I think that the questionnaire should be adjusted ... Maybe the envisaged aspects of analysis could be more interesting emphasized. It is only a thought for you ...

The means are only part of the information within a variable. Do no forget that variability is also important. I do not know what you are measuring, but if you are making comparisons, comparing the means only is incomplete. Check the standard deviations . Also check the distribution (histograms) to se the shape. Two variables may have the same mean, but different variability, and different distribution shape. Good luck.

JT

Unless one is dealing with very polarising issues, means of approximately 3 seem quite normal. As Juan indicated, you should also take the variability into account in the comparisons. However, I want to raise another issue. Were the items meant to be analysed independently or do they form subscales (constructs)? If the latter is true, the analysis of the individual items may not be the optimal approach. This may not lead to a change in the appearance of the results, but I think that it is worth considering.

This is one reason why even-numbered likert-type scales are often used, as participants tend to "hedge" and cluster towards medial responses rather than identify with extremes. That said, there is no one-to-one mapping between any value and the responses on a linguistic scale (even if responses in the questionnaires have numbers above or below each possible response on every question). On the one hand, this makes statistical methods more difficult as any assignment of value is always to some extend arbitrary. On the other, you are not bound by means and variability (the latter is typically derived from the former and neither are, in general, robust measures). Depending upon how your questionnaire was designed (how do questions relate? are some intended to predict or correspond to the responses of others? etc.), you can use different measures. There are a number of more robust measures that use the median, not the mean, as although the median alone tends to be less informative, with the addition of e.g., quartiles or trimming it is often a better measure of location. Additionally, even with the mean you can use transformations/mappings from the 5-point set to a smaller set (using e.g., triangular fuzzy numbers/membership functions) that take into consideration the imprecise, fuzzy "value" of linguistic responses and which can be used to categorize/group responses more accurately than a seemingly more precise method such as exact numerical assignments.

I have encountered this situation a number of times so I had the opportunity to discuss it with my peers in the university. First we defined a Likert sclae and majority of my colleagues agreed that the numerals in the Likert scale are not ratio/interval but nominal. In such scale, the participants are not given the option to answer other than the integers. We therefore agreed that the central measure must not have have decimals, and therefore must be rounded off to whole numbers. Rounding them off to whole number is indispensable because the qualitative descriptions have only equivalent whole numbers and not dong so would mean creating ranges (with decimals) for these qualitative descriptions, which is inconsistent to the original equivalent integers.

I hope this helps Tamer. Ed