Skip to content
Science Journaling Club

Volume 1, Issue 1September 2024

Fall 2024

Chance is the subject. What randomness does to a beneficial mutation, to a small population's odds of lasting the century, and to a survey that comes back empty. It closes on a meta-analysis of the insect decline literature, which turns out to disagree with itself more than either side admits.

This issueSJC 1.1
Published
September 2024
Articles
7
Research
5
Field notes
2
With code
7

Contents

7 articles, in the order the club meant them to be read.

  1. Field noteExplainer

    How to Read a Scientific Paper Without Pretending You Understood It

    Reading primary research is the one skill this club exists to teach. Almost nobody gets taught it. Here is the handout we wish somebody had pushed across the table on day one: what each part of a paper does, why experienced readers go to the figures before the prose, how to check whether the statistics carry the conclusion, and what peer review promises in writing. Honesty demanded that we measure the thing we were describing. With a text-analysis script we wrote ourselves, we scored 67 open-access research papers pulled from Europe PMC and all 298 sections of the club's own 27 published articles on one instrument. The median Results section runs 3,148 words against a median abstract of 206, a compression of 15.3 to one. Register changes harder than length does. Results sections favour reporting vocabulary over interpretation vocabulary 7.58 to one; abstracts flip it, with interpretation winning at 0.75 to one, a swing of about ten times. The popular charge that abstracts inflate their verbs did not survive our own test. Paired within each paper, abstract-minus-Results claim strength came out at +0.33 per thousand words, t = 0.16, n = 67, which is nothing at all. Abstracts do not shout. They change what kind of sentence they are writing, and they take the evidence away with them. Read for that.

    Scientific PracticeRead 37 minSections 13

  2. ResearchOriginal Research

    How Often Does a Good Gene Lose? Measuring the Reach of Genetic Drift

    A mutation that makes its carrier one percent fitter still disappears almost every time it appears. We measured how often. The instrument was a Wright-Fisher simulation the club wrote and ran on a laptop, and no organism entered the study at any point, because here the computation is the experiment. Across a grid of eight selection coefficients (s=0s = 0 to 0.10.1) and five population sizes (N=50N = 50 to 10,00010{,}000) we ran 11,932,191 independent replicates from a single new mutant copy and counted 90,675 fixations, holding the relative standard error to no worse than 5.93% in any of the 40 cells. For N=1000N = 1000 and s=0.01s = 0.01, a single new copy fixed in 1.992% ± 0.063% of replicates. Loss took the other 98.01%. The neutral case recovered the exact answer 1/(2N)1/(2N) at all five population sizes, the one corner of the study where the theory leaves no room for approximation at all, and every 95% interval covered the true value, the largest z score being 1.33. Haldane's 2s2s is excellent while 4Ns≳44Ns \gtrsim 4. Below that it fails badly, falling short of the measured value by 488% at N=50,s=0.001N = 50, s = 0.001; its mean absolute error is 255.6% for 4Ns<14Ns < 1 and 2.5% for 4Ns4Ns between 4 and 20. Kimura's diffusion formula holds to a few percent across the whole grid, then breaks at the opposite edge: at s=0.1s = 0.1 it sits 3.0% above our measurement, a deviation confirmed independently at all five population sizes with a pooled z of −7.33. Haldane's own branching process, taken before the small-ss step that produces 2s2s and solved numerically instead, matches that same measurement to 0.2%, which puts the fault in the diffusion limit rather than in either man's reasoning. Conditional fixation times reproduced the 4N4N neutral prediction to within 5.1% at every size. Seed 20240917.

    Fall 2024Read 31 minSections 12

  3. ResearchOriginal Research

    A Population That Grows on Average Can Still Die Out

    A population in which every individual leaves 1.15 offspring on average grows by 15 per cent a generation and can still go extinct by bad luck alone. We simulated a Galton-Watson branching process and a continuous-time birth-death process, 20,000 replicates per parameter cell, for starting sizes from 2 to 500 and five offspring distributions whose variance ranges from 1.15 to 33.2 with the mean held exactly fixed. The probability of extinction within 100 generations falls below 5 per cent at a starting size of 10.5 under Poisson reproduction, 21.4 under geometric reproduction, and 302.9 under a lottery distribution in which about one individual in twenty-six leaves 30 offspring and the rest leave none. Across the five levels the threshold is close to proportional to the variance: a log-log fit gives a slope of 1.0014 with R² = 0.9994, so the safe size is about 8.8 times the offspring variance at this growth rate. Every simulated cell carries an exact analytic partner, computed along a wholly separate path: we iterate the offspring probability generating function forward, then locate its smallest non-negative fixed point by bisection, and the two answers have to agree. Over the 154 cells with at least ten expected extinctions the mean z-score is −0.048 with a standard deviation of 0.943 and no cell exceeds three standard errors; the two closed forms we can solve independently, 1/m for geometric offspring and the Lambert-W form for Poisson offspring, match our root finder to within 1.3×10−15. Haldane's 1927 approximation for lineage survival, 2s/V, written before anyone could simulate anything, sits within 0.2 to 13.2 per cent of our exact values over a thirty-fold range of variance. The whole study is a computation: the club has no laboratory, observed no animals, and claims nothing about any of them.

    Fall 2024Read 28 minSections 12

  4. ResearchOriginal Research

    Everything a Small Founding Group Loses on the Way

    A founding group of two individuals carries 74.973% ± 0.048% of the source population's heterozygosity and 20.69% of its alleles. Both numbers come out of one run of one program. They are most of what this paper has to say. We simulated founder events by drawing 2k2k gametes from a source population with a fixed and printed allele frequency spectrum, then grew the new population back to a carrying capacity of 1,000 under Wright-Fisher reproduction with K-allele mutation. No organism was involved, and no island. The computation is the experiment. Across fifteen founding sizes from k=2k = 2 to k=500k = 500 at 20,000 replicates each, measured heterozygosity retention matched the closed form Hnew=Hold(1−1/2k)H_{\text{new}} = H_{\text{old}}(1 - 1/2k) with a largest standardised deviation of 2.13 and no cell outside three standard errors. Allele loss matched the binomial prediction ∑i(1−pi)2k\sum_i (1-p_i)^{2k} to a largest deviation of 1.54 standard errors; two founders lost 183.21 of the source's 231 alleles against a prediction of 183.19. Drawing founders without replacement from a finite source of N=60N = 60 reproduced the hypergeometric correction, and at k=Nk = N the founding group lost nothing at all, in every one of 20,000 replicates, to a floating-point difference of 1.11×10−161.11 \times 10^{-16}. Over a grid of eight founding sizes by four growth rates, 5,000 replicates each, simulated gene diversity agreed with an exact recursion at every one of the 32 cells, mean zz −0.004 and standard deviation 0.998. Growth rate mattered more than we expected: two founders growing at 5% a generation held 12.35% of the source's diversity by the time they reached capacity, and at 100% a generation they held 60.15%. The bottleneck signature faded much faster than the loss did. A heterozygosity excess test on 30 individuals at 15 loci detected a five-founder event 67.08% of the time at generation 10 and 6.48% of the time at generation 60, while the population itself was still 29.6% short of the control two hundred generations later. Seed 20241108, 630,000 replicates, 276 seconds.

    Fall 2024Read 29 minSections 12

  5. ResearchMethods

    Absence of Evidence: How Many Surveys Before You Can Say a Species Is Gone

    A survey that finds nothing has not proved anything. We built the detection half of a single-season occupancy model and asked the scheduling question underneath it: how long does a run of blank visits have to be before finding nothing counts as evidence of absence? With a constant per-visit detection probability pp, the probability that an occupied site produces kk blanks is (1−p)k(1-p)^k, so 95 percent confidence needs 14 visits at p=0.20p = 0.20 and 59 at p=0.05p = 0.05. We checked 500,000 simulated survey histories per configuration against that closed form across 76 cells: mean z=+0.0044z = +0.0044, largest ∣z∣=3.32|z| = 3.32, and the eleven independent anchor cells average +0.27+0.27, which is 0.89 standard errors from zero. Varying detection probability costs nothing, or costs everything, depending on where the variation lives. Redrawn from a Beta distribution at every visit it is free, which we prove and then confirm. Drawn once per site and kept, the bill arrives: at mean p=0.20p = 0.20 and Beta concentration κ=5\kappa = 5 (SD 0.163) the requirement rises from 14 visits to 76, a factor of 5.66, and at κ=2\kappa = 2 to 2,369 visits. Twenty blanks that the naive formula scores at a 1.15 percent chance of error actually leave 16.7 percent, and we validated that heterogeneous case three ways, simulation against a Beta-function closed form against tanh-sinh quadrature, agreeing to 5.7 × 10-14. Our first quadrature was wrong by 0.86. We say where. Putting Kéry's published snake detection probabilities into our formula reproduces two of his three survey requirements exactly. Seed 20240921. Everything here is computation, and the club made no observations of anything.

    Fall 2024Read 27 minSections 12

  6. ResearchMethods

    How Many Samples Before Your Species Count Means Anything

    A species count is always short. The four estimators everybody reaches for are supposed to tell you by how much, so we tested them against a truth we knew, because we built it: five community types, each holding exactly 100 species, sampled at ten levels of effort, 6,000 replicate communities per configuration, 300,000 simulated surveys in total. No organism was involved at any stage. On a mildly uneven community (log-normal σ=1.0\sigma = 1.0, Pielou J=0.895J = 0.895) a survey of 500 individuals turned up 86.28 ± 0.05 species, which Chao1 corrected to 95.42 ± 0.09. Still 4.58 species short. Chao2 landed at 94.74, ACE at 95.49. The second-order jackknife, a cruder and older device, overshot to 104.26 and reached the 10 percent band at 106 individuals against Chao1's 268 and Chao2's 302, a gap of nearly three to one. Chao1 and Chao2 run low in all fifty cells of the grid, the jackknives correct past the truth instead, and the ranking tracks the abundance distribution rather than the estimator's reputation. Past a certain effort the correction stops helping: at n = 5,000 in that same community the raw count beat Chao1 on root mean square error, 0.490 against 0.570. Validation: the accumulation curve matched its closed form in all 50 cells, largest ∣z∣=2.62|z| = 2.62, spread of zz across cells 1.01; Hurlbert's hypergeometric rarefaction matched direct subsampling to 0.71 standard errors at worst; and every estimator returned exactly 100.000 species at exhaustive effort. Seed 20241104.

    Fall 2024Read 27 minSections 12

  7. Field noteMeta-analysis

    The Insect Decline Literature Disagrees With Itself, and We Wanted to Know How Badly

    One German study reported a 76.7 percent fall in flying-insect biomass over 27 years [1]. A synthesis of 166 long-term surveys put terrestrial declines near 9 percent per decade while freshwater insects rose about 11 percent per decade, and an erratum later steepened the terrestrial figure [2, 3]. More than 5,300 American time series gave net trends indistinguishable from zero [6], which is a statement about time series and about the units they were expressed in, though almost nobody repeated it that way. All three are competent work. We extracted 11 effect sizes from 6 papers covering 10 monitoring datasets, traced every one to a printed slope or a printed interval in its source, converted them to a common log-linear annual rate, and pooled them under a DerSimonian-Laird random-effects model. The pool comes out at −2.68 percent per year (95% CI −4.34 to −0.99), or −23.8 percent per decade, and the spread around it is the part worth printing: I² = 96.3 percent, with Cochran's Q at 269.96 on 10 degrees of freedom. That second pair of numbers matters more than the first. A 95% prediction interval for the next comparable study runs from −8.14 to +3.10 percent per year, so a new paper reporting insects on the rise would sit comfortably inside everything published to date. We also simulate the objection that monitoring programmes began at unusually good sites, where regression to the mean manufactures a decline nobody caused; under a deliberately unkind version of that model the artefact accounts for roughly 15 percent of our pooled estimate. Real and quantified, and nowhere near large enough to carry the objection. We are a school club reading printed tables. Weight the pooled number below what its confidence interval implies. Weight the heterogeneity above it.

    EcologyRead 41 minSections 13