I'm fitting a mixed model using Standard Least Squares (REML): Y ~ A(Sum) + B(Sum) + C(Sum) + D(Sum) + Subject(&Random)
- ~14,000 obs, ~200 levels of factor A, ~60 levels of C, ~120 subjects
- Highly sparse A×C design: ~97% cells empty (most A-levels appear in only 1 C-level)
When I replicate in Python (statsmodels MixedLM, REML), all 200 LS Means for factor A differ from JMP by an exactly constant offset (0.22). Relative effects and rankings are 100% identical (std of pairwise diff = 0.0004).
Other findings: For a different response variable on the same data with the same model structure (same X matrix), the offset is only 0.006. The problematic response has higher ICC (0.36 vs 0.27).
Questions:
- Does JMP SLS compute the intercept (or LS Means) differently from a standard V⁻¹-weighted GLS when ICC is moderate-to-high in a sparse design?
- Is there a known algorithmic difference between SLS (REML) and the Mixed Model personality for this scenario?
JMP Pro 17, Windows.
