Abstract
<p>Background. PISA 2022 introduced seven scales measuring social and emotional skills. Each has ten items of mixed wording, and each student answers five. The OECD withholds a country's score when any of that country's country-by-language groups has a Cronbach's alpha below 0.60; across the seven scales the rule removed 158 of 513 administered country-by-scale indices. Whether the discarded information was recoverable, and on whom the loss falls, has not been examined.Methods. The OECD scaling and exclusion pipeline was reimplemented on the public-use file and validated against published item parameters, published group alphas and the individual suppression codes. Country-by-language groups were reconstructed with the OECD's own rule, giving 621 group-by-scale units. For every suppressed cell we computed the alpha a same-valence half-scale would have reached in every language group of that country, corrected for estimator bias and bounded by Feldt confidence intervals. This was compared with respondent screening, with a simulation containing no careless respondents, and with how each half's criterion validity varies with the country's rate of extreme invariant responding.Results. In 155 of the 158 suppressed cells the positively worded half alone would have reached 0.60 in every country-by-language group, 150 using the lower bound of the 95% interval. Respondent screening is not equivalent: discarding 15.5% of students recovers 87 cells and discarding 42.3% recovers 122; the positively worded half recovers 155 discarding none. A simulation without careless respondents reproduces the pattern: administered-form alpha falls from 0.775 to 0.511 while the positively worded half holds at 0.795. It also predicts an inverted U in the number of reverse-worded items, which the data confirm. In all seven scales it is the reverse-worded half whose relation to reading achievement weakens as extreme invariant responding rises.Conclusions. The threshold discards information a same-valence scale would have retained, and discards it unequally: the number of scales a country retains correlates 0.866 with its mean reading score, the lowest-performing quarter of countries retaining 1.47 of seven and the highest 6.95. Reporting decisions for mixed-worded scales should be evaluated against what a same-valence subscale would deliver, and the resulting missingness treated as informative.</p>