Research 1 of 3
Priors in the loop
Context
Human-in-the-loop systems rarely behave like textbook Bayesians: feedback is delayed, confounded, and easy to misread. This line of work asks how interface design shapes what people treat as evidence and how quickly beliefs should move.
Hypothesis
Under sparse, noisy feedback, operators overweight recent signals and underweight base rates unless the interface makes uncertainty and sample size salient.
Result
Calmer analytics copy—explicit intervals, likelihood language, and gentle priors in the UI—reduces over-confidence without slowing expert workflows in pilot reviews.
Impact
Informs calmer analytics surfaces across projects: fewer false alarms, a clearer sense of sample size, and copy that resists over-confidence when the model is uncertain.
Concept
Bayes in one line. Let prior odds be \(O_0 = P(H) / P(\neg H)\). After observing data \(D\), posterior odds multiply by the likelihood ratio.
When the ratio \(P(D \mid H) / P(D \mid \neg H)\) is close to \(1\), beliefs should barely move—strong updates require strong evidence or many independent pieces of data.
- Role
- Research, writing
- Year
- 2026
- Focus
- Decision under uncertainty
- Methods
- Literature synthesis, small-n studies
- Stack
-
-
Research 2 of 3
Linearity and limits
Context
Teams often reach for regression because it is legible, not because the world is linear. This thread documents where linear summaries help, where they mislead, and how to communicate curvature and heterogeneity without drowning readers in math.
Hypothesis
Stakeholders routinely read a single OLS slope as causal when interactions and confounding are present but invisible in aggregate charts.
Result
Pairing slopes with residual views, pre-declared scope, and one sensitivity scenario per slide measurably reduced over-interpretation in design critiques.
Impact
Better default for stakeholder-facing charts: explicit scope, residual checks called out, and language that separates prediction from explanation—reusable in decks and specs.
Concept
Ordinary least squares. With responses \(y_i\) and predictors \(x_i\), OLS chooses coefficients that minimize squared error:
The estimated slope \(\hat{\beta}_1\) describes a conditional association in the fitted model. It is a causal effect only under identification assumptions (e.g. no unmeasured confounding) that the regression line alone does not guarantee.
- Role
- Analysis, visualization
- Year
- 2025
- Focus
- Applied statistics
- Methods
- Simulation, robustness checks
- Stack
-
-
Research 3 of 3
Signals in the noise
Context
Surveys, logs, and experiments all inherit measurement error. This note collects heuristics for judging whether a “significant” bump is plausible given base rates, \(n\), and how many hypotheses were quietly tested along the way.
Hypothesis
With moderate \(n\) and many implicit comparisons, nominal \(p \lt 0.05\) thresholds produce a high rate of actionable false positives for product teams.
Result
Workshops that led with effect sizes, intervals, and a simple FDR metaphor cut “ship it” decisions on weak lifts in post-experiment reviews.
Impact
A shared vocabulary for partners: false-discovery risk, effect sizes, and confidence intervals as honest summaries of what the data can support—not just whether a bar “won.”
Concept
Type I and Type II error. Let \(\alpha\) be the chance of rejecting a true null (false alarm), and \(\beta\) the chance of missing a real effect. Power is \(1 - \beta\).
Tightening \(\alpha\) reduces false alarms but lowers power unless \(n\) or effect size increases—so inference is always a deliberate tradeoff, not a checkbox.
- Role
- Methods notes
- Year
- 2024
- Focus
- Probability & inference
- Methods
- Analytical bounds, examples
- Stack
-
-
