There is no universally correct number of respondents. Your sample size is calculated from three things: how large your population is, which analysis you plan to run, and whether your study is quantitative or qualitative. For a survey of a known population, a published table gives you the figure directly. For a regression, the number of predictors decides it. For interviews, saturation decides it. This answers each case, and shows you how to defend the number you land on.
Why Is There No Single Correct Sample Size?
Because “how many respondents” is really three different questions wearing one coat. A survey generalising to a population needs enough respondents that the sample’s percentages are close to the population’s — and that depends on how big the population is. A study testing relationships needs enough cases for the specific test to detect an effect if one exists — and that depends on the test and the number of variables. A qualitative study is not estimating a population value at all, so a large sample buys nothing; depth and saturation matter instead.
This is also why “my senior used 100, so I will use 100” is the wrong reasoning even when 100 turns out to be right. Examiners do not object to your number; they object to a number with no derivation behind it.
How Do You Calculate It for a Survey with a Known Population?
When you can state the size of your population — the 850 students in your faculty, the 320 staff in a company — use the sample-size table published by Krejcie and Morgan in 1970, which remains the standard reference in Malaysian methodology chapters. It assumes a 95% confidence level, a 5% margin of error and maximum variability, and it comes from this formula:
s = X²NP(1 − P) / [d²(N − 1) + X²P(1 − P)], where X² = 3.841, P = 0.5 and d = 0.05.
Because it is a formula, you should recompute the row you need rather than trusting a copy of the table found online — copies drift. The values below were recomputed from the formula for this article and reproduce the published table:
| Population (N) | Required sample (s) |
|---|---|
| 50 | 44 |
| 100 | 80 |
| 200 | 132 |
| 300 | 169 |
| 500 | 217 |
| 1,000 | 278 |
| 2,000 | 322 |
| 5,000 | 357 |
| 10,000 | 370 |
| 100,000 and above | 384 |
Two things this table teaches beyond its numbers. First, the required proportion shrinks fast: you need 88% of a population of 50, but 3.7% of a population of 10,000. Second, the requirement plateaus at 384 — a national study of 33 million people needs no more respondents than a study of 100,000. If someone tells you a Malaysian national survey needs “at least a few thousand”, the table disagrees.
What If You Do Not Know the Population Size?
Plenty of real populations are unknown or uncountable — online shoppers in Selangor, informal traders, users of a service with no register. Two honest routes exist. Use the large-population figure of 384, since the plateau means any sufficiently large unknown population lands there anyway. Or use the requirement of your planned analysis instead, described next, which is often the more useful constraint for an FYP.
What you must not do is invent a population figure to make a table row apply. If the population is unknown, say so in the methodology chapter and state which of the two routes you used — that sentence is a strength, not a weakness.
How Does Your Statistical Test Change the Number?
If your study tests relationships rather than estimating a percentage, the analysis sets the floor. These are the working conventions supervisors most commonly accept, and they should be stated as conventions, not laws:
- Multiple regression: commonly 10 to 15 cases per predictor as a working minimum, so a four-predictor model wants roughly 40 to 60 at absolute minimum and considerably more in practice.
- Comparing groups (t-test or ANOVA): aim for at least 30 in each group rather than 30 overall — a 200-respondent sample split unevenly into a group of 185 and a group of 15 is a small-sample problem wearing a large-sample number.
- Correlation: around 30 pairs to detect a strong relationship, but far more for a weak one, which is why small studies so often report “no significant relationship”.
- Chi-square: driven by cell counts rather than the total — the usual expectation is at least five expected cases in each cell of your table.
- Factor analysis or structural equation modelling: substantially larger, often several hundred; if this is your design, plan the sample before you plan anything else.
The rigorous version of all of these is a power analysis, computed from your chosen test, your alpha (usually .05), your desired power (usually .80) and the effect size you expect from prior studies. Free software such as G*Power does this in a few clicks, and a power analysis reported in your methodology chapter is the strongest possible justification. Which test you are sizing for is a prior decision — our guide to choosing the right statistical test settles that first.
What If Two Methods Give You Different Numbers?
They often will: a population of 200 gives 132 from the table, while a three-predictor regression might suggest 45. Take the larger one. Each method protects against a different failure — the table protects your ability to generalise, the power calculation protects your ability to detect an effect — and satisfying one while failing the other still leaves your study exposed. Report both figures and state that you adopted the higher; examiners read that as understanding rather than indecision.
How Many Interviews Does a Qualitative Study Need?
Qualitative sampling is purposive, not statistical, so no table applies and no percentage of the population is meaningful. The standard is saturation: you keep interviewing until new participants stop generating new codes or themes. In practice, undergraduate qualitative FYPs in Malaysia commonly work with somewhere between six and fifteen participants, and focus-group studies with three to five groups — but those are observations about typical projects, not requirements.
What makes a qualitative sample defensible is evidence rather than a number. Analyse as you collect rather than waiting until the end, keep a log of new codes produced by each interview, and you will be able to write the sentence an examiner wants: “No new codes emerged from the eleventh and twelfth interviews, and data collection was concluded at twelve participants.” That sentence beats any figure borrowed from another study, and it can only be written by someone who was coding along the way.
One caveat worth stating in your chapter: saturation is a claim about your own data and your own coding, so describe how you recognised it rather than simply asserting that it occurred.
How Many Extra Should You Distribute?
Your calculated number is how many usable responses you need, not how many forms to send. Distribute enough to absorb non-response and unusable returns — for student surveys in Malaysia it is common to distribute 30% to 50% above target, and more for online-only distribution, where response rates are usually lower than for administered paper forms. Then report all three numbers in Chapter 4: distributed, returned, usable. Instrument design affects these rates more than students expect, and our guide to designing a questionnaire covers the choices that keep completion high.
What If You Cannot Reach the Number?
This is common and it is survivable, provided you handle it honestly. Do not pad your dataset, duplicate responses, or quietly drop the target from your methodology chapter — all three are detectable and all three are far more damaging than a small sample. Instead: report what you achieved and the response rate; state the consequence plainly, which is usually reduced power and limited generalisability; reframe the study’s claims to match the sample you have, such as one faculty rather than a university; and record the shortfall in your limitations section. Examiners routinely accept a small sample that is honestly bounded, and routinely fail a large one that cannot be accounted for.
How Do You Justify Your Number to an Examiner?
Write three sentences in the sampling section of your methodology chapter: what the population is and how you know its size; which method you used to derive the sample size, named and referenced; and what the resulting target is, with the distribution allowance. For example: “The population comprised 850 final-year students in the faculty. Using the Krejcie and Morgan (1970) table, a minimum sample of 265 was required; 380 questionnaires were distributed to allow for non-response.” Then make sure the number in that sentence, the number in your results chapter and the number in your abstract are the same figure.
A defensible sample size is one you can explain in one sentence. Tesify helps you build a methodology chapter where your population, sampling and analysis stay consistent with each other — and with the results you eventually report. Your thesis, 100% written by you. Used by 9,000+ students. Plan your methodology in Tesify.
FAQ: Sample Size
How many respondents do I need for an undergraduate FYP?
Whatever your population and analysis require. For a faculty-sized population of a few hundred, the table typically lands between 130 and 220; for a relationship-testing study, the number of predictors may set a lower floor. Take the larger of the two.
Is 30 respondents enough?
Only for a pilot test or a very small known population. Thirty is the conventional minimum per group for some tests, not a target for a whole study, and a 30-respondent survey generalising to a faculty will not survive the sampling question.
What is the maximum sample size I ever need?
For estimating a proportion at 95% confidence and a 5% margin, 384. The requirement plateaus there, so a population of 100,000 and a population of 30 million need the same sample.
Do I need to use Krejcie and Morgan specifically?
No. It is the most commonly cited reference in Malaysian theses, but any documented method is acceptable — a power analysis, another published table, or your faculty’s stated convention. What is not acceptable is an undocumented number.
How many respondents for a pilot study?
Typically a small group drawn from the same population but excluded from the main sample. The pilot’s purpose is finding unclear items and testing reliability, not producing findings — the mechanics are covered in our questionnaire design guide.
Does my sample size change if my population is very small?
Yes, sharply upward as a proportion. Below roughly 50 people it is often simplest to survey everyone, which removes the sampling question altogether and is a perfectly respectable design called a census.
What is G*Power and do I need it?
Free software that calculates required sample size from your test, alpha, desired power and expected effect size. You do not have to use it, but a reported power analysis is the strongest justification available and takes minutes once you know your test.
How many participants for a focus group study?
Commonly three to five groups of six to ten participants each, continued until themes repeat. As with interviews, the defensible criterion is saturation rather than the count.
Can I combine a small survey with interviews to compensate?
A mixed-methods design is legitimate and often stronger, but it is not a repair for an underpowered survey. Design it as mixed methods from the start, with a stated rationale, rather than adding interviews after the response rate disappoints.
My response rate is low. Does that invalidate my study?
Not automatically, but it must be reported and discussed, because non-response can bias results if those who replied differ systematically from those who did not. Report the rate, note the risk, and where possible compare respondent demographics with the population.
Should the sample size appear in my abstract?
Yes — the method move of an abstract normally states the design and the number of participants, so a reader can judge the weight of the findings immediately.
Does a bigger sample always make a better thesis?
No. Past the calculated requirement, extra respondents add cost and time while adding little precision, and they can make trivial differences statistically significant. A well-designed instrument delivered to the right calculated number beats a larger careless sample every time.