0
$\begingroup$

I have a data set which is based on 5-point Likert scale. The options are in either [Strongly disagree, ..., Strongly agree] or [Never, ..., Very frequently]. Now I want to map the responses to numbers like [Strongly disagree -> 1, Disagree -> 2, Neutral -> 5, Agree -> 10, Strongly agree -> 17] for running some machine learning algorithms. Is it ok? Am I doing it right?

$\endgroup$
9
  • $\begingroup$ One commonly sees 5-category Likert scale mapped to numbers $-2, -1, 0, 1, 2$ or $1,2,3,4,5.$ Do you have a good reason (before seeing data) to have a difference of $1$ on the numerical scale btw Strongly disagree and Disagree, while having a difference of $7$ btw Agree and Strongly agree? $\endgroup$
    – BruceET
    Oct 4, 2019 at 22:35
  • 1
    $\begingroup$ Without considering the data set, I also think that the options of the answer should be mapped to 1,2,3,4,5. @BruceET $\endgroup$
    – Najat
    Oct 5, 2019 at 7:04
  • $\begingroup$ But when I ran linear regression, the r2 score was really low (like 0.009). Because, for a particular value of the independent variable, multiple values of the dependent variable were present. The dependent variable was calculated by summing multiple attributes of the data set (Likert scale values). So I thought that if the responses of the attributes that I've used to make the dependent variable were mapped to different numbers (like I said), then different options will have different impact on the final dependent variable and finally, may be an improved r2 score can be achieved. @BruceET $\endgroup$
    – Najat
    Oct 5, 2019 at 7:06
  • 1
    $\begingroup$ How important do you think the 'Strongly Agree' answers really are? Important enough to be heavily weighted (17) so you get a nice regression? Or is it more realistic to weight them in the normal way (5) and not get such a nice regression? // Also pay due attention to @ Bernhard's Comment, which I upvoted. $\endgroup$
    – BruceET
    Oct 5, 2019 at 7:13
  • 1
    $\begingroup$ I'm not urging any one assignment. Just trying to make sure you think through what you're doing. // Some people don't trust research done with Likert surveys with categories coded as numerical because the values assigned to categories can make big differences in 'results'. Always try to remember the categories are really ordinal, not numerical. $\endgroup$
    – BruceET
    Oct 5, 2019 at 7:32

1 Answer 1

1
$\begingroup$

A likert scale is not the 5-point answering format. A scale is a sum of such "assigned" numbers and usual practice is to do some thourough investigation (including but not limited to factor analysis) whether the added items belong together in the sense of forming more or less a single factor.

Such sums of multiple items have been used by Rensis Likert in Likert, R. “A technique for the measurement of attitudes”.Archives of Psychology, 1932;Vol 22, No. 140, p 55 https://legacy.voteview.com/pdf/Likert_1932.pdf

Whilst attaching numbers to ordinal data still results in ordinal data that you are not allowed to sum, simply doing so worked well for Likert, as he writes

Although the sigma technique seemed to be quite satisfactory for the intended use, it was decided to try a simpler technique to see if it gave results comparable with the sigma technique. If it did, the simpler method would save considerable work in a general survey type of study of this kind. The simpler technique involved the assigning of values of from 1 to 5 to each of the five different positions on the five-point statements. The ONE end was always assigned to the negative end of the sigma scale, and the FIVE end to the positive end of the sigma scale. After assigning in this manner the numerical values to the possible responses, the score for each individual was determined by finding the average of the numerical values of the positions that he checked. Actually, since the number of statements was the same for all individuals, the sum of the numerical scores rather then the mean was used. The reliability of odds vs. evens for this method yielded essentially the same values as those obtained with the sigma method of scoring. The scores obtained by this method and the sigma method correlated almost perfectly as will be seen in Table IV.

So you are doing something that is not following strict mathematical rules but has worked anyways for some time now in a framework called "Classical Test Theory". Taking the habit of summing ordinal data out of that framework (i. e. scale building etc.) might impose serious problems, so there is no simple saying "Its ok, you're doing it right." Depending on the circumstances it may be the most suitable thing for you to do, but be aware of breaking the rules of mathematics here. The mean of "disagree" and "neutral" is not "disagree and a half"!

Strongly disagree -> 1, Disagree -> 2, Neutral -> 5, Agree -> 10, Strongly agree -> 17

Assigning number 1 to 5 is common practice, using values of 5, 10 and 17 will need some serious justification.

$\endgroup$
2
  • $\begingroup$ Thanks for your answer $\endgroup$
    – Najat
    Oct 4, 2019 at 9:45
  • $\begingroup$ You are welcome. If you consider an answer helpful, feel free to upvote it. If you have further questions, feel free to ask them. Should you feel that one of the answers settles the question, please "accept" that answer. $\endgroup$
    – Bernhard
    Oct 4, 2019 at 11:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.