技能 数据科学 认知理论与模型构建指南

认知理论与模型构建指南

v20260724
cogpsych-theory-and-hypotheses
本指南指导研究人员如何将抽象的认知理论转化为严格的、可证伪的、具有解释性的形式化模型。它强调了在实验设计中,如何构建模型、定义参数意义,并推导出能够区分自身理论与竞争理论的预测,适用于高水平的认知心理学研究。
获取技能
399 次下载
概览

Theory, Models & Hypotheses (cogpsych-theory-and-hypotheses)

Cognitive Psychology rewards a formal account of a cognitive process — a computational or mathematical model whose parameters have interpretable meaning and whose predictions can be fit to data and compared against rival models. The cardinal move here is to turn a verbal theory into a model that makes the experiments discriminating, and to separate predicted (confirmatory) from discovered (exploratory) results.

When to trigger

  • Specifying the theory and the formal/computational model that the experiments will test
  • Deriving the predictions that separate your account from rival models
  • Co-designing the model with the experiments (iterate with cogpsych-study-design)
  • A reviewer said the work is "atheoretical," "the model is just a curve fit," or "your data don't distinguish the accounts"

Build the theory-and-model

  1. State the cognitive theory. What mechanism or representation explains the phenomenon, and why — in words, before equations. Name the rival accounts you intend to adjudicate.
  2. Formalize it. Write the model: its representations, processes, free parameters, and what each parameter means psychologically. A model whose parameters lack interpretation is a red flag here.
  3. Name the rival model(s). Specify the competing account(s) in the same formal language so the comparison is fair (nested or matched-flexibility where possible).
  4. Derive discriminating predictions. Identify the data pattern that the models predict differently — that qualitative or quantitative signature is what your experiments must produce.
  5. Mark prediction status. Separate confirmatory (pre-committed/preregistered) predictions from exploratory model exploration done after seeing data; do not present a post hoc fit as predicted.
  6. State what would disconfirm the model. Which data pattern, or which parameter estimate, would count against your account — this is what makes the model a theory, not a fitting exercise.

Avoiding the "just a curve fit" objection

  • A model that fits anything explains nothing. Show the model is falsifiable (some data it cannot produce) and identifiable (its parameters can be recovered — handoff to cogpsych-data-analysis).
  • Prefer qualitative signatures that one model predicts and the other forbids over a small numerical edge in fit; reviewers trust a crossed prediction more than a smaller AIC.

Worked micro-example — theory to discriminating prediction (illustrative)

A recognition-memory program adjudicating two models, written so prediction status is legible.

Theory:  Recognition reflects a single continuous memory-strength signal;
         the unequal-variance signal-detection (UVSD) model formalizes it.
Rival:   A dual-process account adds a threshold recollection process (DPSD).
Formalization:
         UVSD parameters: d', sigma(old). DPSD parameters: R (recollection),
         d' (familiarity). Both fit the same confidence-ROC data.
Discriminating prediction (confirmatory, preregistered, Exps 1-3):
         The z-ROC slope is < 1 and *linear* under UVSD; DPSD predicts a
         characteristic U-shaped/curved z-ROC. The shape, not the fit index,
         separates them.
Exploratory: any post hoc parameter that improves DPSD fit is reported as
         exploratory, not as a prediction.
Disconfirming: a reliably curved z-ROC across experiments counts against UVSD,
         stated up front.

Theory-stage reviewer pushback and the venue fix

Reviewer pushback Cognitive Psychology fix
"Atheoretical / mechanism unclear" state the mechanism in words, then give the formal model before the experiments
"The model is just a curve fit" show a falsifiable, identifiable model with a crossed qualitative prediction, not only a fit edge
"Your data can't distinguish the accounts" design the discriminating signature into the experiments; formalize both rivals in the same language
"Parameters are uninterpretable" give each free parameter a psychological meaning and a recovery check
"This looks post hoc" mark confirmatory vs. exploratory; pre-commit the model comparison where feasible

Theory calibration anchors

  • The contribution is the model-as-theory, not the experiments alone; experiments earn their place by discriminating models, and the model earns its place by being falsifiable and identifiable.
  • A crossed qualitative prediction (one model predicts a pattern the other forbids) is worth more than a marginal fit advantage; lead with it.
  • Pre-commit the model space and the comparison criteria before fitting where you can; deciding the winning model after seeing the fits is the modeling form of HARKing.
  • Match model flexibility when comparing — a more flexible model that fits better may simply be overfitting; this is why parameter recovery and model recovery matter (cogpsych-data-analysis).

Anti-patterns

  • A verbal theory with no formal model where the phenomenon is plainly formalizable
  • A model with uninterpretable parameters or that cannot fail to fit
  • Comparing models of unequal flexibility without acknowledging it
  • Presenting a post hoc model selection as a predicted result
  • No statement of which data or parameter estimate would disconfirm the account

Output format

【Theory】the mechanism/representation, briefly
【Model】formalization: parameters + their psychological meaning
【Rival(s)】competing account(s) in matched formal language
【Discriminating prediction】the signature that separates the models
【Status】confirmatory (pre-committed) vs exploratory
【Disconfirming evidence】what would count against the model
【Next】cogpsych-literature-positioning

Supplementary resources

信息
Category 数据科学
Name cogpsych-theory-and-hypotheses
版本 v20260724
大小 6.46KB
更新时间 2026-07-28
语言