Hello Beena Lawania: Negative factor loadings can occur if items are worded in the opposite direction to other items. Have you checked that out? It's also important to look at the strength of the loading. If the loading is less than, say, .30, I would definitely consider discarding that item.

Whether negative or positive,the factor loading above 0.5 can be considered for selecting the statements in a scale.

I would agree that having items with loadings greater than 5 can be good for the final version of a scale, but at the initial stage of item selection in factor analyses it could be good to retain some items with lower items in case they end up having higher loadings as some other items are discarded.

I would add that sometimes it's good to have items with loadings down in the .40s if they provide breadth to a construct. If items have only high loadings (all up around, say .70 and higher), there is a risk of redundancy among them and a very narrow construct being measured. Sometimes that could be desirable; sometimes not. Researchers need to consider the nature of the construct they intend to measure, including whether that construct is best considered to be broad or narrow.

Dear Beena Lawania ,

the previous replies are exactly to the point. I just want to add a little bit more context to their replies and provide you with some literature that might help you and that you can consider citing when you write it up.

First, in terms of the sign of a loading Jaspreet Kaur already pointed out that it is not the sign, but the loading that is important. Depending on the psychometric properties of your scale it can be beneficial to include reverse-coded items (i.e., low values should correlate with high values on the other items of the scale). This can have several reasons, for example to see whether participants actually think about the questions or simply "click through". There is interesting research by Rik Pieters out of Tilburg, Bert Weijters out of Ghent and Hans Baumgartner out of Penn State. They deal with a multitude of issues in item response theory, such as extreme response styles, mixing items of different scales, etc. If you go to their webpages you will find several articles that might be interesting to you. Second, it is not only the simple loading that is relevant for your measurement model assessment. Once you move on to structural equation modeling as a means of confirmatory measurement model quality assessment (I am assuming that you have reflective items) there are other issues to consider. You can read up on this in Joe Hair's excellent book "Multivariate Data Analysis". He covers the different logics in a way that is easy to grasp and geared towards applicability.

I hope this helps, please let me know.

Best,

Alex