Increased drought severity tracks warming in the United States' largest river basin.

Across the Upper Missouri River Basin, the recent drought of 2000 to 2010, known as the "turn-of-the-century drought," was likely more severe than any in the instrumental record including the Dust Bowl drought. However, until now, adequate proxy records needed to better understand this event with regard to long-term variability have been lacking. Here we examine 1,200 y of streamflow from a network of 17 new tree-ring-based reconstructions for gages across the upper Missouri basin and an independent reconstruction of warm-season regional temperature in order to place the recent drought in a long-term climate context. We find that temperature has increasingly influenced the severity of drought events by decreasing runoff efficiency in the basin since the late 20th century (1980s) onward. The occurrence of extreme heat, higher evapotranspiration, and associated low-flow conditions across the basin has increased substantially over the 20th and 21st centuries, and recent warming aligns with increasing drought severities that rival or exceed any estimated over the last 12 centuries. Future warming is anticipated to cause increasingly severe droughts by enhancing water deficits that could prove challenging for water management.

In much of the western United States (hereafter “the West”), water demand (i.e., the combination of atmospheric demands, ecological requirements, and consumptive use) is approaching or has exceeded supply, making the threat of future drought an increasing concern for water managers (1–5). Prolonged drought can disrupt agricultural systems and economies (6–9), challenge river system control and navigation (10, 11), and complicate management of sensitive ecological resources (12, 13). Recently, ample evidence has emerged to suggest that the severity of several regional 21st-century droughts has exceeded the severity of historical drought events; these recent extreme droughts include the 2011 to 2016 California drought (14, 15) and the 2000 to 2015 (16, 17) drought in the Colorado River basin.

Conspicuously absent thus far from investigations of recent droughts has been the Missouri River, the longest river in North America draining the largest independent river basin in the United States (18). Similar to California (14) and the Upper Colorado River Basin (16, 17), parts of the early 21st century have been remarkably dry across the Upper Missouri River Basin (UMRB) (19). In fact, our assessment of streamflow for the UMRB suggests that the widespread drought period of 2000 to 2010, termed the “turn-of-the-century drought” by Cook et al. (19), was a period of observationally unprecedented and sustained hydrologic drought likely surpassing even the drought of the Dust Bowl period.

Northern Hemisphere summer temperatures are now likely higher than they have been in the last 1,200 y (20), and the unique combination of recent anomalously high temperatures (20) and severe droughts across much of the West (14, 16, 17) has led numerous researchers to revisit the role of temperature in changing the timing and efficiency of runoff in the new millennium (16, 21–24). Evidence suggests that across much of the West atmospheric moisture demands due to warming are reducing the effectiveness of precipitation in generating streamflow and ultimately surface-water supplies (16, 22–29).

The waters of the Upper Missouri River originate predominantly in the Rocky Mountains of Montana, Wyoming, and Colorado, where high-elevation catchments capture and store large volumes of water as winter snowpack that are later released as spring and early summer snowmelt (9). This mountain water is an important component of the total annual flow of the Missouri, accounting for roughly 30% of the annual discharge delivered to the Mississippi River on average, but ranging between 14% to more than 50% from year to year, most of which is delivered during the critical warm-season months (May through September) (9, 30). Across much of the UMRB, cool-season (October through May) precipitation stored as winter snowpack has historically been the primary driver of streamflow, with observed April 1 snow-water equivalent (SWE) usually accounting for at least half of the variability in observed streamflow from the primary headwaters regions (9). However, since the 1950s, warming spring temperatures have increasingly driven regional snowpack declines that have intensified since the 1980s (31–33). By 2006, these declines amounted to a low snowpack anomaly of unusual severity relative to the last 800 y and spanned the snow-dominated watersheds of the interior West (32). A recent reassessment of snowpack declines across the West by Mote et al. (33) suggests continued temperature-driven snowpack declines through 2016 totaling a volumetric storage loss of between 25 and 50 km3, which is comparable to the storage capacity of Lake Mead (∼36 km3), the United States’ largest reservoir.

Here we examine the extended record (ca. 800 to 2010 CE) of streamflow and the influence of temperature on drought through the Medieval Climate Anomaly, with a focus on the recent turn-of-the-century drought in the UMRB. The role of increasing temperature on streamflow and basin-wide drought is examined in the UMRB over the last 1,200 y by analyzing a basin-wide composite streamflow record developed from a network of 17 tree-ring–based reconstructions of streamflow for major gages in the UMRB (Fig. 1) (34) and an independent runoff-season (March through August) regional temperature reconstruction. We also explore the hydrologic implications (e.g., drought severity and spatial extent) and climatic drivers (temperature and precipitation) of the observed changes in streamflow across the UMRB and characterize shifts in the likelihood of extreme flow levels and reductions in runoff efficiency across the basin.

Fig. 1.

The Missouri River Basin and its subregions. The location of the Missouri River Basin within the continental United States (gray watershed, upper right) and the location of the five hydrologically distinct subregions (colored watersheds) that define the UMRB. Reconstructed gages used to develop the estimate of basin-wide mean annual streamflow are shown as triangles.

Results and Discussion

The Turn-of-the-Century Drought in a Long-Term Context.

Persistent streamflow deficits during the turn-of-the-century drought were greater than those observed at any other time since widespread gaging of streamflow began across the UMRB in the early 20th century (Figs. 2 and 3). However, a more robust understanding of how the turn-of-the-century drought compares to past droughts in the UMRB requires the multicentury perspective provided by paleoclimate and paleohydrologic data. Such datasets provide numerous historical events for comparison and have documented the occurrence of very severe drought events in the American Southwest (35–38), California (39), and the Southern Great Plains (40, 41) during the last millennium. In these other regions, several drought events of the last millennium, often identified as “megadroughts” (19, 39, 40), are unrivaled in recent times in terms of severity and/or duration. Until recently (34), no comparable proxy records of hydrologic drought existed for the UMRB.

Fig. 2.

Streamflow, temperature forcing of streamflow, drought severity, and spatial distributions of the five most severe droughts in the UMRB over the last 1,200 y. The time series of reconstructed streamflow (A), the 10-y spline of temperature forcing of streamflow (B), and decadal-scale drought deficits (C) for the UMRB. The red line in A shows the 10-y cubic smoothing spline of streamflow and the dashed horizontal line shows the long-term mean. Color in B denotes positive (blue) and negative (red) forcings. White hatching in B denotes the period of high uncertainty in the UMRB regional temperature reconstruction. Color in C denotes temperature in the basin over the periods of the various drought events. D shows the spatial distributions of maximum flow deficits during the five major droughts annotated in C; 1930 to 2010 are instrumental data and 800 to 1929 are reconstructed.

Fig. 3.

Climate forcing of streamflow. (A) The relative forcing of precipitation and temperature (arrows) on basin-wide mean annual streamflow (black line) in the UMRB. Colored arrows show the individual relative forcing estimates for each subregion of the UMRB for each year, n = 550. Colors denote which climate variable was more dominant in the combined forcing of streamflow relative to its long-term average influence. The y axis shows the relative magnitude and direction of that combined forcing. The direction of the arrows shows the direction of forcing of the temperature component of that combined forcing (up = supporting and down = suppressing streamflow). All data shown are derived from the 5-y cubic smoothing splines of streamflow and climate data. (B) The relative forcing of temperature on streamflow is determined as the temperature anomaly times its multiple-regression coefficient for predicting streamflow along with precipitation. The black line denotes the mean temperature forcing of temperature on streamflow for all subregions of the UMRB; 1930 to 2010 are instrumental streamflow data and 1900 to 1929 are reconstructed. Climate data are from PRISM.

Using a 1,200-y, basin-wide Upper Missouri River streamflow reconstruction (34), we place the severity and duration of the turn-of-the-century drought in the context of long-term hydroclimatic variability (Fig. 2). The reconstruction skillfully captures the observed variation of streamflow across the basin and was specifically designed for the assessment of drought conditions over time (34). Combining this record (800 to 1929) with the naturalized flow records (1930 to 2010) from across the UMRB, we developed a representative estimate of basin-wide average streamflow spanning 800 to 2010 CE (Fig. 2 and ). We then developed a record of drought deficits focused on decadal-scale variability from the basin-wide flow estimate by first defining drought events as any sequence of two or more years in which the 10-y cubic smoothing spline (42) of reconstructed or observed streamflow anomalies was negative. Drought severity was then quantified as the magnitude of flow deficits over the period of each drought event determined by the value of the spline itself in units of SDs (z-scores) (Fig. 2) (). The duration of each drought event was determined as the number of years the smoothed streamflow anomaly remained negative.

In terms of the most severe flow deficits, the driest years of the turn-of-the-century drought in the UMRB appear unmatched over the last 1,200 y (Fig. 2). Only a single event in the late 13th century rivaled the greatest deficits of this most recent event; however, the lowest point in the spline of streamflow during the 13th century drought was 0.13 SDs (s) higher than that of the turn-of-the-century drought. The robustness of this finding was tested in multiple ways including by using various spline lengths from 5 to 15 y to quantify drought deficits and by comparing the intensity [cumulative deficit/duration (43)] of drought deficits determined from the unsmoothed streamflow record over time. In all cases except for the highest degree of smoothing (>12 y), the deficits during the driest years of the turn-of-the-century drought exceed those of all earlier droughts in the record ().

In terms of duration, however, the 13th century drought was over three times the length of the turn-of-the-century drought, overlapping with the start of the severe and sustained “Great Drought” period in the American Southwest (35, 36) that has been implicated in the abandonment of Anasazi settlements in that region (44). Its length firmly places it in the league of other “megadroughts” reported in North America that are unprecedented in modern times in terms of duration. However, in the six centuries that followed, coincident with the period of the Little Ice Age (ca. 1300s to late 1800s), drought severity in the UMRB was relatively mild. Only a single event in the late 1500s and consistent with the timing of the 16th-century North American megadrought reported by Stahle et al. (41) rivaled the flow deficits common during the period of the Medieval Climate Anomaly. This long hiatus in drought severity was abruptly ended by the onset of the Dust Bowl drought in the 1930s, which produced the fourth-lowest streamflow departure in the 1,200-y record. This severe and sustained drought was followed 70 y later by the largest decadal-scale flow deficits on record during the turn-of-the-century drought. Thus, in terms of drought events with decadal persistence in streamflow, two of the four most severe droughts in the last 1,200 y appear to have occurred within the last century.

To contextualize the spatial extent and magnitude of the turn-of-the-century drought relative to the four other most severe events in the 1,200-y record, we compared the maximum negative departure of the 10-y spline within each subregion during each drought event (Fig. 2). The maximum deficit during the turn-of-the-century drought was more than 1 s below the long-term mean flow level in all subbasins of the UMRB with the most severe flow deficits centered on the Yellowstone basin and the Missouri Mainstem region. Drought severity during the driest decade of the 13th century drought was positioned over the northern tier of the UMRB while the maximum severity of both the 12th-century drought and the Dust Bowl drought was focused over the Missouri Headwaters and Mainstem regions. The fifth-most-severe drought, which occurred during the early 11th century, appears to be the only event to reach peak flow deficits during below-average temperatures (Fig. 2). The spatial distribution of this event was characterized by more moderate flow deficits evident across the entire basin (Fig. 2).

Twentieth- and 21st-Century Streamflow and Climate Relationships.

We explored the 20th- and 21st-century climatic drivers (temperature and precipitation) of the observed changes in UMRB streamflow primarily using naturalized streamflow records compiled from 31 gage records representing nearly every major subbasin in the UMRB. These “naturalized” records represent instrumental-period measurements of streamflow with human influences such as upstream withdrawals, diversions, and reservoir operations removed. Using hierarchical clustering of the streamflow data, five hydrologically distinct subregions within the UMRB are evident: the Northern Tributaries, the Missouri Mainstem region, the Missouri Headwaters, the Yellowstone River, and the Platte River (Fig. 1) (34). We then generated composite records of streamflow for each of these subregions to assess the basin-specific climatic influences on streamflow by averaging standardized reconstructed flow (1900 to 1929) joined to observed (1930 to 2010) flow from those constituent gage records that covered a common period of 1900 through the turn-of-the-century drought (1900 to 2010) (17 records total; Fig. 1). These streamflow records were then compared with subbasin average temperature and precipitation records derived from the 4-km × 4-km gridded Precipitation-elevation Regression on Independent Slopes Model (PRISM) (45) and the Vose et al. (46) climate datasets to examine long-term relationships between hydroclimate and streamflow in the UMRB. Both datasets were used for comparison in these analyses since uncertainties exist in all gridded climate datasets that may affect the results. We also considered that some evidence from the adjacent Columbia River basin and western portions of the UMRB suggests possible underestimation of mountain precipitation in gridded climate datasets during earlier parts of the record (47). Additional analyses using modified precipitation datasets were carried out to estimate how potential underestimation of precipitation could affect our results. See for additional explanation of these uncertainties and supporting analyses. Importantly, the analyses of 20th- and 21st-century climate and streamflow relationships were found to be largely insensitive to the choice of climate dataset used.

The relative influence of temperature on streamflow in each of the five subregions of the UMRB after accounting for the influence of observed precipitation is shown in Fig. 3. Temperature negatively influences streamflow in the UMRB in general, explaining roughly 6% of the variability in instrumental streamflow alone over the 20th and 21st centuries after accounting for precipitation, which explains ∼45%. This estimate is similar to estimates of the influence of temperature on Upper Colorado streamflow (∼6%) (15, 23). In the UMRB, a defined shift in the relative influence of temperature on the generation of streamflow is evident since 1984 (Fig. 3 ). This is consistent with the findings of numerous studies that have identified a distinct shift in the behavior of various biophysical systems ranging from plant phenology to snowmelt timing across North America and beyond that is centered on the mid-1980s (31, 48–51). The year 1984 marks a clear and substantial shift in the negative temperature influence on streamflow relative to preceding decades of the instrumental record and is consistent across all major subbasins of the UMRB (Fig. 3). Prior to 1984, observed streamflow was higher than expected given observed precipitation due to the occurrence of relatively cooler temperatures over this period. However, after 1984, observed precipitation translated to lower than expected streamflow due to warmer temperatures. Consequently, average natural flows across the UMRB declined to levels not seen since the Dust Bowl era by the late 1980s and exceeded those flow reductions by the early 2000s (Fig. 3). The most likely timing of this shift in the UMRB was determined using a series of different length (5 to 30 y) moving-window t tests on the time series of the precipitation-adjusted temperature forcing of streamflow. This produced nested probabilities of the timing of a significant shift in the temperature influence occurring on both short and long timescales, which peak at 1984 (P < 0.001 on average).

The timing of reduced streamflow and the shifting influence of temperature on water supplies in the UMRB mirrors the snowpack declines in the headwaters of the UMRB from the mid-1970s to late 1980s, which have been attributed in large part to rising spring temperatures (31, 33). Declines in both SWE and snow fraction (ratio of snow to total precipitation) then intensified into the early 2000s (9), coincident with the strongest negative temperature forcing and lowest UMRB streamflow (Fig. 3 ) of the turn-of-the-century drought.

Impacts and Potential Mechanisms Underlying the Observed Changes in the Influence of Temperature on UMRB Streamflow.

The 20th- and 21st-century temperature forcing and drought severity records display close synchrony between the influence of warming on streamflow and increasing hydrologic drought severity, suggesting a strong mechanistic link between the two (Fig. 2 ). We investigated both the impacts and potential mechanisms of this linkage over the period of available precipitation data (since 1900) and the highest-quality portion of the reconstructed temperature and streamflow records (since 1800). This allowed us to describe the changes in the likelihood of extreme hydroclimatic conditions within the basin over the last two centuries and the efficiency of streamflow generation relative to changes in temperature since 1900.

The relationship between temperature and streamflow extremes over time was examined by quantifying the likelihood of their cooccurrence by tallying years where standardized values of both temperature and streamflow fall 1 s or greater from their respective long-term means over three time periods. The time periods assessed were 1800 to 2010, 1900 to 2010, and 1984 to 2010, when the temperature–streamflow relationship changed distinctly in the climate–streamflow analysis (Fig. 3). A greater occurrence of hot–dry extremes in the UMRB since 1900 is evident with ∼81% of extreme years falling into this category, while ∼59% of extreme years would be classified as hot–dry years over the period since 1800 (Fig. 4). Restricting the extreme year analysis to only the events since 1900, we find that 53% of extreme years since 1900 were coeval hot–dry years. Since 1984, every extreme year has been a hot–dry year (Fig. 4), representing a substantially greater likelihood of hot–dry extremes across the basin since 1984 relative to the period of 1900 to 2010. The combination of elevated air temperature and low streamflow presents a dual challenge for water managers in the UMRB where both agricultural irrigation demands and in-stream water quality for aquatic species are top management priorities (52). High temperatures and low flows simultaneously increase heat stress and evaporative demand on crops while reducing available water for irrigation. Likewise, low-flow conditions restrict available habitat and exacerbate the risk of excessive water temperatures for aquatic species during anomalously warm years (53).

Fig. 4.

Extremes in the temperature–streamflow relationship. (A) The relationship between temperature and streamflow anomalies for the period 1800 to 2010 (all points, n = 210). Colored points show those years in which both temperature and streamflow values were greater than one SD (s) from their mean levels (extremes). Red points show the extremes recorded from 1900 to 2010. (B) The temperature–streamflow anomaly relationship from 1900 to 2010 (all points, n = 110) and extremes (colored points). Red points show the extremes recorded from 1984 to 2010. Here extremes are defined by the variability in the records from 1900 to 2010 compared with that of 1800 to 2010 shown in A.

Severe drought in the UMRB and elsewhere in the West is primarily the result of regional precipitation deficits with evidence pointing to an increasing temperature influence (15, 22). Since the beginning of the 20th century, UMRB runoff-season temperature has warmed considerably (1.4 ± 0.6 °C) and significantly (P < 0.001) (), and increasing temperature can reduce streamflow by removing water from the landscape via evapotranspiration (ET) before it reaches rivers and streams. On the other hand, warmer air temperatures and meteorological drought are also physically linked because a lack of moisture at the surface limits the conversion of down-welling long-wave radiation to latent heat (evaporation) and increases sensible heating (54). This complimentary relationship (55, 56) between temperature as both a driver of potential ET, and an indicator of actual ET, limits the mechanistic inference that can be gleaned from observed relationships between temperature and streamflow alone. Rather, the role of temperature in streamflow generation may more clearly be inferred directly from the relationship between precipitation and streamflow. This is because, over the long term, precipitation not realized as streamflow leaves the landscape primarily as ET, and potential ET is strongly controlled by temperature (57).

We examined the relationship between UMRB precipitation and streamflow directly by estimating the runoff efficiency (RE) (defined here as the difference between streamflow and precipitation anomalies) (28) across the basin from 1900 to 2010. We found significant evidence for a reduction in RE over the 20th and 21st centuries across the Upper Missouri Basin (Fig. 5 ). In an analysis of water-year flows for each gage record where combined reconstructed and observed records are complete from 1900 to 2010, warmer temperatures and reduced RE are apparent and significantly different (P < 0.001) across the UMRB. These changes are clear when comparing time periods before and after 1984 (Fig. 5 ), as well as when comparing the notable historical droughts of record, with the turn-of-the-century drought exhibiting significantly (P < 0.001) lower REs than even the Dust Bowl drought (Fig. 5). We also carried out an analogous comparison of RE using a monthly water balance model (58), allowing us to directly link estimated changes in ET with changes in RE over time within a framework where closure of the regional water balance is constrained by the physics of the model (Fig. 5 ) (). The 20th- and 21st-century records of temperature, precipitation, ET, and RE from this exercise suggest that the increase in temperature observed over the period is likely responsible for the decrease in RE by way of an increase in ET relative to precipitation (Fig. 5 and ). This in turn points to a likely mechanism for explaining increased drought severity across the UMRB in recent decades (Figs. 2 and 3). For additional estimates of UMRB RE based on different climate datasets see . However, results were found to be largely unaffected by the choice of climate dataset or analytical method.

Fig. 5.

UMRB basin-wide temperature, ET, and RE. Distributions of (A) observed runoff-season (March through August) temperature and (B) RE from 1900 to 1983 (blue) and 1984 to 2010 (red), n = 1870. C shows the distributions of observed RE during the years of the Dust Bowl drought (blue) and turn-of-the-century drought (red), n = 408. Lines show the kernel density estimates of the distributions. D and E show the relationship between aggregated, basin-wide modeled ET and RE with color in D denoting values during the two major droughts of record in the UMRB and color in E denoting values from before and after 1984. Curves on the top and right axes show the kernel density estimates of the distributions of the values within each time period being compared. provides statistics for the time period comparisons based on the full, nonaggregated modeling results.

Conclusions

The 1,200-y drought history of the UMRB suggests that, while the turn-of-the-century drought was shorter in duration than numerous earlier events, it may have exceeded the droughts of both the recent and distant past in terms of severity during the driest years of the drought. Similarly, the late 20th and early 21st century in the basin have been characterized by an increasing frequency of coeval hot–dry years that challenge both supply and demand of surface water resources during a period when numerous persistent low streamflow events have resulted in a general drying of the basin relative to the early and mid-20th century.

In consideration of these results, it is important to note that irreducible uncertainties exist in gridded regional climate datasets over complex terrain (). Additionally, there are inherent limitations to empirical, observational assessments and modeling exercises, such as those employed here, that challenge our ability to accurately and mechanistically define the often-interrelated influences of precipitation and temperature on streamflow and surface water resources in the western United States. For example, uncertainty remains on the exact contribution of temperature versus precipitation to the 20th-century streamflow declines in the Upper Colorado River Basin despite significant research effort (e.g., refs. 15, 23, 51, and 59), although a clearer picture of the growing temperature influence emerges with improved estimates of key water and energy balance drivers developed at geographically relevant scales (29). Nonetheless, the challenges posed by severe drought such as the turn-of-the-century drought in the UMRB only highlight the need for better climate and hydrologic monitoring moving forward. Improved monitoring would allow for advances in hydrological and statistical modeling, further reducing uncertainties around the proximal drivers of extreme drought under warming conditions.

Despite these challenges, the analyses of basin-wide observations from the UMRB presented here suggest that the enhanced drying during the turn-of-the-century drought, and persistent periods of low flow since the mid-1980s, are coincident with observations of warming air temperatures and reduced RE in the basin (i.e., the proportion of precipitation contributing to streamflow) (24, 26, 28), of which the UMRB appears particularly vulnerable (60). Snowpack has historically been the primary driver of streamflow in the UMRB (9), with recent temperature-driven snowpack declines and earlier spring melt-out documented across the Northern Rockies and the West essentially mirroring drought severity in the UMRB (31–33). Thus, it appears likely that the unusual severity of the turn-of-the-century drought reflects multiple complex hydroclimatic influences centered on the vulnerability of this snow-driven water supply to the effects of warming temperatures (16, 22).

Modeling efforts suggest that future reductions in RE can be expected to continue with warming (), as increased temperature contributes directly to the observed changes in precipitation phase (snow to rain) (9), reduction of mountain snowpack development (32), reductions in surface albedo (29), and enhancement of warm-season ET (9, 24, 29). The combination of hydrologic changes, such as reduced RE and an increasing likelihood of hot–dry extreme years, represent significant challenges for water management in the UMRB. Recent trends (9) and projected changes (52) suggest a future that may require the capture and storage of increasingly early snowmelt runoff, with increased risk of either severe flooding or increasingly severe drought for the portion of the Missouri basin lacking significant multiyear storage capacity. Improvements in multiyear to decadal forecasting capabilities made by incorporating temperature information in snowmelt dominated basins (61) combined with implementing subbasin drought plans (52) could result in enhanced infrastructure operation and water allocation during increasingly severe future drought events.

Methods

Naturalized Streamflow and Climate Data.

For the purpose of this study, the UMRB is defined as the region ranging east to west from roughly 105°W longitude to the continental divide, and from north to south from the Milk River in Canada to the South Platte River in Colorado (Fig. 1). An initial collection of 31 naturalized streamflow records for key gaging locations across the UMRB were compiled by Martin et al. (34), representing records deemed to reasonably represent natural flow with limited impacts from human activity (62–66). For the analyses in this study, that dataset was then reduced to only those records used in generating the basin-wide composite streamflow reconstruction and contains a total of 17 streamflow records (Fig. 1) (34).

The climate data used were the 4-km × 4-km gridded monthly temperature and precipitation data from the PRISM dataset (45) for water years (October through September) 1900 through 2010. Analyses of 20th- and 21st-century hydroclimate were also carried out using the nCLIMDIV climate dataset (46) for comparison with results based on PRISM. These analyses and results are discussed in and shown in . It should be noted that uncertainties exist in both natural estimates of streamflow and in gridded climate datasets. These are inherent limitations that result from the difficulties associated with quantifying human modification of gaged streamflow as well as the patterns of weather that occur between station-based measurements in both space and time. Such uncertainties are not explicitly quantified here.

Development of the Basin-Wide Runoff-Season Temperature Reconstruction.

We used the North American tree-ring network from the second phase of the PAGES2k project (67) as our initial set of predictors since it was developed specifically for the reconstruction of regional and global temperatures. This dataset excludes tree-ring chronologies with climate relationships dominated by precipitation or moisture sensitivity, ensuring the records used here primarily reflected temperature conditions and do not overlap with the chronologies used in reconstructing streamflow. Additionally, to ensure fidelity to regional temperatures we screened the full North American network for chronologies within 1,000 km of 110W and 46.75N () that had a positive and significant (P < 0.10) relationship with March through August mean temperature. This selection radius was based on the spatial extent of the Upper Missouri watersheds, previous temperature reconstructions (20, 68), and analyses of observational data (69). We also excluded bristlecone pine chronologies that are known to have a complex and topographically mediated relationship to temperature (70, 71). Applying these criteria resulted in a predictor set of 34 tree-ring chronologies, 10 composed of ring-width measurements and 24 of maximum latewood density. This latter measurement is known to be a better proxy for temperature than ring width alone (20, 72).

PRISM (45) monthly temperature data averaged over the primary snowmelt and ET months of March through August served as the predictand for our reconstruction. This temperature record is the average of the temperature records developed for each of the five UMRB subregions (Fig. 1) and spans the period 1900 to 2014. We used a nested composite-plus-scale method (69, 72, 73) to reconstruct regional mean runoff-season temperature from the network of temperature sensitive tree-ring data ().

Composite Climate Data for Each of the Five Naturalized Flow Regions.

To investigate the relationship between climate and streamflow for each region, we identified the major hydrologic unit (HU) level-8 watersheds that made up the primary drainage area for each regional cluster then averaged the PRISM 4-km gridded climate data. Climate data were averaged for each month of each year for each variable across the HU 8 watersheds falling within each cluster. The HU 8 watersheds used to estimate climate in each cluster are identified in .

Spatial Distributions of the Most Severe Droughts.

In order to characterize the geographic distribution of drought during the most severe events in the record, we first calculated decadal flow deficits separately for each of the five clustered subregions. This provided a record of drought severity for each subregion relative to the long-term variation in streamflow for the region itself. We then assessed the level of flow deficits for each subregion over the driest decade within each of the five major droughts evident in the basin-wide drought severity record. (Fig. 2).

Estimating the Relative Forcing of Precipitation and Temperature on Streamflow.

To estimate the temperature forcing of streamflow since 800 (Fig. 2), we followed the approach of Pederson et al. (31). Using linear regression, we regressed the time series of water-year (prior October through September) streamflow z-scores for each of the five subregions against the time series of mean runoff-season (March through August) temperature z-scores for each subregion. We then multiplied the regression coefficients by the time series of temperature z-scores to estimate the relative forcing of temperature on streamflow.

To estimate both the precipitation and temperature forcing of streamflow since 1900 (Fig. 3), we used an analogous multiple regression (MLR) approach. We regressed the streamflow z-scores for each UMRB subbasin against the time series of total water-year precipitation and mean runoff-season temperature z-scores then multiplied the regression coefficients by the time series of temperature and precipitation z-scores to estimate the relative forcing of each variable on streamflow over time. For each year of the common observational period, this quantifies the relative magnitude and direction of the contributions of temperature and precipitation to streamflow volumes. We quantified the magnitude and sign of the combined forcing of both variables as the sum of the relative forcings (Fig. 3, y axis). We also show which variable is more anomalous in its forcing of streamflow each year relative to its long-term average influence (absolute value of the dominant forcing – absolute value of the subordinate forcing) using color in Fig. 3. The relative influence of temperature alone on streamflow over time after accounting for the influence of precipitation is shown in Fig. 3. We used a series of moving windows from 5 to 30 y preceding and following each year in the temperature-forcing record to identify the most likely year that the apparent shift to persistent negative temperature forcing of streamflow occurred based on a two-tailed t test in which the null hypothesis was that the magnitude of the forcing preceding and following a given year in the record were the same. For this analysis, all data were instrumental in origin except for the 1900 to 1929 streamflow values which were reconstructed from tree rings (34).

It is important to note that the average correlation between temperature and precipitation across the five subregions of the UMRB during the 20th and early 21st century is −0.39. This means that one variable could potentially account for up to roughly 16% of the variability in the other if one variable fully controlled the response of the other. In reality, precipitation can lead to changes in temperature (e.g., sensible versus latent heating) (56) or temperature can drive changes in precipitation (74), and this can happen in various ways that are difficult to quantify. This highlights an important limitation of any MLR analysis in which predictor variables are correlated resulting in a degree of uncertainty that will always exist when trying to quantify the possible effect of a single predictor variable on the response. In this particular case, we based our analyses on the determined relationships between either temperature or precipitation and streamflow, while holding the second predictor variable constant. However, because some information about variability in streamflow is shared by variability in both temperature and precipitation, our estimation of those relationships is somewhat less precise than if temperature and precipitation varied completely independently.

Estimating the Probability of Extremes in the Temperature vs. Streamflow Relationship over Time.

We investigated the occurrence of extremes in the relationship between temperature and streamflow by first identifying years in which both temperature and streamflow values were further than 1 s from their respective means (hereafter “extreme” years). This established four possible conditions in which extreme years could occur, dry–hot, dry–cold, wet–hot, and wet–cold, in terms of temperature and streamflow, respectively. We then calculated the percentage of extreme years falling into each category, carrying out the calculation for the 211-y period from 1800 to 2010, the period since 1900, and the period since 1984. The percentages for the period since 1900 are in reference to extreme years defined by the SD of the full 211 y, while the percentages for the period since 1984 are in reference to extreme years defined over the period since 1900.

Estimating the Change in RE over Time.

Following the approach of Woodhouse and Pederson (28), we estimated RE at every gage in the composite record for every year since 1900 (1900 to 2010) as the difference between standardized streamflow and precipitation. Differences in RE between time periods were assessed using two-tailed t tests.

Data Availability.

The Upper Missouri Basin naturalized streamflow records and tree-ring–based naturalized streamflow reconstructions used in this study are available online from the US Geological Survey (USGS) (https://doi.org/10.5066/P9FC7ILX). The tree-ring chronologies used in the streamflow reconstructions are available online from National Oceanic and Atmospheric Administration National Centers of Environmental Information (https://www.ncdc.noaa.gov/paleo/study/26831). The Upper Missouri Basin runoff-season temperature reconstruction is available from the National Oceanic and Atmospheric Administration National Centers of Environmental Information (https://www.ncdc.noaa.gov/paleo/study/29413). The tree-ring chronologies used in the temperature reconstruction are available online from the PAGES 2K version 2 consortium (https://doi.org/10.1038/sdata.2017.88).