Pair-bonding, Fatherhood, and the Role of Testosterone: A Meta-Analytic Review

Abstract

Males of many species must allocate limited energy budgets between mating and parenting effort. The Challenge Hypothesis provides a framework for understanding these life-history trade-offs via the disparate roles of testosterone (T) in aggression, sexual behavior, and parenting. It predicts that males pursuing mating opportunities have higher T than males pursuing paternal strategies, and in humans, many studies indeed report that men who are fathers and/or pair-bonded have lower T than childless and/or unpaired men. However, the magnitude of these effects, and the influence of methodological variation on effect sizes, have not been quantitatively assessed. We meta-analyzed 114 effects from 66 published and unpublished studies covering four predictions inspired by the Challenge Hypothesis. We confirm that pair-bonded men have lower T than single men, and fathers have lower T than childless men. Furthermore, men more oriented toward pair-bonding or offspring investment had lower T. We discuss the practical meaningfulness of the effect sizes we estimate in relation to known factors (e.g., aging, geographic population) that influence men's T concentrations.

Full text

1
Pair-bonding, Fatherhood, and the Role of Testosterone: A Meta-Analytic Review
Nicholas M. Grebe,1* Ruth E. Sarafin,2* Chance R. Strenth,2* & Samuele Zilioli3,4
1 Department of Evolutionary Anthropology, Duke University, Durham, NC, USA
2 Department of Psychology, University of New Mexico, Albuquerque, NM, USA
3 Department of Psychology, Wayne State University, Detroit, MI, USA
4 Department of Family Medicine and Public Health Sciences, Wayne State University, Detroit, MI,
USA
* Authors share co-first authorship and are listed in alphabetical order.
Send correspondence to Nicholas M. Grebe, Department of Evolutionary Anthropology, Campus
Box 90383, Durham, NC, 27708. Email: nicholas.grebe@duke.edu
31 December 2018
Word Count (text): 10,323

2
Abstract
Males of many species must allocate limited energy budgets between mating and parenting effort.
The Challenge Hypothesis provides a framework for understanding these life-history trade-offs via
the disparate roles of testosterone (T) in aggression, sexual behavior, and parenting. It predicts that
males pursuing mating opportunities have higher T than males pursuing paternal strategies, and in
humans, many studies indeed report that men who are fathers and/or pair-bonded have lower T
than childless and/or unpaired men. However, the magnitude of these effects, and the influence of
methodological variation on effect sizes, have not been quantitatively assessed. We meta-analyzed
114 effects from 66 published and unpublished studies covering four predictions inspired by the
Challenge Hypothesis. We confirm that pair-bonded men have lower T than single men, and fathers
have lower T than childless men. Furthermore, men more oriented toward pair-bonding or offspring
investment had lower T. We discuss the practical meaningfulness of the effect sizes we estimate in
relation to known factors (e.g., aging, geographic population) that influence men’s T concentrations.
Key words: testosterone; pair-bonding; challenge hypothesis; fatherhood; life-history theory; meta-
analysis

3
1. Introduction
The hormone testosterone (T) possesses a wide range of physiological and psychological
functions across vertebrates. However, much scientific and widespread interest in T focuses on its
role in promoting male-typical behavior across species (see Fine, 2017 and Sapolsky, 2017, for two
recent popular science examples). These psychological functions of T may be situated within a
broader theoretical framework regarding the evolutionary biology of the endocrine system. Within
this framework, diverse functions of T are conceived of as intertwined components of an adaptive
resource allocation strategy: given environmental conditions (e.g., abundance of resources, risk of
extrinsic mortality), individual conditions (e.g., mutation load, susceptibility to infection), and finite
resource budgets, an organism must maximize its fitness by modifying the allocation of energy and
effort towards certain classes of activities (e.g., growth, somatic maintenance) at the expense of
others (e.g., reproduction) (Del Giudice, Kaplan, & Gangestad, 2015). Hormones, including T, may
constitute a major biological mechanism by which these coordinated trade-offs are achieved
(Ketterson & Nolan, 1992). In particular, T may underlie variation in male reproductive strategies,
due to its functions as a coordinating biological messenger that facultatively adjusts behavior,
morphology, and physiology to secure reproductive opportunities and, in so doing, increase
reproductive fitness (Wingfield et al., 1990; Gettler et al., 2011).
1.1 Reproductive Fitness Trade-Offs in Males
Reproductive fitness—an individual’s success in passing on genes to the next generation—
can be increased through investment in either mating effort or parenting effort. With mating effort,
individuals emphasize finding and attracting partners and competing with rivals for these mating
opportunities. This strategy, across males of many species, is characterized by increased aggressive
behavior, risk-taking, and investment in costly ornamentation (Archer, 2006; McGlothlin, Jawor, &
Ketterson, 2007; Ligon, Thornhill, Zuk, & Johnson, 1990; Parker, Knapp, & Rosenfield, 2002;

4
Setchell, Smith, Wickings, & Knapp, 2008; Rose, Holaday, & Bernstein, 1971; Muller & Wrangham,
2004). In contrast, individuals allocating more energy to parenting effort invest time and resources
into long-lasting, stable mating relationships and care for offspring, either through provisioning or
direct involvement in child rearing (Ziegler & Snowdon, 2000; Kaplan & Lancaster, 2003;
Fernandez-Duque, Valeggia, & Mendoza, 2009). Although both strategies can increase reproductive
fitness, there is an inherent trade-off between the two, such that allocation of energy toward one
strategy reduces the pool of resources available for involvement in the other.
Relative to females, male vertebrates typically have a smaller obligate investment in
offspring, permitting a higher degree of flexibility in the optimal balance between mating and
parenting investment (Trivers, 1972; for counterarguments, see Kokko & Jennions, 2008). This
balance can vary both between males and within males, depending on ecological conditions, social
context, and genetic variation. Males in a highly ‘fit’ condition (whether due to intrinsic advantages
such as favorable genes, advantageous environmental conditions that result in energetic surpluses,
or both; Maynard-Smith, 1989) may obtain the greatest marginal benefits from investing in mating
effort in the form of mate attraction and rival fighting. However, the costs associated with pursuing
new mates (physical harm, pathogen contraction, etc.) and leaving potential offspring unsupported
might hinder males that exclusively pursue this strategy from achieving optimal reproductive fitness
(Kokko & Jennions, 2008). Similarly, males investing solely in parenting effort may miss mating
opportunities that can result in surviving offspring with little investment on their part. Therefore, as
with many life-history trade-offs, strategies richly embody contingencies: certain aspects of the
internal and external environment favor investment in one kind of strategy, but a shift in these
conditions may lead to an adaptive shift in the balance between mating and parenting effort. These
shifts are neither conscious nor instantaneous; instead, it is thought that they result from
physiological and neuromodulatory effects that unfold over time as regulated by T.

5
1.2 The Challenge Hypothesis
A theoretical model encompassing trade-offs between mating and parenting effort within
males, the ability of males to switch between varying reproductive strategies, and the role of T
during these switch points was first developed in avian seasonal breeders. Wingfield, Hegner, Dufty,
and Ball (1990) found that baseline T levels increased at the beginning of the breeding season, which
appeared to facilitate mate acquisition and territory formation. During confrontations with other
males, T levels surged from the new baseline to the physiological maximum; these surges predicted
increased aggressive behaviors, which aided in defending mates and territory. At the end of the
breeding season, birds’ T decreased as they maintained their pair bonds and provisioned offspring.
In short, these males were shifting between mating-dominant and parenting-dominant strategies,
with T mediating the behavioral changes that reflected these strategies. Wingfield and colleagues
dubbed this framework on T and male reproductive strategies “The Challenge Hypothesis” (CH): in
their formulation, males increase their mating effort in response to mating opportunities and
challenges from other males, and T surges—at multiple timescales—permit this reallocation of
effort. Although the CH was originally formulated to explain within-male behavioral shifts in avian
species, it has since spawned a large body of supporting evidence conducted at multiple levels of
analysis, including between-male comparisons and examinations across multiple animal taxa.
Experimental research in birds comparing T-treated males to controls has corroborated the
hormone’s role in controlling shifts between reproductive strategies. Male sparrows injected with T
competed more with other males and fed their young far less frequently than controls, whereas birds
treated with flutamide, an androgen receptor antagonist, showed the opposite pattern (Hegner &
Wingfield, 1987). Dark-eyed junco males treated with T were more attractive to females, but they
strayed further from the nest after their offspring hatched (Ketterson & Nolan, 1999). Male house
finches with experimentally increased T fed offspring less frequently but sang, an index of mating

6
effort, more frequently than controls (Stoehr & Hill, 2000). Similar effects have been found in
Lapland longspurs, jays, and western screech owls (Hunt, Hahn, & Wingfield, 1999; Vleck & Brown,
1999; Herting & Belthoff, 1997). Importantly, however, predicted associations between T and
mating/parenting effort have not been universally found in birds. Some systematic analyses suggest
that the link between T and parenting behavior in birds may be restricted to certain passerine species
(Hirschenhauser et al., 2003). A recent review of pair-bonding in the zebra finch presents mixed
evidence for associations between T and pair-bonding, suggesting that effects may be context-
specific (Prior & Soma, 2015).
Though parental care is relatively rare in reptiles and fishes, evidence supporting the CH has
also been found in a number of these species. In teleost fishes, rises in T and 11-ketotestosterone
(11-KT) have been associated with the display of dominant behaviors and increases in territoriality
during mating season, but these hormones decrease outside of the mating season or when a male is
providing paternal care, albeit not universally across species (Cardwell et al., 1996; Oliveira, Almada,
& Canario, 1996, Cardwell & Liley, 1991; Francis & Fernald, 1993; Mayer et al., 1993; Kindler et al.,
1990; Sikkel, 1993; see Oliveira, Hirschenhauser, Carneiro, & Canario, 2002 and Hirschenhauser &
Oliveira, 2006 for reviews). A similar pattern has been found in amphibian species (Townsend &
Moger, 1987; Orchinik, Licht & Crews, 1988). In reptiles, T is linked to social rank, dominance,
male-male competition, and aggression (Greenberg & Crews, 1990; Schuett, Harlow, Rose, Van
Kirk, & Murdoch, 1996; Thompson & Moore, 1992). One intriguing recent paper reports that
estradiol-17β, in addition to androgens, increases in response to competition in male cichlids (Scaia
et al., 2018), suggesting a possible role for estrogens in the CH.
As our analysis concerns the role of T in human males, perhaps the most relevant
comparative evidence hails from non-human primates. Unlike birds, many primate species do not
have specific breeding seasons; of those that do, increases in baseline T as predicted by the CH are

7
observed (Dixson, 1987). Yet it appears that non-seasonal breeders might still shift between mating
and parenting effort via the effects of T. Lemurs, mandrills, and chimpanzees exhibit increased T
and aggressive behavior when in the presence of a parous female, and male tamarins exhibit
increased T and arousal behaviors when presented with the scent of an ovulating female, suggesting
that males switch to mating effort when mating opportunities are salient (Cavigelli & Pereira, 2000;
Setchell, et al., 2008; Muller & Wrangham, 2004; Sobolewski, Brown, & Mitani, 2013; Ziegler,
Schultz-Darken, Scott, Snowdon, & Ferris, 2005). Moreover, T levels increase in male tamarins
coinciding with their partners’ ovulation, which may function to increase reproductive success
(Ziegler, Jacoris, & Snowdon, 2004). In group-established male howling monkeys, T levels and
aggression increase with the threat of an outside male (Cristóbal-Azkarate, Chavira, Boeck,
Rodríguez-Luna, & Veál, 2006). As in birds, fathering behaviors in primates are correlated with a
drop in T. Male marmosets and siamangs that carried their offspring and participated in more
paternal care had lower T, suggesting that males switch to parenting effort in these situations
(Nunes, Fite, Patera, & French, 2001; Morino, 2015). Marmoset fathers exposed to the scent of their
infant experienced a drop in T (Prudom, Broz, Schultz-Darken, Ferris, Snowdon, & Ziegler, 2008).
In sum, there is some evidence to suggest that the CH may also apply to non-human primates.
1.3 The Challenge Hypothesis and Humans
Researchers have recently turned to examine the strength of evidence in favor of the CH in
humans (e.g., Archer, 2006; Wingfield, 2017). Several original predictions of the CH concern links
between T and aggression, and in humans, these predictions have been both reviewed (e.g. Carré &
Archer, 2017; Wingfield, 2017) and meta-analyzed (Archer, 2006). However, a number of other
predictions stemming from the CH more broadly concern the role of T in the balance between
mating and parenting effort in men (e.g. Burnham et al., 2003; Gettler et al., 2011); these predictions

8
have not been subjected to a formal meta-analysis. Below, we outline these predictions as
adaptations of the CH for human mating systems.
Human breeding systems are characterized by a wide diversity of mating systems (e.g.,
monogamy, polygyny, polygynandry) as well as configurations of offspring care (e.g., maternal,
paternal, communal). Human sexual activity is not confined to a specific season or particular
window of female sexual receptivity (Thornhill & Gangestad, 2008); consequently, CH-inspired
predictions for humans will differ in some respects from those in other species. However, because
T’s coordination of physiological and psychological effort toward mate acquisition is thought to be
conserved across animal taxa (Roney & Gettler, 2015), some predictions made for humans will
closely resemble those advanced for other species.
1) Men’s baseline T levels will not differ between seasons. They will, however,
covary positively with men’s mating effort. Thus, single men, who are presumably
actively searching for mates, will have higher T than pair-bonded men, who are
less likely to be actively seeking for mates (e.g. Burnham et al., 2003; but see
prediction 2).
2) Within paired men, individuals who report greater commitment or investment in
their current relationship will have lower T concentrations than men who report
less commitment or greater interests in extra-pair sexual opportunities (both
reflections of increased mating effort; e.g. McIntyre et al., 2006).
3) Within single men, individuals who report a greater number of sexual partners,
and those with less restricted sociosexuality (Simpson & Gangestad, 1991)—both
indicative of greater investment in acquiring new mates—will have higher T
concentrations (e.g., Puts et al., 2015).

9
4) Fathers, who are presumed to invest at least some degree of effort in parental
care, will have lower T levels than non-fathers (e.g. Gettler et al. 2011).
5) Fathers with a greater degree of involvement in parenting their offspring will have
lower T levels than fathers with minimal investment in parenting (e.g. Weisman et
al., 2014).
1.4 The Current Analysis
Dozens of empirical studies have investigated the above predictions, and narrative reviews
have, in general, concluded that these predictions are supported by scientific evidence (see, e.g.,
Ellison & Gray, 2009). Nevertheless, a meta-analysis of this literature is timely for several reasons.
First, some scholars argue that only some of these predictions are supported in humans; for
example, Mazur (2017) argues that marriage/pair-bonding, but not fatherhood, should predict a
decrease in T. Our analyses will be able to adjudicate disagreements such as these through a
quantitative analysis of the literature as a whole. Further, the precise effect size of CH-derived
comparisons is unknown. Statistical significance need not imply practical significance, but through a
meta-analysis, we gain the ability to provide an accurate estimate of T differences between groups of
men, which can be compared to other factors known to relate to changes in T, such as aging, certain
medical conditions, or exogenous administration. Lastly, at least two characteristics of the CH
literature in humans present challenges to theoretical interpretation that can be fruitfully addressed
in a meta-analysis. First, there exists no single standard method to analyze whether ‘relationships’ or
‘fatherhood’ predict decreased T—for instance, studies may include or exclude covariates, measure
T from samples taken at various times of day, and may adopt different operational definitions of
‘pair-bonded’. This analytic flexibility has recently been scrutinized in the psychological literature as a
major obstacle to determining the ‘true’ support for an effect (Simmons, Nelson, & Simonsohn,
2011). Second, CH effects, like the vast majority of empirical findings in the social sciences, are

10
disproportionately drawn from WEIRD (Western, Educated, Rich, Industrialized, and Democratic;
Henrich, Heine, & Norenzayan, 2010) populations. In the current review, 73% of the effects in our
dataset come from Western samples. Thus, in our analyses we tested how subjective analytic
decisions and the disproportionate representation of certain populations might affect effect size
estimates, and in so doing, we also provide recommendations for future studies.
2. Methods
2.1 Search Strategy
We located studies through multiple channels, including reference sections of published
articles, online database search engines, and email and personal correspondence with researchers in
this area. Our familiarity with the literature on human male behavioral endocrinology, as well as the
list of studies cited in Gray and Campbell (2009), acted as a starting point for our search. We next
performed searches on Google Scholar and Web of Science using “relationship status testosterone”,
“romantic relationships testosterone”, “human parental testosterone”, and “endocrinology of social
relationships” as search phrases. Lastly, with the goal of locating unpublished data and manuscripts
not identified through these methods, we emailed colleagues known to have conducted human-
subjects research on the behavioral correlates of T (whether or not this research was specifically
framed in terms of the Challenge Hypothesis) and requested data that could be included in the meta-
analysis. We discontinued our literature search in October 2017.
We first restricted our search to studies that assessed relationships between two narrowly-
defined predictor variables (pair-bond status and fatherhood status) and T concentrations (whether
assessed through blood or saliva samples). Studies that assessed T indirectly, such as via assessments
of masculinity, voice pitch, or fluctuating asymmetry were not included. However, a substantial
number of effects pertinent to our predictions assessed continuous characteristics rather than the
binary variables of fatherhood or pair-bond status. In these studies, T level was usually a predictor of

11
behavioral outcomes such as time spent with children, relationship satisfaction, or interest in extra-
pair copulations. We thus included these effects in two additional categories, grouped as pair-bond
behaviors and fathering behaviors. Henceforth, we refer to analyses on pair-bond status and
fatherhood status as our “primary analyses”, and those on pair-bonding behavior and fathering
behaviors as our “secondary analyses”. We also chose to limit our analysis to heterosexual men.
Though previous research suggests that women, but not homosexual men, experience reductions in
T during pair-bonding (van Anders & Watson, 2006; van Anders & Goldey, 2010), a dearth of
studies concerning these populations limit the utility of meta-analyses. Our initial search identified
127 relevant effects from 49 published manuscripts and 31 unpublished effects.
2.2 Inclusionary Criteria
The 127 total effects were reduced to a working data set that balanced the desire to include
as many effects as possible while limiting the dataset to only include effects that would facilitate a
meaningful examination of the CH. Thus, we had a number of criteria that determined which effects
would be included in the analyses:
1. Men’s T concentrations decline across the lifespan (e.g., Kelsey et al., 2014); this presents
a potentially important confound because fathers and men in committed relationships
may be older on average than single and/or childless men. In some cases, effects were
presented in papers both as raw T differences and, separately, adjusting for age (whether
via including it as a covariate in the statistical model, analyzing a cohort of men across
time, or matching paired men and/or fathers with age-matched controls). Whenever
possible, we selected the results controlling for age, as it likely represented the more
accurate estimate of the effect of interest. However, we also included effects that did not
control for age when they were the only estimates available in the manuscript. When
requesting unpublished effects, we asked all authors to share data or unpublished

12
comparisons including age as a covariate. In the results, we compare the strength of age-
controlled compared to non-age-controlled effects.
2. In many cases, we included multiple effects from the same paper. We elected to do this
when effects represented distinct pieces of information despite their non-
independence—this is distinct from the criterion described above, because non-age-
controlled samples provided no additional value when age-controlled comparisons were
available. Most commonly, we included multiple effects from papers when authors
reported one set of results for morning samples, and one for afternoon samples (e.g.,
Berg & Wynne-Edwards, 2001; Gray et al., 2006; Muller et al., 2009). Other manuscripts
reported multiple operationalizations of an effect of interest (e.g., both relationship
‘commitment’ and ‘satisfaction’; Hooper et al. [2011]). We control for the non-
independence of these effects with multilevel analyses (see below).
This reduced set of effects consisted of 114 total effects: 60 for relationship status, 28 for
fatherhood status, 16 for relationship behaviors, and 10 for fathering behaviors. All effects are
described and categorized in a spreadsheet contained in our Supplemental Online Materials (SOM),
and at https://osf.io/4r3a5/.
2.3 Obtaining Effect Sizes and Coding Moderators
Studies reported effects as t-statistics, F-statistics, or Pearson r correlations; unpublished
effects were provided to us as t-statistics, Pearson r correlations, or raw data from which we
calculated t-statistics. In some cases, effects of interest were not reported in the manuscript but were
presented in figures or graphs. For six effects, means and standard deviations/standard errors were
extracted from published figures using an online application
(http://arohatgi.info/WebPlotDigitizer/), which were then used to calculate test statistics. All

13
effects were converted into Fisher’s z-transformation for meta-analytic estimates and transformed
back into Pearson correlations for reporting results.
Pair-bond status. Pair-bond status effects (k = 60, 38 published) were included if they
compared the T levels of two or more groups of men as grouped by pair-bonding status, though the
operational definition of “pair-bonded” differed between studies. In our meta-analytic dataset, 15
considered only married men as pair-bonded, two only considered unmarried men in committed
relationships, and 12 contained a mix of married and unmarried men in the pair-bonded group. For
the remaining 31 effects, the distinction was unclear. However, due to the large global variation in
mating systems, we left it to the original researchers to determine what constituted ‘paired’ vs.
‘unpaired’ and included the study as long as a distinction was made. Twelve effects included fathers
in the comparison and nine did not; it is unclear in the remaining 39 effects. Of the 38 published
effects, 35 collected T via saliva and three via blood (serum or plasma). Forty-four of the 60 effects
(73%) of effects came from Western samples. See SOM for coding.
Pair-bond behavior. Pair-bond behavior effects (k = 16, 12 published) were diverse and
included T’s associations with relationship satisfaction, relationship commitment, relationship
length, interest in novel partners, number of sexual partners, and sociosexual orientation. Effects
were coded such that higher values represented greater mating effort (e.g., less restricted sociosexual
orientation, lower relationship commitment, higher number of sexual partners). Nine of these
effects examined paired men, three examined single men, and three provided insufficient
information on relationship status. Of the effects, several (k = 9) assessed relationship satisfaction
or commitment via an existing measure on relationship quality such as the Investment Model Scale
(IMS, Rusbult, Martz, & Agnew, 1998) or the relationship satisfaction scale (Hendrick, 1988). Others
considered individuals’ interest in extra-pair partners (k = 1), number of sexual partners (k = 2),
mating success (a composite score combining previous sexual experiences such as age of the first

14
intercourse and number of partners; k = 1) or sociosexual orientation (Simpson & Gangestad, 1991;
k = 3). Fourteen of 16 effects (88%) came from Western samples.
Fatherhood status. Fatherhood status effects (k = 28; published = 22) included 27 between-
subjects effects and one within-subjects effect that assessed men’s T levels before and after the birth
of their first child. Of the between-subjects effects, studies differed regarding the pair-bond status of
participants. If paired men do indeed have lower T than single men, and fathers are more likely to be
paired than non-fathers, then this variation might confound any fatherhood status effect. Again,
given the large global variation in relationship and paternity norms, we included the effect if it
separated fathers from non-fathers, regardless of pair-bond status. However, to assess the impact of
a “pair-bond confound”, we coded the extent to which comparisons of fathers to non-fathers also
compared single to paired men. Effects fell into one of four categories: only men (fathers and non-
fathers) with the same relationship status were compared (no confound; k = 6); some but not all
men were compared that had different relationship statuses (partial confound; k = 4); all fathers
were paired, and all non-fathers were unpaired (full confound; k = 4); there was insufficient
information regarding pair-bond status (unknown; k = 14). Of the published effects, 18 collected T
via saliva and four via blood. Seventeen of 28 effects (61%) were drawn from Western samples.
Fathering behavior. Fathering behavior effects included ten published effects. Fathering
behaviors were assessed in variety of ways, including partner reports of involvement (k = 2), a self-
report composite of ‘male parenting effort’ (see Gray et al., 2002; k = 1), experience as a parent (k =
1), reaction to infant cries (k =1), affectionate touch (k = 1), gaze towards infants (k =1), use of
‘motherese’ (high-pitched, rhythmic speech directed toward infants; k = 1), and time spent with the
offspring (k = 1). One effect assessed “caregiving behaviors,” but the authors did not further
operationalize this variable. Eight of these 10 effects came from Western samples.
2.4 Data Analysis Plan

15
All analyses were conducted on Fisher’s z-transformed correlation coefficients. F and t-
statistics were converted to r using formulas in Borenstein, Hedges, Higgins, & Rothstein (2011);
Kendall’s tau values were converted to r using the formula provided by Walker (2003). For the
binary domains of relationship status and fatherhood, r represents the point-biserial correlation
between pair-bond/fatherhood status and T concentrations; for the continuous domains, r
represents the linear association between indices of either ‘pair-bonding behavior’ or ‘fathering
behavior’ and T concentrations. We conducted four sets of analyses, one for each of the domains
identified above: pair-bond status, pair-bond behaviors, fatherhood status, and fathering behaviors.
For each set, we conducted a series of analyses to establish a plausible range of effect sizes, in
recognition of the different strengths and weaknesses that individual techniques possess (Simonsohn
et al., 2014b; McShane et al., 2016) and the lack of consensus regarding how best to correct for bias
in meta-analyses (Carter & McCullough, 2018). The techniques we used for our analyses, and the
accompanying justifications, are detailed below.
2.4.1 Traditional Meta-Analyses
Our traditional meta-analyses were conducted using multilevel modeling, which specified
random effects at two levels: effects nested within studies, and studies within the overall set of
effects. This approach permitted us to include multiple, non-independent effects from the same
study when estimating mean effect size within domains (see Raudenbush & Bryk, 2002); it also
conceives of the true mean r within a given domain as varying over the population of studies.
Finally, this approach also allows for the examination of the effect of study-level moderators in
meta-regressions. Effects were weighted by the inverse of their variance, providing more weight to
more precisely estimated effects in the dataset (Borenstein et al., 2011). All multilevel analyses were
conducted using the ‘metafor’ package (Viechtbauer, 2010) in R version 3.3.1. Because these analyses
utilized multiple effects per study, included both significant and non-significant results, and also

16
included unpublished effects, they have the advantage over other techniques of providing an
estimate based on the largest overall sample size.
Potential sources of bias have been identified in traditional meta-analytic techniques.
Publication bias, the increased chance for statistically significant findings to be published compared
to non-significant effects, is one of the oldest and most well-known sources (Sterling, 1959). Because
of strong incentives to report statistically significant effects, publication bias is likely a ubiquitous
feature of scientific literature (Simonsohn, 2012), which, if left unaddressed, can lead to substantial
overestimates of an underlying effect. Two types of corrections for publication bias are commonly
pursued. First, meta-analysts attempt to reduce the influence of publication bias by seeking out the
entirety of published and unpublished studies to include in analyses. We attempted to do this for our
analyses (see Search Strategy). However, because meta-analysts are unlikely to identify all
unpublished studies to overcome publication bias completely, meta-analyses may still overestimate
effect size. Thus, a second type of correction concerns methods that statistically adjust estimates
(e.g., trim-and-fill; Duval & Tweedie, 2000; PET-PESSE; Stanley & Duocuoliagos, 2014). While
gaining popularity in the meta-analytic literature, these types of analyses may actually lead to
estimates of effect size more biased than those derived from uncorrected analyses, especially in
datasets that fail to conform to idealized assumptions (e.g., homogeneous effects, little to no
publication bias)—and violations of these assumptions are likely very common in real-world datasets
(Ledgerwood, 2016; for simulations and criticism of trim-and-fill, see Terrin et al. [2003] and
Simonsohn et al. [2014b]; for simulations and criticism of PET-PEESE, see Carter et al. [2017] and
Simonsohn [2017]). For this reason, we chose not to calculate estimates using these statistical
corrections for publication bias. At the same time, we acknowledge that we were likely not able to
recover every unpublished effect. Thus, estimates from these analyses may represent slight
overestimates, or at least ‘upper bounds’, of a plausible effect size for a given domain. To help

17
establish a realistic range of effect sizes, we opted to also perform several alternative meta-analytic
analyses that have recently been argued to provide adjusted effect size estimates with minimal bias.
2.4.2 P-curve and Alternative Selection Models: Alternative Forms of Meta-Analysis
Importantly, some features of the source data itself may lead true effect sizes to be
overestimated, and even the inclusion of unpublished effects to ‘open the file drawer’ may not be
sufficient. P-hacking—a term referring to the assortment of subjective, defensible decisions in data
collection and analysis that researchers can exploit to artificially inflate the likelihood of obtaining
statistically significant effects—distorts the literature as a whole by “replacing” null effects with
larger, statistically significant effects (Simonsohn et al., 2014a). Hence, in the presence of p-hacking,
meta-analyses that account for publication bias per se may still detect a “true” effect size greater than
zero when in fact none exists (e.g., Harris, Pashler, & Mickes, 2014).
P-curve is a procedure developed to detect p-hacking (Simonsohn et al. 2014a, 2014b). The
p-curve (Simonsohn et al., 2014a) is a distribution of p-values that are published, statistically
significant (i.e., ranging from ~0 to .05) and in the predicted direction for a given research domain.
The shape of the p-curve provides diagnostic information regarding the evidential value in a set of
studies (versus the influence of p-hacking and publication bias). Given even modest statistical power
(30%), a p-curve examining true effects will be markedly right-skewed, with 43% of p-values under
.01 (e.g., Hung, O’Neill, Bauer, & Köhne, 1997).
Moreover, p-curve analyses can do more than simply detect whether reported effects are real
or spurious; they can also generate estimates of effect size. As such, p-curve constitutes an
alternative or supplement to traditional meta-analysis. P-curve furthermore has several desirable
features relative to traditional meta-analysis. One need not comprehensively sample all reported
effects, or search for unpublished findings; only significant published values are needed. In absence
of p-hacking, p-curve returns unbiased estimates of mean true effect size, unaffected by publication

18
bias. This remains true when the effects included in p-curve analyses are heterogeneous (Simonsohn
et al. 2014a, 2014b; Gervais 2015)
1
. When p-hacking has affected results, p-curve underestimates true
effect size (Simonsohn et al. 2014a, b). As this bias is the opposite direction of the bias in traditional
meta-analysis, P-curve analysis may be a valuable complement as a more ‘lower-bound’ estimate.
All p-curves were run using the latest app on p-curve.com, p-curve 4.05. This app yields Z-
tests for right skew, left skew, and a curve flatter than one with 33% power. Negative Z-values
indicate an effect in the expected direction. We estimated effect sizes with the p-curve method using
the R code reported by Simonsohn et al. (2014b), which yields an estimate by minimizing the
Kolmogorov-Smirnov fit statistic to observed p-values (for details, see Simonsohn et al. 2014b).
Although Simonsohn and collogues expressed these values as Cohen’s d, we convert them to
Pearson’s r to match the estimates from the traditional meta-analysis. Lastly, we also estimated mean
effect size using the p-uniform method (van Assen, van Aert, & Wicherts, 2015). This procedure
employs a model identical to p-curve—and thus generates nearly identical estimates of effect size—
but uses an alternative estimator that has the added benefit of yielding a 95% confidence interval,
which p-curve does not readily provide (for details, see Simonsohn et al., 2014b; van Assen et al.,
2015). All such estimates were generated using the ‘puniform’ R package (van Aert, 2016).
McShane et al. (2016) recently offered a critical discussion of using p-curve as a meta-analytic
tool, emphasizing an important point: p-curve and p-uniform assume that the overall set of effects in
an analysis—published and unpublished, significant and non-significant alike—is homogeneous,
which may not be realistic. Indeed, statistically significant effects (for which p-curve does estimate an
accurate effect size in the face of heterogeneity) are not a random subset of all effects. Most often,
the chance that a finding is deemed significant scales positively with the size of the true effect
1
We note that the p-curve authors and other groups disagree on how well p-curve handles heterogeneity. Disagreements
may stem from differing interpretations on whether an effect size estimate should reflect the average of all possible
conducted studies, or the average of studies included in the p-curve. See http://datacolada.org/67.

19
investigated. Especially when mean true effect size is small, and heterogeneity of true effects in the
population is large, the bias resulting from assuming homogeneity in a set of effects can be
substantial (McShane et al., 2016).
Alternative selection models have been developed that estimate not only a true mean effect
but also the heterogeneity of effects (e.g., the standard deviation of effect size in the population, τ)
based on the variability of observed effect sizes relative to that expected from sampling variability
alone. Simulations show that these selection models can estimate mean true effect size without bias
across a variety of conditions (Hedges & Vevea, 1996; see also McShane et al., 2016). As McShane et
al. (2016) acknowledge, estimated effect sizes generated by heterogeneous selection models on
significant effects only are inherently very unstable. Fortunately, instability lessens when published
non-significant effects are included in the analysis, and thus the most desirable models (1) allow for
the inclusion of published non-significant effects while (2) simultaneously accounting for effect size
heterogeneity (Hedges & Vevea, 2005).
Under publication bias, non-significant effects are presumed to have less chance to appear in
the literature; however, selection models can actually estimate or assume varying chances of non-
significant effects appearing in the literature, relative to significant effects. The effects of publication
bias on estimation can be examined through sensitivity analyses. For instance, one can compare
estimates of mean effects assuming that non-significant effects have 20%, 30%, or 40% the chance
of being published relative to significant effects. We estimated effect sizes using alternative selection
models that included non-significant findings from published studies. We estimated findings
assuming non-significant effects had 20% or 40% the chance of being published compared to
significant findings in the predicted direction. The former level is in a range McShane et al. (2016)
claim is likely representative of most literatures. The latter level could apply to literatures in which
authors can argue that failures to replicate high profile findings can meaningfully contribute to the

20
literature. In these analyses, we also permitted true effect sizes to vary across studies. Hence, we
estimated, within subsets of study populations, the standard deviation of effect sizes τ. These
analyses were performed on the published effects only, as these models account for publication bias
only; including unpublished effects may lead to downwardly biased estimates. All analyses were
performed using the R code provided in the supplementary material from McShane et al. (2016).
2.4.3 Summary of Analyses Performed
For each domain of effects, we performed the same series of analyses. First, we performed a
simple multilevel analysis with random effects at the effect and study level. Second, we assessed
whether publication status moderated mean effect size estimates; specifically, we performed a meta-
regression to estimate whether, consistent with an influence of publication bias, published effects
were significantly larger than unpublished effects in a given domain. We also assessed the impact of
variation in conceptualization and measurement between studies (operationalization of ‘pair-bond
status’, the confounding of pair-bond status with fatherhood status, inclusion of an age covariate,
Western versus non-Western samples) via moderator analyses. Third, we performed p-curve and p-
uniform analyses on a given set of effects. For these analyses, we reduced a set of effects to only
include independent, published, and statistically significant effects in the published direction, as
specified by Simonsohn et al. (2014a, 2014b). In instances where there was more than one
statistically significant, non-independent effect in a domain, we selected the median effect size (see
SOM for the full set of effects used in our analyses). Finally, we estimated average effect size using a
pair of alternative selection models, as outlined above and in McShane et al. (2016). These models
permit effect size heterogeneity and the chance publication of non-significant results. The first
selection model assumes non-significant results have 20% the chance of publication as significant
results, whereas the second model assumes this relative probability to be 40%. Complete R code for
the analyses described above is available at https://osf.io/4r3a5/.

21
3. Results
3.1 Primary Analyses: Pair-bond Status
Multilevel Meta-analysis. The first set of analyses tested the hypothesis that men downregulate
T within the context of a romantic relationship, and thus pair-bonded men should have lower T
concentrations than single men. The overall analysis included 38 published effects from 25 studies
(N = 9,536 data points), and 22 unpublished effects (N = 1,502). A multilevel meta-analysis
including study as a random factor (k = 60) yielded a mean effect size estimate of r = 0.149, 95% CI:
.115:.183. See Figure 1. On average, single men exhibited higher T concentrations than men in
committed relationships.
Moderators. One common argument is that, due to publication bias, unpublished effects are
likely to be weaker than published ones. Unpublished effects in the pair-bond status domain were
estimated to be smaller (unpublished r = .097; published r = .177), and this difference was significant
(p = .039). As testosterone levels decline with age, and men in relationships may be systematically
older than single men, we also tested whether age-controlled effects differed from effects without an
age control. We found no evidence that these two categories of effects differed (age controlled: r =
.156; age not controlled: r = .144; p = .759). The difference in effect sizes between Western and non-
Western samples fell just short of significance (Western r = .130, non-Western r = .207; p = .056).
Finally, we tested whether differing operationalizations of ‘relationship status’ led to
different estimates of its effect on T. When comparing across categories of how studies classified
whether men were ‘paired’ (married only, paired but unmarried, mix of married and unmarried,
undefined), effect sizes ranged from r = .091 -.164; the overall difference between categories failed
to reach significance (p = .787); pairwise comparisons similarly revealed that none of the contrasts
between individual categories approached significance.

22
p-curve and p-uniform. Only independent, statistically significant published effects were used in
the p-curve and p-uniform analyses (k = 18), per the recommendations in Simonsohn et al. (2014a,
2014b). The p-curve for these effects demonstrated significant right skew, consistent with results
containing evidential value (i.e., not being due entirely to p-hacking and publication bias—full p-
curve: Z = -2.61, p = 0.005; half p-curve: Z = -1.99, p = 0.023). P-curve’s estimate of the true mean
effect was r = 0.083. The average effect size estimate from p-uniform was, as expected, nearly
identical: r = 0.078. The confidence interval for these estimates did not reject the null hypothesis of
no effect (95% CI: -.042:.133).
Alternative Selection Models. Finally, we present estimates derived from alternative selection
models. In the overall set of published effects, when heterogeneity was permitted to be non-zero,
and non-significant results had 20% the chance of being published relative to significant results, the
estimated effect size was r = .124. Increasing this probability to 40% raised the estimate to r = .154.
3.2 Secondary Analyses: Pair-bond Behaviors
Multilevel analysis. This analysis included 12 effects from 8 studies (N = 1,388 data points),
and 4 unpublished effects (N = 186). An analysis of this body of effects (k = 16) yielded an average
effect size estimate of r = .215, 95% CI: .138:.291. Unpublished and published effects were
approximately equal in size (effect of publication status: r = .-005, p = .962). Overall, men with
higher T concentrations exhibited more behaviors indicative of an interest in finding and acquiring
new mates (i.e., “mating effort”).
Other moderators. Pair-bond behaviors measured varied across studies. We coded whether the
effect we extracted from a given study pertained to a) paired men’s attitudes or behaviors in their
current relationship (k = 10), b) single men’s sexual behavior or attitudes (k = 3); or c) a mix of the
two (k = 3). Effect sizes ranged from r = .118 -.315. The largest effect sizes pertained to current
relationship attitudes/behaviors (r = .315), whereas single men’s sexual attitudes/behaviors was

23
smaller (r = .195); however, this difference was non-significant (p = .131). Effects from non-
Western samples were larger than those from Western samples (r = .363 and r = .203, respectively),
but with only two non-Western effects, this difference was non-significant (p = .317).
p-curve and p-uniform. The estimated effect size with p-curve (k = 5) was r = .123. Estimates
from p-uniform revealed that the small number of significant published effects led to an imprecise
estimate with a confidence interval overlapping zero: r = .131, 95% CI: -.146:.406. Unlike the
traditional estimates in this domain, these estimates do not find evidence for a robust effect.
However, the p-curves indicated evidential value: full p-curve: Z = -2.14, p = 0.016; half p-curve: Z =
-1.46, p = 0.072.
Alternative Selection Models. A heterogeneous selection model, in which non-significant results
have 20% the chance of being published relative to significant results, estimated the effect size to be
r = .141. Increasing this probability to 40% raised the estimate to r = .169.
3.3 Primary Analyses: Fatherhood Status
Multilevel Meta-analysis. The first set of analyses in the fatherhood domain tested whether
fathers, who are hypothesized to upregulate parenting effort and downregulate mating effort, have
lower T than non-fathers. The overall analysis included 22 published effects from 16 studies (N =
5,223 data points), and 6 unpublished effects (N = 1,091). A multilevel meta-analysis of effects
nested within studies (k = 28) yielded a mean effect size estimate of r = 0.189, 95% CI: .111:.267,
suggesting that fathers overall have lower concentrations of T than non-fathers. See Figure 2.
Moderators. In this domain, unpublished effects were smaller than published effects
(published r = .233; unpublished r = .077), consistent with a set of studies affected by publication
bias. This difference fell short of significance (p = .067).
As with relationship status, age might introduce a confound into comparisons based on
fatherhood status, as fathers tend to be older than non-fathers. However, we found no evidence that

24
age-controlled effects differed from non age-controlled effects (p = .458). Lastly, we examined how
the presence of a “pair-bond confound” moderated effect size estimates (i.e., comparing fathers to
non-fathers may also be comparing paired to single men; see Methods). Estimates varied appreciably
across categories, ranging from r = .155 (when there was insufficient information regarding pair-
bond status) to r = .248 (when there was no confound with pair-bonding), to r = .313 (when
fatherhood was fully confounded with pair-bonding). The existence of strongest effects in the ‘full
confound’ category—in which paired fathers were compared with unpaired non-fathers—is
consistent with expectations that bias inflates the estimates of these effects. However, all possible
pairwise comparisons of confound categories failed to reach statistical significance (all p > .05).
p-curve and p-uniform. The estimated effect size with p-curve using only significant,
independent, published effects in the predicted direction (k = 12) is r = .186. The right skew for the
half p-curve was highly significant, Z = -3.30, p < 0.001, indicating ‘evidential value’ of the body of
studies in this domain. The estimate from p-uniform was very similar and indicated the presence of a
robust effect: r = .185, 95% CI: .037:.339.
Alternative Selection Models. A heterogeneous selection model, in which non-significant results
had 20% the chance of being published relative to significant results, estimated the effect size to be r
= .145. Increasing this probability to 40% raised the estimate to r = .190.
3.4 Secondary Analyses: Fathering Behaviors
Multilevel analysis. This analysis included 11 published effects from six studies (N = 504 data
points). A multilevel analysis yielded an average effect size estimate of r = .334, 95% CI: .244:.424.
p-curve and p-uniform. The estimated effect size with p-Curve (k = 5) was r = .173. Estimates
from p-uniform differed noticeably, once again likely due to the small number of significant
published effects: r = .266, 95% CI: -.364:..573. The p-curves did not indicate evidential value: full p-
curve: Z = -0.87, p = 0.191; half p-curve: Z = 0.13, p = 0.551.

25
Heterogeneous Selection Models. A heterogeneous selection model, in which non-significant
results had 20% the chance of being published relative to significant results, estimated the effect size
to be r = .258. Increasing this probability to 40% raised the estimate to r = .308.
3.5 Summary of estimates
For a summary of effect size estimates by domain, see Figure 3.
4. Discussion
4.1 Summary of Findings
In our meta-analysis, we evaluated evidence for four different predictions derived from the
Challenge Hypothesis: 1) pair-bonded men will have lower T levels than single men; 2) men more
committed and/or invested in their current pair-bonded relationship will have lower T levels than
those less involved; 3) fathers will have lower T levels than non-fathers; and 4) fathers more
involved in parenting activities will have lower T levels than fathers who are less involved. To
establish a plausible range of effect sizes, we used a variety of meta-analytic techniques (traditional
estimates, p-curve/p-uniform, and alternative selection models) and found that each of our
predictions was supported, albeit to varying degrees. In our primary analyses, the aggregate of
evidence suggests that the effect of pair-bond status and fatherhood status are both robust and non-
zero. The effect of pair-bond status (r = .08 - .15) was smaller than the effect of fatherhood (r = .15
- .19). Within the pair-bond status domain, published effects were appreciably larger than
unpublished effects, and non-Western effects were marginally larger than Western effects. Within
the fatherhood domain, published effects, and those confounded with pair-bond status were
appreciably larger than unpublished and non-confounded effects, respectively; however, both
differences fell short of statistical significance.
We also conducted secondary analyses on sets of effects involving correlations of T with
specific pair-bond or fathering behaviors thought to indicate men’s balance between mating and

26
parenting effort. Here, we found larger effects than in our primary analyses: r = .12 - .22 for pair-
bond behaviors and r = .13 - .33 for fathering behaviors. However, given the smaller number of
effects used in these analyses, the uncertainty in these estimates was also higher. Once again, the
aggregate of evidence suggested that these effects were significantly different from zero, though
analyses were less unanimous on this point for fathering behaviors.
4.2 Contextualization of Effect Sizes
In interpreting effect sizes, researchers often turn to Cohen’s (1988; 1992) rules of thumb: r
coefficients of .1, .3, and .5 are translated as small, medium, and large, respectively. By these
conventions, the effect sizes we estimate generally fall between small and medium. Yet, this begs a
fundamental question (as estimations of effect size often do): How should one interpret the practical
significance of a “small” or “medium” effect? To address this point, it may be helpful to turn to
factors—investigated in other literatures—that are known to affect T concentrations. As Cohen
himself noted, “a basis for positing [effect size] which comes from theory or experience should
automatically take precedence” over arbitrary benchmarks (Cohen, 1988; p. 147). With this in mind,
we contextualize our effect size estimates by comparing them to known differences in T as a
function of aging, ethnic or geographic population, and exogenous administration.
Following a peak shortly after puberty, men’s T declines throughout the life course;
however, the steepness of the decline, and whether the decline stops at a certain age, remain a matter
of debate (see Kelsey et al. [2014] and Handelsman et al. [2015] for two treatments of this issue).
Furthermore, as the results of our meta-analysis show, circumstantial and/or social factors have an
appreciable influence on men’s T. Thus, estimates of the size of men’s age-related declines differ
substantially. Kelsey et al. (2014) aggregated data from 13 samples of men (total N = 10,098) to
generate a predictive model for men’s T across the lifespan. We used the data from Kelsey et al. to
determine median T, along with standard deviations, at various adult ages. As illustrative examples,

27
this model predicts an r = .17 decrease when comparing men age 20 and age 30, but only an r = .06
decrease between age 30 to 40. Comparing our meta-analytic estimates to the normative data from
Kelsey et al., the average T difference between a single and a paired man that we report (r ≈ .10)
approximates the average T difference between a 20 year old man and a 25 year old man. The T
difference between fathers and non-fathers (r ≈ .18) is approximately equal to the difference
between a 20 and 32 year old man.
As T secretion appears to increase along a gradient of socioeconomic status, urbanization,
and affluence (Alvarado, 2010; Gray et al., 2006), between-population differences in men’s T
provide another means of contextualizing our estimates. Comparisons of populations of men both
within and between societies tend to show moderate differences that get smaller with increasing
ages. Taking the within-society comparison of black and white American men as an example, one
study with a mean age of 31 found a difference of r = .12 (black men had higher average T; Ettinger
et al., 1997), but a second study with a mean age of 38 only reported a difference of r = .03 (Ellis &
Nyborg, 1992). Our pair-bond status estimate falls near the higher end of this ethnic difference, but
our other estimates all exceed the magnitude of this difference. Moving to between-society
comparisons, Ellison et al. (2002) showed that some geographic populations of men vary
considerably in their average T concentrations. Here, our pair-bond status effect closely resembles
the average difference between American and Nepalese men (higher in American men; r = .10). Our
fatherhood status effect is nearly equal to the difference between Congolese and Nepalese men
(higher in Congolese men; r = .19).
Finally, studies of exogenous T administration provide yet another benchmark. Such effects
generally greatly exceed the estimates we provide that are based on social and/or life history factors.
For instance, 25 mg of a weekly T injection, considered to be a very small dose, has an effect on
endogenous T equal to r = .39 (Bhasin et al., 2001). Larger doses of exogenous T often show effects

28
even larger (e.g., r = .91 for the effect of increasing a weekly injection by 100 mg; see Bhasin et al.,
2001).
In sum, most natural between-population comparisons of men’s T—along with our
estimates that closely resemble these population differences—would be classified conventionally as
“small” to “medium” effect sizes. Yet, given the extensive interest in the scholarly literature on the
consequences of T differences of this size (see, e.g., Kelsey et al., 2014; Alvarado, 2010), we believe
our effect size estimates suggest an appreciable practical (not just statistical) significance.
4.3 Strengths and Limitations
In this meta-analysis, we have provided the first comprehensive quantitative review to date
of several predictions derived from the Challenge Hypothesis in humans. Previous reviews have
either concerned separate predictions of the Challenge Hypothesis (e.g., the link between T and
aggression; Archer, 2006), or been narrative in nature and have relied on published effects only (e.g.,
Gray & Campbell, 2009). Drawing from the development of new theory and techniques from meta-
analytic science, we have provided a plausible range of effect size estimates for predictions that vary
in the methodologies used to test them, and in the power available within the literature to detect true
effects. For our two main predictions, concerning pair-bond and fatherhood status, we aggregated a
large enough set of effects to estimate a tight range of effect sizes, providing strong evidence of real
differences in these domains. For our two secondary analyses, although the overall pattern of results
was consistent with our predictions, the meta-analytic estimates for the pair-bonding behavior and
fathering behavior domain did not contain sufficient statistical power to draw equally firm
conclusions regarding the robustness of these factors in predicting men’s T.
We sent out a widespread call for unpublished effects and obtained dozens of these effects
for our analyses, which we see as a major strength of our meta-analysis. However, several caveats
apply to this point. First, the majority of the unpublished effects (22 out of 30) were in the pair-

29
bond status domain, with only a handful of unpublished effects for fatherhood status or pair-
bonding behaviors (and none for fathering behaviors). Thus, the benefit of reduced bias may be
largely concentrated in the pair-bond status estimates. Second, we have no doubt that we were
unable to recover every unpublished effect. At the very least, we know of several datasets containing
the appropriate variables that we were never able to access through authors. Third, while analyses
comparing published to unpublished effects showed no statistically significant differences, the
overall pattern—with unpublished effects being smaller—goes in the direction one would expect
with publication bias. Perhaps a sample with greater power—i.e., more unpublished effects—would
reveal significant differences.
We have attempted to address these limitations transparently. First, we fully acknowledge the
challenge of generating estimates in a heterogeneous and biased literature. We believe publication
bias has affected the literature on human behavioral endocrinology and the Challenge Hypothesis, as
it has virtually all other fields in psychology (Simonsohn, 2012). For this reason, we made a
concerted decision not to rely on any single procedure in generating our estimates. Meta-analytic
techniques may treat the non-independence of effect optimally (by using multilevel models), or may
treat publication bias optimally (by modeling selection processes). To our knowledge, there is no
technique that does both. Traditional meta-analytic estimates are able to incorporate multiple non-
independent effects in a multilevel analysis, and possess greater precision due to incorporating a
larger number of effects; however, the estimates may still suffer from some degree of upward bias. A
number of techniques have been proposed to account for the bias introduced by heterogeneity,
publication bias, and p-hacking, and we used several of them to help fill out a plausible range of
effect size estimates. Discussion is ongoing regarding the strengths and weaknesses of these
techniques (e.g., McShane et al., 2016, Carter et al., 2017; Nelson, 2018), particularly those that are
newly developed, but through our reading of these discussions, we attempted to select techniques

30
supported by simulations and quantitative demonstrations. More generally, given widespread
disagreement regarding how to best correct for bias in meta-analyses (Carter & McCullough, 2018),
we feel that an approach focused on sensitivity analyses that relies on multiple techniques is superior
to selecting a particular technique and cherry-picking evidence in its favor.
Importantly, the preponderance of evidence from our results suggests that bias is not the sole
source of positive effects in any of the domains we examine. In our mind, this point is also worth
emphasizing: based on our analyses, we believe effects based on the Challenge Hypothesis in men
are both real and affected by selective reporting.
4.4 Suggestions for Future Research
Our meta-analysis has provided what we hope will serve as a useful reference for the average
effect of social relationships on men’s T in several domains. However, numerous avenues still exist
to more precisely tease out the nature of relationships between life-history shifts and men’s T (see
e.g., Zilioli and Bird, 2017). Thanks to the work of biological anthropologists, our analyses contained
a non-trivial number of effects from non-Western populations (e.g., Alvergne, Faurie, & Raymond,
2009; Gettler et al., 2011; Muller et al., 2009; Nansunga et al., 2014); however, much more work
remains to be done to determine the impacts of cross-cultural socioecological variation on hormone-
mediated life-history shifts.
Second, future research in humans may benefit from an increased effort to tease apart the
proximate, biological connections between T and parenting effort. Our analysis, despite finding an
overall negative association between T and the expression of parenting behaviors in men, is unable
to speak to specific mechanisms at play that mediate the theorized trade-off. Non-human animal
research, in its effort to explain complex and heterogeneous findings regarding the link between
paternal behavior and T (Bales & Saltzman, 2016; Hirschenhauser et al., 2003; Wynne-Edwards &
Timonin, 2007), has turned to identifying specific neural mechanisms that regulate paternal behavior.

31
For example, studies from rodents suggest that the neuromodulatory effects of T on reward circuits,
which include brain regions such as the medial preoptic area of the hypothalamus (MPOA) and bed
nucleus of the stria terminalis (BST), are central to the expression of paternal behavior (reviewed in
Bales & Saltzman, 2016). Similar research in humans is just beginning. A preliminary imaging study
of ten human fathers identified numerous prefrontal and subcortical brain regions that associate
with T responses to infants (Kuo et al., 2012). Mascaro, Hackett, and Rilling (2014), in a study of 88
fathers, found that neural responses in a brain region associated with face emotion processing (the
caudal middle frontal gyrus [MFG]) were correlated with concurrent T concentrations. Another
study of 70 fathers found that ventral tegmental area (VTA) activity, while associated with viewing
pictures of their infants, was not correlated with T (Mascaro et al., 2013). Obviously, much more
remains to be done to elucidate the role of T in the paternal brain (Swain et al., 2014; Feldman,
2015), and we strongly encourage future research in this vein.
Relatedly, the role of other neuroendocrine mechanisms in men’s life-histories also remains
ripe for exploration and review. Oxytocin (OT), in particular, is another hormone intimately
involved in the processes of sexual pair-bonding and parenting (Gangestad & Grebe, 2017).
Although recent meta-analyses have failed to establish robust effects of OT on more general
prosocial behavior (e.g., trust; Nave, Camerer, & McCullough, 2015), researchers have found robust
support for OT’s role in social processes specific to romantic relationships and parent-child bonds
(Feldman, 2017; Gangestad & Grebe, 2017). Theoretical frameworks have been developed to jointly
account for the effects of OT and T in social relationships (e.g. the ‘steroid/peptide theory of social
bonds’ from van Anders, Goldey, & Kuo, [2011]; the ‘neuroscience of social decision-making’ from
Rilling & Sanfrey [2011]). These frameworks expand upon the CH and are necessary to advance our
understanding of the behavioral endocrinology of humans’ social relationships. An emblematic
example in this regard concerns paternal protection of infants, which is conventionally considered

32
an example of parenting effort, yet it has been associated with increased T (van Anders et al., 2011).
It may be that interactions with concomitant OT surges contribute to positive correlations between
T and certain aspects of parenting effort. Similar interactive effects may also drive associations
between hormones and men’s mating effort (Mascaro et al., 2014).
Prolactin, too, has been associated with paternal care across diverse animal taxa (Brown,
Murdoch, Murphy, & Moger 1995; Gubernick & Nelson, 1989; Lynn, 2016; Reburn & Wynne-
Edwards, 1999; Saltzman & Ziegler; 2014; Schradin & Anzenberger, 1999). Pioneering correlational
and experimental work in birds has established that prolactin directly contributes to fathering
behaviors such as incubation and offspring provisioning (reviewed in Lynn, 2016). Though
correlational studies from monogamous rodents and New World monkeys are consistent with
findings in birds, experimental work interestingly has failed to find evidence for a direct activational
effect of prolactin on paternal behavior (see Saltzman & Ziegler, 2014). Finally, early evidence
supports a potential role for prolactin in human fatherhood as well, though stronger conclusions
await more definitive evidence (see Gangestad & Grebe, 2017). We agree with van Anders et al.
(2011) that an increased emphasis on studies of prolactin in humans may reveal important roles in
conjunction with T and OT.
For future research in the realms we mention above, we provide two suggestions that may
aid in yielding more robust effects.
First, our analyses focus almost exclusively on cross-sectional, between-subjects comparisons
of T. This was a purposeful choice, motivated by a desire to obtain effects broadly represented
across the literature. At the same time, we acknowledge that a reliance on between-subjects
comparisons as a test of the CH entails paying a cost in construct validity; indeed, theory specifically
pertains to within-individual changes in T in response to life circumstances. Although we attempted to
control for a number of confounds that could introduce error into between-group comparisons,

33
extraneous factors may still have affected the observed ‘baseline’ concentrations of T. Thus, a more
desirable method for conducting future research entails the measurement of within-person
hormonal changes in response to evolutionarily-salient life events. The few longitudinal studies in
humans bolster the conclusion of a T-mediated shift toward parenting effort when men become
fathers (e.g., Storey et al., 2000; Gettler et al., 2011). Future work could complement these findings
by shedding light on the role played by individual differences in the degree to which fathers shift
from mating to parenting effort. Short-term changes in OT, too, have been shown to predict
dynamics of romantic relationships (Grebe et al., 2017) and parental care (e.g., Feldman et al., 2010)
in men. A promising avenue for future longitudinal research would be to clarify how these changes
function as a component of men’s life-history strategies.
Second, the published studies we review are generally greatly underpowered to detect the
average effect sizes that we estimate. For instance, the median sample size in our pair-bond status
domain is 73; this provides just a 17% chance of detecting a true effect of r = .12. Substituting our
various effect size estimates generates a range of power from 10% - 26%. The power to detect our
estimated fatherhood status effect (r = .18) with the median sample size in our dataset (n = 71) is
higher but still far from ideal, at 32% (ranging from 23% to 38%, again depending on effect size
estimate used)
2
. Statistical power at these levels presents problems of both an inflated false negative
rate, and an inflated estimate of effect size (“the winner’s curse”; see Ioannidis, 2008). We
recommend that researchers look to power analysis, rather than previously published papers, when
designing future studies on T and the Challenge Hypothesis. As an example, the sample size
necessary to achieve 80% power (a figure often used as a target when designing new studies) to
detect effects in the range of our fatherhood status estimate is anywhere from 200 – 360
participants, greater than the vast majority of studies in our dataset. We note that this
2
All estimates obtained from G*Power (version 3.1.9.2).

34
recommendation also applies to research in behavioral endocrinology more generally, a field that
often suffers from a reliance on small, underpowered samples to address hypotheses of interest (see
e.g. Walum, Waldman, & Young, 2016).
4.5 Conclusion
Across male members of many species, T has been shown to mediate trade-offs between
mating and parenting effort. The same mediating role has been argued to exist in humans,
manifesting in T levels of single men being generally higher than T levels of pair-bonded men and
fathers, relationship investment and fathering behavior negatively associated with circulating T, and
mate-seeking behaviors positively associated with circulating T. Although these relationships have
been reported in narrative reviews, to date no meta-analytic evidence exists that speaks to the precise
effect size, statistical significance, and practical relevance of these relationships. In the current meta-
analysis, we aggregated findings from 114 effects from 66 published and unpublished studies (for a
total of 19,397 data points from 17,341 individuals) and found that being single was associated with
greater T levels, non-fathers had higher T levels than fathers, mate-seeking behavior was associated
with higher T levels, and fathering behavior was associated with lower T levels. Our effect sizes,
which would be interpreted as somewhere between “small” and “medium” based on convention,
can be fruitfully interpreted in light of the effects of other factors known to relate to T, such as
aging, population differences, and T administration. With these overall associations in mind, we
encourage future research, based on large samples and within-subjects comparisons whenever
possible, that links specific neuroendocrine mechanisms and behaviors to the dynamic nature of
men’s life-histories.

35
References
Alvarado, L. C. (2010). Population differences in the testosterone levels of young men are associated
with prostate cancer disparities in older men. American Journal of Human Biology, 22(4), 449-
455.
Alvergne, A., Faurie, C., & Raymond, M. (2009). Variation in testosterone levels and male
reproductive effort: Insight from a polygynous human population. Hormones and Behavior,
56(5), 491-497.
Archer, J. (2006). Testosterone and human aggression: an evaluation of the challenge hypothesis.
Neuroscience & Biobehavioral Reviews, 30(3), 319-345.
Bales, K. L., & Saltzman, W. (2016). Fathering in rodents: neurobiological substrates and
consequences for offspring. Hormones and Behavior, 77, 249-259.
Berg, S. J., & Wynne-Edwards, K. E. (2001, June). Changes in testosterone, cortisol, and estradiol
levels in men becoming fathers. In Mayo Clinic Proceedings (Vol. 76, No. 6, pp. 582-592).
Elsevier.
Bhasin, S., Woodhouse, L., Casaburi, R., Singh, A. B., Bhasin, D., Berman, N., & Dzekov, J. (2001).
Testosterone dose-response relationships in healthy young men. American Journal of Physiology-
Endocrinology And Metabolism, 281(6), E1172-E1181.
Borenstein, M., Hedges, L. V., Higgins, J. P., & Rothstein, H. R. (2011). Introduction to meta-
analysis. John Wiley & Sons.
Brown, R. E., Murdoch, T., Murphy, P. R., & Moger, W. H. (1995). Hormonal responses of male
gerbils to stimuli from their mate and pups. Hormones and Behavior, 29(4), 474-491.

36
Burnham, T. C., Chapman, J. F., Gray, P. B., McIntyre, M. H., Lipson, S. F., & Ellison, P. T. (2003).
Men in committed, romantic relationships have lower testosterone. Hormones and Behavior,
44(2), 119-122.
Cardwell, J. R. & Liley, N. R. (1991). Hormonal control of sex and colour change in stoplight
parrotfish, Sparisoma viride, Gen. comp. Endocrin, 81, 7-20.
Cardwell, J. R., Sorensen, P. W., Van Der Kraak, G. J., & Liley, N. R. (1996). Effect of dominance
status on sex hormone levels in laboratory and wild-spawning male trout. General and
Comparative Endocrinology, 101(3), 333-341.
Carter, E. C., & McCullough, M. E. (2018). A simple, principled approach to combining evidence
from meta-analysis and high-quality replications. Advances in Methods and Practices in
Psychological Science.
Carter, E., Schönbrodt, F., Gervais, W. M., & Hilgard, J. (2017). Correcting for bias in psychology: A
comparison of meta-analytic methods. Preprint online at:
https://osf.io/preprints/psyarxiv/9h3nu
Cavigelli, S. A., & Pereira, M. E. (2000). Mating season aggression and fecal testosterone levels in
male ring-tailed lemurs (Lemur catta). Hormones and Behavior, 37(3), 246-255.
Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd edition).
Cohen, J. (1992). A power primer. Psychological Bulletin, 112(1), 155.
Cristóbal-Azkarate, J., Chavira, R., Boeck, L., Rodríguez-Luna, E., & Veàl, J. J. (2006). Testosterone
levels of free-ranging resident mantled howler monkey males in relation to the number and
density of solitary males: a test of the challenge hypothesis. Hormones and Behavior, 49(2), 261-
267.

37
Del Giudice, M., Gangestad, S. W., & Kaplan, H. S. (2015). Life history theory and evolutionary
psychology. In D.M. Buss (Ed.), The Handbook of Evolutionary Psychology. Hoboken, NJ: John
Wiley & Sons.
Dixson, A. F., & Lunn, S. F. (1987). Post-partum changes in hormones and sexual behaviour in
captive groups of marmosets (Callithrix jacchus). Physiology & Behavior, 41(6), 577-583.
Duval, S., & Tweedie, R. (2000). Trim and fill: a simple funnel‐plot–based method of testing and
adjusting for publication bias in meta‐analysis. Biometrics, 56(2), 455-463.
Edelstein, R. S., Chopik, W. J., & Kean, E. L. (2011). Sociosexuality moderates the association
between testosterone and relationship status in men and women. Hormones and Behavior,
60(3), 248-255.
Ellis, L., & Nyborg, H. (1992). Racial/ethnic variations in male testosterone levels: a probable
contributor to group differences in health. Steroids, 57(2), 72-75.
Ellison, P. T., & Gray, P. B. (Eds.). (2009). Endocrinology of social relationships. Harvard
University Press.
Ettinger, B., Sidney, S., Cummings, S. R., Libanati, C., Bikle, D. D., Tekawa, I. S., ... & Steiger, P.
(1997). Racial differences in bone density between young adult black and white subjects
persist after adjustment for anthropometric, lifestyle, and biochemical differences. The Journal
of Clinical Endocrinology & Metabolism, 82(2), 429-434.
Farrelly, D., Owens, R., Elliott, H. R., Walden, H. R., & Wetherell, M. A. (2015). The effects of
being in a “new relationship” on levels of testosterone in men. Evolutionary Psychology, 13(1),
147470491501300116.
Feldman, R. (2015). The adaptive human parental brain: implications for children's social
development. Trends in neurosciences, 38(6), 387-399.
Feldman, R. (2017). The neurobiology of human attachments. Trends in Cognitive Sciences, 21(2), 80-99.

38
Feldman, R., Gordon, I., Schneiderman, I., Weisman, O., & Zagoory-Sharon, O. (2010). Natural
variations in maternal and paternal care are associated with systematic changes in oxytocin
following parent–infant contact. Psychoneuroendocrinology, 35(8), 1133-1141.
Fernandez-Duque, E., Valeggia, C. R., & Mendoza, S. P. (2009). The biology of paternal care in
human and nonhuman primates. Annual Review of Anthropology, 38, 115-130.
Fine, C. (2017). Testosterone rex: Myths of sex, science, and society. New York, NY: WW Norton
& Company.
Fleming, A. S., Corter, C., Stallings, J., & Steiner, M. (2002). Testosterone and prolactin are
associated with emotional responses to infant cries in new fathers. Hormones and Behavior,
42(4), 399-413.
Francis, R.C., Fernald, R.D., 1993. Neuroendocrine effects of dominance status in an African
cichlid. XXIII International Ethological Conference (Torremolinos, Spain) Abstracts, 340.
Gangestad, S. W., & Grebe, N. M. (2017). Hormonal systems, human social bonding, and
affiliation. Hormones and Behavior, 91, 122-135.
Gervais, W. Putting PET-PEESE To The Test [Internet]. Lexington, KY: William Gervais; 2015
June 25 [cited 2016 March 30]. Available from:
http://willgervais.com/blog/2015/6/25/putting-pet-peese-to-the-test-1
Gettler, L. T., McDade, T. W., Feranil, A. B., & Kuzawa, C. W. (2011). Longitudinal evidence that
fatherhood decreases testosterone in human males. Proceedings of the National Academy of
Sciences, 108(39), 16194-16199.
Gray, P. B. (2003). Marriage, parenting, and testosterone variation among Kenyan Swahili men.
American Journal of Physical Anthropology, 122(3), 279-286.

39
Gray, P. B., & Campbell, B. C. (2009). Human male testosterone, pair bonding and fatherhood. In
P.T. Ellison & P.B. Gray (Eds.), Endocrinology of Social Relationships. Cambridge, MA: Harvard
University Press.
Gray, P. B., Chapman, J. F., Burnham, T. C., McIntyre, M. H., Lipson, S. F., & Ellison, P. T. (2004).
Human male pair bonding and testosterone. Human Nature, 15(2), 119-131.
Gray, P. B., Ellison, P. T., & Campbell, B. C. (2007). Testosterone and marriage among Ariaal men
of northern Kenya. Current Anthropology, 48(5), 750-755.
Gray, P. B., Kahlenberg, S. M., Barrett, E. S., Lipson, S. F., & Ellison, P. T. (2002). Marriage and
fatherhood are associated with lower testosterone in males. Evolution and Human Behavior,
23(3), 193-201.
Gray, P. B., Parkin, J. C., & Samms-Vaughan, M. E. (2007). Hormonal correlates of human paternal
interactions: A hospital-based investigation in urban Jamaica. Hormones and Behavior, 52(4),
499-507.
Gray, P. B., Yang, C. F. J., & Pope, H. G. (2006). Fathers have lower salivary testosterone levels than
unmarried men and married non-fathers in Beijing, China. Proceedings of the Royal Society of
London B: Biological Sciences, 273(1584), 333-339.
Grebe, N. M., Kristoffersen, A. A., Grøntvedt, T. V., Thompson, M. E., Kennair, L. E. O., &
Gangestad, S. W. (2017). Oxytocin and vulnerable romantic relationships. Hormones and
Behavior, 90, 64-74.
Greenberg, N., & Crews, D. (1990). Endocrine and behavioral responses to aggression and social
dominance in the green anole lizard, Anolis carolinensis. General and comparative Endocrinology,
77(2), 246-255.
Gubernick, D. J., & Nelson, R. J. (1989). Prolactin and paternal behavior in the biparental California
mouse, Peromyscus californicus. Hormones and Behavior, 23(2), 203-210.

40
Handelsman, D. J., Yeap, B. B., Flicker, L., Martin, S., Wittert, G. A., & Ly, L. P. (2015). Age-
specific population centiles for androgen status in men. European Journal of
Endocrinology, 173(6), 809-817.
Harris, C. R., Pashler, H., & Mickes, L. (2014). Elastic analysis procedures: an incurable (but
preventable) problem in the fertility effect literature. Comment on Gildersleeve, Haselton,
and Fales (2014). Psychological Bulletin, 140(5), 1260.
Hedges, L. V., & Vevea, J. L. (1996). Estimating effect size under publication bias: small sample
properties and robustness of a random effects selection model. Journal of Educational and
Behavioral Statistics, 21(4), 299-332.
Hedges, L.; Vevea, J. (2005). Selection method approaches. In H. Rothstein, A. Sutton, A., & M.
Borenstein, (Eds.), Publication bias in meta-analysis: Prevention, assessment and adjustments, pp. 145–
174. Chichester, UK: John Wiley & Sons.
Hegner, R. E., & Wingfield, J. C. (1987). Effects of experimental manipulation of testosterone levels
on parental investment and breeding success in male house sparrows. The Auk, 462-469.
Hendrick, S. S. (1988). A generic measure of relationship satisfaction. Journal of Marriage and the
Family, 93-98.
Henrich, J., Heine, S. J., & Norenzayan, A. (2010). The weirdest people in the world? Behavioral and
Brain Sciences, 33(2-3), 61-83.
Herting, B. L., & Belthoff, J. R. (1997). Testosterone, Aggression, and Territoriality in Male Western
Screech-owls (Otus kennicottue): Results from Preliminary Experiments. United States
Department of Agriculture Forest Service General Technical Report NC, 213-217.
Hirschenhauser, K., & Oliveira, R. F. (2006). Social modulation of androgens in male vertebrates:
meta-analyses of the challenge hypothesis. Animal Behaviour, 71(2), 265-277.

41
Hirschenhauser, K., Winkler, H., & Oliveira, R. F. (2003). Comparative analysis of male androgen
responsiveness to social environment in birds: the effects of mating system and paternal
incubation. Hormones and Behavior, 43(4), 508-519.
Hooper, A. E. C., Gangestad, S. W., Thompson, M. E., & Bryan, A. D. (2011). Testosterone and
romance: the association of testosterone with relationship commitment and satisfaction in
heterosexual men and women. American Journal of Human Biology, 23(4), 553-555.
Hung, H. J., O'Neill, R. T., Bauer, P., & Kohne, K. (1997). The behavior of the p-value when the
alternative hypothesis is true. Biometrics, 11-22.
Hunt, K. E., Hahn, T. P., & Wingfield, J. C. (1999). Endocrine influences on parental care during a
short breeding season: testosterone and male parental care in Lapland longspurs (Calcarius
lapponicus). Behavioral Ecology and Sociobiology, 45(5), 360-369.
Ioannidis, J. P. (2008). Why most discovered true associations are inflated. Epidemiology, 19(5), 640-
648.
Kaplan, H. S., & Lancaster, J. B. (2003). An evolutionary and ecological analysis of human fertility,
mating patterns, and parental investment. Offspring: Human fertility behavior in Biodemographic
Perspective, 170-223.
Kelsey, T. W., Li, L. Q., Mitchell, R. T., Whelan, A., Anderson, R. A., & Wallace, W. H. B. (2014). A
validated age-related normative model for male total testosterone shows increasing variance
but no decline after age 40 years. PloS One, 9(10), e109346.
Ketterson, E. D., & Nolan Jr, V. (1992). Hormones and life histories: an integrative approach. The
American Naturalist, 140, S33-S62.
Ketterson, E. D., & Nolan, Jr, V. (1999). Adaptation, exaptation, and constraint: a hormonal
perspective. The American Naturalist, 154(S1), S4-S25.

42
Kindler, P.M., Philipp, D.P., Gross, M.R., Bahr, J.M., 1989. Serum 11-ketotestosterone and
testosterone concentrations associated with reproduction in male bluegill (Lepomis macrochirus:
Centrarchidae). General and Comparative Endocrinology, 75, 446–453.
Kokko, H., & Jennions, M. D. (2008). Parental investment, sexual selection and sex ratios. Journal of
Evolutionary Biology, 21(4), 919-948.
Kuo, P. X., Carp, J., Light, K. C., & Grewen, K. M. (2012). Neural responses to infants linked with
behavioral interactions and testosterone in fathers. Biological psychology, 91(2), 302-306.
Ledgerwood, A. (2016). Introduction to the special section on improving research practices:
Thinking deeply across the research cycle. Perspectives on Psychological Science, 11(5), 661-663.
Ligon, J. D., Thornhill, R., Zuk, M., & Johnson, K. (1990). Male-male competition, ornamentation
and the role of testosterone in sexual selection in red jungle fowl. Animal Behaviour, 40(2),
367-373.
Lynn, S. E. (2016). Endocrine and neuroendocrine regulation of fathering behavior in
birds. Hormones and Behavior, 77, 237-248.
Maestripieri, D., M. Baran, N., Sapienza, P., & Zingales, L. (2010). Between-and within-sex variation
in hormonal responses to psychological stress in a large sample of college students. Stress,
13(5), 413-424.
Mascaro, J. S., Hackett, P. D., & Rilling, J. K. (2013). Testicular volume is inversely correlated with
nurturing-related brain activity in human fathers. Proceedings of the National Academy of Sciences,
201305579.
Mascaro, J. S., Hackett, P. D., & Rilling, J. K. (2014). Differential neural responses to child and
sexual stimuli in human fathers and non-fathers and their hormonal correlates.
Psychoneuroendocrinology, 46, 153-163.

43
Mayer, I., Rosenqvist, G., Borg, B., Ahnesjo, I., Berglund, A., Schulz, R., 1993. Plasma levels of sex
steroids in three species of pipefish (Syngnathidae). Canadian Journal of Zoology, 71, 1903–1907.
Mazur, A. (2014). Testosterone of young husbands rises with children in the home. Andrology, 2(1),
125-129.
Mazur, A. (2017). Testosterone in biosociology: A memoir. Hormones and Behavior, 92, 3-8.
Mazur, A., & Michalek, J. (1998). Marriage, divorce, and male testosterone. Social Forces, 77(1), 315-
330.
McGlothlin, J. W., Jawor, J. M., & Ketterson, E. D. (2007). Natural variation in a testosterone-
mediated trade-off between mating effort and parental effort. The American Naturalist, 170(6),
864-875.
McIntyre, M., Gangestad, S. W., Gray, P. B., Chapman, J. F., Burnham, T. C., O'rourke, M. T., &
Thornhill, R. (2006). Romantic involvement often reduces men's testosterone levels--but not
always: The moderating role of extrapair sexual interest. Journal of Personality and Social
Psychology, 91(4), 642.
McShane, B. B., Böckenholt, U., & Hansen, K. T. (2016). Adjusting for publication bias in meta-
analysis: An evaluation of selection methods and some cautionary notes. Perspectives on
Psychological Science, 11(5), 730-749.
Muller, M. N., & Wrangham, R. W. (2004). Dominance, aggression and testosterone in wild
chimpanzees: a test of the ‘challenge hypothesis’. Animal Behaviour, 67(1), 113-123.
Muller, M. N., Marlowe, F. W., Bugumba, R., & Ellison, P. T. (2009). Testosterone and paternal care
in East African foragers and pastoralists. Proceedings of the Royal Society of London B: Biological
Sciences, 276(1655), 347-354.

44
Nansunga, M., Manabe, Y. C., Alele, P. E., & Kasolo, J. (2014). Association of testosterone levels
with socio-demographic characteristics in a sample of Ugandan men. African Health Sciences,
14(2), 348-355.
Nave, G., Camerer, C., & McCullough, M. (2015). Does oxytocin increase trust in humans? A critical
review of research. Perspectives on Psychological Science, 10(6), 772-789.
Nelson, L. [71] The (Surprising?) Shape of the File Drawer. [Internet]. Leif Nelson; 2018 April 30
[cited 2018 May 1]. Available from: http://datacolada.org/71
Nunes, S., Fite, J. E., Patera, K. J., & French, J. A. (2001). Interactions among paternal behavior,
steroid hormones, and parental experience in male marmosets (Callithrix kuhlii). Hormones and
Behavior, 39(1), 70-82.
Oliveira, R. F., Almada, V. C., & Canario, A. V. (1996). Social modulation of sex steroid
concentrations in the urine of male cichlid fish Oreochromis mossambicus. Hormones and
Behavior, 30(1), 2-12.
Oliveira, R. F., Hirschenhauser, K., Carneiro, L. A., & Canario, A. V. (2002). Social modulation of
androgen levels in male teleost fish. Comparative Biochemistry and Physiology Part B: Biochemistry
and Molecular Biology, 132(1), 203-215.
Orchinik, M., Licht, P., & Crews, D. (1988). Plasma steroid concentrations change in response to
sexual behavior in Bufo marinus. Hormones and Behavior, 22, 338e350.
package version 0.0.2.
Parker, T. H., Knapp, R., & Rosenfield, J. A. (2002). Social mediation of sexually selected
ornamentation and steroid hormone levels in male junglefowl. Animal Behaviour, 64(2), 291-
298.
Perini, T., Ditzen, B., Fischbacher, S., & Ehlert, U. (2012). Testosterone and relationship quality
across the transition to fatherhood. Biological Psychology, 90(3), 186-191.

45
Pollet, T. V., Cobey, K. D., & van der Meij, L. (2013). Testosterone levels are negatively associated
with childlessness in males, but positively related to offspring count in fathers. PloS One, 8(4),
e60018.
Pollet, T. V., van der Meij, L., Cobey, K. D., & Buunk, A. P. (2011). Testosterone levels and their
associations with lifetime number of opposite sex partners and remarriage in a large sample
of American elderly men and women. Hormones and Behavior, 60(1), 72-77.
Prior, N. H., & Soma, K. K. (2015). Neuroendocrine regulation of long-term pair maintenance in
the monogamous zebra finch. Hormones and Behavior, 76, 11-22.
Prudom, S. L., Broz, C. A., Schultz-Darken, N., Ferris, C. T., Snowdon, C., & Ziegler, T. E. (2008).
Exposure to infant scent lowers serum testosterone in father common marmosets (Callithrix
jacchus). Biology Letters, 4(6), 603-605.
Puts, D. A., Pope, L. E., Hill, A. K., Cárdenas, R. A., Welling, L. L., Wheatley, J. R., & Breedlove, S.
M. (2015). Fulfilling desire: Evidence for negative feedback between men's testosterone,
sociosexual psychology, and sexual partner number. Hormones and Behavior, 70, 14-21.
Raudenbush, S. W., & Bryk, A. S. (2002). Hierarchical linear models: Applications and data analysis
methods (Vol. 1). Thousand Oaks, CA: Sage Publishing.
Reburn, C. J., & Wynne-Edwards, K. E. (1999). Hormonal changes in males of a naturally biparental
and a uniparental mammal. Hormones and Behavior, 35(2), 163-176.
Rilling, J. K., & Sanfey, A. G. (2011). The neuroscience of social decision-making. Annual Review of
Psychology, 62, 23-48.
Roney, J. R., & Gettler, L. T. (2015). The role of testosterone in human romantic relationships.
Current Opinion in Psychology, 1, 81-86.
Rose, R. M., Holaday, J. W., & Bernstein, I. S. (1971). Plasma testosterone, dominance rank and
aggressive behaviour in male rhesus monkeys. Nature, 231, 366-368.

46
Rusbult, C. E., Martz, J. M., & Agnew, C. R. (1998). The investment model scale: Measuring
commitment level, satisfaction level, quality of alternatives, and investment size. Personal
Relationships, 5(4), 357-387.
Sakaguchi, K., Oki, M., Honma, S., & Hasegawa, T. (2006). Influence of relationship status and
personality traits on salivary testosterone among Japanese men. Personality and Individual
Differences, 41(6), 1077-1087.
Sakaguchi, K., Oki, M., Honma, S., Uehara, H., & Hasegawa, T. (2007). The lower salivary
testosterone levels among unmarried and married sexually active men. Journal of Ethology,
25(3), 223-229.
Saltzman, W., & Ziegler, T. E. (2014). Functional significance of hormonal changes in mammalian
fathers. Journal of neuroendocrinology, 26(10), 685-696.
Sapolsky, R. M. (2017). Behave: The biology of humans at our best and worst. London, UK:
Penguin Books.
Scaia, M. F., Morandini, L., Noguera, C., Trudeau, V. L., Somoza, G. M., & Pandolfi, M. (2018). Can
estrogens be considered as key elements of the challenge hypothesis? The case of intrasexual
aggression in a cichlid fish. Physiology & Behavior.
Schradin, C., & Anzenberger, G. (1999). Prolactin, the hormone of paternity. Physiology, 14(6), 223-
231.
Schuett, G. W., Harlow, H. J., Rose, J. D., Van Kirk, E. A., & Murdoch, W. J. (1996). Levels of
plasma corticosterone and testosterone in male copperheads (Agkistrodon contortrix) following
staged fights. Hormones and Behavior, 30(1), 60-68.
Setchell, J. M., Smith, T., Wickings, E. J., & Knapp, L. A. (2008). Social correlates of testosterone
and ornamentation in male mandrills. Hormones and Behavior, 54(3), 365-372.

47
Sikkel, P.C., 1993. Changes in plasma androgen levels associated with changes in male reproductive
behavior in a brood cycling marine fish. General and Comparative Endocrinology. 89, 229–237.
Simmons, J. P., Nelson, L. D., & Simonsohn, U. (2011). False-positive psychology: Undisclosed
flexibility in data collection and analysis allows presenting anything as significant. Psychological
Science, 22(11), 1359-1366.
Simonsohn, U. (2012). It does not follow: Evaluating the one-off publication bias critiques by
Francis (2012a, 2012b, 2012c, 2012d, 2012e, in press). Perspectives on Psychological Science, 7(6),
597-599.
Simonsohn, U. [59] PET-PEESE Is Not Like Homeopathy [Internet]. Uri Simonsohn; 2017 April
12 [cited 2017 October 30]. Available from: http://datacolada.org/59
Simonsohn, U., Nelson, L. D., & Simmons, J. P. (2014). P-curve and effect size: Correcting for
publication bias using only significant results. Perspectives on Psychological Science, 9(6), 666-681.
Simonsohn, U., Nelson, L. D., & Simmons, J. P. (2014). P-curve: a key to the file-drawer. Journal of
Experimental Psychology: General, 143(2), 534.
Simpson, J. A., & Gangestad, S. W. (1991). Individual differences in sociosexuality: evidence for
convergent and discriminant validity. Journal of Personality and Social Psychology, 60(6), 870.
Smith, J. M. (1989). Evolutionary genetics. Oxford, UK: Oxford University Press.
Sobolewski, M. E., Brown, J. L., & Mitani, J. C. (2013). Female parity, male aggression, and the
challenge hypothesis in wild chimpanzees. Primates, 54(1), 81-88.
Stanley, T. D., & Doucouliagos, H. (2017). Neither fixed nor random: Weighted least squares meta‐
regression. Research Synthesis Methods, 8(1), 19-42.
Sterling, T. D. (1959). Publication decisions and their possible effects on inferences drawn from
tests of significance—or vice versa. Journal of the American Statistical Association, 54(285), 30-34.

48
Stoehr, A. M., & Hill, G. E. (2000). Testosterone and the allocation of reproductive effort in male
house finches (Carpodacus mexicanus). Behavioral Ecology and Sociobiology, 48(5), 407-411.
Storey, A. E., Walsh, C. J., Quinton, R. L., & Wynne-Edwards, K. E. (2000). Hormonal correlates of
paternal responsiveness in new and expectant fathers. Evolution and Human Behavior, 21(2), 79-
95.
Swain, J. E., Dayton, C. J., Kim, P., Tolman, R. M., & Volling, B. L. (2014). Progress on the paternal
brain: Theory, animal models, human brain research, and mental health implications. Infant
Mental Health Journal, 35(5), 394-408.
Terrin, N., Schmid, C. H., Lau, J., & Olkin, I. (2003). Adjusting for publication bias in the presence
of heterogeneity. Statistics in Medicine, 22(13), 2113-2126.
Thompson, C. W., & Moore, M. C. (1992). Behavioral and hormonal correlates of alternative
reproductive strategies in a polygynous lizard: tests of the relative plasticity and challenge
hypotheses. Hormones and Behavior, 26(4), 568-585.
Thornhill, R., & Gangestad, S. W. (2008). The evolutionary biology of human female sexuality.
Oxford, UK: Oxford University Press.
Townsend, D. S., & Moger, W. H. (1987). Plasma androgen levels during male parental care in a
tropical frog (Eleutherodactylus). Hormones and Behavior, 21(1), 93-99.
Trivers, R. (1972). Parental investment and sexual selection (Vol. 136, p. 179). Cambridge, MA:
Biological Laboratories, Harvard University.
van Aert, R. (2016). puniform: Meta-analysis methods correcting for publication bias. R
van Anders, S. M., & Goldey, K. L. (2010). Testosterone and partnering are linked via relationship
status for women and ‘relationship orientation’ for men. Hormones and Behavior, 58(5), 820-
826.

49
van Anders, S. M., & Watson, N. V. (2006). Relationship status and testosterone in North American
heterosexual and non-heterosexual men and women: Cross-sectional and longitudinal data.
Psychoneuroendocrinology, 31(6), 715-723.
van Anders, S. M., & Watson, N. V. (2007). Testosterone levels in women and men who are single,
in long-distance relationships, or same-city relationships. Hormones and Behavior, 51(2), 286-
291.
van Anders, S. M., Goldey, K. L., & Kuo, P. X. (2011). The steroid/peptide theory of social bonds:
integrating testosterone and peptide responses for classifying social behavioral
contexts. Psychoneuroendocrinology, 36(9), 1265-1275.
van Anders, S. M., Hamilton, L. D., & Watson, N. V. (2007). Multiple partners are associated with
higher testosterone in North American men and women. Hormones and Behavior, 51(3), 454-
459.
van Assen, M. A., van Aert, R., & Wicherts, J. M. (2015). Meta-analysis using effect size distributions
of only statistically significant studies. Psychological Methods, 20(3), 293.
Viechtbauer, W. (2010). Conducting meta-analyses in R with the metafor package. Journal of Statistical
Software, 36(3), 1-48.
Vleck, C. M., & Brown, J. L. (1999). Testosterone and social and reproductive behaviour in
Aphelocoma jays. Animal Behaviour, 58(5), 943-951.
Walker, D. A. (2003). JMASM9: converting Kendall’s tau for correlational or meta-analytic analyses.
Journal of Modern Applied Statistical Methods, 2(2), 26.
Walum, H., Waldman, I. D., & Young, L. J. (2016). Statistical and methodological considerations for
the interpretation of intranasal oxytocin studies. Biological Psychiatry, 79(3), 251-257.
Wingfield, J. C. (2017). The challenge hypothesis: where it began and relevance to humans. Hormones
and Behavior, 92, 9-12.

50
Wingfield, J. C., Hegner, R. E., Dufty Jr, A. M., & Ball, G. F. (1990). The" challenge hypothesis":
theoretical implications for patterns of testosterone secretion, mating systems, and breeding
strategies. The American Naturalist, 136(6), 829-846.
Wynne-Edwards, K. E., & Timonin, M. E. (2007). Paternal care in rodents: weakening support for
hormonal regulation of the transition to behavioral fatherhood in rodent animal models of
biparental care. Hormones and Behavior, 52(1), 114-121.
Ziegler, T. E., & Snowdon, C. T. (2000). Preparental hormone levels and parenting experience in
male cotton-top tamarins, Saguinus oedipus. Hormones and Behavior, 38(3), 159-167.
Ziegler, T. E., Jacoris, S., & Snowdon, C. T. (2004). Sexual communication between breeding male
and female cotton‐top tamarins (Saguinus oedipus), and its relationship to infant care. American
Journal of Primatology, 64(1), 57-69.
Ziegler, T. E., Schultz-Darken, N. J., Scott, J. J., Snowdon, C. T., & Ferris, C. F. (2005).
Neuroendocrine response to female ovulatory odors depends upon social condition in male
common marmosets, Callithrix jacchus. Hormones and Behavior, 47(1), 56-64.
Zilioli, S., & Bird, B. M. (2017). Functional significance of men’s testosterone reactivity to social
stimuli. Frontiers in Neuroendocrinology, 47, 1-18.
Zilioli, S., Caldbick, E., & Watson, N. V. (2014). Testosterone reactivity to facial display of emotions
in men and women. Hormones and Behavior, 65(5), 461-468.

51
Figure 1. Forest plot of pair-bond status effects. Brackets represent 95% confidence
intervals for individual effects. Width of diamond represents the 95% confidence interval for
the overall effect size estimate.

52
Figure 2. Forest plot of fatherhood status effects. Brackets represent 95% confidence
intervals for individual effects. Width of diamond represents the 95% confidence interval for
the overall effect size estimate. Effect for Storey et al. (2000) not depicted.
3
3
One effect (Storey et al., 2000) is not shown in this figure. This study was a within-subjects study of three men before
and after becoming fathers. The small sample size leads to an impossible confidence interval (r = -1.10 - 2.91) that
interferes with visualization of the remaining effects.

53
Figure 3. Dot plot summarizing the estimates of effect size estimates by domain. Brackets
represent 95% confidence intervals for each analysis type. Confidence interval omitted for
Fathering Behavior
p
-curve/
p
-uniform estimate (-.364:.573).