Picking the wrong statistical test can quietly ruin an otherwise solid research project. In population and family health studies, where data often ranges from yes/no responses on contraceptive use to precise measurements of birth weight, the choice of test is never just a technicality. It is dictated by something more fundamental: the scale of measurement on which your variable sits. Stanley Smith Stevens laid out four such scales – nominal, ordinal, interval, and ratio – and each one permits only certain mathematical operations. Once you understand which tests are admissible for which scale, your analysis becomes both statistically defensible and meaningful for real-world health decisions.
Table of Contents
- Why the measurement scale dictates the test
- Statistical tests for nominal scales
- The chi-square test
- What you can and cannot do with nominal data
- Statistical tests for ordinal scales
- Mann-Whitney U test
- Wilcoxon signed-rank test
- Kruskal-Wallis H test
- Spearman’s rank correlation
- Statistical tests for interval and ratio scales
- The t-test
- The Z-test
- ANOVA
- Pearson’s correlation and linear regression
- The assumptions behind parametric tests
- A quick guide to matching tests with scales
- Common mistakes researchers make
Why the measurement scale dictates the test
A measurement scale is not just a label for your data. It tells you which arithmetic operations make sense. You can count categories of religion, but you cannot calculate their “average.” You can rank patient satisfaction from “very poor” to “excellent,” but the gap between “good” and “very good” is not necessarily the same as the gap between “poor” and “fair.” These properties decide whether a test like the mean – and the parametric machinery built around it – is even meaningful.
The scales are hierarchical. Nominal data only names categories. Ordinal data ranks them. Interval data has equal gaps between values but no true zero. Ratio data has all the above plus a meaningful zero, which allows ratios like “twice as many.” As you can read in this clear breakdown of the four scales, this order matters because higher scales unlock more powerful statistical operations.
Researchers often split tests into two broad families: parametric and non-parametric. Parametric tests assume the data follows a known distribution (typically normal) and require interval or ratio measurement. Non-parametric tests make far fewer assumptions and are designed for nominal or ordinal data, or for situations where parametric assumptions break down. A recent biostatistics review warns that misusing these tests is widespread and leads to flawed conclusions – a serious concern when the findings inform maternal health policy or family planning programmes.
Statistical tests for nominal scales
Nominal scales sort observations into categories with no inherent order. Think of variables like blood group (A, B, AB, O), method of contraception used (pills, IUD, condoms, sterilisation), or religion. You cannot rank these, only count how many cases fall into each bucket. Therefore, only tests that work with frequencies are admissible.
The chi-square test
The chi-square (ฯยฒ) test is the workhorse for nominal data. It examines whether two categorical variables are associated. For example, you might ask: is the type of contraceptive method used (nominal) associated with the woman’s place of residence – urban or rural (also nominal)? The test compares observed frequencies with what would be expected if there were no relationship. According to guidance in StatPearls, the chi-squared test typically requires a sample size of more than 60 subjects with at least five expected counts in each cell of the contingency table.
Two close relatives fill in where chi-square cannot. Fisher’s exact test is used with small samples or when expected cell counts fall below five. The McNemar test is for paired nominal data – for instance, comparing whether the same group of women used a particular family planning service before and after an awareness campaign.
What you can and cannot do with nominal data
You can calculate frequencies, proportions, percentages, and identify the mode. You cannot compute a mean or median. Visualisation is restricted to bar charts and pie charts. Any attempt to apply a t-test or correlation to raw nominal data is statistically meaningless.
Statistical tests for ordinal scales
Ordinal scales rank observations but do not guarantee equal spacing between ranks. Common examples in health research include Likert-scale responses on patient satisfaction (very dissatisfied to very satisfied), severity of illness (mild, moderate, severe), socio-economic status categories, and pain ratings. Because the distances between ranks are not uniform, calculating a mean is conceptually shaky. Non-parametric tests that rely on ranks rather than raw values are the appropriate choice.
Mann-Whitney U test
The Mann-Whitney U test (also called the Wilcoxon rank-sum test) compares two independent groups on an ordinal outcome. It is the non-parametric counterpart of the independent samples t-test. As described in this guide to biostatistics, the test compares ranks of data rather than raw values, assessing whether one group tends to have systematically higher or lower ranks than the other. A typical application: comparing satisfaction ratings with antenatal care services between women attending public versus private health facilities.
Wilcoxon signed-rank test
When the data is paired – say, the same patients rating their quality of life before and after a community health intervention – the Wilcoxon signed-rank test is appropriate. It is the non-parametric equivalent of the paired t-test.
Kruskal-Wallis H test
To compare three or more independent groups on an ordinal outcome, researchers use the Kruskal-Wallis H test, which extends the Mann-Whitney logic to multiple groups. For instance, comparing perceived barriers to family planning across women from three different income strata.
Spearman’s rank correlation
For relationships between two ordinal variables, Spearman’s rank correlation measures the strength and direction of association. It works on ranks, making it ideal when you want to know, for example, whether higher education levels (ordinal) correlate with more positive attitudes toward delayed marriage (also ordinal).
Statistical tests for interval and ratio scales
Interval and ratio scales provide numerical data with meaningful distances between values. Interval scales lack a true zero (temperature in Celsius is the classic example). Ratio scales include an absolute zero, enabling statements like “twice as heavy” – height, weight, age, blood pressure, haemoglobin levels, fertility rates, and infant mortality rates all live here. Because the data is continuous and arithmetic operations are valid, the full machinery of parametric statistics becomes available, provided assumptions are met.
The t-test
The t-test compares means and comes in three flavours. The one-sample t-test checks whether a sample mean differs from a known population value – useful when comparing the mean birth weight in a district hospital against the national average. The independent samples t-test compares means of two unrelated groups, such as haemoglobin levels in pregnant women who received iron supplementation versus those who did not. The paired t-test compares the same group before and after an intervention, like blood pressure before and after a lifestyle counselling programme.
The Z-test
The Z-test is structurally similar to the t-test but is used when the sample size is large (typically n > 30) and the population standard deviation is known. In demographic research with large datasets – such as the National Family Health Survey – the Z-test is often applicable when comparing proportions or means across large samples.
ANOVA
When more than two groups need comparison, Analysis of Variance (ANOVA) tests whether the means differ significantly. For example, comparing average ages at first childbirth across four geographic regions. Performing multiple t-tests instead of ANOVA inflates the chance of false positives, which is a common pitfall flagged in comparisons of parametric and non-parametric procedures.
Pearson’s correlation and linear regression
To examine linear relationships between two continuous variables, use Pearson’s product-moment correlation. A classic example is the relationship between maternal age and birth weight. Linear regression goes further, modelling how one continuous variable predicts another and adjusting for confounders – for instance, predicting infant mortality rates from female literacy levels across districts.
The assumptions behind parametric tests
Parametric tests are powerful but come with strings attached. The four classical assumptions, as outlined in resources on statistical test assumptions, are: the dependent variable should be measured on an interval or ratio scale; observations should be independent; the data should be approximately normally distributed; and the variances across groups should be roughly equal (homogeneity of variance). When these assumptions are seriously violated, falling back to non-parametric alternatives is the safer route. Tools like the Shapiro-Wilk test help check normality before proceeding with a parametric analysis.
A quick guide to matching tests with scales
Here is a practical summary you can use as a mental checklist when planning your analysis:
Nominal data: Chi-square test of independence, Fisher’s exact test, McNemar test. Use these to test associations between categorical variables or compare proportions.
Ordinal data: Mann-Whitney U test (two independent groups), Wilcoxon signed-rank test (paired groups), Kruskal-Wallis H test (three or more groups), Spearman’s rank correlation (association between two ordinal variables).
Interval and ratio data (assumptions met): t-test, Z-test, ANOVA, Pearson’s correlation, and regression analysis. These provide the most statistical power when assumptions hold.
Interval and ratio data (assumptions violated): Switch to the non-parametric equivalent – Mann-Whitney U instead of independent t-test, Wilcoxon signed-rank instead of paired t-test, Kruskal-Wallis instead of one-way ANOVA, and Spearman’s correlation instead of Pearson’s.
Common mistakes researchers make
A few errors crop up repeatedly in published health research. Treating a Likert-scale item as if it were continuous and running a t-test on it is common but contested. Applying parametric tests without checking normality, especially in small samples, leads to misleading p-values. Using multiple t-tests instead of ANOVA when comparing several groups inflates Type I error. And reporting a “statistically significant” finding without considering whether the underlying scale even permits the operation can result in conclusions that look scientific but are mathematically hollow.
The lesson is straightforward. Before you open SPSS, R, or any statistical software, identify the measurement scale of every variable in your dataset. Match the scale to the admissible tests. Then verify the assumptions of the chosen test. This three-step discipline is what separates rigorous population and family health research from the kind of analysis that generates headlines but does not survive scrutiny.
What do you think? If you were studying contraceptive preferences across different educational backgrounds in your state, which combination of measurement scales and statistical tests would you choose first – and why? Have you ever come across a published study where the test seemed mismatched with the data type?
Leave a Reply