Planned contrasts

Description

Tests a small set of theory-driven comparisons among group means with contrast weights summing to zero, instead of an omnibus F followed by post hoc tests.

Each research question is written as a contrast: a weighted sum of group means whose coefficients sum to zero, with zero for groups not involved. 'New design versus the average of the two old ones' becomes weights of 1, -0.5 and -0.5. The contrast is tested with a t or F using the within-groups error, and reported with a confidence interval and standardised effect size. With k groups at most k - 1 linearly independent contrasts exist; with equal group sizes two contrasts are orthogonal when the cross-products of their coefficients sum to zero. Planned contrasts need no significant omnibus F and are more powerful than post hoc tests, which dilute power when ineffective groups are added. Contrasts are not robust to unequal variances even with equal n, so a separate-variance t is used when variances may differ, and the family of contrasts needs a decision about error-rate control.

Attributes

Loss type
Latent
Stage
Evaluation
Effort
Hours

Steps

  1. Before seeing data, write each question as contrast coefficients that sum to zero, with zero for groups not involved.
  2. Check the set: at most k - 1 contrasts, and with equal n confirm orthogonality by checking that coefficient cross-products sum to zero.
  3. Estimate each contrast value and its standard error from group means and the within-groups variance.
  4. Test with t or F; use a separate-variance t if group variances may differ.
  5. Report the confidence interval and a standardised effect size for each contrast.
  6. Decide the family of contrasts and apply multiple-comparison control where several are tested.

Sources

  • Designing Experiments and Analyzing Data (Maxwell et al., 2018)
  • Statistical Reasoning in the Behavioral Sciences (King et al., 2011)
  • Quantitative Psychological Research (Clark-Carter, 2004)
  • Design and Analysis of Experiments, Vol. 1 (Hinkelmann & Kempthorne, 2008)
  • Psychology Statistics for Dummies (Hanna & Dempster, 2012)

Also known as

  • Planned contrasts (pairwise and complex comparisons)
  • Contrast analysis
  • Individual comparisons of means
  • Complex comparisons
  • Planned comparisons
  • A priori comparisons
  • Orthogonal contrasts

Categories

  • Analytics & Behaviour
  • Statistics & Modelling