Tag: statistics

Three questions decide your statistical test: what your research question is asking, what type of data your dependent variable is, and how many groups you are comparing. Answer those three in order and the choice narrows to one or two tests. Choosing the test first and forcing the question to fit is the error that sends most methodology chapters back for revision.

This guide works through those three questions and gives you a decision table you can check your own study against.

Question 1: What is your research question actually asking?

Statistical tests answer four broad kinds of question, and almost every FYP falls into one of them:

• Is there a difference between groups? Do male and female students differ in academic stress? Did the intervention group outperform the control group?
• Is there a relationship between variables? Does study time relate to results? Does service quality relate to satisfaction?
• Does one thing predict another? Do compensation and work environment predict employee performance?
• Is there an association between categories? Is faculty membership associated with preferred learning mode?

Write your research question out and underline the verb. Words like differ, compare and effect of an intervention point to difference tests. Words like relationship, association and correlate point to correlation. Words like predict, influence and determine point to regression.

Question 2: What type of data is your dependent variable?

This is the question students most often get wrong, and it controls everything downstream.

• Nominal. Unordered categories: faculty, gender, employment status, yes or no answers.
• Ordinal. Ordered categories without equal intervals: ranking, satisfaction levels, a single Likert item.
• Interval or ratio (continuous). Numbers where the distance between values is meaningful: age, income, test score, or a summed scale score.

One point causes endless confusion in FYP work: a single Likert item is ordinal, but a summed or averaged scale built from several Likert items measuring the same construct is routinely treated as continuous in the social sciences. That is why studies using validated multi-item instruments can legitimately run correlation and regression on scale scores. If you are analysing one item on its own, treat it as ordinal.

Question 3: How many groups or variables are involved?

For difference questions, count the groups and note whether they contain the same people:

• Two independent groups. Different people in each, for example male and female students.
• Two related measurements. The same people measured twice, for example before and after training.
• Three or more independent groups. For example students from four faculties.

For prediction questions, count your independent variables: one means simple regression, more than one means multiple regression.

The decision table

When do you need the non-parametric alternative?

The tests in the fourth column assume, among other things, that your continuous outcome is reasonably normally distributed within groups. When that assumption fails badly, you move to the alternative in the fifth column, which works on ranks instead of raw values.

You would typically switch when a normality test on your data indicates a clear departure from normality, when your sample is small, or when your dependent variable is genuinely ordinal rather than continuous. The trade-off is real: non-parametric tests are more robust but generally have less power to detect an effect that exists.

Two cautions worth knowing. First, normality tests become very sensitive in large samples and can flag trivial departures, so inspect a histogram rather than relying on the test alone. Second, the assumption concerns the distribution within groups, not the shape of your whole dataset lumped together.

What else must you check before reporting?

1. Independence of observations. Each participant contributes one response, unless you are deliberately using a paired or repeated-measures design.
2. Homogeneity of variance for t-tests and ANOVA. Most software reports this alongside the test, and there are corrected versions to use when it fails.
3. Linearity for correlation and regression. A scatterplot answers this in seconds and can reveal a strong curved relationship that a correlation coefficient would report as near zero.
4. Multicollinearity for multiple regression. Predictors that are too highly correlated with each other make individual coefficients unstable and hard to interpret.
5. Expected cell counts for chi-square. When expected counts fall too low, the test becomes unreliable and Fisher’s exact test is the usual remedy.

Report these checks in your results chapter. Examiners look for them, and a study that reports a failed assumption and explains how it was handled reads as more competent than one that quietly ignores it.

Two worked examples

Example 1. “Is there a significant difference in academic stress between first-year and final-year students?” The question asks about difference. The dependent variable is a summed stress scale, so continuous. There are two independent groups. That gives an independent samples t-test, or Mann-Whitney U if the stress scores are badly skewed.

Example 2. “Do service quality and price fairness influence customer loyalty?” The verb is influence, so this is prediction. The outcome is a continuous loyalty scale. There are two predictors. That gives multiple linear regression, with linearity and multicollinearity checked before the results are interpreted.

The mistake that causes most revisions

Running a test that does not match the research question. A study whose research question asks whether a relationship exists, but whose analysis chapter reports group comparisons, will be sent back regardless of how correctly the arithmetic was done.

The fix is a single sentence in your methodology chapter that connects the two explicitly: “Because research question two asks whether service quality predicts customer loyalty, multiple linear regression was used, with loyalty as the dependent variable and service quality and price fairness as predictors.” That sentence tells the examiner you chose deliberately rather than by habit.

Does the level of your degree change the test?

No. The same tests are available at every level, and a well-executed t-test at doctoral level is not a weakness if it answers the question. What rises with the level is how thoroughly you justify the choice and how carefully you discuss its limitations. The differences between the levels are set out in the comparison of FYP, thesis and dissertation in Malaysia.

Once the test is settled, it needs to be written into your methodology chapter alongside your design, population and instrument. The guide on how to write the methodology chapter covers that structure step by step.

Writing up your analysis so it reads as a decision

Choosing correctly is half the work. The other half is writing the analysis so a reader can follow why each step happened, from assumption checks through to interpretation, without the chapter contradicting itself.

Tesify helps you draft and structure the methodology and results chapters so that the test you chose, the assumptions you checked, and the way you report the outcome stay consistent throughout. Tesify does not run your statistics; you do that in your own software and remain responsible for every number.

Draft your methodology and results chapters in Tesify

Frequently asked questions

Can I use a t-test for three groups by running it three times?

You should not. Running repeated pairwise tests inflates the chance of finding a significant difference by chance. Use one-way ANOVA for three or more groups, then a post-hoc test to identify which specific pairs differ.

What is the difference between correlation and regression?

Correlation measures the strength and direction of a relationship without assigning direction of influence. Regression models one variable as an outcome predicted by others and gives you an equation. If your hypothesis says one thing affects another, regression matches it better.

Can I treat Likert data as continuous?

A summed or averaged multi-item scale is commonly treated as continuous in social science research. A single Likert item is ordinal. State plainly in your methodology chapter which you are doing and cite the convention you are following.

How do I know whether my data is normally distributed?

Combine evidence rather than relying on one signal: inspect a histogram, look at skewness and kurtosis values, and run a normality test. In larger samples the formal test can flag departures too small to matter, so the visual check carries real weight.

What if my results are not significant?

Report them exactly as they are and discuss why. A non-significant result is a legitimate finding. What will damage you in the viva is being unable to explain it, or worse, adjusting data to produce significance, which is a serious breach of academic integrity.

Do I need to report effect size?

It is increasingly expected and always strengthens your discussion. Significance tells you whether an effect is likely to be real; effect size tells you whether it is large enough to matter practically. Most software reports it, sometimes only when you enable the option.

Which software should I use?

Use whatever your university licenses and your supervisor recognises. All mainstream statistical packages produce identical results for these standard tests because the underlying formulas are the same. Familiarity matters more than features when a deadline is close.

My sample is only 30 respondents. Which tests can I use?

Standard tests still run, but they have limited power to detect effects, and normality assumptions matter more at small sample sizes. Non-parametric alternatives are often the safer choice. State the sample size as a limitation and avoid overstating what your results establish.

Can I use chi-square with continuous data?

Not directly. Chi-square works on frequency counts in categories. Some students convert continuous data into categories to force it, which throws away information and is generally discouraged. Use a test suited to continuous data instead.

What is the difference between one-tailed and two-tailed tests?

A two-tailed test looks for a difference in either direction; a one-tailed test looks in a specified direction only. Use one-tailed only when theory genuinely justifies predicting the direction in advance, and state that justification. Two-tailed is the safer default and the usual expectation.