Warm springs alter timing but not total growth of temperate deciduous trees

As the climate changes, warmer spring temperatures are causing earlier leaf-out1–3 and commencement of CO2 uptake1,3 in temperate deciduous forests, resulting in a tendency towards increased growing season length3 and annual CO2 uptake1,3–7. However, less is known about how spring temperatures affect tree stem growth8,9, which sequesters carbon in wood that has a long residence time in the ecosystem10,11. Here we show that warmer spring temperatures shifted stem diameter growth of deciduous trees earlier but had no consistent effect on peak growing season length, maximum growth rates, or annual growth, using dendrometer band measurements from 440 trees across two forests. The latter finding was confirmed on the centennial scale by 207 tree-ring chronologies from 108 forests across eastern North America, where annual ring width was far more sensitive to temperatures during the peak growing season than in the spring. These findings imply that any extra CO2 uptake in years with warmer spring temperatures4,5 does not significantly contribute to increased sequestration in long-lived woody stem biomass. Rather, contradicting projections from global carbon cycle models1,12, our empirical results imply that warming spring temperatures are unlikely to increase woody productivity enough to strengthen the long-term CO2 sink of temperate deciduous forests. Warmer spring temperatures affect the timing of stem diameter growth of temperate deciduous trees but have little effect on annual growth.

In recent decades, tree growth in Earth's forests has more than offset losses from deforestation and other disturbances, such that a net forest CO 2 sink of approximately 1.6 Gt carbon per year offsets approximately 20% of anthropogenic emissions 13 , dramatically slowing the pace of atmospheric CO 2 accumulation and climate change. Of this important carbon sink, approximately 47% occurs in temperate forests 13 , with temperate deciduous forests sequestering more than 0.3 Gt carbon per year 14 . The future behaviour of this carbon sink will have an important yet uncertain role in influencing atmospheric CO 2 and climate change [15][16][17] .
In temperate deciduous forests, spring warming generally lengthens the period over which trees have photosynthetically active leaves 1,3,18 and that over which the ecosystem is a net CO 2 sink 1, 18 . On the basis of these observations, current terrestrial ecosystem models represent warm spring temperatures and longer growing seasons as contributing positively to annual gross primary productivity (GPP) and net CO 2 uptake (that is, net ecosystem exchange (NEE)) 2,12,15 . However, the long-term persistence of this extra carbon in the ecosystem, and the associated negative feedback to climate change 1 , will depend on the extent to which it is allocated to woody growth and therefore resides in the ecosystem for decades to centuries 10,11 , as opposed to being rapidly released back to the atmosphere through respiration 19,20 . Model representation of carbon allocation to stem growth-or woody aboveground net primary productivity (ANPP woody ) on the ecosystem level-does not capture known decoupling of stem growth from photosynthate production 9,17,21 . As a result, the consequences of rising spring temperatures on stem growth may not be accurately represented in models 9,17 . Understanding the sensitivity of ANPP woody to spring temperatures is central to predicting the future of the temperate deciduous forest carbon sink.
Most studies on tree stem growth responses to warmer spring temperatures have focused on boreal or temperate conifers, which tend to respond to warmer spring temperatures with an earlier start to growth 22,23 and increased annual growth in mesic climates 24,25 . By contrast, little evidence exists as to how stem growth and ANPP woody respond to warmer spring temperatures in deciduous forests 8,9 . Close coordination of budburst and initiation of xylogenesis 9 suggest that warm spring temperatures should shift the onset of growth earlier alongside observed advances in leaf phenology [1][2][3] . However, earlier initiation of growth would not necessarily translate to earlier, faster or greater stem growth; rather, stem growth is dependent on environmental conditions on hourly to daily timescales 26,27 , and annual growth is more closely linked to conditions within the growing season than to growing season length 28 , GPP 21 or NEE 21 . Furthermore, growth of broadleaf deciduous trees may be sink-saturated 17,29 , such that longer growing seasons with more carbon fixation do not necessarily augment growth 21,30,31 . Tree-ring records, which can be used to examine annual growth but not growth seasonality, reveal that growth of temperate deciduous trees tends to be most sensitive to temperature or potential evapotranspiration between late spring and early summer 32,33 , with some evidence that warmer spring temperatures may have a modest positive effect on growth 25,33 . These observations do not necessarily align with the finding that warming spring temperatures increase annual forest CO 2 uptake in temperate deciduous forests 1,18 . Characterizing responses of stem growth to warming spring temperatures is critical to bridging this conceptual disconnect and understanding how forest biomass growth is likely to change as the climate warms.
Here, we evaluate how spring temperatures affect the timing, rates and annual increments of stem diameter growth of temperate deciduous trees across eastern North America. To test whether warmer spring temperatures advance the timing and extend the duration of stem diameter growth, we used dendrometer band measurements on 440 trees across two mid-latitude forests. To test whether warm spring temperatures consistently increased annual radial increments, we analysed 207 tree-ring chronologies from 108 forests.

Dendrometer band analysis
Using dendrometer band measurements taken throughout multiple growing seasons at the Smithsonian Conservation Biology Institute (SCBI; VA, USA; n = 119 trees from 2011 to 2020) and Harvard Forest (MA, USA; n = 321 trees from 1998 to 2003), we fit a logistic growth model 34 to determine the days of year (DOYs) when 25%, 50% and 75% annual diameter growth were achieved (DOY 25 , DOY 50 and DOY 75 ), the peak growing season length (L pgs = DOY 75 − DOY 25 ), the maximum growth rate (g max ) and the total annual increment in diameter at breast height (∆DBH; Fig. 1). This analysis was performed separately for ring-porous and diffuse-porous species (Extended Data Table 1), which differ in the seasonal timing of growth 27,35 (Extended Data Table 2 and Extended Data Fig. 1). These milestones in stem growth were compared with canopy foliage phenology (measured at the ecosystem level via remote sensing).
Both canopy foliage phenology and the timing of stem growth shifted earlier as spring temperatures increased (Fig. 2). We found a consistent effect of temperature (T max or T min ) throughout the spring, but the strongest effects on the timing of stem growth were found using T max during a critical temperature window (CTW). The CTW was identified by assessing the correlation between weekly T max and DOY 25 for all combinations of consecutive weeks from 1 January to mean DOY 25  Day of year at which 25% of growth is achieved DOY 50 Day of year at which 50% of growth is achieved DOY 75 Day of year at which 75% of growth is achieved L PGS Peak growing season length (DOY 75 − DOY 25 ) Growth rate g max Maximum growth rate Annual growth ΔDBH Annual growth from dendrometer band RWI Ring width index from tree-ring chronologies  Fig. 2). The CTW was defined as the weeks for which this correlation was strongest, and mean T max during this period (CTW T max ) was used as our independent variable. For ring-porous and diffuse-porous species at both sites, warmer CTW T max resulted in earlier achievement of seasonal milestones. Consistent with findings from previous studies, leaf phenological milestones advanced at both sites ( Fig. 2a Table 2).

Article
Whereas the length of time between canopy greenup and senescence (that is, the day when greenness dropped below 90% of its peak) increased with the CTW T max of the porosity group containing the dominant canopy species at each site (diffuse porous at SCBI and ring porous at Harvard Forest; Fig. 2a,b), there was no consistent lengthening of L pgs ( Fig. 1 and Extended Data Figs. 3 and 4).
In contrast to the pronounced effects of CTW T max on the timing of growth, its effects on g max and ∆DBH were inconsistent and usually weak ( Fig. 1 and Extended Data Figs. 3 and 4). Specifically, g max , which occurred on average within 5 days of DOY 50 , displayed extremely small negative changes (Harvard Forest) or changes in opposite directions (SCBI) in relationship to CTW T max for ring-porous and diffuse-porous species. ∆DBH displayed a weak positive relationship (+0.003-0.008 cm per °C) with CTW T max at SCBI and a weak negative relationship (−0.02 to 0.003 cm per °C) at Harvard Forest (Extended Data Fig. 3).

Tree-ring analysis
To understand how annual radial stem growth increments have responded to spring temperatures at the centennial scale, we analysed tree-ring chronologies of 12 species at SCBI 33 and four species at Harvard Forest (Extended Data Table 1), plus an additional 191 chronologies from 106 sites (Fig. 3, Extended Data Table 3 and Extended Data Fig. 5). In total, our analysis included 207 chronologies representing 24 broadleaf species at 108 sites distributed from Alabama (34.35° N latitude) to Michigan (45.56° N latitude) and spanning a 15 °C range in April T max . Across all chronologies, the standardized ring-width index (RWI) was significantly (95% confidence interval did not include 0) positively correlated with April T max for only 2% of chronologies: 1 of 142 ring-porous and 4 of 66 diffuse-porous species-site combinations (Extended Data Table 3). There were even fewer significant positive correlations with March and May T max : 0% and 1%, respectively (Extended Data Table 3). By contrast, RWI was frequently negatively correlated with T max during peak growing season months (May-August), with significant correlations for 52% ( Mean DOY 25 , DOY 50 and DOY 75 were estimated using the Bayesian model visualized, with confidence intervals, in Extended Data Fig. 3. Mean T max was calculated for each xylem architecture-site combination over the respective CTW, then turned into a ratio and assigned a colour on a gradient in which the coldest year in the sample is blue and the warmest is red. Leaf phenology years are coloured according to the CTW T max of the porosity group containing the dominant canopy species at each site (diffuse porous at SCBI and ring porous at Harvard Forest).
combinations for ring-porous and diffuse-porous species, respectively. T min generally exhibited weaker relationships to annual growth than T max , with few significant correlations between spring T min and RWI (Extended Data Fig. 6).
To test whether warm spring temperatures might result in storage of non-structural carbohydrates that would augment growth the following year 36 , we extended the analysis to examine correlations between RWI and T max in the previous year (Extended Data Fig. 7). This revealed little effect of previous spring temperatures on annual growth, with significant positive correlations of RWI to previous March or April T max for 5 of 142 ring-porous chronologies and to previous April or May T max for 7 of 66 diffuse-porous chronologies.
To test whether there may be an enhancement of growth by warmer spring temperatures that was offset by the negative effect of high summer temperatures, we tested for the joint effects of April and June-July T max on RWI. Results were qualitatively similar to the univariate correlations ( Fig. 3), with significant (P ≤ 0.05) positive correlations to April T max for only 4% of chronologies and significant negative correlations with June-July T max for 77% of chronologies (Extended Data Table 3).

Discussion
Together, our results demonstrate that warmer spring temperatures in the temperate deciduous forests of eastern North America advance the timing of stem diameter growth but have little effect on annual increments (Figs. 1-3). The observed advance in stem growth under warmer spring temperatures parallels advances observed for canopy foliage phenology 1,3 (Fig. 2a,b) and NEE 1,3 . However, inconsistent with  1  3  5  7  9  11  13  15  17  19  21  23  25  27  29  31  33  35  37  39  41  43  45  47  49  51  53  55  57  59  61  63  65  67  69  71  73  75  77  79  81  83  85  87  89  91  93  95  97  99  101  103  105  107  109  111  113  115  117  119  121  123  125  127  129  131  133  135  137  139  Article the concept that an earlier start to growth would increase ANPP woody , we demonstrate that warmer spring temperatures either hasten the deceleration of stem expansion or otherwise fail to translate extended growing seasons into biologically significant increases in stem growth (Fig. 1), and thereby have negligible effect on total annual growth for most species and locations (Fig. 3). Our observations suggest that the deceleration of stem expansion, which occurs in mid-summer near the time of peak canopy greenness 3,28 (Extended Data Fig. 1), is driven by cues other than photosynthate limitation, such as water stress 21,26,28 , nutrient limitation 37 , day length 28 or sink saturation 21,29 . This adds to a growing body of evidence for a sink limitation of stem growth 17,21,31 , in which global change factors known to enhance photosynthesis, such as longer growing seasons or elevated levels of CO 2 , do not cause a corresponding increase in stem growth 19,28 .
Combined with widespread observations that warming spring temperatures tend to lengthen the season of CO 2 uptake 3,18 and increase net annual CO 2 uptake 1,3-7 , our findings imply a lengthening of the period from peak stem growth to the cessation of CO 2 uptake by the ecosystem and an increase in carbon allocated to functions other than stem expansion in the current or following year. It remains theoretically possible that warm spring temperatures could augment ANPP woody , which, although routinely calculated based on stem growth, can be partially decoupled from it through differences in wood density or carbon content 21 . Extra carbon fixed in years with warm spring temperatures could potentially be allocated to the formation of more carbon-dense wood, either through enhanced cell wall thickening (a process that lags behind stem expansion 38 ) or to a higher ratio of high-density latewood to lower-density earlywood. However, existing evidence indicates that vessel features are most strongly controlled by summer drought stress in the previous (earlywood) or current (latewood) year, whereas warm spring temperatures have a neutral or negative effect on the width of latewood [39][40][41] . Thus, it is unlikely that warm spring temperatures have a positive effect on total carbon content of annual rings or ANPP woody .
The fate of any additional carbon fixed during years with warm spring temperatures remains unresolved, but possible destinations-including respiration, non-structural carbohydrate storage and production of foliage, reproductive structures, roots 30 or root exudates-generally have shorter residence times than woody growth. Indeed, when GPP of a mature forest was increased through experimental enrichment of CO 2 , ANPP woody remained unchanged, whereas additional carbon was released back to the atmosphere on relatively short timescales through enhanced respiration 19 . Consistent with this, it has been observed that carbon gains from an earlier spring can be offset through autumn or winter respiration 20 , although even the carbon in shorter-lived pools would often be carried over into the following year 42 . Thus, observed augmentation of NEE by warm spring temperatures 1,6,7 is likely to be compensated by increased respiration in subsequent years.
It is possible that as spring warming continues, forests will adjust to directional changes in growing season length with an enhancement of ANPP woody . Across latitudinal gradients, warmer spring temperatures are associated with earlier leaf-out 43 and longer growing seasons, which in turn are correlated with greater tree growth 44 , ANPP woody (ref. 45 ) and NEE 46 . Thus, warming spring temperatures are likely to increase the biophysical potential for annual tree growth. If extra photosynthate made available through a growing difference between GPP and ANPP woody is allocated to functions that relieve limitations on woody growth-for example, by enhancing nutrient and water acquisition through enhanced allocation to roots 30,47 -it is possible that warming spring temperatures could ultimately increase ANPP woody through indirect mechanisms. Understanding how warming spring temperatures are influencing carbon allocation within ecosystems remains a key outstanding question.
Regardless of the influence of spring temperatures on carbon cycling within the ecosystem, our results clearly demonstrate that the dominant effects of temperature on deciduous tree growth occur not in the spring, but during the peak growing season of the current or sometimes previous year (Fig. 3 and Extended Data Fig. 7), when increased atmospheric demand associated with high temperatures can limit both leaf-level gas exchange and stem growth 21,26,28,48 . Indeed, the timing of peak growth in June and July (Extended Data Table 2 and Extended Data Fig. 1) coincides with the timing of the greatest sensitivity of annual growth to T max (Fig. 3 and Extended Data Table 3). This finding is consistent with numerous tree-ring studies demonstrating strong sensitivity of growth to drought stress or high temperatures during the peak growing season 24,32,33,44 . Warm spring temperatures may also amplify summer drought stress in some times and places, effectively cancelling out any positive effects of an extended growing period 2,49,50 . Although such an interaction was unlikely to have had a major role within the scope of our dendrometer band study, given relatively mesic conditions and lack of significant correlation between spring temperatures and summer drought stress (see Methods), our tree-ring analysis does reveal a higher frequency of negative than positive correlations of annual growth to spring temperatures, particularly for ring-porous species in cooler climates ( Fig. 3 and Extended Data Table 3). Thus, warm spring temperatures can have a net negative effect on growth, particularly when water is limiting 25 .
As spring temperatures become increasingly warmer, growing seasons will start earlier. However, barring rapid acclimation of temperate deciduous forests to the warming conditions, advancement in the timing of stem growth (Fig. 1) is unlikely to provide a sustained augmentation of carbon sequestration in woody biomass and ensuant negative climate change feedback that is anticipated in most climate forecasting models 1,2,12,31 . Rather, the dominant effect of rising temperatures on temperate deciduous forest woody productivity will be a negative effect of high summer temperatures 15 (Fig. 3), which constitutes a positive feedback to climate change.

Online content
Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-022-05092-3.

Dendrometer band analysis
Dendrometer band measurements were collected at SCBI 51  summer SPEI values (4-month value of August) were similar to the average climate before the study period . Specifically, average spring T max was 10.9 ± 1.5 °C before the study period (range: 8.0-13.2) and 11.2 ± 1.0 °C (range: 10.1-12.2) during the study period, whereas the summer SPEI was 0.1 ± 0.9 (range: −1.8 to 1.7) before the study period and 0.2 ± 0.9 (range: −1.0 to 1.1) during. The driest summer during the study period (1999) had the fifth lowest SPEI value (−1.0) in the period 1970-2003, with precipitation of 52 mm per month in June-August compared to average monthly precipitation of 100 mm or more 35 .
Metal dendrometer bands were installed on 941 trees within the SCBI and Harvard Forest ForestGEO plots. Bands were placed on dominant species, including two diffuse and two ring-porous species at SCBI and eight diffuse and three ring-porous species at Harvard Forest (Extended Data Table 1). Although we do not estimate the ages of the trees in our sample, bands at both sites were placed on individuals of differing sizes in an attempt to measure trees across a range of ages. Bands were measured with a digital calliper approximately every 1-2 weeks within the growing season from 2011 to 2020 at SCBI and 1998 to 2003 at Harvard Forest. The number of bands measured at each site fluctuated slightly as trees were added or dropped from the census (for example, because of tree mortality). Across years, the number of bands sampled averaged 129 (range: 91-138) at SCBI and 717 (range: 700-755) at Harvard Forest.
Measurements were timed to begin before the beginning of spring growth and to continue through the cessation of growth in the fall. At SCBI, the median start date was 14 April, which was adjusted forward when early leaf-out of understory vegetation was observed, with the earliest start date being 30 March (in 2020). Measurements were continued through to fall leaf senescence, with the median end date being 17 October and the latest end date being 26 November (in 2012). At Harvard Forest, all measurements from 1998 were dropped because of a late start date (26 May). Among the remaining years, the median start date was 21 April and the median end date of 27 October. 1999 was an anomalous year in which initial measurements were taken on 5 January, but were not taken again until 15 April. The latest end date was 11 November 2002. In our analysis, each band year was treated independently, with no data overlap from one year to the next.
The raw dendrometer band data were screened to remove records or entire tree-years that were inappropriate for our analysis because of reduced reliability of predicted growth in the modelled curves. Specifically, we removed tree-years with small or negligible total growth (∆DBH ≤ 0.005 cm; SCBI = 26, Harvard Forest = 253) and tree-years in which the first intra-annual measurement was later than the first spring survey (trees that were missed in the initial census; SCBI = 22, Harvard Forest = 8). In total, this process removed 309 of the 2,701 available tree-year records for 2011-2020 at SCBI and 1999-2003 at Harvard Forest.
We fit a five-parameter logistic growth model 34 to dendrometer band data from each tree-year to define stem growth milestones and growth rates (Fig. 1). In particular, we modelled the observed DBH on a given DOY (that is, Julian days) as: Here, L and K are lower and upper asymptotes of the model. DOY ip is the DOY during which the inflection point in growth rate occurs, r shapes the slope of the curve at the inflection point, and θ is a tuning parameter controlling the slope of the curve towards the upper asymptote. This allows an asymmetric fit to the data, in which the onset of growth can be estimated independent of the cessation of growth. When θ = 1, g max occurs on DOY ip . The model outputs two additional variables, a and b, which represent the beginning and end DBH in each model year and are constrained by the first and last dendrometer band measurements. The model was fit in R v4.0 using the functions developed in the Rdendrom package 34 . These functions take the time series of manual dendrometer band measurements and return maximumlikelihood-optimized values of the above five parameters that best predict DBH for each DOY. An advantage of this approach is that short-term shrinkage and swelling associated with rain events 34,56 and measurement errors show up as residual variation and do not unduly influence the parameters of interest. Having fit the model for each band year of data, we then modelled DBH using these optimal parameter values in our logistic growth model and extracted the intra-annual growth variables of interest (Fig. 1).
After fitting the growth model, we removed tree-years with poor fits. Models were judged to be poorly fit if modelled growth parameters were outliers, which were commonly indicative of unrealistic fits (for example, growth occurring outside the growing season or over a very short period) and underlain by very slow tree growth or poor data records that passed the initial screening (described above). Modelled fits for tree-years were removed under two conditions: (1) g max was 2.5 or more standard deviations away from the mean for each site-xylem architecture group combination (SCBI = 3, Harvard Forest = 11), and (2) timing variables (DOY ip , DOY 25 , DOY 50 and DOY 75 ) were 2.5 or more standard deviations away from the means for their site, xylem architecture group and year (SCBI = 74, Harvard Forest = 101). In total, this process removed 189 of the 2,392 tree-year records deemed appropriate for analysis, leaving a total of 2,203 tree-years included in the final analysis (Extended Data Table 1). At both sites, the tree-years removed through this method were proportional to the original sample size, indicating that no species or size class was disproportionately removed compared with others. This process was repeated using 2 and 3 standard deviations as the cut-off for defining outliers, yielding qualitatively similar results.
Canopy foliage phenology data for the years 2001-2018 were extracted for SCBI and Harvard Forest from the MCD12Q2 V6 Land Cover Dynamics product (that is, MODIS Global Vegetation Phenology product) 57 via Google Earth Engine. For each year at each site, we extracted data from the pixel (500-m resolution) containing the centre of each forest plot. Using the daily MODIS 2-band EVI2 data, the product yields the timing of phenometrics (vegetation phenology) over each year, including timing of greenup, mid-greenup, peak and senescence, as used in this study. Data points were included in the analysis if they were flagged as 'good' or 'best' quality.
For the dendrometer band and leaf phenology analyses, climate data corresponding to the measurement periods were obtained from local weather stations at each focal site. For SCBI, weather data were obtained from a meteorological tower adjacent to the ForestGEO plot, via the ForestGEO Climate Data Portal v1.0 (https://forestgeo.github.io/ Climate/) 58 . The R package climpact (see www.climpact-sci.org) 59 was used to plot temperatures for visual inspection and to identify readings that were more than 3 standard deviations away from yearly means, which were labelled as outliers and removed from the dataset. Gaps in the SCBI meteorological tower data were subsequently filled using temperature readings obtained from a National Center for Environmental Information (NCEI) weather station located in Front Royal, VA, USA (https://www.ncdc.noaa.gov/cdo-web/datasets/GHCND/stations/ GHCND:USC00443229/detail). Daily temperature records for Harvard Forest, which had already been gap-filled on the basis of other local records, were obtained from the Harvard Forest weather station 60,61 . For each site, we used records of daily maximum (T max ) and minimum (T min ) temperatures. SPEI 55 values were obtained from the ForestGEO Climate Data Portal v1.0 (https://forestgeo.github.io/Climate/) 55,58,62 .
The CTW (Fig. 1), defined as the period over which T max was most strongly correlated with DOY 25 , was determined using the R package climwin 63 . This package tests the correlation between one or more predictor climate variable and a biological outcome variable over all consecutive time windows within a specified timeframe. It does so by reporting the correlation and ∆AICc, the difference in Akaike information criterion corrected for small sample size relative to a null model for each window. Here, we tested for correlation between temperature predictor variables (T max and T min ) and biological outcome variable DOY 25 over the timeframe from 1 January to the mean DOY 25 for the species group (by xylem porosity) and site (Extended Data Table 2). The time period yielding the lowest ∆AICc was selected as the CTW. To avoid spurious correlations that could occur using temperature data at the daily resolution, we ran this analysis with weekly resolution, using temperatures averaged over weekly time periods. Because T max proved to have a generally stronger influence over DOY 25 and other growth parameters, we focused on this variable in our ultimate model, as opposed to T min . We defined CTW for DOY 25 , as opposed to other parameters describing the timing of growth, because spring temperatures should have the most direct influence on this variable.
To ensure that patterns were robust under an alternative definition of CTW, and to parallel the monthly time windows used in our tree-ring analysis (detailed below; Fig. 3 and Extended Data Figs. 6 and 7), we also ran analyses in which we fixed the CTW identified by climwin to be the month with the most days in the CTW (Extended Data Table 2) for each critical window. The months identified were March and April for ring-porous and diffuse-porous species at SCBI, respectively, and April and May for ring-porous and diffuse-porous species at Harvard Forest, respectively.
Correlation between the dendrometer band-derived growth parameters (DOY 25 , DOY 50 , DOY 75 , L pgs , g max and ∆DBH; Fig. 1) and CTW T max (at weekly or monthly resolution, as described above) were assessed using a linear mixed model in a hierarchical Bayesian framework. Analyses were run for both T max and T min , with qualitatively similar results, but we present only results for T max , which had an overall stronger correlation with growth parameters. Mixed-effect models were used to test the response of growth parameters to fixed effects of xylem porosity and mean T max (or T min ) during the CTW, along with random effects of species and of individual tree. We ran separate models for each site, and for the response of all growth parameters to T max (or T min ). This mixed-effect model was run within a hierarchical Bayesian framework and fit using the rstanarm version 2.21.3 R interface to the Stan programming language 64,65 . In all cases, unless otherwise specified, all prior distributions were set to be the weakly informative defaults.
To rule out the possibility that observed patterns were strongly influenced by summer drought, we examined the relationship between spring temperatures and summer SPEI indices. Linear models were run with 4-month, 6-month and 12-month SPEI values of June, July and August versus April T max to determine whether warm spring temperatures were associated with greater summer drought stress in our dataset. No significant correlations were found (all P > 0.05).

Tree-ring analysis
We analysed tree-ring records for 108 sites, including our focal sites. All cores had been previously collected, cross-dated and measured using standard collection and processing methodologies 66,67 .
Dominant tree species were cored at both SCBI 33,51 and Harvard Forest 3,68,69 following sampling designs that covered a broad range of DBH. We analysed records for the ring-porous and diffuse-porous species at each site (Extended Data Table 1), but excluded semi-ring-porous species (for example, Juglans nigra L. at SCBI) and conifers (for example, Tsuga canadensis at Harvard Forest). We studied a total of 976 cores, which included 12 species at SCBI and four species at Harvard Forest (Extended Data Table 1).
The tree-ring records from our focal sites were complemented with a much larger collection spanning 106 deciduous and mixed forest sites in eastern North America 32,70,71 . For the majority of sampled populations (that is, site-species combinations), sampling focused on canopy trees (typically more than 20 trees per population) 32,70,71 , whereas approximately 15% of the total 207 chronologies came from plot-level collections in which trees above a certain diameter (typically 10-cm DBH) were censused and cored 33,69 . Again, analyses were limited to broadleaf deciduous species with clearly defined xylem porosity (that is, excluding semi-ring porous).
For each species-site combination, we converted tree-ring records into the dimensionless RWI to emphasize interannual variability associated with climate 72 . A two-thirds n spline was applied to each core using ARSTAN V49_1b to produce standardized ring-width series; n is the number of years in each series 72,73 . An adaptive power transformation, a process that also stabilizes the variance over time 74 , was used to minimize the influence of outliers in all series. Low series replication, often in the earliest portions of a chronology collection, can also inflate the variance of tree-ring records 75 . The one-thirds spline method was chosen when replication in the inner portion of each chronology (the earliest approximately 30-50 years of each record depending on the full chronology length) was less than three trees. When replication was greater than n = 3 trees, we used the average correlation between raw ring-width series (rbar) method. The robust biweight mean chronology (RWI) for each species-site combination was calculated from the ring-width indices following variance stabilization 73 . We defined chronology start year (Extended Data Table 1) as the year in which subsample signal strength passed a threshold of subsample signal strength = 0.8, or where 80% or more of the population signal was captured in the chronology.
For the analysis of correlation between RWI and climate variables, we obtained monthly T max and T min data for 1901-2019 from CRU v.4.04 (ref. 54 ). Correlations between monthly climate and RWI were assessed in R v 4.0 (ref. 76 ) using the packages dplR 77 and bootRes 78 . Reported correlations and significance were determined using bootstrapped confidence intervals. Summary figures were created using the package dplR 77 (Fig. 3 and Extended Data Figs. 6 and 7).
Our analysis focused on assessing correlations of RWI to months spanning January to September of the current year (presented in Fig. 3 and Extended Data Fig. 6). To test for potential lag effects of spring temperatures on growth the following year, we also ran a version of the analysis extending back to include climate of every month of the previous year (Extended Data Fig. 7). Correlations and significance levels for months March-August are given in Supplementary  Table 1.
We used a multivariate model to test for joint effects of April and summer T max on RWI. We focused on April to represent spring temperatures because it was the month with the greatest overall alignment with the CTWs identified in the dendrometer band analysis and had the highest rate of positive correlations with RWI (Extended Data Table 3). We began by testing univariate correlations of T max over three summer windows: June, June-July and May-August. Having determined that, among these, June-July explained the most variation, we then analysed the joint effects of April T max and June-July T max on RWI for each chronology independently using the base lm() function in R. Slopes and P values for each chronology are given in Supplementary Table 1. Although some models may have benefitted from data transformations, we determined that assumptions of normality and homoscedasticity were sufficiently met for the purposes of this analysis.

Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
Original tree cores are archived at the institutions of various members of the author team (Harvard Forest, SCBI, Indiana University and University of Idaho) and will be made available on reasonable request.