is lower? does past experience correlate with experimentation? More experimentation with unfamiliar input among those with high expected returns; not much action on knowledge or confidence, etc.

Introduction MahindiMaster Data & Experimental design Results 6 / 25

Aside on pH and lime

pH affects plant growth – maize likes it slightly acidic but...

if pH is too low, fertilizer will have little to no effect

• ver half of soils in Kenya are low-responsive to nitrogen (Kihara 2016)

→ Can farmers discover this fact by (virtually) experimenting in MM?

Introduction MahindiMaster Data & Experimental design Results 7 / 25

Information interventions

Do information interventions work? Null effects for many information-only interventions: migration (Bryan et al., 2014), college decisions (Bettinger et al., 2012), water purification (Ashraf et al., 2013), student loan take-up (Booij et al., 2012) But some effective interventions: HIV risk info (Dupas, 2011) and earnings info for college majors (Wiswall and Zafar, 2015) What do ineffective info interventions get wrong?

is the info not useful/specific enough? lack of updating? (behavior? over-confidence?) (Dessà and Zhao (2018) already had the information?

Introduction MahindiMaster Data & Experimental design Results 8 / 25

Meet MahindiMaster: behind the scenes

1 Use DSSAT to simulate maize growth 2 Input soil samples from each farmer’s field and construct 3 weather scenarios

based on historical weather in the area

3 Three different fertilizer choices (decreasing order of familiarity): DAP, CAN, lime 4 Discretize fertilizer application rates → a menu of options Introduction MahindiMaster Data & Experimental design Results 9 / 25

Meet MahindiMaster: UI

Wanted to let farmers “query” the model in an interactive, experiential fun way

Introduction MahindiMaster Data & Experimental design Results 10 / 25

Meet MahindiMaster: UI

Wanted to let farmers “query” the model in an interactive, experiential fun way

Introduction MahindiMaster Data & Experimental design Results 10 / 25

Meet MahindiMaster: UI

Wanted to let farmers “query” the model in an interactive, experiential fun way

Introduction MahindiMaster Data & Experimental design Results 10 / 25

Meet MahindiMaster: UI

Wanted to let farmers “query” the model in an interactive, experiential fun way

Introduction MahindiMaster Data & Experimental design Results 10 / 25

Meet MahindiMaster: UI

Farmers play in seasons/rounds, with limited choice at first:

• nly DAP (first 3 rounds) [25kg intervals ∈ [0, 125]]

then CAN introduced [25kg intervals ∈ [0, 125]] finally lime available after 5 rounds [250kg intervals from ∈ [0, 2000]]

Introduction MahindiMaster Data & Experimental design Results 11 / 25

Sample: background

Introduction MahindiMaster Data & Experimental design Results 12 / 25

Sample: background

Introduction MahindiMaster Data & Experimental design Results 13 / 25

Sample: MM pilot

Introduction MahindiMaster Data & Experimental design Results 14 / 25

Data sources (I)

1 3 rounds of panel data from earlier RCT (2013, 2015, 2016)

prior experience with fertilizer input use & yields (soil samples collected in 2013)

2 New soil samples collected in Oct 2016

Introduction MahindiMaster Data & Experimental design Results 15 / 25

Data sources (II)

3 Pre- and post-game data

subjective yield beliefs (no fertilizer, with "typical" fertilizer, with fertilizer + lime) allocate a budget between DAP, CAN and lime farming knowledge quiz (→ confidence) [only pre-]

Get moments of distribution of beliefs by fitting lognormal CDF using nonlinear least squares Confidence: guesses about how many questions they got correct 4 Interactions with MahindiMaster

# of rounds # of rounds they experiment with different fertilizers

Introduction MahindiMaster Data & Experimental design Results 16 / 25

Estimation

Posti = α + βPrei + γTraiti + ǫi (1) Posti = α + βPrei + γ1pHi + γ2pH2

i + ui

(2) Posti = α + βPrei + φk

5

• k=1

pHk

i + ui

(3) pHk

i : dummy variables indicating farmer i’s pH is in one of five ranges of pH

(pH < 5.5, pH∈ (5.5, 6), pH∈ (6, 6.5), pH∈ (6.5, 7), &r pH> 7)

Introduction MahindiMaster Data & Experimental design Results 17 / 25

Descriptive statistics: farmer characteristics

Table: Farmer Characteristics Mean Std Deviation Min Max

• No. of seasons used fertilizer (long rains)

2.06 2.29 5

• No. of seasons used hybrids (long rains)

2.10 2.15 5 Uses any DAP in a “normal” year 0.49 0.50 1 Uses any CAN in a “normal” year 0.38 0.49 1 Uses any lime in a “normal” year pH of sampled plot 6.45 0.70 4.93 8.65 CEC of sampled plot 25.3 16.3 5.91 68.6 Sampled plot size (acres) 1.15 0.92 0.12 5

• No. farming quiz questions correct

2.93 1.12 6

• No. farming quiz questions=don’t know

4.87 2.04 9 Overconfident (0/1) 0.59 0.49 1

Introduction MahindiMaster Data & Experimental design Results 18 / 25

Descriptive statistics: game play

Table: Game play

Mean Std Deviation Min Max Amount of DAP (kg) across rounds 31.8 17.1 7.14 95 Amount of CAN (kg) across rounds 23.8 14.2 85 Amount of Lime (kg) across rounds 118.2 95.1 611.1 Yields (kg/acre) obtained in game 1343.8 498.4 419.3 2765.3 Share of rounds with DAP 0.78 0.23 0.20 1 Share of rounds with CAN (conditional on avail.) 0.81 0.25 1 Share of rounds with lime (conditional on avail.) 0.60 0.31 1 Share of rounds with no fertilizer 0.046 0.062 0.25 Random rainfall scenario (1=poor, 2=normal, 3=good) 2.01 0.30 1.25 2.80 Chosen rainfall scenario (1=poor, 2=normal, 3=good) 2.70 0.44 1 3 Rounds played 10.6 2.93 1 19 N 175

Introduction MahindiMaster Data & Experimental design Results 19 / 25

Descriptive statistics: subjective expectations

Table: Subjective expectations pre- and post-game

Mean (pre) Mean (post) t-statistic Mean (no fertilizer) 3.31

• Mean (fertilizer)

5.47 8.64 7.95 Mean (fertilizer + lime) 5.51 9.74 9.06 CV (fertilizer) 0.64 0.64 1.55 CV (fertilizer and lime) 0.64 0.64 1.72 N 175

Introduction MahindiMaster Data & Experimental design Results 20 / 25

Descriptive statistics: fertilizer orders

Table: Fertilizer orders

Mean (pre) Mean (post) t-statistic DAP (kg) 36.9 26.5

• 6.45

CAN (kg) 40.1 40.5

• 0.13

Lime (kg) 54.9 96.8 2.26 DAP (KHS), pre 2448.1 1713.1

• 5.92

CAN (KHS), pre 1965.2 2050.6 0.53 Lime (KHS), pre 345 612.9 2.19

Introduction MahindiMaster Data & Experimental design Results 21 / 25

• Amt. (kg)
• Amt. (kg)

Share Share (post-pre) (post-pre) pH

• 431.68***
• 61.60***
• 1877.06

(152.85) (23.27) (1433.93) pH2 28.06** 4.06** 119.28 (11.35) (1.74) (104.39) 5.5 < pH <= 6

• 33.43
• 5.83

288.33 (43.31) (6.24) (416.17) 6 < pH <= 6.5

• 128.53***
• 17.86***
• 278.88

(39.65) (5.89) (396.65) 6.5 < pH <= 7

• 133.37***
• 18.20***
• 414.52

(40.63) (6.03) (419.64) pH > 7

• 153.79***
• 20.71***
• 450.95

(40.03) (5.99) (408.56) Kg lime (pre) 0.07 0.06 (0.08) (0.08) Lime share (pre) 0.07 0.05 (0.10) (0.11) Intercept 1696.28*** 201.23*** 239.34*** 28.15*** 7350.35 486.67 (511.75) (37.59) (77.55) (5.61) (4880.31) (379.58) R2 0.16 0.18 0.13 0.15 0.05 0.07 N 167 167 167 167 167 167 Mean of dep. var: 98.41 98.41 13.78 13.78 270.24 270.24

Results: pH and lime orders

Figure: Average share of post-game order allocated to inputs, by soil pH

10 20 30 40 50 60 70

Average share of post-game order to DAP

pH <= 5.5 5.5 < pH <= 6 6 < pH <= 6.5 6.5 < pH <= 7 pH > 7

pH DAP (A)

10 20 30 40 50 60 70

Average share of post-game order to CAN

pH <= 5.5 5.5 < pH <= 6 6 < pH <= 6.5 6.5 < pH <= 7 pH > 7

pH CAN (B)

10 20 30 40 50 60 70

Average share of post-game order to lime

pH <= 5.5 5.5 < pH <= 6 6 < pH <= 6.5 6.5 < pH <= 7 pH > 7

pH Lime (C)

Introduction MahindiMaster Data & Experimental design Results 23 / 25

Results: do some farmers respond more to the intervention?

DAP CAN Lime DAP CAN Lime (post-pre) (post-pre) (post-pre) (post-pre) (post-pre) (post-pre)

• No. questions correct
• 146.28*

141.73 147.41***

• 179.00**

147.71 87.20 (75.59) (90.86) (53.17) (89.97) (105.05) (71.11)

• No. questions=‘Don’t know’
• 32.13

5.89

• 58.53

(55.50) (62.14) (43.64) DAP value (Pre)

• 0.78***
• 0.79***

(0.15) (0.15) CAN value (Pre)

• 0.97***
• 0.97***

(0.20) (0.20) Lime value (Pre)

• 0.87***
• 0.89***

(0.10) (0.10) Intercept 1600.28*** 1585.49*** 178.10 1868.42*** 1543.86** 641.24 (472.20) (569.52) (168.53) (632.92) (672.79) (398.78) R2 0.13 0.15 0.33 0.13 0.15 0.34 N 158 158 158 158 158 158

Introduction MahindiMaster Data & Experimental design Results 24 / 25

Conclusions

Smallholders face many constraints, but learning about returns is fundamental We think our results are encouraging but many issues open questions remain How to scale cost-effectively? Coordinating with input suppliers?

Introduction MahindiMaster Data & Experimental design Results 25 / 25