Abstract: Abstract
In personality measurement, some examinees show the tendency of simply agreeing with the scale items regardless of the item content. This tendency is called ‘acquiescence’ by Cronbach (1946, 1950), and ‘yea- or nay-saying’ by Couch and Keniston (1960). To diminish this kind of response bias, negatively worded items are adopted. Negatively worded items require more attention to item content, and also can balance the impact of yea- or nay-saying on overall test scores.
But studies on the change of item narrative direction have provided inconsistent findings. Whereas some studies show that the use of negatively worded items can increase reliability and validity (Sandoval and Lambert, 1978; Schriesheim and Hill, 1981; Williams et al., 2001), others show that the use of negatively worded items can confound factor structures (Weems et al., 2001) and may result in a separate factor ‘irrelevant to the trait being measured’ (Ibrahim, 2001). When both positively and negatively worded items are used to measure a uni-dimensional trait, factor analysis will exact two factors, one is related to positively worded items and the other is related to negatively worded items.
In this study, Chinese EPQ and NEO-FFI were used as personality measures, but items narration directions were changed in both measures. In one form (formⅠ) of EPQ(or NEO-FFI), half of items were positively worded, and half of the items negatively worded. In another form (formⅡ) of EPQ(or NEO-FFI), positive items in formⅠwere converted into negative items, and negative items into positive ones. These two test forms of EPQ(or NEO-FFI) were administered to auniversity student group with the time interval of two weeks. Thus we obtained the data from 363 university students who took positively worded EPQ and negatively worded EPQ, and the data from 412 students who took positively worded NEO-FFI and negatively worded NEO-FFI.
The result showed that positively worded sub-scales of EPQ and NEO-FFI were a little more reliable than their negative counterparts, α coefficients for three positive EPQ sub-scales (E, P, N) were 0.80, 0.42, 0.75, for three negative EPQ sub-scales were 0.82, 0.45, 0.81; α coefficients for five positive NEO-FFI sub-scales (N ,E, O, A, C) are 0.82, 0.65, 0.69, 0.64, 0.75, and for five negative NEO-FFI sub-scales were 0.77, 0.62, 0.68, 0.64, 0.67.
Confirmatory Factor Analyses showed that negative EPQ and NEO-FFI exhibited better fit than the positive ones: for negative EPQ, model c2=2456.22, RMSEA=0.046, NNFI=0.87, SRMR=0.069, for positive EPQ c2=2402.96, RMSEA=0.045, NNFI=0.85, SRMR=0.072; for negative NEO-FFI c2=2868.20, RMSEA=0.066, NNFI=0.84, SRMR=0.078, for positive NEO-FFI c2=3952.15, RMSEA=0.075, NNFI=0.79, SRMR=0.086.
Based on classmate nominations as criterions, sub-scales of negative EPQ and NEO-FFI showed higher correlations with criterions than positive ones. When all positive and negative items of EPQ(or NEO-FFI) factor analyzed in one CFA model, the fit measures did not become significantly lower than the positive or negative item model, and loadings of the positive and negative items were highly consistent. These results suggest that positive and negative forms of an item measure the same construct.
In conclusion, positive worded items may produce higher reliability, but lower construct validity and criterion-related validity than negative worded ones do. Negatively and positively worded items measure the same construct, but positively worded items may be more affected by response set bias

Key words: positively and negatively worded items, response bias, EPQ, NEO-FFI