Supporting InformationMarlon et al. 10.1073/pnas.1112839109SI TextData Analysis Methods. Historical records. Historical data are fromReynolds and Pierson (Fig. S1, 1). The whole-U.S. estimates sug-gest that about 60% of the original sawtimber stand at 1700 wasdestroyed by either natural or human causes and 40% was used.Of the 40% used, about half was for fuelwood. The trends in useand destruction, which are for the United States as a whole,indicate that farm clearing accelerated earliest, but occurredprimarily in the East (2). Extraction of wood for lumber and fuel-wood was more important in the West, and both acceleratedrapidly in the 1800s. Fuelwood use peaked around 1875 CE,where lumber extraction continued to rise rapidly until the early1900s, after which it declined rapidly.

Estimates of billions of board feet (bf) destroyed by fire wereprovided for the United States as a whole. In order to obtain anestimate of fire destruction solely for the western United States,we scaled the national data by a factor of 0.14 based on the pro-portion of the original stand estimates assigned to the westernsubregions, which were comprised of the North and South Pacificand Northern and Southern RockyMountains. A simple rescalingof the burned timber based on the proportion of bf in western vs.eastern regions does not take important factors such as areaburned into account (e.g., in terms of stand acreage, 18% wasassigned to the west vs. 82% to the east), however our focus hereon temporal trends in burning should not be strongly affected byscaling factors (Fig. S1).

We also digitized stand-origin age data from Hicke et al. (3) tovalidate the Reynolds and Pierson historical data. These datawere derived from forest-inventory databases for the westernUnited States and include information on the age of the inven-toried forest stands. These stand-origin ages illustrate the pulse ofregeneration accompanying the forest loss reported by Reynoldsand Pierson (Fig. S1, green curve).

Compositing, smoothing and bootstrapping the tree-ring fire-scar,charcoal influx and charcoal-peak/fire-episode data.Because our ob-jective is to focus on region-wide, multi-decadal-to-century timescale variations in fire, we adopted a data-analytical approachthat involved the construction of smooth composite curves basedon the individual datasets described below. The construction ofcomposite curves allows the common underlying variation amongindividual records to emerge in the composite. If there were nounderlying common variation, then the variation in the individualrecords would “wash out” in the composite, leaving no clearlong-term trend. Conversely, if there is some common patternof temporal variation in the data, the individual records will re-inforce one another, and that pattern will emerge in the compo-site. Because we are focusing on decadal and longer time-scalevariations, temporally smoothing the data is also appropriate.In the approach we implement here, we adopted the locallyweighted regression approach, as described by Cleveland (32)and Loader (7), using an appropriate model-fitting procedurefor each dataset. This approach simultaneously composites thedata from individual records and smooths them.

Because the data points are irregularly distributed around theregion, an obvious question can be asked about the potential forindividual records or groups of records to disproportionatelydominate the composite. To examine the impact of individualrecords (in the case of the fire-scar, charcoal-influx, and char-coal-peak/fire-episode data), or individual grid points (in the caseof the temperature and DAI data) we adopted a bootstrapping-by-site (as opposed to bootstrapping-by-samples) with replace-

ment approach. The width of the bootstrap 0.05 and 0.95 confi-dence intervals constructed using 1,000 replications of samplingwith replacement, provides a visual assessment of the impact ofindividual records on the composite curve. Marlon et al. (31) dis-cuss further the general smoothing and bootstrapping approachesused here, and compare the local regression approach with alter-natives such as smoothing splines.

To permit comparisons among the resulting composite curves,we obtained fitted values at even 10-year “target points.” The se-lection of this particular interval is not intended as a statement ofour belief in the underlying temporal resolution of the individualrecords, but instead as a practical step in the comparison of theindividual composite curves.

Tree-ring fire-scar data. Fire-scar data are from the InternationalMultiproxy Paleofire Database (URL: http://www.ncdc.noaa.gov/paleo/impd/paleofire.html). The dataset includes 1.37 million re-cording tree rings, which contain 52,667 individual scars, of which50,456 (96%) were identified in particular as fire scars. Thesedata are shown plotted in latitudinal order in Fig. S2, where eachrow represents a site, with a tick mark plotted each year, in red ifany trees were noted as having been scarred in a particular year,and in gray otherwise. Because the number of sites recording fireactivity varies over time, the proportion of sites recording fire dur-ing each year (Fig. S3) was calculated to serve as a proxy of changesin fire frequency through time for the entire region. The plots showthat the fire-scar data becomes sparser going back in time, parti-cularly prior to 1500 CE. The scatter plot of proportion of record-ing sites with scars by year shows distinct patterns prior to this time,related to the small number of sites in the dataset.

Because the probability of a tree recording a fire greatly in-creases after its initial fire-scar event (becoming a “recordertree”), an increase in time alone may lead to an increasing num-ber of recording trees, and that in turn increases the likelihoodthat any given fire will be recorded (4). Analysis of samplingmethods and fire-scar accumulation curves by dendrochronolo-gists, however, shows that there is a threshold of trees (e.g.,n ¼ 50) beyond which results are robust with respect to addi-tional sampling (5). Fires in the IMDP dataset were identifiedby individual scars at each site, but in the absence of additionaldetailed analyses of the sampling methods at each site, we rely onevidence that the accumulation of fire scars from spatial networksof tree-ring records provides a reasonable proxy for region-widefire history [e.g., (6)].

The proportion of sites recording fires was summarized using alocal regression (logit) model [locfit, (7)]. We used the binomiallocal likelihood family (appropriate for data that are proportions),the tricube weight function, and nearest-neighbor fractions of 0.1and 0.25. To determine the range of years over which the local re-gression provides an adequate summary of the data, we performedthe standard diagnostic analyses of residuals, which confirmed that1500 CE was an appropriate cutoff for plotting and examining thesummary.

In order to ensure that our summary of the tree-ring data isrobust, we compared local regression models of the IMPD data-set using both one or more and two or more scars as criteria. Wedid this for the region as a whole (Fig. S4, top box) and for thenorth (Fig. S4 middle box) and the south (Fig. S4 bottom box).Using two or more scars slightly reduces the overall proportion ofsites recording a fire, but all three summaries yield similarpatterns and demonstrate that the trends (including a dip in fireactivity ca. 1815) are robust. We also compared analyses based on

Marlon et al. www.pnas.org/cgi/doi/10.1073/pnas.1112839109 1 of 9

all injuries vs. only those injuries identified as fire scars, and theresults were nearly indistinguishable.

Examination of individual fire-scar and charcoal records re-veals the spatial and temporal heterogeneity in the data as wellas the key differences between the two types of records (Fig. 1).For example, the fire-scar records highlight increased fire fre-quencies at southern vs. northern sites (Fig. 1C), whereas regularsampling of the continuously varying charcoal accumulationshighlights the dynamic nature of fire on centennial to millennialscales (Fig. 2C). The fire-scar data indicate how synchrony in fireregimes can occur across wide areas in specific years (Fig. S2),but even then not all sites will burn. Similarly, charcoal recordsindicate that peaks and troughs in biomass burning at the regionalscale (Fig. 2D) are not supported by similar trends at all sites(Fig. 1B), but rather indicate intervals when many sites showsimilar patterns.

Our analysis of fire scars (Fig. S2) did not address synchroneityamong episodes of burning, which may be caused by climate pat-terns that promote multiple large fires. Patterns of synchrony donot necessarily correspond with overall fire activity. For example,some studies have shown that fires were larger and regionally moresynchronous apparently due to fuel build-up during cold decades,when fire frequency is reduced (8, 9). Other studies, however, haveindicated that increased synchrony occurs under increased drought(10). Regardless of the occurrence of regionally synchronous fireyears underlying the fire records, the trends in overall rates ofburning from both fire-scar and charcoal data show decadal-to-centennial trends that are consistent with climate reconstructions.

Charcoal influx data. Each of the 69 charcoal records was trans-formed and standardized using a protocol detailed in Power etal. (11) before compositing into a regional charcoal-based indexof biomass burning. The transformation, standardization, andsummarization methods are designed to highlight the long-termtrends in the charcoal data that are characteristic of multiple re-cords. Data were smoothed using local regression using the tri-cube weight function with 25- and 100-year window (half) widths.Confidence intervals for each composite curve were generated bya bootstrap resampling (with replacement) of individual sites (notsamples) over 1,000 replications. Bootstrap confidence intervalsfor each target point were taken as the 5th and 95th percentiles ofthe 1,000 fitted values for each target point. This bootstrappingstrategy reveals the uncertainty in the smoothed (fitted) valuesrelated to the inclusion or not of individual recording sites,and provides a way of gauging the robustness of the fitted curveto the particular distribution of sites in the dataset.

Charcoal decomposition.Charcoal records with data for at least thepast 200 y (i.e., with intact core tops) and with at least 17 samplesper 1,000 y were selected for further processing to identify indi-vidual fire episodes (i.e., one fire event or a series of individualfire events that occurred over several years). Forty-one recordsmet these criteria (Table S1). The selected records were analyzedwith CharAnalysis (12). All records were interpolated to theirmedian sample resolution and were not transformed prior to de-composition. Low-frequency (background) charcoal accumula-tion rates (CHAR) were estimated using a lowess smootherrobust to outliers. The background smoothing window-widthswere determined by testing alternative window widths and com-paring their sensitivity analysis results produced by CharAnalysis.Smoothing window widths that produced a high signal-to-noiseindex and a high goodness-of-fit index for the noise distributionwere selected individually for each record. High-frequencyCHAR was calculated using the ratio of interpolated to back-ground CHAR. Threshold values for peak identification werelocally defined and based on a percentile cutoff of a noise distri-bution determined by a Gaussian mixture model. The selection of

final peaks were based on those in the 99th percentile of the noisedistribution.

Charcoal peak/fire episode frequency. The time series of charcoalpeaks were summarized using a kernel-density estimator (13),with bandwidths (25 and 50 y) selected to be comparable withthe window widths used for summarizing the other data. We usedthe same bootstrapping strategy as for the charcoal influx data.Units of the peak-density curve are number of peaks per year perrecord.

Differences between fire-scar and charcoal data. Fire-scar and char-coal records are complementary across forest types and fire re-gimes, but also with regard to temporal coverage (4). Fire-scarrecords generally span centuries and are annually resolved, butthe records “fade” before 1500 CE, whereas charcoal records spanmillennia and integrate information about fire occurrence overyears, decades, or centuries, depending on the rate of sedimenta-tion in the record. In addition, charcoal records contain noise re-lating to taphonomic and site-specific processes that can makeidentification of biomass burning trends at a single site problematic.

Fire-scar and charcoal records differ in several important re-spects, however, including their distribution across forest types,registration of fire events, temporal length, and temporal resolu-tion (Fig. 1 B and C). Fire-scar records are most prevalent inareas with low effective moisture, such as in dry ponderosa pineforests, because the typically low fuel loads promote low intensityburns that frequently scar trees [e.g., (14–16)] but produce lowamounts of charcoal that are difficult to detect in the lake-sedi-ment record. In contrast, charcoal records from forests with higheffective moisture, such as mixed conifer or subalpine fir forests,are often better resolved because such forests are typically char-acterized by complex vegetation structures and abundant, contin-uous fuels that support infrequent, high-severity fires. Amplecharcoal is produced during such forest fires, which is incorpo-rated into sediments during and after the fire (17–20).

Lake-sediment records of charcoal are restricted to regionswith lakes, and in particular those with continuous deposition ofa sufficient rate to ideally provide decadal sampling resolutions.Because local changes that affect vegetation, lake characteristics,erosion patterns, or topography can alter trends in charcoal accu-mulation, absolute trends in background charcoal at a single lakedo not necessarily reflect changes in the local fire regime. By ana-lyzing trends at multiple sites, however, shared characteristics ofmacroscopic charcoal accumulation rates (CHAR, or charcoal in-flux) can be used to infer regional-scale changes in fire activity (21,22). In western forests, where there has been little apparent changein vegetation composition over the last 2,000 y, there is increasingevidence that trends in charcoal influx from nearby sites are cor-related with area burned over time (23, 24). When sedimentationrates are high enough (ideally>1 mm∕y) and sediment records aresampled continuously, discrete pulses of charcoal can be distin-guished from background charcoal variations and used to inferchanges in the frequency of fires (12, 25).

Multiple fire-scar and charcoal records thus jointly provideinformation about widespread variations in fire frequency andbiomass burning (10, 26). Although charcoal records are less tem-porally and spatially precise than fire-scar records, charcoal datacan provide a continuous record of burning over millennia. Thelong-term context provided by such data is especially relevant inlight of the magnitude of projected climate changes, which exceedthose experienced in recent millennia.

Climate data and treatment. Temperature data were obtained fromMann et al. (27) (Fig. 2E). The temperature time series were ex-pressed as anomalies from a 1960–1990 CE long-term mean. Thedrought reconstruction (Fig. 2F) uses updated data from Cook et

Marlon et al. www.pnas.org/cgi/doi/10.1073/pnas.1112839109 2 of 9

al. (28) to calculate a region-wide drought-area index (i.e., theproportion of grid cells with Palmer Drought-Severity Index[PDSI] values less than −1.0). Both series were smoothed usingthe same approach taken with the charcoal data (i.e., lowesssmoothing based on a 100-y window width), and bootstrap con-fidence intervals were also calculated in a similar fashion, by re-sampling the individual grid-cell time series, as opposed toindividual data values. As is the case for the charcoal data, thisapproach allows the influence of individual grid points or groupsof them on the composite curve to be assessed.

Relationship between biomass burning (charcoal influx) and climate.We used a generalized additive model [Generalized AdditiveModel (GAM), (29, 30)] to describe the dependence of biomassburning on climate, as represented by the summary curves forcharcoal influx, mean annual temperature anomalies [(MAT),(27)] and drought-area index (DAI, 28). We selected this parti-cular modeling framework because it permits the developmentand display of a smooth function that links a particular response(here, charcoal influx) to a small number of predictor variables(e.g., MATand DAI). In fitting GAM “surfaces,” there is a trade-off between the smoothness of the surface and its fit to individualdata points, but this can be controlled by constraining thesmoothness of the fitted surface. The summary curves were eval-uated at 10-year intervals, and the time span of data used to cali-brate the model was 500–1800 CE, thereby avoiding theconfounding effects of settlement and fire suppression.

We used the R package mgcv (29) to fit surfaces to the charcoaland climate data using the tensor-product smoother with thin-plateregression splines because the climate data have different units ofmeasurement, and we suspected that the resulting surfaces wouldbe anisotropic (see ref. 29 ch. 5 for a discussion of the practicalconsiderations of GAM fitting). MATand DAI are also correlated(a situation known as “concurvity” in the GAM literature, r ¼ 0.62for data from 500 to 1800 CE). Although concurvity did not createany noticeable numerical problems here, it did lead to the possi-bility of identifying multiple models with roughly the same good-ness-of-fit, and so we examined a variety of models that includedone or the other of the climate variables, as well as interactionsbetween the two. To understand the trade-off between good-ness-of-fit and the complexity of the model, we examined the shape

of the fitted surface and summary statistics such as the adjusted R2

and Akaike Information Criterion values for individual models, aswell as F-statistics for analysis-of-variance comparisons of a parti-cular model with a “null” model consisting of only the mean valueof the charcoal data.

The fitted response surface (Fig. S5) shows a general increasein predicted z-scores of transformed charcoal influx with increas-ingMATand DAI, a relationship that is also evident in time seriesplots of the data (Fig. 2). The model (with equivalent degrees offreedom of 4, 2, and 9 for the MAT, DAI, and interaction terms,respectively) fits the data well (R2 ¼ 0.85; F ¼ 43.0; p < 0.000),and the residuals to not have any striking features. The surface isrelatively smooth, with a slight concavity centered around valuesof MAT ¼ 0.5 and DAI ¼ 0.35, values that mainly occur between600 and 800 CE. This feature could be removed by increasing thesmoothness of the interaction term, but we elected to keep it in toincrease the generality of the fitted surface.

The composite curves used in fitting the GAMs here are quitesmooth [i.e., the charcoal, temperature, and drought area serieswere smoothed using local regression using the tricube weightfunction with a 100-year window (half) width], which raises thepossibility of observing spurious correlation among them, or ob-taining inflated estimates of the goodness-of-fit. To examine theimpact of pseudoreplication, we refit the model to the compo-site-curve values sampled at 40-year intervals, the practical upperlimit of sample time step for fitting the model (Fig. S6). The modelfits the subset data about as well as the 10-year time step data(R2 ¼ 0.81; F ¼ 10.22; p < 0.000), and captures all of the impor-tant features in the predicted charcoal influx series, including thedivergence between predicted and observed values after thelate 1800s.

A key feature of predictions by the model after the calibrationinterval end (1800 CE) is the divergence between the decreasingcharcoal influx values and the increasing predicted values, begin-ning in the late 1800s (Fig. 2C; Fig. S7 bottom). To visually assessthe extent to which these diverging predictions represented “out-of-sample” extrapolations, we examined the scatter plot of MATand DAI values (Fig. S7, top). This plot suggests that only the1990 and 2000 CE values of MAT lie substantially outside ofthe main cloud of points, and that the predicted values for thesetimes represent at most mild extrapolations.

1. Reynolds RV, Pierson AH (1941) The saw timber resource of the United States,1630-1930. Forest Survey Release 53 (USDA Forest Service, Washington, DC).

2. Ramankutty N, Foley JA (1999) Estimating historical changes in land cover: NorthAmerican croplands from 1850 to 1992. Global Ecol Biogeogr 8:381–396.

3. Hicke JA, Jenkins JC, Ojima DS, Ducey M (2007) Spatial patterns of forest character-istics in the western United States derived from inventories. Ecol Appl 17:2387–2402.

4. Swetnam TW, et al. (2009) Multimillennial fire history of the giant forest, SequoiaNational Park, U.S.A. Fire Ecology 5:120–148.

5. Van Horne ML, Fulé PZ (2006) Comparing methods of reconstructing fire historyusing fire scars in a southwestern United States ponderosa pine forest. CanadianJournal for Forest Resources 36:855–867.

6. Farris CA, Baisan CH, Falk DA, Yool SR, Swetnam TW (2010) Spatial and temporalcorroboration of a fire-scar-based fire history in a frequently burned ponderosa pineforest. Ecol Appl 20:1598–1614.

7. Loader C (1999) Local Regression and Likelihood (Springer, New York).8. Swetnam TW, Baisan CH (2003) Tree-ring reconstructions of fire and climate history in

the Sierra Nevada and southwestern United States. Fire and climate in temperateecosystems of the western Americas, eds Veblen TT, Baker WL, Montenegro G, Swet-nam TW (Springer-Verlag, NY), pp 158–195.

9. Trouet V, Taylor AH,Wahl ER, Skinner CN, Stephens SL (2010) Fire-climate interactionsin the American West since 1400 CE. Geophys Res Lett 37:L04702.

10. Kitzberger T, Brown PM, Heyerdahl EK, Swetnam TW, Veblen TT (2007) ContingentPacific-Atlantic ocean influence on multicentury wildfire synchrony over westernNorth America. Proc Nat'l Acad Sci USA 104:543–548.

11. Power MJ, Marlon JR, Bartlein PJ, Harrison SP (2010) Fire history and the Global Char-coal Database: a new tool for hypothesis testing and data exploration. PalaeogeogrPalaeocl 291:52–59.

12. Higuera P, Gavin D, Bartlein P, Hallett D (2010) Peak detection in sediment-charcoalrecords: impacts of alternative data analysis methods on fire-history interpretations.Int J Wildland Fire 19:996–1014.

13. Venables WN, Ripley BD (2002)Modern applied statistics with S (Springer Verlag, NY)pp 495.

14. Grissino-Mayer HD, Swetnam TW (2000) Century-scale climate forcing of fire regimesin the American Southwest. Holocene 10:213–220.

15. Heyerdahl EK, Brubaker LB, Agee JK (2001) Factors controlling spatial variation inhistorical fire regimes: a multiscale example from the interior West, USA. Ecology82:660–678.

16. Morgan P, Heyerdahl EK, Gibson CE (2008) Multiseason climate synchronized forestfires throughout the 20th-century, Northern Rockies, USA. Ecology 89:717–728.

17. Brunelle A, Whitlock C, Bartlein P, Kipfmueller K (2005) Holocene fire and vegetationalong environmental gradients in the Northern RockyMountains.Quaternary Sci Rev24:2281–2300.

18. Long CJ, Whitlock C, Bartlein PJ (2007) Holocene vegetation and fire history of theCoast Range, western Oregon, USA. Holocene 17:917–926.

19. Gavin DG, et al. (2007) Forest fire and climate change in western North America: in-sights from sediment charcoal records. Frontiers in Ecology and the Environment5:499–506.

20. Whitlock C, Skinner CN, Bartlein PJ, Minckley T, Mohr JA (2004) Comparison of char-coal and tree-ring records of recent fires in the eastern Klamath Mountains, Califor-nia, USA. Canadian Journal of Forest Research 34:2110–2121.

21. Carcaillet C, et al. (2002) Holocene biomass burning and global dynamics of the car-bon cycle. Chemosphere 49:845–863.

22. Power MJ, et al. (2008) Changes in fire regimes since the Last Glacial Maximum: anassessment based on a global synthesis and analysis of charcoal data. Climate Dy-namics 30:887–907.

23. Higuera PE, Peters ME, Brubaker LB, Gavin DG (2007) Understanding the origin andanalysis of sediment-charcoal records with a simulation model. Quaternary Sci Rev26:1790–1809.

24. Higuera PE, Whitlock C, Gage JA (2010) Linking tree-ring and sediment-charcoal re-cords to reconstruct fire occurrence and area burned in subalpine forests of Yellow-stone National Park, USA. Holocene 21:327–341.

25. Gavin DG, Hu FS, Lertzman K, Corbett P (2006) Weak climatic control of stand-scalefire history during the late Holocene. Ecology 87:1722–1732.

Marlon et al. www.pnas.org/cgi/doi/10.1073/pnas.1112839109 3 of 9

26. Marlon J, et al. (2009) Wildfire responses to abrupt climate change in North America.Proc Nat'l Acad Sci USA 106:2519–2524.

27. Mann ME, et al. (2009) Global signatures and dynamical origins of the Little Ice Ageand Medieval Climate Anomaly. Science 326:1256–1260.

28. Cook ER, Woodhouse CA, Eakin CM, Meko DM, Stahle DW (2004) Long-term ariditychanges in the western United States. Science 306:1015–1018.

29. Wood SN (2006)Generalized AdditiveModels: An Introduction with R (Chapman andHall/CRC Press, FL).

30. Hastie TJ, Tibshirani R (1990) Generalized Additive Models (Chapman and Hall, FL).

31. Marlon J, et al. (2008) Climate and human influences on global biomass burning over

the past two millennia. Nature Geosci 1:697–701.

32. Cleveland WS, Devlin SJ (1988) Locally weighted regression: an approach to regres-

sion analysis by local fitting. Journal of the American Statistical Association

83:596–610.

1,600 1,650 1,700 1,750 1,800 1,850 1,900 1,950 2,000Year (A.D.)

0

100

200

300

400

500

600

U.S

.saw

timbe

rlo

ss(b

illio

nsbo

ard

feet

) Natural Fires

Human Fires

Woods Clearing

Insects & Disease

Woods Waste

Total Destruction

Stand Origins

0

5

10

15A

reaofS

tandO

rigins(M

ha)

Source: Reynolds, R.V., and A.H. Pierson (1941)The saw timber resource of the United States, 1630–1930.Forest Survey Release 53. USDA For. Serv., Washington, DC.

Hicke, J.A., J.C. Jenkins, D.S. Ojima and M. Ducey (2007)Spatial patterns of forest characteristics in the western UnitedStates derived from forest inventories. Ecological Applications17:2387-2402

Historical Variations in Forest Disturbance

and Western U.S. Stand Origins

Fig. S1. Historical “drain” on the saw-timber resources of the United States 1630–1930 CE (1). Natural (orange circles) and human-caused fires (red circles) aredistinguished from lumber (diamonds), fuelwood (triangles), other disturbances such as insect infestations and wind throw (vertical bars), and from farmclearing (open squares). Stand-origin ages from Hicke et al. (3) are plotted in green, and show the reforestation following the widespread disturbanceof the 1800s.

Marlon et al. www.pnas.org/cgi/doi/10.1073/pnas.1112839109 4 of 9

no scarscar

2,0001,8001,6001,4001,2001,000Year

0

50

100

150

200

250

300

350

(Sou

th)

Site

s in

Lat

itudi

nal O

rder

(N

orth

)

Western United StatesIMPD Fire Scars

36 N @ Site 100, 42 N @ Site 240

Fig. S2. Fire-scar records from 359 sites spanning the past 1,000 y and arranged in latitudinal order from north (top) to south (bottom). (Site 100 is at 36 N andsite 240 is at 42 N) Each site is depicted by a horizontal gray bar; red tick marks represent individual fire scars. Scattered gray tick marks to the left of each barmark the beginnings of individual tree-ring records. Years when fire was widespread are visible as vertical red streaks in the data. The exclusion and eventualsuppression of fire is evidenced after ca. 1900 CE by a sharp decline in fire scars. A greater number of fire scars in the bottom half of the graph indicatesincreased fire frequencies among southern sites. Few fire scars at the middle sites reflect both fewer years of widespread fire compared to the southwest andalso the fact that in a mixed fire regime stand-replacing fires destroy some of the evidence of previous fires.

0

100

200

300

Num

ber

ofS

ites

800 600 400 200 0yrs BP

1,000 1,200 1,400 1,600 1,800 2,000Year (AD)

0

0.05

0.1

0.15

0.2

0.25

0.3

0.4

0.5

Pro

port

ion

ofR

ecor

ding

Site

sw

ithS

cars

Number of Sites

Number of Siteswith Fire Scars

IMPD Fire-Scar Data -- Western U.S.

n=359

(All Sites West of 104oW)

[Local Logit Regression]

Fig. S3. Fire history from the IMPD fire-scar dataset since 1200 CE showing the proportion of sites recording a fire scar during the past 1,000 y (gray dots, topbox), the number of sites each year with recording trees (gray dots, bottom box), and the number of sites each year that have fire scars (red dots). The trend inthe proportion of sites recording a fire scar is indicated by a thick red line created by a locally fitted binomial (logit) model. Thin red lines are 95% confidencelimits.

Marlon et al. www.pnas.org/cgi/doi/10.1073/pnas.1112839109 5 of 9

800 600 400 200 0yrs BP

1,000 1,200 1,400 1,600 1,800 2,000Year (CE)

0

0.05

0.1

0.15

0.2

Pro

port

ion

ofR

ecor

ding

Site

sw

ithS

cars

0

0.05

0.1

0.15

0.2

0

0.05

0.1

0.15

0.2

IMPD Fire-Scar Data -- Western U.S.

n=359

All Sites West of 104oW, Fire Scars Only

[Local Logit Regression]

Sites West of 104oW &North of 42oNFire Scars Only

Sites West of 104oW &South of 42oNFire Scars Only

n=240

n=119

> 1 Scar

> 1 Scar

> 1 Scar

> 2 Scars

> 2 Scars

> 2 Scars

BW = 0.10BW = 0.25Bootstrap C.I.

Fig. S4. Comparison of fire history reconstructions from the IMPD fire-scar dataset since 1200 CE using ≥1 (top set of curves in each box) and ≥2 (bottom set ofcurves in each box) scars to fit a local regression model using a bandwidth of 10 (black line) and 25 (thick red line) years for (i) the western United States as awhole; (ii) only sites north of 42 °N; and (iii) only sites south of 42 °N. Thin red lines are 95% confidence limits.

Marlon et al. www.pnas.org/cgi/doi/10.1073/pnas.1112839109 6 of 9

Fig. S5. GAM-fitted response surface showing the relationship between the summary curve of standardized and transformed charcoal influx values and thesummary curves of MAT and DAI. Gray dots show the individual data points for the calibration interval (500–1800 CE).

600 800 1,000 1,200 1,400 1,600 1,800 2,000Year A.D.

-0.5

-0.25

0

0.25

0.5

Z-S

core

sof

Tra

nsfo

rmed

Cha

rcoa

lInf

lux

ObservedPredicted (500-1,800 AD calibration interval)

Western U.S. Biomass Burning

Fig. S6. Observed (red) and predicted (dark purple) values of the standardized and transformed charcoal influx. The corresponding values for a subsampled(40-y) dataset are plotted in pink (observed) and light purple (predicted), respectively, offset by −0.10 Z-score units for clarity.

Marlon et al. www.pnas.org/cgi/doi/10.1073/pnas.1112839109 7 of 9

-0.4 -0.2 0 0.2 0.4

Temperature Anomalies (C)

0.3

0.32

0.34

0.36

0.38

0.4

Dro

ught

-Are

aIn

dex

500

520 540

560

580600620

640660680

700720740760

780800820840860880900

920

940

960

980

1000

1020

1040

1060

108011001120

1140

1160

11801200

12201240

1260

1280

1300

1320

13401360

1380

1400

142014401460

1480

1500

1520

1540

15601580

1600162016401660

168017001720

1740

176017801800

18201840

186018801900

19201940

1960

1980

2000

-0.4

-0.2

0

0.2

0.4

Obs

erve

dZ

-Sco

res

ofT

rans

form

edIn

flux

-0.4 -0.2 0 0.2 0.4

Predicted Z-Scores of Transformed Influx

500 520540 560580600620

640660680

700720740760

780800820840860880900

920940

960

980

1000102010401060 108011001120

114011601180

1200

1220

1240

126012801300

13201340

13601380

1400142014401460

1480

1500

1520

1540

1560

1580160016201640

1660

1680

1700

17201740

1760

1780

180018201840

1860

1880

1900

1920

1940

1960 19802000

Fig. S7. Scatter plot of MAT and DAI, labeled by observation year (top). The plot shows that only the values for 1990 and 2000 (and perhaps 1890) CE lieoutside of the main clound of points, and would generate mild extrapolations. Scatter plot of observed and predicted standardized and transformed charcoalinflux (bottom), showing the divergence of observed and predicted (by climate) values in the 20th century.

Table S1. Site names and locations from GCD v1 and additional records fromcoauthors that are included in the charcoal analyses

Site ID Site name Peaks Latitude Longitude Elevation (m)

1 Abalone Rocks Marsh no 33.9564 −119.9767 12 Alamo Bog yes 35.9111 −106.5838 2,6303 Baker Lake yes 45.8919 −114.2619 2,3004 Barrett Lake yes 37.5957 −119.0066 2,8165 Battle Ground Lake yes 45.8000 −122.4917 1546 Beaver Lake yes 44.9175 −123.3050 697 Beef Pasture no 37.4167 −108.1500 3,0608 Bluff Lake yes 41.3467 −122.5575 1,9219 Bolan Lake yes 42.0200 −123.4600 1,63710 Burnt Knob yes 45.7044 −114.9867 2,25011 Cambell Lake yes 41.5333 −123.1050 1,75012 Cedar Lake no 41.2075 −122.4954 1,74013 Chihuahueños Bog yes 36.0472 −106.5083 2,92514 Coast Trail Pond yes 37.9853 −122.8000 23015 Como Lake no 37.5500 −105.5000 3,523

Marlon et al. www.pnas.org/cgi/doi/10.1073/pnas.1112839109 8 of 9

Site ID Site name Peaks Latitude Longitude Elevation (m)

16 Crane Lake no 36.7167 −112.2167 2,59017 Crater Lake yes 41.3836 −122.5778 2,28818 Crevice Lake yes 45.0000 −110.5780 1,71319 Cygnet Lake yes 44.6628 −110.6161 2,53020 Dead Horse Lake yes 42.5606 −120.7781 2,24821 DeHerrera Lake yes 37.7417 −107.7083 3,34322 East Lake yes 37.1779 −119.0277 2,86323 East Sooke Fen no 48.3519 −123.6817 15524 Five Lakes no 48.0819 −118.9294 78025 Forest Pond 1 no 43.4717 −109.9389 2,79726 Forest Pond 2 no 43.4389 −109.9500 2,76627 Foy Lake yes 48.1658 −114.3592 1,00628 Frying Pan no 38.6167 −111.6667 2,72029 Glenmire Lake yes 37.9933 −122.7770 39930 Head Lake no 37.7000 −105.5000 2,30031 Hidden Lake yes 40.5100 −106.6100 2,71032 Hoodoo Lake yes 46.3206 −114.6503 1,77033 Hunter’s Lake no 37.6083 −106.8444 3,51634 Jicarita Bog yes 36.0708 −105.5889 3,20735 Lake Oswego yes 45.4111 −122.6661 3036 Lily Lake yes 41.9758 −120.2097 2,04237 Little Lake yes 44.1680 −123.5835 70338 Little Molas Lake yes 37.7417 −107.7083 3,37039 Lost Lake yes 45.8242 −123.5792 44940 Lower Gaylor Lake yes 37.9086 −119.2863 3,06241 Martins Lake no 47.7138 −123.5400 1,41542 Montezuma Well no 34.0000 −112.0000 1,12543 Moose Lake no 47.8830 −123.3500 1,50844 Mumbo Lake yes 41.1913 −122.5092 1,86045 Park Pond 1 no 43.4681 −109.9594 2,70546 Park Pond 2 no 43.4500 −109.9408 2,71447 Park Pond 3 no 43.4592 −109.9200 2,73948 Patterson Lake yes 41.3867 −120.2236 2,74349 Pintlar Lake yes 45.8406 −113.4403 1,92150 Pixie Lake no 48.5964 −124.1967 7051 Porter Lake no 44.4478 −123.2428 7352 Posy Lake no 37.9500 −111.7000 2,65353 Reservoir Lake no 45.1300 −113.4600 2,16154 San Joaquin Marsh no 33.6583 −117.8583 255 Sanger Lake yes 41.9021 −123.6465 1,54756 Siesta Lake no 37.8500 −119.6667 2,43057 Slough Creek Pond yes 44.9183 −110.3467 1,88458 Swamp Lake no 37.9500 −119.8167 1,55459 Taylor Lake CA yes 41.3614 −122.9675 1,97960 Taylor Lake OR yes 46.1006 −123.9067 661 Three Creeks yes 44.0990 −121.6273 1,99662 Todd Lake yes 44.0277 −121.6846 1,87563 Trail Lake yes 44.2800 −110.1700 2,36264 Tumalo Lake yes 44.0218 −121.5438 1,53665 Upper Rocket Creek Bog no 47.0405 −115.8888 1,71066 Upper Squaw lake yes 42.0333 −123.0150 93067 Warner Lake no 44.2461 −122.9575 59068 Whyac Lake no 48.6722 −124.8444 1569 Wildcat Lake yes 37.9684 −122.7851 67

Marlon et al. www.pnas.org/cgi/doi/10.1073/pnas.1112839109 9 of 9