# Are ratio, interval, ordinal and nominal variables nested?

It seems to me that ratio, interval, ordinal and nominal variables nested in the following sense:

• A ratio variable is also interval, ordinal and nominal.

• An interval variable is also ordinal and nominal.

• An ordinal variable is also nominal

Are the above statements accurate? If not, are there any counter-examples where one or more of the above statements will not hold?

• I think it's a little odd to say an interval scale is nominal. For example, weight in pounds is interval and can hypothetically be any positive real number, which is an uncountable set, so it couldn't possibly have a 1-to-1 correspondence with a discrete set of names. Feb 2, 2012 at 4:29
• Well, suppose two people weigh 145 lb (ignoring decimals etc). Then, you could categorize the two people in the category of 'people weighing 145 lb' which would be a nominal scale as it identifies these two people with the 'name' of 'people weighing 145 lb'. By the way, isn't weight a ratio variable (you mentioned that it is interval)? Feb 2, 2012 at 4:39
• Weight is a ratio and an interval variable. The point is, many interval variables take on values that are real numbers, and are not restricted to integers (or anything with a 1-1 correspondence with integers, such as when you shave off a certain number of decimal places). In that case, the interval variable is not nominal. Feb 2, 2012 at 5:32
• These categories of measurement were originally developed to help guide people towards appropriate forms of analysis. From this point of view, discussions like the present one over what category is appropriate become counterproductive. If the type of the measurement is not perfectly clear, then you need a different way altogether to decide how to proceed with the analysis of your data. One reason we cannot be much more helpful here is that this question is really too abstract and open-ended to suit our format.
– whuber
Feb 2, 2012 at 16:47

Leaping in here, I generally disagree with the proposition.

Let's look at how the categories are defined, by accepted convention.

There is the nominal category, which includes variables such as Sex, Marital Status, Religion, etc. Here we have discrete categories, where there is no ordering to them. If we gave them numbers, we could swap those numbers around however we liked because they make no difference to the underlying features which define those categories. This swapping is not a feature of the other 3 categories, therefore nominal data are quite special.

Next, we have the ordinal category. This is where order is important, for example the top 3 placings in a race, but the spacing between the categories is not consistent. The time to cross the finish line between 1st and 2nd could be 10ms, then 3rd might cross 500ms later. Or we might split income into quartiles, where the range of income in each quartile is not consistent.

Some people do not distinguish much between nominal and ordinal variables, it really depends on the analysis one wishes to do. Both can be analysed by nonparametric tests like chi-square, and the interpretation does not depend on category ordering. However, for analyses like ordinal logistic regression, it does matter that one has ordinal variables.

Coming to interval and ratio variables, I will treat these the same because my argument here does not depend on whether there is a true zero point or not (so yes, in practice there is no difference in statistical treatment of interval and ratio variables for at least the main tests). For both interval and ratio variables, the key feature is that their values have true mathematical meaning, so we can do various types of operations on them, e.g. take differences, linear or nonlinear transforms, and the results of these operations are meaningful. Whereas for nominal or ordinal variables, you can't halve "Sex" or square "Income quantile" - this would be a nonsensical action.

The key features of these two variables should not be confused with convenience of measurement or convenience of analysis: while we may reduce height to m/cm or ft/in with our rulers, height can keep being measured at finer and finer distinctions - it is only the limits of our measuring tool (and the issue of significant places) that makes some interval and ratio variables appear to be categorical. Because interval and ratio variables are essentially infinitely divisible, I disagree that they can be considered ordinal or nominal. Sometimes we choose to make them categorical, e.g. by using age bands in questionnaires, but the underlying variable (in this case, age) is not categorical.

• Thank you for a thoughtful reply. I believe, though, that the "infinite divisibility" characterization fails to capture the essential distinction for statistics, which is that numerical differences (or ratios, respectively) are considered meaningful for interval and ratio variables, whereas they are not for ordinal variables. To add further to the confusion, I would add in the spirit of EDA that the nominal-ordinal-ratio distinction is not inherent in the data but is actually a modeling decision: nonlinear re-expression of values exemplifies the flexibility of this point of view.
– whuber
Feb 2, 2012 at 15:14
• @whuber, ah true. I forgot to add that in! /blush Feb 2, 2012 at 16:42
• No need for embarrassment: very few answers are perfect at the start. Comments are primarily a tool to promote ongoing improvement of answers. Please feel free to edit your reply if there are changes you would like to make.
– whuber
Feb 2, 2012 at 16:43

I also think "no".

You can say these types are in a hierarchy of sorts because (for example)

• "an interval variable can be recoded as an ordinal variable if you are prepared to lose some of its information" and
• "an ordinal variable can be treated as a nominal variable if you are prepared to lose some of its information"...
• etc.

However, my second dash point above (for example) is very different from "an ordinal variable is also a nominal variable" as per the original question. A hierarchy of sorts is not the same as being "nested".

Yes, you are correct. It's better to think of the levels of measurement as describing properties of data, rather than types of data.

All data that exhibit ratio properties also exhibit interval, ordinal, and nominal properties. And so forth.

It makes no difference that (e.g.) interval or ratio data can take on an infinite number of possible values — any individual datum will have a specific value, which can be considered in virtue of its ordinal properties or nominal properties.