In 1972, world-famous biologist Richard Lewontin published an analysis of human classification. Using alleles for 17 genes related to blood type (a state of the art technique at the time), he found that 85.4% of variability existed within racial populations, 8.3% of variation between groups within racial groups, and 6.3% of variation was between racial groups (Lewontin, 1972, pp. 396-397). Because so much of the variation was at the individual level, Lewontin concluded that,

Human racial classification is of no social value and is positively destructive of social and human relations. Since such racial classification is now seen to be of virtually no genetic or taxonomic significance either, no justification can be offered for its continuance

Lewontin (1972, p. 397)

Lewontin’s logic is superficially persuasive, and often people use it to argue against the scientific use of racial group classifications or the study of racial differences (e.g., Kaplan, 2015; Sternberg et al., 2005). I even remember reading it in my introductory psychology textbook two decades ago.

Over the years, people have criticized Lewontin’s (1972) logic. Most damaging is the fact that Lewontin did not take into account the correlations among alleles; taking this into consideration when calculating between-group differences. For this reason, Edwards (2003) calls the argument “Lewontin’s fallacy,” a term that has stuck.

The Statistical Interpretation

In statistical terms, Lewontin (1972) calculated an effect size, and the 6.3% of variation between racial groups is equivalent to an eta-squared effect size of .063. Likewise, the 8.3% of variation between groups within a racial group is equivalent to an eta-squared value .083. The eta-squared effect size is a measure of the total sample variance that can be explained by average group differences in the independent variable (Warne, 2021).

This statistical viewpoint is important for understanding Lewontin’s conclusion:

. . . human races and populations are remarkably similar to each other, with the largest part by far of human variation being accounted for by the differences between individuals.

Lewontin (1972, p. 397).

The reality is that the majority of variation will be due to individual-level differences every time eta-squared is less than .50. which I have pointed out before. There is nothing damning about this fact. In the social sciences the vast majority of eta-squared values are less than .50, and a high level of individual-level of variation is never of any concern. Lewontin (1972) stated a trivial statistical fact as if it were a profound truth.

The non sequitur

As if that were not enough of a problem, Lewontin then came to the conclusion that because group membership explained comparatively little variation that racial classifications were unjustified and not useful.

This is a non sequitur. Just because groups explain a low level of variation does not mean the groups are not useful ways of classifying people to understand them. For example, schools account for 10-15% of variation in students’ academic achievement (Detterman, 2016; Marks, 2015). This does not mean that schools are not important or that using school-level data to understand achievement is not justified.

Sorry, Lewontin. No one buys your non sequitur any more.

Similarly, in a meta-analysis of the relationship between psychosis and violence, Douglas et al. (2009, p. 690) found a median odds ratio of 1.71, indicating that a person with schizophrenia had a 71% higher odds of committing violence than a person in the general population. This is the mathematical equivalent of an eta-squared of .021, meaning that 2.1% of variance in violent behavior is between these groups (i.e., people with schizophrenia and non-diagnosed people). By Lewontin’s logic, this means that categorizing individuals as (1) people with schizophrenia and (2) non-diagnosed people of have “. . . virtually no . . . taxonomic significance either, [and] no justification can be offered for its continuance.”

The flaw in this logic is obvious. Using classrooms, schools, or diagnostic categories can be useful and justified, even with very small effect sizes. Knowing that people with schizophrenia have nearly twice as high odds of committing violence is very important information, despite the very small effect size. Likewise, it is still useful to compare schools and classroom averages (even if most variation is at the individual level) because most interventions are administered at the school and classroom levels. Moreover, the small effect sizes do not invalidate the existence of these groups. They are still real groups of people, even if they have little between-group variance. What’s true in these examples can be true about human racial and ethnic populations, too.

Lewontin’s argument relies on specious reasoning and improper interpretation of statistics. The reality is that small effect sizes can exist with real groups and that the amount of variance that group membership explains (as quantified by the eta-squared effect size) says little — if anything — about the usefulness of those groups. Nobody who makes Lewontin’s argument should be taken seriously.


Detterman, D. K. (2016). Education and intelligence: Pity the poor teacher because student characteristics are more significant than teachers or schools. The Spanish Journal of Psychology, 19, Article E93.

Edwards, A. W. F. (2003). Human genetic diversity: Lewontin’s fallacy. BioEssays, 25(8), 798-801.

Kaplan, J. M. (2015). Race, IQ, and the search for statistical signals associated with so-called “X”-factors: environments, racism, and the “hereditarian hypothesis”. Biology & Philosophy, 30(1), 1-17.

Lewontin, R. C. (1972). The apportionment of human diversity. In T. Dobzhansky, M. K. Hecht, & W. C. Steere (Eds.), Evolutionary biology (Vol. 6, pp. 381-398). Springer.

Marks, G. N. (2015). The size, stability, and consistency of school effects: Evidence from Victoria. School Effectiveness and School Improvement, 26(3), 397-414.

Sternberg, R. J., Grigorenko, E. L., & Kidd, K. K. (2005). Intelligence, race, and genetics. American Psychologist, 60(1), 46-59.

Warne, R. T. (2021). Statistics for the social sciences: A general linear model approach (2nd ed.). Cambridge University Press.