1. Introduction
Precision agriculture technology (PAT) is considered an important economic factor for the future of agricultural efficiency, increasing production or allocating inputs more sustainably (Bullock et al., Reference Bullock, Ruffo, Bullock and Bollero2009; Kolady et al., Reference Kolady, Van der Sluis, Uddin and Deutz2021; Tey and Brindal, Reference Tey and Brindal2012). Previous studies have examined farmers’ adoption decisionsFootnote 1 and economic outcomes.Footnote 2 However, the question of how PATs impact farm financial returns is unclear (Griffin et al., Reference Griffin, Shockley and Mark2018; Lowenberg-DeBoer, Reference Lowenberg-DeBoer2019). Even at the individual operation level, Farmers themselves often struggle to identify the financial returns of their technology investments. (Thompson et al., Reference Thompson, Bir, Widmar and Mintert2019). This reality is likely driven by PATs that provide difficult-to-measure indirect effects on production (Griffin et al., Reference Griffin, Shockley and Mark2018; Shockley et al., Reference Shockley, Dillon, Stombaugh and Shearer2012) and the synergistic effects that arise when technologies are used together (Lambert et al., Reference Lambert, Paudel and Larson2015; Miller et al., Reference Miller, Griffin, Bergtold, Ciampitti and Sharda2017). Further adding complexity to the task of identifying the returns to PAT, farmers are likely to have unobservable characteristics that influence both the adoption decision and their financial outcomes (DeLay et al., Reference DeLay, Thompson and Mintert2022; McFadden et al., Reference McFadden, Rosburg and Njuki2022).
This study uses a two-stage data envelopment analysis (DEA) to capture both the direct and indirect effects of PAT on farm efficiency in generating gross revenue. Specifically, instead of examining agricultural production output as in previous studies (e.g. DeLay et al., Reference DeLay, Thompson and Mintert2022; McFadden et al., Reference McFadden, Rosburg and Njuki2022), we establish cost-minimizing frontiers for farm gross revenue.Footnote 3 This modeling choice captures the combined direct effects and indirect effects, like changes in field efficiency (Shockley et al., Reference Shockley, Dillon and Stombaugh2011), convenience increasing farmers’ capacity for other management tasks (Daberkow and McBride, Reference Daberkow and McBride2003; Thompson et al., Reference Thompson, Bir, Widmar and Mintert2019), and shortening information transfer from principles to non-owner operators (Deininger and Byerlee, Reference Deininger and Byerlee2012). Additionally, we capture the indirect effect of using PATs together, or in a bundle. Farmers often use multiple technologies together (Miller et al., Reference Miller, Griffin, Ciampitti and Sharda2019), and the degree to which different technologies may generate financial returns depends on using outside information (Bullock et al., Reference Bullock, Ruffo, Bullock and Bollero2009; Swinton and Lowenberg-DeBoer, Reference Swinton and Lowenberg-DeBoer1998). Measuring changes in cost efficiency for generating gross revenue allows us to capture the effects of bundling and the complementary effects of information generated and used by different PATs.
This study aims to estimate the within-farm effects of PAT use on farms’ efficiency in generating gross revenue. We use an unbalanced panel of Kansas farms and a two-stage DEA, with two-way fixed effects regressions, to account for unobservable characteristics that may influence both PAT adoption and revenue efficiency, such as management ability. Specifically, we compute yearly cost-minimizing frontiers and measure each farm’s relative efficiency. Because efficiency scores are bounded between 0 and 1, a two-way fixed effects model estimated with ordinary least squares (OLS) does not fully account for this censoring. However, our objective is to identify within-farm changes associated with PAT adoption, and the inclusion of binary adoption indicators is best suited to the linear fixed effects framework. While nonlinear estimators, such as a Tobit with Mundlak corrections, are often applied to censored outcomes, they can obscure the variation we seek to exploit and substantially reduce statistical power. In our sample, observations at the bounds of 0 and 1 are rare, so the risk of bias from using OLS is limited. Further, McDonald (Reference McDonald2009) show that DEA scores are not censored, but fractional, and second stage OLS with White (Reference White1980) robust standard errors is a consistent estimator of the second stage effects. Taken together, this modeling choice provides a tractable and interpretable way to assess the within-farm efficiency effects of PAT use.
To our knowledge, this study is the first to use an unbalanced panel and farm efficiency in relation to gross revenue to evaluate the effects of PATs. Previous studies have used a similar empirical approach to analyze returns to PAT (DeLay et al., Reference DeLay, Thompson and Mintert2022; McFadden et al., Reference McFadden, Rosburg and Njuki2022), but frequently use physical output, like yield, obscuring any potential indirect effects of PATs on financial performance. The dataset includes detailed financial information on 570 unique farms and 7,245 observations over the 21 year period from 2002 to 2022. Our empirical approach first establishes the cost-minimizing frontier for farm gross revenue. Then using the annual efficient frontier, we compute each farm’s annual relative efficiency. Second, the farms’ relative efficiency scores are the dependent variable for two-way fixed effects regressions and the farms’ reported bundle of PATs are the independent variables. Hence, the role of year-specific price and weather in farm efficiency is removed. Consequently, our results are estimates of the within-farm change in efficiency related to the farm’s collection of PATs. An additional benefit of our DEA analysis is that we are able to compute the effects of PAT on overall, pure technical, scale, and allocative efficiency, further characterizing the role of PAT in farm efficiency. Additionally, we estimate the impacts of PAT on farms at various points on the conditional farm efficiency quantiles to examine how the observable characteristics of farms impact the effect of PATs.
It is important to note that our estimates should be interpreted as the correlation between the farms’ reported presence of PATs and measures of farm efficiency. Because PAT is often included in new machinery purchases, our methodology does not exclude the possibility that any changes in efficiency could be attributed to the updated machinery and not the inherent technology. However, given the diversity of our data, the choice to use gross revenue as our output and total assets (including machinery) as an input in our DEA, and the likely presence of classical measurement error due to weather and price variability, we believe that our results are still conservatively indicative of the role of PATs on the efficiency of Kansas farms. Additionally, our results are inherently time averages, which potentially obscures the any effects of PAT that might accrue over time, as discussed by Lowenberg-DeBoer (Reference Lowenberg-DeBoer2003). Despite these limitations, we believe our results related to PAT use on Kansas farms can help inform farm management decision for a broader range of farm operations.
There are two key findings. First, the majority of examined PAT bundles are not associated with an increase in farm overall efficiency. We examine seventeen different bundles of PAT, and only two – guidance alone and yield monitor and grid soil sampling are associated with an increase. Each of these PATs are associated with an increase in allocating inputs in a more cost-effective manner, but guidance is additionally associated with achieving an optimal farm size. Second, less efficient farms are able to experience efficiency gains relative to their more efficient peers. In contrast to the previously mentioned results, five of the seventeen examined bundles are associated with efficiency increases. Of these bundles, they include multiple technologies, suggesting that less efficient farms may benefit from using multiple PATs together. These findings highlight that returns to PAT are likely nuanced and that further research is needed to understand how new agricultural technology can improve efficiency for heterogeneous operations.
This study provides insights for farmers and PAT manufacturers, retailers. First, for farmers, they can carefully evaluate the ways in which potential technologies may impact their operation, without the pressure to adopt quickly, as their does not appear to be a financial motivation for well-managed farms to adopt PATs. Second, for manufacturers and retailers, our results and the fact that previous literature documents uncertainty around financial returns (Griffin et al., Reference Griffin, Shockley and Mark2018; Lowenberg-DeBoer, Reference Lowenberg-DeBoer2019; Thompson et al., Reference Thompson, Bir, Widmar and Mintert2019) suggest that considerable effort should be spent quantifying how farms can benefit from these technologies.
The remainder of this study is organized as follows. In the next section, we discuss our research design, two-stage DEA analysis, which combines DEA analysis, two-way fixed effects OLS estimation, and the unbalanced panel of Kansas farms from the Kansas Farm Management Association (KFMA). We then discuss our results, starting with our analysis at the mean, followed by a discussion of the results estimated at the lowest, median, and highest conditional quartile of farm efficiency. We then conclude, summarizing our results and discussing policy implications and future research.
2. Research design
To examine how PAT impacts farm efficiency, controlling for unobservable characteristics of farm managers, this study employs a two-stage DEA research design. In the first stage, cost frontiers are formed for each production year using DEA. Although recent papers have used stochastic frontier analysis for a similar research aim (i.e. DeLay et al., Reference DeLay, Thompson and Mintert2022; McFadden et al., Reference McFadden, Njuki and Griffin2023), our analysis uses DEA for three reasons. First, DEA allows for multiple outputs. As we will describe in the Data subsection, our sample of farms contains integrated crop and livestock operations. The presence of these two enterprises may impact a farm’s ability to benefit from PAT. For example, guidance may be used to simplify the operator’s task when applying manure from a livestock enterprise, but simultaneously performing an operation in a manner that maximizes the fertilizer value, increasing the efficiency for crop production. Second, DEA follows all theoretical considerations that a cost frontier must adhere to: symmetry, quasi-concavity, homogeneity, monotonicity, and differentiability. Third, the established cost frontiers can be used to calculate common efficiency metrics such as overall, pure technical, allocative, and scale efficiency. We use these metrics to characterize the changes in farm efficiency related to PAT use. In our second stage, the mentioned farm efficiency metrics (overall, pure technical, allocative, and scale) are used as dependent variables in a farm fixed effects regression. The independent variables are binary indicators for the farms’ reported bundle of PATs. As a result, our estimates identify the changes in farms’ gross revenue efficiency associated with PATs.
2.1. Stage one: Data envelopment analysis
The input-oriented minimum cost DEA model is represented by the series of linear programming equations below:
$ MC(y_{i},x^*_{im},w_{i}) = Min\sum _{m=1}^M w_{im}^{'}x_{im}^* $
subject to:
$ \sum _{i=1}^I z_{i}x_{im} \leq x_{qm}^{*} \forall m=1,\ldots, M $
$ \sum _{i=1}^I z_{i}y_{ik} \geq y_{qk} \forall k=1,\ldots, K $
$ \sum _{i=1}^I z_{i} = 1 \ (for \ VRS) $
$ (z_{1},\ldots, z_{I}) \geq 0, $
where MC(y i ,x im *,w i ) is the minimum cost to produce the output, w im are the input prices for farm i with respect to input m, x qm * are the cost-minimizing inputs for farm q with respect to input m, x im are actual inputs used by farm i with respect to input m, z i measures the weight of each farm when forming the frontier, and y ik is the output k for farm i. All variables are assumed to be strictly positive and z is greater than or equal to zero. This model assumes variable returns to scale technology, which is imposed by constraining all z i ’s to sum to one.
Establishing the annual minimum cost frontier provides a standardized efficiency metric for farms. Additionally, inherent in DEA is the ability to calculate various efficiency measures, like overall, pure technical, allocative, and scale efficiency. We use the relative differences in the estimated changes across these efficiency metrics to examine the channels through which PATs impact efficiency.
Overall efficiency is a measure of the optimal cost under constant returns to scale, relative to the actual cost the farm incurred. The formula for overall efficiency is:
where OE
i
is the Overall Efficiency for farm i in a given year,
$\sum _{m=1}^M w_{im}^{'}x_{im}^{CRS}$
is the optimal cost under constant returns to scale for farm i, and
$\sum _{m=1}^M w_{im}^{'}x_{im}$
is the actual cost incurred by farm i. Overall efficiency ranges from zero to one.
Scale efficiency is a measure of the optimal scale under constant returns to scale, relative to the actual cost the farm incurred. The formula for scale efficiency is:
where SE
i
is the Scale Efficiency for farm i in a given year,
$\sum _{m=1}^M w_{im}^{'}x_{im}^{*}$
is the optimal cost under variable returns to scale for farm i, and
$\sum _{m=1}^M w_{im}^{'}x_{im}^{CRS}$
is the minimum cost for farm i under constant returns to scale. Scale efficiency scores range from zero to one and are a measure of the optimum scale of the firm. A scale efficiency of 1 indicates the firm is at constant returns to scale, a score less than 1 indicates the farm is not at the optimal scale.
In addition to (1), we estimate pure technical efficiency with the following series of linear programming equations:
$ Min \ \lambda _{i} $
$ subject \ to \!: $
$ \sum _{i=1}^{I} z_{i}x_{im} \leq \lambda _{i}x_{im} \ for \ m=1,\ldots, M $
$ \sum _{i=1}^{I} z_{i}y_{ik} \geq y_{qk} \ for \ k=1,\ldots, K $
$ \sum _{i=1}^{I} z_{k}= 1 \ (for \ VRS) $
$ (z_{1},\ldots, z_{I}) \geq 0, $
where λ i is the measure of pure technical efficiency for the model. Pure technical efficiency is a measure of how far off the cost frontier a firm is or how efficiently a firm takes inputs and converts them into outputs.
Allocative efficiency is a measure of the optimal input mix for the farm. It measures the distance between the optimal input mix and the input mix the farm actually used. The formula for allocative efficiency is:
$A{E_i} = {{\sum\nolimits_{m = 1}^M {} {w_{im}}x_{im}^*} \over {\sum\nolimits_{m = 1}^M {} {w_{im}}{\lambda _i}{x_{im}}}},$
where AE
i
is the allocative efficiency for farm i,
$\sum _{m=1}^{M} w_{im}x_{im}^{*}$
is the optimal cost under variable returns to scale for farm i, and
$\sum _{m=1}^{M} w_{im}\lambda _{i}x_{im}$
is the actual cost incurred multiplied by pure technical efficiency (λ
i
) for farm i. Allocative efficiency ranges from zero to one.
2.2. Stage two: farm fixed effects regression
We use the four efficiency measures, (overall, pure technical, scale, and allocative efficiency) from the previously mentioned DEA and the within-farm variation in bundles of PAT to identify the effects on efficiency using the following equation:
$ {\rm Efficiency}_{it}=\alpha + \beta {1\,} {\rm Bundle}_{it}' + \delta _{i} + \tau _{t} + \varepsilon _{it}, $
where
${\rm Bundle}_{it}^{\prime} $
is a vector of dummy variables corresponding to PATs bundles for farm i in time t, δ
i
is the farm fixed effect, and τ
t
is the year fixed effect. We define Efficiency
it
with the output from the previous described DEA. Specifically, we are interested in overall efficiency, pure technical efficiency, scale efficiency, and allocative efficiency. Equation (6) is estimated using OLS so that fixed effects may be used to control for time invariant farm characteristics. Robust standard errors are used when estimating this model. This approach does not take into account the possible censorship of the DEA efficiency measures at 0 or 1. However, McDonald (Reference McDonald2009) argues that tobit models are inappropriate to analyze the efficiency scores from DEA models as they are generated from a fractional process, not a censoring process. Additionally, within our data, the censorship issue is minimal in most of the years across the four efficiency measures (see Figure 2).
Our aim is that β captures the within-farm effect of each combination of PATs on farm efficiency. Our DEA efficiency measures are generated separately for each year, controlling for stochastic elements which are common to the year, like weather and prices. Furthermore, τ t controls for any additional year-specific variation common among farms, yet not captured in the DEA analysis. There is a potential that a farm may experience an idiosyncratic shock that would be reflected in their efficiency, such as region specific weather and prices. However, there is no reason to believe that this idiosyncratic shock follows a pattern over several periods. Thus, although we do not address the potential classic measurement error, we believe that our estimates of β are conservative.
${\rm Bundle}_{it}^{\prime} $
is defined by dummy variables for important PAT and the interactions of these technologies. We include yield monitors, guidance (both lightbar and automated), section control, grid soil sampling, variable rate fertilizer, and variable rate seeding. Each of these technologies are represented with a dummy variable, which takes the value of one if the farm reports using this technology and zero otherwise. The full complement of interaction possibilities with technology is included. For example, we estimate the effect of guidance alone, guidance bundled with each other technology as a pair, guidance bundled with every combination of three technologies, guidance bundled with every combination of four technologies, guidance bundled with every combination of five technologies, and the full bundle of all six technologies. Hence,
${\rm Bundle}_{it}^{\prime} $
is a vector of 56 potential bundles of PATs.
2.3. Data
We use an unbalanced panel of Kansas farms from the KFMA to examine the degree to which the use of PAT influences farm efficiency between 2002 and 2022. The panel consists of 7,245 observations from 570 unique farms. Farms that are part of the KFMA data opt-in by becoming members of the management association. In general these types of farm management datasets represent “larger farms” and a higher percentage of crop vs. livestock operations (H. Kuethe et al., Reference Kuethe, B.Briggeman, Paulson and Katchova2014). The reporting farms average almost 13 years within the panel. This data has been used by several studies that examine the economics of PAT. For example, a subset of current data was used by Miller et al. (Reference Miller, Griffin, Ciampitti and Sharda2019) to estimate the propensity to adopt precision agriculture as bundles. Additionally, Ofori et al. (Reference Ofori, Griffin and Yeager2020) use a subsample of the same data to analyze the time to adoption of Kansas farms. Most pertinent to our study, Dhoubhadel (Reference Dhoubhadel2020) use a subsample of the same data to show that PAT technologies are not associated with long-term impacts on farm profitability.
Farmers participating in KFMA report both production and financial characteristics. We use their financial metrics for our DEA analysis. We define the farms’ output as the gross revenue generated by crop production and livestock production. We define farm inputs as the dollar amounts spent on land, labor, capital, and other variable expenses. All expenses are on an accrual basis. Other variable expenses represent the value of all crop and livestock expenses for the year. Capital is the beginning inventory of all capital assets used on the farm. Labor is the annual dollar amount recorded for labor along with any unpaid operator and/or family labor expense. Unpaid operator and/or family labor is an imputed measure based on the average family living expense over the past three years for the state of Kansas. Land expense is the cash rental expense plus the opportunity cost of the owned land which was calculated using USDA NASS county-level cash rental rates (U.S. Department of Agriculture, National Agriculture Statistics Service, 2024). We retain any farms that reported positive input and output values for analysis.
Table 1 reports the annual means for our two outputs and the four inputs on the represented KFMA farms. The mean livestock receipts show that the average farm in our sample generates less than 20% of their revenue from livestock. Our sample contains many integrated crop and livestock farms, further validating our choice to model two outputs with DEA. The role of variability in weather and prices is exhibited in the mean outputs. There is a steady growth in the value of production, as we would expect, but years like 2009, 2014, 2015, and 2022 show an annual decrease in average crop receipts, and 2006, 2008, 2009, 2012, 2013, 2015, 2016, 2018, 2019, and 2022 show an annual decrease in average livestock receipts. Similar trends exist for the four inputs, land, labor, capital, and other variable expenses.
Table 1. Mean of outputs and inputs for data envelop analysis for farms of KFMA 2002 to 2022 (in US dollars)
Source: KFMA 2002 to 2022.
In addition to financial and production records, KFMA also collects information on the use of PAT. Farms voluntarily report this information starting in 2015. The farmer discloses whether they use a specific PAT, the year of first use, and the year of last use or when whether they discontinue the use. In 2015, farms reported their PAT use going back to 2002. As a result, we construct a backward looking representation of the farms’ use of PAT starting in 2002. We focus our analysis on the use of yield monitor, guidance, section control, grid soil sampling, variable rate fertilizing, and variable rate seeding. Figure 1 plots the growth of reported use of PAT in the KFMA subsample. These trends are similar to the patterns reported by USDA (McFadden et al., Reference McFadden, Rosburg and Njuki2022) and Purdue University and CropLife magazine (Erickson et al., Reference Erickson, Lowenberg-DeBoer and Bradford2017).
Figure 1. Precision agriculture usage on farms in the Kansas farm management data 2002 to 2022.
Table 2 shows the frequency with which each technology bundle occurs in the data. Thirty percent of the farm observations indicate that they did not use any of the technologies during the duration of the study. For clarity, this result could represent a farm that has never used PATs or a farm with PATs in recent years and none in the early years of our analysis. The most common technology reported was the use of guidance, with 16.2% of the farm observations indicating that during a given year, guidance was the only technology they had used. We believe that the prevalence of common bundles of technology is suggestive of the perceived gains of technology (Thompson et al., Reference Thompson, Bir, Widmar and Mintert2019) and a common adoption path (Miller et al., Reference Miller, Griffin, Bergtold, Ciampitti and Sharda2017). We choose to limit our analysis to bundles of PAT that account for at least 0.5% of the observations. As a result, Grid Soil Sampling (GRID) is the bundle with the fewest observations included in our analysis.
Table 2. Precision agriculture technology bundle frequency in KFMA 2002 to 2022. Bundles highlighted in light gray are used in our analysis
3. Results
3.1. DEA
Figure 2 reports box and whisker plots of the estimated efficiency scores from our DEA analysis for each year. The top right panel plots the pure technical efficiency as computed by (4) for each year in our data. The top left panel plots the overall efficiency computed by (2). In our sample, the median farms exhibited overall efficiency of 0.2 to 0.35 during our time period. In contrast to the overall efficiency, the median farm in each year exhibited pure technical efficiency between 0.6 and 0.8. The bottom left panel plots the scale efficiency as computed by (3). The scale efficiency exhibits more variation by year, relative to the other efficiency metrics. Lastly, the bottom right panel plots the allocative efficiency as computed by (5). The estimated allocative efficiency of the median farm in each year is in the range of 0.3 to 0.6.
Although the efficiency scores are bounded in [0,1], Figure 2 demonstrates that the distribution of our estimated scores is concentrated well within the interior of this range, with only a negligible share of farms at the boundary. This mitigates the typical concern that censoring at 0 or 1 dominates the regression and biases the results. As Greene (Reference Greene2018) notes, the bias of OLS under censoring is attenuating and proportional to the mass at the boundary. Given the interior concentration of our scores, any attenuation bias is expected to be minimal. Further, our regressors are binary indicators that vary within farms over time. In this setting, the two-way fixed effects OLS estimator has a natural interpretation as the within-farm mean difference in efficiency associated with switching a binary variable from zero to one, holding the shared year effects constant. This interpretation does not rely on distributional assumptions about the dependent variable, making two-way fixed effects a practical and transparent modeling choice for our specific aim. Lastly, as mentioned previously, McDonald (Reference McDonald2009) show OLS with White (Reference White1980) robust standard errors is a consistent second stage estimator for two-stage DEA analysis.
Figure 2. Box and whisker plot for estimated farm efficiency in minimizing the cost of generating gross revenue from KFMA farms from 2002 to 2022.
3.2. Two-way fixed effects
Table 3 reports the two-way fixed effects OLS estimates of the within-farm effects of reported use of PATs on farm efficiency. For brevity, we display coefficients only for the technology bundles highlighted in Table 2; the full set of year fixed effect estimates is reported in Appendix Table A1. The columns of Table 3 correspond to the dependent variable: pure technical efficiency, overall efficiency, scale efficiency, and allocative efficiency. Each coefficient reflects the estimated within-farm change in efficiency when the corresponding technology or bundle is reported in use, relative to periods when the same farm reports no PATs. For example, farms reporting guidance exhibit a 2.2 percentage point higher overall efficiency (column 2) compared to periods when the same farm reports no PATs.
Table 3. Two-way fixed effects estimates of the effects of precision agriculture technology on within Kansas farm efficiency 2002 to 2022
Huber-White heteroskedastic robust standard errors, clustered at the panel level, in parentheses.
* p < 0.10, ** p < 0.05, * * * p < 0.010.
Each coefficient is relative to the farm having no reported Precision Agriculture Technology.
Although several technologies and bundles show coefficients different from zero, the main conclusion of our estimates is that PATs are not an important driver of farm efficiency. Of the seventeen technologies and bundles, only two – guidance (GDNC) and the yield monitor and grid soil sampling bundle (YM × GRID) – have coefficients statistically different from zero at the 1 percent level. Guidance is associated with a 2.2 percentage point increase in overall efficiency, while yield monitor and grid soil sampling are associated with a 3.9 percentage point increase.
These findings are interesting due to the distinction between embodied knowledge technologies and information-intensive technologies discussed by Griffin et al. (Reference Griffin, Miller, Bergtold, Shanoyan, Sharda and Ciampitti2017) and Miller et al. (Reference Miller, Griffin, Ciampitti and Sharda2019). Guidance, which generates value without additional external information, represents an embodied technology, while yield monitors and grid soil sampling inherently provide information that must be incorporated into decision-making to generate value. In our results, one embodied technology (guidance) and one information-intensive bundle (YM × GRID) are associated with efficiency gains. Thus, we cannot conclude that either category is systematically more effective.
Looking more closely at guidance, the overall efficiency gains appear to stem from improvements in both scale efficiency and allocative efficiency. Gains in scale efficiency may reflect reduced overlap and better utilization of machinery, an effect often emphasized in the PAT literature (Griffin et al., Reference Griffin, Shockley and Mark2018). Increases in allocative efficiency suggest improved site-specific use of inputs. In contrast, the gains from the yield monitor and grid soil sampling bundle are concentrated in allocative efficiency, with farms reporting this bundle exhibiting a 7.7 percentage point increase. This result aligns with the expectation that information-intensive technologies improve the cost-effectiveness of input use through site-specific management.
While other technologies and bundles are not associated with changes in overall efficiency, several are linked to improvements in specific efficiency components. For pure technical efficiency, the bundle of yield monitor, guidance, section control, grid soil sampling, and variable rate seeding is associated with a 6.9 percentage point increase, suggesting greater output with given inputs. For scale efficiency, the bundles of guidance and section control (+3.1 percentage points) and yield monitor, guidance, section control, and variable rate seeding (+4.2 percentage points) show significant gains, likely reflecting improved capacity to utilize machinery at optimal scale. For allocative efficiency, yield monitor with guidance (+2.5 percentage points) and the larger bundle of yield monitor, guidance, section control, grid soil sampling, variable rate seeding, and fertilizer (+5.2 percentage points) are both associated with improvements in cost-effective input use.
Table 3 reports the estimated average within-farm change in efficiency associated with reported use of PATs, relative to periods with no PATs. Because farms may differ in their capacity to realize efficiency gains, we also estimate equation (6) using quantile regression. Quantile regression, a generalization of median regression, minimizes absolute error and allows us to examine the conditional effects of PAT use at different points of the efficiency distribution. We estimate the model at the lower quartile, median, and upper quartile of overall farm efficiency.
Table 4 reports the least absolute deviation two-way fixed effects estimates of the within-farm change in overall efficiency associated with PAT use. The first column corresponds to the least efficient 25 percent of farms, the second to the median, and the third to the most efficient 25 percent. As with the OLS estimates, coefficients reflect the change in efficiency associated with reporting a given technology or bundle, relative to no PATs.
Table 4. Two-way fixed effects least absolute difference estimates of the effects of precision agriculture technology adoption on Kansas farm efficiency 2002 to 2022 (Overall efficiency)
Standard errors in parentheses.
* p < 0.10, ** p < 0.05, * * * p < 0.010.
The results suggest that less efficient farms experience the greatest gains from PATs. For farms in the lowest quartile, five of the seventeen technology bundles are associated with statistically significant gains:
-
Guidance (GDNC),
-
Yield monitor × grid soil sampling (YM × GRID),
-
Yield monitor × guidance × section control × grid soil sampling (YM × GDNC × SC × GRID),
-
Yield monitor × guidance × section control × grid soil sampling × variable rate fertilizer (YM × GDNC × SC × GRID × VRF), and
-
Yield monitor × guidance × section control × grid soil sampling × variable rate fertilizer × variable rate seeding (YM × GDNC × SC × GRID × VRF × VRS).
Estimated gains range from 2.1 percentage points for guidance to 5.9 percentage points for YM × GRID. Guidance and yield monitors appear in four of the five significant bundles, underscoring their central role in driving efficiency improvements.
Near the median of the distribution, only guidance is associated with overall efficiency gains, and no technologies are associated with gains in the most efficient quartile. Together with the OLS results in Table 3, these findings suggest that the efficiency benefits of PATs accrue primarily to less efficient farms. Moreover, the quantile regression indicates that the largest gains for these farms often arise from using multiple technologies in combination. Five bundles are associated with efficiency gains that are at least weakly significant, and all include at least five technologies. Our analysis does not allow us to isolate which specific technology drives these gains, but the evidence suggests synergies among technologies play an important role for less efficient farms.
An important caveat to our analysis concerns the learning curve associated with PAT adoption. Farmers may not realize efficiency gains immediately; in the early years, adoption could be characterized by adjustment costs and incomplete utilization of the technology. Over time, as farmers accumulate experience and develop the skills needed to manage information-intensive tools, the potential efficiency benefits are more likely to be realized. Our analysis does not capture this dynamic process, but the results nonetheless provide an important first step in contextualizing the role of PATs in agricultural production.
4. Conclusion
Economists are perplexed by farmers’ pattern of PAT adoption, potentially due to difficulty in quantifying financial returns. This study uses within-farm variation to identify the effect of PATs on farm efficiency. The results suggest that the use of PATs are not broadly associated with gains in efficiency. However, farms reporting using guidance alone or yield monitors and grid soil sampling are estimated to increase overall efficiency relative to periods when they did not report using PATs. In further analysis, we find that less efficient farms are associated with gains from an expanded set PATs and that gains are associated with PAT bundles of multiple technologies. Our results suggest that financial returns to PATs are still difficult to identify. This fact may be an important driver of the adoption puzzle.
Although this study incorporates various aspects of farm management, there are some limitations. First, although the KFMA data allow for an evaluation of within-farm variation, the counterfactual farm without PATs and one with all potential bundles is unlikely. These technologies are likely to be adopted sequentially, layering benefits in subsequent periods. A future analysis could use individual farm data to evaluate the learning process and the efficiency gains of PATs. Second, our methodology is based on the assumption that weather-related variation and marketing effects are random. Although these assumptions are valid, this modeling choice introduces classical measurement error and the implied attenuation bias. As a result, our estimated relationships are likely muted. Future studies may be able to use a more comprehensive set of financial, digital production, and weather records to identify the desired effects. Third, we use DEA to produce the dependent variable, but PAT is an inherent input. Future studies could create a variable related to PATs that could be used in the DEA. Despite these limitations, we believe that our estimates are an important contribution to the literature.
An additional challenge for this study is the fact that the PAT landscape is ever evolving. This study uses PATs that have existed for decades. As a result, the definition of technology bundles excludes potential interaction effects with information synthesizing technologies such as farm management software or data analytics service providers. In addition, these bundles are likely to represent heterogeneous effects for farms of various scope and scale. As farm management datasets continue to incorporate the potential to observe technology engagement, economists will be better able to identify the factors related to heterogeneous effects. Furthermore, McFadden et al. (Reference McFadden, Njuki and Griffin2023) details the interplay of the broader definition of digital agriculture. Future studies could work to understand how farmers use digital agriculture technologies alongside established PATs.
This study has implications for farmers. First, farms and specifically efficient farms may not benefit financially from adopting PATs. If a farmer experiences utility gains from the adoption of PATs, as described in DeLay et al. (Reference DeLay, Thompson and Mintert2022), then we may see further adoption. Second, inefficient farms should consider the adoption of PATs. In contrast to the findings for efficient farms, inefficient farms are likely to experience additional gains from technology adoption.
The findings of this study have implications for equipment manufacturers and policy makers. Equipment manufacturers should focus on identifying inefficient farms and develop PAT platforms for this group. These producers experience a disproportionate increase in efficiency and, as a result, may be motivated consumers. In addition, policymakers, especially those focused on environmental impact, should also be interested in this inefficient group. PAT will allow this group to use their inputs more efficiently. Collectively our results suggest that these two stakeholder groups will benefit by exploring how farm efficiency heterogeneity impacts their goals.
Data availability statement
The farm management database used in this research from Kansas Farm Management Association is available upon request.
Author contributions
Conceptualization, C.F, B.B, J.I, M.B; Methodology, C.F, B.B; Formal Analysis, C.F, B.B; Data Curation, J.I; Writing-Original Draft, C.F, B.B; Writing-Review and Editing, J.I, M.B; Supervision, C.F, B.B, J.I, M.B; Funding, C.F, B.B, J.I, M.B.
Financial support
This research received no specific grant from any funding agency, commercial or not-for-profit sectors.
Competing interests
Authors Chad Fiechter, Brady Brewer, Jeniffer Ifft, and Michael Boehlje declare none.
AI Declaration
AI editing assistant Writefull was used in generating the manuscript.
Appendix A
Table A1. Two-way fixed effects estimates of the effects of precision agriculture technology on within Kansas farm efficiency 2002 to 2022
Huber-White heteroskedastic robust standard errors, clustered at the panel level, in parentheses.
* p < 0.10, ** p < 0.05, *** p < 0.010.
Each coefficient is relative to the farm having no reported Precision Agriculture Technology
Table A2. Two-way fixed effects least absolute difference estimates of the effects of precision agriculture technology on within Kansas farm efficiency 2002 to 2022 (Overall Efficiency)
Standard errors in parentheses.
*p < 0.10, ** p < 0.05, *** p < 0.010.
Table A3. KFMA precision agriculture technology data collection (in US dollars)
Source: KFMA 2002–2022
Open access
