Functional Data Analysis

Description

Treats each densely sampled record, such as a cursor path or continuous rating, as a smooth curve and analyses curves and their derivatives with functional means, registration, PCA and regression.

Many UX measurements are curves: cursor and stylus trajectories, continuous real-time ratings, pupil and other physiological traces, daily usage over a trial. Functional data analysis represents each record as a smooth function built from basis functions (B-splines are the flexible default) with a roughness penalty, rather than as many separate points. Smoothness makes derivatives estimable, and using velocity or acceleration is what sets the approach apart: differencing raw data would magnify noise. Data resolution limits which derivatives can be trusted. Curves vary in amplitude and in phase (the timing of features), so registration aligns landmarks by smooth time-warping before averaging; otherwise the mean curve blurs features nobody actually showed. Further tools include functional principal components to find the main modes of variation, functional correlation and regression relating curves to each other or to scalar outcomes, and differential-equation models of the dynamics.

Attributes

Loss type
Unmigrated
Stage
Evaluation
Effort
Weeks

Steps

  1. Check resolution: enough samples per curve feature and adequate signal-to-noise for the derivatives you plan to use (often only the first).
  2. Choose a basis (for example B-splines) and a roughness penalty, selecting smoothing by generalised cross-validation, and convert each record into a function.
  3. Register the curves to remove phase variation by aligning landmarks or by continuous time-warping.
  4. Compute functional means and variance, and derivatives such as velocity and acceleration.
  5. Apply functional PCA to find the main modes of variation, and functional regression to relate curves to design condition or to scalar outcomes.
  6. Report curve-level results with pointwise or simultaneous confidence bands, and check that conclusions survive alternative smoothing and registration choices.

Sources

  • The SAGE Handbook of Quantitative Methods in Psychology (Millsap & Maydeu-Olivares, 2009)
  • A User's Guide to Principal Components (Jackson, 1991)
  • Introduction to Nonparametric Regression (Takezawa, 2006)

Also known as

  • Curve data analysis
  • Functional principal components analysis
  • Curve registration
  • Functional regression

Categories

  • Analytics & Behaviour
  • Statistics & Modelling