Answers are grounded in this paper and cite their sources.
It is common practice in data analysis to split measurements that were taken on a continuous scale into two categories such as “low” and “high.” Yet many researchers are not aware of the costs and negative consequences when continuous variables are dichotomized.1 One research team estimates that for normally distributed data, one would need 158 observations of a dichotomized variable to get the same information as only 100 observations of a continuous variable.2 This article illustrates the particular costs of loss of statistical power and decreased effect size through an example using simulated data. Suppose we are examining the relationship between age and functional impairment in a certain population, where functional impairment is assessed using measurements on a continuous scale, standardized so that the population mean is 100. Figure 1 shows a scatterplot of the relationship between age and functional impairment in a hypothetical random sample of 50 patients from the target population. The Pearson's correlation coefficient here is calculated to be r = .31, which is typically considered to be a medium effect. The 95% confidence interval is calculated to be (0.04, 0.54), and in a test of the hypothesis that the true population correlation is zero (vs non-zero) the P value for this sample is .027, which is typically considered to be statistically significant. Using regression to fit a linear model using age to explain functional impairment yields an intercept value of 81.0 and a regression coefficient of 0.36. This regression coefficient tells us that we expect the functional impairment score to increase by 0.36 points for every extra year of a patient's age. Put more interpretably: we expect patients' functional impairment score to rise by about 3.6 points over every decade. The R-squared value for this relationship is 0.097 (found by squaring the correlation r), which suggests that about 9.7% of the variability in functional impairment in this group of patients can be explained by their age. Some researchers might be inclined to categorize these measurements into groups, such as “low functional impairment” and “high functional impairment,” rather than leaving them as numbers on a continuous scale. This dichotomization can take on different forms, but the most common is to use either a “median split” (where the results are dichotomized around the median of the data) or to split the results into two groups based on a cut-point with clinical relevance. Yet this would lead us to lose information about the true relationship between the two variables, which often translates into a loss of statistical power as well as a decreased effect size, as will be seen later. Suppose that in our hypothetical study we decide to split our patient sample according to whether they are below 50 years of age or 50 years and above. Figure 2 shows the data under this new dichotomization. We can see that the fine-grained information about the relationship between social support and functional impairment has been lost. The mean functional impairment score in the older group is 101.0 (SD = 14.1), compared to a mean score of 96.4 (SD = 19.4) in the younger group. The sample mean difference is 4.8 points, with a 95% confidence interval of (−4.8, 14.5); a t-test of the hypothesis that the true mean difference is zero (vs non-zero) gives a t-statistic of 1.01 with 44° of freedom and P value of .319, which is not considered to be statistically significant. The Cohen's d effect size for this difference is calculated to be 0.29, which is typically considered to be a small effect. Similarly, the point-biserial correlation (the correlation between a dichotomous variable and a continuous variable) is calculated to be 0.14, which is typically considered to be a small effect, with only about 2% of the variability in these patients' functional impairment can be explained by whether they are younger than 50 years of age or not. We can therefore see that by dichotomizing one variable, the size of the effect dropped from medium to small, and the P value decreased in strength, from a significant result of .027 to a nonsignificant result of .319. We also lost explanatory value of the dichotomized variable: With the continuous measure, almost 10% of the variability in the patients' functional impairment scores could be explained by their raw age in years; with the dichotomized measure into “younger” and “older,” this R-squared value dropped to under 2%. With the dichotomization we have also lost valuable interpretative information. For example, a linear regression using the continuous variable tells us that every 10-point increase in a patient's age translates to an expected 3.6-point increase in functional impairment. The linear regression equation also lets us say, for example, that a 65-year-old patient would be expected to have a functional impairment score of about 104, whereas a 45-year-old patient would be expected to have a functional impairment score of 97 points. Dichotomizing age eliminates this fine-grained relationship and essentially treats a 51-year-old the same as an 81-year-old, and a 29-year-old the same as a 49-year-old. We can say that the functional impairment of patients 50 or older is on average about five points higher than of those younger than 50, but with this it could misleadingly appear that patients' functional impairment changes immediately after age 50. A clinician might be more interested in how steeply the functional impairment of their patients is expected to increase as they age year by year. Some researchers might also be tempted to dichotomize the other variable as well, classifying patients as either “low functional impairment” and “high functional impairment.” This also leads to a loss of information, as illustrated next. Suppose we used a functional impairment score of 100 to split patients into either low- or high-functional impairment groups, based on recommendations for clinical interpretation of the instrument. Table 1 shows the contingency table breakdown of the low/high and younger/older combinations for the two variables. Figure 3 shows a mosaic plot, where the size of each box is proportional to the number of patients in that group. This would be the fully dichotomized version of Figure 1. Here we can see that the box containing younger patients and low functional difficulty is larger than the others, indicating that most patients fell into this classification. A chi-square test of association between the two variables gives a test statistic of 0.33, with a P value of .567, which is not statistically significant. The Cramer's V effect size is 0.08, which is considered to be a negligible effect. We can therefore see that by dichotomizing the second variable, our P value has dropped further in significance, from .03 to .32 to .57, and our effect size has dropped from medium to small to negligible. We have also lost the ability to interpret the nuanced relationship between the two variables. Experts generally recommend avoiding dichotomization of continuous variables in data analysis, as it can lead to loss of statistical power, inappropriate effect size, and loss of explanatory information. In rare situations, such as when there are two distinct taxonomy classes underlying the continuous measure, a dichotomization is appropriate on theoretical grounds. In most cases, however, continuous variables should be left as numbers, not categories.
No takes yet. Share an insight, caveat, or question.
Regina Nuzzo (2019) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: