Research design · Questionnaire

How to design a survey questionnaire for your thesis: from research model to Likert scale

A bad questionnaire means bad data, and no analysis can rescue that. A 6-step process from research model to a sound questionnaire, plus the mistakes that quietly sink Cronbach’s alpha.

⏱ ~6 min readReferences: Dillman et al. (2014) · Hair et al. (2019)Updated:

You collect 300 responses and only then notice a construct in your model has no questions. Cronbach’s alpha collapses because of one double-barreled item. Respondents abandon the survey at question 40. All of these are questionnaire design errors, and no analysis can repair them once the data is in. This guide walks through six steps from research model to a sound survey questionnaire, along with the most common mistakes.

Diagram of the four questionnaire parts: introduction, screening question, main scales and demographics
The four parts of a survey questionnaire.

1What a good survey questionnaire contains

① Introduction

Who you are, why you are asking, how long it takes, and a promise of confidentiality and research-only use.

② Screening question

Confirms the right respondents (“Have you shopped online in the past 6 months?”). A “no” ends the survey.

③ Main scales

Likert items measuring each variable in your model, grouped by construct, with brief instructions.

④ Demographics

Gender, age, income and so on. Placing them last reduces drop-off, though placing them first is also accepted.

2Six steps from research model to questionnaire

🧭 The six steps
  1. List the concepts in your model, each latent variable with a clear working definition.
  2. Adapt validated scales from earlier research, and keep a table of variable, original item and source.
  3. Translate and adjust if needed: translate, then have someone back-translate to check the meaning, and tailor the wording to your context.
  4. Choose the scale and codes: one Likert format for all items in a scale, and assign variable codes now (SAT1, SAT2…).
  5. Add screening and attention checks, such as “For this item, please select ‘Agree’”, which help you remove careless responses.
  6. Run a pilot with about 20–30 people from your target group: fix confusing items, time the survey and run a first Cronbach’s alpha.
✦ How many items per construct?

At least 3 observed items per construct, usually 3 to 6. One spare is your insurance for when an item has to go after reliability or factor analysis.

3How to write Likert items that hold up

According to Dillman, Smyth & Christian (2014), a good question is one that everyone understands in the same way. A quick comparison:

AvoidPrefer
“The price is fair and the staff are friendly.” (double-barreled)Two items: “The price is fair.” / “The staff are friendly.”
“Don’t you agree the service is excellent?” (leading)“The quality of service meets my expectations.”
“I am not dissatisfied.” (double negative)“I am satisfied with the service.”
“The app has optimal UX.” (jargon)“The app is easy to use.”
“I always shop online.” (absolute)“I often shop online.”
  • Write in the first person, as statements (“I…”), so respondents rate their agreement.
  • Keep to one idea per item, under about 20 words, in words your respondents use.
  • Use reversed items only if the original scale has them, and remember to reverse-code them before analysis.

45-point or 7-point Likert scale?

A 5-point Likert scale and an example of splitting a double-barreled item into two
A 5-point Likert scale and how to fix a double-barreled item.
Code5-point Likert (most common)
1Strongly disagree
2Disagree
3Neutral
4Agree
5Strongly agree
  • 5 points: easy to answer and easy to compare with earlier studies. The safe choice for most theses.
  • 7 points: finer distinctions, suits respondents used to surveys, and is the right choice if the original scale uses it.
  • Never mix 5 and 7 points within one scale.
🔢
Related
Data coding and variable names in SPSS
Likert codes, reverse-coded items and value labels.
→

5Putting your questionnaire on Google Forms properly

  • Use a linear scale or a multiple-choice grid for Likert items, with both ends labelled.
  • Mark scale items as required to avoid missing data.
  • Split sections by construct, and use “Go to section based on answer” for the screening question.
  • Limit free-text “Other…” options, which are painful to code later.
  • Keep question text short: it becomes the column name when you export.
🧪 Test the form before launch

Make a copy of the form and fill it in to check the section logic, required items and the exported file. For long, multi-section forms, the form filling tool can run test submissions against that copy. Test data is only for checking the form: delete it before launch and never use it as research data.

610 common questionnaire design mistakes

MistakeWhat it does to your analysis
No items for a variable in the modelThe related hypothesis cannot be tested
Only 1–2 items per constructUnstable alpha; CFA/SEM struggles to identify
Double-barreled or double-negative itemsLow item-total correlation, lower alpha
Near-identical items across two constructsCross-loadings in EFA, high HTMT
Reversed items not flaggedForgotten recoding → low or negative alpha
Mixing 5-point and 7-point scalesConstruct means cannot be combined
No screening questionWrong respondents, biased results
Too long (over 15–20 minutes)Drop-off and straight-lining near the end
Overlapping income or age ranges (“5–10”, “10–15”)One person fits two groups
Skipping the pilotErrors found when it is too late

7Save time on the first draft

Turning a model into a structured list of questions is usually the slowest part. The Light Data’s AI questionnaire builder takes the model you draw (or a saved .json file), drafts Part I, general information, and Part II, a 5-point Likert scale for each construct, and exports a Word file for you to keep editing.

⚠️ Use AI responsibly

Treat an AI-drafted questionnaire as a first draft: check it against the original scales, adjust the wording for your respondents and verify every citation yourself before using it.

🎯 Checklist before launch
  • Every variable in the model has at least 3 items from a scale with a clear source.
  • Items are short, single-idea and not leading; reversed items are flagged.
  • There is a screening question and an attention check; it takes under 15 minutes.
  • You have run a pilot and checked the file Google Forms exports.
👥
Read next
How many respondents are enough?
Hair, Tabachnick, Yamane and Cochran formulas with examples.
→

Frequently asked questions

Should I use a 5-point or 7-point Likert scale?

Five points is the most common: easy to answer and easy to compare with earlier studies. Seven points discriminates more finely. What matters is using one format consistently within a scale, and keeping the original format if you adapt an existing scale.

How many questions does each construct need?

At least 3 items per construct so that Cronbach’s alpha, EFA and CFA run reliably; 3 to 6 is typical. Include one spare, since you may have to drop items after testing.

Can I write my own scale?

Prefer adapting a validated scale from earlier research, translating and adjusting it for your context, and citing the source. A scale you build yourself needs qualitative work and a pilot test to show it is reliable.

Do I need a pilot test?

Yes. Send the questionnaire to about 20–30 people from your target group to catch confusing items and overly long completion times, and run a first Cronbach’s alpha before the main survey.

Can I use a questionnaire drafted by AI in my thesis?

Use it as a first draft. AI speeds up writing items from your model, but you still need to check them against the original scales, adjust the wording and verify every citation yourself.