This post is an addendum to our post: The Google Memo: What Does the Research Say About Gender Differences. Please read the original post first, before consulting this one.

In this addendum we focus on the Greater Male Variability Hypothesis – the idea that men are more variable than women on a variety of abilities, interests, and personality traits – and the possibility that males are overrepresented in the upper and lower tails of such distributions. This hypothesis was first proposed by Ellis over 100 years ago, in 1894. It is also the hypothesis that Lawrence Summers was referring to in 2005 when, at the National Bureau of Economic Research Conference, he weighed in on the gender gap in STEM professions. Like Damore’s memo, Summers’ comments spurred controversy.

As we noted previously we think the central empirical claim of Damore’s memo was: “When addressing the gap in representation in the population, we need to look at population level differences in distributions.” Damore then attempted to clarify his position with this figure:

Yet, research on the Greater Male Variability Hypothesis suggests that Damore’s figure does not correctly capture the state of affairs for some variable. This because on a variety of abilities, interests, and personality traits, there is no mean difference, but males are more variable than females, and they are more likely to be overrepresented in the tails of a distribution. To visually demonstrate this we present Figure 1 from Hyde and Mertz (2009), which displays theoretical distributions for males and females and takes a closer look at the upper right tail of this theoretical distribution:

In Hyde and Mertz’s Figure 1, green represents the theoretical distribution for females, orange represents the theoretical distribution for males, and brown represents the area of overlap between the 2 theoretical distributions. Hyde and Mertz also present a magnified image of the upper tail (3.8 to 4.2 standard deviations above the mean) where green represents females, red represents males, and brown represents the overlap of the 2 theoretical distributions. The Greater Male Variability Hypothesis would also predict a similar ratio of males to females in the lower tail (3.8 to 4.2 standard deviations below the mean). In other words, according to the Greater Male Variability Hypothesis males are expected to be overrepresented in the upper and lower tails.

Greater male variability, particularly in spatial abilities, mechanical reasoning, and mathematics, is likely to be relevant to the hiring practices of tech companies in general, and Google in particular. Simply put, Google is one of the most successful companies in the world. Because of this, they are likely to be hiring employees from the upper tail of distributions measuring abilities that the company considers relevant and important. Furthermore, because Google is a tech company, it is likely that spatial abilities, mechanical reasoning, and mathematics skills are important for success, and are likely to play some role in the interview and screening process.

Thus, if males are overrepresented in the upper tail of the distributions for spatial abilities, mechanical reasoning, and mathematics, it would be possible for Google to end up hiring more males and, at the same time, not be discriminatory in their hiring practices. This is because the pool of potentially qualified applicants may contain more males.

For this post we set aside a discussion over what factors may produce greater male variability to focus on research that investigates this hypothesis with a variety of measures. We adopt the practice from our initial blog post. Findings that support the Greater Male Variability Hypothesis and emphasize potential biological factors for this variability are presented in green font, while findings that contradict the Greater Male Variability Hypothesis and emphasize cultural factors are presented in red font. (The idea here is that the currently greater male variability may be relevant to current hiring outcomes, but if the variability difference could be reduced to zero in a decade or two, then it counts against those who argue that biology may place a constraint on achieving future gender parity).

We have also emphasized other important areas of the text in bold font. We do this to draw attention to other important findings that may not be directly relevant to the Greater Male Variability Hypothesis, but are relevant to the controversy over the Damore memo. Finally, beneath each article citation we list the measure(s) employed by each study. We do this to demonstrate that a variety of abilities, interests, and personality traits have been subjected to empirical tests of the Greater Male Variability Hypothesis.

The Research:

Our Conclusions about the Greater Male Variability Hypothesis:

1. On average, male variability is greater than female variability on a variety of measures of cognitive ability, personality traits, and interests. This means men are more likely to be found at both the low and high end of these distributions (see Halpern et al., 2007; Machin & Pekkarinen, 2008 and, especially, the supplementary materials; for an ungated summary click here). This finding is consistent across decades.
2. The gender difference in variability has reduced substantially over time within the United States and is variable across cultures. It is clearly responsive to social and cultural factors (see Hyde & Mertz, 2009; Wai et al., 2010); Educational programs can be effective. It is also clear that there are cultural/societal influences, as the male:female variability ratios can vary considerably across cultures (e.g., Machin & Pekkarinen, 2008).
3. While the gender difference in the male:female ratio for the upper tail of the distribution of math test scores (SAT, ACT) narrowed considerably in the United States in the 1980s, it appears to have remained steady since the early 1990s. This can be seen visually in Figure 1 from Wai et al. (2010):
   • Therefore at the top end of any distribution of test scores where men have higher variability, we’d expect men to make up more than 50% of the upper end of the tail. Thus, any company drawing from the top 5% is likely to find a pool that contains more males. As one goes further out into the tail (i.e. becomes even more selective) the gender tilt becomes larger

1. Further compounding the gender tilt: the women in this elite group generally have much better verbal skills than the men in that elite group (see Reilly, 2012). This means that these women may be better employees than men who match them on quantitative skills, but because they have such superior verbal skills they have more choices available to them when selecting a profession.

Our Revised Conclusions About the Damore Memo

We maintain that the research findings are complicated. This is evident in both this post and our original one. There are many abstracts containing both red and green text, and some of the top researchers in psychology are represented on both sides of the debate. Furthermore, many of the experts have concluded that:

… early experience, biological factors, educational policy, and cultural context affect the number of women and men who pursue advanced study in science and math and that these effects add and interact in complex ways. There are no single or simple answers to the complex questions about sex differences in science and mathematics (Halpern et al., 2009).

In light of of the research on the Greater Male Variability Hypothesis however, we have revised our original conclusions:

1. Gender differences in math/science ability, achievement, and performance are small or nil. (See especially the studies by Hyde; see also this review paper by Spelke, 2005). There are two exceptions to this statement:
   • Men (on average) score higher than women on most tests of spatial abilities, but the size of this advantage depends on the task and varies from small to large (e.g., Lindberg et al., 2010). There is at least one spatial task that favors females (spatial location memory; see e.g., Galea & Kimura, 1993; Kimura, 1996; Vandenberg & Kuse, 1978). Men also (on average) score higher on mechanical reasoning and tests of mathematical ability, although this latter advantage is small. Women get better grades at all levels of schooling and score higher on a few abilities that are relevant to success in any job (e.g., reading comprehension, writing, social skills). Thus, we assume that this one area of male superiority is not likely to outweigh areas of male inferiority to become a major source of differential outcomes.
   • There is good evidence that men are more variable on a variety of traits, meaning that they are over-represented at both tails of the distribution (i.e., more men at the very bottom, and at the very top), even though there is no gender difference on average. Thus, the pool of potentially qualified applicants for a company like Google is likely to contain more males than females. To be clear, this does not mean that males are more “suited” for STEM jobs. Anyone located in the upper tail of the distributions valued in the hiring process possesses the requisite skills. Although there may be fewer women in that upper tail, the ones who are found there are likely to have several advantages over the men, particularly because they likely have better verbal skills.
2. Gender differences in interest and enjoyment of math, coding, and highly “systemizing” activities are large. The difference on traits related to preferences for “people vs. things” is found consistently and is very large, with some effect sizes exceeding 1.0. (See especially the meta-analyses by Su and her colleagues, and also see this review paper by Ceci & Williams, 2015).
3. Culture and context matter, in complicated ways. Some gender differences have decreased over time as women have achieved greater equality, showing that these differences are responsive to changes in culture and environment. But the cross-national findings sometimes show “paradoxical” effects: progress toward gender equality in rights and opportunities sometimes leads to larger gender differences in some traits and career choices. Nonetheless, it seems that actions taken today by parents, teachers, politicians, and designers of tech products may increase the likelihood that girls will grow up to pursue careers in tech, and this is true whether or not biology plays a role in producing any particular population difference. (See this review paper by Eagly and Wood, 2013).

In conclusion, based on the meta-analyses we reviewed and the research on the Greater Male Variability Hypothesis, Damore is correct that there are “population level differences in distributions” of traits that are likely to be relevant for understanding gender gaps at Google and other tech firms. The differences are much larger and more consistent for traits related to interest and enjoyment, rather than ability. This distinction between interest and ability is important because it may address one of the main fears raised by Damore’s critics: that the memo itself will cause Google employees to assume that women are less qualified, or less “suited” for tech jobs, and will therefore lead to more bias against women in tech jobs. But the empirical evidence we have reviewed should have the opposite effect. Population differences in interest and population differences in variability of abilities may help explain why there are fewer women in the applicant pool, but the women who choose to enter the pool are just as capable as the larger number of men in the pool. This conclusion does not deny that various forms of bias, harassment, and discouragement exist and may contribute to outcome disparities, nor does it imply that the differences in interest are biologically fixed and cannot be changed in future generations.

If our conclusions are correct then Damore was drawing attention to empirical findings that seem to have been previously unknown or ignored at Google, and which might be helpful to the company as it tries to improve its diversity policies and outcomes.

We again thank Alice Eagly for her helpful input.