2. Background and literature review

As discussed in the previous section, cover crops provide a number of advantageous properties, however, these benefits are predicated on proper management including selecting appropriate species, seeding rates, planting period, and termination methods (Wayman et al., Reference Wayman, Cogger, Benedict, Burke, Collins and Bary2015). Failure to do so may lead to the cover crop preemptively competing with the cash crop by immobilizing nutrients and temporarily lowering soil moisture prior to planting the cash crop (Alonso-Ayuso, Gabriel, and Quemada, Reference Alonso-Ayuso, Gabriel and Quemada2014). Consequently, an improperly managed cover crop practice (and in some cases even a well-managed cover crop) may increase the risk of lower yieldsFootnote ³ (Deines et al., Reference Deines, Guan, Lopez, Zhou, White, Wang and Lobell2023; Garba, Bell, and Williams, Reference Garba, Bell and Williams2022; Miner et al., Reference Miner, Delgado, Ippolito and Stewart2020). This has historically led to guidelines for cover crop use that must be followed if producers wish to also maintain crop insurance coverage.

Past literature has posited that the complexity of these guidelines may have prevented farmers from adopting cover crops for fear of inadvertently violating the guidelines and losing crop insurance eligibility (Connor, Rejesus, and Yasar, Reference Connor, Rejesus and Yasar2021). As part of the 2018 Farm Bill, further guidance was established for insured producers to explicitly address concerns related to insurability of cash crops that follow a cover crop (Natural Resource Conservation Service, 2019).Footnote ⁴ Consequently, the primary mechanism by which crop insurance participation potentially serves as a disincentive to cover crop adoption (i.e., losing eligibility by inadvertently violating guidelines) was largely eliminated following the implementation of the 2018 Farm Bill. However, cover crops may also diminish the negative effects of adverse weather on crop yields (Kaye and Quemada, Reference Kaye and Quemada2017) which means that cover crops and crop insurance potentially both serve as risk management tools to some degree (Aglasan et al., Reference Aglasan, Rejesus, Hagen and Salas2024; O’Connor, Reference O’Connor2013; Won et al., Reference Won, Rejesus, Goodwin and Aglasan2024).Footnote ⁵ Moreover, past literature also notes the role that crop insurance coverage may play on cover crop use via a traditional moral hazard hypothesis in which an insured farmer reduces “effort” in farm production by not planting cover crops (Connor, Rejesus, and Yasar, Reference Connor, Rejesus and Yasar2021). Thus, it is possible to motivate a negative relationship between cover crop use and crop insurance participation through both a “redundant” risk management strategy perspective and a moral hazard point of view (Connor, Rejesus, and Yasar, Reference Connor, Rejesus and Yasar2021).Footnote ⁶

Although the potential negative relationship between cover crop use and crop insurance participation is often mentioned in existing literature, studies that empirically address the association between the two are scarce. Lee and McCann (Reference Lee and McCann2019) make use of the 2012 USDA ARMS to estimate a probit model to identify factors that influence the adoption of cover crops among 1,712 soybean producers. Using a binary indicator for enrollment within the Federal Crop Insurance Program (FCIP), they find no significant effect between crop insurance participation and cover crop adoption. Fleckenstein et al. (Reference Fleckenstein, Lythgoe, Lu, Thompson, Doering, Harden, Getson and Prokopy2020) conducted a primarily qualitative analysis through a series of interviews and elicited concerns related to cover crop adoption among Midwestern row crop producers. They generally conclude that crop insurance requirements do not serve as a major deterrent to cover crop adoption. Connor, Rejesus, and Yasar (Reference Connor, Rejesus and Yasar2021) use county level data spanning 2006–2015 on Indiana corn and soybean producers to examine the relationship between crop insurance participation and cover crop adoption rates. Using panel fixed effects and instrumental variables techniques, they find a statistically significant and negative relationship between county-level crop insurance participation rate and the county’s share of acreage devoted to cover crops. However, they note the magnitude of the effect is small, and negligible from an economic perspective. Thompson et al. (Reference Thompson, Reeling, Fleckenstein, Prokopy and Armstrong2021) utilize a novel survey data set consisting of 719 US Midwestern corn producers and make a notable contribution by recognizing that cover crop intensity (measured by share of farm acres with cover crops planted) and the binary decision to adopt cover crops on the farm may be subject to separate decision-making processes. They model cover crop use as a function of observable variables using a double hurdle model and find that the factors associated with cover crop adoption (i.e., extensive margin) are potentially different from the determinants of the extent of cover crop implementation (i.e., the intensive margin).Footnote ⁷ With respect to crop insurance’s influence on cover crop use, they find no statistically significant effect between cover crop use (at either the intensive or extensive margin) and a binary variable indicating when survey respondents reported that crop insurance requirements limited their ability to implement cover crops.

3. Data

The data set used for this study comes from USDA’s Agricultural Resource Management Survey (ARMS). As a nationally representative survey, ARMS serves as USDA’s primary source of information on the financial condition, production practices, and resource use of US farm businesses and the economic well-being of farm households. The survey is administered using several phases – a sample screener phase [Phase I], a field-level phase [Phase II], and a farm-level phase [Phase III], and targets about 5,000 fields and 30,000 farms each yearFootnote ⁸ . In this study, we pool data from Phase III of the ARMS, which has cover crop information for the following years: 2012 and 2015–2021. Crop insurance expendituresFootnote ⁹ , which we use as a proxy for crop insurance participation (or demand), was first cleaned to remove any observations with unrealistically high per acre crop insurance expenditures over $500 per acre.Footnote ¹⁰ Cover crop use was measured as the share of operating acres planted with cover crops. Values above one, which indicates that there are more cover crop acres than total farm operation acres, were dropped from the sample. The final cleaned sample consisted of 119,166 farm level observations.Footnote ¹¹

Descriptive statistics are reported in Table 1. The average crop insurance expenditure in the sample was $7.34, however, this value includes many zero values of farms that did not purchase crop insurance. For the 42 percent of farm operators that had some level of FCIP participation (N = 49,687), the average expenditure was approximately $17.59 per acre. The average share of acres planted with cover crops by a farm was 3 percent, but as with crop insurance expenditures, this metric is influenced by many zero values. Among the 12 percent of respondents that planted cover crops, on average 27 percent of their operating acres were planted with cover crops. Conservation payments averaged $3.43 per acre in our sample. Farm operator characteristics gathered indicate that 6 percent of farms are run by female operators while 8 percent of farms are beginner farms (anyone that has operated a farm or ranch for less than 10 years). The mean age of an operator from the sample was approximately 60 years old. Farm ownership responses indicate that 50 percent of the farms in the sample were partly owned by the operator, 10 percent were rented, with the rest being fully owned by the operator. Twenty-nine percent of operators reported being a college graduate. The average farm was just over 1300 acres and allocated an average of 55 percent of that acreage to crop production. Fifty percent of farms indicated raising cattle on their farm. To account for the relative differences in moisture, temperature, and evapotranspiration across regions, we use monthly values of the Palmer Drought Severity Index (National Center for Environmental Information 2024) or “PDSI” averaged over 1990–2021 at the county level.Footnote ¹² The average farm in the dataset was located in a county that had a PDSI of 0.36 on a scale ranging from −10 (most dry) to 10 (most wet).

Table 1.

Descriptive statistics

Preconstructed farm typology indicators in the ARMS dataset are used to control for farm type and are based on farm typology definitions developed by USDA Economic Research Service (Hoppe and MacDonald, Reference Hoppe and MacDonald2013). Retired farms, which are defined as having a gross farm cash income of less than $250,000 and an operator that reported being retired but continuing to engage in farming on a small scale, made up 4 percent of our sample. Residential (or “lifestyle” farms) made up 17 percent of our sample and are defined as those with GCFI less than $250,000 and have operators that reported a primary occupation other than farming. Nonfamily farms, which made up 5 percent of our sample, are defined as any farm (regardless of GCFI) where the operator and persons related to the operators do not own a majority of the business. The remaining farm classifications are used to characterize family farms (i.e. any farm where the majority of the business is owned by the operator or persons related to the operator) and are based on GCFI. Small family farms are defined in two sub groups; “lower-sales” farms (21 percent of our sample) are those with GCFI less than $100,000 while “higher-sales” farms (10 percent of our sample) are those with GCFI between $100,000 and $250,000. Large family farms (12 percent of our sample) are defined based on a GCFI between $250,000 and $500,000 and very large family farms (31 percent of our sample) are defined as any farm with a GCFI of more than $500,000.

4. Empirical model

The empirical model underlying each of our regressions is defined as:

(1) $$C{C_{it}} = \alpha + \gamma\ asinh\left( {{E_{it}}} \right) + \beta {X_{it}} + \theta + \lambda + {\epsilon _{it}}$$

where the variable CC _it measures the percentage (%) of acres with cover crops planted as a share of total operating acres on farm i in crop year t. The key independent variable capturing FCIP participation, E _it , is crop insurance expenditures measured as the premium per acre paid by the farm ($/acre). The vector X _it contains other observable control variables that potentially influence the decision to adopt cover crops on the farm (extensive margin), as well as the farm acreage where cover crops are used. Because crop insurance expenditures (E _it ) contain zero values (indicating no crop insurance was purchased) and also tends to follow a right-skewed distribution, we use the inverse hyperbolic sine transformation which is denoted as $\textit{asinh}(.\text)$ (Bellemare and Wichman, Reference Bellemare and Wichman2020).Footnote ¹³ The terms θ and λ represent year and state fixed effectsFootnote ¹⁴ respectively.

The model structure that is used to estimate equation (1) is motivated by the large share of observations for which no cover crops were planted, meaning a value of zero is recorded for the dependent variable in this case. To explicitly model the fact that many farms choose a corner solution (i.e., not to plant cover crops on the farm), we make use of a double hurdle modeling structure (Cragg, Reference Cragg1971). In our empirical context, the first hurdle indicates the decision to initially adopt a cover crop practice. If a farmer does decide to plant a cover crop, the second hurdle characterizes the acreage allocated to cover crops. The first hurdle is estimated using a probit model. More formally, the first hurdle to be estimated is defined as:

(2) $$d_{it}=\beta _{1}+\beta _{2}\textit{asinh}\left(E_{it}\right)+\beta _{3}X_{it}+\theta +\lambda +v_{it}$$

where d _it is a binary variable equal to one if farmer i planted cover crops in crop year t, zero otherwise. The second hurdle is functionally equivalent to the specification defined by equation (1) and is estimated via a truncated normal regression.

5. Identification strategy

As previously mentioned, a primary concern for identification is the fact that both cover crop use and crop insurance participation both potentially affect production risk, to varying degrees. Consequently, the underlying risk profile of a particular producer (i.e. the various moments that characterize their yield distribution) is likely to influence their decisions pertaining to participation in both planting of cover crops and purchase of crop insurance. Given that “risk” is not directly observable, a canonical case of endogeneity persists (i.e., due to omitted variables). Further, recent empirical evidence suggests that cover crops potentially reduce the occurrence and magnitude of crop insurance losses in the FCIP for perils related to extreme moisture levels (i.e. both drought and excess moisture) (Chen et al., Reference Chen, Rejesus, Aglasan, Hagen and Salas2022; Won et al., Reference Won, Rejesus, Goodwin and Aglasan2024). Consequently, the existing literature suggests that cover crops may directly attenuate the losses that crop insurance are designed to provide financial protection against leading to a case of simultaneity where increased cover crop use may lead to crop insurance being perceived as a redundant risk management strategy.

To address endogeneity in our empirical framework, we follow past literature and define an instrumental variable that utilizes the premium rating framework inherent to the FCIP as a way to generate exogenous variation in premium rates which in turn directly influence crop insurance demand (Tsiboe and Turner, Reference Tsiboe and Turner2023; Woodard and Yi, Reference Woodard and Yi2020). A number of policy parameters are internally defined each year by RMA using historical data from the FCIP. Notably, even though these parameters are defined using historic data (data which any individual producer may have contributed to), a single producer has negligible influence on their own premium rate due to the large number of producers that that contribute to the historic data series (Tsiboe and Turner, Reference Tsiboe and Turner2023).

Temporal and spatial variation in policy parameters is further made exogenous by smoothing operations that RMA applies to limit sharp discontinuities in premium rates across adjoining county boundaries and catastrophic loading which is done to account for extreme tail events being sparsely represented in historic data (Coble et al., Reference Coble, Knight, Goodwin, Miller, Rejesus and Duffield2010). In effect, this means that variation in these policy parameters that govern premium rates shift the premium rate curve in a way that is exogenous to current individual farmer demand for insurance (i.e., which is also exogenous to the omitted “inherent farm risk” variables in the error term of equation 1). For a more detailed discussion on crop insurance rating methods and how they produce an instrumental variable the is exogenous to producer crop insurance purchases, see the appendix.

Specifically, we make use of two policy parameters to define our instrument. The first being the county base rate α _65ct which represents the starting premium for an individual choosing a 65%Footnote ¹⁵ coverage level in county c and crop year t. Notably, α _65ct , represents the premium rate before any characteristics of an individual producer are used to customize the premium.Footnote ¹⁶ The second is the catastrophic loading factor δ _ct which as noted above is used to attenuate sampling error for sparsely observed catastrophic events. Our final instrument is defined as τ _65ct = α _65ct + δ _ct which represents the initial premium faced by a producer, prior to any information about a producer’s past production experience being used in the rating process. We refer to this instrument as the “initial rate.”

A major advantage of using the initial rate as an instrument over other common instrumentation strategies that make use of changes in subsidy rates (DeLay, Reference DeLay2019; Miao, Reference Miao2020; Yu, Smith, and Sumner, Reference Yu, Smith and Sumner2018) is that the initial rate provides extensive spatial variation which is lacking when using the instruments based on subsidy rates which only vary annually. The initial rate calculated for our sample had a mean value of $0.11 (reported in Table 1). In other words, on average, a producer in our sample faced an initial premium of $0.11 per dollar of insured liability before any further actuarial adjustments were made to their final premium quote.Footnote ¹⁷

Following previous literature, we incorporate our instrumental variables into the double hurdle framework by using a control function approach (Ricker-Gilbert, Jayne, and Chirwa, Reference Ricker-Gilbert, Jayne and Chirwa2011; Verkaart et al., Reference Verkaart, Munyua, Mausch and Michler2017). Implementation of the control function approach entails estimating the following reduced form model of crop insurance expenditures as a first stage, where X _it contains other observable covariates that act as controls and τ _65ct represents our excluded instrument which captures the producers’ relevant initial premium rate:

(4) $${\rm asinh}\ (E_{it})=\alpha +\beta X_{it}+\varphi \tau _{65ct}+\theta +\lambda +\epsilon _{it}$$

The residuals from estimation of equation (4) are then included as a covariate for estimation of both equations that define the double hurdle model (in equation (2) and (3)). Table S2 in the appendix reports the results for estimation of the first stage equation (4) for the entire sample as well as samples characterized by before and after the 2018 Farm Bill. In each of these estimations, the initial premium rates is highly correlated with crop insurance expenditures.Footnote ¹⁸ Further, the relationship between the initial rate and crop insurance expenditures is negative indicating that the change in expenditures due to increases in the initial rate is due to a reduction in demand for insurance, rather than a higher initial rate increasing expenditures through purchase of the same level of insurance at a higher cost.

7. Robustness checks

Our key independent variable, FCIP expenditures per acre, is both right skewed and includes many zero values making application of an inverse hyperbolic sine (IHS) transformation an appropriate choice. However, the IHS transformation is not scale invariant (Aihounton and Henningsen, Reference Aihounton and Henningsen2020; Bellemare and Wichman, Reference Bellemare and Wichman2020). For a variable x, the IHS function has little to no effect on values of x near zero and converges to log(2x) as x approaches infinity (Aihounton and Henningsen, Reference Aihounton and Henningsen2020). Thus the effect of the IHS transformation on the original data values is different based on the scale of the data (a choice that is potentially arbitrary). As a robustness check we re-compute our results after re-scaling FCIP expenditures per acre up and down by a factor of 10. Additionally, we estimate a set of results after instead applying a log transformation after adding 1 to avoid taking the log of zero.

Marginal effects for these estimations are reported in the appendix in figures S1-S3. Overall, our findings related to the relationship between various empirical specifications is preserved across all data transformations. Using OLS, 2SLS, or a standard double hurdle model (without instruments) all produce marginal effects that are biased toward zero and remain economically negligible. Applying the control function approach to the double hurdle model suggests qualitatively equivalent results relative to our main specification. Despite the differences in models being preserved, scaling our primary dependent variable does alter the magnitude of the estimated marginal effects. Scaling up by a factor of 10 results in most estimated marginal effects decreasing in magnitude slightly while scaling down by a factor of ten slightly increases the magnitude of the effect. However, the changes in magnitude of the marginal effects when taking the log of the dependent variable are minute.

Recognizing that a substantial number of farms in our sample report zero FCIP expenditures, as an additional robustness check, we estimate an alternative specification that keeps only observations that report positive crop insurance expenditures which restricts the sample to those farms that are already participating in the FCIP. Marginal effects estimated on this sub-sample of the data are reported in figure S4. Results from our preferred model (DH with instrumental variables) again suggest qualitatively equivalent results across pre and post 2018 Farm Bill subsets, however, estimation on the full data set (of FCIP participants) produces null effects at both margins of cover crop use. Notably, the magnitude of the marginal effect at the extensive margin prior to the 2018 Farm Bill is substantially larger among existing FCIP participants which is consistent with the idea that concerns about FCIP eligibility were a potential deterrent to cover crop adoption prior to the 2018 Farm Bill.

Finally, we incorporate the subsidy-based instrument proposed by (Yu, Smith, and Sumner, Reference Yu, Smith and Sumner2018) into our analysis by constructing an alternative instrument incorporating both the average subsidy rate for 65 and 75% yield protection policies (which we denote as S65 and S75 respectively) and the initial rate (τ _65ct ) as described in a previous section. The variables s65 and s75 are weak instruments (as indicated by 2SLS diagnostic tests) when used alone in our empirical setting, thus, we interact the average of the subsidy rates with our existing initial rate (i.e. τ _65ct [(s65+s75)/2] ). Marginal effects from using this alternatively defined instrument are reported in figure S5 and show qualitatively equivalent results to our primary specification.