Auto-Research PipelineA durable, human-gated pipeline for computational research

← All examples

Research run510551359416· 24 October 2026

Does intraspecific competition exceed interspecific competition in a desert rodent community, and is interspecific competition stronger between body-size-similar species?

CompleteArchived runEcology & Evolutionary Biology · Modelled on Ecology Letters

Paper

Evidence for self-limitation, but not limiting similarity, in a desert rodent community

Abstract

Coexistence requires self-limitation; limiting similarity predicts that similar-sized species compete most. We tested both with 13,534 captures of 11 rodent species on eight Portal plots without rodent removal (1977–2002), fitting a Lotka–Volterra model with a body-size kernel by two-stage least squares (2SLS) with lagged-density instruments against census error. Intraspecific point estimates were positive for all 11 species, with intervals excluding zero for seven; 2SLS coefficients were 2.24-fold smaller than OLS (median), consistent with census-error bias. The pre-specified self-limitation rule held in the primary model (mean relative interspecific effect 0.016 [−0.400, 0.389]) and in 9 of 12 alternatives, failing in the three least precise, including second-lag instruments (upper limit 2.145). An i.i.d. Sargan test rejected joint instrument validity; dependence-robust tests did not, except, marginally, one allowing correlation only within plot-censuses. In the primary model, competitive effects declined toward smaller, not larger, competitors, contrary to limiting similarity; second-lag instruments reversed this.

Keywordscoexistence theorycompetition coefficientslimiting similaritybody sizecensus errordensity dependencedesert rodentsPortal Project

§Ask about this paper

Grounded in this run

Questions are answered from this run's record — the paper, the executed analysis, the data provenance and the referee report. Nothing else is consulted.

This run is finished, so the recorded paper cannot be rewritten. Questions are still answered from its record.

Pipeline

1.Literature review

Attempt 3Done

Literature review complete (live:agent:claude-opus-5-5): 10 key papers, 42 grounded sources, 6 open gaps.

Summary

Coexistence theory states the question precisely: a set of competitors coexists stably only if each species limits itself more than it limits the others, that is, if intraspecific exceeds interspecific competition when both are expressed per unit of the same competitor (Chesson 2000). The best quantitative test is for plants, where intraspecific competition averaged four to five times interspecific competition among pairs with competitive effects, and most remaining pairs combined self-limitation with interspecific facilitation (Adler et al. 2018). For animals there is no comparable synthesis. Decompositions of community time series attribute most population regulation to intraspecific density dependence and little to interspecific interactions (Mutshinda et al. 2009), but such estimates are fragile: census error and the openness of counted subpopulations inflate apparent intraspecific density dependence and attenuate between-species effects, biasing analyses toward the very conclusion being tested (Freckleton et al. 2006). The second half of the question invokes limiting similarity: if resource-utilisation curves overlap more for similar species, competition coefficients should decline with niche distance (MacArthur & Levins 1967). Desert granivorous rodents are the textbook case: species of similar body mass (ratio < 1.5) co-occur less often than expected by chance, a pattern that weakens when other guilds are included (Bowers & Brown 1982). Modern theory warns that species differences also generate fitness differences, so similarity need not intensify competition (Mayfield & Levine 2010), and in forest trees interaction strength followed trait hierarchy rather than trait similarity (Kunstler et al. 2012). At Portal, Arizona, long-term press experiments showed that kangaroo rats suppress small granivores but not insectivorous Onychomys, and that the three Dipodomys species interact asymmetrically (Brown & Munger 1985). Mark-recapture comparisons of the 11 commonest species showed correlated year-to-year fluctuations, implying shared responses to rainfall and food, and dispersal that declines with body size (Brown & Zeng 1989). What is missing is a population-dynamic estimate, for a whole animal community, of how per-capita conspecific and heterospecific effects compare, and whether heterospecific effects scale with body-size similarity once competitive hierarchy, shared environmental forcing and observation error are dealt with. The Portal monitoring data (monthly censuses on fenced plots, 1977-2002) allow this, provided analyses respect the plot treatment histories: several plots labelled 'Control' in the teaching database received seed additions, annual-plant removals or ant removals, although no rodent was ever excluded from them.

Key papers (10)

  1. Mechanisms of Maintenance of Species Diversity. Chesson, P. - Annual Review of Ecology and Systematics 31: 343-366, 2000.
    Defines the condition tested here: stable coexistence requires each species to limit itself more than its competitors (alpha_jj > alpha_ij), with niche overlap and fitness differences as the two components.
  2. Competition and coexistence in plant communities: intraspecific competition is stronger than interspecific competition. Adler, P.B., Smull, D., Beard, K.H. et al. - Ecology Letters 21: 1319-1329, 2018.
    The benchmark synthesis: in plants intraspecific competition averaged four- to five-fold interspecific competition among competitive pairs, and 93% of the other pairs combined self-limitation with interspecific facilitation.
  3. The Limiting Similarity, Convergence, and Divergence of Coexisting Species. MacArthur, R. & Levins, R. - The American Naturalist 101: 377-385, 1967.
    Source of the prediction that competition coefficients decline with niche distance, which motivates testing whether interspecific competition is stronger between body-size-similar species.
  4. Opposing effects of competitive exclusion on the phylogenetic structure of communities. Mayfield, M.M. & Levine, J.M. - Ecology Letters 13: 1085-1093, 2010.
    Argues that species differences raise both stabilising niche differences and fitness differences, so similarity need not mean stronger competition; frames similarity versus hierarchy.
  5. Competitive interactions between forest trees are driven by species' trait hierarchy, not phylogenetic or functional similarity: implications for forest community assembly. Kunstler, G., Lavergne, S., Courbaud, B. et al. - Ecology Letters 15: 831-840, 2012.
    Template estimator: interaction coefficients modelled as functions of absolute (similarity) and signed (hierarchy) trait differences in a growth model; found hierarchy, not similarity, drove 275 tree interactions.
  6. Body Size and Coexistence in Desert Rodents: Chance or Community Structure?. Bowers, M.A. & Brown, J.H. - Ecology 63: 391-400, 1982.
    Classic evidence for size-structured competition in desert granivores: species with body-mass ratios < 1.5 co-occur less than expected; the pattern weakens when guilds are pooled.
  7. Experimental Manipulation of a Desert Rodent Community: Food Addition and Species Removal. Brown, J.H. & Munger, J.C. - Ecology 66: 1545-1563, 1985.
    Portal press experiments on 0.25-ha plots: removing Dipodomys increased small granivores but not insectivorous Onychomys; D. spectabilis removal caused reciprocal shifts in D. merriami and D. ordii; responses were lagged and asymmetric.
  8. Comparative Population Ecology of Eleven Species of Rodents in the Chihuahuan Desert. Brown, J.H. & Zeng, Z. - Ecology 70: 1507-1525, 1989.
    Mark-recapture demography of the 11 commonest Portal rodents: positively correlated year-to-year fluctuations (shared environmental forcing) and dispersal inversely related to body size (plot-scale turnover).
  9. What drives community dynamics?. Mutshinda, C.M., O'Hara, R.B. & Woiwod, I.P. - Proceedings of the Royal Society B 276: 2923-2929, 2009.
    Animal community time series: dynamics dominated by environmental fluctuations, populations regulated by intraspecific interactions, interspecific interactions broadly minor.
  10. Census error and the detection of density dependence. Freckleton, R.P., Watkinson, A.R., Green, R.E. & Sutherland, W.J. - Journal of Animal Ecology 75: 837-851, 2006.
    Shows that sampling error and open subpopulations make log-growth-on-density regressions spuriously density dependent, the bias that favours finding intra > inter and that this study must correct.

Methods in use

  • Phenomenological competition coefficients from discrete-time Lotka-Volterra, Ricker or Beverton-Holt models of per-capita growth on conspecific and heterospecific densities; comparison through alpha_ij/alpha_jj, niche overlap and fitness ratios, reported with 95% intervals and as the fraction of pairs with stronger intra- than interspecific effects (Chesson 2000; Adler et al. 2018).
  • Multispecies time-series models (first-order multivariate autoregressive or Gompertz) and variance decompositions of community dynamics into environmental, intraspecific and interspecific components (Mutshinda et al. 2009).
  • Corrections for census error and open subpopulations when estimating density dependence, since uncorrected log-growth regressions inflate intraspecific effects (Freckleton et al. 2006).
  • Trait-based competition kernels in which interaction coefficients depend on absolute (similarity) and signed (hierarchy) trait differences, estimated jointly from growth data (Kunstler et al. 2012).
  • Null-model tests of body-size ratios and co-occurrence in rodent assemblages (Bowers & Brown 1982).
  • Long-term press experiments with semipermeable exclosures compared against unmanipulated plots (Brown & Munger 1985), and mark-recapture comparisons of demography and dispersal (Brown & Zeng 1989).

Open gaps

  • No population-dynamic estimate exists for a whole desert rodent community of whether each species limits itself more than it limits heterospecifics; animal evidence on intra- versus interspecific competition is thin and mixed.
  • The body-size similarity hypothesis for desert rodents rests on co-occurrence and size-ratio patterns and on removals of a few large species; per-capita competition has not been related to body-size difference across many species pairs.
  • Similarity and competitive hierarchy make different predictions but have not been separated for animals in a dynamic framework; with few species, size difference and competitor size are nearly collinear, which needs explicit treatment.
  • Census error inflates intraspecific and attenuates interspecific estimates, biasing comparisons toward intra > inter; an analysis must correct it (for example with lagged abundances as instruments) and show how much the correction matters.
  • Shared responses to rainfall and food make heterospecific abundances covary positively; designs that compare plots within the same census can remove community-wide forcing, but plot-scale densities in open populations still mix competition with movement.
  • Analyses of the Portal teaching data must respect plot treatment histories: some 'Control' plots received seed additions, annual-plant removals or ant removals even though rodents were never excluded.
Grounded search sources (42) · mode: live:agent:claude-opus-5-5

2.Ideation

Attempt 2Done

3 candidate ideas generated; selected: “Self-limitation versus size-structured interspecific competition in the Portal rodent community: a trait-kernel Lotka-Volterra model corrected for census error”.

  1. Idea 1Selected

    Self-limitation versus size-structured interspecific competition in the Portal rodent community: a trait-kernel Lotka-Volterra model corrected for census error

    Hypothesis. H1: every species limits itself more than it limits heterospecifics - the community mean of q_ij = alpha_ij / a_j (the per-capita effect of species j on species i divided by j's effect on itself, both per unit of j's metabolic biomass) is below 1, with the upper 95% bootstrap limit < 1. H2 (limiting similarity): interspecific competition per unit of competitor metabolic biomass declines with the absolute log body-mass difference between focal and competitor (kernel slope k_sim < 0), holding competitive hierarchy (signed log mass difference) and feeding guild constant.

    Dataset. Portal Project long-term rodent monitoring, Chihuahuan Desert portal_rodent_surveys

    Test plan, novelty and risk

    Test plan. Data: portal_rodent_surveys captures, Oct 1977-Dec 2002. Sample filter: the eight plots from which no rodent species was ever excluded, restricted to periods without food manipulation according to the Portal treatment history - plots 4, 8, 11, 12, 14, 17 from 1 Oct 1977 and plots 2, 22 from 1 Jan 1988 (seed additions and annual-plant removals before then); ant removals on plots 4, 8, 12, 17 are absorbed by plot x phase fixed effects (phase split at 1 Jan 1988). Trapping dates are merged into monthly censuses; abundance N = captures per trapping night per plot; species = rodents with >= 250 captures in the sample; body mass m = median weight of all weighed captures; competitor density E_j = N_j (m_j / m_DM)^0.75 (metabolic D. merriami-equivalents). Model (stacked over focal species i, discrete-time Lotka-Volterra/Ricker): ln(1+N_i,t+1) - ln(1+N_i,t) = plot-phase_i + census_i - a_i E_i,t - sum_j alpha_ij E_j,t, with alpha_ij = alpha0_i + k_sim |ln m_j - ln m_i| + k_hier (ln m_j - ln m_i) + k_guild [same guild]; species x census fixed effects remove community-wide rainfall and seed pulses. Estimator: two-stage least squares with densities at t-1 instrumenting densities at t, to remove the census-error bias that inflates intraspecific and attenuates interspecific coefficients; Sanderson-Windmeijer first-stage F reported. Inference: block bootstrap over calendar years (999 resamples) for 95% intervals; 999 species-label permutations of the size-difference matrix for k_sim. Measured: a_i per species; kernel coefficients; mean and median q_ij and fraction of ordered pairs with q < 1; D. merriami-D. ordii pair ratios. Refutation: H1 is refuted if the mean q interval includes or exceeds 1; H2 is refuted if k_sim >= 0, and judged unsupported unless its interval lies below 0, the permutation p < 0.05 and k_sim stays negative in five variants (OLS, t-2 instruments with a Sargan test of t-1 versus t-2, per-capita currency, granivores only, hindfoot length as size axis).

    Novelty. Body-size-structured competition in desert rodents has been inferred from co-occurrence and size ratios and from removals of a few large species, never from per-capita competition coefficients estimated for a whole community. This study estimates the intra/interspecific contrast in the unit-consistent form that coexistence theory requires, separates similarity from competitive hierarchy as done for trees but not for animals, and corrects the census-error bias that favours finding intraspecific dominance.

    Expected outcome. Strong self-limitation (mean q well below 1), as in plant syntheses and animal time-series decompositions. For H2 a modest negative similarity slope is plausible if seed-size partitioning tracks body size, but similar-sized species may also be near-equivalent competitors, so the similarity term may be weak relative to hierarchy and sensitive to specification.

    Main risk. Plot-scale densities in open populations mix competition with movement and turnover, inflating intraspecific coefficients; lagged-density instruments are invalid if census errors or density effects persist across censuses; with 11 species, size difference and competitor size are nearly collinear; plot-specific resource pulses are not removed by census fixed effects; ant-removal and unmanipulated plots are pooled.

  2. Idea 2

    Experimental release of small granivores after kangaroo-rat exclusion as a function of body-size similarity to Dipodomys

    Hypothesis. Species whose body mass is closer to that of the excluded kangaroo rats show larger proportional increases after Dipodomys exclusion: the log response ratio of abundance on exclusion versus open plots declines with |ln m_species - ln m_Dipodomys|.

    Dataset. Portal Project long-term rodent monitoring, Chihuahuan Desert portal_rodent_surveys

    Test plan, novelty and risk

    Test plan. Use the January 1988 reassignment in which plots 6, 13, 18, 20 became kangaroo-rat exclosures, with plots 2, 4, 8, 11, 12, 14, 17, 22 as open plots, in a before-after-control-impact design on per-species mean captures per plot-census (1985-1987 versus 1990-2002). Estimate per-species BACI log response ratios with year-block bootstrap intervals and regress them on log mass difference to the Dipodomys mean mass, with a species-label permutation test. Refuted if the slope is >= 0 or its interval spans 0.

    Novelty. Turns the repeated Portal exclusion experiment into a dose-response test of size similarity instead of a single treatment contrast.

    Expected outcome. Large releases of Chaetodipus and Perognathus, none in Onychomys; the size-similarity slope is uncertain because only about seven species respond.

    Main risk. The 1988 exclosure plots received seed additions and plant removals before 1988, confounding the before period; only one competitor group is manipulated; few species and lagged responses limit power.

  3. Idea 3

    Are fluctuations of body-size-similar rodents more compensatory?

    Hypothesis. Pairwise correlations between detrended plot-level abundance fluctuations are more negative (compensatory) for species pairs of similar body mass within the granivore guild.

    Dataset. Portal Project long-term rodent monitoring, Chihuahuan Desert portal_rodent_surveys

    Test plan, novelty and risk

    Test plan. On unmanipulated plot-periods, compute monthly plot-level abundances, remove species x census and species x plot means, and correlate residual fluctuations for each species pair; regress pair correlations on |ln mass difference| and same-guild membership with a species-label (Mantel-type) permutation test and year-block bootstrap. Refuted if correlations do not become more negative as size difference shrinks.

    Novelty. Links the compensatory-dynamics literature to the size-similarity hypothesis using within-census spatial covariation.

    Expected outcome. Mostly weak positive or near-zero correlations because shared resource pulses dominate; at most a weak similarity signal.

    Main risk. Covariance is not competition: shared habitat responses and sampling noise drive correlations, and synchrony cannot separate per-capita effects or their direction.

Ranking rationale

Idea 0 answers both parts of the question with per-capita competition coefficients in the form coexistence theory requires, uses the whole 11-species community, removes community-wide environmental forcing with census fixed effects and corrects census error with lagged instruments; its risks are explicit and testable with the planned variants. Idea 1 has experimental leverage but manipulates only one competitor group, and the exclusion plots carry a pre-1988 seed-addition history that confounds the baseline. Idea 2 is simple but cannot distinguish competition from shared responses or recover per-capita effects.

3.Experimentation

Attempt 4Done

Experiment succeeded in 1 round (52.879s).

Reviewer feedback applied: Independent referee (Claude Fable 5.1): major_revision. Required changes: (1) [text] Abstract, 4.2, 4.3 and Discussion: state that the pre-specified H1 rule (upper limit of mean q < 1) holds in the primary model and the OLS, per-capita, granivore and hindfoot variants but fails with t-2 instruments (upper limit 2.145), label the median and fraction-of-pairs criteria as post hoc summaries, and delete or qualify 'The robust pattern here is strong relative self-limitation' and 4.2's 'H1's robustness therefore rests on the t-2 variant ... still support it'. (2) [text] Abstract: replace 'Intraspecific coefficients were positive for all species' with 'point estimates were positive for all 11 species, with intervals excluding zero for seven'; add that the Sargan test rejected joint validity of the t-1 and t-2 instruments; and change 'Ignoring census error inflated intraspecific coefficients 2.24-fold' to '2SLS intraspecific coefficients were 2.24-fold smaller than OLS (median), consistent with census-error bias', since the gap is not uniquely attributable to census error and 2SLS is itself biased if the instruments are invalid. (3) [analysis] Replace the i.i.d. Sargan statistic with an overidentification test that allows dependence: a Hansen J with clusters (years or censuses), or the Sargan J referred to its year-block bootstrap distribution; report it in Table 4 and rewrite the Discussion sentence 'at least one instrument set is invalid' in its light. (4) [analysis] Test delayed density dependence: fit Eq. 1 with E(j,p,t-1) terms added, instrumented by t-2 and t-3 densities, and report a_i, k_sim, k_hier and the q summaries; cite Lima et al. (2008) as motivation. In Methods, state that the t-1 instrument requires both serially independent sampling errors and no direct effect of N(t-1) on growth from t to t+1. (5) [analysis] Add a variant with plot-by-calendar-year (or plot-by-five-year-period) effects, or plot-specific linear trends within phase, to absorb plot-specific habitat change (cite Christensen et al. 2018); report kernel slopes and q summaries in Table 5. (6) [analysis] Report the two arms of the fitted kernel with bootstrap intervals, the slope for larger competitors (k_sim + k_hier) and for smaller competitors (k_hier - k_sim), for the primary model and each variant; then rewrite 4.3 and the Discussion so that the flat larger-competitor arm (0.004) is stated, acknowledge that k_hier's interval includes zero and reverses sign under OLS and t-2, and temper 'larger competitors tended to be stronger' and 'the hierarchy that dominated tree competition'. (7) [analysis] Report a sensitivity of the ln(1+N) specification to the offset (for example ln(0.5+N) and ln(2+N), or captures per census instead of per night): the range of a_i, k_sim and the q summaries; justify the chosen offset in Methods. (8) [analysis] Determine whether the 763 species-less records mark zero-capture plot-nights; if so, use them to define sampled plot-nights and re-run; if not, state that trapped plots with no captures are coded as unsampled, count how many plot-censuses this affects, and discuss the bias toward weaker self-limitation of rare species. Justify the row-selection rule (focal caught on any sample plot at t-1 or t) and report the primary estimates with the rule relaxed to presence at t-1 only. (9) [analysis] Report a variant without the ant-removal plots (11 and 14 throughout; 2 and 22 from 1988), or with ant-removal-by-density interactions, with its interval widths, since half the control plot-years had ants removed throughout. (10) [text] Rename 3.2 'Simulation Protocol' to 'Resampling and inference', drop the disclaimer sentence, split 4.2 into two paragraphs (coefficients; ratios and the size-matched comparison), and remove the section numerals, which Ecology Letters does not use. (11) [text] Cite Sanderson & Windmeijer (2016) for the conditional F statistic and Brown et al. (2004) for the 3/4 exponent; add Ives et al. (2003) and Knape & de Valpine (2012) where interaction estimation from counts with observation error is motivated; add Heske et al. (1994), Ernest & Brown (2001), Lima et al. (2008) and Christensen et al. (2018) where Portal competition dynamics, the C. baileyi colonisation that yields only 608 rows for PB, and the assumption of constant coefficients over 25 years are discussed; and Barabas et al. (2018) for the q_ij framing. All are retrieved sources. (12) [text] Discussion: give the interval for the non-competitive fraction (66.4%, 95% CI 33.6-80.0) and state that mean alpha_ij intervals exclude zero for only C. penicillatus and D. spectabilis, so 'most pairs combine self-limitation with interspecific facilitation' is a point-estimate statement. (13) [analysis] Name in the Table 4 note the two kernel-implied ordered pairs with q >= 1; from Table 3 and Eq. 3 these are D. merriami and D. spectabilis acting on C. baileyi (q about 1.4 and 1.3), to be confirmed from the fitted matrix, and relate them in the Discussion to the colonisation of Portal by C. baileyi (Ernest & Brown 2001). (14) [text] Methods: state the source of the guild assignments (the Portal species table) and that P. eremicus and R. megalotis are omnivores classed as granivores; state that the 305 censuses are defined on all 24 plots; state that 581 identified captures of other species in the sample are not modelled as competitors; state that body masses are medians over the whole database rather than the sample.
Succeeded1 round · exit 0 · 52.879s

Data provenance

Datasetportal_rodent_surveys

Portal Project long-term rodent monitoring, Chihuahuan Desert

Source
R package 'ratdat' (Portal Project Teaching Database), mirrored by the Rdatasets archive; original data from the Portal Project, near Portal, Arizona
Citation
Ernest, M., Brown, J., Valone, T. and White, E.P. (2018) Portal Project Teaching Database. figshare. DOI: 10.6084/m9.figshare.1314459 -- the citation the project's own README instructs users to give, and the source of every dataset in the R package 'ratdat'. Peer-reviewed data papers for the full research database: Ernest, S.K.M., Valone, T.J. and Brown, J.H. (2009) Ecology 90(6), 1708, DOI: 10.1890/08-1222.1; and Ernest, S.K.M. et al. (2016) Ecology 97(4), 1082, DOI: 10.1890/15-2115.1.
Licence
Creative Commons Zero (CC0 1.0) public domain dedication -- the figshare record, the PortalData LICENSE file and the 'ratdat' package DESCRIPTION all state CC0.
File
https://vincentarelbundock.github.io/Rdatasets/csv/ratdat/surveys.csv
Observations
35,549
Variables (10)
record_id, month, day, year, plot_id, species_id, sex, hindfoot_length, weight, date
Units
Each row is one captured individual, not a site-visit summary. hindfoot_length is in millimetres and weight in grams -- the two standard body-size measures. species_id is a two-letter code (for example 'DM' for Dipodomys merriami, the dominant kangaroo rat) across 48 codes, and sex is 'M' or 'F'. plot_id identifies one of 24 fenced plots under different rodent-exclusion treatments. month, day and year give the trapping session, and the loader adds a parsed 'date' column.
Retrieved
24 Oct 2026, 10:37
Checksum
sha256 0e7b92161a10b1fc…

Real, publicly available data — downloaded and checksum-verified before the analysis ran, never simulated.

Result (RESULT_JSON)

{
  "n_records_total": 35549,
  "n_records_missing_species": 763,
  "n_censuses_total": 305,
  "sample_plots": [
    2,
    4,
    8,
    11,
    12,
    14,
    17,
    22
  ],
  "n_sample_records": 14158,
  "n_captures_analysed_species": 13534,
  "n_plot_censuses_sampled": 2139,
  "n_transitions_lag1": 2081,
  "n_transitions_lag2": 2055,
  "n_stacked_rows_main": 16025,
  "n_stacked_rows_lag2": 15823,
  "n_years_bootstrap_blocks": 26,
  "species": [
    "PF",
    "RM",
    "PP",
    "PE",
    "OT",
    "PB",
    "OL",
    "DM",
    "DO",
    "DS",
    "NL"
  ],
  "guild": {
    "PF": "granivore",
    "RM": "granivore",
    "PP": "granivore",
    "PE": "granivore",
    "OT": "insectivore",
    "PB": "granivore",
    "OL": "insectivore",
    "DM": "granivore",
    "DO": "granivore",
    "DS": "granivore",
    "NL": "folivore"
  },
  "captures": {
    "PF": 372,
    "RM": 386,
    "PP": 1083,
    "PE": 268,
    "OT": 927,
    "PB": 789,
    "OL": 405,
    "DM": 6180,
    "DO": 1578,
    "DS": 1048,
    "NL": 498
  },
  "body_mass_g_median": {
    "PF": 8,
    "RM": 10,
    "PP": 17,
    "PE": 22,
    "OT": 24,
    "PB": 31,
    "OL": 32,
    "DM": 44,
    "DO": 50,
    "DS": 124,
    "NL": 164
  },
  "hindfoot_mm_median": {
    "PF": 16,
    "RM": 17,
    "PP": 22,
    "PE": 20,
    "OT": 20,
    "PB": 26,
    "OL": 21,
    "DM": 36,
    "DO": 36,
    "DS": 50,
    "NL": 32
  },
  "mean_N_per_plot_census": {
    "PF": 0.17345,
    "RM": 0.17952,
    "PP": 0.50491,
    "PE": 0.12366,
    "OT": 0.43128,
    "PB": 0.36863,
    "OL": 0.18724,
    "DM": 2.87728,
    "DO": 0.73516,
    "DS": 0.48597,
    "NL": 0.23142
  },
  "intra_a_iv": {
    "PF": [
      0.52337,
      -0.0445,
      1.04202
    ],
    "RM": [
      1.25611,
      0.47169,
      2.14959
    ],
    "PP": [
      0.22481,
      -0.3589,
      0.5283
    ],
    "PE": [
      0.42177,
      0.07898,
      0.88806
    ],
    "OT": [
      0.14296,
      -0.27589,
      0.48121
    ],
    "PB": [
      0.08017,
      0.04652,
      0.15732
    ],
    "OL": [
      0.25682,
      -0.00677,
      0.46473
    ],
    "DM": [
      0.03313,
      0.01166,
      0.05736
    ],
    "DO": [
      0.08177,
      0.04956,
      0.12067
    ],
    "DS": [
      0.03941,
      0.02164,
      0.06416
    ],
    "NL": [
      0.06878,
      0.03652,
      0.12065
    ]
  },
  "intra_a_ols": {
    "PF": 1.21818,
    "RM": 1.28896,
    "PP": 0.48564,
    "PE": 0.62688,
    "OT": 0.67183,
    "PB": 0.20763,
    "OL": 0.52948,
    "DM": 0.10618,
    "DO": 0.16252,
    "DS": 0.09476,
    "NL": 0.15403
  },
  "median_ratio_ols_to_iv_intra": 2.23956,
  "inter_mean_iv": {
    "PF": [
      0.0193,
      -0.01385,
      0.05152
    ],
    "RM": [
      -0.00629,
      -0.03743,
      0.01612
    ],
    "PP": [
      -0.04005,
      -0.1254,
      -0.00757
    ],
    "PE": [
      0.00053,
      -0.03528,
      0.02835
    ],
    "OT": [
      -0.0195,
      -0.05493,
      0.00574
    ],
    "PB": [
      0.02801,
      -0.04415,
      0.07998
    ],
    "OL": [
      -0.01638,
      -0.05527,
      0.01146
    ],
    "DM": [
      -0.00832,
      -0.06636,
      0.02783
    ],
    "DO": [
      -0.01255,
      -0.05921,
      0.01745
    ],
    "DS": [
      -0.03034,
      -0.0644,
      -0.00928
    ],
    "NL": [
      -0.03218,
      -0.08554,
      0.00003
    ]
  },
  "inter_size_matched_same_guild": {
    "PF": [
      0.01709,
      -0.05738,
      0.06886
    ],
    "RM": [
      -0.00668,
      -0.0681,
      0.03563
    ],
    "PP": [
      -0.03419,
      -0.13557,
      0.008
    ],
    "PE": [
      0.01035,
      -0.02394,
      0.03975
    ],
    "OT": [
      0.00008,
      -0.04051,
      0.03497
    ],
    "PB": [
      0.04526,
      -0.01996,
      0.09462
    ],
    "OL": [
      0.00981,
      -0.03055,
      0.0484
    ],
    "DM": [
      0.01925,
      -0.0162,
      0.04517
    ],
    "DO": [
      0.01929,
      -0.00201,
      0.03695
    ],
    "DS": [
      0.03512,
      0.00217,
      0.08533
    ],
    "NL": [
      0.05409,
      0.00984,
      0.11093
    ]
  },
  "intra_minus_size_matched": {
    "PF": [
      0.50628,
      -0.05375,
      0.98994
    ],
    "RM": [
      1.26279,
      0.49191,
      2.17064
    ],
    "PP": [
      0.259,
      -0.24933,
      0.55955
    ],
    "PE": [
      0.41143,
      0.08288,
      0.86459
    ],
    "OT": [
      0.14288,
      -0.27585,
      0.48461
    ],
    "PB": [
      0.03491,
      -0.03135,
      0.13553
    ],
    "OL": [
      0.24702,
      -0.04633,
      0.46648
    ],
    "DM": [
      0.01388,
      -0.00413,
      0.0444
    ],
    "DO": [
      0.06248,
      0.03678,
      0.10331
    ],
    "DS": [
      0.0043,
      -0.04499,
      0.04662
    ],
    "NL": [
      0.01469,
      -0.05656,
      0.09096
    ]
  },
  "q_effect_of_DM_on_DO_rel_DM_on_DM": [
    0.42716,
    -0.40722,
    1.12531
  ],
  "q_effect_of_DO_on_DM_rel_DO_on_DO": [
    0.24192,
    -0.14119,
    0.61289
  ],
  "n_rows_by_focal_species": {
    "PF": 1057,
    "RM": 1252,
    "PP": 1346,
    "PE": 1084,
    "OT": 1887,
    "PB": 608,
    "OL": 1343,
    "DM": 2081,
    "DO": 2059,
    "DS": 1513,
    "NL": 1795
  },
  "abs_dlnm_min_max": [
    0.03175,
    3.02042
  ],
  "n_species_intra_gt_size_matched_ci": 3,
  "kernel_main": {
    "k_sim": [
      -0.01803,
      -0.03496,
      -0.00527
    ],
    "k_hier": [
      0.02217,
      -0.00009,
      0.05884
    ],
    "k_guild": [
      0.01351,
      -0.02075,
      0.05128
    ]
  },
  "kernel_ols": {
    "k_sim": -0.00568,
    "k_hier": -0.00643,
    "k_guild": 0.00627
  },
  "perm_p_k_sim_two_sided": 0.022,
  "perm_null_sd_k_sim": 0.00705,
  "n_permutations": 999,
  "q_mean": [
    0.01614,
    -0.40018,
    0.38947
  ],
  "q_median": [
    -0.0323,
    -0.14448,
    0.04576
  ],
  "turnover_breakdown_fraction": 0.61053,
  "frac_pairs_q_lt_1": [
    0.98182,
    0.88182,
    1
  ],
  "frac_pairs_inter_competitive": [
    0.33636,
    0.2,
    0.66364
  ],
  "q_mean_ols": 0.03646,
  "min_sw_F_main": 416.9,
  "min_sw_F_lag2": 24.1,
  "sargan_lag1_lag2": {
    "J": 53.84,
    "df": 25,
    "p": 0.000698,
    "n_rows": 15823
  },
  "variants": {
    "ols": {
      "k_sim": [
        -0.00568,
        -0.00987,
        -0.00177
      ],
      "k_hier": [
        -0.00643,
        -0.01053,
        -0.00287
      ],
      "k_guild": [
        0.00627,
        -0.0001,
        0.01317
      ],
      "q_mean": [
        0.03646,
        0.02234,
        0.04617
      ],
      "q_median": [
        0.02632,
        0.01571,
        0.03353
      ],
      "frac_q_lt1": [
        1,
        1,
        1
      ]
    },
    "lag2_instruments": {
      "k_sim": [
        0.01128,
        -0.02207,
        0.07879
      ],
      "k_hier": [
        -0.02036,
        -0.05,
        0.0076
      ],
      "k_guild": [
        -0.01688,
        -0.1066,
        0.0461
      ],
      "q_mean": [
        0.12814,
        -3.57234,
        2.14535
      ],
      "q_median": [
        0.06761,
        -0.00324,
        0.13113
      ],
      "frac_q_lt1": [
        0.98182,
        0.76364,
        1
      ]
    },
    "per_capita": {
      "k_sim": [
        -0.01427,
        -0.0367,
        -0.00219
      ],
      "k_hier": [
        0.02424,
        0.00463,
        0.05283
      ],
      "k_guild": [
        0.01651,
        -0.02671,
        0.06777
      ],
      "q_mean": [
        -0.03629,
        -1.14617,
        0.42832
      ],
      "q_median": [
        -0.04961,
        -0.19235,
        0.06709
      ],
      "frac_q_lt1": [
        0.98182,
        0.87273,
        0.99091
      ]
    },
    "granivores_only": {
      "k_sim": [
        -0.02478,
        -0.06409,
        0.02387
      ],
      "k_hier": [
        0.01593,
        -0.03377,
        0.0689
      ],
      "q_mean": [
        0.10178,
        -0.27218,
        0.33094
      ],
      "q_median": [
        0.03573,
        -0.09056,
        0.13328
      ],
      "frac_q_lt1": [
        0.96429,
        0.85536,
        1
      ]
    },
    "hindfoot_size_axis": {
      "k_sim": [
        -0.02678,
        -0.07623,
        0.01671
      ],
      "k_hier": [
        0.03813,
        -0.01759,
        0.13585
      ],
      "k_guild": [
        0.00011,
        -0.03838,
        0.02509
      ],
      "q_mean": [
        0.02116,
        -0.51259,
        0.20532
      ],
      "q_median": [
        -0.03187,
        -0.13903,
        0.06166
      ],
      "frac_q_lt1": [
        0.98182,
        0.90409,
        1
      ]
    }
  },
  "bootstrap_B_main": 999,
  "bootstrap_B_variants": 299,
  "seed": 20261024,
  "h1_intra_exceeds_inter": true,
  "h2_similarity_main": true,
  "h2_sign_robust_all_variants": false,
  "hypothesis_supported": false,
  "revision": {
    "checks": {
      "species_less": {
        "n": 763,
        "with_sex_or_size": 0,
        "plot_dates": 762,
        "plot_dates_without_identified_capture": 754,
        "in_sample": 43,
        "sample_plot_censuses_defined_only_by_them": 41
      },
      "censuses_with_sample_records": 298,
      "eligible_plot_censuses_unrecorded": 5,
      "unrecorded_census_dates": [
        "1980-05-25"
      ],
      "plots_of_24_recorded_in_those_censuses": [
        5
      ],
      "nights_per_plot_census": {
        "1": 2123,
        "2": 16
      },
      "other_identified_captures_in_sample": 581,
      "n_transitions_lag3": 2029
    },
    "overid_lag1_lag2": {
      "sargan_J_iid": 53.84,
      "df": 25,
      "n_rows": 15823,
      "hansen_J_by_cluster": {
        "census": {
          "J": 28.25,
          "p": 0.296,
          "n_clusters": 285
        },
        "plot_year": {
          "J": 24.72,
          "p": 0.478,
          "n_clusters": 186
        },
        "plot_census": {
          "J": 38.18,
          "p": 0.0444,
          "n_clusters": 2055
        }
      },
      "sargan_J_yearblock_bootstrap": {
        "p": 0.215,
        "mcse_p": 0.013,
        "B": 999,
        "n_blocks": 26,
        "J_boot_median": 36.97,
        "J_boot_95th": 83.67
      }
    },
    "arms_h1_rule_mcse": {
      "primary": {
        "arm_larger": [
          0.00414,
          -0.01825,
          0.0363
        ],
        "arm_smaller": [
          0.0402,
          0.01211,
          0.08846
        ],
        "h1_rule_met": true,
        "mcse_q_mean_upper": 0.08418,
        "mcse_k_sim_upper": 0.00057
      },
      "ols": {
        "arm_larger": [
          -0.01211,
          -0.0158,
          -0.00781
        ],
        "arm_smaller": [
          -0.00076,
          -0.00809,
          0.00598
        ],
        "h1_rule_met": true,
        "mcse_q_mean_upper": 0.00094,
        "mcse_k_sim_upper": 0.00059
      },
      "lag2_instruments": {
        "arm_larger": [
          -0.00908,
          -0.058,
          0.07908
        ],
        "arm_smaller": [
          -0.03164,
          -0.08772,
          -0.00151
        ],
        "h1_rule_met": false,
        "mcse_q_mean_upper": 2.10188,
        "mcse_k_sim_upper": 0.01032
      },
      "per_capita": {
        "arm_larger": [
          0.00997,
          -0.01545,
          0.03882
        ],
        "arm_smaller": [
          0.03851,
          0.01481,
          0.08184
        ],
        "h1_rule_met": true,
        "mcse_q_mean_upper": 1.30526,
        "mcse_k_sim_upper": 0.00163
      },
      "granivores_only": {
        "arm_larger": [
          -0.00886,
          -0.02737,
          0.0113
        ],
        "arm_smaller": [
          0.04071,
          -0.05124,
          0.13272
        ],
        "h1_rule_met": true,
        "mcse_q_mean_upper": 0.39728,
        "mcse_k_sim_upper": 0.01102
      },
      "hindfoot_size_axis": {
        "arm_larger": [
          0.01135,
          -0.02161,
          0.07366
        ],
        "arm_smaller": [
          0.06491,
          -0.02782,
          0.21002
        ],
        "h1_rule_met": true,
        "mcse_q_mean_upper": 0.2152,
        "mcse_k_sim_upper": 0.00395
      }
    },
    "pairs_q_ge_1": [
      {
        "focal": "PB",
        "competitor": "DM",
        "q": [
          1.40976,
          -0.44038,
          2.95333
        ],
        "alpha_ij": [
          0.04671,
          -0.0123,
          0.09316
        ],
        "a_j": [
          0.03313,
          0.01166,
          0.05736
        ]
      },
      {
        "focal": "PB",
        "competitor": "DS",
        "q": [
          1.29395,
          -0.03987,
          2.85144
        ],
        "alpha_ij": [
          0.051,
          -0.00192,
          0.09819
        ],
        "a_j": [
          0.03941,
          0.02164,
          0.06416
        ]
      }
    ],
    "new_variants": {
      "plot_by_period": {
        "k_sim": [
          -0.02457,
          -0.05473,
          -0.00829
        ],
        "k_hier": [
          0.03157,
          0.00267,
          0.09999
        ],
        "k_guild": [
          0.01579,
          -0.03537,
          0.0716
        ],
        "q_mean": [
          -0.01988,
          -0.481,
          0.45602
        ],
        "q_median": [
          -0.05685,
          -0.24558,
          0.02548
        ],
        "frac_q_lt1": [
          0.98182,
          0.86364,
          1
        ],
        "arm_larger": [
          0.00701,
          -0.0247,
          0.06728
        ],
        "arm_smaller": [
          0.05614,
          0.02266,
          0.14081
        ],
        "h1_rule_met": true,
        "mcse_q_mean_upper": 0.73516,
        "mcse_k_sim_upper": 0.00625,
        "a_i": {
          "PF": 0.59042,
          "RM": 1.41692,
          "PP": 0.29172,
          "PE": 0.46694,
          "OT": 0.16599,
          "PB": 0.09642,
          "OL": 0.28443,
          "DM": 0.05044,
          "DO": 0.10274,
          "DS": 0.04631,
          "NL": 0.09325
        },
        "median_ratio_a_to_primary": 1.17505,
        "n_rows": 16025
      },
      "plot_trends": {
        "k_sim": [
          -0.02387,
          -0.04948,
          -0.00699
        ],
        "k_hier": [
          0.03095,
          0.00396,
          0.08645
        ],
        "k_guild": [
          0.01442,
          -0.02704,
          0.05292
        ],
        "q_mean": [
          -0.01233,
          -0.38314,
          0.27823
        ],
        "q_median": [
          -0.05644,
          -0.17829,
          0.02354
        ],
        "frac_q_lt1": [
          0.98182,
          0.89091,
          1
        ],
        "arm_larger": [
          0.00708,
          -0.02413,
          0.04814
        ],
        "arm_smaller": [
          0.05481,
          0.02158,
          0.12389
        ],
        "h1_rule_met": true,
        "mcse_q_mean_upper": 0.42055,
        "mcse_k_sim_upper": 0.00792,
        "a_i": {
          "PF": 0.5503,
          "RM": 1.39515,
          "PP": 0.27739,
          "PE": 0.45927,
          "OT": 0.15952,
          "PB": 0.11631,
          "OL": 0.32671,
          "DM": 0.05062,
          "DO": 0.09762,
          "DS": 0.05172,
          "NL": 0.09066
        },
        "median_ratio_a_to_primary": 1.23387,
        "n_rows": 16025
      },
      "offset_0.5": {
        "k_sim": [
          -0.02578,
          -0.04864,
          -0.00408
        ],
        "k_hier": [
          0.03208,
          0.00039,
          0.09642
        ],
        "k_guild": [
          0.02192,
          -0.02505,
          0.08253
        ],
        "q_mean": [
          0.00673,
          -0.44775,
          0.56158
        ],
        "q_median": [
          -0.03717,
          -0.15936,
          0.04928
        ],
        "frac_q_lt1": [
          0.98182,
          0.87273,
          1
        ],
        "arm_larger": [
          0.00629,
          -0.02864,
          0.06097
        ],
        "arm_smaller": [
          0.05786,
          0.01241,
          0.1336
        ],
        "h1_rule_met": true,
        "mcse_q_mean_upper": 0.71409,
        "mcse_k_sim_upper": 0.00166,
        "a_i": {
          "PF": 0.76652,
          "RM": 1.81449,
          "PP": 0.27431,
          "PE": 0.60304,
          "OT": 0.21381,
          "PB": 0.10204,
          "OL": 0.36669,
          "DM": 0.03994,
          "DO": 0.109,
          "DS": 0.05637,
          "NL": 0.10102
        },
        "median_ratio_a_to_primary": 1.42976,
        "n_rows": 16025
      },
      "offset_2": {
        "k_sim": [
          -0.01185,
          -0.02118,
          -0.00293
        ],
        "k_hier": [
          0.01419,
          0.00343,
          0.03469
        ],
        "k_guild": [
          0.00774,
          -0.01458,
          0.03351
        ],
        "q_mean": [
          0.02516,
          -0.43606,
          0.31999
        ],
        "q_median": [
          -0.03078,
          -0.132,
          0.03756
        ],
        "frac_q_lt1": [
          0.98182,
          0.89091,
          1
        ],
        "arm_larger": [
          0.00234,
          -0.01065,
          0.02142
        ],
        "arm_smaller": [
          0.02604,
          0.01048,
          0.05191
        ],
        "h1_rule_met": true,
        "mcse_q_mean_upper": 0.28912,
        "mcse_k_sim_upper": 0.00085,
        "a_i": {
          "PF": 0.32806,
          "RM": 0.80759,
          "PP": 0.17358,
          "PE": 0.27386,
          "OT": 0.0878,
          "PB": 0.05816,
          "OL": 0.16743,
          "DM": 0.02581,
          "DO": 0.05693,
          "DS": 0.02557,
          "NL": 0.04349
        },
        "median_ratio_a_to_primary": 0.64929,
        "n_rows": 16025
      },
      "row_rule_t1_only": {
        "k_sim": [
          -0.01936,
          -0.03716,
          -0.00502
        ],
        "k_hier": [
          0.0261,
          0.00544,
          0.06212
        ],
        "k_guild": [
          0.01871,
          -0.01638,
          0.04891
        ],
        "q_mean": [
          0.00241,
          -0.61846,
          0.32665
        ],
        "q_median": [
          -0.04448,
          -0.13065,
          0.03356
        ],
        "frac_q_lt1": [
          0.98182,
          0.87682,
          1
        ],
        "arm_larger": [
          0.00674,
          -0.01724,
          0.03591
        ],
        "arm_smaller": [
          0.04546,
          0.01847,
          0.09253
        ],
        "h1_rule_met": true,
        "mcse_q_mean_upper": 3.03499,
        "mcse_k_sim_upper": 0.00108,
        "a_i": {
          "PF": [
            0.72833,
            0.44512,
            1.4734
          ],
          "RM": [
            1.33496,
            0.42957,
            2.34189
          ],
          "PP": [
            0.2131,
            -0.60575,
            0.50834
          ],
          "PE": [
            0.51766,
            0.30804,
            0.98657
          ],
          "OT": [
            0.16157,
            -0.15924,
            0.50162
          ],
          "PB": [
            0.08268,
            0.05279,
            0.17194
          ],
          "OL": [
            0.25437,
            -0.04921,
            0.48061
          ],
          "DM": [
            0.0328,
            0.00829,
            0.05765
          ],
          "DO": [
            0.08533,
            0.04795,
            0.12554
          ],
          "DS": [
            0.04058,
            0.02153,
            0.06699
          ],
          "NL": [
            0.06981,
            0.03779,
            0.12458
          ]
        },
        "median_ratio_a_to_primary": 1.03132,
        "n_rows": 14064
      },
      "no_ant_removal_plots": {
        "k_sim": [
          -0.00493,
          -0.04325,
          0.02252
        ],
        "k_hier": [
          0.01069,
          -0.02658,
          0.06257
        ],
        "k_guild": [
          0.00783,
          -0.03976,
          0.07918
        ],
        "q_mean": [
          0.02264,
          -1.2434,
          4.20146
        ],
        "q_median": [
          -0.02778,
          -0.13433,
          0.06972
        ],
        "frac_q_lt1": [
          0.97273,
          0.79909,
          1
        ],
        "arm_larger": [
          0.00576,
          -0.0618,
          0.06711
        ],
        "arm_smaller": [
          0.01562,
          -0.0264,
          0.07693
        ],
        "h1_rule_met": false,
        "mcse_q_mean_upper": 2.39673,
        "mcse_k_sim_upper": 0.02282,
        "a_i": {
          "PF": [
            1.2649,
            -2.43131,
            8.30503
          ],
          "RM": [
            1.18486,
            0.09973,
            2.71039
          ],
          "PP": [
            0.26032,
            -1.3728,
            0.71421
          ],
          "PE": [
            0.27986,
            -0.83483,
            1.1023
          ],
          "OT": [
            0.15909,
            -1.3722,
            0.91835
          ],
          "PB": [
            0.12183,
            -0.00418,
            0.42301
          ],
          "OL": [
            0.30784,
            -0.78016,
            0.63975
          ],
          "DM": [
            0.07,
            0.03671,
            0.12624
          ],
          "DO": [
            0.04192,
            -0.07273,
            0.11875
          ],
          "DS": [
            0.03584,
            -0.00678,
            0.09316
          ],
          "NL": [
            0.07411,
            -0.00029,
            0.17748
          ]
        },
        "median_ratio_a_to_primary": 1.11282,
        "n_rows": 7118
      }
    },
    "delayed_dd": {
      "n_rows": 15625,
      "min_sw_F": 1.1,
      "sw_F_k_sim_k_hier": [
        13.5,
        16.9
      ],
      "n_regressors_sw_F_below_10": 19,
      "n_regressors": 50,
      "all_instrumented_min_max_sw_F": [
        0.01,
        0.42
      ],
      "a_lag0": {
        "PF": [
          0.92264,
          -0.17104,
          2.14029
        ],
        "RM": [
          0.56454,
          -1.02038,
          2.64854
        ],
        "PP": [
          0.92956,
          -0.18391,
          1.42638
        ],
        "PE": [
          1.30106,
          -0.05502,
          2.26202
        ],
        "OT": [
          0.28831,
          0.07199,
          0.69493
        ],
        "PB": [
          -0.00856,
          -0.44956,
          0.29926
        ],
        "OL": [
          0.56662,
          -0.52553,
          1.29846
        ],
        "DM": [
          0.00374,
          -0.04435,
          0.06089
        ],
        "DO": [
          0.08495,
          -0.00033,
          0.28009
        ],
        "DS": [
          0.04162,
          -0.02559,
          0.11252
        ],
        "NL": [
          0.10323,
          0.03769,
          0.25346
        ]
      },
      "a_lag1": {
        "PF": [
          -0.08689,
          -0.37791,
          0.20915
        ],
        "RM": [
          0.14223,
          -0.24575,
          0.51917
        ],
        "PP": [
          -0.20921,
          -0.38213,
          0.20239
        ],
        "PE": [
          -0.23962,
          -0.46039,
          0.17287
        ],
        "OT": [
          -0.01888,
          -0.10572,
          0.04458
        ],
        "PB": [
          0.03763,
          -0.07448,
          0.21442
        ],
        "OL": [
          -0.066,
          -0.21957,
          0.25596
        ],
        "DM": [
          0.01775,
          -0.01369,
          0.04503
        ],
        "DO": [
          0.00259,
          -0.0789,
          0.04612
        ],
        "DS": [
          -0.00212,
          -0.04141,
          0.04445
        ],
        "NL": [
          -0.00988,
          -0.04627,
          0.015
        ]
      },
      "kernel_lag0": {
        "k_sim": [
          0.01667,
          -0.01718,
          0.04052
        ],
        "k_hier": [
          -0.02946,
          -0.05761,
          0.00954
        ],
        "k_guild": [
          -0.02107,
          -0.09515,
          0.0378
        ]
      },
      "kernel_lag1": {
        "k_sim": [
          -0.01052,
          -0.02027,
          0.00241
        ],
        "k_hier": [
          0.01328,
          -0.00213,
          0.02352
        ],
        "k_guild": [
          0.00551,
          -0.0106,
          0.02937
        ]
      },
      "q_mean_lag0": [
        0.23927,
        -2.02878,
        2.51978
      ],
      "q_median_lag0": [
        0.0846,
        0.0017,
        0.18351
      ],
      "frac_q_lt1_lag0": [
        0.92727,
        0.82727,
        0.99091
      ],
      "q_mean_total": [
        0.20526,
        -0.79936,
        1.5326
      ],
      "q_median_total": [
        0.07178,
        0.00246,
        0.17744
      ],
      "frac_q_lt1_total": [
        0.93636,
        0.84045,
        1
      ],
      "arm_larger": [
        -0.01279,
        -0.04366,
        0.02763
      ],
      "arm_smaller": [
        -0.04613,
        -0.08822,
        0.02033
      ],
      "h1_rule_met": false,
      "mcse_q_mean_upper": 5.56516,
      "mcse_k_sim_upper": 0.00437,
      "h1_rule_met_total": false,
      "n_a_lag1_ci_above_0": 0,
      "n_a_lag1_ci_below_0": 0
    },
    "ci_widths": {
      "primary": {
        "k_sim": 0.02969,
        "k_hier": 0.05893,
        "q_mean": 0.78965
      },
      "no_ant_removal_plots": {
        "k_sim": 0.06577,
        "k_hier": 0.08915,
        "q_mean": 5.44486
      }
    },
    "h1_rule_tally": {
      "primary": true,
      "alternatives_met": 9,
      "alternatives": 12,
      "failed": [
        "lag2_instruments",
        "no_ant_removal_plots",
        "delayed_dd"
      ]
    },
    "bootstrap_B_new": 299
  },
  "runtime_s": 52.3
}
Experiment script · experiment_v1.py
"""Intra- vs interspecific competition and body-size similarity, Portal desert rodents (real data, 1977-2002).
Executed design:
- Data: Portal Project Teaching Database captures via dataset.load(); one row per capture.
- Plots: the 8 plots where no rodent was ever excluded, minus food-manipulated periods (Portal treatment
  history): plots 4, 8, 11, 12, 14, 17 from 1 Oct 1977 (after set-up); plots 2 and 22 from 1 Jan 1988 (seed
  additions to Jul 1985, annual-plant removals to Dec 1987). Ant removals on 4, 8, 12, 17 are absorbed by
  plot x phase fixed effects (phase split at 1 Jan 1988, when Portal treatments were reassigned).
- Censuses: dates <=3 days apart, or <=21 days apart when the later date samples only plots not yet trapped,
  form one monthly census. N = captures per trapping night on a plot (mean over nights in the census).
  Species: the rodents with >=250 captures in this sample. Body mass m = median weight (g) of all weighed captures.
- Model, stacked over focal species i (discrete-time Lotka-Volterra / Ricker):
    ln(1+N_i,t+1) - ln(1+N_i,t) = plot-phase_i + census_i - a_i E_i,t - sum_{j!=i} alpha_ij E_j,t + e
    alpha_ij = alpha0_i + k_sim*|ln m_j - ln m_i| + k_hier*(ln m_j - ln m_i) + k_guild*[same feeding guild]
  E_j = N_j * (m_j / m_DM)^0.75 (metabolic D. merriami-equivalents). Rows for focal i: censuses in which i was
  caught on any sample plot at t-1 or t. Fixed effects absorbed exactly (census demeaning + plot-phase dummies).
- Estimator: 2SLS, densities at t-1 instrument densities at t (sampling error); Sanderson-Windmeijer F.
- Inference: block bootstrap over calendar years (B=999, seed 20261024); 999 species-label permutations of the
  size-difference matrix for k_sim (two-sided p).
- Q1 (intra > inter): q_ij = alpha_ij / a_j (j's effect on i relative to its effect on itself); supported if the
  upper 95% limit of mean q < 1. Q2 (similarity): supported if the k_sim 95% interval lies below 0, permutation
  p < 0.05, and k_sim < 0 in every variant.
- Variants (B=299): OLS; instruments at t-2 (+ Sargan test of t-1 and t-2 jointly); per-capita currency;
  granivores only; hindfoot length as the size axis.
Revision 1 (referee changes 3-9, 13). Every number of version 1 reproduces exactly: the new analyses run after it and
draw only from their own seeded streams (np.random.default_rng([SEED, k])).
- Species-less records: checked whether they mark trapped plot-nights without captures (they count as trapping effort).
- Overidentification allowing dependence: Hansen J with census clusters (two-step GMM), and the Sargan J referred to
  its year-block bootstrap distribution with moments recentred at the full-sample estimate (999 replicates).
- Kernel arms per unit signed log-mass difference: larger competitors k_sim + k_hier, smaller ones k_hier - k_sim.
- Ordered pairs with q >= 1 in the fitted matrix; Monte Carlo SE of bootstrap limits (order-statistic interval).
- New variants (B=299 each): plot x five-year-period effects; plot-specific linear trends within phase; ln(c+N) with
  c = 0.5 and 2; focal rows selected on presence at t-1 only; without the ant-removal plots 4, 8, 12, 17.
- Delayed density dependence (B=299): E(t-1) terms added as predetermined regressors and E(t) instrumented by densities
  at t-2 and t-3; instrumenting E(t-1) too is reported as an identification check (conditional F).
- New analyses use the same 2SLS computed from cross-products (fast=True) to fit the 120 s budget.
"""
import json, time, warnings
import numpy as np, pandas as pd
from scipy import stats
from dataset import load

T0 = time.time(); SEED = 20261024; rng = np.random.default_rng(SEED)
warnings.filterwarnings('ignore', message='Mean of empty slice')
df = load()
print(f"loaded {df.shape[0]} records; missing species_id {int(df.species_id.isna().sum())}; missing weight {int(df.weight.isna().sum())}")

# --- censuses ---
on = df.groupby('date').plot_id.apply(set).to_dict(); cmap, cid, cur, last = {}, 0, set(), None
for dt in sorted(on):
    gap = None if last is None else (dt - last).days
    if last is not None and gap > 3 and not (gap <= 21 and not (on[dt] & cur)):
        cid, cur = cid + 1, set()
    cur |= on[dt]; last = dt; cmap[dt] = cid
df['census'] = df.date.map(cmap); C = cid + 1

# --- sample ---
PLOTS = [2, 4, 8, 11, 12, 14, 17, 22]
keep = (df.plot_id.isin([4, 8, 11, 12, 14, 17]) & (df.date >= '1977-10-01')) | (df.plot_id.isin([2, 22]) & (df.date >= '1988-01-01'))
d = df[keep]
caps = d.species_id.value_counts()
GUILD = {'DM': 'granivore', 'DO': 'granivore', 'DS': 'granivore', 'PB': 'granivore', 'PP': 'granivore', 'PF': 'granivore',
         'PE': 'granivore', 'PM': 'granivore', 'RM': 'granivore', 'OT': 'insectivore', 'OL': 'insectivore', 'NL': 'folivore'}
wt = df.groupby('species_id').weight.median(); hfm = df.groupby('species_id').hindfoot_length.median()
SP = sorted([s for s in caps.index if caps[s] >= 250], key=lambda s: wt[s]); S = len(SP)
mass = wt[SP].values; lm = np.log(mass); lhf = np.log(hfm[SP].values)
GS = np.array([[float(GUILD[a] == GUILD[b]) for b in SP] for a in SP])
W_E = (mass / wt['DM']) ** 0.75
print(f"sample records {len(d)}; species {SP}; captures of these {int(caps[SP].sum())} of {int(d.species_id.notna().sum())} identified")

nights = d.groupby(['census', 'plot_id', 'date']).size().index
pc = (d[d.species_id.isin(SP)].groupby(['census', 'plot_id', 'date', 'species_id']).size().unstack(fill_value=0)
      .reindex(columns=SP, fill_value=0).reindex(nights, fill_value=0).groupby(level=[0, 1]).mean())
N = np.full((C, len(PLOTS), S), np.nan)
N[pc.index.get_level_values(0), [PLOTS.index(p) for p in pc.index.get_level_values(1)]] = pc.values
ok = np.isfinite(N[:, :, 0]); site = np.nansum(N, 1) > 0
cc, kk = np.nonzero(ok[1:-1] & ok[:-2] & ok[2:]); cc += 1
cdate = pd.Series(d.groupby('census').date.min()).reindex(range(C))
phase = (cdate.values[cc] >= np.datetime64('1988-01-01')).astype(int); PP_ = kk * 2 + phase
year = pd.DatetimeIndex(cdate.values[cc]).year.values
X = N[cc, kk]; Y = np.log1p(N[cc + 1, kk]) - np.log1p(X); Z1 = N[cc - 1, kk]
ok2 = np.zeros(len(cc), bool); ok2[cc >= 2] = ok[cc[cc >= 2] - 2, kk[cc >= 2]]
Z2 = np.where(ok2[:, None], N[np.maximum(cc - 2, 0), kk], np.nan)
print(f"plot-censuses sampled {int(ok.sum())}; transitions with t-1 {len(cc)}; with t-2 {int(ok2.sum())}")
# revision: t-3 densities, ln(c+N) responses, plot x five-year-period groups, time within phase, plots without ant removal
ok3 = np.zeros(len(cc), bool); ok3[cc >= 3] = ok[cc[cc >= 3] - 3, kk[cc >= 3]]
Z3 = np.where(ok3[:, None], N[np.maximum(cc - 3, 0), kk], np.nan)
YC = {c: np.log(c + N[cc + 1, kk]) - np.log(c + X) for c in (0.5, 2.0)}
GPP = kk * 5 + np.digitize(year, [1983, 1988, 1993, 1998]); TR = (cdate.values[cc] - np.datetime64('1977-10-01')) / np.timedelta64(1, 'D') / 365.25
NOANT = np.nonzero(np.isin(np.array(PLOTS)[kk], [2, 11, 14, 22]))[0]
na = df.species_id.isna(); gpd = df.assign(na=na).groupby(['plot_id', 'date']).agg(n_na=('na', 'sum'), n_id=('species_id', 'count')); gna = gpd[gpd.n_na > 0]
elig = cdate.values[:, None] >= np.array([np.datetime64('1988-01-01' if p in (2, 22) else '1977-10-01') for p in PLOTS])[None, :]
unrec = elig & ~ok & ok.any(1)[:, None]; uc = np.unique(np.nonzero(unrec)[0])
checks = {'species_less': {'n': int(na.sum()), 'with_sex_or_size': int(df.loc[na, ['sex', 'hindfoot_length', 'weight']].notna().any(axis=1).sum()),
                           'plot_dates': int(len(gna)), 'plot_dates_without_identified_capture': int((gna.n_id == 0).sum()), 'in_sample': int(d.species_id.isna().sum()),
                           'sample_plot_censuses_defined_only_by_them': int(ok.sum() - d[d.species_id.notna()].groupby(['census', 'plot_id']).ngroups)},
          'censuses_with_sample_records': int(ok.any(1).sum()), 'eligible_plot_censuses_unrecorded': int(unrec.sum()),
          'unrecorded_census_dates': [str(cdate[c].date()) for c in uc], 'plots_of_24_recorded_in_those_censuses': [int(df[df.census == c].plot_id.nunique()) for c in uc],
          'nights_per_plot_census': {str(k): int(v) for k, v in d.groupby(['census', 'plot_id']).date.nunique().value_counts().items()},
          'other_identified_captures_in_sample': int(d.species_id.notna().sum() - caps[SP].sum()), 'n_transitions_lag3': int((ok2 & ok3).sum())}
print(f"checks {checks}")

def regs(M, i, w, Dm, Gm, spp, cols):
    E = M * w; o = [j for j in spp if j != i]; Eo = E[:, o]
    out = [E[:, i], Eo.sum(1)]
    if 'sim' in cols: out.append(Eo @ Dm[i, o])
    if 'hier' in cols: out.append(Eo @ Gm[i, o])
    if 'guild' in cols: out.append(Eo @ GS[i, o])
    return np.column_stack(out)

def absorb(M, gc, gp, tr=None):  # census demeaning, then plot-group dummies (and plot-group trends if tr) partialled out
    n = np.bincount(gc)[:, None]; dm = lambda A: A - (np.stack([np.bincount(gc, A[:, j]) for j in range(A.shape[1])], 1) / n)[gc]
    M = dm(M); G = gp.max() + 1
    P = (np.eye(G)[gp] - (np.bincount(gc * G + gp, minlength=len(n) * G).reshape(len(n), G) / n)[gc])[:, 1:]  # = dm(dummies)
    if tr is not None: P = np.column_stack([P, dm(np.eye(G)[gp] * tr[:, None])])
    if P.shape[1] == 0: return M
    return M - P @ np.linalg.lstsq(P, M, rcond=None)[0]

def stack(rows, cid, w, Dm, Gm, Zs, spp, cols, Xs=None, Yv=None, gp=None, tr=None, t1=False):
    Xs = [X] if Xs is None else Xs; Yv = Y if Yv is None else Yv; gp = PP_ if gp is None else gp
    F, K = len(spp), len(cols); ys, xs, zs, used, rr = [], [], [], [], []
    for a, i in enumerate(spp):
        sel = site[cc[rows] - 1, i] if t1 else site[cc[rows] - 1, i] | site[cc[rows], i]
        for Z in Zs + Xs[1:]: sel = sel & np.isfinite(Z[rows, 0])
        if sel.sum() < 40: continue
        r = rows[sel]; h = 2 + K
        M = absorb(np.column_stack([Yv[r, i]] + [regs(V[r], i, w, Dm, Gm, spp, cols) for V in Xs + Zs]),
                   pd.factorize(cid[sel])[0], pd.factorize(gp[r])[0], None if tr is None else tr[r])
        blocks = []
        for q in range(len(Xs) + len(Zs)):
            B = M[:, 1 + h * q: 1 + h * (q + 1)]; Q = np.zeros((len(r), 2 * F + K)); Q[:, a] = B[:, 0]; Q[:, F + a] = B[:, 1]; Q[:, 2 * F:] = B[:, 2:]
            blocks.append(Q)
        nx = len(Xs); ys.append(M[:, 0]); xs.append(np.hstack(blocks[:nx])); zs.append(np.hstack(blocks[nx:]) if Zs else blocks[0]); used.append(a); rr.append(r)
    keepc = [k for k in range(2 * F + K) if k >= 2 * F or k % F in used]
    x = np.vstack(xs); z = np.vstack(zs); xk = [q * (2 * F + K) + k for q in range(len(Xs)) for k in keepc]
    zk = [q * (2 * F + K) + k for q in range(max(1, len(Zs))) for k in keepc]
    return np.concatenate(ys), x[:, xk], z[:, zk], xk, np.concatenate(rr)

def fit(rows=None, cid=None, w=W_E, Dm=None, Gm=None, Zs=(Z1,), spp=None, cols=('sim', 'hier', 'guild'), iv=True, diag=False, ret=False, fast=False, **kw):
    rows = np.arange(len(cc)) if rows is None else rows; cid = cc[rows] if cid is None else cid
    spp = list(range(S)) if spp is None else spp; L = lm if Gm is None else Gm
    Dm = np.abs(lm[None, :] - lm[:, None]) if Dm is None else Dm; Gm = L[None, :] - L[:, None]
    y, x, z, keepc, rr = stack(rows, cid, w, Dm, Gm, list(Zs), spp, cols, **kw)
    if ret: return y, x, z, rr
    out = np.full(len(kw.get('Xs') or [X]) * (2 * len(spp) + len(cols)), np.nan)
    if fast:  # revision analyses only: the same 2SLS from cross-products, b = (X'Pz X)^-1 X'Pz y
        zx = z.T @ x; A_ = np.linalg.solve(z.T @ z, zx); out[keepc] = -np.linalg.solve(zx.T @ A_, A_.T @ (z.T @ y)); return out
    xh = z @ np.linalg.lstsq(z, x, rcond=None)[0] if iv else x
    b = np.linalg.lstsq(xh, y, rcond=None)[0]; out[keepc] = -b
    if not diag: return out
    e = y - x @ b; sw = []
    for k in range(x.shape[1]):
        o = [m for m in range(x.shape[1]) if m != k]; u = x[:, k] - x[:, o] @ np.linalg.lstsq(xh[:, o], x[:, k], rcond=None)[0]
        uh = z @ np.linalg.lstsq(z, u, rcond=None)[0]; sw.append((uh @ uh) / (z.shape[1] - x.shape[1] + 1) / (((u - uh) @ (u - uh)) / (len(u) - z.shape[1])))
    J = len(e) * (1 - ((e - z @ np.linalg.lstsq(z, e, rcond=None)[0]) ** 2).sum() / (e ** 2).sum()); fit.sw = sw
    return out, len(y), float(min(sw)), float(J), z.shape[1] - x.shape[1]

def derived(b, spp=None, cols=('sim', 'hier', 'guild'), size=None):
    spp = list(range(S)) if spp is None else spp; F = len(spp); s = lm if size is None else size
    a, a0 = b[:F], b[F:2 * F]; kv = dict(zip(cols, b[2 * F:]))
    A = np.array([[a[p] if i == j else a0[p] + kv.get('sim', 0) * abs(s[j] - s[i]) + kv.get('hier', 0) * (s[j] - s[i]) + kv.get('guild', 0) * GS[i, j]
                   for j in spp] for p, i in enumerate(spp)])
    off = ~np.eye(F, dtype=bool); Q = A / np.diag(A)[None, :]
    return dict(q_mean=np.nanmean(Q[off]), q_median=np.nanmedian(Q[off]), frac_q_lt1=np.nanmean(Q[off] < 1), frac_inter_competitive=np.nanmean(A[off] > 0), Q=Q,
                inter_mean=np.nanmean(np.where(off, A, np.nan), 1), size_matched=a0 + kv.get('guild', 0), A=A)

years = np.unique(year); yrows = {y: np.nonzero(year == y)[0] for y in years}
def boot(B, gen=None, base=None, **kw):  # gen: random stream (default: the original one); base: subset of transitions
    g = rng if gen is None else gen; yr = yrows if base is None else {y: v[np.isin(v, base)] for y, v in yrows.items()}
    out = []
    for _ in range(B):
        pick = g.choice(years, len(years), replace=True)
        out.append(fit(np.concatenate([yr[y] for y in pick]), np.concatenate([cc[yr[y]] + C * m for m, y in enumerate(pick)]), **kw))
    return np.array(out)
ci = lambda v: [float(np.nanpercentile(v, 2.5)), float(np.nanpercentile(v, 97.5))]
r5 = lambda v: None if v is None or not np.isfinite(v) else round(float(v), 5)
tri = lambda est, bs: [r5(est)] + [r5(x) for x in ci(bs)]
def mcse(v, p):  # Monte Carlo SE of a bootstrap percentile: half-width of the order-statistic 95% interval / 1.96
    v = np.sort(np.asarray(v, float)[np.isfinite(v)]); B = len(v); h = 1.96 * np.sqrt(B * p * (1 - p))
    return r5((v[min(int(np.ceil(B * p + h)), B - 1)] - v[max(int(np.floor(B * p - h)), 0)]) / 3.92)
def extras(bv, bb, dbb, F):  # kernel arms (per unit signed log-mass difference), H1 rule, Monte Carlo SE of the deciding limits
    qm = [x['q_mean'] for x in dbb]
    return {'arm_larger': tri(bv[2 * F] + bv[2 * F + 1], bb[:, 2 * F] + bb[:, 2 * F + 1]), 'arm_smaller': tri(bv[2 * F + 1] - bv[2 * F], bb[:, 2 * F + 1] - bb[:, 2 * F]),
            'h1_rule_met': bool(ci(qm)[1] < 1), 'mcse_q_mean_upper': mcse(qm, 0.975), 'mcse_k_sim_upper': mcse(bb[:, 2 * F], 0.975)}

# --- primary fit ---
b, n_main, sw_main, _, _ = fit(diag=True); B_MAIN = 999
bs = boot(B_MAIN); dv = derived(b); dvb = [derived(x) for x in bs]
print(f"main fit: stacked rows {n_main}; min SW F {sw_main:.1f}; bootstrap done {time.time() - T0:.1f}s")
b_ols = fit(iv=False); dv_ols = derived(b_ols)
perm = []
for _ in range(999):
    p = rng.permutation(S); lp = lm[p]; perm.append(fit(Dm=np.abs(lp[None, :] - lp[:, None]))[2 * S])
perm = np.array(perm); k_sim = b[2 * S]; p_perm = float((np.sum(np.abs(perm) >= abs(k_sim)) + 1) / (len(perm) + 1))
print(f"permutation done {time.time() - T0:.1f}s")

# --- variants ---
GR = [i for i in range(S) if GUILD[SP[i]] == 'granivore']
VAR = {'ols': dict(iv=False), 'lag2_instruments': dict(Zs=(Z2,)), 'per_capita': dict(w=np.ones(S)),
       'granivores_only': dict(spp=GR, cols=('sim', 'hier')), 'hindfoot_size_axis': dict(Dm=np.abs(lhf[None, :] - lhf[:, None]), Gm=lhf)}
variants = {}; ext = {'primary': extras(b, bs, dvb, S)}; kn = ['sim', 'hier', 'guild']
for name, kw in VAR.items():
    bv = fit(**kw); bb = boot(299, **kw); F = len(kw.get('spp', range(S))); cols = kw.get('cols', ('sim', 'hier', 'guild'))
    sz = lhf if name.startswith('hindfoot') else None
    dvv = derived(bv, kw.get('spp'), cols, sz); dbb = [derived(x, kw.get('spp'), cols, sz) for x in bb]
    variants[name] = {**{f'k_{c}': tri(bv[2 * F + m], bb[:, 2 * F + m]) for m, c in enumerate(cols)},
                      **{k: tri(dvv[k], [x[k] for x in dbb]) for k in ('q_mean', 'q_median', 'frac_q_lt1')}}
    ext[name] = extras(bv, bb, dbb, F)
_, n2, sw2, _, _ = fit(Zs=(Z2,), diag=True); _, nj, _, J, dfj = fit(Zs=(Z1, Z2), diag=True)
print(f"variants done {time.time() - T0:.1f}s")

# --- revision: overidentification allowing dependence (instruments at t-1 and t-2 jointly) ---
yj, xj, zj, rj = fit(Zs=(Z1, Z2), ret=True)
ej = yj - xj @ np.linalg.lstsq(zj @ np.linalg.lstsq(zj, xj, rcond=None)[0], yj, rcond=None)[0]
ZX0, ZY0 = zj.T @ xj, zj.T @ yj; dfo = zj.shape[1] - xj.shape[1]; hansen = {}
for lab, key in (('census', cc[rj]), ('plot_year', kk[rj] * 10000 + year[rj]), ('plot_census', kk[rj] * 1000 + cc[rj])):
    gcl = pd.factorize(key)[0]; H = np.zeros((gcl.max() + 1, zj.shape[1])); np.add.at(H, gcl, zj * ej[:, None]); Wc = np.linalg.pinv(H.T @ H)
    g2 = ZY0 - ZX0 @ np.linalg.solve(ZX0.T @ Wc @ ZX0, ZX0.T @ Wc @ ZY0); Jh = float(g2 @ Wc @ g2)  # two-step GMM, cluster-robust weight
    hansen[lab] = {'J': round(Jh, 2), 'p': float(f"{stats.chi2.sf(Jh, dfo):.3g}"), 'n_clusters': int(gcl.max() + 1)}
Js = len(ej) * (1 - ((ej - zj @ np.linalg.lstsq(zj, ej, rcond=None)[0]) ** 2).sum() / (ej ** 2).sum())
yu = pd.factorize(year[rj])[0]; U = yu.max() + 1; G0 = zj.T @ ej
ZZ, ZX, ZY, XX, XY, YY, NN = [np.array(v) for v in zip(*[(zj[m].T @ zj[m], zj[m].T @ xj[m], zj[m].T @ yj[m], xj[m].T @ xj[m], xj[m].T @ yj[m], yj[m] @ yj[m], m.sum())
                                                         for m in (yu == u for u in range(U))])]
gJ = np.random.default_rng([SEED, 10]); Jb = []
for _ in range(999):  # recentred moments: E*[Z*'(y* - X* b_hat)] = G0
    wt_ = np.bincount(gJ.integers(0, U, U), minlength=U).astype(float); A_ = np.tensordot(wt_, ZX, 1); W_ = np.linalg.pinv(np.tensordot(wt_, ZZ, 1))
    c_ = np.tensordot(wt_, ZY, 1) - G0; bj_ = np.linalg.solve(A_.T @ W_ @ A_, A_.T @ W_ @ c_); m_ = c_ - A_ @ bj_
    ee = np.tensordot(wt_, YY, 1) - 2 * bj_ @ np.tensordot(wt_, XY, 1) + bj_ @ np.tensordot(wt_, XX, 1) @ bj_
    Jb.append(np.tensordot(wt_, NN, 1) * (m_ @ W_ @ m_) / ee)
pJb = (1 + int(np.sum(np.array(Jb) >= Js))) / 1000
overid = {'sargan_J_iid': round(float(Js), 2), 'df': int(dfo), 'n_rows': int(len(ej)),
          'hansen_J_by_cluster': hansen,
          'sargan_J_yearblock_bootstrap': {'p': round(pJb, 4), 'mcse_p': round(float(np.sqrt(pJb * (1 - pJb) / 999)), 4), 'B': 999, 'n_blocks': int(U),
                                           'J_boot_median': round(float(np.median(Jb)), 2), 'J_boot_95th': round(float(np.percentile(Jb, 95)), 2)}}
print(f"overidentification {overid} {time.time() - T0:.1f}s")

# --- revision: new variants, each on its own stream ---
NEWV = {'plot_by_period': dict(gp=GPP), 'plot_trends': dict(tr=TR), 'offset_0.5': dict(Yv=YC[0.5]), 'offset_2': dict(Yv=YC[2.0]),
        'row_rule_t1_only': dict(t1=True), 'no_ant_removal_plots': dict(rows=NOANT)}
newv = {}
for k_, (name, kw) in enumerate(NEWV.items()):
    bk = {a: v for a, v in kw.items() if a != 'rows'}; bv = fit(fast=True, **kw); bb = boot(299, gen=np.random.default_rng([SEED, k_ + 1]), base=kw.get('rows'), fast=True, **bk)
    dvv = derived(bv); dbb = [derived(x) for x in bb]
    newv[name] = {**{f'k_{c}': tri(bv[2 * S + m], bb[:, 2 * S + m]) for m, c in enumerate(kn)},
                  **{q: tri(dvv[q], [x[q] for x in dbb]) for q in ('q_mean', 'q_median', 'frac_q_lt1')}, **extras(bv, bb, dbb, S),
                  'a_i': {s: tri(bv[p], bb[:, p]) if name in ('row_rule_t1_only', 'no_ant_removal_plots') else r5(bv[p]) for p, s in enumerate(SP)},
                  'median_ratio_a_to_primary': r5(np.median(bv[:S] / b[:S])),
                  'n_rows': int(len(fit(ret=True, **kw)[0]))}
    print(f"{name}: k_sim {newv[name]['k_sim']} q_mean {newv[name]['q_mean']} {time.time() - T0:.1f}s")

# --- revision: delayed density dependence. E(t-1) terms added as predetermined regressors (valid if sampling errors are
# serially independent, the primary model's own assumption); E(t) instrumented by densities at t-2 and t-3. Instrumenting
# E(t-1) as well (by t-2 and t-3 only) is reported as an identification check.
_, _, swA, _, _ = fit(diag=True, Xs=[X, Z1], Zs=(Z2, Z3)); swA_max = float(max(fit.sw))
KD = dict(Xs=[X, Z1], Zs=(Z1, Z2, Z3)); K3 = 2 * S + 3
_, nd, swd, _, _ = fit(diag=True, **KD); swD = np.array(fit.sw); bd = fit(fast=True, **KD); bdb = boot(299, gen=np.random.default_rng([SEED, 9]), fast=True, **KD)
d0, dT = derived(bd[:K3]), derived(bd[:K3] + bd[K3:]); d0b, dTb = [derived(x[:K3]) for x in bdb], [derived(x[:K3] + x[K3:]) for x in bdb]
dd = {'n_rows': int(nd), 'min_sw_F': round(swd, 1), 'sw_F_k_sim_k_hier': [round(float(swD[2 * S]), 1), round(float(swD[2 * S + 1]), 1)],
      'n_regressors_sw_F_below_10': int((swD < 10).sum()), 'n_regressors': int(len(swD)),
      'all_instrumented_min_max_sw_F': [round(swA, 2), round(swA_max, 2)], 'a_lag0': {s: tri(bd[p], bdb[:, p]) for p, s in enumerate(SP)},
      'a_lag1': {s: tri(bd[K3 + p], bdb[:, K3 + p]) for p, s in enumerate(SP)},
      'kernel_lag0': {f'k_{c}': tri(bd[2 * S + m], bdb[:, 2 * S + m]) for m, c in enumerate(kn)},
      'kernel_lag1': {f'k_{c}': tri(bd[K3 + 2 * S + m], bdb[:, K3 + 2 * S + m]) for m, c in enumerate(kn)},
      **{f'{q}_lag0': tri(d0[q], [x[q] for x in d0b]) for q in ('q_mean', 'q_median', 'frac_q_lt1')},
      **{f'{q}_total': tri(dT[q], [x[q] for x in dTb]) for q in ('q_mean', 'q_median', 'frac_q_lt1')},
      **extras(bd[:K3], bdb[:, :K3], d0b, S), 'h1_rule_met_total': bool(ci([x['q_mean'] for x in dTb])[1] < 1),
      'n_a_lag1_ci_above_0': int(sum(ci(bdb[:, K3 + p])[0] > 0 for p in range(S))), 'n_a_lag1_ci_below_0': int(sum(ci(bdb[:, K3 + p])[1] < 0 for p in range(S)))}
print(f"delayed DD: kernel {dd['kernel_lag0']} q_mean lag0 {dd['q_mean_lag0']} total {dd['q_mean_total']} {time.time() - T0:.1f}s")

# --- revision: ordered pairs with q >= 1 in the fitted matrix (focal i, competitor j: alpha_ij / a_j) ---
pairs = [{'focal': SP[i], 'competitor': SP[j], 'q': tri(dv['Q'][i, j], [x['Q'][i, j] for x in dvb]), 'alpha_ij': tri(dv['A'][i, j], [x['A'][i, j] for x in dvb]),
          'a_j': tri(b[j], bs[:, j])} for i in range(S) for j in range(S) if i != j and dv['Q'][i, j] >= 1]
wid = lambda t: r5(t[2] - t[1]); na_ = newv['no_ant_removal_plots']
widths = {'primary': {'k_sim': wid(tri(b[2 * S], bs[:, 2 * S])), 'k_hier': wid(tri(b[2 * S + 1], bs[:, 2 * S + 1])), 'q_mean': wid(tri(dv['q_mean'], [x['q_mean'] for x in dvb]))},
          'no_ant_removal_plots': {'k_sim': wid(na_['k_sim']), 'k_hier': wid(na_['k_hier']), 'q_mean': wid(na_['q_mean'])}}
rule = {**{k: v['h1_rule_met'] for k, v in ext.items() if k != 'primary'}, **{k: v['h1_rule_met'] for k, v in newv.items()},
        'delayed_dd': dd['h1_rule_met'] and dd['h1_rule_met_total']}
h1_rule = {'primary': ext['primary']['h1_rule_met'], 'alternatives_met': int(sum(rule.values())), 'alternatives': len(rule), 'failed': [k for k, v in rule.items() if not v]}

# --- summaries ---
kn = ['sim', 'hier', 'guild']; kmain = {f'k_{c}': tri(b[2 * S + m], bs[:, 2 * S + m]) for m, c in enumerate(kn)}
q_tri = tri(dv['q_mean'], [x['q_mean'] for x in dvb])
h1 = q_tri[2] is not None and q_tri[2] < 1
iDM, iDO = SP.index('DM'), SP.index('DO'); offm = ~np.eye(S, dtype=bool); Dabs = np.abs(lm[None, :] - lm[:, None])
nrow = {s: int(((site[cc - 1, p] | site[cc, p]) & np.isfinite(Z1[:, 0])).sum()) for p, s in enumerate(SP)}
h2_main = kmain['k_sim'][2] < 0 and p_perm < 0.05
h2_robust = all(v['k_sim'][0] < 0 for v in variants.values())
print("species  mass  a_i(IV) [95% CI]            a_i(OLS)  mean inter")
for p, s in enumerate(SP):
    lo, hi = ci(bs[:, p]); print(f"{s:6s} {mass[p]:6.0f} {b[p]:8.4f} [{lo:8.4f},{hi:8.4f}] {b_ols[p]:8.4f} {dv['inter_mean'][p]:8.4f}")
print(f"kernel {kmain}; perm p {p_perm:.3f}; q_mean {q_tri}; H1 {h1} H2 main {h2_main} robust {h2_robust}")
res = {
    'n_records_total': int(len(df)), 'n_records_missing_species': int(df.species_id.isna().sum()), 'n_censuses_total': int(C),
    'sample_plots': PLOTS, 'n_sample_records': int(len(d)), 'n_captures_analysed_species': int(caps[SP].sum()),
    'n_plot_censuses_sampled': int(ok.sum()), 'n_transitions_lag1': int(len(cc)), 'n_transitions_lag2': int(ok2.sum()),
    'n_stacked_rows_main': int(n_main), 'n_stacked_rows_lag2': int(n2), 'n_years_bootstrap_blocks': int(len(years)),
    'species': SP, 'guild': {s: GUILD[s] for s in SP}, 'captures': {s: int(caps[s]) for s in SP},
    'body_mass_g_median': {s: r5(m) for s, m in zip(SP, mass)}, 'hindfoot_mm_median': {s: r5(np.exp(h)) for s, h in zip(SP, lhf)},
    'mean_N_per_plot_census': {s: r5(np.nanmean(N[:, :, p])) for p, s in enumerate(SP)},
    'intra_a_iv': {s: tri(b[p], bs[:, p]) for p, s in enumerate(SP)}, 'intra_a_ols': {s: r5(b_ols[p]) for p, s in enumerate(SP)},
    'median_ratio_ols_to_iv_intra': r5(np.median(b_ols[:S] / b[:S])),
    'inter_mean_iv': {s: tri(dv['inter_mean'][p], [x['inter_mean'][p] for x in dvb]) for p, s in enumerate(SP)},
    'inter_size_matched_same_guild': {s: tri(dv['size_matched'][p], [x['size_matched'][p] for x in dvb]) for p, s in enumerate(SP)},
    'intra_minus_size_matched': {s: tri(b[p] - dv['size_matched'][p], bs[:, p] - np.array([x['size_matched'][p] for x in dvb])) for p, s in enumerate(SP)},
    'q_effect_of_DM_on_DO_rel_DM_on_DM': tri(dv['Q'][iDO, iDM], [x['Q'][iDO, iDM] for x in dvb]), 'q_effect_of_DO_on_DM_rel_DO_on_DO': tri(dv['Q'][iDM, iDO], [x['Q'][iDM, iDO] for x in dvb]),
    'n_rows_by_focal_species': nrow, 'abs_dlnm_min_max': [r5(Dabs[offm].min()), r5(Dabs[offm].max())],
    'n_species_intra_gt_size_matched_ci': int(sum(np.nanpercentile(bs[:, p] - np.array([x['size_matched'][p] for x in dvb]), 2.5) > 0 for p in range(S))),
    'kernel_main': kmain, 'kernel_ols': {f'k_{c}': r5(b_ols[2 * S + m]) for m, c in enumerate(kn)},
    'perm_p_k_sim_two_sided': round(p_perm, 4), 'perm_null_sd_k_sim': r5(perm.std()), 'n_permutations': int(len(perm)),
    'q_mean': q_tri, 'q_median': tri(dv['q_median'], [x['q_median'] for x in dvb]), 'turnover_breakdown_fraction': r5(1 - q_tri[2]), 'frac_pairs_q_lt_1': tri(dv['frac_q_lt1'], [x['frac_q_lt1'] for x in dvb]),
    'frac_pairs_inter_competitive': tri(dv['frac_inter_competitive'], [x['frac_inter_competitive'] for x in dvb]),
    'q_mean_ols': r5(dv_ols['q_mean']), 'min_sw_F_main': round(sw_main, 1), 'min_sw_F_lag2': round(sw2, 1),
    'sargan_lag1_lag2': {'J': round(J, 2), 'df': int(dfj), 'p': float(f"{stats.chi2.sf(J, dfj):.3g}"), 'n_rows': int(nj)},
    'variants': variants, 'bootstrap_B_main': B_MAIN, 'bootstrap_B_variants': 299, 'seed': SEED,
    'h1_intra_exceeds_inter': bool(h1), 'h2_similarity_main': bool(h2_main), 'h2_sign_robust_all_variants': bool(h2_robust),
    'hypothesis_supported': bool(h1 and h2_main and h2_robust),
    'revision': {'checks': checks, 'overid_lag1_lag2': overid, 'arms_h1_rule_mcse': ext, 'pairs_q_ge_1': pairs, 'new_variants': newv,
                 'delayed_dd': dd, 'ci_widths': widths, 'h1_rule_tally': h1_rule, 'bootstrap_B_new': 299},
    'runtime_s': round(time.time() - T0, 1)}
print('RESULT_JSON: ' + json.dumps(res))
Listing. experiment_v1.py

4.Analysis

Attempt 4Done

Analysis complete: experiment succeeded; inconclusive; 61 key numbers.

Reviewer feedback applied: Independent referee (Claude Fable 5.1), round 2: minor_revision. Required changes: (1) [text] Abstract, final sentence: restrict the claim to the primary model and note the reversal, for example 'In the primary model, competitive effects declined toward smaller, not larger, competitors, which limiting similarity does not predict; both arms reversed under second-lag instruments.' (2) [text] Abstract: acknowledge the marginal rejection, for example 'tests robust to dependence within censuses or years did not (one allowing only within-plot-census correlation rejected marginally)', trimming elsewhere to stay within 150 words. (3) [text] Table 5: enter 'not computed' in the blank MC SE cell of the 'Delayed density, summed' row, or compute it. Keep the analysis record consistent with the paper.
InconclusiveExperiment succeeded

Results and analysis

The revised script ran successfully (exit 0) and reproduced every earlier number exactly. H1 meets its pre-specified rule in the primary model: the mean ratio of a species' per-biomass effect on heterospecifics to its effect on itself was 0.016 (95% CI -0.400 to 0.389; Monte Carlo SE of the upper limit 0.084). The rule held in 9 of 12 alternative specifications and failed in the three with the widest intervals: instruments at t-2 (upper limit 2.145), the four plots without ant removal (4.201) and the model with delayed density terms (2.520). Because this mean of ratios is heavy-tailed when bootstrap intraspecific coefficients of rare species approach zero, the Monte Carlo SE of its upper limit was 0.29-5.57 in 10 of the 12 alternatives, so several verdicts are not Monte Carlo-stable; the post hoc median ratio had an upper limit of at most 0.184, and at least 92.7% of pairs had q < 1, in every specification. The i.i.d. Sargan rejection of joint t-1 and t-2 instrument validity (J = 53.84, p = 0.000698) did not survive dependence: Hansen J was 28.25 with census clusters (p = 0.296) and 24.72 with plot-year clusters (p = 0.478), and the year-block bootstrap gave p = 0.215; only clustering confined to species within a plot-census rejected (p = 0.0444). No delayed intraspecific coefficient had an interval excluding zero, but that model was weakly identified. H2 fails its pre-specified rule, as before. In the primary model the fitted kernel is one-sided: per unit biomass, effects rose as competitors approached the focal species' size from below (smaller-competitor arm 0.0402, 95% CI 0.0121 to 0.0885) but were flat among larger competitors (0.0041, -0.0183 to 0.0363), so the negative similarity slope reflects weak effects of smaller competitors, not limiting similarity; both arms reversed sign with t-2 instruments. Plot x period effects, plot trends, both offsets and the t-1 row rule left these conclusions unchanged. The joint hypothesis is not supported: self-limitation holds under the primary identification but not in every alternative, and body-size similarity does not structure competition.

Key numbers

MetricValueInterpretation
records loaded / records with missing species_id35549 / 763Whole teaching database.
species-less records: with sex or size / plot-dates / plot-dates without an identified capture / in sample / sample plot-censuses defined only by them0 / 762 / 754 / 43 / 41They carry no measurement and are almost always the only record of their plot-night: placeholders for trapped plot-nights without captures. The analysis counts them as trapping effort, so these plot-censuses enter with zero density for every species.
eligible sample plot-censuses with no record although another sample plot was trapped; census date; plots of 24 recorded in that census5; 1980-05-25; 5All five fall in one partial census in which most plots have no record, so they were most likely not trapped; trapped-but-empty plots are not coded as unsampled.
trapping censuses (all 24 plots) / censuses with sample-plot records305 / 298Censuses are formed from trapping dates on all 24 plots.
nights per sampled plot-census (1 / 2)2123 / 16N is almost always an integer count of captures, so an offset of one capture is natural; per-census units would change N in only 16 plot-censuses.
sample plots2, 4, 8, 11, 12, 14, 17, 22No rodent ever excluded; 4, 8, 11, 12, 14, 17 from Oct 1977, 2 and 22 from Jan 1988; ants removed on 4, 8, 12, 17.
records in sample / captures of the analysed species / other identified captures not modelled14158 / 13534 / 581Rarer taxa do not enter competitor densities.
plot-censuses sampled / transitions with t-1 / with t-2 / with t-2 and t-32139 / 2081 / 2055 / 2029Plot-census transitions t -> t+1 with the instrument censuses available.
stacked focal-species rows (primary / t-2 / delayed model / t-1 rule / no ant-removal plots)16025 / 15823 / 15625 / 14064 / 7118Estimation samples.
rows by focal speciesPF 1057; RM 1252; PP 1346; PE 1084; OT 1887; PB 608; OL 1343; DM 2081; DO 2059; DS 1513; NL 1795Fewest rows for C. baileyi, which colonised Portal in 1995.
species and guildPF granivore; RM granivore; PP granivore; PE granivore; OT insectivore; PB granivore; OL insectivore; DM granivore; DO granivore; DS granivore; NL folivoreGranivores as flagged in the Portal rodent species table, which includes the omnivorous P. eremicus and R. megalotis; two insectivorous Onychomys; folivorous Neotoma.
median body mass (g; all weighed captures in the database)PF 8.0; RM 10.0; PP 17.0; PE 22.0; OT 24.0; PB 31.0; OL 32.0; DM 44.0; DO 50.0; DS 124.0; NL 164.0Size axis spans 8-164 g.
median hindfoot length (mm)PF 16.0; RM 17.0; PP 22.0; PE 20.0; OT 20.0; PB 26.0; OL 21.0; DM 36.0; DO 36.0; DS 50.0; NL 32.0Alternative size axis.
captures by speciesPF 372; RM 386; PP 1083; PE 268; OT 927; PB 789; OL 405; DM 6180; DO 1578; DS 1048; NL 498D. merriami dominates the sample.
mean captures per plot-censusPF 0.17345; RM 0.17952; PP 0.50491; PE 0.12366; OT 0.43128; PB 0.36863; OL 0.18724; DM 2.87728; DO 0.73516; DS 0.48597; NL 0.23142Plot-scale abundances are small; only D. merriami averages > 1.
range of |ln mass difference| across pairs0.03175 to 3.02042From near-identical pairs to a 20-fold mass difference.
intraspecific coefficient a_i (2SLS, per D. merriami-equivalent per plot)PF 0.52337 [95% CI -0.0445, 1.04202]; RM 1.25611 [95% CI 0.47169, 2.14959]; PP 0.22481 [95% CI -0.3589, 0.5283]; PE 0.42177 [95% CI 0.07898, 0.88806]; OT 0.14296 [95% CI -0.27589, 0.48121]; PB 0.08017 [95% CI 0.04652, 0.15732]; OL 0.25682 [95% CI -0.00677, 0.46473]; DM 0.03313 [95% CI 0.01166, 0.05736]; DO 0.08177 [95% CI 0.04956, 0.12067]; DS 0.03941 [95% CI 0.02164, 0.06416]; NL 0.06878 [95% CI 0.03652, 0.12065]Point estimates positive for all 11 species; intervals exclude 0 for 7 (not PF, PP, OT, OL).
intraspecific coefficient a_i (OLS)PF 1.21818; RM 1.28896; PP 0.48564; PE 0.62688; OT 0.67183; PB 0.20763; OL 0.52948; DM 0.10618; DO 0.16252; DS 0.09476; NL 0.15403OLS exceeds 2SLS for every species.
median ratio OLS / 2SLS intraspecific2.239562SLS coefficients are 2.24-fold smaller than OLS (median), consistent with census-error bias; the gap is not uniquely attributable to census error.
mean kernel-implied interspecific coefficient on focal iPF 0.0193 [95% CI -0.01385, 0.05152]; RM -0.00629 [95% CI -0.03743, 0.01612]; PP -0.04005 [95% CI -0.1254, -0.00757]; PE 0.00053 [95% CI -0.03528, 0.02835]; OT -0.0195 [95% CI -0.05493, 0.00574]; PB 0.02801 [95% CI -0.04415, 0.07998]; OL -0.01638 [95% CI -0.05527, 0.01146]; DM -0.00832 [95% CI -0.06636, 0.02783]; DO -0.01255 [95% CI -0.05921, 0.01745]; DS -0.03034 [95% CI -0.0644, -0.00928]; NL -0.03218 [95% CI -0.08554, 3e-05]Intervals exclude 0 (negative) only for C. penicillatus and D. spectabilis.
interspecific coefficient of a size-matched same-guild heterospecific (alpha0_i + k_guild)PF 0.01709 [95% CI -0.05738, 0.06886]; RM -0.00668 [95% CI -0.0681, 0.03563]; PP -0.03419 [95% CI -0.13557, 0.008]; PE 0.01035 [95% CI -0.02394, 0.03975]; OT 8e-05 [95% CI -0.04051, 0.03497]; PB 0.04526 [95% CI -0.01996, 0.09462]; OL 0.00981 [95% CI -0.03055, 0.0484]; DM 0.01925 [95% CI -0.0162, 0.04517]; DO 0.01929 [95% CI -0.00201, 0.03695]; DS 0.03512 [95% CI 0.00217, 0.08533]; NL 0.05409 [95% CI 0.00984, 0.11093]Hypothetical most-similar competitor implied by the kernel.
intraspecific minus size-matched same-guild coefficientPF 0.50628 [95% CI -0.05375, 0.98994]; RM 1.26279 [95% CI 0.49191, 2.17064]; PP 0.259 [95% CI -0.24933, 0.55955]; PE 0.41143 [95% CI 0.08288, 0.86459]; OT 0.14288 [95% CI -0.27585, 0.48461]; PB 0.03491 [95% CI -0.03135, 0.13553]; OL 0.24702 [95% CI -0.04633, 0.46648]; DM 0.01388 [95% CI -0.00413, 0.0444]; DO 0.06248 [95% CI 0.03678, 0.10331]; DS 0.0043 [95% CI -0.04499, 0.04662]; NL 0.01469 [95% CI -0.05656, 0.09096]Interval excludes 0 for 3 species (RM, PE, DO).
mean q_ij = alpha_ij / a_j (pre-specified H1 statistic)0.01614 [95% CI -0.40018, 0.38947]; Monte Carlo SE of upper limit 0.08418Pre-specified rule (upper limit < 1) met in the primary model; the mean averages competitive and facilitative pairs.
median q_ij (post hoc)-0.0323 [95% CI -0.14448, 0.04576]Post hoc summary, robust to near-zero intraspecific coefficients.
fraction of ordered pairs with q < 1 (post hoc)0.98182 [95% CI 0.88182, 1.0]Kernel-implied; 108 of 110 pairs.
fraction of ordered pairs with competitive (positive) interspecific coefficient0.33636 [95% CI 0.2, 0.66364]Non-competitive fraction 66.4% [95% CI 33.6, 80.0], a point-estimate statement.
ordered pairs with q >= 1 (focal <- competitor: q; alpha_ij; a_j)PB <- DM: q 1.40976 [95% CI -0.44038, 2.95333]; alpha 0.04671 [95% CI -0.0123, 0.09316]; a_j 0.03313 [95% CI 0.01166, 0.05736]; PB <- DS: q 1.29395 [95% CI -0.03987, 2.85144]; alpha 0.051 [95% CI -0.00192, 0.09819]; a_j 0.03941 [95% CI 0.02164, 0.06416]D. merriami and D. spectabilis acting on C. baileyi; both intervals include 1.
D. merriami on D. ordii / D. ordii on D. merriami, relative to effect on self0.42716 [95% CI -0.40722, 1.12531] / 0.24192 [95% CI -0.14119, 0.61289]Kernel-implied; only the second has an upper limit below 1.
mean q under OLS0.03646Same H1 conclusion without the census-error correction.
turnover breakdown fraction (1 - upper limit of mean q)0.61053Sensitivity value for uniform non-competitive turnover across species.
minimum Sanderson-Windmeijer conditional first-stage F (t-1 / t-2 instruments)416.9 / 24.1Instruments strong at t-1 and adequate at t-2.
overidentification, t-1 and t-2 instruments jointly (df 25, n 15823)i.i.d. Sargan J 53.84 (p 0.000698); Hansen J census clusters 28.25 (p 0.296, 285 clusters); plot-year clusters 24.72 (p 0.478, 186); plot-census clusters 38.18 (p 0.0444, 2055); Sargan J vs year-block bootstrap p 0.215 (Monte Carlo SE 0.013, B 999; bootstrap median 36.97, 95th percentile 83.67)The i.i.d. rejection does not survive dependence across plots within censuses or within years; it survives marginally only when correlation is confined to species within a plot-census. No evidence that the instrument sets identify different coefficients, though non-rejection does not establish validity.
delayed density model: identificationE(t-1) also instrumented: conditional F 0.01 to 0.42; E(t-1) predetermined: minimum conditional F 1.1, 19 of 50 regressors below 10, k_sim 13.5, k_hier 16.9Instrumenting the delayed terms leaves the model unidentified; with them predetermined it is identified but weak for several small species.
delayed intraspecific coefficients a_i(t-1)PF -0.08689 [95% CI -0.37791, 0.20915]; RM 0.14223 [95% CI -0.24575, 0.51917]; PP -0.20921 [95% CI -0.38213, 0.20239]; PE -0.23962 [95% CI -0.46039, 0.17287]; OT -0.01888 [95% CI -0.10572, 0.04458]; PB 0.03763 [95% CI -0.07448, 0.21442]; OL -0.066 [95% CI -0.21957, 0.25596]; DM 0.01775 [95% CI -0.01369, 0.04503]; DO 0.00259 [95% CI -0.0789, 0.04612]; DS -0.00212 [95% CI -0.04141, 0.04445]; NL -0.00988 [95% CI -0.04627, 0.015]No interval excludes zero (0 of 11): no detectable delayed intraspecific density dependence, in a weakly identified model.
immediate intraspecific coefficients in the delayed modelPF 0.92264 [95% CI -0.17104, 2.14029]; RM 0.56454 [95% CI -1.02038, 2.64854]; PP 0.92956 [95% CI -0.18391, 1.42638]; PE 1.30106 [95% CI -0.05502, 2.26202]; OT 0.28831 [95% CI 0.07199, 0.69493]; PB -0.00856 [95% CI -0.44956, 0.29926]; OL 0.56662 [95% CI -0.52553, 1.29846]; DM 0.00374 [95% CI -0.04435, 0.06089]; DO 0.08495 [95% CI -0.00033, 0.28009]; DS 0.04162 [95% CI -0.02559, 0.11252]; NL 0.10323 [95% CI 0.03769, 0.25346]Identified from t-2 and t-3 variation; imprecise for small species.
delayed model kernel (immediate; delayed)k_sim 0.01667 [95% CI -0.01718, 0.04052], k_hier -0.02946 [95% CI -0.05761, 0.00954], k_guild -0.02107 [95% CI -0.09515, 0.0378]; delayed k_sim -0.01052 [95% CI -0.02027, 0.00241], k_hier 0.01328 [95% CI -0.00213, 0.02352]Immediate slopes reverse sign, as with t-2 instruments.
delayed model q summaries (immediate; immediate + delayed)mean 0.23927 [95% CI -2.02878, 2.51978] (Monte Carlo SE of upper limit 5.56516), median 0.0846 [95% CI 0.0017, 0.18351], pairs q < 1 0.92727 [95% CI 0.82727, 0.99091]; summed mean 0.20526 [95% CI -0.79936, 1.5326] (Monte Carlo SE of upper limit not computed), median 0.07178 [95% CI 0.00246, 0.17744], pairs q < 1 0.93636 [95% CI 0.84045, 1.0]Pre-specified H1 rule fails by both; post hoc summaries remain far below 1.
variant OLSk_sim -0.00568 [95% CI -0.00987, -0.00177]; k_hier -0.00643 [95% CI -0.01053, -0.00287]; larger-competitor arm -0.01211 [95% CI -0.0158, -0.00781]; smaller-competitor arm -0.00076 [95% CI -0.00809, 0.00598]; mean q 0.03646 [95% CI 0.02234, 0.04617] (Monte Carlo SE of upper limit 0.00094); median q 0.02632 [95% CI 0.01571, 0.03353]; pairs q < 1 1.0 [95% CI 1.0, 1.0]Pre-specified H1 rule met.
variant t-2 instrumentsk_sim 0.01128 [95% CI -0.02207, 0.07879]; k_hier -0.02036 [95% CI -0.05, 0.0076]; larger-competitor arm -0.00908 [95% CI -0.058, 0.07908]; smaller-competitor arm -0.03164 [95% CI -0.08772, -0.00151]; mean q 0.12814 [95% CI -3.57234, 2.14535] (Monte Carlo SE of upper limit 2.10188); median q 0.06761 [95% CI -0.00324, 0.13113]; pairs q < 1 0.98182 [95% CI 0.76364, 1.0]Pre-specified H1 rule not met.
variant per-capita currencyk_sim -0.01427 [95% CI -0.0367, -0.00219]; k_hier 0.02424 [95% CI 0.00463, 0.05283]; larger-competitor arm 0.00997 [95% CI -0.01545, 0.03882]; smaller-competitor arm 0.03851 [95% CI 0.01481, 0.08184]; mean q -0.03629 [95% CI -1.14617, 0.42832] (Monte Carlo SE of upper limit 1.30526); median q -0.04961 [95% CI -0.19235, 0.06709]; pairs q < 1 0.98182 [95% CI 0.87273, 0.99091]Pre-specified H1 rule met.
variant granivores onlyk_sim -0.02478 [95% CI -0.06409, 0.02387]; k_hier 0.01593 [95% CI -0.03377, 0.0689]; larger-competitor arm -0.00886 [95% CI -0.02737, 0.0113]; smaller-competitor arm 0.04071 [95% CI -0.05124, 0.13272]; mean q 0.10178 [95% CI -0.27218, 0.33094] (Monte Carlo SE of upper limit 0.39728); median q 0.03573 [95% CI -0.09056, 0.13328]; pairs q < 1 0.96429 [95% CI 0.85536, 1.0]Pre-specified H1 rule met.
variant hindfoot size axisk_sim -0.02678 [95% CI -0.07623, 0.01671]; k_hier 0.03813 [95% CI -0.01759, 0.13585]; larger-competitor arm 0.01135 [95% CI -0.02161, 0.07366]; smaller-competitor arm 0.06491 [95% CI -0.02782, 0.21002]; mean q 0.02116 [95% CI -0.51259, 0.20532] (Monte Carlo SE of upper limit 0.2152); median q -0.03187 [95% CI -0.13903, 0.06166]; pairs q < 1 0.98182 [95% CI 0.90409, 1.0]Pre-specified H1 rule met.
variant plot x five-year-period effectsk_sim -0.02457 [95% CI -0.05473, -0.00829]; k_hier 0.03157 [95% CI 0.00267, 0.09999]; larger-competitor arm 0.00701 [95% CI -0.0247, 0.06728]; smaller-competitor arm 0.05614 [95% CI 0.02266, 0.14081]; mean q -0.01988 [95% CI -0.481, 0.45602] (Monte Carlo SE of upper limit 0.73516); median q -0.05685 [95% CI -0.24558, 0.02548]; pairs q < 1 0.98182 [95% CI 0.86364, 1.0]; median a_i ratio to primary 1.17505Pre-specified H1 rule met.
variant plot-specific linear trends within phasek_sim -0.02387 [95% CI -0.04948, -0.00699]; k_hier 0.03095 [95% CI 0.00396, 0.08645]; larger-competitor arm 0.00708 [95% CI -0.02413, 0.04814]; smaller-competitor arm 0.05481 [95% CI 0.02158, 0.12389]; mean q -0.01233 [95% CI -0.38314, 0.27823] (Monte Carlo SE of upper limit 0.42055); median q -0.05644 [95% CI -0.17829, 0.02354]; pairs q < 1 0.98182 [95% CI 0.89091, 1.0]; median a_i ratio to primary 1.23387Pre-specified H1 rule met.
variant ln(0.5 + N)k_sim -0.02578 [95% CI -0.04864, -0.00408]; k_hier 0.03208 [95% CI 0.00039, 0.09642]; larger-competitor arm 0.00629 [95% CI -0.02864, 0.06097]; smaller-competitor arm 0.05786 [95% CI 0.01241, 0.1336]; mean q 0.00673 [95% CI -0.44775, 0.56158] (Monte Carlo SE of upper limit 0.71409); median q -0.03717 [95% CI -0.15936, 0.04928]; pairs q < 1 0.98182 [95% CI 0.87273, 1.0]; median a_i ratio to primary 1.42976Pre-specified H1 rule met.
variant ln(2 + N)k_sim -0.01185 [95% CI -0.02118, -0.00293]; k_hier 0.01419 [95% CI 0.00343, 0.03469]; larger-competitor arm 0.00234 [95% CI -0.01065, 0.02142]; smaller-competitor arm 0.02604 [95% CI 0.01048, 0.05191]; mean q 0.02516 [95% CI -0.43606, 0.31999] (Monte Carlo SE of upper limit 0.28912); median q -0.03078 [95% CI -0.132, 0.03756]; pairs q < 1 0.98182 [95% CI 0.89091, 1.0]; median a_i ratio to primary 0.64929Pre-specified H1 rule met.
variant focal rows on presence at t-1 onlyk_sim -0.01936 [95% CI -0.03716, -0.00502]; k_hier 0.0261 [95% CI 0.00544, 0.06212]; larger-competitor arm 0.00674 [95% CI -0.01724, 0.03591]; smaller-competitor arm 0.04546 [95% CI 0.01847, 0.09253]; mean q 0.00241 [95% CI -0.61846, 0.32665] (Monte Carlo SE of upper limit 3.03499); median q -0.04448 [95% CI -0.13065, 0.03356]; pairs q < 1 0.98182 [95% CI 0.87682, 1.0]; median a_i ratio to primary 1.03132Pre-specified H1 rule met.
variant without ant-removal plots 4, 8, 12, 17k_sim -0.00493 [95% CI -0.04325, 0.02252]; k_hier 0.01069 [95% CI -0.02658, 0.06257]; larger-competitor arm 0.00576 [95% CI -0.0618, 0.06711]; smaller-competitor arm 0.01562 [95% CI -0.0264, 0.07693]; mean q 0.02264 [95% CI -1.2434, 4.20146] (Monte Carlo SE of upper limit 2.39673); median q -0.02778 [95% CI -0.13433, 0.06972]; pairs q < 1 0.97273 [95% CI 0.79909, 1.0]; median a_i ratio to primary 1.11282Pre-specified H1 rule not met.
offset sensitivity: a_i range across speciesln(0.5+N): 0.03994 to 1.81449; ln(1+N): 0.03313 to 1.25611; ln(2+N): 0.02557 to 0.80759Coefficients scale with the offset (median ratio 1.43 and 0.65), while the scale-free q summaries barely move.
t-1 row rule: intraspecific coefficientsPF 0.72833 [95% CI 0.44512, 1.4734]; RM 1.33496 [95% CI 0.42957, 2.34189]; PP 0.2131 [95% CI -0.60575, 0.50834]; PE 0.51766 [95% CI 0.30804, 0.98657]; OT 0.16157 [95% CI -0.15924, 0.50162]; PB 0.08268 [95% CI 0.05279, 0.17194]; OL 0.25437 [95% CI -0.04921, 0.48061]; DM 0.0328 [95% CI 0.00829, 0.05765]; DO 0.08533 [95% CI 0.04795, 0.12554]; DS 0.04058 [95% CI 0.02153, 0.06699]; NL 0.06981 [95% CI 0.03779, 0.12458]Close to the primary estimates (median ratio 1.03); intervals exclude 0 for 8 of 11.
no ant-removal plots: intraspecific coefficientsPF 1.2649 [95% CI -2.43131, 8.30503]; RM 1.18486 [95% CI 0.09973, 2.71039]; PP 0.26032 [95% CI -1.3728, 0.71421]; PE 0.27986 [95% CI -0.83483, 1.1023]; OT 0.15909 [95% CI -1.3722, 0.91835]; PB 0.12183 [95% CI -0.00418, 0.42301]; OL 0.30784 [95% CI -0.78016, 0.63975]; DM 0.07 [95% CI 0.03671, 0.12624]; DO 0.04192 [95% CI -0.07273, 0.11875]; DS 0.03584 [95% CI -0.00678, 0.09316]; NL 0.07411 [95% CI -0.00029, 0.17748]Intervals exclude 0 for 2 of 11; far less precise.
interval widths, primary vs without ant-removal plotsk_sim 0.02969 vs 0.06577; k_hier 0.05893 vs 0.08915; mean q 0.78965 vs 5.44486Dropping half the plots widens the k_sim interval 2.2-fold and the mean-q interval 6.9-fold.
pre-specified H1 rule across alternative specificationsmet in 9 of 12; failed: lag2_instruments, no_ant_removal_plots, delayed_ddThe three failures are the specifications with the widest mean-q intervals.
k_sim (similarity slope, per unit |ln mass difference|)-0.01803 [95% CI -0.03496, -0.00527]Negative in the primary model; Monte Carlo SE of upper limit 0.00057.
k_hier (hierarchy slope, per unit ln competitor/focal mass)0.02217 [95% CI -9e-05, 0.05884]Interval touches zero; negative under OLS and t-2 instruments.
k_guild (same feeding guild)0.01351 [95% CI -0.02075, 0.05128]No detectable guild effect.
kernel arms, primary (per unit signed ln mass difference)larger competitors (k_sim + k_hier) 0.00414 [95% CI -0.01825, 0.0363]; smaller competitors (k_hier - k_sim) 0.0402 [95% CI 0.01211, 0.08846]In the primary model, flat for larger competitors; effects fall as competitors become smaller than the focal species: the similarity signal reflects weak effects of smaller competitors. Both arms reverse under t-2 instruments (smaller-competitor arm -0.03164 [95% CI -0.08772, -0.00151]).
permutation test of k_sim (two-sided p; null SD; permutations)p = 0.022; SD 0.00705; 999Observed k_sim is extreme relative to random reassignment of body sizes among species.
kernel OLS (k_sim, k_hier, k_guild)-0.00568, -0.00643, 0.00627Hierarchy sign reversed without the census-error correction.
decision flagsH1 primary True; H2 primary True; H2 sign-robust across the five pre-specified variants False; joint hypothesis_supported FalsePre-specified rules in the script.
bootstrap replicates (primary / variants / new analyses) / seed / runtime (s)999 / 299 / 299 / 20261024 / 52.3Year-block bootstrap over 26 calendar years; new analyses on separate seeded streams.

Caveats

  • The pre-specified H1 statistic, the mean of q_ij, is a mean of ratios: when bootstrap intraspecific coefficients of rare species approach zero it becomes heavy-tailed, so with 299 replicates the Monte Carlo SE of its upper limit was 0.29-5.57 in 10 of 12 alternative specifications and several rule verdicts are not Monte Carlo-stable. The median and the fraction of pairs with q < 1 are post hoc summaries.
  • Identification rests on lagged densities. Dependence-robust overidentification tests did not reject the joint validity of t-1 and t-2 instruments, but non-rejection does not establish validity, and such tests cannot detect a violation shared by both instrument sets. The specifications identified from t-2 or later (t-2 instruments; delayed density model) reverse the kernel slopes and fail the H1 rule.
  • Delayed density dependence was not detected, but the test is weak: instrumenting the delayed terms left the model unidentified, and with them predetermined 19 of 50 coefficients had conditional F below 10.
  • Plots are open 0.25-ha areas: intraspecific coefficients mix competition with movement and turnover; the turnover breakdown value (0.611) assumes uniform turnover across species.
  • With 11 species, absolute size difference and competitor size are nearly collinear; the kernel is linear with pooled slopes, and pair-level coefficients are implied by it, not estimated freely.
  • Two-thirds of kernel-implied interspecific effects are negative, but mean interspecific intervals exclude zero only for C. penicillatus and D. spectabilis; the facilitation reading is a point-estimate statement.
  • Half the sample plot-years had ants removed; dropping those plots leaves estimates too imprecise to confirm or contradict the pooled result (mean-q interval 6.9-fold wider).
  • Single site and the teaching subset ending in 2002; coefficients are assumed constant over 25 years despite community change (C. baileyi colonised in 1995 and contributes 608 rows), although plot x period effects and plot trends leave the conclusions unchanged.
  • Abundance is captures per trapping night without detection modelling; body masses are medians over the whole database, including juveniles; guilds follow the Portal species table, which classes the omnivorous P. eremicus and R. megalotis as granivores; 581 identified captures of rarer taxa are not modelled as competitors.

5.Paper writing

Attempt 8Done

Manuscript drafted: “Evidence for self-limitation, but not limiting similarity, in a desert rodent community” (≈ 5871 words).

Reviewer feedback applied: Independent referee (Claude Fable 5.1), round 2: minor_revision. Required changes: (1) [text] Abstract, final sentence: restrict the claim to the primary model and note the reversal, for example 'In the primary model, competitive effects declined toward smaller, not larger, competitors, which limiting similarity does not predict; both arms reversed under second-lag instruments.' (2) [text] Abstract: acknowledge the marginal rejection, for example 'tests robust to dependence within censuses or years did not (one allowing only within-plot-census correlation rejected marginally)', trimming elsewhere to stay within 150 words. (3) [text] Table 5: enter 'not computed' in the blank MC SE cell of the 'Delayed density, summed' row, or compute it. Keep the analysis record consistent with the paper.

≈ 5871 words · AI-drafted manuscript · not peer-reviewed

Evidence for self-limitation, but not limiting similarity, in a desert rodent community

Abstract

Coexistence requires self-limitation; limiting similarity predicts that similar-sized species compete most. We tested both with 13,534 captures of 11 rodent species on eight Portal plots without rodent removal (1977–2002), fitting a Lotka–Volterra model with a body-size kernel by two-stage least squares (2SLS) with lagged-density instruments against census error. Intraspecific point estimates were positive for all 11 species, with intervals excluding zero for seven; 2SLS coefficients were 2.24-fold smaller than OLS (median), consistent with census-error bias. The pre-specified self-limitation rule held in the primary model (mean relative interspecific effect 0.016 [−0.400, 0.389]) and in 9 of 12 alternatives, failing in the three least precise, including second-lag instruments (upper limit 2.145). An i.i.d. Sargan test rejected joint instrument validity; dependence-robust tests did not, except, marginally, one allowing correlation only within plot-censuses. In the primary model, competitive effects declined toward smaller, not larger, competitors, contrary to limiting similarity; second-lag instruments reversed this.

Keywords: coexistence theory; competition coefficients; limiting similarity; body size; census error; density dependence; desert rodents; Portal Project

Introduction

Stable coexistence requires each species to limit itself more than it limits its competitors (Chesson 2000; Barabás et al. 2018). In terrestrial plants, intraspecific competition averaged four to five times interspecific competition in pairs where both were competitive, and most other pairs combined self-limitation with interspecific facilitation (Adler et al. 2018). For animals, interaction strengths can be estimated from multispecies time series (Ives et al. 2003), and decompositions attribute regulation mainly to intraspecific density dependence (Mutshinda et al. 2009). But census error enters both a density and its subsequent change, inflating apparent intraspecific density dependence (Freckleton et al. 2006), and allowing for it turned strong into weak density dependence across hundreds of population series (Knape & de Valpine 2012). The same error attenuates interspecific effects, so naive estimates are biased toward the hypothesis under test.

Limiting-similarity theory predicts that competition declines with niche distance (MacArthur & Levins 1967). Desert granivorous rodents are the classic case: species of similar mass co-occur less often than expected, a pattern that fades when other guilds are included (Bowers & Brown 1982). Yet species differences can enlarge fitness as well as niche differences, so similar species need not compete most (Mayfield & Levine 2010), and interactions among trees followed trait hierarchy rather than similarity (Kunstler et al. 2012). At Portal, Arizona, removing kangaroo rats (Dipodomys) increased small granivores but not insectivorous Onychomys (Brown & Munger 1985), a response that persisted for 13 years and recurred in a replicate experiment (Heske et al. 1994). Kangaroo-rat dynamics there are nonlinear and driven by summer rainfall and competition (Lima et al. 2008); the community reorganised repeatedly (Christensen et al. 2018), and Chaetodipus baileyi, colonising in 1995, compensated for missing kangaroo rats (Ernest & Brown 2001). Common species fluctuate in synchrony and dispersal declines with size (Brown & Zeng 1989), so shared responses can mask competition and movement can mimic density dependence.

We estimated per-capita competition among the 11 common nocturnal rodents on Portal plots never subjected to rodent removal, comparing plots within monthly censuses and instrumenting densities with lagged values. H1 (self-limitation): a species' per-capita effect on heterospecifics, relative to its effect on itself, is on average below one; H1 is refuted if the 95% interval of the mean ratio reaches one. H2 (limiting similarity): interspecific competition per unit of competitor metabolic biomass declines with the absolute log body-mass difference, holding competitive hierarchy and feeding guild constant; H2 is supported only if the slope's 95% interval lies below zero, a permutation test gives p < 0.05 and the slope stays negative in five pre-specified sensitivity analyses.

Material and Methods

We analysed the surveys table of the Portal Project Teaching Database (Ernest et al. 2018), compiled from long-term monitoring near Portal, Arizona (Ernest et al. 2009, 2016), distributed in the R package 'ratdat' and mirrored by Rdatasets (https://vincentarelbundock.github.io/Rdatasets/csv/ratdat/surveys.csv; CC0 licence; retrieved 24 October 2026; SHA-256 0e7b92161a10b1fc…). Its 35,549 records are individual captures on 24 fenced 0.25-ha plots (Brown & Munger 1985) from 1977 to 2002, giving date, plot, species code, sex, hindfoot length (mm) and body mass (g).

Following the plot treatment history (Ernest et al. 2009, 2016), we kept the eight plots from which no rodent was ever excluded, outside periods of food manipulation: plots 4, 8, 11, 12, 14 and 17 from 1 October 1977 and plots 2 and 22 from 1 January 1988 (earlier seed additions and annual-plant removals). Ants were removed on plots 4, 8, 12 and 17; plot-by-phase effects, with phases split at 1 January 1988 when treatments were reassigned, absorb these level differences. The sample held 14,158 records. Trapping dates no more than 3 days apart, or up to 21 days apart when the later date sampled only untrapped plots, formed one monthly census (305 censuses, defined on all 24 plots). Abundance N was captures per trapping night on a plot; 2,123 of the 2,139 sampled plot-censuses had a single night, so N is almost always a count.

The 763 records without a species code carry no sex or measurement and, on 754 of the 762 plot-dates where they occur, are the only record: they mark trapped plot-nights without captures. A plot-census counted as sampled if it had any record, so the 41 sampled plot-censuses resting on such records alone enter with zero density for every species. Only five eligible plot-censuses had no record while another sample plot was trapped, all in one census (25 May 1980) in which 5 of the 24 plots had records, so trapped plots without captures are coded as zeros, not as unsampled.

We analysed the 11 species with at least 250 captures (13,534 captures; Table 2); 581 identified captures of rarer taxa were not modelled as competitors. Guilds follow the granivore flag of the rodent species table distributed with the Portal database, which, like the removal experiments (Heske et al. 1994), counts the omnivorous Peromyscus eremicus and Reithrodontomys megalotis as granivores; the non-granivores form an insectivorous Onychomys pair and the folivorous Neotoma albigula. Body mass m was the median over all weighed captures of a species in the whole database, not only the sample. The sampled plot-censuses gave 2,081 transitions from census t to t + 1 on plots also sampled at t − 1.

Model Description

For focal species i on plot p, growth between consecutive censuses follows a discrete-time Lotka–Volterra (Ricker) model on ln(1 + N):

ln⁡(1+Ni,p,t+1)−ln⁡(1+Ni,p,t)=μi,p,ϕ+τi,t−aiEi,p,t−∑j≠iαijEj,p,t+εi,p,t(1)\ln(1+N_{i,p,t+1})-\ln(1+N_{i,p,t})=\mu_{i,p,\phi}+\tau_{i,t}-a_i E_{i,p,t}-\sum_{j\neq i}\alpha_{ij}E_{j,p,t}+\varepsilon_{i,p,t} \tag{1}

The offset accommodates zeros; as N is almost always an integer count, one capture, the smallest non-zero count, is the natural offset, and we also report offsets of 0.5 and 2. Densities are metabolic biomass in D. merriami equivalents, since metabolic rate scales with mass to the 3/4 power (Brown et al. 2004),

Ej,p,t=Nj,p,t (mj/mDM)3/4,(2)E_{j,p,t}=N_{j,p,t}\,(m_j/m_{DM})^{3/4}, \tag{2}

and interspecific coefficients follow a trait kernel of the form used for trees by Kunstler et al. (2012),

αij=α0,i+ksim ∣ln⁡mj−ln⁡mi∣+khier (ln⁡mj−ln⁡mi)+kguild Gij,(3)\alpha_{ij}=\alpha_{0,i}+k_{sim}\,|\ln m_j-\ln m_i|+k_{hier}\,(\ln m_j-\ln m_i)+k_{guild}\,G_{ij}, \tag{3}

where G<sub>ij</sub> = 1 within a feeding guild. Eq. 3 has two arms: per unit Δ ln m = ln m<sub>j</sub> − ln m<sub>i</sub>, α<sub>ij</sub> changes by k<sub>sim</sub> + k<sub>hier</sub> among competitors larger than the focal species and by k<sub>hier</sub> − k<sub>sim</sub> among smaller ones. Limiting similarity predicts a negative larger-competitor arm and a positive smaller-competitor arm (k<sub>sim</sub> < 0); a hierarchy of larger, stronger competitors predicts two equal positive arms (k<sub>hier</sub> > 0), identified because α<sub>0,i</sub> is focal-specific. Following Chesson (2000), as reviewed by Barabás et al. (2018), self-limitation relative to effects on others is

qij=αij/aj,(4)q_{ij}=\alpha_{ij}/a_j, \tag{4}

the effect of j on i divided by the effect of j on itself, both per unit of j's biomass; q<sub>ij</sub> < 1 means j limits itself more than it limits i. Plot-by-phase effects μ absorb persistent plot and treatment differences, and species-specific census effects τ absorb forcing common to a census, such as rainfall-driven seed pulses, so coefficients are identified from differences among plots within a census (Table 1).

Table 1. State variables and parameters.

SymbolMeaningUnitsJustification
N<sub>i,p,t</sub>captures of i per trapping nightindividuals plot⁻¹ night⁻¹local abundance on a 0.25-ha plot
E<sub>j,p,t</sub>metabolic biomass (Eq. 2)D. merriami equivalents plot⁻¹energy demand scales with mass<sup>3/4</sup>
m<sub>i</sub>median body massgsize axis of seed partitioning
a<sub>i</sub>intraspecific coefficientper D. merriami equivalent per census intervalself-limitation
α<sub>ij</sub>, α<sub>0,i</sub>interspecific coefficient; focal baselineas a<sub>i</sub>effect of j on i; sensitivity of i
k<sub>sim</sub>, k<sub>hier</sub>similarity and hierarchy slopesas a<sub>i</sub>, per unit of |Δ ln m| and of Δ ln mpredicted < 0 and > 0
k<sub>sim</sub> + k<sub>hier</sub>; k<sub>hier</sub> − k<sub>sim</sub>kernel arms: slope of α<sub>ij</sub> for larger and for smaller competitorsas a<sub>i</sub>, per unit Δ ln msimilarity: < 0 and > 0; hierarchy: both > 0
k<sub>guild</sub>same-guild shiftas a<sub>i</sub>diet overlap
μ<sub>i,p,φ</sub>, τ<sub>i,t</sub>plot-by-phase and census effectslog growthplot and treatment differences; census-wide forcing
q<sub>ij</sub>relative interspecific effect (Eq. 4)dimensionless< 1: j limits itself more than i

Notes: Census interval: time between consecutive monthly censuses. Δ ln m = ln m<sub>j</sub> − ln m<sub>i</sub>. Guilds as in Table 2.

Resampling and inference

Eq. 1 is a one-step model conditioned on observed densities from October 1977 to December 2002. Intervals are the 2.5th and 97.5th percentiles of block-bootstrap replicates that resample the 26 calendar years with replacement, keeping a year's plots and censuses together, with distinct census effects for repeated years; a focal species with fewer than 40 rows is omitted from that replicate. The primary model used 999 replicates and each sensitivity analysis 299. For decisive limits we report the Monte Carlo standard error (MC SE: half-width of the order-statistic 95% interval divided by 1.96), and for bootstrap p-values √(p(1 − p)/B). The similarity slope was also compared with 999 permutations of body mass among species in the similarity term. All resampling was seeded (20261024); the analyses added in revision drew from separate streams.

Analysis

Equations 1–3 were stacked over the 11 focal species (16,025 rows), with focal-specific a<sub>i</sub> and α<sub>0,i</sub>, shared k terms and fixed effects absorbed exactly. Rows were limited to censuses in which the focal species was caught on some sample plot at t − 1 or t, because while a species was absent from every plot, as C. baileyi was before 1995, its rows carry no information on its own density dependence; as presence at t enters the response, we also selected on presence at t − 1 alone. Sampling error biases ordinary least squares (OLS) toward stronger intraspecific and weaker interspecific effects (Freckleton et al. 2006), so we used two-stage least squares (2SLS), instrumenting each density term at t by the same term at t − 1. This is valid only if sampling errors are independent between censuses and N(t − 1) has no direct effect on growth from t to t + 1. We report the minimum conditional first-stage F over the 25 regressors (Sanderson & Windmeijer 2016).

The pre-specified H1 statistic is the mean of q<sub>ij</sub> over the 110 ordered pairs, H1 being supported if its upper 95% limit is below one. Because a mean of ratios becomes unstable when a bootstrap a<sub>j</sub> approaches zero, we added two post hoc summaries, the median q<sub>ij</sub> and the fraction of pairs with q<sub>ij</sub> < 1, and report the fraction with competitive α<sub>ij</sub> > 0. Pair-level quantities are implied by the fitted kernel, not estimated freely. If the same fraction f of every a<sub>j</sub> were non-competitive turnover on these open plots, every q<sub>ij</sub> would scale by 1/(1 − f), so H1 survives while f is below one minus the upper limit of mean q (the turnover breakdown fraction). We also compare a<sub>i</sub> with the implied coefficient of a hypothetical same-guild heterospecific of identical mass (α<sub>0,i</sub> + k<sub>guild</sub>).

We tested the t − 1 and t − 2 instruments jointly with the i.i.d. Sargan statistic; with Hansen's J from two-step generalized method of moments, its weights robust to correlation within censuses, plot-years or plot-censuses; and with the Sargan statistic referred to its year-block bootstrap distribution, moments recentred at the full-sample estimate (999 replicates). Gestation and recruitment could give density a delayed effect at monthly resolution, and Portal kangaroo-rat dynamics are nonlinear and rainfall-driven (Lima et al. 2008), so we also added the E(t − 1) terms as predetermined regressors, valid if sampling errors are serially independent, and instrumented E(t) by densities at t − 2 and t − 3.

Five sensitivity analyses were pre-specified: OLS; t − 2 instruments (2,055 transitions); per-capita currency (E = N); granivores only (guild term dropped); and log median hindfoot length as the size axis. Six were added in revision: plot-by-five-year-period effects (1977–82, 1983–87, 1988–92, 1993–97, 1998–2002) and plot-specific linear trends within phase, because vegetation diverged among Portal plots within a decade (Heske et al. 1994) and the community reorganised repeatedly (Christensen et al. 2018); offsets of 0.5 and 2; rows selected on presence at t − 1; and the four plots without ant removal (11 and 14 throughout, 2 and 22 from 1988).

Results

Sample

Body mass spanned 8–164 g; D. merriami dominated (6,180 captures) and was the only species averaging more than one capture per trapping night per plot-census (2.877), while C. baileyi contributed the fewest rows (608; Table 2).

Table 2. Species analysed, ordered by body mass.

CodeSpeciesGuildMass (g)Hindfoot (mm)CapturesMean NRows
PFPerognathus flavusgranivore8163720.1731,057
RMReithrodontomys megalotisgranivore10173860.1801,252
PPChaetodipus penicillatusgranivore17221,0830.5051,346
PEPeromyscus eremicusgranivore22202680.1241,084
OTOnychomys torridusinsectivore24209270.4311,887
PBChaetodipus baileyigranivore31267890.369608
OLOnychomys leucogasterinsectivore32214050.1871,343
DMDipodomys merriamigranivore44366,1802.8772,081
DODipodomys ordiigranivore50361,5780.7352,059
DSDipodomys spectabilisgranivore124501,0480.4861,513
NLNeotoma albigulafolivore164324980.2311,795

Notes: Guild: granivore flag of the Portal rodent species table; non-granivores split into insectivores (Onychomys) and a folivore. Mass and hindfoot: medians over all weighed or measured captures in the database. Captures: sample plot-periods. Mean N: captures per trapping night per sampled plot-census. Rows: focal-species rows in the stacked estimation (sum 16,025).

Self-limitation (H1)

Intraspecific point estimates were positive for all 11 species, with intervals excluding zero for seven; per unit of biomass they were largest for small species (R. megalotis 1.256) and smallest for kangaroo rats (D. merriami 0.033; Table 3). OLS intraspecific coefficients exceeded the 2SLS estimates for every species (median ratio 2.24), consistent with census-error bias, although 2SLS would itself be biased if the instruments were invalid; instruments were strong (minimum conditional F = 416.9). Kernel-implied interspecific coefficients were far smaller and mostly negative: only 33.6% (95% CI 20.0–66.4%) of ordered pairs were competitive (Table 4).

The mean ratio of a species' effect on heterospecifics to its effect on itself was 0.016 [−0.400, 0.389], so H1 meets its pre-specified rule (MC SE of the upper limit 0.084); this mean averages competitive and facilitative pairs. By the post hoc summaries, the median ratio was −0.032 [−0.144, 0.046] and 98.2% (88.2–100.0%) of kernel-implied pairs had q < 1; the exceptions were D. merriami and D. spectabilis acting on C. baileyi (q = 1.410 [−0.440, 2.953] and 1.294 [−0.040, 2.851]). D. ordii limited itself more than it limited the similar-sized D. merriami (0.242 [−0.141, 0.613]); the reverse was imprecise (0.427 [−0.407, 1.125]). Only for R. megalotis, P. eremicus and D. ordii did the intraspecific coefficient clearly exceed that of a hypothetical size-matched same-guild heterospecific; for the others this partly reflects imprecision, and the comparison is hypothetical for N. albigula, which has no guild partner. Under uniform turnover, H1 would fail only above 61.1% turnover.

Table 3. H1: intraspecific and implied interspecific coefficients by focal species (per D. merriami equivalent per plot per census interval).

Codea<sub>i</sub> 2SLS [95% CI]a<sub>i</sub> OLSMean α<sub>ij</sub> [95% CI]Size-matched α* [95% CI]a<sub>i</sub> − α* [95% CI]
PF0.523 [−0.045, 1.042]1.2180.019 [−0.014, 0.052]0.017 [−0.057, 0.069]0.506 [−0.054, 0.990]
RM1.256 [0.472, 2.150]1.289−0.006 [−0.037, 0.016]−0.007 [−0.068, 0.036]1.263 [0.492, 2.171]
PP0.225 [−0.359, 0.528]0.486−0.040 [−0.125, −0.008]−0.034 [−0.136, 0.008]0.259 [−0.249, 0.560]
PE0.422 [0.079, 0.888]0.6270.001 [−0.035, 0.028]0.010 [−0.024, 0.040]0.411 [0.083, 0.865]
OT0.143 [−0.276, 0.481]0.672−0.020 [−0.055, 0.006]0.000 [−0.041, 0.035]0.143 [−0.276, 0.485]
PB0.080 [0.047, 0.157]0.2080.028 [−0.044, 0.080]0.045 [−0.020, 0.095]0.035 [−0.031, 0.136]
OL0.257 [−0.007, 0.465]0.529−0.016 [−0.055, 0.011]0.010 [−0.031, 0.048]0.247 [−0.046, 0.466]
DM0.033 [0.012, 0.057]0.106−0.008 [−0.066, 0.028]0.019 [−0.016, 0.045]0.014 [−0.004, 0.044]
DO0.082 [0.050, 0.121]0.163−0.013 [−0.059, 0.017]0.019 [−0.002, 0.037]0.062 [0.037, 0.103]
DS0.039 [0.022, 0.064]0.095−0.030 [−0.064, −0.009]0.035 [0.002, 0.085]0.004 [−0.045, 0.047]
NL0.069 [0.037, 0.121]0.154−0.032 [−0.086, 0.000]0.054 [0.010, 0.111]0.015 [−0.057, 0.091]

Notes: a<sub>i</sub>: Eq. 1. Mean α<sub>ij</sub>: mean of the kernel-implied coefficients (Eq. 3) of the 10 heterospecifics on species i. α* = α<sub>0,i</sub> + k<sub>guild</sub>, a hypothetical same-guild heterospecific of identical mass. Intervals: 999 year-block bootstrap replicates (OLS without intervals).

The pre-specified rule held in 9 of 12 alternative specifications (Table 5): OLS (mean q = 0.036), per-capita currency, granivores only, the hindfoot size axis, plot-by-period effects, plot trends, both offsets and rows selected on presence at t − 1. It failed in the three with the widest intervals: t − 2 instruments (upper limit 2.145), the plots without ant removal (4.201) and the delayed-density model (2.520 for immediate, 1.533 for summed coefficients). With 299 replicates the upper limit of this mean of ratios was itself uncertain (MC SE 0.29–5.57 in 10 of the 12 alternatives), so several verdicts may not be stable. The post hoc summaries were stable: the upper limit of the median ratio never exceeded 0.184, and at least 92.7% of pairs had q < 1 in every specification. Intraspecific coefficients scaled with the offset (0.040–1.814 for 0.5 and 0.026–0.808 for 2, against 0.033–1.256) while the ratios barely moved, and selection on presence at t − 1 changed them little (median ratio 1.03; Table S2).

The overidentification evidence against the instruments did not survive allowing for dependence across plots or censuses (Table 4). The i.i.d. Sargan test rejected the joint validity of t − 1 and t − 2 instruments (J = 53.84, df = 25, p = 0.000698), but Hansen's J did not reject with weights robust to correlation within censuses (28.25, p = 0.296) or plot-years (24.72, p = 0.478), nor did the Sargan statistic against its year-block bootstrap distribution (p = 0.215, MC SE 0.013); weights robust only to correlation among species within a plot-census gave J = 38.18 (p = 0.0444). In the delayed-density model no delayed intraspecific coefficient had an interval excluding zero (Table S1), but identification was weak (minimum conditional F = 1.1; 19 of 50 regressors below 10), and with E(t − 1) also instrumented the model was unidentified (conditional F 0.01–0.42).

Table 4. Community-level results and instrument diagnostics of the primary model (2SLS, t − 1 instruments).

QuantityEstimate [95% CI]
H1: mean q<sub>ij</sub>, 110 ordered pairs (pre-specified)0.016 [−0.400, 0.389]; MC SE of upper limit 0.084
Median q<sub>ij</sub> (post hoc)−0.032 [−0.144, 0.046]
Kernel-implied pairs with q<sub>ij</sub> < 1 (post hoc)98.2% [88.2, 100.0]
Kernel-implied pairs with competitive α<sub>ij</sub> > 033.6% [20.0, 66.4]
D. merriami on C. baileyi, relative to D. merriami on itself1.410 [−0.440, 2.953]
D. spectabilis on C. baileyi, relative to D. spectabilis on itself1.294 [−0.040, 2.851]
D. merriami on D. ordii, relative to D. merriami on itself0.427 [−0.407, 1.125]
D. ordii on D. merriami, relative to D. ordii on itself0.242 [−0.141, 0.613]
Turnover breakdown fraction0.611
Mean q<sub>ij</sub> under OLS; median ratio of OLS to 2SLS a<sub>i</sub>0.036; 2.24
H2: k<sub>sim</sub>, per unit |Δ ln m|−0.0180 [−0.0350, −0.0053]; MC SE of upper limit 0.0006
H2: permutation test of k<sub>sim</sub> (two-sided p; null SD)p = 0.022; 0.0071
k<sub>hier</sub>, per unit Δ ln m0.0222 [−0.0001, 0.0588]
k<sub>guild</sub>0.0135 [−0.0208, 0.0513]
Kernel arm, larger competitors (k<sub>sim</sub> + k<sub>hier</sub>)0.0041 [−0.0183, 0.0363]
Kernel arm, smaller competitors (k<sub>hier</sub> − k<sub>sim</sub>)0.0402 [0.0121, 0.0885]
Minimum conditional first-stage F (t − 1; t − 2 instruments)416.9; 24.1
Overidentification, t − 1 and t − 2 jointly: i.i.d. Sargan JJ = 53.84, p = 0.000698
Hansen J, census clusters (285)J = 28.25, p = 0.296
Hansen J, plot-year clusters (186)J = 24.72, p = 0.478
Hansen J, plot-census clusters (2,055)J = 38.18, p = 0.0444
Sargan J against its year-block bootstrap distributionp = 0.215 (MC SE 0.013); bootstrap median 36.97, 95th percentile 83.67

Notes: q<sub>ij</sub> = α<sub>ij</sub>/a<sub>j</sub> (Eq. 4), implied by the fitted kernel (Eq. 3); the two ordered pairs with q<sub>ij</sub> ≥ 1 in the fitted matrix are D. merriami and D. spectabilis acting on C. baileyi. Breakdown fraction: 1 minus the upper limit of mean q. Slopes and arms: units of α per unit log-mass difference (pairwise range 0.032–3.020). Intervals: 999 year-block bootstrap replicates; permutation: 999 reassignments of mass. MC SE: Monte Carlo standard error. Overidentification: df = 25, n = 15,823; Hansen J from two-step GMM with cluster-robust weights; bootstrap with recentred moments, 999 replicates over 26 years.

Table 5. H1 under the primary model and 12 alternative specifications.

SpecificationMean q [95% CI]MC SERuleMedian q [95% CI]Pairs q < 1
Primary0.016 [−0.400, 0.389]0.084met−0.032 [−0.144, 0.046]98.2% [88.2, 100.0]
Pre-specified
OLS0.036 [0.022, 0.046]0.001met0.026 [0.016, 0.034]100.0% [100.0, 100.0]
t − 2 instruments0.128 [−3.572, 2.145]2.102failed0.068 [−0.003, 0.131]98.2% [76.4, 100.0]
Per-capita currency−0.036 [−1.146, 0.428]1.305met−0.050 [−0.192, 0.067]98.2% [87.3, 99.1]
Granivores only0.102 [−0.272, 0.331]0.397met0.036 [−0.091, 0.133]96.4% [85.5, 100.0]
Hindfoot size axis0.021 [−0.513, 0.205]0.215met−0.032 [−0.139, 0.062]98.2% [90.4, 100.0]
Added in revision
Plot × period effects−0.020 [−0.481, 0.456]0.735met−0.057 [−0.246, 0.025]98.2% [86.4, 100.0]
Plot trends within phase−0.012 [−0.383, 0.278]0.421met−0.056 [−0.178, 0.024]98.2% [89.1, 100.0]
ln(0.5 + N)0.007 [−0.448, 0.562]0.714met−0.037 [−0.159, 0.049]98.2% [87.3, 100.0]
ln(2 + N)0.025 [−0.436, 0.320]0.289met−0.031 [−0.132, 0.038]98.2% [89.1, 100.0]
Rows on presence at t − 10.002 [−0.618, 0.327]3.035met−0.044 [−0.131, 0.034]98.2% [87.7, 100.0]
Without ant-removal plots0.023 [−1.243, 4.201]2.397failed−0.028 [−0.134, 0.070]97.3% [79.9, 100.0]
Delayed density, immediate0.239 [−2.029, 2.520]5.565failed0.085 [0.002, 0.184]92.7% [82.7, 99.1]
Delayed density, summed0.205 [−0.799, 1.533]not computedfailed0.072 [0.002, 0.177]93.6% [84.0, 100.0]

Notes: Rule: pre-specified H1 criterion, upper 95% limit of mean q below 1; MC SE: Monte Carlo standard error of that limit. Median q and pairs with q < 1 are post hoc. Intervals: 999 (primary) or 299 year-block bootstrap replicates. Delayed density: E(t − 1) added as predetermined regressors, E(t) instrumented at t − 2 and t − 3; summed adds the delayed to the immediate coefficients (Table S1). Rows: 16,025 unless stated; 15,823 (t − 2); 14,064 (presence at t − 1); 7,118 (without ant-removal plots); 15,625 (delayed density). Mean-q interval width 0.790 (primary) and 5.445 (without ant-removal plots).

Size structure of interspecific competition (H2)

In the primary model, interspecific competition per unit of competitor biomass declined with absolute log mass difference (k<sub>sim</sub> = −0.0180 [−0.0350, −0.0053]; permutation p = 0.022), but the arms show where (Table 6). Among competitors larger than the focal species the kernel was flat (0.0041 [−0.0183, 0.0363] per unit Δ ln m), whereas among smaller ones effects fell steeply as competitors became smaller (0.0402 [0.0121, 0.0885]). The hierarchy slope was positive but its interval included zero (k<sub>hier</sub> = 0.0222 [−0.0001, 0.0588]), and it reversed sign under OLS (−0.0064 [−0.0105, −0.0029]) and with t − 2 instruments (−0.0204 [−0.0500, 0.0076]); guild had no detectable effect. A flat larger-competitor arm and a positive smaller-competitor arm excluding zero recurred with per-capita currency, plot-by-period effects, plot trends, both offsets and selection on presence at t − 1. Under OLS the shape inverted, effects falling among larger competitors (−0.0121 [−0.0158, −0.0078]); with t − 2 instruments both arms reversed (smaller-competitor arm −0.0316 [−0.0877, −0.0015]) and the similarity slope became positive (0.0113 [−0.0221, 0.0788]), as in the delayed-density model (0.0167 [−0.0172, 0.0405]). Among the five pre-specified sensitivity analyses the similarity slope was negative in four, with intervals excluding zero only under OLS and per-capita currency, so H2 fails its pre-specified rule.

Table 6. H2: similarity and hierarchy slopes and the two arms of the fitted kernel under each specification.

Specificationk<sub>sim</sub>k<sub>hier</sub>Larger-competitor armSmaller-competitor arm
Primary−0.0180 [−0.0350, −0.0053]0.0222 [−0.0001, 0.0588]0.0041 [−0.0183, 0.0363]0.0402 [0.0121, 0.0885]
Pre-specified
OLS−0.0057 [−0.0099, −0.0018]−0.0064 [−0.0105, −0.0029]−0.0121 [−0.0158, −0.0078]−0.0008 [−0.0081, 0.0060]
t − 2 instruments0.0113 [−0.0221, 0.0788]−0.0204 [−0.0500, 0.0076]−0.0091 [−0.0580, 0.0791]−0.0316 [−0.0877, −0.0015]
Per-capita currency−0.0143 [−0.0367, −0.0022]0.0242 [0.0046, 0.0528]0.0100 [−0.0155, 0.0388]0.0385 [0.0148, 0.0818]
Granivores only−0.0248 [−0.0641, 0.0239]0.0159 [−0.0338, 0.0689]−0.0089 [−0.0274, 0.0113]0.0407 [−0.0512, 0.1327]
Hindfoot size axis−0.0268 [−0.0762, 0.0167]0.0381 [−0.0176, 0.1359]0.0114 [−0.0216, 0.0737]0.0649 [−0.0278, 0.2100]
Added in revision
Plot × period effects−0.0246 [−0.0547, −0.0083]0.0316 [0.0027, 0.1000]0.0070 [−0.0247, 0.0673]0.0561 [0.0227, 0.1408]
Plot trends within phase−0.0239 [−0.0495, −0.0070]0.0310 [0.0040, 0.0865]0.0071 [−0.0241, 0.0481]0.0548 [0.0216, 0.1239]
ln(0.5 + N)−0.0258 [−0.0486, −0.0041]0.0321 [0.0004, 0.0964]0.0063 [−0.0286, 0.0610]0.0579 [0.0124, 0.1336]
ln(2 + N)−0.0119 [−0.0212, −0.0029]0.0142 [0.0034, 0.0347]0.0023 [−0.0107, 0.0214]0.0260 [0.0105, 0.0519]
Rows on presence at t − 1−0.0194 [−0.0372, −0.0050]0.0261 [0.0054, 0.0621]0.0067 [−0.0172, 0.0359]0.0455 [0.0185, 0.0925]
Without ant-removal plots−0.0049 [−0.0433, 0.0225]0.0107 [−0.0266, 0.0626]0.0058 [−0.0618, 0.0671]0.0156 [−0.0264, 0.0769]
Delayed density, immediate0.0167 [−0.0172, 0.0405]−0.0295 [−0.0576, 0.0095]−0.0128 [−0.0437, 0.0276]−0.0461 [−0.0882, 0.0203]

Notes: Arms: change in α<sub>ij</sub> per unit Δ ln m among competitors larger (k<sub>sim</sub> + k<sub>hier</sub>) or smaller (k<sub>hier</sub> − k<sub>sim</sub>) than the focal species; a positive arm means effects increase with competitor size. Hindfoot: per unit log hindfoot difference. k<sub>guild</sub>: Table S3. Interval widths of k<sub>sim</sub> and k<sub>hier</sub>: 0.0297 and 0.0589 (primary), 0.0658 and 0.0892 (without ant-removal plots). Tables only: the kernel's shape is given by its arms rather than plotted.

Discussion

By its pre-specified rule, self-limitation held in the primary model and in 9 of 12 alternative specifications but failed in the three least precise. The rule rests on a mean of ratios whose denominators approach zero in some bootstrap replicates, and the summaries that favoured self-limitation in every specification are post hoc. Relative self-limitation is therefore supported under the primary identification rather than established robustly. It parallels the plant result of Adler et al. (2018) and time-series decompositions dominated by intraspecific regulation (Mutshinda et al. 2009), but is necessary for stable coexistence, not sufficient (Chesson 2000; Barabás et al. 2018); we did not estimate fitness differences or invasion growth rates. The direction of the 2.24-fold gap between OLS and 2SLS coefficients is what census error predicts (Freckleton et al. 2006; Knape & de Valpine 2012), though the gap is not uniquely attributable to it.

The instrument problem suggested by the i.i.d. Sargan rejection did not survive tests allowing for dependence across plots within a census, within plot-years or within years; only a version allowing correlation solely among species within a plot-census rejected, marginally. The data therefore do not show that either instrument set is invalid. Non-rejection does not establish validity, however, and an overidentification test cannot detect a violation shared by both sets, such as a direct effect of N(t − 1) that also biases the t − 2 instruments. We detected no delayed intraspecific density dependence, but the test was weak, and specifications identified from second or third lags gave less precise estimates with reversed kernel slopes; these data do not settle which identification is closer to the truth.

In the primary model competition was size-structured, but not as limiting similarity predicts. The negative similarity slope comes from the smaller-competitor arm: competitors much smaller than the focal species had weak effects per unit biomass, whereas larger competitors had effects that did not vary with their size. That is a hierarchy among smaller competitors that levels off at the focal species' size, not the symmetric decline with size difference that limiting similarity predicts (MacArthur & Levins 1967) or that size ratios among desert granivores suggest (Bowers & Brown 1982). It only partly resembles the hierarchy among trees (Kunstler et al. 2012), since the overall hierarchy slope had an interval including zero and reversed under OLS and second-lag instruments. With 11 species, size difference and competitor size are nearly collinear, and if size differences both reduce niche overlap and create fitness differences, similarity need not have a large or stable net effect (Mayfield & Levine 2010). For D. merriami and D. ordii, which shifted reciprocally after removals (Brown & Munger 1985), the kernel-implied ratios ranked only one direction.

The two pairs with q ≥ 1 involve C. baileyi, which colonised Portal in 1995 and, on plots without kangaroo rats, used most of the resources they had left (Ernest & Brown 2001). That, by point estimates, D. merriami and D. spectabilis would limit C. baileyi more than themselves fits this overlap in resource use, but the estimates rest on 608 rows and both intervals include one and zero. Two-thirds of kernel-implied pairs were non-competitive (66.4%, 95% CI 33.6–80.0%), echoing the facilitation Adler et al. (2018) found among plants, but mean implied interspecific coefficients excluded zero only for C. penicillatus and D. spectabilis, so 'most pairs combine self-limitation with facilitation' describes point estimates. Facilitative estimates may also reflect plot-specific resource pulses not removed by census effects, a bias favouring H1.

Limitations are substantial. Plots are open, so intraspecific coefficients include movement and resident turnover; dispersal declines with size at Portal (Brown & Zeng 1989), which may explain the large per-biomass coefficients of the smallest species, and the 61% breakdown value assumes uniform turnover. Half the sample plot-years had ants removed, and the plots without ant removal alone were too few to confirm or contradict the pooled result: intervals were 2.2 times wider for the similarity slope and 6.9 times for mean q. Coefficients were assumed constant over 25 years at one site despite community reorganisation (Christensen et al. 2018), although plot-by-period effects and plot trends left the conclusions unchanged; abundance is an index without detection modelling; body masses include juveniles. The format carries tables only, so the kernel's shape is given by its arms rather than plotted. Multispecies autoregressive models (Ives et al. 2003) fitted with explicit observation error (Knape & de Valpine 2012), and marked-individual demography, are the next tests.

For conservation and management, competition inferred from monitoring counts needs explicit treatment of census error, and body-size similarity is an unreliable guide to which residents will compete most with a colonist or gain from a decline. Self-limitation was the more consistent of the two patterns, but it held under the primary identification rather than under every alternative, and whether it secures coexistence also depends on fitness differences we did not estimate.

Acknowledgements; Author Contributions; Data Availability Statement

Acknowledgements. We thank the Portal Project for collecting and openly publishing its rodent census data. No specific funding supported this work.

Author Contributions. Designed, analysed and written by an AI research model (Claude Opus 5.5) in an automated pipeline that executed the analysis script on the archived data; the Portal Project collected the data.

Data Availability Statement. CC0 Portal Project Teaching Database (Ernest et al. 2018; https://doi.org/10.6084/m9.figshare.1314459), surveys table retrieved 24 October 2026 (SHA-256 0e7b92161a10b1fc…); the analysis script and run record are archived with this study.

References

Adler, P.B., Smull, D., Beard, K.H., Choi, R.T., Furniss, T., Kulmatiski, A. et al. (2018) Competition and coexistence in plant communities: intraspecific competition is stronger than interspecific competition. Ecol. Lett., 21, 1319–1329. https://doi.org/10.1111/ele.13098

Barabás, G., D'Andrea, R. & Stump, S.M. (2018) Chesson's coexistence theory. Ecol. Monogr., 88, 277–303. https://doi.org/10.1002/ecm.1302

Bowers, M.A. & Brown, J.H. (1982) Body size and coexistence in desert rodents: chance or community structure? Ecology, 63, 391–400. https://doi.org/10.2307/1938957

Brown, J.H., Gillooly, J.F., Allen, A.P., Savage, V.M. & West, G.B. (2004) Toward a metabolic theory of ecology. Ecology, 85, 1771–1789. https://doi.org/10.1890/03-9000

Brown, J.H. & Munger, J.C. (1985) Experimental manipulation of a desert rodent community: food addition and species removal. Ecology, 66, 1545–1563. https://doi.org/10.2307/1938017

Brown, J.H. & Zeng, Z. (1989) Comparative population ecology of eleven species of rodents in the Chihuahuan Desert. Ecology, 70, 1507–1525. https://doi.org/10.2307/1938209

Chesson, P. (2000) Mechanisms of maintenance of species diversity. Annu. Rev. Ecol. Syst., 31, 343–366. https://doi.org/10.1146/annurev.ecolsys.31.1.343

Christensen, E.M., Harris, D.J. & Ernest, S.K.M. (2018) Long-term community change through multiple rapid transitions in a desert rodent community. Ecology, 99, 1523–1529. https://doi.org/10.1002/ecy.2373

Ernest, M., Brown, J., Valone, T. & White, E.P. (2018) Portal Project Teaching Database. figshare. https://doi.org/10.6084/m9.figshare.1314459

Ernest, S.K.M. & Brown, J.H. (2001) Delayed compensation for missing keystone species by colonization. Science, 292, 101–104. https://doi.org/10.1126/science.292.5514.101

Ernest, S.K.M., Valone, T.J. & Brown, J.H. (2009) Long-term monitoring and experimental manipulation of a Chihuahuan Desert ecosystem near Portal, Arizona, USA. Ecology, 90, 1708. https://doi.org/10.1890/08-1222.1

Ernest, S.K.M., Yenni, G.M., Allington, G., Christensen, E.M., Geluso, K., Goheen, J.R. et al. (2016) Long-term monitoring and experimental manipulation of a Chihuahuan desert ecosystem near Portal, Arizona (1977–2013). Ecology, 97, 1082. https://doi.org/10.1890/15-2115.1

Freckleton, R.P., Watkinson, A.R., Green, R.E. & Sutherland, W.J. (2006) Census error and the detection of density dependence. J. Anim. Ecol., 75, 837–851. https://doi.org/10.1111/j.1365-2656.2006.01121.x

Heske, E.J., Brown, J.H. & Mistry, S. (1994) Long-term experimental study of a Chihuahuan Desert rodent community: 13 years of competition. Ecology, 75, 438–445. https://doi.org/10.2307/1939547

Ives, A.R., Dennis, B., Cottingham, K.L. & Carpenter, S.R. (2003) Estimating community stability and ecological interactions from time-series data. Ecol. Monogr., 73, 301–330. https://doi.org/10.1890/0012-9615(2003)073[0301:ECSAEI]2.0.CO;2

Knape, J. & de Valpine, P. (2012) Are patterns of density dependence in the Global Population Dynamics Database driven by uncertainty about population abundance? Ecol. Lett., 15, 17–23. https://doi.org/10.1111/j.1461-0248.2011.01702.x

Kunstler, G., Lavergne, S., Courbaud, B., Thuiller, W., Vieilledent, G., Zimmermann, N.E. et al. (2012) Competitive interactions between forest trees are driven by species' trait hierarchy, not phylogenetic or functional similarity: implications for forest community assembly. Ecol. Lett., 15, 831–840. https://doi.org/10.1111/j.1461-0248.2012.01803.x

Lima, M., Ernest, S.K.M., Brown, J.H., Belgrano, A. & Stenseth, N.C. (2008) Chihuahuan Desert kangaroo rats: nonlinear effects of population dynamics, competition, and rainfall. Ecology, 89, 2594–2603. https://doi.org/10.1890/07-1246.1

MacArthur, R. & Levins, R. (1967) The limiting similarity, convergence, and divergence of coexisting species. Am. Nat., 101, 377–385. https://doi.org/10.1086/282505

Mayfield, M.M. & Levine, J.M. (2010) Opposing effects of competitive exclusion on the phylogenetic structure of communities. Ecol. Lett., 13, 1085–1093. https://doi.org/10.1111/j.1461-0248.2010.01509.x

Mutshinda, C.M., O'Hara, R.B. & Woiwod, I.P. (2009) What drives community dynamics? Proc. R. Soc. B, 276, 2923–2929. https://doi.org/10.1098/rspb.2009.0523

Sanderson, E. & Windmeijer, F. (2016) A weak instrument F-test in linear IV models with multiple endogenous variables. J. Econom., 190, 212–221. https://doi.org/10.1016/j.jeconom.2015.06.004

Supporting Information

Table S1. Delayed density dependence: coefficients on densities at t (immediate) and at t − 1 (delayed).

Code or slopeImmediate [95% CI]Delayed [95% CI]
PF0.923 [−0.171, 2.140]−0.087 [−0.378, 0.209]
RM0.565 [−1.020, 2.649]0.142 [−0.246, 0.519]
PP0.930 [−0.184, 1.426]−0.209 [−0.382, 0.202]
PE1.301 [−0.055, 2.262]−0.240 [−0.460, 0.173]
OT0.288 [0.072, 0.695]−0.019 [−0.106, 0.045]
PB−0.009 [−0.450, 0.299]0.038 [−0.074, 0.214]
OL0.567 [−0.526, 1.298]−0.066 [−0.220, 0.256]
DM0.004 [−0.044, 0.061]0.018 [−0.014, 0.045]
DO0.085 [0.000, 0.280]0.003 [−0.079, 0.046]
DS0.042 [−0.026, 0.113]−0.002 [−0.041, 0.044]
NL0.103 [0.038, 0.253]−0.010 [−0.046, 0.015]
k<sub>sim</sub>0.0167 [−0.0172, 0.0405]−0.0105 [−0.0203, 0.0024]
k<sub>hier</sub>−0.0295 [−0.0576, 0.0095]0.0133 [−0.0021, 0.0235]
k<sub>guild</sub>−0.0211 [−0.0952, 0.0378]0.0055 [−0.0106, 0.0294]

Notes: E(t − 1) terms added to Eq. 1 as predetermined regressors; E(t) instrumented by densities at t − 2 and t − 3; 15,625 rows; 299 year-block bootstrap replicates. Minimum conditional first-stage F 1.1; 19 of 50 regressors below 10; k<sub>sim</sub> and k<sub>hier</sub> 13.5 and 16.9. Instrumenting E(t − 1) as well: conditional F 0.01–0.42 (unidentified).

Table S2. Intraspecific coefficients a<sub>i</sub> under the specifications added in revision.

CodePrimaryPresence at t − 1 [95% CI]Without ant-removal plots [95% CI]ln(0.5 + N)ln(2 + N)Plot × periodPlot trends
PF0.5230.728 [0.445, 1.473]1.265 [−2.431, 8.305]0.7670.3280.5900.550
RM1.2561.335 [0.430, 2.342]1.185 [0.100, 2.710]1.8140.8081.4171.395
PP0.2250.213 [−0.606, 0.508]0.260 [−1.373, 0.714]0.2740.1740.2920.277
PE0.4220.518 [0.308, 0.987]0.280 [−0.835, 1.102]0.6030.2740.4670.459
OT0.1430.162 [−0.159, 0.502]0.159 [−1.372, 0.918]0.2140.0880.1660.160
PB0.0800.083 [0.053, 0.172]0.122 [−0.004, 0.423]0.1020.0580.0960.116
OL0.2570.254 [−0.049, 0.481]0.308 [−0.780, 0.640]0.3670.1670.2840.327
DM0.0330.033 [0.008, 0.058]0.070 [0.037, 0.126]0.0400.0260.0500.051
DO0.0820.085 [0.048, 0.126]0.042 [−0.073, 0.119]0.1090.0570.1030.098
DS0.0390.041 [0.022, 0.067]0.036 [−0.007, 0.093]0.0560.0260.0460.052
NL0.0690.070 [0.038, 0.125]0.074 [0.000, 0.177]0.1010.0430.0930.091
Median ratio to primary11.031.111.430.651.181.23

Notes: Point estimates for the offset and plot-effect variants; intervals from 299 year-block bootstrap replicates. Offsets change the scale of the response, so coefficients scale with the offset while the ratios q<sub>ij</sub> do not.

Table S3. Same-guild shift k<sub>guild</sub> under each specification.

Specificationk<sub>guild</sub> [95% CI]
Primary0.0135 [−0.0208, 0.0513]
OLS0.0063 [−0.0001, 0.0132]
t − 2 instruments−0.0169 [−0.1066, 0.0461]
Per-capita currency0.0165 [−0.0267, 0.0678]
Hindfoot size axis0.0001 [−0.0384, 0.0251]
Plot × period effects0.0158 [−0.0354, 0.0716]
Plot trends within phase0.0144 [−0.0270, 0.0529]
ln(0.5 + N)0.0219 [−0.0251, 0.0825]
ln(2 + N)0.0077 [−0.0146, 0.0335]
Rows on presence at t − 10.0187 [−0.0164, 0.0489]
Without ant-removal plots0.0078 [−0.0398, 0.0792]
Delayed density, immediate−0.0211 [−0.0952, 0.0378]

Notes: Units of α. Not estimated in the granivores-only variant.

6.Review

Attempt 2Done

Automated review: Accept (novelty 3, rigor 4, clarity 4, reproducibility 5).

Accept
Novelty 3/5Rigor 4/5Clarity 4/5Reproducibility 5/5

Refereed by Claude Fable 5.1, a different model from the one that wrote the paper (Claude Opus 5.5).

Final round. The three text changes requested in round 2 are made as specified and verified against the manuscript and the unchanged record: the Abstract's size-structure claim is confined to the primary model with the second-lag reversal noted, the marginal plot-census rejection is acknowledged, and the blank Table 5 cell is filled. Every number still traces to RESULT_JSON, all citations are retrieved sources with resolving DOIs, and the venue conventions are met (149-word abstract, hypothesis-driven Letter, parameters with units, variability across replicates, ecological interpretation, first person plural). The remaining limitations (no figure, 299-replicate verdicts, modest generality) are stated in the paper and do not require changes. Accept.

Strengths

  • The three round-2 changes are made exactly as required and checked against the record. The Abstract's last sentence now reads 'In the primary model, competitive effects declined toward smaller, not larger, competitors, contrary to limiting similarity; second-lag instruments reversed this', which matches Table 6 (smaller-competitor arm 0.0402 [0.0121, 0.0885] primary; -0.0316 [-0.0877, -0.0015] under t-2). The instrument sentence names the one marginal exception (plot-census clusters, p = 0.0444). Table 5's summed delayed-density cell reads 'not computed', consistent with the script, which computes the Monte Carlo SE only from the immediate-coefficient replicates.
  • The Abstract stays at 149 words; naming only the second-lag failure (upper limit 2.145) among 'the three least precise' is accurate, and the Results and Table 5 name all three (2.145, 4.201, 2.520/1.533).
  • RESULT_JSON is unchanged from round 2 and every number in the text, Tables 1-6 and Tables S1-S3 still traces to it; the dataset, source, hash, sample filters and design match the executed script and the provenance exactly.
  • All 14 round-1 required changes remain met: dependence-robust overidentification tests, H1 judged by its pre-specified rule across 12 alternatives with Monte Carlo error and post hoc summaries labelled, kernel arms in every specification, the species-less-record check, delayed-density, plot-trend, offset, row-rule and ant-removal variants, the nine added citations (22 DOIs resolve), and the venue structure.
  • The paper now claims exactly what the data support: self-limitation under the primary identification and in 9 of 12 alternatives but not robustly; no limiting similarity; a smaller-competitor decline in the primary model that reverses under second-lag instruments.

Weaknesses

  • No figure; the kernel arms in Table 6 substitute, and the limitation is stated. The format cannot supply one.
  • The pre-specified H1 verdicts in several alternatives rest on 299 replicates with large Monte Carlo error; the manuscript says so, and the author declined 999 replicates to keep the shared script identical across editions, which is a reasonable trade-off given the statistic's intrinsic instability.
  • The conceptual advance is modest: the main result is qualified rather than general, and H2 fails; the contribution is a carefully identified, honestly reported estimate of community-wide competition coefficients corrected for census error.

Required changes

—

Record

§Event timeline

  1. Run createdQuestion: “Does intraspecific competition exceed interspecific competition in a desert rodent community, and is interspecific competition stronger between body-size-similar species?' language=en discipline=ecology venue='Ecology Letters” · output language: English · models: agent:claude-opus-5-5
  2. Stage started · Literature reviewAttempt 1 · agent:claude-opus-5-5
  3. awaiting_agent · Literature reviewsearch response for literature_review: literature_review-search-1.response.json
  4. Resume requested
  5. Stage started · Literature reviewAttempt 2 · agent:claude-opus-5-5
  6. Search complete · Literature review42 sources · 15,105 characters of notes
  7. awaiting_agent · Literature reviewcomplete_json response for literature_review: literature_review-complete_json-1.response.json
  8. Resume requested
  9. Stage started · Literature reviewAttempt 3 · agent:claude-opus-5-5
  10. Search complete · Literature review42 sources · 15,105 characters of notes
  11. Stage completed · Literature review
  12. Stage started · IdeationAttempt 1 · agent:claude-opus-5-5
  13. awaiting_agent · Ideationcomplete_json response for ideation: ideation-complete_json-1.response.json
  14. Resume requested
  15. Stage started · IdeationAttempt 2 · agent:claude-opus-5-5
  16. Awaiting approval · Ideation
  17. Approved · IdeationIdea 1 selected
  18. Stage started · ExperimentationAttempt 1 · agent:claude-opus-5-5
  19. Dataset provisioned · Experimentationportal_rodent_surveys · 35,549 observations · 10 variables · sha256 0e7b92161a10
  20. awaiting_agent · Experimentationcomplete response for experimentation: experimentation-complete-1.response.txt
  21. Resume requested
  22. Stage started · ExperimentationAttempt 2 · agent:claude-opus-5-5
  23. Dataset provisioned · Experimentationportal_rodent_surveys · 35,549 observations · 10 variables · sha256 0e7b92161a10
  24. Experiment round · ExperimentationRound 1: exit 0; result parsed (37.255s)
  25. Stage completed · Experimentation
  26. Stage started · AnalysisAttempt 1 · agent:claude-opus-5-5
  27. awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
  28. Resume requested
  29. Stage started · AnalysisAttempt 2 · agent:claude-opus-5-5
  30. Stage completed · Analysis
  31. Stage started · Paper writingAttempt 1 · agent:claude-opus-5-5
  32. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  33. Resume requested
  34. Stage started · Paper writingAttempt 2 · agent:claude-opus-5-5
  35. Stage completed · Paper writing
  36. Stage started · ReviewAttempt 1 · agent:claude-opus-5-5
  37. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  38. Resume requested
  39. Stage started · ReviewAttempt 2 · agent:claude-opus-5-5
  40. Awaiting approval · Review
  41. Sent back for revision · ReviewRe-running from Paper writing: “Referee recommendation: minor_revision. Make every required change below using only results already in the run record (no new analyses or numbers). Required changes, quoted from the referee:\n1. Label ”
  42. Stage started · Paper writingAttempt 3 · agent:claude-opus-5-5
  43. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  44. Resume requested
  45. Stage started · Paper writingAttempt 4 · agent:claude-opus-5-5
  46. Stage completed · Paper writing
  47. Stage started · ReviewAttempt 1 · agent:claude-opus-5-5
  48. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  49. Resume requested
  50. Stage started · ReviewAttempt 2 · agent:claude-opus-5-5
  51. Awaiting approval · Review
  52. Approved · Review
  53. Run completereport=/Users/haodingresearch/Desktop/auto-research/runs/studies-opus55/ecology-en/runs/510551359416/REPORT.md
  54. Reopened after sign-off · ReviewReopened after sign-off; re-running from Review: “Independent review: the paper is re-refereed by an independent referee (Claude Fable 5.1) that took no part in writing it.”
  55. Stage started · ReviewAttempt 3 · agent:claude-fable-5-1
  56. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  57. Resume requested
  58. Stage started · ReviewAttempt 4 · agent:claude-fable-5-1
  59. Awaiting approval · Review
  60. Sent back for revision · ReviewRe-running from Experimentation: “Independent referee (Claude Fable 5.1): major_revision. Required changes: (1) [text] Abstract, 4.2, 4.3 and Discussion: state that the pre-specified H1 rule (upper limit of mean q < 1) holds in the pr”
  61. Stage started · ExperimentationAttempt 3 · agent:claude-opus-5-5
  62. Dataset provisioned · Experimentationportal_rodent_surveys · 35,549 observations · 10 variables · sha256 0e7b92161a10
  63. awaiting_agent · Experimentationcomplete response for experimentation: experimentation-complete-1.response.txt
  64. Resume requested
  65. Stage started · ExperimentationAttempt 4 · agent:claude-opus-5-5
  66. Dataset provisioned · Experimentationportal_rodent_surveys · 35,549 observations · 10 variables · sha256 0e7b92161a10
  67. Experiment round · ExperimentationRound 1: exit 0; result parsed (52.879s)
  68. Stage completed · Experimentation
  69. Stage started · AnalysisAttempt 1 · agent:claude-opus-5-5
  70. awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
  71. Resume requested
  72. Stage started · AnalysisAttempt 2 · agent:claude-opus-5-5
  73. Stage completed · Analysis
  74. Stage started · Paper writingAttempt 5 · agent:claude-opus-5-5
  75. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  76. Resume requested
  77. Stage started · Paper writingAttempt 6 · agent:claude-opus-5-5
  78. Stage completed · Paper writing
  79. Stage started · ReviewAttempt 1 · agent:claude-fable-5-1
  80. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  81. Resume requested
  82. Stage started · ReviewAttempt 2 · agent:claude-fable-5-1
  83. Awaiting approval · Review
  84. Sent back for revision · ReviewRe-running from Analysis: “Independent referee (Claude Fable 5.1), round 2: minor_revision. Required changes: (1) [text] Abstract, final sentence: restrict the claim to the primary model and note the reversal, for example 'In t”
  85. Stage started · AnalysisAttempt 3 · agent:claude-opus-5-5
  86. awaiting_agent · Analysiscomplete_json response for analysis: analysis-complete_json-1.response.json
  87. Resume requested
  88. Stage started · AnalysisAttempt 4 · agent:claude-opus-5-5
  89. Stage completed · Analysis
  90. Stage started · Paper writingAttempt 7 · agent:claude-opus-5-5
  91. awaiting_agent · Paper writingcomplete response for paper_writing: paper_writing-complete-1.response.txt
  92. Resume requested
  93. Stage started · Paper writingAttempt 8 · agent:claude-opus-5-5
  94. Stage completed · Paper writing
  95. Stage started · ReviewAttempt 1 · agent:claude-fable-5-1
  96. awaiting_agent · Reviewcomplete_json response for review: review-complete_json-1.response.json
  97. Resume requested
  98. Stage started · ReviewAttempt 2 · agent:claude-fable-5-1
  99. Awaiting approval · Review
  100. Approved · Review
  101. Run completereport=/Users/haodingresearch/Desktop/auto-research/runs/studies-opus55/ecology-en/runs/510551359416/REPORT.md