Technology Systems in Crop Agriculture: Case of Complementarity - - PowerPoint PPT Presentation

Technology Systems in Crop Agriculture: Case of Complementarity between Soybean Seed and Tillage Intensity

Ed Perry, Graduate Student Iowa State University, GianCarlo Moschini, CARD & Iowa State University, David Hennessy

11/24/2015

Work partly funded by NIFA grant Perry’s fellowship funded by NIFA National Needs

Issue

• Row crop agriculture in United States has seen many

innovations in recent decades. These include – Capital intensification & mechanical innovation, substituting out labor – Data as inputs & outputs – Rotation simplification & regional specialization – Reduced till & no till cultivation – Genetically modified seeds

• Our interest is in the last two and how in one sense

they interact

• Our viewpoint, not new, is that cropping practices

come as systems and not as stand-alone practices

2

GM Seeds of interest

• Focus: glyphosate tolerant (GT) soybeans
• Can apply herbicide over the crop, killing other plants on

contact

• Reduce costs in form of

– Other pesticides – Spray runs so lower labor, fuel, machinery depreciation – Management time & effort, scouting for weeds, etc.

• Yield impact now negligible
• Seeds more expensive
• Environmental issues: More glyphosate used. Less use of
• ther toxic herbicides. Weed resistance developing.

Impacts on conservation tillage

3

4

From Fernandez-Cornejo et al. Amber Waves, March 2014

Cultivation Practice of Interest

• Conservation tillage (CT): any

soil cultivation method that leaves previous year's crop residue on fields before and after planting next crop. Reduces soil erosion & nutrient runoff, enhances carbon sequestration. No till: plant crops directly into debris from previous crop Reduced till: some seedbed preparation. Cultivation, plowing, disking, etc. kills weeds mechanically Conservation tillage is easier with roundup- ready crops

5

Newly planted soybeans in corn residue. Photo courtesy USDA NRCS

Research question

• Do glyphosate tolerant (GT) soybean seed and

reduced/no till cropping practices complement?

• If they do then soybean seed contribute to some

environmental benefits that may go away upon – Burdensome regulation on GM seed industry – Impediments to export markets for GM products – Poor management of gene flow and resistance dynamics – Market restructuring such that seed prices rise

• Further implication: policies promoting CT would

promote GT soybean and vice versa

6

Prior research

• Positive correlation between GT crops & CT

commonly agreed. Nature of correlation not understood

• Cotton:

– Roberts et al. (2006), Frisvold et al. (2009), Kalaitzandonakes & Suntornpithug (2003) conclude in favor of complementarity – Banerjee et al. (2009) fail to reject null hypothesis that CT and GT cotton are independent

• Soybeans:

– Fernandez-Cornejo et al. (2002) + Fernandez-Cornejo et

• al. (2013) support causal relationship between CT & GT

– Fernandez-Cornejo et al. (2003) partially reject the presence of complementarities

7

State of findings: aggregation & temporal variation

• The above provides qualified evidence in favor of

complementarity between GT crops & CT. However, data limitations and methodological assumptions restrict generality of findings

• Because of its nature, complementarity is best studied at

individual choice level. Three of papers cited above (Roberts et al., Frisvold et al., Fernandez-Cornejo et al. 2013) use state-level data

• The three studies using farm-level data (Fernandez-

Cornejo et al. 2003, Kalaitz. & Suntorn., & Banerjee et

• al. 2009) have single cross-section

8

State of findings: methodology & choice set

• Regarding methodology, two important features for

identification of complementarity have been neglected by previous studies

• appropriate test for complementarity requires a choice-

set defined over all possible combinations of the available practices (Gentzkow 2007)

• grower facing choice between two binary technologies,

should be modelled as choosing between four technology systems. Otherwise, as is with the bivariate probit or logit models, complementarity is either ruled

• ut or inadequately characterized (Miravate and Pernías

2010, Gentzkow 2007)

9

State of findings: methodology & correlation

• Also, need to allow for possibility that unobserved

returns are correlated across practices

• Clustering (or lack thereof) of the observed practices

may be due to correlated unobserved tastes or attributes, rather than complementarity

• Restricting unobserved returns across practices to be

uncorrelated—as done by nearly all existing studies dealing with (GT, CT) complementarity—can lead to accepting complementarity when absent, or rejecting it when present (Athey & Stern 1998; Cassiman & Veuglers 2006)

10

Our contribution

Novelties are

• data used, considerably more extensive than others,
• econometric methodology applied.

– Based on a structural model that includes all four decision combinations. Typically a grower is held to make two simultaneous, albeit distinct, adoption

• decisions. Then complementarity isn’t directly estimated

and results can be difficult to interpret – Controls for correlation induced by unobserved heterogeneity by estimating full covariance matrix of individual random effects

• auxiliary findings developed during course of econometric

identification

11

Data

A representative farm-level dataset from survey company GfK spanning 1998–2011 and containing seed and tillage choices of 29,518 soybean growers (GT soybeans were commercialized in 1996). See http://www.gfk.com/us Not a balanced panel dataset, but it does contain repeated

• bservations over time for a subset of individuals. About

43% of farmers sampled in a year are re-sampled next

• year. For many farmers we observe whether tillage choice

changed upon switching to GT soybeans, helping to sort

• complement. from correl. among unobserved returns

Education is one type of unobserved factors. More schooled producers may use both CT and GT soybeans, so unconditional correlation between practices would exceed correlation conditioned on education

12

13

Data, Cont’d

• Data are designed to be representative at CRD level.

Including multiple fields, sample contains 82,056 farm- field-year observations across 235 CRDs in 31 states, largest soybean states being most heavily represented.

• Data coming from GfK surveys include tillage & seed

choices, seed & herbicide prices, and farm size variable

• Each field is identified as using one of “Intensive Till.,”

“Conservation Tillage,” “No-Till.” In baseline specification, we treat intensive till. as distinct, and combine the others into the CT category

14

15

20 30 40 50 60 70 80 90 100 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011

Conservation Tillage GT Soybeans No-Till

Figure 1. Conservation Tillage and GT Adoption Rates for U.S. Soybeans (percent of acres)

16

Table 1. Distribution of Tillage and Seed Systems (% of obs.) System 1998- 2001 2002-’06 2007-’11 1998-’11 (CV,IT) 20.73 6.34 2.26 10.18 (GT,IT) 21.53 30.41 29.38 27 (CV,CT) 20.3 6.63 3.01 10.35 (GT,CT) 37.44 56.61 65.34 52.47 Observations 28,701 29,240 24,115 82,056

Variables

17

= = = = =

, , , ,

year price of CV soybean year price of GT soybean year price of herbicide used on CV soybean year price of herbicide on GT soybean dummy variable, 1 if growing more

CV t GT t CV t GT t it

p t p t r t r t Size = = = = =

• avg. January soybean futures price for November

than 500 acres diesel fuel price index index measuring soil erodability drought severity index simple time tren contract d

t t i it t

Fuel Futures EI Palmer Trend

Variables, cont’d

18

ν = time-invariant, practice-specific normally distributed

• unobservables. They represent individual characteristics

we do not observe, such as land quality, which may affect the returns to the differen

i

t practices. We allow for these to be correlated across systems.

τ

ε =

,

 system-specific IID type I extreme value errors. They are residual in that they allow growers with the same characteristics and environment to still choose a different system.

s

d d itf

19

Figure 2. U.S. Soybean Seed Prices, 1998-2011 ($/50lb)

5 10 15 20 25 30 35 40 45 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Conventional Seed GT Seed

20

Figure 3. U.S. Soybean Herbicide Prices, 1998-2011

0.2 0.4 0.6 0.8 1 1.2 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Glyphosate Price Index Conventional Price Index

Conceptual model

• We implement a variant of the mixed logit model similar

to Gentzkow’s (2007) framework.

21

Let soybean growers be indexed by {1,..., }, year by {1,..., }, field Formal unit of analysis is a by {1,..., }. In a given year and field, grower chooses seed farm-field-year com and bin t

• n.

i ati

it

i N t T f F ∈ ∈ ∈ llage type: ( , ) {( , ),( , ),( , ),( , )}.

s

d d CV IT GT IT CV CT GT CT

τ ∈Ω ≡

Profit is given by For each of his/her field in each time period, grower chooses system such t ( , ). ( , ) ( , ) ( , ) ,( , ) ( h , ) at ,

itf s itf s itf s s it s s

d d d d d d d d d d d d

τ τ τ τ τ τ

π π π ′ ′ ′ ′ ′ ′ > ∈Ω ≠ ∀

Reference conceptual model

22

[ ( , ) ( , )] [ ( , ) ( , )] If so then marginal profit from CT choice is larger in pres Profit function is supermodular (i.e., com ence of GT than abs plementary) i e f n

itf itf

GT CT GT IT CV CT CV IT γ π π π π ≡ − − − ≥ t GT

π β β β β β β β β β β β ν ν ε = + + + + + + + + + + + + +

, 1 , 2 , 3 3 4 5 6 7 , 8 8

( , ) ( ) ( ) ;

CV IT CV IT itf CV t CV t it IT IT IT IT t t i it CV IT CV IT CV IT t i i itf

CV IT p r Size Fuel Futures EI Palmer Trend π β β β β β β β β β β β ν ν ε = + + + + + + + + + + + + +

, 1 , 2 , 3 3 4 5 6 7 , 8 8

( , ) ( ) ( ) ;

GT IT GT IT itf GT t GT t it IT IT IT IT t t i it GT IT GT IT GT IT t i i itf

GT IT p r Size Fuel Futures EI Palmer Trend

Non-estimable reference conceptual model

23

π β β β β β β β β β β β ν ν ε = + + + + + + + + + + + + +

, 1 , 2 , 3 3 4 5 6 7 8 8 ,

( , ) ( ) ( ) ;

CV CT CV CT itf CV t CV t it CT CT CT t t i CT CV CT it t CV CT CV CT i i itf

CV CT p r Size Fuel Futures EI Palmer Trend π β β β β β β β β β β β ν ν ε = + + + + + + + + + + + + +

, 1 , 2 , 3 3 4 5 6 7 8 8 ,

( , ) ( ) ( ) .

GT CT GT CT itf GT t GT t it CT CT CT t t i CT GT CT it t GT CT GT CT i i itf

GT CT p r Size Fuel Futures EI Palmer Trend

Estimated model

• Not all of the parameters in four equations are
• identified. Only parameters that contribute to differences

in per acre returns are estimable (Train 2009)

• To clarify identified parameters, and how the model

nests a test for complementarity, we normalize returns relative to a base system, which is taken to be (CV, IT)

24

τ τ

π π π π ≡ − = ( , ) ( , ) ( , ); ( , ) Define

itf s itf s itf itf

d d d d CV IT CV IT

Model’s meat

25

π β β β β β ν ε π β β β β β β β ν ε π π = + − + − + + + + = + + + + + + + + =

1 , , 2 , , 3 8 3 4 5 6 7 8

( , ) ( ) ( ) ; ( , ) ; ( , ) ( ,

GT itf GT t CV t GT t CV t GT GT GT GT it t i itf CT CT CT CT itf it t t CT CT CT CT CT i it t i itf itf itf

GT IT p p r r Size Trend CV CT Size Fuel Futures EI Palmer Trend GT CT GT IT

( ) ( ) ( ) ( )

γ γ

π γ ε β β β γ β β β β ε ε ε ε ε + + + ≡ − ≡ − − − ≡ − − −

, , , , , , , , , ,

) ( , ) ; ; ; .          

itf itf GT IT CV IT GT GT CT GT IT GT IT CV IT GT CT GT IT CV CT CV IT itf itf itf itf itf

CV CT

26

Table 2. Regressor Summary Statistics

Variable Mean S.D. Min 0.25 Median 0.75 Max Size (>500 acres) 0.33 0.47 1 1 Futures ($/bu) 7.3 2.78 4.48 5.2 6.37 9.6 13.13 Fuel Price Index 49.96 24.33 19.6 29 43.8 65.4 91.2 Erodiblity Index 8.36 9.49 0.29 2.67 5.2 11.3 192.07 Palmer’s Z- Index 0.29 2.47

• 4.93 -1.46
• 0.11

1.48 11.84 Seed Price ($/50lb bag) 8.98 1.92 6.34 7.46 8.67 9.84 12.41 Herbicide Price Index

• 0.28

0.2

• 0.65 -0.42
• 0.26
• 0.1

Decomposition between complementarity and common correlation

27

Sum captures whether GT soybeans and CT complement. To see this, note that, in terms of the un-normalized returns: ( , ) ( , ) ( , ) ( , This revisits a ). n e

itf itf itf itf itf itf

GT CT GT IT CV CT CV IT

γ γ

γ ε γ ε π π π π + + = − − +     arlier equation. However, now the relation is adjusted for presence of unobserved heterogeneity. Complementarity can vary over the population through .

itf γ

ε

[ ] 0, we can interpret as complementarity. Comment: Our modeling approach has constrained to be independent of prac Becaus tices e .

itf

E

mean

γ

ε γ γ =

Identification

• Model is formally identified: more moments than
• parameters. However, we want more than that in seeking

to separate complementarity from covariance parameters

28

, , ,

Estimating accounts for unobserved factors that contribute to both and

GT CT GT IT CV CT itf itf

σ π π

2 , 2 ,

To control for the correlation induced by unobserved heterogeneity, we allow for correlation between and . Specifically, we ( , ) ~ (0, ); . assume that

GT i CT i GT GT CT GT CT i i GT CT CT

N ν ν σ σ ν ν σ σ   Σ Σ =      

Identification of beta parameters & of decomposition

• Beta parameters identified through use of

– seed & herbicide price differences (in GT choice) – fuel, EI, Palmer index, futures (in tillage choice)

29

{ }

,

Precise identification of complementarity and covariance parameters requires additional sources of variation and information beyond the basic formal requ 0, 0 hard to irements distinguish fro . m

GT CT

σ γ > <

{ }

,

0,

GT CT

σ γ < >

• Decomposition identification is lar

argely through panel

• data. If adoption of GT soybeans & CT are correlated

because of high then adoption of one should not depend on adoption of other for a given individual

σ

, GT CT

Estimation

30

. For given realizatio Rewrite normalized eq n ,probability uations as ( = observation, = sys

• f choosing system is

tem) . ( ; )

j j j i itf k k k itf i it

j j j j j itf itf i itf j i x j itf x k

x j e L e i j

β ν β ν

π β ν ε ν ν θ

+ + ∈Ω

= + + =

∑

{ }

the probability of is produc Let set of actual cho t of corresponding lo ices in datase gits: t. Given , ( ; ) ( ; )

i i

i itf j i i j itf j

j L L

ζ ζ

ζ ν ζ ν θ ν θ

∈

≡ = =∏

Estimation, again

31

1

Unconditional probability is found by integrating over random vector , with parameters yet to be ide Integral simulated via Halton draws. 1 ( ; ) ( ) ; ( ; ). Then f ntified i . nd ˆ

i i i i

R r r

R P L f dv P L R v

ζ ζ ζ ζ

ν θ ν ν θ θ

=

= ≈ ∑

∫

1

1 arg max ln ( ; ) . Done by Stata. Identification is relative to a scaling parameter

• n extreme value distribution

i

R r i r

L R

ζ θ

ν θ

=

  =    

∑ ∑

32

Table 3. Simulated Maximum Likelihood Results

Variables Coefficient Standard Error GT Variables Constant 1.597*** (0.206) Seed Price

• 0.326***

(0.025) Herbicide Price

• 0.984***

(0.157) Size 0.119*** (0.042) Trend 0.442*** (0.010) CT Variables Constant

• 0.571***

(0.132) Size 0.285*** (0.056) Futures

• 0.026**

(0.012) Fuel Price 0.007*** (0.002) Palmer Drought Index

• 0.024**

(0.010) Erodibility Index 0.079*** (0.012) Trend 0.044*** (0.009)

33

Table 3. Simulated Maximum Likelihood Results, Cont’d Variables Coefficient Std Error 2.220*** (0.1097) 3.919*** (0.2225) 0.309*** (0.0846) 0.461*** (0.0405)

Notes: based on 82,056 observations. Std. errors clustered at the CRD level. Except for covariance parameters, coefficients identified relative to , the scale parameter for . Covariance parameters are identified relative to . ***Significant at the 1% level. **Significant at the 5% level.

σ 2

GT

σ 2

CT

σ

, GT CT

γ

φ φ 2

ε j

itf

34

Table 4. Average Marginal Effects GT(%) CT(%) Seed Price

• 1.14 D
• 0.08 I

Herbicide Price

• 0.11 D
• 0.01 I

Soy Futures

• 0.002 I
• 0.04 D

Fuel Price 0.002 I 0.07 D Palmer Z

• 0.0004 I
• 0.01 D

Erodibility Index 0.01 I 0.10 D Notes: The reported effects are elasticities; i.e., the % change in the probability of adopting GT (CT) given a 1% change in the respective variable. See text for additional discussion.

D = Direct Effect I =Indirect Effect.

35

Table 5. Robustness Checks: Alternative Estimates for Gamma

Alternative Specifications Coefficient

• Std. Error

Include Herbicide Price in CT Variables

0.414*** (0.0395)

No Correlation:

0.585*** (0.0322)

Ignore Panel Aspect of Data

1.361** (0.6699)

Basic Logit

0.547*** (0.0333)

Restrict Sample to Central Corn Belt Only1

0.304*** (0.0519)

No-Till or Till for Tillage Choice2

0.651*** (0.0414) ***Significant at 1% level. **Significant at 5% level.

1Includes IA, IL, and IN, for which there are 26,304 obs. in all. 2This variation specifies the tillage choice as being between no-

till or a positive amount of tillage (rather than between CT and IT).

γ

, GT CT

σ =

WTP for system benefits

• Divide coefficient by coefficient on seed price

gives WTP per bag.

• For the complementarity coefficient, this gives

$1.41/bag

• At 1.2 bags/acre WTP for synergies is about

$1.69/acre

36

37

Table 6. Tillage Predicted Adoption Rates (percent of acres)

• Cons. Till Predicted

Rates No-Till Predicted Rates With GT Without GT Diff. With GT Without GT Diff.

1998

53.9 50.6 3.3 31.0 27.0 4.0

2000

57.6 53.3 4.2 35.8 30.5 5.4

2002

59.9 54.7 5.3 38.6 31.8 6.8

2004

62.0 56.0 6.0 40.4 32.6 7.7

2006

66.6 60.7 5.9 48.0 39.8 8.2

2008

68.7 62.7 6.0 48.8 40.2 8.6

2010

68.7 62.6 6.1 48.1 39.4 8.7

2011

69.4 63.3 6.2 49.4 40.5 8.9

Application: soil erosion

• Environment motivates interest in CT/NT. We focus on a

small part of this, the implied impact of GT soybeans, through NT adoption, on soil erosion.

• Data from Montgomery (2007, p. 13270). Median erosion

rate under IT about 1.5 mm/yr., roughly 20 times 0.08 mm/year CT median erosion rate. Difference about 6.8 tons/acre per year. Table 6 percent differentials, implies a mean soil loss reduction of 27 million tons/yr.

• For context, total U.S. cropland soil erosion in 2007 about

1.9 bill. tons. USDA/NRCS (2009) estimate benefits $4.93/ton water quality improvements, $1.93/ton saved

• fertilizer. Thus, value of soil saving benefits $189 mill./yr.

38

Other applications?

• Complementary packages in cropping? Model can be

extended to more inputs, but curse of dimensionality

• A case where there is likely substitution is monoculture

and nitrogen use. This can be handled, but only when there are two practice choices

• In animal agriculture, move to industrialization entails

many practice changes, including perhaps scale, intensity, genetics.

39

Questions?

40