MH4: APPLICATION OF THE RASCH MODEL TO THE SF36 MENTAL HEALTH 5 ITEM SCALE (MH5)

Abstract

OBJECTIVES: To verify whether the SF36 Mental Health 5 item scale (MH5), an instrument commonly used in mental health research, fulfills criteria of unidimensionality and invariant measurement specified by item response theory (IRT). METHODS: As part of a survey of health care needs among university students in Geneva, Switzerland, 1257 respondents (64% of eligible persons) filled the MH5 (version 2). Each item was scored on a 5-level frequency scale (“never” to “all of the time”). We analyzed these data using both the traditional method (summative scoring), and the polytomous one-parameter IRT (or Rasch) model. RESULTS: In traditional analysis, the MH5 scale performed as expected (single factor, Cronbach alpha 0.85, mean 67.1, standard deviation 17.6, range 0–100). Rasch analysis revealed the good fit of all item characteristic curves. Threshold locations for feeling “nervous” (mental health logits: −3.2, −0.7, 2.0, 3.3), “down in the dumps” (−4.0, −1.7, 0.4, 1.5), “calm and peaceful” (−3.0, −0.6, 0.9, 6.0), “downhearted and blue” (−4.2, −1.6, 1.0, 3.0), and “happy” (−3.1, −0.9, 0.7, 4.2) were all ordinal, consistently with theory, and spread widely over a span of 10.2 logits. While classic MH5 scores and Rasch-based scores were closely correlated (r = 0.98), their relationship was S-shaped: the intervals, in mental health logits, between MH5 scores 0–10, 10–20, etc, up to 90–100 were: 2.23, 1.22, 1.00, 0.90, 0.85, 0.83, 0.90, 1.09, 1.54, and 3.21. CONCLUSIONS: The five MH5 items fit the Rasch model very well. Classic summative scoring tends to compress the measurement scale at its extremities, by assigning scores that are too low at the high end, and too high at the low end. Rasch scoring may render MH5 more sensitive to changes among well populations and among the very sick.