Upon the commercial introduction of Bacillus thuringiensis (Bt) cotton in 1996, the US Environmental Protection Agency mandated an Insect Resistance Management (IRM) program for the purpose of preserving the benefits of this technology’s insect protection by delaying resistance. According to that mandate, growers planting Bt cotton were required to follow the IRM practices designed to keep some lepidopteran populations from being exposed to the Bt protein. Thus, a refuge of conventional, non-Bt cotton had been planted along with Bt cotton. Other than Bt (insect-resistant) cotton, genetically modified (GM) types of cotton include herbicide-resistant (HR) and stacked-gene (SG), the latter combining the traits of insect resistance with herbicide resistance.

On another front, conservation tillage—and no-till practices in particular—has increased over the past few years (Figures 1 and 2). Conservation tillage (either no-till or reduced tillage) is reported to reduce soil erosion by preventing reduction in crop residue on the soil surface (Harper, 1996). Yet, despite the apparent advantages of conservation-tillage (CT) practices in reducing soil erosion, soil degradation, runoff, and in improving soil quality (Edwards, 1995; Sandretto, 1997), some farmers adopt no-till or minimum-till, and others do not adopt CT practices at all.

The use of CT practices may be even more important in cotton production than in other row crop production because of the minimal amount of residue left on the soil surface. Crop residues after planting averages 3% for cotton compared with 29% for corn (US Department of Agriculture [USDA], 1997). The use of HR or SG seed technology indirectly reduces the use of residual herbicides and increases profit potential of the farmer (Marra, Pardey, & Alston, 2003).

Past studies have attempted to show the relationship between adoption of GM—specifically HR—seed and conservation tillage and to determine the causal effects between them. For example, Fernandez-Cornejo and McBride’s (2002) cross-sectional study using USDA’s Agricultural Resource Management Survey (ARMS) data for 1997 investigated a potential simultaneous relationship between HR soybean seed and CT practices using two simultaneously estimated binomial probit models; the study compared those results with two single-equation probit models. Their study suggested that accounting for simultaneity was important for the no-till decision but not for the seed-use decision. In a more recent study with time-series data for 1992-2004, Roberts, English, Gao, and Larson (2006) used a Bayesian analysis and a two-equation simultaneous logit analysis to ascertain that introduction of HR cottonseed in Tennessee increased the probability that farmers would adopt conservation tillage, and farmers who had previously adopted CT practices were more likely to adopt HR cottonseed. They concluded that the simultaneous adoption of conservation tillage and HR cottonseed reduced soil erosion and residual herbicide use, as well as increased profit. However, with ARMS data for 2003, Banerjee et al. (2007) found no evidence of a simultaneous relationship between these two adoption decisions.

The overall objective of this study is to identify factors that led to the adoption of GM cotton in the United States. Specifically, the study seeks to identify farm and farmer characteristics driving the production of GM cotton in the United States in general. In particular, it studies the impact (or otherwise) of conservation tillage (i.e., no-till and reduced till: ridge-till, strip-till, and mulch-till) and conventional cotton used as refuge on the adoption of GM cottonseed across the United States and within selected regions.

Data and Methods

This study used two definitions of genetic modification of cottonseed adoption: one in which adoption of either Bt or HR cottonseed was considered as GM cottonseed adoption (denoted as BH); and another in which adoption of Bt, HR, or SG cottonseed adoption was considered as GM cottonseed adoption (denoted as BHS). The survey respondents were classified as adopters of a particular type of GM cotton according to these definitions. The “yes” responses received for these definitions (as percentages of the sample used in the regressions) were 65% and 86%, respectively. Two single-equation binary logit models (Maddala, 1983) were estimated using the same set of independent variables, except for the relevant adoption variable.

The ARMS is a random sample of cotton farms across the United States. The USDA Economic Research Service (ERS) used data from the 2003 ARMS—particularly the Cotton Production Practices and Costs Report and the Cotton Costs and Returns Report—to estimate the two binary logit models with the two definitions of genetic modification. The dependent variable for each equation indicated the adoption of GM cotton vis-à-vis its non-adoption (i.e., adoption of conventional cotton). The explanatory variables for each equation included

• demographic information of the survey respondents/farmers: age, education level, gross annual farm income, and length of tenure in cotton farming;
• farm characteristics: farm size, farm labor expense, refuge size in field surveyed, percentage of farm acres in cotton, if they grew genetically engineered cotton in the previous year (2002), if any part of their surveyed field was declared “highly erodible land” (HEL) by the Natural Resources Conservation Service (NRCS), if they no-tilled in 2002, cotton acres harvested, and percentage of cotton acres harvested using one or more CT equipment; and
• region-specific dummy variables: one for each cotton-producing region, with the 17 cotton-producing states being divided into five regions.

with the cumulative distribution function given by

where β' represents the vector of parameters associated with the factors x.

Assuming the probability that farmer n will choose to produce cotton using a particular technology (BH or BHS seed) is equal to the proportion of cotton farmers using that technology, the individual empirical models to be estimated may be specified as

BH = β0 + β1CONSTILL + β2BH02 + β3LABEXP + β4CTA + β5CTP        + β6HEL + β7REFUGE + β8YIELD + β9PNT + β10AGE + β11EDU        + β12TEN + β13GFI + βi REGIONj + ε T,

(3)

BHS = γ0 + γ1CONSTILL + γ2BHS02 + γ3LABEXP + γ4CTA + γ5CTP          + γ6HEL + γ7REFUGE + γ8YIELD + γ9PNT + γ10AGE + γ11EDU          + γ12TEN + γ13GFI + γi REGIONj + ε G ,

(4)

where

CONSTILL = dummy variable indicating whether or not CT practices were adopted;

BH02 = dummy variable indicating if Bt or HR cotton was grown in the prior year, 2002;

BHS02 = dummy variable indicating if Bt, HR, or SG cotton was grown in the prior year, 2002;

LABEXP = total labor expense on the farm ($100,000);

CTA = total harvested cotton acres (dryland and irrigated) on the farm (100 acres);

CTP = percentage of cotton acres on the farm;

HEL = dummy variable indicating if NRCS classified any part of the field surveyed as HEL;

REFUGE = dummy indicating if conventional, refuge cotton was used in the survey year, 2003;

YIELD = lint yield in pounds per acre harvested or expected in 2003;

PNT = dummy variable indicating if no-till was used in the surveyed field in the year before the survey, 2002;

AGE = age of the principal farm operator;

EDU = dummy variable indicating the surveyed farm operator’s education level (whether or not college graduate);

TEN = length of tenure (in years) of the operator household for the surveyed field;

GFI = gross farm income in 2003 ($100,000);

REGION = set of four dummies:

REGION1 includes Alabama, Florida, and Georgia;

REGION2 includes North Carolina, South Carolina, and Virginia;

REGION3 includes Arkansas, Louisiana, Mississippi, Missouri, and Tennessee and is excluded as the reference region to compare other regions with;

REGION4 includes Texas, Oklahoma, and Kansas; and

REGION5 includes Arizona, California, and New Mexico; with

ε_T and ε_G= random error terms;

β’s and γ’s = parameters to be estimated; and

the subscript n for the n^th farmer is suppressed for clarity.

To determine the quantitative effects of the right-side variables on each technology, marginal effects were calculated. Marginal effects (Greene, 1993; Maddala, 1983) measure changes in the probability of adopting each technology due to given changes in the relevant explanatory variables (Liao, 1994; Long, 1997). Marginal effects of continuous variables were calculated at the means of the data.¹ For dummy variables, a value of 0 was used if the mean was less than 0.5, and a value of 1 was used if the mean was greater than or equal to 0.5 (Obubuafo, Gillespie, Paudel, & Kim, 2006; Schlotzhauer, personal communication, 2006).²

The ARMS uses a complex stratified, multiphase, nonrandom survey design that may render naïve standard errors obtained by classical statistical algorithms invalid insofar as inferences on point estimates are concerned. Each observation in the ARMS represents a number of similar farms based on factors such as land use, farm size, etc.; the particular number is the survey weight or survey expansion factor. In effect, this is the inverse of the probability that the surveyed farm is selected to be surveyed (El-Osta, Mishra, & Morehart, 2007, p. 292). In order to alleviate the possibility of such a bias in measurement, the National Agricultural Statistics Service (NASS) has established standards that allow valid inferences based on the entire population. These standards were used in this analysis. In particular, “full-sample weights” were used to calculate means, parameter estimates, and marginal effects, and “replicate weights” were used in the “delete-a-group jackknife” procedure to calculate variances and, hence, standard deviations of all point estimates (Ahearn, El-Osta, & Deewbre, 2006; Dubman, 2000; El-Osta, Mishra, & Ahearn, 2004; El-Osta, Mishra, & Morehart, 2007; Kott, 1997a, 1997b; Lambert, Sullivan, & Claasen, 2007; Lohr, 1999). The samples used in the analysis included 537 and 898 respondents, respectively, for BH and BHS technologies, which, when properly expanded using survey weights, yielded a population of 9,128 and 12,126 farm operator households, respectively (Tables 1 and 2).

Results

Results (as given by the maximum likelihood estimates from Equations 3 and 4) indicate conservation tillage (CONSTILL) did not affect adoption of GM cotton with either definition for the study year (2003) at the 5% level; however, it did seem to cause a reduction in the likelihood of BH adoption (Equation 3) at the 10% level (Table 3). The latter counter-intuitive result brings to light one of the issues regarding the use of cross-sectional data. Instead of establishing causality, effects, determinants, or impacts, our regression results may be read as implying correlations, relationships, and associations. The “correlations” presented, however, assume ceteris paribus conditions. The weak negative correlation becomes weaker and no more significant even at the 10% level when SG cotton adopters are included in the mix, as apparent from the BHS adoption model (Equation 4) estimation. It is plausible that farmers who had not adopted CT practices before the introduction of BH seed might have first adopted BH seed and then decided that they could improve efficiency by adopting CT practices sometime during or before 2003. A similar study in a later year might make this variable positive and significant.

Whether GM cotton was used in the prior year (2002) positively affected the adoption of GM cotton in the study year at the 1% level, using both definitions. Percentage of cotton acres on the farm (CTP) also positively affected both adoption decisions at the 1% level. The variable for Region 1 was significant at the 5% level and negative in the BH equation, with a marginal effect of -0.173. This indicates that, estimated in isolation, this region (southeastern states of AL, FL, and GA) was 17.3% less likely to adopt BH relative to cotton farmers in Region 3 (AR, LA, MS, MO, TN).

Previous experience of no-till (PNT) in the BH equation and education level (EDU) in the BHS equation were significant only at the 10% level, with EDU having an unexpected negative sign.

Labor expense, cotton acres on farm, HEL, refuge size, yield, tenure, gross farm income, and age were not significant factors in determining adoption of GM technologies in cotton production for the study year.

Conclusions

Recent studies involving adoption of CT practices and HR cotton have attempted to determine if the adoption of CT practices impacts the adoption of GM cottonseed and/or vice versa. If CT adoption impacts GM cottonseed adoption, then adoption of CT practices can indirectly affect residual herbicide use and farm profits. The current study using cross-sectional ARMS data does not provide strong evidence of such a correlation with either of the two definitions of genetic modification used. Refuge (conventional, non-GM) cotton also did not affect GM cottonseed adoption using any definition in this study. However, the limitations of a cross-sectional study using complex data such as from the ARMS are apparent.

Endnotes

¹ In sufficiently “large samples,” marginal effects calculated by averaging the individual marginal effects at each observation (Bell, Roberts, English, & Park, 1994; Neter, Wasserman, & Kutner, 1983; Pindyck & Rubinfeld, 1976) would give the same results as obtained here from the means of the data (Greene, 1993, p. 876) by adding an observation with all means and calculating the marginal effects at that point.

² Anderson and Newell (2003) have developed a novel way of simplifying the calculation of marginal effects in logit and probit models (making them a function of only the estimated constant term) and their associated asymptotic variances by normalizing the explanatory variables at any desired value.

References

Ahearn, M., El-Osta, H.S., & Deewbre, J. (2006). The impact of coupled and decoupled government subsidies on off-farm labor participation of U.S. farm operators. American Journal of Agricultural Economics, 88, 393-408.

Anderson, S., & Newell, R.G. (2003, December). Simplified marginal effects in discrete choice models. Economics Letters, 81(3), 321-326.

Banerjee, S.B., Martin, S.W., Roberts, R.K., Larson, J.A., Hogan, R., Johnson, J.L., et al. (2007, February). Adoption of conservation-tillage practices in cotton production. Paper presented at the Southern Agricultural Economics Association annual meeting, Mobile, AL.

Bell, C.D., Roberts, R.K., English, B.C., & Park, W.M. (1994). A logit analysis of participation in Tennessee’s Forestry Stewardship Program. Journal of Agricultural and Applied Economics, 26, 463-472.

Dubman, R. (2000). Variance estimation with USDA’s farm costs and returns surveys and Agricultural Resource Management Survey (US Department of Agriculture [USDA] Staff Report AGES 00–01). Washington, DC: USDA, Economic Research Service (ERS).

Edwards, W.M. (1995). Effects of tillage and residue management on water for crops. In R.L. Blevins & W.C. Moldenhauer (Eds.), Crop residue management to reduce erosion and improve soil quality: Appalachia and northeast region (Conservation Research Report No.41).Washington, DC: USDA, Agricultural Research Service.

El-Osta, H.S., Mishra, A.K., & Ahearn, M. (2004). Labor supply by farm operators under ‘decoupled’ farm program payments. Review of Economics of the Household, 2(4), 367-385.

El-Osta, H.S., Mishra, A.K., & Morehart, M.J. (2007). Determinants of economic well-being among U.S. farm operator households. Agricultural Economics, 36, 291-304.

Fernandez-Cornejo, J., & McBride, W.D. (2002). Adoption of bioengineered crops (Agricultural Economics Report No. 810). Washington, DC: USDA, ERS.

Greene, W.H. (1993). Econometric analysis (3^rd Ed.) New York, NY: Macmillan Publishing Company.

Harper, J.K. (1996). Economics of conservation tillage (Pennsylvania State University College of Agricultural Sciences Cooperative Extension Conservation Tillage Series #6). University Park, PA: Penn State University, College of Agricultural Sciences. Available on the World Wide Web: http://pubs.cas.psu.edu/FreePubs/pdfs/uc130.pdf.

Kott, P.S. (1997a, May 14). Statistical analysis with a delete-a-group jackknife (NRED Seminar Series). Washington, DC: USDA, National Agricultural Statistics Service (NASS).

Kott, P.S. (1997b). Using the delete-a-group jackknife variance estimator in NASS surveys (Research Report RD-98-01). Washington, DC: USDA, NASS.

Lambert, D.M., Sullivan, P., & Claasen, R. (2007, April). Working farm participation and acreage enrollment in the Conservation Reserve Program. Journal of Agricultural and Applied Economics, 39(1), 151-169.

Liao, T.F. (1994). Interpreting probability models: Logit, probit, and other generalized linear models (Sage University Paper QASS Series No. 07-101). Thousand Oaks, CA: Sage Publications.

Lohr, S. (1999). Sampling: Design and analysis. Cincinnati, OH: Duxbury Press.

Long, J.S. (1997). Regression models for categorical and limited dependent variables. Thousand Oaks, CA: Sage Publications.

Maddala, G.S. (1983). Limited-dependent and qualitative variables in econometrics. Cambridge, UK: Cambridge University Press.

Marra, M.C., Pardey, P.G., & Alston, J.M. (2003). The payoffs to transgenic field crops: An assessment of the evidence. AgBioForum, 5(2), 43-50. Available on the World Wide Web: http://www.agbioforum.org/v5n2/v5n2a02-marra.htm.

Neter, J., Wasserman, W., & Kutner, M.H. (1983). Applied linear regression models. Homewood, IL: Richard D. Irwin, Inc.

Obubuafo, J., Gillespie, J., Paudel, K., & Kim, S.A. (2006, February 5-8). Knowledge, application and adoption of best management practices by cattle farmers under the environmental quality incentives program—A sequential analysis. Paper presented at the Southern Agricultural Economics Association annual meeting, Orlando, FL.

Pindyck, R.S., & Rubinfeld, D.L. (1976). Econometric models and economic forecasts. New York, NY: McGraw Hill, Inc.

Roberts, R.K., English, B.C., Gao, Q., & Larson, J.A. (2006). Simultaneous adoption of herbicide-resistant and conservation-tillage cotton technologies. Journal of Agricultural and Applied Economics, 38(3), 629-643.

Sandretto, C. (1997). Crop residue management (Agricultural Resources and Environmental Indicators, Agricultural Handbook 712, pp. 155-174). Washington, DC: USDA, ERS.

US Department of Agriculture, Economic Research Service. (1997). Agricultural resources and environmental indicators, 1996-97 (Agricultural Handbook 712). Washington, DC: Author.

Acknowledgements

The authors would like to acknowledge Dr. Dayton M. Lambert of the University of Tennessee in Knoxville, TN, for his useful suggestions, and Dr. Robert Dubman of the US Department of Agriculture Economic Research Service and Dr. Tommy Gregory of the US Department of Agriculture National Agricultural Statistics Service for their support of this research. In addition, the anonymous reviewers are also acknowledged. This research was supported in part by the Delta Research and Extension Center of Mississippi State University in Stoneville, MS. Usual disclaimers apply.