Ruth,

the issue of whether or not to offer a midpoint has been disputed for decades, as you already found out. The original meaning of the midpoint on a Likert-type scale is neutral or indifferent, e.g., “neither agree nor disagree”. However, in several research contexts the midpoint has been found to have two different meanings, i.e., true neutral meanings and meanings that refer more to a kind of non-response, such as “don’t know”, “undecided” or “no opinion”. Studies have also shown that respondents use the neutral category more often when “don’t know” is not offered as an alternative. Therefore the researcher cannot differentiate between true neutral responses and non-responses. When no midpoint is offered, one can expect that the respondents will randomly use the category above or below the assumed midpoint.

References:

Baumgartner, H., & Steenkamp, J.-B. E. M. (2001). Response styles in marketing research. Journal of Marketing Research, 38, 143-156.

Krosnick, J. A., & Fabrigar, L. R. (1997). Designing rating scales for effective measurement in surveys. In L. Lyberg, P. P. Biemer, M. Collins, E. de Leeuw, C. S. Dippo, N. Schwarz, & D. Trewin (Eds.), Survey measurement and process quality. New York: Wiley.

HTH

Karin

Here is a link to a more recent discussion of this and a number of issues by two of the best known authors in this field:

https://pprg.stanford.edu/wp-content/uploads/2010-Handbook-of-Survey-Research.pdf

Pros: You get to pretend that this response (like the others) enables you to measure a response

Cons: You get to pretend that this response (like the others) enables you to measure a response

Every measurement in every science involves error. Much of statistics and data analysis comes from applied science and statistical mechanics developed in the 19th and early 20th centuries, and in order to deal with measurement error and deficiencies in measurement capacity (such as systems to complex for anything other than a statistical description). Modern Likert-type response data uses such statistics (Pearson's product-moment linear correlation coefficient, (M)ANOVA, least-squares regression, etc.), but the logic is reversed. Instead of trying to take a number of measurements of a system and derive the most accurate or minimize measurement error for a series of measurement values, the "measures" are treated as infinitely precise such that the loss of information to subjective interpretations, middle response bias, etc., is dealt with theoretically (via a priori methods, namely experimental design) rather than quantitatively and for each measurement (each response to each item) the error isn't just ignored entirely but assumed to be non-existence. The statistics are then used to take these measurements and, instead of minimizing the over-all error (as non-exists by assumption), this massive pool of measurement uncertainty is treated as an ideal measurement for a "property" of population that wasn't measured.

In addition to what Karin said above (which I agree with), I can say that the main cons for the mid point is the respondents who feel that they obliged to give answer for the items in the scale but they are undecided and there is no time to think. However, I think the Pros of keeping the midpoint is much more. It is normal response for human beings to be in the middle for some actions or behaviors. If midpoint is deleted, I dont go with FORCING rspondents to stand in one side of the question!! Further, when we run inferential statitics, the DV will be the total score of the scale which will reflect the overall stand for each respondent.

Thank you, Colleagues, for your thoughtful answers.

I find that each answer that has been provided is very compelling, and, perhaps because I am a pragmatist, convincing - as I can see both sides of each argument. Karin accurately pointed out the dilemma that I am facing in making the decision to include or not include the midpoint response, i.e., the respondent choosing the response that would normally flank either side of the midpoint. The Krosnick and Presser (2010) article that David provided is very enlightening, and has me thinking that this is a dilemma that will not be easily solved, and the debate will go on forever.

I noted with interest Andrew Messing's opening statement that a benefit of the midpoint is that "You get to pretend that this response (like the others) enables you to measure a response". I think that perhaps a major consideration that will not allow this debate to be easily solved is rooted in the inherent problem of self-reporting. It seems to me that, as already noted in extant literature, with or without the mid-point, the very nature of self-reporting threatens to bias studies... hence, the idea of pretending that the mid-point - or any other choice, when you think of it, enables one to measure a response is a sobering thought, the consequences of which are too unsettling to consider...

Finally, Professor Ahmad also made a good point when he observed that it seems to be the nature of human beings to be in the middle for some actions or behaviors (after all, reality is not starkly black or white, but there are many shades of grey in between!). When all is said and done, you have all helped to alert me to the fact that this is not a simple question, and there might never be definitive answers; the answers might be rooted more deeply in one's philosophical stance  than I could have imagined!

Of course, all of this now has me wondering if the mid-point might also be a way of unintentionally introducing subjectivity into what is supposed to be an objective process...  Any thoughts?

Ruth

I am strongly agreed with the very Interesting respondents answers please.

Ruth:  I think you have answered your own question in summarizing the primary purpose of the mid-point as a way to intentionally - or unintentionally introducing subjectivity into a supposedly objective process.  Might this be an unintended consequence of using or having such a mid-point in the Likert scale?

I think there is need to keep the mid point. There are statements on which you are neutral and keep yourself on the mid point. As given above in a suggestion, that the respondent put themselves in the neutral category when they do not have an answer, it means that the investigator is not in a position to make the point clear. In this situation, we can have a separate column "No Response" which should not be included while we calculate the scale value.

Dear Randhir Singh Poswal:

Let us imagine that I'm responding to the statement "Pluto is not a planet" and my possible answer choices include "strongly disagree", "disagree", "neutral", "agree", "strongly agree". Inevitably, each response will be coded with an integer 1 through 5, with the middle response being 3. Now, let us imagine that (as is true) I couldn't care less. So I respond with 3. Let us imagine that someone else does care, but really is neutral in that she neither agrees nor disagrees because she finds arguments from both sides compelling. Both of our answers are treated as exactly the same, and even worse both of us are treated (usually) as agreeing exactly one unit greater than those who respond "disagree", and precisely one unit less than those who respond "agree". What has this middle response measured? After all, it is certainly not the case that I disagree more than anybody who has disagreed or agree less than anybody who agrees, as I don't care. My friend, the neutral one, neither disagrees nor agrees, but the sale doesn't reflect this. It's an ordered scale of agreement. That's the continuum.

A middle response isn't neutral just because you label it, because inevitably this becomes the number 3 an is plugged into some statistical model or test in which it is treated as having a value only in relationship to the numbers 1, 2, 4, 5. (assuming a 5 point scale as this generalizes to more points with a middle response). My scale is actually essentially a rating of agreement that is centered at 0, such that the two responses left of center are negative agreement and the two on the right are positive agreement and the middle is a value of 0 agreement.

Think about information theory for a moment. That is, consider how we treat information mathematically such that we can measure information lost in a signal transmission, encoded in language, etc. Having a 0 response is like having no information, and treating the 0 response as 3 is just shifting values, not changing the fact that it isn't a neutral response (not to mention the number of other motivations like not caring that respondents have for opting for the middle).

It doesn't really matter if you account for middle response bias, choose the right number of points, do or do not label every response, etc., unless you have some justification for asserting that particular values you assign responses are actually measuring (and measuring with sufficient accuracy) the information the participant's response actually was by the value they gave. This is never true, because all possible answers involve information loss we can't measure and that's without factoring in how the participant interpreted the question and how the intepreted the answer/number they circled, checked, etc. However, the problem with a middle response is that mathematically the scale is always a continuum, and not only does even a purely neutral response not belong on such a continuum (only 0 knowledge/information), but this is the option most likely to be interpreted differently or selected as an opt out. The fact that there actually are people who are neutral simply means that they are forced to select an answer that doesn't correspond exactly to their idealized/"true" response. But this is true for all responses anyway. You don't fix it by treating a middle value as neutral linguistically only to treat it differently in the only way that matters (i.e., in whatever analyses you use on the data).

Thanks Andrew. I do agree with you. This situation is not only with the middle value, it is with the continuum too. How difficult it is for any human being to exactly  measure his/her position on a continuum. Most of the respondents put themselves on a particular point and that point may not necessarily reflect their actual position.

In my opinion the issue is whether we are trying to detect 1. a true moderate level of agreement on something we've thought about, possibly a lot (e.g. for a lot of people, attitude towards some aspect of abortion policy, where a 3 out of 5 truly means they place themselves midway between the two groups of people who have the most extreme views), or 2. the view of someone who has neither a strong nor a weak level of agreement because they haven't given it a moment's thought (e.g, for my 'level of agreement with Japan's newest industrial policy', the correct response would be 'Don't Know' or 'No opinion') or 3. someone to whom the question has little if any applicability (e.g. the perceptual question that "People treated me differently when I was pregnant than when I was not" - for people who have never been pregnant, the correct response is 'not applicable'. )

Depending on the specific nature of the question and the target population, this is an excellent topic to be addressed in a questionnaire pre-test where you administer a questionnaire to a focus group and then ask them later what they thought the questions meant and whether their responses support your assumptions of what they mean.

If you code 1, 2, 3, 4, 5 as valid responses, and 8 = don't know/no opinion, and 9 = NA, you at least have a pragmatic framework for data management that partially solves some of the problems noted above, although the fact remains that measurement always involves some error ranging from the theoretical to the clerical.

Not providing a middle category is sometimes justified because it 'gets people off the fence', which it does.  But if people truly are a 3 out of 5 on something, forcing them to respond to an adjacent category that does not capture their view deliberately introduces measurement error.  

My professor in social psychology had conducted research comparing responses when allowing for a mid-point and not, and he showed that excluding the mid-point made the sample seem to be more conservative. In other words, when forced to use one or another directions, participants took more often the conservative direction.

This observation is in accordance with the tendency of people to censor themselves due to social desirability. Responding at the mid-point then must be correlated to other measurable dimensions of attitudes. There are some lines of research that routinelly omit the mid-point, such as Deconchy's work on ideology.

I would not think that "forcing individuals to respond to an adjacent category that does not capture their view deliberately introduces measurement error". A-B scales do the same in an even more absolute way, and they are considered psychometrically sound. Furthermore, 2 alternative forced choice and 4 alternative forced choice are considered as objective response techniques in the context of psychophysics.

Hi Nikolaos

Thank you, there has been some research in social psychology that I as an epidemiologist should learn more about.  My point was that if a person can truly hold a neutral position on a scale, then forcing the person to an adjacent category is necessarily causing measurement error.  If we posed the question Rate your body image: 'much too skinny', 'a little too skinny', 'a little too fat', 'much too fat', and some people truly occupy a position where they are comfortable with their body image as being 'just about right', how would the resulting distribution not be fundamentally wrong?

Hello Mark, I see your point. I will argue that the neutral position should be tapped by a number of items rather than a single item. For a single item you are absolutely justified to think that omitting the mid-point would be biased. Nonetheless, Likert scales are summations of repeated ordinal measures of the same thing. When those measures (the items) are balanced for and against the attitude object, then the overall mean should lie in the middle point, because the participant will have answered half of the items in one direction and half in another, with comparable intensity. (This is related to another question of the same researcher I answered today, regarding reverse items.)

I agree 100% if you are able to ask multiple attitude questions about a single topic.  This gets back to Ruth's original question about developing a scale.  If she will be asking multiple questions about a single area of agreement, averaging over multiple forced choices may be justifiable.  This is where the psychological and epidemiological approaches diverge.  As an epidemiologist (unless it was a primary outcome variable), I would question the added information value of all those questions (versus 1or 2 or maybe 3), and whether that intensity of measurement is worth any missing values or losses to follow-up it may cause versus a shorter scale.  (Also, I'm genuinely curious whether forcing people to agree or disagree too many times to something they are truly neutral about might not actually cause frustration and loss of participation, particularly as we are not able to explain the underlying rationale to them.)  

Well, Likert scale refers explicitly to the sum or average of a number of questions. When measuring attitudes this is the most common approach, and it is well documented in psychometrics.

Speaking of single ordinal items, I would be more approving of a mid-point for all the reasons you mention. I would then use ordinal data analysis procedures in order to find out whether the mid-point correlates with other clusters of measures.

You must carefully consider your research objectives and the likely data analyses. Likert scales are used extensively to survey opinion/attitudes, but are also used to construct situational judgment tests (SJTs) and personality inventories. At least for SJTs, you should consider the impact of individual differences in terms of rating scatter (the tendency to provide ratings with a high level of variance), elevation (the tendency to bias one's ratings toward either end of the scale) and shape (the extent one's set of ratings correlate with a standard).

There is also evidence that minorities tend to provide ratings with higher levels of scatter than non-minorities. So a failure to control scatter can exasperate adverse impact at least for SJT applications. Although I expect that there are analogous considerations for more conventional opinion surveys.  Most Likert scales reflect the simple summation of Likert ratings, but this is not necessarily advisable. For additional information:

McDaniel, M., Psotka, J., Legree, P., Yost, A. P. &  Weekley, J. A.  (2011). Toward an Understanding of Situational Judgment Item Validity and Group Differences. Journal of Applied Psychology, 96, 321-336.

Legree, P.J, Psotka, J., Robbins, J., Roberts, R. D., Putka, D. J., Mullins, H. M. (2014).  Profile Similarity Metrics as an Alternate Framework to Score Rating-Based Tests: MSCEIT Reanalyses. Intelligence, 47,159-174.