You built a 20-item questionnaire, collected 200 responses, and opened your data file. Every variable is a number between 1 and 5. Now what? The answer depends on whether you treat those numbers as labels for ordered categories or as points on a continuous ruler, and that single decision changes which statistical tests are valid, which descriptive statistics to report, and whether your committee will accept your results.
What a Likert Scale Actually Is
Rensis Likert introduced the format in 1932 as a way to measure attitudes. The original design asked respondents to rate their agreement with a statement on a five-point scale: strongly disagree (1), disagree (2), neutral (3), agree (4), strongly agree (5). The idea was simple. Instead of asking a single yes/no question about whether nurses enjoyed their work, you could ask 10 or 15 related statements and sum the responses into a total score that captured attitude with more precision than any single question could.
Today, variations are everywhere. Seven-point scales, four-point scales with no neutral option, frequency scales (never to always), satisfaction scales (very dissatisfied to very satisfied). The common thread is that responses are ordered categories. A "4" always means more agreement than a "3." But how much more? That is where the debate starts.
The Ordinal vs. Interval Debate
Strictly speaking, Likert data is ordinal. The distance between "strongly disagree" and "disagree" is not guaranteed to equal the distance between "agree" and "strongly agree." A respondent who circles 4 instead of 3 has expressed more agreement, but you cannot say they expressed exactly one unit more. This is the textbook position, and it is technically correct.
In practice, the field is split. Norman (2010) published a widely cited analysis reviewing decades of evidence and concluded that parametric tests (t-tests, ANOVA, Pearson correlation) are surprisingly robust with Likert-type data. Simulation studies show that even with five-point scales and moderate skew, the Type I error rate stays close to the nominal .05. The F-test, in particular, tolerates ordinal input well when sample sizes exceed 30 per group.
The practical rule most supervisors accept: treat individual Likert items as ordinal, but treat summed or averaged scales (5+ items measuring one construct) as approximately interval when the distribution is roughly symmetric. A single item scored 1 to 5 should get a Mann-Whitney U test, not an independent-samples t-test. A 12-item satisfaction scale averaged into a composite score between 1.0 and 5.0? That can be analyzed with parametric methods if skewness stays within ±1.0 and the sample is at least 30.
Designing Good Likert Items
Bad items produce bad data, and no statistical technique can fix that after the fact. We have reviewed questionnaires from over 300 projects, and the same design errors appear repeatedly.
Each item should measure exactly one concept. "My supervisor provides clear instructions and supports my professional development" is a double-barreled item. A respondent might agree with the first half and disagree with the second, but the scale forces a single answer. Split it into two separate statements. Every double-barreled item you keep introduces measurement noise that will lower your Cronbach's alpha.
Avoid double negatives. "I do not feel that my workload is unreasonable" forces the respondent to parse two layers of negation before answering. The cognitive load leads to random responses, especially in long questionnaires. Write it as "My workload is reasonable" and save the mental gymnastics.
Balance your wording direction. If all 15 items are phrased positively ("I enjoy my work," "I feel valued," "I am satisfied with my pay"), respondents tend to agree with everything out of habit. This is acquiescence bias, and it inflates your means without reflecting genuine attitudes. Include 4 to 6 negatively worded items ("I often feel bored at work") and reverse-code them during data preparation. A well-balanced 10-item scale might have 6 positive and 4 negative items.
Coding and Preparing Likert Data in SPSS
Once data collection is finished, preparation follows a specific sequence. In SPSS, define your variables as numeric with values labeled (1 = Strongly Disagree through 5 = Strongly Agree). Set the measurement level to "Ordinal" for individual items.
Reverse coding comes next. If item 7 reads "I find my job boring" on a satisfaction scale where high scores mean high satisfaction, a respondent who strongly agrees (5) is actually expressing low satisfaction. You need to flip the coding: 1 becomes 5, 2 becomes 4, 3 stays 3, 4 becomes 2, 5 becomes 1. In SPSS syntax, the command is straightforward:
COMPUTE Item7_R = 6 - Item7.
The formula (number of response options + 1) minus the original score works for any scale length. For a 7-point scale, use 8 minus the score. After reverse coding, compute your scale composite. If you have 10 items measuring job satisfaction (with items 4, 7, and 9 already reverse-coded), create the mean:
COMPUTE Satisfaction_Mean = MEAN(Item1, Item2, Item3, Item4_R, Item5, Item6, Item7_R, Item8, Item9_R, Item10).
Use MEAN rather than a manual sum divided by 10. The MEAN function in SPSS handles missing values gracefully; it computes the average of available items rather than returning a missing value if even one item is blank. For a 10-item scale, SPSS will compute the mean as long as the respondent answered at least 1 item, though you should set a threshold (typically 80% of items answered) to avoid computing means from too few responses. You can upload your data to Academic Stats Agent to automate this entire preparation step, including reverse coding and reliability checks.
Checking Reliability Before Analysis
Before running any hypothesis tests, confirm that your scale items measure the same construct consistently. Cronbach's alpha is the standard check. The widely accepted minimum is α ≥ .70 (George & Mallery, 2003). Published psychological instruments typically report values between .80 and .90.
Run the reliability analysis on each subscale separately. A 30-item questionnaire with three 10-item subscales needs three separate alpha values, not one overall coefficient. If you compute a single alpha across all 30 items measuring different constructs (say, anxiety, depression, and stress), the result will be misleading because the items are not supposed to correlate perfectly across constructs.
When alpha falls below .70, check the "alpha if item deleted" column in your output. If removing one item raises alpha from .65 to .74, and that item is theoretically weak (for example, a commute question in a job satisfaction scale), dropping it is defensible. But do not remove items purely to chase a higher number. Every deletion must have a substantive reason your committee can accept.
Choosing the Right Statistical Test
The test you pick depends on whether you are analyzing a single Likert item or a summed/averaged scale. The table below summarizes the most common scenarios.
| Data Type | Comparison (2 Groups) | Comparison (3+ Groups) | Descriptive Stats |
|---|---|---|---|
| Single Likert item (ordinal) | Mann-Whitney U | Kruskal-Wallis H | Median, IQR, frequencies |
| Summed/averaged scale (interval-like) | Independent t-test | One-way ANOVA | Mean, SD |
| Paired Likert item (ordinal, pre-post) | Wilcoxon signed-rank | Friedman test | Median, IQR |
| Paired scale (interval-like, pre-post) | Paired t-test | Repeated-measures ANOVA | Mean, SD |
For descriptive statistics, report the median and interquartile range (IQR) for individual items, and the mean and standard deviation (SD) for composite scales. A common thesis table might show: "Item 3 ('I feel supported by my team'), Mdn = 4.00, IQR = 3.00–5.00." For the overall scale: "Team Support Scale (8 items), M = 3.72, SD = 0.68, α = .81."
The Most Common Mistake
Running an independent-samples t-test on a single 5-point Likert item. We see this in roughly 1 out of every 5 thesis drafts that come through our service. The student compares males and females on a single item ("I am satisfied with my university experience," scored 1–5) and reports t(198) = 2.14, p = .034. The problem is that a single item with five possible values violates the continuous-variable assumption of the t-test. The correct test is Mann-Whitney U, which compares rank distributions rather than means. The result might look nearly identical (U = 4,230, p = .031), but the methodological foundation is sound, and no reviewer will question it.
The reverse mistake also happens, though less often: applying nonparametric tests to a well-constructed composite scale with 12 items and a roughly normal distribution. If your averaged scale produces values like 3.42, 2.85, 4.17 and skewness is −0.38, a t-test is perfectly appropriate and will give you more statistical power than a Mann-Whitney.
Key takeaway: Treat individual Likert items as ordinal (report medians, use Mann-Whitney or Kruskal-Wallis). Treat composite scales of 5+ items as interval when the distribution is roughly normal (report means, use t-tests or ANOVA). Always check Cronbach's alpha (α ≥ .70) before running hypothesis tests. Reverse-code negatively worded items before computing scale scores. And never run a t-test on a single 5-point item.