Abstract

The official thesis abstract (verbatim).

Fields work in very different ways — but fail in similar ways. This thesis covers some of my work in epidemiology, psychology, and machine learning with the common thread of shared methodological issues. I identify frameworks which fail to cover actual practice, practices which fail to live up to normative principles, and propose practices which are sometimes able to address some failures at some cost.

In epidemiology I take a Bayesian approach to infectious disease modelling and infer the effect of entire populations wearing face masks during the Covid pandemic — with the key caveat that this is an observational study. I identify a ubiquitous methodological mistake (using mandate timings as a proxy for wearing behaviour, when these are, surprisingly, not strongly correlated).

In psychology I synthesise theories of the replication crisis and report on initiating a large (n = 1931) dataset of replication studies covering the original effect sizes, replication effect sizes, and both raw and recalculated statistics. These nonrandom data still give some insight into post-crisis psychology, confirming past results showing that even the sign of effects is often not replicable. I note a pattern of considerable ‘shrinkage’ in effect sizes between the original study and their replications.

In machine learning I trace recent methodological changes and provide a novel analysis of roughly forty ways that ML evaluations are often misleading.

Each chapter contains a self-contained background section for its respective field. I conclude with lessons for each field from the other two.

“an impressive work of scholarship… The epidemiology has had broad influence, including on policy discussion at a national level… the machine learning chapter is an admirable feat of investigation and synthesis.” — Conor Houghton

Author’s gloss: Broadly it is about epistemics: why we can’t learn that much from one study, or from many studies. In Newell’s typology the thesis “contradicts existing knowledge; thoroughly explores an area; provides empirical data; and produces a negative result.” A Gelmanian work — Gelman is cited 74 times, in every chapter. Contribution: ~3500 hours on the whole PhD (226 on the writeup).

Full text

The thesis runs to ~165,000 words across four chapters. Rather than inline it all (its chapters are the substance of several other entries in this folder), this section reproduces the thesis’s unique synthesis material — the structure, the per-paper contributions, and the cross-field conclusion — and cross-links each chapter to the corresponding paper. Full PDF linked in the frontmatter.

Table of contents

• Front matter: Contributions; Glossary; Software
• 1. Introduction
• 2. Bayes: How some epidemiologists learn from data (epidemiology)
  • Background: Bayesian inference and workflow
  • Background: Epidemiology and Bayes
  • Application: Modelling nonpharmaceutical interventions and mask-wearing
  • Constituent papers: inferring-npi-effectiveness, npi-robustness, mask-wearing, seasonal-covid, parallel-reweighted-wake-sleep
• 3. Crisis: How psychologists learn from data (psychology)
  • Background: the workflow of frequentist inference
  • Literature review: The replication crisis
  • ‘Do more replications’
  • The FORRT Replications and Reversals database
  • Against binary science
  • Constituent papers: tracking-replications, replication-database, replications-reversals-social-science
• 4. Bitter: How ML people learn from data (machine learning)
  • Background: ML workflow
  • Background: The deep learning revolutions
  • Questionable research practices in ML
  • Constituent papers: questionable-practices-ml, ten-hard-problems, decision-trees-misspecification, ilp-safety
• Position paper (outro)
• Conclusion

Contributions

The thesis draws on the following published work. Bullet points denote Gavin’s contributions; * denotes equal authorship.

As first or senior (last) author:

1. Mask wearing in community settings reduces SARS-CoV-2 transmission (2022, PNAS) — Gavin Leech*, Charlie Rogers-Smith*, et al. → mask-wearing.md
   • Noted the global pre-mandate voluntary spike in mask wearing
   • Identified a ubiquitous methodological mistake (the mandate timing proxy)
   • First international analysis of masks with a random sample of self-reported mask-wearing data; constructed and validated a global dataset from multiple sources
   • First Bayesian hierarchical model for mask wearing, adapting an existing NPI model
   • Functional-form modelling of mask-wearing effects
   • Lead writer
2. Ten Hard Problems in Artificial Intelligence We Must Get Right (2024, in review at ACM Computing Surveys) → ten-hard-problems.md
   • Lead writer; sole author on the capabilities, alignment, opportunities, access, and meaning sections
   • Characterised early deep learning and the ‘large scale era’
   • 10 diagrams and informal models (AI governance field; two decompositions of the alignment problem; the family tree of AlexNet and GPT-2)
3. Tracking replications in the social, cognitive, and behavioural sciences (2024, in review at Nature Human Behaviour) → tracking-replications.md
   • Initiated the project alone with 53 famous effects that fail to replicate
   • Helped organise crowdsourcing of the full dataset of 1932 experiments
   • Designed analyses of the resulting nonrandom sample
   • Sole writer on first draft
4. Questionable practices in machine learning (2024, in review at JMLR) → questionable-practices-ml.md
   • Taxonomy of questionable practices; collated published and unpublished examples
   • Generated possible solutions; literature review relating ML to metascience
   • Lead writer

As other author:

1. Inferring the effectiveness of government interventions against COVID-19 (2020, Science) → inferring-npi-effectiveness.md
   • Wrote most of the first draft; literature review of semi-mechanistic models; helped set the epidemiological priors; one among many authors on NPI data collection
2. How Robust are the Estimated Effects of Nonpharmaceutical Interventions against COVID-19? (2020, NeurIPS spotlight) → npi-robustness.md
   • Formalised model assumptions; worked on theorems 1 and 2; writing on final draft; diagrams (model variations)
3. Seasonal variation in SARS-CoV-2 transmission in temperate climates (2022, PLOS Computational Biology) → seasonal-covid.md
   • Insolation analysis; helped formalise the model; diagrams (key figure 2); wrote half of the first draft
4. Massively Parallel Reweighted Wake-Sleep (2023, UAI) → parallel-reweighted-wake-sleep.md
   • Sole implementation of an earlier branch of the project; editing on final draft
5. The Replication Database: Documenting the Replicability of Psychological Science (2024, Journal of Open Psychology Data) → replication-database.md
   • Initiated one of the source datasets; editing on final draft

Conclusion

The thesis’s de facto research questions and their answers:

Epidemiology.

• Why do studies of individual mask-wearing conflict with population-level studies of mask mandates? Because a key assumption of the mandate-timing method was violated: most of the wearing uptake was pre-mandate in most regions, so the binary ‘after mandate enforcement’ variable is a poor proxy for wearing levels — those studies are incoherent and to be deprecated.
• What did mass mask-wearing achieve during the first year of the Covid pandemic? Some evidence of a substantial ~25% [6%, 43%] reduction in $R_t$ for a fully-masked population (slightly lower in practice at 70–85% wearing). The estimate is time-bounded (pre-saturation of vaccination/immunity) and reflects median mask quality (mostly cloth or surgical masks).
• How does modelling seasonality change estimates of government-intervention effects? Even a single scalar (the amplitude of the annual sine wave in Europe) greatly improved model fit, implying seasonality is a significant but easily-estimated confounder. On 2020 data, transmission varies 42% [25%, 53%] from peak winter to peak summer.

Psychology.

• What could explain the replication crisis? Dozens of possible contributors; evidence given for each, but no synthesis or structural model.
• How do effect sizes change under replication? The nonrandom sample doesn’t admit a proper answer, but descriptively there was an average ‘shrinkage’ of $d = 0.34$ from originals to replications.
• Does crowdsourcing replications help? In the weak sense of being a tolerable way to collect information (independent double entry gives a <1% error rate); not justified as counteracting the replication crisis.
• Would foregrounding estimation over discovery help? Some pathologies can be seen as optimising a binary ‘discover a qualitative effect’ ($p<0.05$) objective; but little was done to justify that shifting to estimation would be better under equal adversarial pressure.

Machine learning.

• How has ML changed in the last decade? Extensively — deep learning has taken over large parts of the field. Besides the belated triumph of neural networks, six other huge methodological shifts: the scaling era; end-to-end learning; standard training frameworks; pretraining and transfer; prompting as research; and the privatisation and turn to secrecy.
• In what ways can an evaluation metric be misleading? In at least 43 ways.

As a whole. The thesis’s strength is that it collates work which led somewhere: the Covid papers were used in (at least) UK and Czech policy decisions; the psychology replication collection has some chance of being a standard reference with ongoing crowdsourcing; the QRP work has been used at Ofcom and as part of an evaluation checklist in at least one industrial lab. Its weakness is that it isn’t theoretically deep: the problems are not common to all three fields (besides underreporting, the ur-problem which hides all other problems). The original plan was to compare the evidential standards of the three fields and unify Bayesian inference, hypothesis testing, and statistical learning.