There are some topics for which neutral is not a plausible answer, and your response options should not include that. If you're asking about opinions about capital punishment or a single-payer insurance model, I would use questions with an even number of response options."TEACHER provides feedback that helps me to understand how to improve." Students might have reason to agree or not agree to that sentiment. I could also picture myself as a student not having an opinion either way for some courses.
The idea is that if neutral is plausible as someone's "true score," thenthereshould be an option for that. If that is not included, then you may be introducing measurement error by forcing them to pick a direction.Furthermore, in some surveys by not providing a Neutral option you might have more non-responses (some who truly feel neutral might not answer some questions). For course feedback, I would lean toward an odd number of responses but without much conviction. I could see it either way (I more agree than disagree??). It would be a worthy pilot study to see whethertheformat changeaffects things like reliability or response rates for course evals at your school.
As far as the number of response options goes, a common recommendation I have heard is to use between 4 and 7 options. More than that is beyond the number of bits of information that people typically can keep in working memory. Less than that results in too much loss of information due to categorization (see below). Shaw et al. (1987) give a pretty straightforward look at the effect of just categorizing a continuous variable, representing the effect of using an ordinal scale to measure a hypothetical, continuous latent trait. If you categorize a normally distributed variable into 3 bins, the r-square between X and the discretized version of X is 48%. You're losing over half of the variability. For 4, 5, 6, & 7 bins, the r-squares are 67, 77, 83, and 87% respectively. If you conceptualize the items as individual predictors used to estimate each person's latent true score, those are the maximum bivariate r-squares. Other aspects of measurement error would further attenuate the bivariate relationships.
Shaw, D. G., Huffman, M. D., & Haviland, M. G. (1987). Grouping continuous data in discrete intervals: Information loss and recovery. Journal of Educational Measurement, 24(2), 167-173.------------------------------
Robert Pearson
Asst. Professor
Grand Valley State University
------------------------------
Original Message:
Sent: 03-24-2017 09:30
From: Heather Rollins
Subject: question about survey questions
I suspect that the "gut feel" for an odd number of choices is a psychological thing, along the lines of having to choose "random" numbers:
https://blogs.msdn.microsoft.com/shawnhar/2009/12/17/the-psychology-of-randomness/
That said, I always like to have a neutral or N/A option on surveys that I fill out. Sometimes the particular aspect in the question is something that isn't a high priority for me, and so I haven't paid attention to it. And sometimes (thinking especially of customer satisfaction surveys at a hotel) it is genuinely not applicable.
Best,
Heather
--
Heather Rollins
Analysis Manager, Web Teks
Graduate student in math, UWF