MerQuR+Statistics Glossary

Statistics Glossary → Fisher's LSD (unadjusted pairwise comparisons)

Fisher's LSD (unadjusted pairwise comparisons)concept

A post-hoc method that compares pairs of groups after ANOVA with t tests using the pooled error variance, without any multiple-comparison adjustment.

In the LSD (least significant difference) method, each pair of groups gives t = (ȳᵢ - ȳⱼ) / √(MSE·(1/nᵢ + 1/nⱼ)), where MSE is the error mean square from the ANOVA and the degrees of freedom are the error degrees of freedom. In Fisher's 'protected' approach, pairwise comparisons are carried out only if the overall F test is significant. This protection controls the familywise error when all means are equal.

But when some means are equal and others differ, the protection weakens, and with four or more groups the familywise error rate can clearly exceed α; with exactly three groups, protected LSD does control it. LSD is the most powerful of the methods but also the least conservative; with many groups, adjusted methods such as Tukey, Holm or Bonferroni are preferred. The report should state clearly that unadjusted comparisons were used.

An everyday example

A bakery compares liking scores for three recipes (A, B, C) and the ANOVA is significant. LSD shows that the 0.8-point difference between A and B is significant. Had ten recipes been compared instead of three, there would be 45 pairwise comparisons, and without adjustment the chance of finding at least one 'significant' difference by luck alone would be very high.

In MerQur

Available as the LSD choice of the 'Post-hoc test' option in one-way ANOVA, two-way ANOVA and ANCOVA; LSD (unadjusted p) is also among the choices of 'EMM / simple-effect adjustment' in two-way and mixed-design ANOVA and of 'Pairwise comparisons' and the multi-factor adjustment options in repeated measures ANOVA. In the nested mixed model it appears under 'Upper-group comparisons' and in the crossed mixed model under 'Fixed-factor comparisons' as 'Unadjusted (LSD)'.

Analyses where this concept appears

One-Way ANOVATwo-Way ANOVARepeated Measures ANOVAMixed-Design ANOVA, Split-PlotANCOVANested Mixed Model (Nested LMM, Multi-Level Hierarchical Mixed Model)Crossed Mixed Model (Crossed LMM, Crossed Random Effects Mixed Model)

Related concepts

Post hoc tests (pairwise comparisons)Multiple comparisons and the familywise error rate (FWER)