On the timing of relevant weather conditions in agriculture

Journal of the Agricultural and Applied Economics AssociationEarly View ARTICLEOpen Access On timing relevant weather conditions in agriculture Zhiyun Li, Corresponding Author Li [email protected] Charles H. Dyson School Management, Cornell University, Ithaca, New York, USA Correspondence NY 14850, USA. Email: [email protected] for more papers by this authorAriel Ortiz-Bobea, Ariel Ortiz-Bobea USASearch author First published: 28 August 2022 https://doi.org/10.1002/jaa2.21AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text full-text accessPlease review our Terms Conditions Use check box below share version article.I have read accept Wiley Online Library UseShareable LinkUse link a article with your friends colleagues. Learn more.Copy URL Share linkShare onFacebookTwitterLinked InRedditWechat Abstract A growing literature is analyzing effects on agricultural outcomes. In these studies, constructing variables requires researchers define “season,” period over which are considered outcome. We explore consequences assuming an incorrect season such analyses. show simulations that imposing introduces nonclassical measurement error regressors, bias unknown directions. propose approach recover “true” underlying apply it US state-level panel Total Factor Productivity. find accounting seasonality can lead substantially different estimates. 1 INTRODUCTION rapidly documenting micro- macro-level impacts shocks wide range economic outcomes goal characterizing potential climate change. These studies explored high temperature energy demand (Dirks et al., 2015), labor productivity (Cachon al. 2012), mortality (Deschênes & Greenstone, 2011), cognitive performance (Zivin 2018), industrial output (Dell 2012; Zhang exports (Jones Olken, 2010), conflict (Burke 2014; Hsiang Burke, 2011, 2013), migration (Carleton Hsiang, 2016; Cattaneo 2019; Stern, corporate earnings (Addoum 2019). There even larger exploring extreme crop yields (Schlenker Roberts, 2009; Tack 2015) also aggregate like gross domestic product (GDP) 2015; Dell 2012) or total factor (TFP) (Liang 2017; 2018, 2021). seemingly innocuous but important empirical challenge choice “season” time window variables. Weather data available at much higher temporal frequency (e.g., daily) than variable interest annual). As result, typically assume observations construct regression analysis. Researchers often necessary background make reasonable assumptions about issues crop-specific studies. For instance, common spanning April through September when corn Midwest. However, modeling choices difficult justify assessing across countries states where climatic vary significantly dealing GDP TFP, do not well-defined “seasons.” Farmers’ responses could inconsistency between researchers’ prior knowledge true season. farmers may adjust their inputs switch other types production adapt fluctuations. cases, uncommon address problem using annual Specifically, they consider entire calendar year as capture all plausibly parsimonious way. To knowledge, there little formal guidance regarding its implication change impact assessments remains largely unexplored. This explores seasons differ from generating process (DGP) estimating see “season selection” issue. some terminology clarify exposition. refer DGP “correct” hand, we “selected” referring one assumed researcher. Naturally, selected be correct depending whether matches DGP. “relevant” occur within season, “irrelevant” occurring outside season.11 Note multiple rather comprised consecutive months year. TFP sensitive parts since mixtures components. The ideal tackle question would examine selecting intra-annual combinations (s) (are) certain part current work take several weeks run, only accounts Applying simulation will increase computational exponentially because algorithm must search prohibitive number seasonal combinations. Due computation complexity, focus single Future efficient model selection procedure. illustrate example. Suppose researcher selects maximum regressor July, then essentially irrelevant into variable. If due inclusion (occurring July) uncorrelated July), introduced classical regressor. well known, leads attenuation bias. same year, particularly adjacent periods, correlated. suggests thus nature conduct simple exercise based data. Essentially, want measured difference within-variations variables—defined season—is correlated case, nonclassical. results Figure 1. Each 20 panels shows correlation (Tmax, Tmean, Tmin, Precipitation, shown column) aggregated hypothetical (Winter, Spring, Summer, Fall, Annual, row) errors aggregating 78 defined (corresponding each colored cell panel). White cells figure indicate no while red blue negative positive correlations, respectively. 1Open viewerPowerPoint Correlation demeaned constructed alternative seasons. Correlations computed lower 48 1961–2004 period. maximum, minimum, mean precipitation average monthly values throughout paper. removed location fixed effect them compute correlations Winter Tmax, top left panel) possible (each indicated main text, panel, vertical axis indicates starting month horizontal ending highlight solid black circle Interestingly, most either blue, suggesting biases direction. Most winter fall tend coldest means value Tmax closer warm winters falls. Thus, greater correspond smaller (Tmax − season) negative. later issue formally conceptual framework. issue, tractable data-driven Our guiding principle fit econometric should improve closely That is, fit, whereas incorporating deteriorate it. strategy consists grid identify best-fitting constitutes estimate measure obtained cross-validation minimize out-of-sample mean-squared (MSE), criterion captures tradeoff goodness complexity (Arlot Celisse, 2010). evaluate Monte Carlo simulation. generate consistent particular seasonality. effective recovering cases type does match DGP.22 select Tmin. uniform warming scenario differs quantify practical direction, suggested error. Finally, identification real-world setting. analyze fluctuations 1960 2004. line simulations, successfully recovers well-known sector various country dominated field crops. Importantly, substantially, approach. Although ignore application, remain makes clear difference. paper mainly contributes two strands literature. First, econometrics approaches agriculture. has focused cross-sectional approaches, measuring adaptability Di Falco 2011; Just, 2013; Moore Lobell, Schlenker Taraz, 2017), understanding properties tradeoffs existing models (Carter 2018; Cui Ghanem Smith, 2021; Newell aggregation raise overlooked regard, contribute consequence arbitrary windows how Second, Studies area arising inaccuracies self-reported survey applies settings, especially Arthi Bound Krueger, 1991; French, 2004), consumer behavior (Gibson Kim, 2010; Gibson development (Baird Özler, Beegle Desiere Jolliffe, health (Carletto Larsen 2019), political science (Imai (Abay relatively less studied. temporally introduce document causal estimation study. proposed limitations reader keep mind. primarily geared toward identifying restricted matters. considering competing (temperature vs. vapor pressure deficit, precipitation). fact reflect extended accommodate expanding under consideration clearly requirements. explicitly functional form issues. analysis quadratic form, too strong assumption. Again, expanded demand. Third, account nonadditivity It known varying harnessing within-season nuances Despite limitations, think step right direction proposes improvement relative practices. hope study motivates others develop issues.33 Possible strategies include LASSO (least absolute shrinkage operator) MARS (multivariate adaptive splines), potentially simultaneously. 2 GRID SEARCH APPROACH TO RECOVER THE TRUE SEASON relying matching provide superior fit. build idea out large set “optimal” effectively T∗. estimator 10-fold year-block cross-validation, whole years sampled together block. seems suitable applications spatial low serial three steps: Perform candidate “seasons” store MSE model. 2. Visualize every grid. visualization helps better appreciate overall outcome variable.44 restrict ourselves periods contiguous 12-month possibilities. One conceivably extend previous years. 3. Select corresponding best fit.55 reveal if portion simulation, T∗ realistic setting short set. 3 DATA 3.1 Data sources processing rely growth rates states, obtain USDA Economic Research Service (ERS).66 Unfortunately, been updated after 2004 critical source information, Farm Labor Survey, was discontinued. methodology described ERS website (www.ers.usda.gov/data-products/agricultural-productivity-in-the-us/methods/). Link data: www.ers.usda.gov/data-products/agricultural-productivity-in-the-us/ divided input. obvious given encompassing outputs inputs. 44-year Climate Hub regions (see Supporting Information: A1). both application. fine-scale gridded PRISM Group Oregon State University (http://prism.oregonstate.edu). group provides continental 1895 4-km resolution. study, minimum (Tmin, Tmax) overlapping Because state, activities appropriate therefore state level cropland weights counts USDA's Crop Layer (30m) 2008–2014. depict A2. 3.2 Summary statistics summary United States 1961 Table A1. expected, considerably exhibiting slightly conditions. Also, summers appear wetter. summarizes dependent variable, rate. rate around 1.4% sample period, substantial regional variation. grew modest 0.8% 0.7% Northern Plains Southern Plains, contrast, faster stable 2.0% 1.3% Pacific Northwest Southwest region, standard deviations Eastern irrigation region (1961–2004) Region/TFP (%) Max Min Mean SD 15.3 −9.5 2.0 5.0 40.8 −76.4 0.8 11.0 Midwest 28.6 −40.2 1.5 9.6 Northeast 25.4 −18.7 1.6 6.3 17.0 −19.1 1.3 5.3 19.4 −27.6 0.7 7.3 Southeast 25.8 −31.8 8.0 National 1.4 7.9 Abbreviation: productivity. 4 SIMULATION 4.1 Description here interested resulting practice estimates terms magnitude Another effectiveness priori purpose, context obviously control 10 (2000–2010). express general (1)where generated i t, DGPs . term drawn distribution U (0;1) normally distributed N (0;1). choose parameters () = (12; −0.5), response function “inverted U” shape peaking W 12. Increases beyond point detrimental simplicity, parameter (measured °C) mm). subsequently expressed (2)where arbitrarily T need those represents term. “baseline” excludes purpose comparing four DGP.77 amounts × DGPs, adds 80 simulations. addition, perform 450 iterations iteration, deriving sampling regressions 780 351,000 30 million 4.2 Simulated Here Tmax). this, 1°C simulated estimated compare computing “bias ratio” deviates fractional terms. (3) results. T∗, ratio close 0, gray areas “close” exhibit small ratios. encouraging information what might be. plenty 2, indicating decisions underestimation projected warming. nonnegligible decision over- under-estimation While States, patterns arise elsewhere ways. way determine 2Open column used estimation. row corresponds etc.). color 4.3 now Tmin A3 summer (July–September).88 Winter, similar summer. Figures A4–A7). diagonal zero. off-diagonal panels, Selecting itself, Not surprisingly, very biases, case. 4.4 Evaluating directions virtually unpredictable. Therefore, help fruitful. useful selection. result average, correctly Visually seen coincidence square symbols panel. words, seasons, sample. density meaning always variability. 3Open Columns bot