I am doing CFA in lavaan (3-factor model, all factors correlated). The fit of this model is quite bad and I wanted to improve it. I realized, that the full questionnaire contains four negatively-keyed items, and that they are all loaded under one factor (F3). So I wondered if the model fit might be bad because of the potential influence of this "method factor." I decided to allow residuals of these items to correlate with each other. The model fit improved immediately, but factor loadings of the items in F3 are really low now (0.2-0.5).
Could anyone tell, why this happened? Does it mean, that the respecified model is bad, or does it confirm the hypothesis, that there is some bias due to presence of negatively-keyed items?