blog – Examined Algorithms

Measurement, Latent Factors and the Garden of Forking Paths

Confirmatory Factor Analysis and Structural Equations in PyMC

cfa

sem

measurment

Choices are made as we step down the garden of forking paths in the course of any analysis. Often, in the process of frantic search we cling too strongly to the first…

Prompting the Void: On Sustained Attention and Goal Discovery

A case study in iterative refinement

AI workflow

writing

data science

collaboration

AI-assisted work is often sold on the generative possibilities. The quality comes from diagnostic reflection and iterative refinement. This essay tracks the development of a technical post through 21 commits over 15 days, recording every prompt, every correction, and every modelling decision. The pattern that emerged was consistent: first drafts were competent but inert, and every useful advance began with a diagnosis the tool did not originate. The final product is stronger for the combination, but the process revealed something about the nature of the work too. The goal of the project was discovered through the iteration and refined. The exploratory phase of any project is improved by sustained attention. Optimising for a known goal is premature when the goal itself is what the work discovers.

The Architecture of Focus: Attention in Transformers and Discrete Choice

How Transformer Attention and Econometric ‘Consideration Sets’ share a mathematical skeleton

NLP

discrete choice

econometrics

deep learning

PyMC

Every act of attention is an act of exclusion. To focus on one thing is to ignore another. When the mechanism operates invisibly in systems that make consequential decisions…

Condorcet Juries and Incentivised Legibility

Sensitivity Analysis with Item Responses

theory

simulation

causal inference

sensitivity analysis

irt

You don’t have 30 employees. You have three, rehired ten times. This is a paradox of corporate organisation.

The Journey Is the Model: Dynamic Path Analysis in PyMC

Causal Inference with Time-Varying Effects

path-models

sem

causal inference

If you look to Odysseus on the morning the gates of Troy fell, he is well set up for a happy journey home. He is the architect of victory, his ships are loaded with spoils…

Heuristics in Latent Space: VAEs and Bayesian Inference

pytorch

variational auto-encoders

amortized bayesian inference

Missing data imposes two costs: the expense of collecting new observations and the risk of distorting the underlying structure when we construct plausible replacements. In job satisfaction surveys, non-response patterns can conceal meaningful latent relationships. This study compares statistical and machine learning approaches for imputation, focusing on their ability to preserve these latent structures while producing realistic reconstructions.

Freedom, Hierarchies and Confounded Estimates

TWFE

mundlak

hierarchical models

DiD

Two Way Fixed Effects (TWFE) regression models are often used in Differences-in-Differences designs to estimate treatment effects while accounting for variation due to group…

GAMs and GPs: Flexibility and Calibration

probability

generalised additive models

gaussian processes

Flexible Spline models risk overfit and need to be carefully calibrated against real data. Hierarchical models impose structure and aid calibration. We combine them.