Question Taxonomy

Question Taxonomy

Overview

The type of research question determines the appropriate analytical approach. Misclassifying a question leads to methods that don't answer what you actually want to know.

Primary Question Types

1. Prediction

"Can we accurately forecast outcome Y given information X?"

• Goal: Maximize predictive accuracy for new, unseen cases
• Focus: Model performance (discrimination, calibration)
• Coefficients: Not the primary interest; may be uninterpretable
• Causation: Not claimed or required

Examples:

• "Can we predict ICU mortality at admission?"
• "Which patients will develop AKI in the next 24 hours?"
• "What is the risk of readmission for this patient?"

Appropriate methods: Logistic regression, random forest, gradient boosting, neural networks, penalized regression

Key considerations:

• Features must be available at prediction time (no future information leakage)
• Validation on truly held-out data is essential
• Calibration matters for clinical use, not just discrimination

2. Inference (Association)

"Is X associated with Y, accounting for confounders?"

• Goal: Estimate the relationship between exposure and outcome
• Focus: Effect size, confidence interval, statistical significance
• Coefficients: Primary interest; must be interpretable
• Causation: Claimed only with strong design (RCT, quasi-experimental)

Examples:

• "Is vasopressor choice associated with mortality?"
• "Do patients with diabetes have longer ICU stays?"
• "Is early mobilization associated with reduced delirium?"

Appropriate methods: Linear/logistic regression, Cox regression, GEE, mixed models

Key considerations:

• Confounder identification and adjustment
• Model specification (correct functional form)
• Effect modification / interaction

3. Causal Inference

"Does X cause Y?"

• Goal: Estimate the causal effect of an intervention
• Focus: What would happen if we changed X?
• Requires: Strong assumptions (exchangeability, positivity, consistency)
• Gold standard: Randomized controlled trial

Examples:

• "Does early intubation reduce mortality compared to delayed intubation?"
• "What is the effect of a restrictive transfusion strategy on outcomes?"

Appropriate methods: IPTW, propensity score matching, instrumental variables, difference-in-differences, regression discontinuity

Key considerations:

• Causal assumptions must be explicitly stated and defended
• Unmeasured confounding is always a threat in observational data
• Sensitivity analyses for violations are essential

4. Description

"What is the distribution/frequency of X?"

• Goal: Characterize a population or phenomenon
• Focus: Summary statistics, patterns, trends
• No exposure-outcome relationship tested

Examples:

• "What is the 30-day mortality rate in sepsis patients?"
• "How has ICU length of stay changed over time?"
• "What are the most common diagnoses in our cohort?"

Appropriate methods: Descriptive statistics, data visualization, trend analysis

Key considerations:

• Clearly define the population
• Report variability (SD, IQR, range), not just central tendency
• Consider selection bias in who enters the cohort

5. Clustering / Subgroup Discovery

"Are there natural subgroups within this population?"

• Goal: Identify latent structure or phenotypes
• Focus: Group membership, cluster characteristics
• Unsupervised: No predefined outcome

Examples:

• "Are there distinct sepsis phenotypes?"
• "Can we identify patient subgroups with different treatment responses?"

Appropriate methods: K-means, hierarchical clustering, latent class analysis, Gaussian mixture models

Key considerations:

• Cluster validity (internal and external)
• Clinical interpretability of clusters
• Stability across samples

Secondary Distinctions

Exploratory vs. Confirmatory

Cross-sectional vs. Longitudinal

Time-to-Event vs. Binary/Continuous

Decision Aid: What Type is My Question?

START: What do you want to learn?
│
├─ "How well can we forecast Y?"
│   └─ → PREDICTION
│
├─ "Is X related to Y?"
│   ├─ "...and I want to claim X causes Y"
│   │   └─ → CAUSAL INFERENCE (requires strong assumptions)
│   └─ "...adjusting for confounders"
│       └─ → INFERENCE (association)
│
├─ "What does Y look like in this population?"
│   └─ → DESCRIPTION
│
└─ "Are there natural groups in my data?"
    └─ → CLUSTERING

Common Misclassifications

Prediction disguised as inference

• Red flag: "We found that feature X has high importance in predicting Y, therefore X is associated with Y"
• Problem: Variable importance ≠ causal effect; prediction models optimize for accuracy, not interpretability

Association claimed as causation

• Red flag: "Patients who received treatment X had lower mortality, so X reduces mortality"
• Problem: Without randomization or causal methods, confounding by indication is likely

Description without context

• Red flag: "Mortality was 25% in our cohort"
• Problem: Without comparison group or benchmark, the number is uninterpretable

Implications for Analysis Planning