Likert scales, done properly.

Strictly speaking, a Likert scale is the sum or average of several agree-disagree items measuring one construct; a single row is a Likert item. The distinction is not just etiquette. Single items are noisy and ordinal; multi-item scales are where the format earns its reputation for reliability. If a construct matters to your study, measure it with three to five items, not one, and check that they hang together before you sum them.

Five or seven points. Below five, you throw away real variance. Above seven, respondents cannot reliably tell adjacent points apart, so the added precision is cosmetic. Seven suits engaged, literate audiences; five is the safer default for everyone else. Whatever you pick, keep it identical across every battery in the survey, because respondents calibrate once and reuse the calibration.

Keep the midpoint, usually. Genuinely neutral respondents exist. Remove the midpoint and they do not become opinionated, they pick a side at random, or quit. The honest use of an even-point forced choice is a construct with no defensible neutral state, and that decision belongs in your method section.

Label every point, with words. Fully labeled scales beat endpoint-only scales on reliability, and the labels must be symmetric: as many agree options as disagree options, with mirrored intensity.

Note what the example does not do: no lopsided sets like one disagree option against three agree options, no intensity jumps like Slightly straight to Extremely, and the midpoint says neither, not Don't know. If some respondents may lack the experience to rate the item, give them a separate Not applicable option outside the scale instead of letting them camp on the midpoint.

Likert items usually ship as a matrix: statements down the side, the scale across the top. Grids are efficient and dangerous in equal measure, because the format invites straight-lining, racing down one column without reading. The defenses are structural. Keep batteries to five to eight rows before a break. Randomize the row order so fatigue lands on a different item for each respondent, a technique covered in the randomization guide, and never randomize the scale points themselves. On long batteries, add one instructed-response row as an attention check.

Check the phone rendering before launch. A seven-point matrix that is pleasant on a laptop can become a horizontal-scrolling wall on a 375 pixel screen, and mobile respondents will be a third to a half of your sample whether you planned for them or not.

In Plumeform, the Likert matrix is a question type with per-row requirements, row randomization with the scale order fixed, and automatic straight-liner flagging: a respondent who gives the same answer down a battery of three or more rows is flagged in results and exports, next to the speeder and attention-check flags.

A single Likert item is ordinal data. Medians, distributions, and proportions, such as the share who agree or strongly agree, are always defensible; a mean of one item is a convention many fields tolerate and some reviewers do not. Summed multi-item scales are conventionally treated as interval, and means, correlations, and regressions on them are standard practice. Two habits keep you out of trouble: look at the full distribution before summarizing anything, because a polarized 1-and-5 split and a uniform spread can share a mean, and decide your analysis plan before the data arrives, not after.