Analysis Design
Reference for the design-phase decisions that precede model specification: what is this analysis trying to do, how will it be evaluated, what domain conventions apply, and what structural questions about the data-generating process should the candidate models address.
These are the choices a competent statistician makes once per analysis, before
any individual model is specified. For specifying generative models within a
structural question, see generative-model-design.
Key Practices
- Assume loudly. When the user's prompt lacks a specific question, synthesize a purpose from the data's nature and state it explicitly. A sharp answer to an assumed question beats a generic answer to no question.
- Validation discipline. Standard observation-level LOO is only appropriate for i.i.d. data. For grouped data it interpolates within known groups; for temporal data it overstates future predictive performance. Pick the hold-out scheme that matches the prediction target.
- Domain over generic. Do not default to generic GLMs when the domain has mechanistic structure — use the domain's standard response model.
Analysis purpose
Determine the goal type from the user's prompt and the data's nature:
- Descriptive (default for minimal prompts). Characterize the data-generating process with honest uncertainty.
- Inferential. Estimate a specific causal or conditional effect. Prioritize unconfounded estimation of target parameters; do not add flexible structures that absorb the estimand's signal.
- Predictive. Maximize out-of-sample performance; parameter interpretability is secondary.
Define 1-3 key quantities of interest and state what adequate means — when is the model good enough to stop?
Validation strategy
Choose the hold-out scheme based on the data's dependence structure:
| Data structure | Standard LOO valid? | Recommended hold-out | |---|---|---| | i.i.d. | Yes | Standard observation-level LOO | | Grouped (predicting new obs in known groups) | Yes | Standard LOO | | Grouped (generalizing to new groups) | No | Leave-out-group CV (entire groups out) | | Temporal | No (overstates predictive performance) | Leave-future-out (LFO-CV) or rolling-origin |
State the chosen scheme up front. Downstream critique and selection trust this declaration when deciding whether ELPD rankings are interpretable.
Domain context
Identify the scientific domain and its canonical modeling frameworks. Different domains have different standard response models, parameterizations, and baseline components — e.g., pharmacokinetics defaults to ODE compartmental models, psychophysics to psychometric functions, epidemiology to SIR/SEIR. Use the domain's canonical framework as a starting point rather than defaulting to a generic GLM.
If no strong domain conventions are recognizable from training or the EDA, state so explicitly and proceed with empirical-first design.
Structural questions
Extract 2-3 contrastive structural questions — each pits two explanations of the data-generating process against each other, framed in terms of the key quantities of interest and domain theory where possible.
Good question shape:
- "Does response variance differ by group (hierarchical) or is it pooled?"
- "Does the temporal trend follow domain X's mechanistic ODE, or is a flexible spline sufficient?"
- "Is the apparent zero-inflation a true two-stage process, or thin-tailed observation noise?"
Bad question shape:
- "What model fits best?" (not contrastive)
- "Should we use Student-t?" (just a likelihood swap, not structural)
Each question should be answerable by comparing models with different structural commitments.
Output
The analysis design produces four sections — analysis purpose (with key quantities of interest), validation strategy, domain context, structural questions — that downstream model specification depends on. Be precise; vague analysis design propagates as vague modeling.