# Reverse scoring when question is stated in a negative fashion

In my survey, I had a question that asked respondents why they did NOT engage in a particular behavior (x). Some of the answers were stated in a positive way, such as "because of age" while some were stated in a negative way "I did NOT know how to do x".

I am trying to reverse code because I am using these statements in a factor analysis and I would like the means that I compute to make sense (so that higher values indicate higher agreement with a positively stated statement). The problem is that when I run Cronbach's alpha test, the values I get make sense when I do NOT reverse code and they are negative when I DO reverse code.

I am wondering if it is because my question was stated in a negative way to start with ("why did you not engage in x behavior?")? Myself and my supervisor feel that in this case it doesn't really make sense to reverse code, but I'm wondering how that will affect the means that I calculate? I will be using these means subsequently in a logistic regression.

• What are the possible answers to the questions? Are they multiple choice? Are they nominal? Jun 27, 2013 at 20:20
• How do you create factors? With what kind of other factors/components are these responses to engaging in $x$ compared? What I am imagining now is that positive scores in a question 1 (e.g. did you engage in behaviour $x$) are gonna be compared with negative scores in a follow up question 1a (I did not lnow how to do $x$). Aug 27, 2018 at 7:21
• What Cronbach's alpha test are you using? You got a negative value (or you mean 'negative' as in a 'no presence')? Could you write out the equation that you used and what the data is that you put into it. Aug 27, 2018 at 7:33

You should probably first look at the loadings in the factor analysis before any recoding. If they are negative (or equivalently, if they are positive while positively-worded question have negative loadings), then reverse scoring makes sense. Otherwise, it means that this particular item was not understood in the way you expected. In any case, you don't need to recode anything before factor analysis and the results will help you decide what makes sense or not.

Reverse scoring an item loading negatively is, however, essential to compute alpha. Here as well, I would look at the result of the factor analysis to decide what to recode. I don't think that arbitrarily recoding items to obtain an attractive coefficient is appropriate. If you get some nonsensical alpha (e.g. negative), you might consider removing the item or forgetting about alpha entirely. Note that Cronbach alpha has many limitations and that small samples can produce negative values even for perfectly good scales.

Incidentally, simply taking the mean is not the only way to create a composite score.