genetic modification and yield risk in corn: parametric and non-parametric analysis
DESCRIPTION
genetic modification and yield risk in corn: parametric and non-parametric analysis. Elizabeth Nolan University of sydney Paulo santos Monash university. Bt corn. First genetically modified traits approved at end of 1996 Introduced commercially for 1997 - PowerPoint PPT PresentationTRANSCRIPT
1
E L I Z A B E T H N O L A NU N I V E R S I T Y O F S Y D N E Y
PA U L O S A N T O SM O N A S H U N I V E R S I T Y
GENETIC MODIFICATION AND YIELD RISK IN CORN: PARAMETRIC AND NON-PARAMETRIC ANALYSIS
2
Bt corn
First genetically modified traits approved at end of 1996 Introduced commercially for 1997
In 2012, corn hybrids with at least one GM trait were planted in over 88% of the
crop area in the United States Soil bacterium Bacillus thuringiensis is toxic to lepidopterous insects
European Corn Borer (CB) (1996) Corn Rootworm (RW) (2002)
allows for an almost complete control of the European corn borer and corn rootworm Superior to that of previously used techniques Previous damage control technologies:
up to 80% against first generation corn borer 67% against second generation borer 63% in the case of corn rootworm.
3
Pesticides and risk
complete evaluation of the impact of new technologies requires explicit recognition of risk
Debate about how pesticides affect risk Risk reducing (Feder) Risk increasing if output uncertainty is the dominant cause of
randomness (Pannell (1991) and Horowitz and Lichtenberg (1993))
The impact of GM traits is, therefore, an empirical question
4
Objectives
Identify effect of GM traits on production risk 1. flexible moments approach of Antle to analyse higher moments
variance of a distribution does not distinguish between upside risk and downside risk parametric
2. stochastic dominance Can compare distributions of yields originated by different technologies rank them in terms of desirability under minimal assumptions regarding decision makers’ preferences non-parametric
Complementary methods parametric approach allows quantification of the contribution of different input factors to differences
in yield distributions , but requires distributional assumptions. Stochastic dominance does not require distributional assumptions, not possible to quantify the
influence of specific inputs on yield distributions.
5
Data
Results of experimental field trials of corn hybrids submitted by corn
breeders to the Agricultural Extension Services of ten universities from
1997-2009 Illinois, Indiana, Iowa, Kansas, Minnesota, Missouri, Nebraska, Ohio, South Dakota and
Wisconsin Range of locations important for results Choice of period
Observations 147,790 individual trials 8,423 hybrids 339 locations 430 companies
6
CRDs represented in dataset compared with planted acres
Planted Acres by County
Source USDA NASS 2011
(http://www.nass.usda.gov/Charts_and_Maps/Crops_
County/cr-pl.asp)
Crop Reporting Districts represented in the
dataset
7
Advantages of data
Advantages of experimental data standardised trials avoid problems of identification but recognise that experimental yields higher than on farm yields
Provide details of agronomic practices information about the traits present in each hybrid Information on weather conditions wide variety of production conditions
over period since introduction of GM traits up to 2009 Spatial variability compensates for relatively short temporal range
8
Independent variables
GM traits (combinations) Site details and agronomic practices
plant density soil type cultivation type (conventional versus minimum or no till) previous crop early or late trial irrigated or dryland nitrogen application in lbs/ac
Weather conditions monthly rainfall April to September average minimum and maximum temperatures April to September
Year by location (CRD) interaction terms
9
Empirical method (parametric)
Large unbalanced panel dataset Individual corn hybrids are the cross sectional elements
Use both fixed effects and the Hausman-Taylor random effects specification to estimate a linear production function
Square (cube) residuals to obtain variance (skewness) Regress variance and skewness on the individual inputs
(including combinations of GM traits) to find marginal variance and skewness for each input.
10
Effect of GM traits on yield variance and skewness (HT)
VARIABLES Mean Variance Skewness
Corn borer resistance (CBO) 6.85*** 82.86*** -2,982*** (0.58) (14.64) (1018)
Rootworm resistance (RWO) 6.86** -16.46 1472(2.79) (71.49) (5710)
Herbicide tolerance (Ht) 0.88 44.17* -1682(0.95) (24.61) (1818)
CB and Ht 8.001*** 139.5*** -6,442*** (0.76) (21.42) (1726)
CB and RW 12.37*** 289.2*** -9,389*** (2.00) (48.78) (3398)
RW and Ht 14.67*** 177.3*** -5418(1.96) (50.13) (3640)
CB, RW and Ht 14.78*** 94.21*** -5,760** (0.84) (25.79) (2237)
Observations 147790 147790 147790Number of hybrids 8423 8423 8423
GM traits (Conventional as base)
Hausman Taylor
11
Margnal variance and skewness (FE)VARIABLES Variance Skewness
Corn borer resistance (CB) 86.44*** -3,225***(14.72) (1021)
Rootworm resistance (RW) -17.74 1648(71.82) (5727)
Herbicide tolerance (Ht) 41.37* -1426(24.73) (1823)
CB and Ht 141.0*** -6,256***(21.5) (1731)
CB and RW 289.1*** -9,006***(49.04) (3407)
RW and Ht 181.5*** -4572(50.38) (3651)
CB, RW and Ht 97.95*** -5,342**(25.85) (2244)
Observations 147790 147790Number of hybrids 8423 8423
GM traits (Conventional as base)
12
Results of parametric analysis
Marginal variance for most GM traits (and their
combinations) is positive presence of GM leads to an increase in variance
HTO only weakly statistically significant RWO not statistically significant
Most of the GM trait combinations have a statistically
significant negative effect on skewness Increase in downside risk.
RWO and HTO and RWHT not statistically significant effect.
13
Stochastic dominance
GM traits also have an important impact on mean
yield possible that decision makers are willing to trade the increase in
variability and downside risk (which they may dislike) with the increase in mean (which they may like) and still be
better-off
Use stochastic dominance to take into account these
simultaneous changes.
14
Ranking new technologies
If a new technology, for example GM traits, is superior according to first order stochastic dominance (FOD) criterion will be selected by any risk-averse or risk neutral firm will be chosen by farmers who always prefer higher expected
return to lower If the new technology is superior according to the
second order stochastic dominance (SOD) criterion will be selected by those farmers who prefer higher return to
lower and are also strictly risk averse
15
Figure 3.2 Number of trials by GM category
0
2000
4000
6000
8000
10000
12000
14000
1997 1999 2001 2003 2005 2007 2009 2011
CB only RW only Ht only CBHt
CBRW RWHt CBRWHt Total conventional
Trials by type of GM hybrids by year
16
SD for sub groups of GM hybrids
Divide data into subsetsanalyse for each trait within specific CDFs for
unconditional yield for the sub setsCBHT hybrids first order dominate conventional hybrids
produce more than conventional hybrids under all conditions effect of increased mean yield more than compensates for the increase in
variance and downside risk in terms of producers’ welfare
CBO, CBRW, and CBRWHT hybrids No dominant strategy
17
Stochastic Dominance Results
Trait Period Statistically significantly
different distributions
FOD SOD
CBO 2001-2005 Yes No NoHtO 2002-2007 No No NoCBHt 1999-2009 Yes YesCBRW 2005-2008 Yes No NoRWHt 2005-2008 No No YesCBRWHt 2005-2009 Yes No No
18
0.2
.4.6
.81
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320Yield bu/ac
Conventional hybrids CBRWHt hybrids
Conventional compared with CBRWHt hybrids 2005-2009
19
Results
Most combination of traits lead to increased variance exception is rootworm resistance by itself, which has a negative
coefficient, but is not statistically significant Downside risk increases with the presence of CBO,
CBHT, CBRW, and CBRWHT RWO and HTO, by themselves and in combination have no
statistically significant effect on downside risk However, CBHT first order dominates conventional for
the period 1999-2009
Conclusion
Pest resistance traits appear to lead to an increased variability of yield
Farmers may be more concerned about downside risk our results show that in most trait combinations downside risk is
increased for GM hybrids Results differ from those of other recent studies
Data include observations from more marginal corn producing states
20
21
Number of observations below cross (SOD)
StateFrequency Percent Frequency Percent Frequency Percent
Illinois 194 5.11 3 0.52 1 0.35Indiana 240 6.32 5 0.87 2 0.7Iowa 101 2.66 0 0 0 0Kansas 515 13.56 70 12.13 12 4.2Minnesota 44 1.16 1 0.17 1 0.35Missouri 1,120 29.5 268 46.45 150 52.45Nebraska 565 14.88 140 24.26 89 31.12Ohio 370 9.74 33 5.72 4 1.4South Dakota 174 4.58 8 1.39 8 2.8Wisconsin 474 12.48 49 8.49 19 6.64Total 3,797 100 577 100 286 100
CBRW if yield<67 CBRWHT if yield<52CBO if yield<103
22
More detailed resultsEstimator used in first step Fixed Effects Hausman Taylor
VARIABLES Variance Skewness Mean Variance Skewness Mean
GM traits (Conventional as base) Corn borer resistance (CB) 86.44*** -3,225*** N/A 82.86*** -2,982*** 6.85*** (14.72) (1,021) N/A (14.64) (1,018) (0.58) Rootworm resistance (RW) -17.74 1,648 N/A -16.46 1,472 6.86** (71.82) (5,727) N/A (71.49) (5,710) (2.79) Herbicide tolerance (Ht) 41.37* -1,426 N/A 44.17* -1,682 0.88 (24.73) (1,823) N/A (24.61) (1,818) (0.95) CB and Ht 141.0*** -6,256*** N/A 139.5*** -6,442*** 8.001*** (21.50) (1,731) (21.42) (1,726) (0.76) CB and RW 289.10*** -9,006*** N/A 289.2*** -9,389*** 12.37*** (49.04) (3,407) (48.78) (3,398) (2.001) RW and Ht 181.5*** -4,572 N/A 177.3*** -5,418 14.67*** (50.38) (3,651) (50.13) (3,640) (1.96) CB, RW and Ht 97.95*** -5,342** N/A 94.21*** -5,760** 14.78***
(25.85) (2,244) (25.79) (2,237) (0.84)
Plant density -1.431 -1,076*** 2.253*** -0.0708 -1,175*** 2.23***
(2.098) (257.7) (0.0520) (2.101) (256.9) (0.051)
No min till 43.96*** -1,481 -12.80*** 37.78** -1,304 -12.93***
(15.94) (1,958) (0.395) (15.96) (1,952) (0.39)
Irrigated -92.82*** 10,972*** 28.80*** -95.11*** 11,454*** 29.19***
(20.55) (2,523) (0.509) (20.58) (2,515) (0.50)
Early -56.32*** 1,798 2.747*** -58.54*** 1,512 -0.11
(9.469) (1,093) (0.261) (9.475) (1,090) (0.24)
Previous crop: Corn 169.4*** -11,288*** -10.76*** 166.2*** -11,638*** -11.57***
-14.32 -1,755 (0.356) -14.33 -1,750 (0.35)
Previous crop: Wheat 199.2*** -11,310*** -13.97*** 201.5*** -10,985*** -14.64***
(20.15) (2,473) (0.500) (20.17) (2,466) (0.49)
Previous crop: Alfalfa 108.60*** -1,883 2.434*** 95.81*** -772.9 1.42**
(27.29) (3,348) (0.677) (27.32) (3,338) (0.67)
Previous crop: Other 347.40*** -24,428*** -15.17*** 337.7*** -23,819*** -14.81***
(29.21) (3,586) (0.724) (29.25) (3,575) (0.717)
Nitrogen in lbs/ac -0.603*** 11.34 0.102*** -0.541*** 9.472 0.108***
(0.121) (14.86) (0.00301) (0.12) (14.82) (0.003)
Soil type: Clay 204.9*** -2,333 -5.293*** 212.8*** -2,414 -5.61***
(57.05) (7,007) (1.413) (57.12) (6,987) (1.39)
Soil type: Silty clay loam -97.20*** 5,465*** -0.399 -94.94*** 5,393*** -0.008
(11.35) (1,392) (0.281) (11.36) (1,388) (0.28)
Soil type: Clay loam -133.3*** 2,930 -0.930** -141.1*** 3,113* -1.45***
(15.31) (1,878) (0.380) (15.32) (1,873) (0.37)
Soil type: Loam -42.10** -1,472 -4.695*** -33.35* -1,578 -4.99***
(18.80) (2,308) (0.466) (18.82) (2,301) (0.46)
Soil type: Sandy loam 364.2*** -17,917*** 2.948*** 383.4*** -19,377*** 3.41***
(18.66) (2,292) (0.463) (18.69) (2,285) (0.46)
Soil type: Sand 657.90*** -13,935** -4.461*** 695.4*** -17,436*** -4.49***
(55.23) (6,782) (1.368) (55.30) (6,762) (1.35)
Constant -1,164** 178,162*** 111.3***
-1,437*** 194,387*** 106.2***
(538.10) (66,016) (12.44) (538.70) (65,825) (13.16)
Observations 147,790 147,790 147,790 147,790 147,790 147,790 Number of hybrids 8,423 8,423 8,423 8,423 8,423 8,423 Note 1: Asterisk (*), double asterisk (**) and triple asterisk (***) denote variables significant at 10%, 5% and 1% respectively. Standard errors in parentheses Note 2: Weather variables and a set of CRD by year interactions were included in the estimation, but are not reported.
23
Summary statistics
24
Number of trials of hybrids by year and state
Year IL IN IA KS MN MO NE OH SD WI Total1997 1189 981 3693 642 823 1190 1139 1004 535 2146 133421998 1069 3245 668 789 308 1169 955 590 2063 108561999 2095 3409 621 993 1223 1149 967 634 2159 132502000 1810 1626 3575 555 985 334 1332 853 556 1997 136232001 1739 1710 3321 671 859 1168 1087 844 593 1767 137592002 1302 1629 505 697 1201 1010 844 481 1765 94342003 1630 1155 466 735 1389 996 888 522 1797 95782004 2005 1341 672 931 1468 1149 1010 731 1818 111252005 1925 1471 2214 679 836 1479 1043 941 494 1803 128852006 1816 1196 2607 702 1190 1825 1023 838 640 1682 135192007 1778 1160 2810 932 1296 1529 1352 1215 588 2205 148652008 2020 1470 2587 1029 1039 1585 1201 1053 472 1779 142352009 1565 1241 2397 1028 940 1589 1185 1435 420 1669 13469
Total 21943 14980 29858 9170 12113 16288 14835 12847 7256 24650 163940
25
Hausman Taylor
Fits random effects model Some covariates correlated with the unobserved effects But not with idiosyncratic error Estimate
yit = x′1itβ1 + x′2itβ2 + z′1i γ1 + θi + μit
where x1it is a matrix of variables that are time varying and uncorrelated with θi,,
x2it is a matrix of variables that are time varying and are correlated with θi
and z1i is a matrix of variables that are time invariant and uncorrelated with θi (in this
case, the various combinations of GM traits)
Hausman and Taylor show that we can use x1it, z1i , x2it - x2i and x1i as instrumental
variables
26
Stochastic dominance for subsets of traits and combinations of traits by period
27
0.2
.4.6
.81
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320Yield bu/ac
Conventional hybrids CBO hybrids
CBO hybrids compared with conventional hybrids 2001-2005
28
0.2
.4.6
.81
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320Yield bu/ac
Conventional hybrids Hto hybrids
Conventional compared with Hto hybrids 2002-2007
29
0.2
.4.6
.81
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320Yield bu/ac
Conventional hybrids CBHt hybrids
Conventional compared with CBHt hybrids 1999-2009
30
0.2
.4.6
.81
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320Yield bu/ac
Conventional hybrids CBRW hybrids
Conventional compared with CBRW hybrids 2005-2008
31
0.2
.4.6
.81
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320Yield bu/ac
Conventional hybrids RWHt hybrids
Conventional compared with RWHt hybrids 2005-2008
32
Production function
When the disturbance h(X)ε enters the production function in an additive way Allows for the possibility of increasing, decreasing or constant marginal risk
(Just and Pope 1978). Can express function as:
yit = f(Xit) + uit = f(Xit) + h(Zit) εit where
yit is output
we assume that E(εit) = 0, var(εit) = 1.
f(Xit) is the deterministic component of production (representing the conditional mean of
production) as a function of the independent variables and uit is the stochastic component (representing its conditional variance), which can be rewritten
as a function of input use h(Zit).
33
Description of probability distributions
conditional mean and variance functions are not sufficient for a description of a stochastic production function. risk not only equivalent to output variance (uit
2)
Antle (1983) proposed the flexible moments approach shows that consistent estimates of all central moments can be
obtained econometrically without imposing arbitrary restriction on the moments of the distribution.
0.0
02.0
04.0
06.0
08.0
1.0
12.0
14
20 40 60 80 100 120 140 160 180 200 220 240 260 280 300Linear estimates of yield
Conventional hybrids using FE GM hybrids using FE
Conventional hybrids using HT GM hybrids using HT
Kernel density plots for linear estimates of GM and conventional hybrids, FE and HT