Application Of Bias Correction Research Articles

Evaluation of bias correction methods for wave modeling output

Models that seek to predict environmental variables invariably demonstrate bias when compared to observations. Bias correction (BC) techniques are common in the climate and hydrological modeling communities, but have seen fewer applications to the field of wave modeling. In particular there has been no investigation as to which BC methodology performs best for wave modeling. This paper introduces and compares a subset of BC methods with the goal of clarifying a “best practice” methodology for application of BC in studies of wave-related processes. Specific focus is paid to comparing parametric vs. empirical methods as well as univariate vs. bivariate methods. The techniques are tested on global WAVEWATCH III historic and future period datasets with comparison to buoy observations at multiple locations. Both wave height and period are considered in order to investigate BC effects on inter-variable correlation. Results show that all methods perform uniformly in terms of correcting statistical moments for individual variables with the exception of a copula based method underperforming for wave period. When comparing parametric and empirical methods, no difference is found. Between bivariate and univariate methods, results show that bivariate methods greatly improve inter-variable correlations. Of the bivariate methods tested the copula based method is found to be not as effective at correcting correlation while a “shuffling” method is unable to handle changes in correlation from historic to future periods. In summary, this study demonstrates that BC methods are effective when applied to wave model data and that it is essential to employ methods that consider dependence between variables.

Historical ocean reanalyses (1900–2010) using different data assimilation strategies

Two historical ocean reanalyses that cover the period from 1900 to 2010 are performed. One ocean reanalysis (CHOR_AS) assimilates both sea‐surface temperature (SST) and vertical profile data with a 3D‐Var assimilation scheme, while the other one (CHOR_RL) assimilates vertical profile data, but is nudged to monthly Hadley Centre global sea Ice coverage and SST (HadISST) reconstructed fields. The reanalyses are thus designed to investigate the impact of different strategies for constraining the SST in long‐term assimilation experiments and the feasibility of assimilating subsurface observations in the same kind of experiments. The atmospheric forcing comes from the ensemble mean of the Twentieth‐Century Reanalysis version 2 (20CRv2). Due to biases being particularly large at high latitudes, corrections are applied to the atmospheric forcing, based on the comparison with ECMWF's ERA‐Interim during the overlapped period. In general, these ocean reanalyses capture trends and variability of key ocean parameters. A warm bias of SST in CHOR_AS in the first half of the century leads to incorrect long‐term trends. However, SST anomalies in both CHOR_AS and CHOR_RL are realistically captured. Heat contents at 0–300, 0–700 and 0–2000 m have an increasing trend in both CHOR_RL and CHOR_AS, consistent with available datasets for the second half of the century. In order to reproduce realistic mass transport, mixed‐layer depth and eddy kinetic energy, it is necessary to assimilate vertical profile data, although this leads to some discontinuities in the reanalysis time series. Based on these findings, we suggest ideas for the requirements of future historical ocean reanalyses, such as related to the necessity of applying bias correction and minimizing the effects of the observing network discontinuities. Meanwhile, the comparison between CHOR_RL and CHOR_AS shows that both direct assimilation and the nudging scheme have their advantages and disadvantages. A combination of these two schemes is desirable in future work.

Bias correcting precipitation forecasts to improve the skill of seasonal streamflow forecasts

Abstract. Meteorological centres make sustained efforts to provide seasonal forecasts that are increasingly skilful, which has the potential to benefit streamflow forecasting. Seasonal streamflow forecasts can help to take anticipatory measures for a range of applications, such as water supply or hydropower reservoir operation and drought risk management. This study assesses the skill of seasonal precipitation and streamflow forecasts in France to provide insights into the way bias correcting precipitation forecasts can improve the skill of streamflow forecasts at extended lead times. We apply eight variants of bias correction approaches to the precipitation forecasts prior to generating the streamflow forecasts. The approaches are based on the linear scaling and the distribution mapping methods. A daily hydrological model is applied at the catchment scale to transform precipitation into streamflow. We then evaluate the skill of raw (without bias correction) and bias-corrected precipitation and streamflow ensemble forecasts in 16 catchments in France. The skill of the ensemble forecasts is assessed in reliability, sharpness, accuracy and overall performance. A reference prediction system, based on historical observed precipitation and catchment initial conditions at the time of forecast (i.e. ESP method) is used as benchmark in the computation of the skill. The results show that, in most catchments, raw seasonal precipitation and streamflow forecasts are often more skilful than the conventional ESP method in terms of sharpness. However, they are not significantly better in terms of reliability. Forecast skill is generally improved when applying bias correction. Two bias correction methods show the best performance for the studied catchments, each method being more successful in improving specific attributes of the forecasts: the simple linear scaling of monthly values contributes mainly to increasing forecast sharpness and accuracy, while the empirical distribution mapping of daily values is successful in improving forecast reliability.

Technical Note: The impact of spatial scale in bias correction of climate model output for hydrologic impact studies

Abstract. Statistical downscaling is a commonly used technique for translating large-scale climate model output to a scale appropriate for assessing impacts. To ensure downscaled meteorology can be used in climate impact studies, downscaling must correct biases in the large-scale signal. A simple and generally effective method for accommodating systematic biases in large-scale model output is quantile mapping, which has been applied to many variables and shown to reduce biases on average, even in the presence of non-stationarity. Quantile-mapping bias correction has been applied at spatial scales ranging from hundreds of kilometers to individual points, such as weather station locations. Since water resources and other models used to simulate climate impacts are sensitive to biases in input meteorology, there is a motivation to apply bias correction at a scale fine enough that the downscaled data closely resemble historically observed data, though past work has identified undesirable consequences to applying quantile mapping at too fine a scale. This study explores the role of the spatial scale at which the quantile-mapping bias correction is applied, in the context of estimating high and low daily streamflows across the western United States. We vary the spatial scale at which quantile-mapping bias correction is performed from 2° ( ∼ 200 km) to 1∕8° ( ∼ 12 km) within a statistical downscaling procedure, and use the downscaled daily precipitation and temperature to drive a hydrology model. We find that little additional benefit is obtained, and some skill is degraded, when using quantile mapping at scales finer than approximately 0.5° ( ∼ 50 km). This can provide guidance to those applying the quantile-mapping bias correction method for hydrologic impacts analysis.

Comparing bias correction methods in downscaling meteorological variables for a hydrologic impact study in an arid area in China

Abstract. Water resources are essential to the ecosystem and social economy in the desert and oasis of the arid Tarim River basin, northwestern China, and expected to be vulnerable to climate change. It has been demonstrated that regional climate models (RCMs) provide more reliable results for a regional impact study of climate change (e.g., on water resources) than general circulation models (GCMs). However, due to their considerable bias it is still necessary to apply bias correction before they are used for water resources research. In this paper, after a sensitivity analysis on input meteorological variables based on the Sobol' method, we compared five precipitation correction methods and three temperature correction methods in downscaling RCM simulations applied over the Kaidu River basin, one of the headwaters of the Tarim River basin. Precipitation correction methods applied include linear scaling (LS), local intensity scaling (LOCI), power transformation (PT), distribution mapping (DM) and quantile mapping (QM), while temperature correction methods are LS, variance scaling (VARI) and DM. The corrected precipitation and temperature were compared to the observed meteorological data, prior to being used as meteorological inputs of a distributed hydrologic model to study their impacts on streamflow. The results show (1) streamflows are sensitive to precipitation, temperature and solar radiation but not to relative humidity and wind speed; (2) raw RCM simulations are heavily biased from observed meteorological data, and its use for streamflow simulations results in large biases from observed streamflow, and all bias correction methods effectively improved these simulations; (3) for precipitation, PT and QM methods performed equally best in correcting the frequency-based indices (e.g., standard deviation, percentile values) while the LOCI method performed best in terms of the time-series-based indices (e.g., Nash–Sutcliffe coefficient, R2); (4) for temperature, all correction methods performed equally well in correcting raw temperature; and (5) for simulated streamflow, precipitation correction methods have more significant influence than temperature correction methods and the performances of streamflow simulations are consistent with those of corrected precipitation; i.e., the PT and QM methods performed equally best in correcting flow duration curve and peak flow while the LOCI method performed best in terms of the time-series-based indices. The case study is for an arid area in China based on a specific RCM and hydrologic model, but the methodology and some results can be applied to other areas and models.

Bridging the gap between climate models and impact studies: the FORESEE Database.

Studies on climate change impacts are essential for identifying vulnerabilities and developing adaptation options. However, such studies depend crucially on the availability of reliable climate data. In this study, we introduce the climatological database called FORESEE (Open Database for Climate Change Related Impact Studies in Central Europe), which was developed to support the research of and adaptation to climate change in Central and Eastern Europe: the region where knowledge of possible climate change effects is inadequate. A questionnaire‐based survey was used to specify database structure and content. FORESEE contains the seamless combination of gridded daily observation‐based data (1951–2013) built on the E‐OBS and CRU TS datasets, and a collection of climate projections (2014–2100). The future climate is represented by bias‐corrected meteorological data from 10 regional climate models (RCMs), driven by the A1B emission scenario. These latter data were developed within the frame of the ENSEMBLES FP6 project. Although FORESEE only covers a limited area of Central and Eastern Europe, the methodology of database development, the applied bias correction techniques, and the data dissemination method, can serve as a blueprint for similar initiatives.

Multi‐GCM by multi‐RAM experiments for dynamical downscaling on summertime climate change in Hokkaido

AbstractThe experiments with three general circulation models (GCMs) by three regional atmospheric models (RAMs) for the dynamical downscaling (DDS) have been performed to evaluate the uncertainty in the global warming response during summertime in Hokkaido, Japan. The results of a 10‐year RAM integration nested into GCM under present or future climate conditions were synthesized after applying bias correction. For the target decades during which the global‐mean temperature increases by 2 K in each GCM, the DDS results indicate that surface air temperature and precipitation mostly depend on the GCM imposed as the lateral boundary condition.

Applying bias correction for merging rain gauge and radar data

Weather radar provides areal rainfall information with very high temporal and spatial resolution. Radar data has been implemented in several hydrological applications despite the fact that the data suffers from varying sources of error. Several studies have attempted to propose methods for solving these problems. Additionally, weather radar usually underestimates or overestimates the rainfall amount. In this study, a new method is proposed for correcting radar data by implementing the quantile mapping bias correction method. Then, the radar data is merged with observed rainfall by conditional merging and kriging with external drift interpolation techniques. The merging product is analysed regarding the sensitivity of the two investigated methods to the radar data quality. After implementing bias correction, not only did the quality of the radar data improve, but also the performance of the interpolation techniques using radar data as additional information. In general, conditional merging showed greater sensitivity to radar data quality, but performed better than all the other interpolation techniques when using bias corrected radar data. Furthermore, a seasonal variation of interpolation performances has in general been observed. A practical example of using radar data for disaggregating stations from daily to hourly temporal resolution is also proposed in this study.

How well do general circulation models represent low‐frequency rainfall variability?

AbstractGeneral circulation models (GCMs) provide reliable simulations of global‐ and continental‐scale atmospheric variables, yet have limited skill in simulating variables important for water resource management at regional to catchment scales. GCM simulations suffer from a range of uncertainties leading to transient (changing over time) and systemic (consistent over time) biases in the output when compared to observed records. An important GCM bias in managing water resources infrastructure is the underrepresentation of low‐frequency variability a characteristic central to the simulation of floods and droughts. This study presents a performance metric, the aggregated persistence score (APS), which is used to assess the reliability of GCMs in simulating low‐frequency rainfall variability. The APS identifies regions where GCMs poorly represent the amount of variability seen in the observed precipitation. This study calculated the APS at monthly aggregations for GCM precipitation as well as GCM precipitation that was bias‐corrected to better represent low‐frequency variability. It was found that there were (1) large spatial variations in the skill of GCMs to capture observed rainfall persistence, (2) widespread undersimulation of rainfall persistence characteristics in GCMs, and (3) substantial improvement in rainfall persistence after applying bias correction.

Does bias correction increase reliability of flood projections under climate change? A case study of large rivers in Germany

ABSTRACTThere is a large uncertainty associated with flood projections driven by different climate scenarios. The bias‐corrected regional climate scenarios are widely used to drive hydrological models in climate impact studies, but there are also doubts and questions about the application of bias correction (BC) methods. This study aims to investigate the performance and impacts of BCs on flood projections in Germany. The distribution mapping method was applied to correct the climate data from the regional climate models (RCMs) CCLM (Cosmo‐Climate Local Model) and REMO (REgional MOdel) developed in Germany. The results show that BC can effectively reduce bias in the simulated average annual discharge, but the uncertainty of simulated floods remains due to the imperfect correction of extreme precipitations. About 75% of the change directions in the 50‐year flood discharge remain the same before and after the BC was used. The relatively short control period of 40 years and the assumption of stationarity of the BC method are two important and problematic issues for flood projections. Hence, it is difficult to prove that BC can increase reliability of flood projections. The direct use of RCM outputs for the control and scenario periods may still be useful for flood impact studies. In addition, a bilateral analysis of RCM and hydrological model performance involving the meteorologists and hydrologists could be helpful for reducing the bias of the RCM outputs in the future.

A trend-preserving bias correction – the ISI-MIP approach

Abstract. Statistical bias correction is commonly applied within climate impact modelling to correct climate model data for systematic deviations of the simulated historical data from observations. Methods are based on transfer functions generated to map the distribution of the simulated historical data to that of the observations. Those are subsequently applied to correct the future projections. Here, we present the bias correction method that was developed within ISI-MIP, the first Inter-Sectoral Impact Model Intercomparison Project. ISI-MIP is designed to synthesise impact projections in the agriculture, water, biome, health, and infrastructure sectors at different levels of global warming. Bias-corrected climate data that are used as input for the impact simulations could be only provided over land areas. To ensure consistency with the global (land + ocean) temperature information the bias correction method has to preserve the warming signal. Here we present the applied method that preserves the absolute changes in monthly temperature, and relative changes in monthly values of precipitation and the other variables needed for ISI-MIP. The proposed methodology represents a modification of the transfer function approach applied in the Water Model Intercomparison Project (Water-MIP). Correction of the monthly mean is followed by correction of the daily variability about the monthly mean. Besides the general idea and technical details of the ISI-MIP method, we show and discuss the potential and limitations of the applied bias correction. In particular, while the trend and the long-term mean are well represented, limitations with regards to the adjustment of the variability persist which may affect, e.g. small scale features or extremes.

Bias Correction Techniques to Improve Air Quality Ensemble Predictions: Focus on O3 and PM Over Portugal

Five air quality models were applied over Portugal for July 2006 and used as ensemble members. Each model was used, with its original set up in terms of meteorology, parameterizations, boundary conditions and chemical mechanisms, but with the same emission data. The validation of the individual models and the ensemble of ozone (O3) and particulate matter (PM) is performed using monitoring data from 22 background sites. The ensemble approach, based on the mean and median of the five models, did not improve significantly the skill scores due to large deviations in each ensemble member. Different bias correction techniques, including a subtraction of the mean bias and a multiplicative ratio adjustment, were implemented and analysed. The obtained datasets were compared against the individual modelled outputs using the bias, the root mean square error (RMSE) and the correlation coefficient. The applied bias correction techniques also improved the skill of the individual models and work equally well over the entire range of observed O3 and PM values. The obtained results revealed that the best bias correction technique was the ratio adjustment with a 4-day training period, demonstrating significant improvements for both analysed pollutants. The increase in the ensemble skill found comprehends a bias reduction of 88 % for O3, and 92 % for PM10, and also a decrease in 23 % for O3 and 43 % for PM10 in what concerns the RMSE. In addition, a spatial bias correction approach was also examined with successful skills comparing to the uncorrected ensemble for both pollutants. © 2013 Springer Science+Business Media Dordrecht.

Impact of Bias-Correction Type and Conditional Training on Bayesian Model Averaging over the Northeast United States

Abstract The performance of a multimodel ensemble over the northeast United States is evaluated before and after applying bias correction and Bayesian model averaging (BMA). The 13-member Stony Brook University (SBU) ensemble at 0000 UTC is combined with the 21-member National Centers for Environmental Prediction (NCEP) Short-Range Ensemble Forecast (SREF) system at 2100 UTC. The ensemble is verified using 2-m temperature and 10-m wind speed for the 2007–09 warm seasons, and for subsets of days with high ozone and high fire threat. The impacts of training period, bias-correction method, and BMA are explored for these potentially hazardous weather events using the most recent consecutive (sequential training) and most recent similar days (conditional training). BMA sensitivity to the selection of ensemble members is explored. A running mean difference between forecasts and observations using the last 14 days is better at removing temperature bias than is a cumulative distribution function (CDF) or linear regression approach. Wind speed bias is better removed by adjusting the modeled CDF to the observation. High fire threat and ozone days exhibit a larger cool bias and a greater negative wind speed bias than the warm-season average. Conditional bias correction is generally better at removing temperature and wind speed biases than sequential training. Greater probabilistic skill is found for temperature using both conditional bias correction and BMA compared to sequential bias correction with or without BMA. Conditional and sequential BMA results are similar for 10-m wind speed, although BMA typically improves probabilistic skill regardless of training.

An improved standardization procedure to remove systematic low frequency variability biases in GCM simulations

The quality of the absolute estimates of general circulation models (GCMs) calls into question the direct use of GCM outputs for climate change impact assessment studies, particularly at regional scales. Statistical correction of GCM output is often necessary when significant systematic biasesoccur between the modeled output and observations. A common procedure is to correct the GCM output by removing the systematic biases in low‐order moments relative to observations or to reanalysis data at daily, monthly, or seasonal timescales. In this paper, we present an extension of a recently published nested bias correction (NBC) technique to correct for the low‐ as well as higher‐order moments biases in the GCM‐derived variables across selected multiple time‐scales. The proposed recursive nested bias correction (RNBC) approach offers an improved basis for applying bias correction at multiple timescales over the original NBC procedure. The method ensures that the bias‐corrected series exhibits improvements that are consistently spread over all of the timescales considered. Different variations of the approach starting from the standard NBC to the more complex recursive alternatives are tested to assess their impacts on a range of GCM‐simulated atmospheric variables of interest in downscaling applications related to hydrology and water resources. Results of the study suggest that three to five iteration RNBCs are the most effective in removing distributional and persistence related biases across the timescales considered.

Methods to apply probabilistic bias analysis to summary estimates of association

Bias analysis methods are developed for application to 2 x 2 tables, which may be crude or stratified data. Methods for application to associations adjusted for multiple covariates, such as associations from regression modeling, are rarely seen. We have developed probabilistic methods to evaluate bias from disease misclassification or an unmeasured confounder that can be applied to adjusted estimates of association. Rather than applying bias correction methods that rearrange data within 2 x 2 tables, we have applied them to bias factors directly. We illustrate the methods by application to two pharmacoepidemiology problems. In example one, the adjusted odds ratio associating glucocorticoid use with the rate of basal cell carcinoma was 1.15 (95%CI 1.07, 1.25). With bias analysis to account for differential disease misclassification, the median odds ratio was 1.32 and the 95% simulation limits were 1.16 and 1.56. In example two, the adjusted odds ratio associating concomitant use of clopidogrel and proton pump inhibitors with recurrent myocardial infarction was 1.21 (95%CI 0.90, 1.61). With bias analysis to account for confounding by smoking, which was unmeasured, the median odds ratio was 1.15 with 95% simulation interval 0.85 to 1.55. Methods to apply probabilistic bias analysis to adjusted estimates of association can be implemented if a bias factor can be calculated directly from the bias model. This strategy requires that the bias is independent of confounding by measured variables, or requires that the dependence be incorporated into the bias model, as illustrated in an extension of the second example.

A robust method to estimate the intracranial volume across MRI field strengths (1.5T and 3T)

As population-based studies may obtain images from scanners with different field strengths, a method to normalize regional brain volumes according to intracranial volume (ICV) independent of field strength is needed. We found systematic differences in ICV estimation, tested in a cohort of healthy subjects (n=5) that had been imaged using 1.5T and 3T scanners, and confirmed in two independent cohorts. This was related to systematic differences in the intensity of cerebrospinal fluid (CSF), with higher intensities for CSF located in the ventricles compared with CSF in the cisterns, at 3T versus 1.5T, which could not be removed with three different applied bias correction algorithms. We developed a method based on tissue probability maps in MNI (Montreal Neurological Institute) space and reverse normalization (reverse brain mask, RBM) and validated it against manual ICV measurements. We also compared it with alternative automated ICV estimation methods based on Statistical Parametric Mapping (SPM5) and Brain Extraction Tool (FSL). The proposed RBM method was equivalent to manual ICV normalization with a high intraclass correlation coefficient (ICC=0.99) and reliable across different field strengths. RBM achieved the best combination of precision and reliability in a group of healthy subjects, a group of patients with Alzheimer's disease (AD) and mild cognitive impairment (MCI) and can be used as a common normalization framework.

Improving assimilation of SeaWiFS data by the application of bias correction with a local SEIK filter

Ocean-biogeochemical models show typically significant errors in the representation of chlorophyll concentrations. The model state can be improved by the assimilation of satellite chlorophyll data with algorithms based on the Kalman filter. However, these algorithms do usually not account for the possibility that the model prediction contains systematic errors in the form of model bias. Accounting explicitly for model biases can improve the assimilation performance. To study the effect of bias estimation on the estimation of surface chlorophyll concentrations, chlorophyll data from the Sea-viewing Wide Field-of-view Sensor (SeaWiFS) are assimilated on a daily basis into the NASA Ocean Biogeochemical Model (NOBM). The assimilation is performed by the ensemble-based SEIK filter combined with an online bias correction scheme. The SEIK filter is simplified here by the use of a static error covariance matrix. The performance of the filter algorithm is assessed by comparison with independent in situ data over the 7-year period 1998–2004. The bias correction results in significant improvements of the surface chlorophyll concentrations compared to the assimilation without bias estimation. With bias estimation, the daily surface chlorophyll estimates from the assimilation show about 3.3% lower error than SeaWiFS data. In contrast, the error in the global surface chlorophyll estimate without bias estimation is 10.9% larger than the error of SeaWiFS data.

PODEX - PhOtometric Data EXtractor

New reduction tools are required to handle the increasing amount of photometric CCD data within a reasonable time. podex is an IDL based tool developed to optimize the time needed for extracting light curves out of photometric CCD data. The GUI based semi-automatic reduction tool provides the possibility to apply bias correction, flat fielding, and background modeling to the raw data. After extracting magnitudes even in crowded fields, a color–dependent extinction correction and a conversion to relative magnitudes is performed. The output of podex are completely reduced light curves of an arbitrary number of stars on the CCD images. As the conditions between various nights and observing runs change continuously and sometimes dramatically, full interactivity throughout the entire reduction procedure via a powerful graphical user interface is a big advantage over all black-box procedures known to us.

THE IMPORTANCE OF CONSIDERING PREDICTION VARIANCE IN ANALYSES USING PHOTOGRAMMETRIC MASS ESTIMATES

AbstractDevelopment and application of photogrammetric mass‐estimation techniques in marine mammal studies is becoming increasingly common. When a photogrammetrically estimated mass is used as a covariate in regression modeling, the error associated with estimating mass induces bias in regression statistics and decreases model explanatory power. Thus, it is important to understand and account for prediction variance when addressing ecological questions that require use of estimated mass values. In a simulation study based on data collected from Weddell seals, we developed regression models of pup weaning mass as a function of maternal postparturition mass where maternal mass was directly measured and second where maternal mass was photogrammetrically estimated. We demonstrate that when estimated mass was used, the regression coefficient was biased toward zero and the coefficient of determination was 30% less than the value obtained when using maternal postparturition mass obtained from direct measurement. After applying bias correction procedures, however, the regression coefficient and coefficient of determination were within 2% of their true values. To effectively use photogrammetrically estimated masses, prediction variance should be understood and accounted for in all analyses. The methods presented in this paper are effective and simple techniques to explore and account for prediction variance.

Simulation of Crop Yields Using ERA-40: Limits to Skill and Nonstationarity in Weather–Yield Relationships

Abstract Reanalysis data provide an excellent test bed for impacts prediction systems, because they represent an upper limit on the skill of climate models. Indian groundnut (Arachis hypogaea L.) yields have been simulated using the General Large-Area Model (GLAM) for annual crops and the European Centre for Medium-Range Weather Forecasts (ECMWF) 40-yr reanalysis (ERA-40). The ability of ERA-40 to represent the Indian summer monsoon has been examined. The ability of GLAM, when driven with daily ERA-40 data, to model both observed yields and observed relationships between subseasonal weather and yield has been assessed. Mean yields were simulated well across much of India. Correlations between observed and modeled yields, where these are significant, are comparable to correlations between observed yields and ERA-40 rainfall. Uncertainties due to the input planting window, crop duration, and weather data have been examined. A reduction in the root-mean-square error of simulated yields was achieved by applying bias correction techniques to the precipitation. The stability of the relationship between weather and yield over time has been examined. Weather–yield correlations vary on decadal time scales, and this has direct implications for the accuracy of yield simulations. Analysis of the skewness of both detrended yields and precipitation suggest that nonclimatic factors are partly responsible for this nonstationarity. Evidence from other studies, including data on cereal and pulse yields, indicates that this result is not particular to groundnut yield. The detection and modeling of nonstationary weather–yield relationships emerges from this study as an important part of the process of understanding and predicting the impacts of climate variability and change on crop yields.