Across 1,175 tree-ring sites, an earlier moderate drought did not detectably toughen trees

A common intuition holds that stress hardens: a tree that came through a dry spell should cope better with the next one. We tested that idea on 1,175 tree-ring sites in the contiguous United States, 65 species and 3,780 site-drought events, with the test written down and hash-locked before the run. A moderate drought four to eleven years earlier moved log resilience to a later severe drought by +0.039 (95% interval -0.011 to +0.090), inside a time-shift null whose 95th percentile was 0.065 (p 0.18).c1 The preregistered steeling test failed. A severe prior drought lowered growth in the drought year (resistance -0.087, p 0.006) and raised recovery after it by +0.096, so resilience barely moved.c2 Part of that recovery gain follows from how the indices are built. A rerun with drought taken from 0.5 degree gridded data on 1,097 sites gave +0.025 (-0.009 to +0.059), below its own time-shift 95th percentile of 0.038 (p 0.14).c3 The interval still allows a gain of up to about 9 percent in resilience, so small conditioning effects are not ruled out.

The question comes from a broad idea about stress and adaptation: that a moderate earlier stress leaves an organism better prepared for a harder one. Applied to trees, we call this steelingDefinition: steeling psychology and stress biology, here applied to trees The idea that a moderate earlier stress leaves an organism better able to cope with a later, harder one. Here, that a tree which lived through a moderate drought loses less growth to a later severe drought. Excludes recovery from the same drought, and any gain that follows only from how the growth indices are computed. full definition. Trees make a good test, because each annual ring records growth, and public archives hold thousands of sites with records over a century long.

Forest ecology points mostly the other way. Anderegg and colleagues [1] examined stem growth after severe drought at 1,338 forest sites worldwide and reported a widespread drought legacyDefinition: drought legacy forest ecology Growth that stays below its expected level for one or more years after a drought ends. Excludes the loss of growth in the drought year itself. full definition: growth stayed low for one to four years after the drought, most often in dry ecosystems and in the pine family. Anderegg and colleagues [2] later asked whether forests acclimate to repeated droughts or become more vulnerable, and concluded that successive droughts are generally more damaging, with conifers the most affected. We read that paper at the level of its publisher page and press coverage.

The usual measures of drought response come from Lloret, Keeling and Sala [3], who proposed resistance, recovery and resilience indices from ring growth around a low-growth episode in ponderosa pine. Schwarz and colleagues [4] reviewed 82 studies that used these indices and listed cases where they mislead: a lagged growth minimum, depressed growth in the reference years before the drought, and repeated droughts. Their review bears on our result, because resilience equals resistance times recovery, so a weak drought-year ring raises recovery by arithmetic alone.

Our searches, run on 9 November 2026, found these and related studies on legacies and repeated drought. We did not find a study of this exact contrast, moderate earlier drought against later resilience, scored against a time-shifted drought record. That is a search result, not proof that none exists.

The design was written as a preregistrationDefinition: preregistration research method (metascience) A dated written plan for a study, fixed and hash-locked before any result is seen: the question, the data, the statistic, the controls, the forecasts and the kill conditions. Excludes a plan revised after looking at the outcome, and a protocol written only for the record after the run. full definition and hash-locked on 29 October 2026 at 22:54 UTC, before any hypothesis value was computed. A smoke run (a short trial run that checks the code executes) at small sizes before the lock printed design counts only: 1,175 sites, 3,780 events, 65 species, 177 climate divisions, and the prior-drought group sizes. No coefficient was printed or read.

The locked file covers four tests, A to D. Tests B to D are simulations of a loaded rod, a small neural network and a Bayesian learner, and they say nothing about trees. This paper reports Test A, the only test on real data. Across all four tests the run returned 22 verdictsDefinition: verdict research method (our usage) The pass or fail of one preregistered rule set against the run's numbers, read off by code. Excludes an interpretation written after reading the numbers. full definition, 10 passing and 12 failing.

Test A had four preregistered predictions, each a set of rules that must all hold:

The forecast written before the run: A1 weak or null, A2 likely positive because severe drought leaves legacies, and A3 failing.

Rings: standard chronologiesDefinition: standard chronology dendrochronology One yearly ring-width index for a site, made by removing each tree's age trend and averaging the trees, scaled so that an ordinary year is near 1. Excludes raw ring widths of single trees, and density or isotope records. full definition for the contiguous United States from the International Tree-Ring Data Bank [5], held by NOAA's National Centers for Environmental Information as public data with a request to cite the contributors. Ring width only. Years from 1896 with at least 5 samples where a sample count exists; index values at or below 0 dropped.

Drought, main test: NOAA nClimDiv divisional Palmer Drought Severity IndexDefinition: Palmer Drought Severity Index (PDSI) climatology and hydrology A monthly index of soil-moisture supply against demand built from temperature and precipitation. 0 is normal, -2 is moderate drought and -3 or lower is severe drought. Excludes a direct measurement of soil water, and drought indices built on precipitation alone. full definition, version 1.0.0 built on 6 October 2026 [6], US federal data with no restriction on use. Each site was placed in its climate divisionDefinition: climate division US climatology (NOAA) One of the regions into which NOAA splits the contiguous United States for climate records, each with one averaged monthly series per variable. Excludes a single weather station, and a regular grid cell. full definition by its coordinates, and the yearly drought value is the June to August mean of that division. The final sample is 1,175 sites, 65 species and 177 divisions, with 579 division drought years and 3,780 site-drought events.

Drought, rerun: the self-calibrating PDSIDefinition: self-calibrating PDSI (scPDSI) climatology A version of the Palmer index whose constants are fitted to each location's own climate, so that the thresholds mean similar rarity everywhere. Here taken from the CRU TS gridded product at 0.5 degree. Excludes the original Palmer index with fixed constants, and station measurements. full definition from CRU TS 4.10 at 0.5 degree, 1903 to 2025 [7] [8], released under the Open Database License 1.0 with attribution to the Climatic Research Unit, University of East Anglia. Each site takes the nearest cell with complete June to August data in at least 60 years, within 1 degree. This gave 1,097 sites, 64 species, 505 cells and 2,549 events.

A division averages over a large area, and a site sits at one point in it. The rerun exists to pay that debt. A 0.5 degree cell is closer to the site and is still not a station.

A focal year $t$ is a division year with June to August PDSI at or below -3. An event needs ring growth in years $t-3$ to $t+3$, no other severe year in $t-3$ to $t-1$ or $t+1$ to $t+3$, and drought data for the exposure window $t-11$ to $t-4$.

With $g_t$ the ring index in the drought year, $G_{\mathrm{pre}}$ the mean of the three years before and $G_{\mathrm{post}}$ the mean of the three years after, the Lloret indicesDefinition: Lloret indices dendroecology Three ratios of ring growth around a drought year: resistance is the drought-year growth over the mean of the three years before, recovery is the mean of the three years after over the drought-year growth, and resilience is the mean after over the mean before. Excludes measures of tree death, and growth responses that unfold over more than three years. full definition are resistance, recovery and resilience:

All three are analysed as natural logarithms. The product in eq. 1 matters later: a lower drought-year ring lowers $R_t$ and raises $R_c$ by the same factor.

The exposure is the lowest June to August PDSI in years $t-11$ to $t-4$. An event is in the none group if that minimum is above -2, moderate if it lies above -3 and at or below -2, and severe if it is at or below -3. The groups held 986, 941 and 1,853 events. The dose is the larger of 0 and minus that minimum.

For event $i$ at site $s$ in year $t$:

where $M_i$ and $S_i$ mark the moderate and severe groups, $D_t$, $D_{\mathrm{pre}}$ and $D_{\mathrm{post}}$ are the drought-year PDSI and the mean PDSI of the three years before and after, and ${\alpha }_s$ and ${\tau }_t$ are site and year fixed effectsDefinition: fixed effects econometrics and statistics One free constant per site (or species) and one per calendar year in a regression, so that differences between sites and between years cannot be mistaken for the effect of the variable under test. Excludes random effects, which treat those constants as draws from one distribution. full definition. Species are constant within a site, so site effects absorb them; a model with species and year effects was reported as secondary. Standard errors are clusteredDefinition: clustered standard errors econometrics Standard errors computed so that events in the same climate division may be correlated with each other, which widens the interval when they are. Excludes ordinary standard errors that treat every event as independent. full definition by climate division. Resistance and recovery were fitted with the same model as secondary outcomes. The coefficient under test is ${\beta }_m$.

In the rerun, the preregistered primary model used species and year effects in place of site effects, with the same covariates and the same clustering by division.

For prediction A3, resilience with site, year and covariate effects removed and no group terms was set against the dose, and the two-lines test was applied. Its statistic is the smaller of the rising line's z and minus the falling line's z, so a large positive value means a rise and then a fall.

Test A had no planted checkDefinition: planted check research method; called a positive control in experimental science Data built so that the right answer is known, run through the same analysis before the real data, to show that the analysis can see the effect when it is there. Excludes any check made on the real data, which cannot show what the analysis misses. full definition: no synthetic forest with a known steeling effect was built and run through the model. The placebo window serves as the design check, together with three controlsDefinition: control statistics and experimental design A version of the data with the structure under test removed (for example shuffled or pooled), scored the same way as the real data, so the real score has something to beat. Excludes a comparison with a different dataset or with a published number. full definition.

Each p value from draws is the share of draws at or above the real value, with one added to the count and to the number of draws:

where $k$ is the number of draws at or above the real value and $n$ the number of draws. The real value is compared with the null distributionDefinition: null distribution statistics The spread of scores a control gives over many random draws. Its 95th percentile is the bar a real score must clear. Excludes a theoretical distribution assumed rather than drawn from the data. full definition's 95th percentile. The rerun used the same time-shift null on each grid cell's record, 200 draws.

Table 1 gives the estimates, and Figure 1 draws the main ones. A moderate drought four to eleven years earlier moved log resilience to a later severe drought by +0.039 (95% interval -0.011 to +0.090), inside a time-shift null whose 95th percentile was 0.065 (p 0.18).c1 Its raw p was 0.128 and its adjusted q 0.166. A severe prior drought gave +0.009 (-0.039 to +0.057). Moderate minus severe was +0.030 (-0.014 to +0.074, q 0.214).

Figure 2 sets the moderate-drought estimate against the three null ranges. It sits inside the time-shift range in both runs and above the label-permutation range (95th percentile 0.020, p 0.004). The label permutation keeps each site's drought timing, so beating it says less, and the preregistered rule used the time shift.

The components explain the null in resilience. A severe prior drought lowered growth in the drought year (resistance -0.087, p 0.006) and raised recovery after it by +0.096, so resilience barely moved.c2 By eq. 1, a lower drought-year ring raises recovery by the same factor it lowers resistance, so the recovery gain is partly arithmetic and is not evidence of conditioning.

The placebo window was clean: drought after the event gave +0.006 (p 0.86).c4 No rise-then-fall shape appeared with dose. The two lines fell and then rose, neither significantly (z -0.89 and +0.81), and the statistic of -0.89 sat below its shuffle 95th percentile of 1.26. The mean residual resilience by number of moderate-drought years in the window was -0.007 for none (1,705 events), +0.005 for one (1,268), +0.008 for two (643), -0.019 for three (103) and +0.059 for four (44). These five groups hold 3,763 of the 3,780 events; the result file lists no group for the other 17.

With species and year effects in place of site effects, the moderate prior drought gave +0.037 (0.000 to +0.074, p 0.047), and severe +0.035 (p 0.089). The two did not differ (+0.002), so under that model any prior drought went with slightly higher resilience, with no edge for a moderate one. The rerun tested that sign.

The rerun, preregistered and locked on 30 October 2026 at 03:20 UTC, took drought from the nearest 0.5 degree cell of the CRU self-calibrating PDSI. Its kill conditionDefinition: kill condition research method (our usage, after Popper's falsification) A rule written in the preregistration that, when met, decides against the claim with no room for judgement after the run. Excludes reasons found after the run for setting a result aside. full definition fires when the 95% interval of the species-effects estimate includes 0 or the estimate sits at or below the time-shift 95th percentile. Table 2 gives its estimates.

A rerun with drought taken from 0.5 degree gridded data on 1,097 sites gave +0.025 (-0.009 to +0.059), below its own time-shift 95th percentile of 0.038 (p 0.14).c3 The kill condition fired on both of its rules, and the adjusted q was 0.174. The small positive sign under division drought with species effects did not survive the finer drought data. Resistance fell and recovery rose, as in the main run.

Table 3 sets the forecasts written before the main run against the outcomes.

The test asked whether moderate drought conditions trees for a later severe one. It found no effect beyond what a drought record moved in time also produces, in two runs with two drought products. That is an honest nullDefinition: honest null research method (our usage) A working test that found no effect beyond a stated bound, a forecast that failed, or an error of our own found and fixed, reported with the same weight as a positive result. Excludes a void run, and a result left unpublished because it was negative. full definition: a working test, a clean placebo, and an estimate inside its null.

The result fits the direction of Anderegg and colleagues [2], who reported that repeated droughts tend to hurt more. Our severe-prior estimates of resistance fit the legacy picture of Anderegg and colleagues [1]: a tree that has been through a severe drought lays down a narrower ring in the next one.

The recovery gain is where a careless reading would find steeling. Recovery rose by about as much as resistance fell. Schwarz and colleagues [4] warned that Lloret's indices mislead when growth in the reference years is depressed or droughts repeat, and eq. 1 shows the simplest version of the problem. A study that reported recovery alone would see trees "bouncing back" better after earlier drought. The resilience ratio, which cancels the drought-year ring, shows no net gain.

The label-permutation null is the trap on the other side. The estimate beat it with p 0.004. That null keeps every drought date in place, so it compares events within a site's own history and does not remove effects tied to when droughts fell. We fixed the time-shift null as binding before the run, and it was not beaten.