Drought assessment has been outpaced by climate change: empirical arguments for a paradigm shift

The study addresses how assumptions of climate stationarity embedded in U.S. drought monitoring and water resource management bias contemporary assessments under rapid climate change. National drought monitoring (USDM) synthesizes multiple datasets to classify drought severity, triggering substantial federal disaster relief. Many operational drought metrics standardize anomalies using historical probability distributions, with recommendations historically favoring long periods of record (~60+ years) to reduce parameter uncertainty. However, in a non-stationary climate, the choice of reference period critically affects perceived drought severity. The authors hypothesize that using long, period-of-record climatologies biases standardized drought metrics relative to contemporary conditions, and that shorter, contemporary climatologies (e.g., 30-year moving windows) better represent present drought risks.

Prior work established best practices for computing drought metrics like the Standardized Precipitation Index (SPI), often recommending long records (≈60+ years) to stabilize parameter estimates, a position reflected in World Meteorological Organization guidance. In contrast, climate science standard practice uses 30-year climate normals updated each decade to reflect variability and change. Milly et al. (2008) argued that stationarity is no longer a valid assumption for water management. Numerous drought indicators (SPI, SPEI) and anomaly frameworks (z-scores, empirical percentiles) rely on historical distributions and thus may inherit biases under non-stationarity. This paper builds on these insights by quantitatively examining how non-stationarity alters fitted probability distributions and the resulting standardized drought indices.

Data and study scope: The authors analyzed precipitation from GHCN stations across the contiguous U.S., focusing on summer (June 1–August 31). Stations included those operating in 1950 or earlier with complete records through 2020 to enable both long period-of-record and contemporary windows. Completeness windows were aligned to timescales: 30-, 60-, and 90-day SPI required continuous data back to 30, 60, or 90 days prior to each date of interest. Drought metric and distribution fitting: SPI was computed at 30-, 60-, and 90-day timescales by aggregating daily precipitation to the target accumulation and fitting two-parameter gamma distributions to the aggregated series. Parameter estimation used unbiased probability-weighted moments (L-/t-moments) and the L-moment method. The cumulative distribution function (CDF) of aggregated precipitation was transformed to a standard normal variate to obtain SPI. Analyses were implemented in R (RNOAA for data access; impod for distribution fitting). Moving-window vs. period-of-record reference frames: For contemporary assessments, 30-year moving windows (e.g., 1991–2020 for year 2020) were contrasted with long, time-integrated period-of-record climatologies (>70 years where available). Bias was computed as the median daily difference (POR SPI minus contemporary 30-year SPI) over June–August during selected intervals, including episodes where POR SPI indicated very dry conditions (SPI < −2). Monte Carlo simulations: Two sets of simulations evaluated SPI error as a function of climatology length (number of observations/years): - Stationary climate: Random samples were drawn from a single known gamma distribution. Gamma parameters were re-estimated via L-moments for sample sizes ranging from small (≈2) up to 100 observations, and the fitted CDF and SPI were compared to the true generating distribution. Across 1,000 iterations, convergence of parameter estimates and declines in absolute CDF error and absolute SPI error were summarized. Reported results for one generating pair (true rate ≈0.03–0.05, shape ≈2.35–2.5) showed rapid error declines up to ≈30 observations, with median absolute SPI errors of roughly 0.16 (30 obs), 0.11 (60 obs), and 0.08 (90 obs). - Non-stationary climate: Generative distributions varied over time using observed 30-year moving-window gamma parameters from 11 GHCN sites for specific dates (e.g., August 1) and timescales (30, 60, 90 days). For each climatology length, SPI was estimated by fitting a gamma distribution to random samples drawn from the time-varying generative process; the absolute SPI error was computed relative to the known SPI for the most recent distribution (operationally analogous to "today"). Results were summarized by medians and interquartile ranges over 1,000 simulations per configuration and across sites and timescales. Spatial bias assessment: Using 1,934 GHCN sites, the study mapped SPI bias (POR vs. contemporary 30-year SPI) during summers (1991–2020), including analysis of periods where POR SPI < −2 to assess bias under extreme dryness. Hotspot regions and timescale dependencies (30-, 60-, 90-day) were identified. Data/code: All data are publicly available; derived data archived in Zenodo (accession 504780). Code is available at https://github.com/iholyman/drought-year-sensitivity.

- In stationary simulations, uncertainty in gamma parameter estimates and absolute SPI error declined rapidly with increasing sample size, with median absolute SPI errors around 0.16 (30 observations), 0.11 (60), and 0.08 (90), confirming that longer records reduce error under stationarity. - In non-stationary simulations based on observed moving-window parameters, absolute SPI error did not consistently decrease with longer climatologies. At some sites, error increased substantially with longer records, reflecting dilution by outdated conditions. Example: Clemson University, SC (USC00381770), 30-day SPI for August 1 showed median absolute SPI error ≈0.18 (IQR ≈0.227) with 30 observations, increasing to ≈0.49 (±0.131) with 90 observations. - The relationship between SPI error and climatology length is site- and timescale-specific under non-stationarity; minima sometimes occurred below 30 years or above 30 years depending on local climate change velocity. - Across 1,934 GHCN sites, long period-of-record SPI often exhibited a dry bias relative to contemporary 30-year SPI, especially during very dry POR conditions (SPI < −2). During such periods, a dry bias occurred at 65.1% of stations (30-day), 61.2% (60-day), and 56.5% (90-day). - Spatially, dry-bias hotspots were prominent in the southwestern and southeastern U.S., with bias magnitude generally larger at shorter accumulation timescales (e.g., 30-day). - Overall, where climate is changing rapidly, longer, period-of-record reference frames can substantially mischaracterize contemporary drought severity compared to a 30-year moving window.

The findings directly address the hypothesis that non-stationarity undermines the reliability of long period-of-record climatologies for contemporary drought assessment. In rapidly changing climates, longer records incorporate conditions that are now less probable, biasing standardized indices such as SPI. This has practical implications because U.S. drought classifications trigger significant federal assistance; biased assessments risk misalignment between emerging drought risk and resource allocation. The study shows that a 30-year moving window often balances statistical stability with representativeness of current conditions. The non-stationarity issue extends beyond SPI to other metrics (e.g., SPEI), z-scores, and empirical percentiles that rely on historical distributions; temperature- and evaporative demand–driven metrics may be even more consistently biased due to strong, unidirectional trends. The authors advocate aligning operational drought monitoring with contemporary climate normals and considering advanced, locally adaptive statistical models where feasible.

Drought assessment practices that assume stationarity and favor very long climatologies are increasingly inconsistent with contemporary climate realities. Empirical analyses and simulations demonstrate that, under non-stationarity, longer period-of-record reference frames can increase error and introduce systematic bias, whereas a 30-year moving window often provides a more accurate portrayal of present drought risk. The authors recommend standardizing drought monitoring around contemporary, regularly updated climatologies and engaging the drought and water resources communities to modernize practices. Future research should evaluate adaptive, covariate-informed, or time-varying distribution models (e.g., GAMs, climate index–conditioned PDFs), extend analyses to additional variables (temperature, evapotranspiration, streamflow, snowpack), and assess operational impacts of adopting non-stationary methods.

- Analyses focus primarily on precipitation and SPI; other hydroclimatic variables and compound indices may exhibit different sensitivities. - Non-stationary simulations were based on observed moving-window parameters from a limited set of sites (11) for detailed simulations, which may not capture all regional behaviors, though broader spatial bias mapping used many stations. - Relationships between optimal climatology length and error are site- and timescale-specific; a uniform window may not be optimal everywhere. - The moving-window approach, while practical, still lags real-time changes and may not fully capture rapidly evolving distributions; more sophisticated time-varying models could perform better but are more complex to operationalize. - Data completeness requirements and station selection criteria may influence spatial coverage and generalizability; summer-focused analyses may not represent other seasons.