Scale length and midpoint change the result
Aliases: number of scale points · midpoint effect · even-numbered scale
What it is
The number of scale points and whether a midpoint is offered change the distribution the same people produce, even when the stem is fixed. Five, seven, or eleven points look like finer resolution; they also change how neutrality can be expressed, how hard the scale forces a lean, and how many answers pile at the poles. An even scale with no midpoint forces every respondent to take a side; true neutrality, “don’t know,” and refusal get squeezed into a weak positive or negative. Agreement rates and means then depend partly on how the boxes are drawn, not only on the attitude.
Why it happens
Respondents map a fuzzy internal state onto a finite grid. Fewer boxes merge nearby states and raise apparent agreement. More boxes leave some people unused middle ticks, while others write tiny distinctions as attitude differences. A midpoint carries at least three meanings: genuine neutrality, no attitude yet, and unwillingness to take a side. Three kinds of people press the same number and are later treated as one attitude. An even scale removes that exit and encodes “no position” as a slight lean, so opinion looks more polarized. Endpoint labels (very / extremely) further decide how many people will use the full range, which pulls the mean. Scores for the same construct on different grids are therefore not directly comparable.
Studying it
Randomly assign scale length or midpoint presence inside one population, and compare means, variances, occupancy of the midpoint (or the two central boxes), and occupancy of the poles. The question is not which grid is “more accurate,” but how sensitive the distribution is to the grid. Cognitive probes should separate midpoint choosers who are neutral, who do not know, and who are skipping difficulty. Comparisons across time or products must lock the same length and midpoint rule. If history used five points and the new study uses seven, report non-comparability; do not stretch scores linearly and call it a trend.
Where it stops holding
Expert raters are more stable on long scales; end users on mobile usually cannot use eleven distinguishable steps, so extra points become noise. A satisfaction tracker that has used five points for years will manufacture a fake change if it suddenly switches to seven “for more resolution.” Some populations avoid poles culturally; high midpoint use is not automatically a neutral attitude. A two-point yes/no item has no midpoint problem and crushes a continuous attitude into a binary, which is a different information loss. Midpoint effects are not option-order effects, and they are not mixed positive/negative item sets.
Applying it
- Lock one length and midpoint rule inside a program, and write it into the metric definition. A change is a new metric and needs a new baseline.
- When “neutral” must be separated from “cannot answer,” offer “not applicable / don’t know” besides the midpoint; do not let those meanings share the middle number.
- If a decision uses the share who took a side, declare in advance whether the scale is even; if it is, the report must say that neutrality was forcibly allocated.
- Run a small paired comparison of old and new grids. If the mean or “positive share” moves only because the boxes changed, do not write it up as a product change.