Obviously: just reverse one of the two scales (or insert cross-references).

It is not clear what "numerical conversion" means...

No need to reverse one of them, since statistics such as the t-Test simply compare means -- they have no "knowledge" of the direction in which a scale is scored. For measures such as correlation, you will get a negative coefficient, which is just as meaningful as a positive one.

I'm all about 'brief' and therefore prefer David's solution, above.

Further to my last post. let's remember that in applying a parametric statistical test (like t-test) to ordinal Likert data, we are going with much expert opinion to get a 'good enough' or 'near as dammit' result; it's still not quantitative precision but it'll do.

It depends what you mean by 'compare'. If you want to do a correlation, to see whether there is a meaningful difference between the two data sets, you don't need to do any number conversion, because (as David says) all that will happen is that you get a negative correlation. So long as you INTERPRET this as a positive correlation, you'll be fine. Personally I would do a number conversion, by recoding one set of data with the same values as are used in the other (this is very straightforward in SPSS, and I imagine in other packages) as this is tidier and will mean you have one less complication to think about when interpreting your data.

If the word familiarity evaluation is done on two independent group of participant, with one group the scale rating familiarity from 1 (familiar) to 5 (unfamiliar) and the other group just reverse of it, then the values from these two groups cannot be compared directly. Use the technique suggested by Oluwaseun Adeyemo for conversion. On the other hand, if the word familiarity evaluation is based one group and certain words rated familiarity from 1 (familiar) to 5 (unfamiliar) and the other words from 1 (unfamiliar) to 5 (familiar), then the values can be compared directly.

I'm puzzled by the previous answers. Since both Likert answers are to the same question, just coded differently, of course you can compare them the same as Likert scales coded in the same manner. You just reverse the coding in one of the answers, so both are Likert responses with the same mapping of response to number. You can choose to make 1 represent either familiar or unfamiliar and recode the other Likert scale response to be consistent.

The appropriate statistical test depends on what you want to know. Likert scales are ordinal variables. We know that a 2 response means less familiarity (assuming "familiar" is coded as 1) than a 1 response, but we don't know that the difference between a 2 and a 1 is the same as the difference between a 3 and a 2. Nevertheless, for simple situations like this, many people compare the means of the Likert responses and use a t-test for statistical significance. It's wrong in theory, but works well in practice much of the time if your interest is only ranking words or people based on familiarity.

If you have a more complex hypothesis, such that some words or people give a lot of 3s while other words or people give lots of 1s and 5s, you'll have to specify your alternative hypothesis, which in turn will suggest a null hypothesis, which will then result in a nonparametric test.

Aaron Brown, reversing one of the items may make it easier to interpret some types of results but it will not affact any of the statisticAl analyses

I'm not sure what you mean. Suppose the two words (assuming we are comparing familiarity of two different words among a group of people) have identical scores, say 20% familiar, 40% somewhat familiar, 40% neither familiar nor unfamiliar. On one scale that's 20% 1s, 40% 2s and 40% 3s; on the other it's 20% 5s, 40% 4s and 40% 3s. If we test to see if people are equally familiar with both words, we will conclude "yes" if we reverse one of the scales, "no" if we do not.

Let’s suppose that you have four items, with 3 ranked from 1 (familiar) to 5 (unfamiliar), and 1 ranked from 1 (unfamiliar) to 5 (familiar). Suppose one person always answers familiar, with a mean of (1+1+1+5)/4=2 and another always answers unfamiliar, with a mean of (5+5+5+1)/4=4. Thus, the difference between the means is 2, while it really should be 4 if every item is ranked in the same way (1+1+1+1)/4=1 vs. (5+5+5+5)/4=5. So the ordering does make a difference, and you should convert the scores so all have the same orderings (e.g., change 5 to 1, and 4 to 2, for those in which 5=familiar). If you are only looking at one item at a time, then the ordering doesn’t matter because each item will be scored consistently, but if you are looking across items, then the ordering matters.

There is a big gap between practice and theory, and somehow in these societies, theory is an ideal, and practice is an arbitrary process.

Zhao Quanbei Bradford Chaney Aaron Brown David L Morgan Kandasamy Ravichandran

Practically speaking, yes, we do this all the time by having reverse items (which we do to reduce "flat-lining", i.e. people getting into a pattern/habit of checking off the same response across items) and then without hesitation reverse back the reversed items, so we can compare responses across items. As well, it's not atypical to "score" by summing across Likert items with the same number of response options (e.g. 5-point) to obtain "area scores" so...I don't see a problem, pragmatically. What can get a bit more problematic is, for example, directly comparing without calibrating how people answer using the same Likert scale (e.g. 5-point) but in different cultures, say, different countries where certain cultures may be less prone to answering on the "extremes", while other cultures may be more "polite" thus tend to be biased towards more positive responses.

The Likert scale is a scale that converts qualitative adjectives into quantitative, and at first it was 6 options, but it was gradually changed to 5 options and it can be collected based on its frequency.

There is an open question as to how to interpret the numbers. If you revise them so that 1 and 5 always mean the same thing, then the numbers are at least ordinal; i.e., moving up from 1 to 5 reflects a progressive move in one direction that is defined consistently for all items. Strictly speaking, to calculate means results in imposing the assumption that all intervals are equal (e.g., that the distinction between 1 and 2 has the same meaning as the distinction between between 2 and 3). That may or may not be true, but people often calculate means anyway (and there may be some degree to which, when we assign numbers visibly on the questionnaire, those numbers themselves affect how people perceive the differences between categories). Making assumptions that simplify the data is often useful. Still, when we see a number, we often attach more precision to it than we would to a word such as "familiar." As Randall Chadwick notes, there are variations in how people respond to the same set of categories. It is good to remember that the distinctions between categories is not as precise as we often imagine.

Exactly Here, we should have an interpretative view assuming that the data is relative, and according to the studied society and its differences, it can have a significant variance.

Through mathematical transformation, it can be compared

Yes sir It can be done as an absolute value In addition, the researcher should not be limited to small amounts and should also use the creative mind.

For me, it is a "NO" Familiarity should have same point on both instruments that is, either 5 or 1. In my opinion, i think 5 is a better option. If you wish to have more of familiarity in your first instrument and more of unfamiliarity in your second instrument, you can pattern your questions in a

way that will sway the respondents' responses to having your preferred point.

Two Likert scales with opposite values cannot be directly compared numerically.

Hello The Likert scale has different domains, six of which are divided into five, and the way it is arranged depends on the researcher whether it is left or right, the best statistical method is the weighted average.

good luck.

prof.latifi