01 · The Question
If Categories Are Easier to Interpret, Why Not Create Them?
Suppose participants' ages range from 18 to 72 years. You could analyze age in its original form, or divide participants into groups such as 18–29, 30–44, 45–59, and 60 or older. Similarly, a researcher might turn examination scores into pass/fail, blood pressure into normal/high, or an attitude score into low/moderate/high.
The resulting categories can certainly be easier to describe. But statistical convenience comes with a trade-off: observations that originally had different values are now treated as belonging to the same group, while observations separated by a cutoff may be treated as categorically different even when their original values were almost identical.
The question is therefore not whether a continuous variable can be categorized. It can. The more important question is whether doing so helps answer the research question enough to justify the information that is discarded.
03 · What You Need to Know
Categorization Changes the Variable, Not Just Its Label
What happens when a continuous variable is categorized?
A continuous variable represents quantitative values along a continuum. Categorization replaces those individual values with membership in a limited number of groups.
Suppose age is recorded in years and then classified as:
- 18–29 = younger;
- 30–49 = middle;
- 50 or older = older.
The original variable preserves distinctions such as 22 versus 28, 31 versus 47, and 51 versus 68. The categorized version does not. Every participant within a category receives the same group designation.
The new variable is therefore not simply the original variable with nicer labels. It has a different information structure. The distinction between continuous and categorical variables has practical consequences for summaries, models, and interpretation.
Dichotomization is the most extreme form of categorization
Dichotomization means dividing a continuous variable into only two groups. Examples include high/low, pass/fail, positive/negative, or above/below a selected threshold.
Altman and Royston have highlighted the statistical costs of this practice. When continuous observations are reduced to two categories, information about differences among individuals within each category is discarded. The resulting analysis may consequently have less statistical power than one that appropriately retains the continuous variable.
The apparent simplicity can therefore be expensive. Two participants with scores of 49 and 50 might be placed on opposite sides of a cutoff, while participants scoring 50 and 90 are treated identically as members of the same category.
Cut points create boundaries that may not exist in the phenomenon
Imagine a stress scale ranging from 0 to 100. A researcher defines scores below 50 as “low stress” and scores of 50 or above as “high stress.”
Someone scoring 49 and someone scoring 50 are nearly identical on the original scale, yet categorization places them in different groups. Meanwhile, people scoring 50 and 95 are placed together.
If there is no substantive reason to believe that 50 represents a genuine threshold, the distinction is imposed by the researcher rather than discovered in the phenomenon.
Continuous representation
Preserves differences among values throughout the observed range.
Categorical representation
Retains group membership but discards distinctions among values within each category.
Data-driven cutoffs are particularly risky
A more serious problem occurs when researchers try several cut points and select the one that produces the strongest association, smallest p-value, or clearest group difference.
The chosen threshold is then partly determined by random features of the observed sample. Altman and colleagues have warned against selecting “optimal” cut points for continuous prognostic variables because this process can exaggerate apparent effects and produce results that do not replicate well in new data.
If a threshold is chosen after examining outcomes, that decision should not be presented as though it had been established beforehand on substantive grounds.
Watch Out
Do not search across multiple cut points and report only the threshold that produces the most favorable result. The apparent distinction may be a consequence of data-driven selection rather than a stable feature of the underlying relationship.
Median splits do not create naturally meaningful groups
A common strategy is to divide participants at the sample median into “low” and “high” groups. This guarantees two reasonably sized groups, but it does not establish that the median represents a meaningful substantive threshold.
The cutoff can also change from one sample to another. A participant classified as “high” in one study might be classified as “low” in another simply because the samples have different distributions.
Median splitting therefore answers a sample-relative classification question rather than preserving the original quantitative meaning of the variable.
Quantiles have descriptive uses, but they still discard information
Researchers also divide variables into tertiles, quartiles, or quintiles. These groups can be useful for descriptive presentation, particularly when showing how outcomes vary across a distribution.
However, quantile categories are determined by the sample distribution rather than by natural boundaries in the underlying construct. Two individuals with almost identical measurements can fall into adjacent quantiles, while substantial variation may remain hidden within a wide category.
Using categories for a descriptive table does not necessarily mean that the categorized version should replace the continuous variable in statistical modeling.
Established thresholds are different from arbitrary cutoffs
There are situations in which a threshold has substantive meaning. Clinical guidelines may define categories used for diagnosis, treatment decisions, or risk management. Educational systems may establish formal passing scores. Policy rules may determine eligibility according to specified thresholds.
If the research question concerns those categories themselves, analyzing category membership may be entirely appropriate.
For example, if a scholarship is available only to applicants whose income falls below an official eligibility threshold, the threshold has an institutional consequence. Researchers interested in eligibility may reasonably study the corresponding categorical variable.
Even in such cases, the original continuous measurement can remain valuable. A threshold may be meaningful for a decision while the underlying relationship between the measurement and an outcome remains continuous.
Keeping a variable continuous does not require assuming a straight-line relationship
Researchers sometimes categorize a predictor because they worry that treating it continuously assumes that every one-unit increase has the same effect throughout its range.
That concern can be legitimate, but categorization is not the only solution.
Regression models can represent nonlinear relationships using approaches such as polynomial terms, splines, or other flexible functions. These methods can preserve the continuous information while allowing the association between X and Y to change across the range.
The real choice is therefore not simply “linear continuous variable or categories.” Modern statistical modeling provides other possibilities.
Categorization can sometimes help communication without replacing the primary analysis
Categories may make a graph, table, or explanation easier for a nontechnical audience to understand. Researchers can sometimes use categories descriptively while retaining the original continuous variable for the primary statistical analysis.
For example, a paper might show descriptive outcomes by age bands for readability while modeling age continuously using an appropriately specified functional form.
This separates two legitimate purposes: communication and statistical estimation.
Creating categories changes how the variable should be described
Once a continuous variable has been categorized, the derived variable should be described according to its new structure.
Age in years may be continuous or treated as quantitative. Age grouped into 18–29, 30–44, and 45 or older is an ordinal categorical variable because the categories have a meaningful order. A pass/fail variable derived from an examination score is binary categorical.
The binary, nominal, ordinal, and continuous distinction therefore applies to the representation actually used in the analysis, not merely to the underlying characteristic from which it was derived.
04 · A Practical Example
What Gets Lost When a Score Becomes “Low” or “High”?
Hypothetical Example
Categorizing an academic-engagement score
A researcher measures academic engagement on a validated scale ranging from 20 to 100. The researcher wants to examine whether engagement predicts course completion and considers dividing the score at 60 into “low engagement” and “high engagement.”
Original measurement Students have scores such as 41, 52, 59, 60, 72, and 91. Their relative positions along the scale remain available.
After dichotomization Scores from 20 to 59 become “low,” while scores from 60 to 100 become “high.”
Information lost A student scoring 20 and another scoring 59 are now treated as members of the same category, while students scoring 59 and 60 are treated as different groups.
Better question Unless 60 has an established substantive meaning relevant to the research question, the researcher should consider retaining the original score and modeling its relationship with completion appropriately, including possible nonlinearity.
If 60 were an externally established threshold used for a real institutional decision, reporting results around that classification might be useful. But that practical classification would not erase the information contained in the underlying score.
06 · What This Means for You
When Should You Keep the Original Variable?
Begin by asking why you want categories. “It will make the table easier” and “this threshold determines an actual clinical decision” are very different justifications.
A simple decision framework
If there is no established substantive reason for a cutoff
Usually retain the continuous variable and model its relationship appropriately.
If a validated clinical, legal, policy, or institutional threshold is central to the research question
A categorical representation may be substantively meaningful, while retaining the original measurement where useful.
If you want categories only because you suspect nonlinearity
Consider flexible modeling of the continuous variable rather than automatically creating groups.
If categories are useful only for presentation
Use them descriptively if helpful, but consider keeping the continuous measurement for the primary analysis.
If you selected a cutoff after inspecting the outcome
Treat the result cautiously and disclose how the threshold was chosen rather than presenting it as predetermined.
If categorization is genuinely necessary, document the thresholds, their source, whether they were specified before examining the data, and what happened to observations exactly at the boundaries. These details are small until someone tries to reproduce the analysis. Then, as often happens in methods sections, the small details suddenly acquire tenure.
