4  Interpretation methods ✎ Very rough draft

4.1 Interpretation of p-values

  • Intuitions for p-values
  • Alpha, power, diagnosticity refreshers
    • Distribution of unbiased p-values under null/alternative
  • p-hacking and its impact on p-values and effect sizes
  • Methods that aren’t useful for trustworthiness assessment but are good for intuitions: distribution of biased p-values, p-curve, z-curve
  • Too good to be true in the context of power vs significance rate
  • Peeking and trends towards significance

4.2 Misinterpretation of p-values

  • Non-significant p-values as evidence for the null
  • Difference between significant and non-significant is not itself significant
  • Simply claiming differences with non-significant p-values

4.3 Misinterpretation of specific tests

  • Analysis-claim misalignment in general; most of the below are a specific class of this
    • Any test that exists can and will be misaligned somewhere
    • Nature and magnitude of misalignment, e.g., CLPM lends itself to causal interpretation but can’t give it
    • Goals confusion gives rise to misalignment: description, prediction, influence
    • The demand for tests is understandable; “Hot Right Now” stats that are poorly understood as a general problem

4.3.1 Efficacy claims

  • Efficacy claims from anything other than a post-intervention comparison with a control group (pre-post, two interventions, etc.)

4.3.2 MANOVA

  • The “Protected F-test” logic: the belief that a significant omnibus MANOVA “protects” subsequent univariate ANOVAs from Type I error inflation

4.3.3 Mediation and PROCESS macro

  • Mediation
  • PROCESS macro — “That’s a lot to process”

4.3.4 CLPM and variants

  • Cross-lagged panel models and causal interpretation limitations

4.3.5 Bifactor models in SEM

  • Real-world data rarely perfectly follows proportionality constraints, so the less-restricted bifactor model will statistically outperform the higher-order model in nearly every CFA application, regardless of the “true” structure
  • Bifactor model is more robust to minor misspecifications because its direct paths to items allow it to “soak up” variance that a more constrained higher-order model cannot

4.3.6 ANOVA

  • Hidden multiplicity in ANOVA
  • Lack of power in higher order interactions (problems with both false positives and false negatives)

4.3.7 Other commonly misused methods

  • Stepwise regression
  • Post hoc power analysis
  • “Little Jiffy” Factor Analysis (PCA with Varimax & Kaiser Criterion of Eigenvalues > 1; default in SPSS)
  • Choosing covariates based on significance or univariate associations
  • Latent class growth models (note: very different from latent growth curve models)
  • Magnitude-Based Inference
  • ANCOVA applied to observational data — Lord’s paradox
  • Propensity Score Matching
  • Partial Least Squares SEM
  • Statistical confusion of assumption and inference / affirmation of the consequent in analyses

4.4 Easter egg research

  • Creating groups via clustering then testing for differences between them
  • Assuming differences between groups and then reifying the groups
  • K-means or Latent Class Analysis to create groups, followed by t-tests/ANOVA on the same variables to “validate” the groups
  • Analyses that assume the number of clusters and then discover them (e.g., Latent Profile Analysis)

4.5 Conditioning on post-treatment effects

  • Common cause problems

4.6 Causal claims, implications, and assumptions

  • Causality tests don’t exist, only causal assumptions
  • Causal assumptions do exist and should be clarified
  • Causal claims in non-experiments
  • Causal claims in experiments that aren’t experiments: pre-post change, follow-up comparisons against LOCF
  • The litmus test of “is this only interesting if it’s causal, yet it’s described as non-causal evidence”

4.7 Other interpretation issues

  • Confusing IV and DV
  • Table 2 fallacy