How to Analyze Questionnaire Data for Your Medical Thesis

Q: Can I use mean and standard deviation for Likert scale data?

Strictly speaking, Likert items are ordinal data, so median and IQR are the appropriate descriptive statistics. However, when a scale has 5 or more response categories and a composite score is computed from multiple items (n > 30), many researchers report means and SDs and treat the composite as approximately continuous. When in doubt, report both the median/IQR and mean/SD, and use non-parametric inferential tests.

Q: What is Cronbach's alpha and when do I need it?

Cronbach's alpha measures the internal consistency reliability of a multi-item scale how well all items measure the same underlying construct. You need it whenever you combine multiple questionnaire items into a total score or subscale score. Alpha 0.70 is generally acceptable, 0.80 is good, and 0.90 is excellent (though above 0.95 may indicate redundant items).

Q: How do I handle missing data in questionnaire analysis?

First, determine whether data are missing at random (MAR) or not at random (MNAR). For small amounts of missing data (<5% per item), mean substitution or listwise deletion is acceptable. For larger amounts, consider multiple imputation. Never ignore missing data always report the amount of missing data per variable in your results.

Questionnaires are among the most widely used data collection instruments in medical research from needs assessments and KAP studies (Knowledge, Attitudes, Practices) to validated patient-reported outcome measures. Yet they are also among the most frequently misanalyzed. The reason: questionnaire data is mostly ordinal, and most researchers incorrectly apply tests designed for continuous, normally distributed outcomes. This guide shows you the correct approach, step by step.

Step 1 Understand Your Questionnaire Data Types

Before you run a single test, classify every variable in your questionnaire into one of four data types. This determines everything that follows.

Nominal (Categorical)

Unordered categories with no inherent ranking. Arithmetic operations have no meaning.

Examples: Gender (male/female), specialty (cardiology/surgery/GP), disease diagnosis, yes/no questions

Ordinal

Ordered categories where the distances between levels are not equal or not known. This includes ALL Likert items.

Examples: Likert scale (1=Strongly Disagree to 5=Strongly Agree), satisfaction (poor/fair/good/excellent), pain severity (mild/moderate/severe)

Continuous (Scale)

Measured on a continuous scale where equal differences mean equal amounts. Mean and SD are appropriate.

Examples: Age, BMI, weight, total score of a validated instrument (if treated as continuous)

Composite Score

Sum or average of multiple Likert items measuring one construct. Often treated as approximately continuous when n is large and the scale is validated.

Examples: Total knowledge score (sum of 10 items), PHQ-9 depression score, SF-36 domain scores

Most important ruleIndividual Likert items are ordinal. Do NOT calculate means and standard deviations for individual Likert items, and do NOT use t-tests or ANOVA on them. Use medians, frequencies, and non-parametric tests instead.

Step 2 Data Coding and Cleaning

Before analysis, every questionnaire response must be converted to a number and checked for errors. This preparation stage often called data management takes as much time as the analysis itself, and is just as important.

Coding Your Variables

Create a codebook that maps every response to a numeric value. For example:

// Codebook example Gender: 1 = Male, 2 = Female Specialty: 1 = Internal Medicine, 2 = Surgery, 3 = Pediatrics, 4 = Other Q1 (Likert): 1 = Strongly Disagree, 2 = Disagree, 3 = Neutral, 4 = Agree, 5 = Strongly Agree Knowledge Q: 0 = Wrong answer, 1 = Correct answer

Reverse Coding Negatively Worded Items

When a questionnaire includes negatively worded items (e.g., "I do NOT feel confident performing this procedure"), the response scale is reversed relative to the other items. A response of "Strongly Agree" (5) to this item actually indicates low confidence. Before computing a total score, reverse-code these items:

For a 5-point Likert scale: Reversed score = (Maximum + Minimum) Original score = 6 original score

So a score of 5 becomes 1, 4 becomes 2, 3 stays 3, etc. Your questionnaire's manual or the published validation paper will tell you which items to reverse.

Handling Missing Data

Missing responses are common in questionnaire studies. Your approach depends on how much data is missing:

< 5% missing per item: Acceptable to use mean substitution (replace missing with the item mean for that respondent's group) or listwise deletion (exclude participants with any missing data)
515% missing: Use multiple imputation; report sensitivity analyses
> 15% missing: Serious concern; investigate the cause; report clearly in limitations

Always report the number and percentage of missing responses per variable in your results table.

Step 3 Descriptive Statistics

Describe your sample and each variable before any inferential tests. The correct descriptive statistic depends on data type:

Nominal variables Frequency table

Report n and percentage for each category. Example: "126 (63%) participants were female." Include all categories even if n = 0.

Ordinal / Likert items Median + IQR + Frequencies

Report the median, interquartile range (IQR = 25th75th percentile), and the percentage in each response category. Example: "Median awareness score = 3 (IQR: 24); 47 (23.5%) Strongly Agree." Never report only the mean for individual Likert items.

Continuous variables Mean + SD (if normal) or Median + IQR (if not)

Test normality with Shapiro-Wilk (n < 50) or KS test. Age and BMI are often approximately normal; report Mean +/- SD. Scores, lengths of stay, and lab values are frequently non-normal; report Median (IQR).

Composite scores Check normality first

Sum all correctly coded items for the relevant subscale. Run the normality test. If approximately normal, report Mean +/- SD and use parametric tests. If not normal, report Median (IQR) and use non-parametric tests.

Step 4 Reliability Analysis (Cronbach's Alpha)

If your questionnaire uses a multi-item scale where several questions combine to measure a single construct (e.g., 10 knowledge questions combined into a "total knowledge score") you must report Cronbach's alpha before using the composite score in any analysis.

Cronbach's alpha (+/-) measures internal consistency reliability: how well all items in a scale measure the same underlying construct. It ranges from 0 to 1.

Alpha Value	Interpretation	Acceptable for publication?
+/- 0.90	Excellent internal consistency	Yes but check if items are redundant (above 0.95 often indicates item redundancy)
+/- = 0.800.89	Good internal consistency	Yes ideal range for most medical scales
+/- = 0.700.79	Acceptable internal consistency	Yes acceptable for exploratory research
+/- = 0.600.69	Questionable	Borderline report with caution; may need to revise or remove items
+/- < 0.60	Poor	No items do not sufficiently measure the same construct

When to report Cronbach's alphaRequired whenever you sum or average multiple items into a composite score. Not needed for individual items or demographic variables. If using a validated questionnaire (e.g., PHQ-9, GAD-7, SF-36), report the alpha from your sample alongside the value from the original validation study.

Step 5 Inferential Tests: Choosing the Right Test

This is where most errors occur. Match the correct test to your research question and data type using the guide below.

Comparing questionnaire scores (ordinal/Likert) between 2 independent groupsExample: Is knowledge score different between male vs female physicians?

Mann-Whitney U

Comparing questionnaire scores between 3+ independent groupsExample: Does attitude score differ across 4 medical specialties?

Kruskal-Wallis

Comparing questionnaire scores before and after an intervention (same subjects)Example: Did awareness scores improve after a training workshop?

Wilcoxon Signed-Rank

Association between two categorical variablesExample: Is there an association between specialty and willingness to refer patients for mental health care?

Chi-Square

Correlation between two ordinal or non-normal variablesExample: Is there a correlation between knowledge score and years of experience?

Spearman

Predicting a binary outcome from multiple questionnaire variablesExample: Which factors predict willingness to receive HPV vaccination (yes/no)?

Logistic Regression

The Likert controversyThere is ongoing debate about whether Likert scales can be analyzed as continuous. The majority position: individual Likert items are ordinal use non-parametric tests. However, if a scale has 5+ response options AND items are summed into a validated composite score AND n > 30 per group AND the composite is approximately normal, treating the composite as continuous is defensible. When in doubt, run both parametric and non-parametric tests if conclusions are the same, report the parametric result with a note; if conclusions differ, report both.

Step 6 Reporting Results in Your Thesis

Your thesis methods section and results section need to present the analysis systematically. Here is the standard structure for questionnaire-based medical research:

Methods Section Should Include:

Description of the questionnaire instrument (validated tool or self-designed; if validated, cite the original validation paper and report the alpha from the validation study)
Scoring method (how individual items combine into composite scores)
Definition of the primary outcome measure
The specific statistical tests used for each research objective, with justification
The significance threshold (+/- = 0.05, two-tailed)
Software used (SPSS version X, or StatClinic)

Results Section Should Include:

Table 1: Sociodemographic characteristics (frequency/% for categorical, Mean+/-SD or Median(IQR) for continuous)
Table 2: Responses per questionnaire item frequency and percentage in each category, and median (IQR) for each item
Table 3: Composite score distribution Cronbach's alpha, Mean+/-SD or Median(IQR), and total score distribution
Table 4: Inferential analysis test statistic, p-value, and effect size for each comparison or correlation

Example of correctly written results: "The median total knowledge score was 14 (IQR: 1117) out of 20. Male physicians had significantly higher knowledge scores than female physicians (Median: 15 vs 13; U = 1,842, p = 0.003, r = 0.22)."

Common Mistakes to Avoid

Calculating means for Likert items ordinal data; use medians and percentages for description, Mann-Whitney/Kruskal-Wallis for comparison
Using a t-test or ANOVA on raw Likert scores these tests assume continuous, normally distributed data
Omitting Cronbach's alpha required for any composite scale; thesis examiners and peer reviewers will notice its absence
Analyzing a poorly designed scale if your questionnaire items don't all load on the same construct (verified by factor analysis), combining them into a total score is meaningless
Not accounting for multiple comparisons if you compare scores across 6 specialties and then compare each specialty pair, apply Bonferroni correction or Dunn's post-hoc test
Reporting only p-values without effect sizes for Mann-Whitney U, report r = Z/N; for Chi-Square, report Cram(c)r's V or Phi; for Spearman, r is the effect size

Frequently Asked Questions

Can I use mean and standard deviation for Likert scale data?+

Strictly speaking, Likert items are ordinal so median and IQR are appropriate descriptive statistics. However, when a validated composite scale has 5+ response options and is computed from multiple items (with n > 30), many researchers report means and SDs and treat the composite as approximately continuous. When in doubt, report both the median/IQR and mean/SD, and use non-parametric inferential tests.

What is Cronbach's alpha and when do I need it?+

Cronbach's alpha measures internal consistency reliability how well all items measure the same construct. You need it whenever you combine multiple questionnaire items into a total or subscale score. Alpha 0.70 is acceptable, 0.80 is good. Always report it for your specific sample, not just the value from the published validation study.

What test should I use to compare questionnaire scores between two groups?+

For ordinal questionnaire scores (individual Likert items or non-normal composite scores), use the Mann-Whitney U test for two independent groups, or the Wilcoxon Signed-Rank test for paired comparisons. For normally distributed composite scores with n > 30 per group, an independent t-test or paired t-test is acceptable.

How do I handle missing data in questionnaire analysis?+

For < 5% missing per item: mean substitution or listwise deletion is acceptable. For 515%: use multiple imputation. For > 15%: investigate the cause and report in limitations. Always report the number and percentage of missing responses per variable. Never simply ignore missing data without documenting it.

Should I combine Likert items into a total score or analyze them separately?+

Analyze items separately for descriptive purposes (frequency table per item). For inferential testing, combine validated items into subscale or total scores to increase reliability and statistical power. Only combine items that measure the same construct confirmed by Cronbach's alpha 0.70.

Need help analyzing your study?

Use StatClinic AI Statistical Assistant free Chi-Square, Mann-Whitney, Spearman Correlation, and more. No SPSS required.

Use StatClinic AI Statistical Assistant