Why do rape victimization rates vary across studies? A meta-analysis examining moderating variables

Abstract

Research studies have identified multiple study- and sample-related factors that predict variation in the proportion of participants who report experiences of rape (non-consensual oral, anal, or vaginal penetration obtained by force, threat of force, and/or victim incapacitation). The magnitude of variation introduced by these methodological variables is often unclear, which can complicate attempts to compare findings across research studies. With the goal of identifying and quantifying sources of variation, we conducted a meta-analysis that compared rates of rape experienced by women in the United States during adolescence or adulthood. 6391 research articles were evaluated for inclusion and 84 studies (89 independent samples) met inclusion criteria. Results of a random-effects meta-analysis found that an average of 17.0 % (95 % CI [15.7 %, 18.3 %]) of participants across samples reported experiences of rape in adolescence or adulthood. The mean participant age, source of the sample, perpetration tactics included in the measure, and interaction between sample source and perpetration tactics each predicted significant variation in the proportion of victims identified. Participant recruitment method, publication year, and the earliest age included in the reference period did not predict significant variation. These findings clarify the impact of methodological variables on observed victimization rates and provide context that can inform comparisons across sexual victimization studies.

1. Introduction

Rape, defined as forced or incapacitated penetration of or by a sexual organ (Department of Justice, 2012), is a major public health issue with long-lasting consequences for victims. The National Intimate Partner and Sexual Violence Study found that 12.3 % of women had experienced forcible rape and 8.0 % had experienced incapacitated rape in their lifetimes (Black et al., 2011). Although all forms of sexual violence cause substantial harm, relative to forms of violence accomplished via other tactics (e.g., coercion) or involving non-penetrative acts only (e.g., groping), women who have experienced rape are particularly likely to report negative physical and mental health effects (Turchik & Hassija, 2014; Ullman et al., 2007).

Rape victimization rates obtained in the wider literature vary substantially, due in part to measurement and methods differences across studies (Muehlenhard et al., 2017; Rennison & Addington, 2014). Clear and specific prevalence data are needed to inform policy and practice, and quantifying the impact of methodological variables could allow for more precise cross-study comparisons. Focusing such an analysis on a single type of sexual violence, such as rape, may be especially useful in illuminating sources of variation that are attributable to methodological differences rather than variation in the underlying prevalence of different assaultive experiences. Therefore, the purpose of the current study is to identify variables that predict variation in rape prevalence rates among women in the United States and quantify the impact of those variables. This focus is not intended to prioritize rapes perpetrated against women over rapes perpetrated against people of other genders nor rapes perpetrated in the United States over those perpetrated in other countries. Rather, this focus was guided by the substantial differences in rates of rape perpetrated against people of different genders and in different countries, and the need to define a scope of inquiry narrow enough that potentially subtle sources of methodological variation could be identified.

We begin by reviewing likely sources of variation, focusing specifically on rape when possible and considering variation in sexual victimization rates more broadly when those are the comparisons available in the research literature. We then introduce meta-analysis as a method that can be used to address this research question and discuss the advantages to such an approach.

1.1. Sources of variation in rates of rape and sexual assault

Potential sources of variation include who is part of the sample, how participants are invited into the sample and asked about their experiences, and what experiences participants are asked about. Different populations experience distinct risks for and dynamics of rape, and who comprises a study’s sample will therefore inform the experiences identified. For example, closed communities, such as universities and the military, have internal sexual assault response systems as well as cultural norms that may uniquely encourage or discourage victimization. Universities have frequently been criticized for institutional factors that encourage rape and other forms of sexual assault (Mohler-Kuo et al., 2004; Stotzer & MacCartney, 2016) and multiple studies have found high rates of rape in college samples (Fisher et al., 2000; Krebs et al., 2009; Mohler-Kuo et al., 2004). However, research directly comparing students and similarly-aged non-students has indicated that college students may actually be at lower risk for rape than their non-student peers (Sinozich & Langton, 2014). Samples drawn from military settings have consistently reported elevated rates of victimization, both among veterans (Barth et al., 2016; Campbell & Raja, 2005) and new recruits (McWhorter et al., 2009). Heavy alcohol use has been noted in both military and university settings (Castro et al., 2015; Krebs et al., 2009), potentially creating uniquely hospitable environments for rape and sexual assault.

How victimization is measured in a research study reflects both how participants are recruited into a study and how they are asked about rape experiences once they have agreed to participate. A study’s recruitment process introduces potential participants to a study, and if that process is differently successful at recruiting individuals who have and have not been assaulted, the victimization rate identified in the study would be impacted by the recruitment strategy used. Rosenthal and Freyd (2018) evaluated this potential for self-selection bias by comparing rape prevalence rates between two samples, one of which was recruited with and one without knowledge that the study was about sexual assault. They found comparable rates of rape in the two samples, suggesting that transparency in recruitment language did not bias the results. However, recruitment language may not be equally relevant to all recruitment methods. Studies that solicit participation through advertisements or sign-up sheets typically rely on a description of the study to encourage participation, and the language used to describe the study may therefore be relevant. Group-based recruitment methods, in which everyone in a pre-existing group is asked to participate, frequently produce participation rates above 80 % (e.g., Campbell et al., 2008; Humphrey & White, 2000), creating an environment in which the vast majority of people participate regardless of how the study is described and thereby limiting the potential impact of recruitment language. Recruitment language may also be less relevant to studies that provide the data collection instrument to potential participants during recruitment (e.g., mailed surveys), as potential participants have full knowledge of the study’s contents when deciding whether to take part. In sum, these recruitment methods have different levels of transparency and researcher contact, and different selection and refusal procedures, as well. No study, to our knowledge, has evaluated these methods for their potential impact on obtained rates of rape, specifically, or sexual victimization, more broadly.

Once participants are recruited into a study, they must answer questions to indicate whether they have experienced rape. A large research literature has established that many women do not recognize their victimization as “rape” even if their experience reflects the legal definition. This phenomenon, known as “unacknowledged rape,” (Koss et al., 1988; Littleton et al., 2007), illuminates the importance of using what are known as “behaviorally specific questions,” which describe the acts and tactics that comprise a type of sexual victimization and ask participants whether they have had that experience, with no requirement that they recognize their experience by its legal term. For example, one of the behaviorally specific questions used to measure rape in an early version of the widely-used Sexual Experiences Survey (SES; Koss et al., 1987) asked participants whether they have had “sexual intercourse when you didn’t want to because a man¹ threatened or used some degree of physical force (twisting your arm, holding you down, etc.) to make you”. This behaviorally specific question can be contrasted with the broadly-worded screening question, “Have you ever been raped?” asked for comparison purposes in the same survey. Behaviorally specific questions have been consistently found to elicit a more complete recounting of participants victimization experiences than broadly-worded screening questions, with one quasi-experimental study finding that use of behaviorally specific questions produced a rape estimate over nine times higher than what was identified using non-behaviorally specific questions (Fisher, 2009). A meta-analysis conducted by Wilson and Miller (2016) found that roughly six out of ten women rape survivors did not acknowledge their rape in those terms, and they recommended that researchers use behaviorally specific questions to maximize inclusion of these survivors in prevalence estimates.

Finally, what experiences of victimization are documented in a study depends on the acts and tactics included in the study’s operationalization of victimization, as well as the time period over which victimization is assessed. Some measures assess victimization over relatively short periods of time, such as six months to a year, whereas others measure victimization over the course of life stages (e.g., childhood, adolescence, adulthood) or the entire lifespan. Advocates for shorter reference periods argue that they reduce measurement error (Cantor & Lynch, 2000; Daigle et al., 2016) whereas proponents of life stage measurements emphasize the long-lasting impact of sexual assault and the importance of understanding the percentage of a population who may be dealing with its effects (Bachar & Koss, 2001; Koss et al., 2007). The range of acts included in a study’s operationalization of sexual victimization may include non-contact offenses (i.e., exhibitionism, voyeurism), contact but non-penetrative offenses (i.e., groping), and/or penetrative offenses (i.e., oral, vaginal, or anal penetration). Penetrative offenses are often also distinguished by whether penetration was attempted or completed. The operationalization of perpetration tactics may vary to include non-consent, coercion, victim incapacitation, and threats or use of physical force (Muehlenhard et al., 2017).

Across these domains, it would be expected that more expansive operationalizations would result in higher reported rates of victimization. Inquiring about victimization over one’s lifetime, for example, is likely to identify higher rates than limiting the inquiry to the previous year. The comparability of reference periods reflecting subtler differences, however, is difficult to ascertain. The SES (Koss et al., 1987; Koss et al., 2007) was written to ask about assaultive experiences since age 14, but some researchers alter the measure to ask about assaultive experiences beginning at other points in adolescence (e.g., age 16, Schumm et al., 2006; age 17, Brown et al., 2005). Does this type of change result in notably different prevalence estimates? To our knowledge, that question has not been answered.

The acts and tactics included in a given study’s operationalization of sexual victimization would be expected to impact the resulting prevalence rate, as well. Krebs et al. (2017) set out to quantify those differences and found that when college women were screened for unwanted or nonconsensual sexual contact, 10.3 % reported such victimization during an academic year. When the scope was expanded to explicitly include sexual coercion (i.e., continued verbal pressure, obtaining sexual contact through lies, manipulation, or social threats), the victimization figure increased to 14.1 %. Nearly a third (32.4 %) of women participants reported victimizing experiences when the scope was defined to include unwanted or nonconsensual sexual contact, coercion, and sexual harassment.

Other studies have focused on rape victimization rates, specifically, and compared the occurrence of rape perpetrated by force or threats of force (“forcible rape”) to rape perpetrated through victim incapacitation (“incapacitated rape”). In one such study, Kilpatrick et al. (2007) compared the lifetime prevalence of forcible and incapacitated rape² in two samples of women, one a general community sample and the other a sample of women enrolled in college. Among the general household sample, they found higher rates of forcible rape (14.6 %) compared to incapacitated rape (5.0 %). In contrast, similar rates of forcible rape (6.4 %) and incapacitated rape (6.4 %) were reported by the college student sample. Carey et al. (2015) found an even higher prominence of incapacitated rape among college women when the measurement period was restricted to participants’ first year of college, with 15 % of women reporting attempted or completed incapacitated rape and 9 % reporting attempted or completed forcible rape. Multiple sample- and methods-related variables appear to impact the relative prevalence of forcible and incapacitated rape, and analyses that evaluate the prevalence associated with different combinations of these variables could further clarify these relationships.

1.2. Identifying sources of variation in observed prevalence rates

Identifying differences in study methods and samples that account for variation in sexual victimization rates has been recognized as an important topic of research (Muehlenhard et al., 2017; Rennison & Addington, 2014). Studies that manipulate some design characteristics while holding others constant can be used to understand the effects of these manipulations on observed prevalence rates. Several studies have assessed prevalence in two populations using the same methods. Sinozich and Langton (2014) compared rape victimization rates among college students and similarly-aged non-students, for example, and found a higher rate of rape in the non-student sample. Their operationalization of victimization did not explicitly include incapacitation, however, and the applicability of their conclusions to prevalence rates that include both forcible and incapacitated rape is unclear. Kilpatrick et al. (2007) compared a student and community sample across multiple operationalizations of rape, but the large mean age difference between the samples made it difficult to distinguish risk associated with being in college with risk from being college-age. Very few studies have assessed the role of changes in methods on observed prevalence rates or have simultaneously manipulated more than one methods-related variable at a time. In a notable exception, McCallum and Peterson (2017) used an experimental design to assess the impact of researcher contact, research setting, and mode of inquiry on the proportion of victims identified, but this created eight experimental conditions, significantly limiting statistical power. As a result, it was not possible to assess interactions between methodological variables in relation to prevalence rates.

Meta-analysis may be uniquely well-suited to evaluate the impact of differences in methods and samples on observed rape prevalence rates. The goal of meta-analysis is to synthesize results across a wide range of research literature such that results of any one study can be interpreted in context of the others (Borenstein et al., 2009). Due to the diversity of studies included in meta-analyses, there are likely to be a variety of sample- and methods-related factors that can be coded and quantitatively compared. Furthermore, because meta-analyses can generate a single pooled statistic of interest (i.e., “effect size”) across all included studies, they benefit from substantially more statistical power than what is obtainable within a single study (Cohn & Becker, 2003; Lipsey & Wilson, 2001).

1.3. Current study

In sum, the extant literature has identified numerous differences in studies and samples that predict variation in reported rates of sexual victimization, broadly, and in rape, specifically. Due to the increase in statistical power and diversity of studies typically included, meta-analysis may be a promising method through which the impact of study- and sample-related differences can be further explored. Such an exploration could be especially effective if it focused on a specific type of sexual victimization in order to limit the true heterogeneity that would be expected across a wider range of victimization experiences. The purpose of the current study, therefore, was to conduct a meta-analysis to examine how study- and sample-related variables predict variation in the observed prevalence of rape.

Two research questions guided this study: 1) How much variation is observed in the proportion of women in the United States who are identified as rape victims in research studies, and 2) to what extent do study- and sample-related variables predict this variation? Based on previous research and the availability of study- and sample-related details in published articles, the source of a study’s sample, perpetration tactics included in the operationalization of rape, youngest age included in the reference period, mean participant age, participant recruitment method, and year of publication were examined as potential sources of variation. No a priori hypotheses were made.

2. Method

2.1. Identification and selection of studies for inclusion

A systematic review of the published literature was conducted in ProQuest on 6/12/2016 using the PsycINFO and PsycARTICLES databases. These databases were chosen for their coverage of research in psychology as well as related fields including medicine, psychiatry, nursing, sociology, education, etc., as well as a function through which returned results could be limited to empirical studies (discussed below). The search was designed such that that the words “rape” or “sexual assault” had to appear anywhere in the article other than full text (e.g., title, abstract, keywords). These search terms were chosen in consultation with experts in the field who felt it would be exceedingly rare for an article that reported rape prevalence in adolescence or adulthood to not use the words “rape” or “sexual assault” in the article abstract, title, or keywords. It was anticipated that the vast majority of articles returned with these search terms would not be relevant to a meta-analysis on rape victimization prevalence (i.e., returned articles would include those on rape myths, rape laws, rape perpetration, etc.) but in consultation with multiple reference librarians, the first author determined that more restrictive search terms would miss relevant articles. Included articles that used the word “sexual assault” rather than “rape” were verified to meet this study’s operational definition of rape.

Results were restricted to those written in English and published after January 1, 1980, both to facilitate a manageable sample of articles to review and to correspond to the time frame during which reviewers began to use behaviorally-specific questions to assess rape prevalence (see Cook et al., 2011, and Krebs, 2014, for reviews). To obtain a manageable sample size, we also set the restrictions that the study be peer-reviewed and that it be classified by ProQuest as an empirical study. The “empirical study methodology” restriction was particularly useful in that it allowed us to reduce the number of irrelevant results our search would return while allowing us to use broad search terms (i.e., “rape” and “sexual assault”) to ensure that we were capturing as many relevant articles as possible. After deleting duplicates, the sampling frame contained 5289 articles to screen for inclusion in the meta-analysis. With the goal of expanding the sample prior to publication, a second search was conducted on 2/12/2020. The parameters of the second search were identical to the first with the exception that the date range was set to return only articles published 12/31/2015–2/12/2020. After deleting duplicates from within that sample, the second sampling frame contained 1228 articles. Once duplicates were deleted between the two samples (n = 126), 6391 unique articles were screened for inclusion.

The first step in the article screening process was identifying those that reported the statistic we would be using as our effect size. For our study, the effect size was the proportion of women participants in a sample who had experienced rape (i.e., completed oral, vaginal, or anal penetration by an object or body part obtained through force, threat of force, and/or victim incapacitation) in adolescence or adulthood. We therefore excluded studies that measured rape over other time periods (e.g., annual incidence rates, childhood victimization rates) as one would not expect them to be comparable to rape prevalence rates specific to adolescence and adulthood. We also excluded studies that restricted their measure of rape to specific situations (e.g. the number of women who had been raped specifically by an intimate partner, the number of women who had been raped specifically while at work), as one would expect those rates to be substantially lower than the number of women who had experienced rape in any circumstances during adolescence or adulthood. Additional inclusion criteria were that the data be collected in the United States, the sample be comprised of adult women who did not need a guardian’s consent to participate, and participants not have been recruited on the basis of victimization history or membership in a specifically high-risk group (e.g., involvement in sex work, incarcerated, currently receiving in-patient psychiatric care). Finally, studies were excluded if they did not use behaviorally specific questions to screen for rape, based on the robust rape acknowledgment literature that has found non-behaviorally specific questions to be associated with drastic underestimates.³ For article screening purposes, behaviorally specific questions were operationalized as those that described the acts and/or tactics associated with the study’s definition of rape rather than asking participants directly whether they had been raped. Collectively, these inclusion criteria were used to produce a final sample of articles in which the primary observable differences between studies were the variables that would be tested as moderators in the analysis.

The first set of articles (n = 5289) was screened by the first author and four undergraduate research assistants. The first author screened approximately 85 % of the articles, and the remaining 15 % were screened by either two undergraduate research assistants or one research assistants and the first author. The second set of articles (n = 1228) was screened by the by the first and third authors, with each author screening 35 % and 65 % of the articles, respectively. The inter-screener agreement rate regarding whether the article met inclusion criteria was over 95 % for both the first and second sets of articles. Any disagreements between research assistants were resolved by the first author or, when the first author was one of the disagreeing screeners, by the full research team. Each excluded article was assigned a code denoting one of the inclusion criteria that it did not meet. Of the 6517 articles reviewed, 84 articles reporting data from 89 independent samples met inclusion criteria. The article identification and screening process is documented in Fig. 1.

Fig. 1. Flow diagram of the article search and screening process.

Note. Exclusion codes were not mutually exclusive and were applied based on the first reason for exclusion identified by the coder.

2.2. Analysis

The analyses were conducted using Comprehensive Meta-Analysis software (Version 3; Borenstein et al., 2013). A random-effects model was used, as is the recommended approach when sample or study differences suggest that there is true variation from study to study (Borenstein et al., 2009). Because many of the observed proportions were expected to fall below 0.2 or above 0.8, we used a logit transformation of the raw proportion (Lipsey & Wilson, 2001) and inverse log-transformed the intercepts and regression coefficients after analysis for ease of interpretation. Reported results include the summary effect size and mean 95 % confidence interval (CI), I² (a measure of relative heterogeneity, i.e., the proportion of observed variance that represents variance in true effects rather than variance due to sampling error), Cochran’s Q (an indication of heterogeneity), as well as τ and τ² (measures of the random-effect variance component).

A conditional model that included the following moderators was also estimated: the year the article was published (integer variable, range 1987–2020), participant recruitment method (0 = sign-up, 1 = group, 2 = mailing, 3 = other/missing), the youngest age included in the reference period (integer variable, range 13–18), mean participant age⁴ (continuous variable, range 18.2–41.0), perpetration tactics (0 = force only, 1 = force and victim incapacitation), and sample source (0 = college, 1 = community, 2 = military). The sample source variable reflected where participants were recruited from, not where the rape had occurred. Studies were coded as “military sample source,” for example, if they were recruited through their association with the military (e.g., new recruit training, veterans services) with no assumptions made about the context in which any disclosed rape took place. Based on the findings reported by Kilpatrick et al. (2007), we also tested the interaction between sample source (college/community) and perpetration tactics (force only/force and incapacitation) to evaluate whether a study’s inclusion of incapacitation impacted college and community samples differently. Comprehensive Meta-Analysis software uses listwise deletion based on missing moderator data, which reduced the sample of the conditional model to 83 articles reporting data from 87 independent samples. All included studies and their moderator values are listed in Appendix A.

3. Results

Across the 89 independent samples included in unconditional model, between 4.6 % and 48.9 % of participants reported completed rape in adolescence or adulthood (i.e., individual study effect sizes ranged from 4.6 % to 48.9 %). The pooled effect size across studies was 17.0 % (95 % CI [15.7 %, 18.3 %]). The included studies had a large amount of heterogeneity (Q(88) = 1489.70, p < .001) and variance in true effects relative to sampling error (I² = 94.09 %), indicating that substantial variation existed which could potentially be accounted for by study moderators. The random-effects variance component was also substantial (τ = 0.42, τ² = 0.17). Sensitivity analysis, which removes one study at a time and assesses the change in pooled effect size, revealed only marginal changes (pooled prevalence rate varied from 16.8 % to 17.2 %), indicating that no one study had an undue effect on the overall finding.

Publication bias was explored using both Begg and Mazumdar’s Rank Correlation Test and Duval and Tweedie’s Trim and Fill method.⁵ The Rank Correlation Test value was −0.14 (p = .04), indicating a lack of symmetry across effect sizes included in the analysis. Duval and Tweedie’s Trim and Fill method found 18 fewer studies than would be expected to the right of the mean. Publication bias is usually evidenced by missing studies to the left of the mean, as smaller effect sizes are typically less likely to be published (Dickersin, 2005). In the current analysis, there was instead under-representation of small-to-medium sized studies with large effect sizes.

There was little change in the model statistics when the sample was restricted to the 87 independent samples that had complete data on all moderators and were thus included in the conditional model. The pooled effect size of this reduced sample was 16.9 % (95 % CI [15.6 %, 18.3 %]) and the model retained significant heterogeneity (Q(86) = 1487.18, p < .001), inconsistency (I² = 94.22 %,), and variation in true effects (T = 0.42, T² = 0.18).

The Q-value for the model that included moderators indicated that at least one of the moderators was related to the effect size (Q = 68.89, df = 10, p < .001). The goodness of fit statistics confirmed that unexplained variance remained when moderators were included (Q = 793.46, df = 76, p < .001) and there continued to be a large amount of variance in true effects relative to sampling error (I² = 90.42 %) and random-effects variance (T = 0.35, T² = 0.12). The R² analog for this model was 0.33, indicating that 33 % of the variance in true effects was explained by the included moderators. Participant recruitment method (Q = 3.77, p = .29), the youngest age included in the reference period (OR = 0.95, p = .12), and publication year (OR = 1.00, p = .57) did not predict significant variation, but mean participant age (OR = 1.04, p < .01), perpetration tactics (OR = 1.94, p < .001), and sample source (Q = 10.11, p < .01) each predicted significant variation in the effect size when all other variables were held constant, indicating that studies that included incapacitation as a perpetration tactic and reported on samples with a higher mean participant age tended to find a larger proportion of rape victims in their sample (see Table 1 for effect size and confidence interval by group). Examination of the sample source variable by category indicated that while there was no significant difference between college and community samples (OR = 1.57, p = .08), there was a significant difference between college and military samples (OR = 2.21, p < .01), with studies based on military samples finding a significantly higher proportion of victims. The interaction term that evaluated the impact of perpetration tactics in college compared to community samples was also significant (OR = 0.42, p < .01), indicating that the exclusion of incapacitation from a study’s operationalization of rape affects college and community samples differently. More specifically, while studies based on community samples did not find a lower victimization rate when perpetration tactics were restricted to use of force compared to when they included victim incapacitation, such a restriction did predict a significantly lower victimization rate in college student samples.

Table 1.

Effect size (i.e., proportion of women participants in the United States reporting completed rape in adolescence or adulthood) and 95 % confidence intervals (CIs) by group.

Moderator variable	# samples in each condition	Effect size in each condition	95 % CI
Sample source
College	71	0.16	[0.15, 0.17]
Community	14	0.18	[0.15, 0.22]
Military	4	0.36	[0.28, 0.45]
Mean participant age
≤Median (<19.7 yrs.)	42	0.15	[0.13, 0.17]
>Median (≥19.7 yrs.)	45	0.19	[0.17, 0.21]
Missing	2	0.21	[0.19, 0.23]
Perpetration tactics
Force only	19	0.14	[0.10, 0.17]
Force and incapacitation	70	0.18	[0.17, 0.19]
Lower-bound age cutoff
14 years	74	0.17	[0.15, 0.18]
15 years	0	–	–
16 years	3	0.25	[0.23, 0.27]
17 years	3	0.14	[0.07, 0.27]
18 years	9	0.17	[0.13, 0.22]
Participant recruitment method
Sign-up	57	0.16	[0.15, 0.18]
Group	15	0.19	[0.15, 0.23]
Mailing	11	0.20	[0.15, 0.26]
Other/missing	6	0.16	[0.14, 0.19]
Year published
≤Median (2008 or earlier)	13	0.18	[0.16, 0.20]
>Median (2009 or later)	46	0.16	[0.15, 0.18]
Community × perpetration tactics
College sample
Force only	13	0.10	[0.08, 0.12]
Force and incapacitation	58	0.18	[0.16, 0.19]
Community sample
Force only	4	0.18	[0.13, 0.25]
Force and incapacitation	10	0.18	[0.14, 0.23]

4. Discussion

Measuring the prevalence of rape requires navigating a multitude of methodological and conceptual choice points that reflect who is recruited into a study, what questions are asked, and how participants are asked those questions. There is no need for uniformity in these decisions, but untangling their impact may allow for clearer interpretations of results and comparisons across research studies. In turn, this added clarity may increase the degree to which research is able to inform policy and practice. Parsing these impacts within individual studies is challenging, however, as few studies directly manipulate their methods in order to understand the impact of these manipulations on observed prevalence rates. Study participants also experience multiple methods-related factors simultaneously, making it difficult to consider any one factor outside the context of other study characteristics. The current study responded to these challenges by testing the impact of methodological variables on rape prevalence rates using meta-analysis.

Across the 89 independent samples and 57,019 participants included in the unconditional analysis, 17 % of participants reported a history of rape in adolescence or adulthood. There was substantial variation across studies, with effect sizes ranging from 4.6 % to 48.9 %. The vast majority of studies included in this analysis did not use probability sampling methods; therefore, these results do not offer a generalizable prevalence statistic but rather a statistical “birds eye view” of the research on rape in the United States and the factors that can explain variation in that literature. Such a view identified participants’ mean age, the perpetration tactics included in a study’s operationalization of rape, and the source of a study’s sample as variables that predict variation in the proportion of victims identified. In addition to these main effects, the interaction between sample source and perpetration tactics was also significant, such that including victim incapacitation in the operationalization of rape predicted higher rates of victimization in college, but not community, samples.

This interaction aligns with Kilpatrick et al.’s (2007) finding that while the prevalence of forcible rape was higher in a community compared to a college sample, the prevalence of incapacitated rape was higher among college students. The current analysis establishes that this relationship persists outside of direct comparisons within a single study and predicts variation in the peer-reviewed research literature as a whole. Furthermore, the significance of the interaction term when mean participant age is held constant supports arguments made by Krebs et al. (2009), Mohler-Kuo et al. (2004) and others that incapacitated rape appears to be associated with college life and not just being college-age.

Mean participant age was also a significant predictor of variation, such that participants’ probabilities of disclosing rape was 4 % higher with each one-year increase in a study’s mean participant age. The age of study participants has been recognized as a central risk factor for victimization when measuring shorter-term incidence rates (Breiding, 2014). Less attention has been paid to the ways in which participants’ mean age may affect longer-term prevalence rates and these results suggest that even small differences may correspond to noticeable variation across studies. Interestingly, while increases in mean participant age predicted higher rates of victimization, the age that demarcated the start of adolescence did not. Both of these variables inform the number of years included in a reference period, and future research should further explore the effects of expanding one or both ends of the reference period range.

Neither participant recruitment method nor the year the article was published predicted significant variation in the proportion of victims identified. Recruitment method categories included advertisements or sign-up sheets, recruiting participants from pre-existing groups in which all members were asked to participate, and unsolicited mailings in which surveys were sent to individuals who had not previously agreed to participate in the research. These methods offer different amounts of information about the content of a study at the time of recruitment, and the non-significance of participant recruitment method in our analysis offers tentative support for Rosenthal and Freyd’s (2018) finding that knowing a study was about sexual assault did not affect the participation of victims and non-victims differently.

This study’s findings must be understood in the context of its limitations, the first of which is the inclusion criteria for the meta-analysis. The peer-review search restriction reduced the number of articles to be reviewed from 12,720 to a more feasible 6517 but excluding non-peer reviewed studies can bias the summary effect size if obtaining a significant effect is related to the likelihood of publication. Because prevalence rates are not conceptualized in terms of statistical significance, it is unlikely that the rape prevalence rate directly impacted the likelihood of publication, but nevertheless, the summary effect size should be understood specifically as a representation of the peer-reviewed literature. There were other inclusion criteria that reflected our study’s focus on rape prevalence among women in the United States, and while those inclusion criteria allowed us to identify reasonably comparable effect sizes, they do limit the generalizability of our results such that they do not reflect the experiences of men, gender diverse communities, individuals living in countries other than the United States, individuals who have experienced sexual assault other than rape, and certain communities disproportionately impacted by rape, including people who are homeless, involved in the sex industry, or receiving inpatient psychiatric care. There may be other variables that predict rape prevalence rate variation in these situations or communities and exploring those variables is an important area for future work. While these limitations were anticipated, our inclusion criteria may have limited our sample in unanticipated ways, as well. The relatively smaller number of military and community samples, for example, may have been related to studying rape specifically in adolescence and adulthood; it is possible that studies based on college students are particularly likely to assess this time frame, whereas community and military studies may have been more likely to use lifetime or other time-limited measures. Regardless of the reasons, the unequal breakdown of college, community, and military samples is a limitation, particularly as it relates to the identification of interactions between variables. Finally, there are limitations related to the research literature itself. The mean percentage of white participants in studies that reported this information was 77 %, with twenty-five studies based on samples that were >90 % white. This disproportionate focus on the experience of white participants highlights the need for research that reflects a fuller spectrum of survivors’ identities.

In spite of these limitations, our findings have implications for research, policy, and practice. Our analysis revealed that excluding incapacitation as a perpetration tactic affects college and community samples differently, such that only including forcible rape is likely to underestimate rape prevalence among college students more drastically than it would in a community sample. This has particularly strong implications for comparing victimization rates among students and non-students, as was carried out by Sinozich and Langton (2014). The conclusion of their research was that the rate of rape and sexual assault was higher for nonstudents than for similarly-aged students, but our findings suggest that not explicitly asking about incapacitated rape may have impacted the college sample more strongly than the non-college sample, unintentionally creating a comparison that better reflected the experiences of non-students than students. Rape victimization measures and the resulting prevalence data may be used to allocate resources, determine the effectiveness of prevention programming, or screen for trauma in the context of mental health services. In each of these instances, accurate data is crucial, and our findings emphasize the importance of explicitly including incapacitation in rape victimization measures used with college students, particularly if the findings are to be compared to rape prevalence rates among non-students.

More broadly, our findings highlight that measurement decisions may impact subgroups of survivors differently and encourage exploration of other sample × method interactions. One such interaction that has benefitted from study in recent years is that of participant gender, sexual identity, and victimization screening measure. While research has found that the SES is a valid tool for use with heterosexual, lesbian, and bisexual women (Canan et al., 2020), Anderson et al. (2018) have found that it has substantially lower validity when assessing victimization among heterosexual men. Heterosexual male victims have particularly low rates of rape acknowledgment (Artime et al., 2014; Reed et al., 2020), and it is possible that new tools may need to be developed to better capture rape in this demographic. As the sexual violence literature expands beyond its traditional focus on heterosexual women, meta-analysis may become a viable strategy for exploring these and other sample × method interactions. However, we must emphasize that the utility of meta-analysis in identifying these interactions relies on the amount of sample- and methods-related detail included in published articles. Many studies we reviewed did not report contextual variables (e.g., data collection year, recruitment language) or reported them in such a way that it was difficult to extract the level of detail needed for nuanced comparisons (e.g., reporting that a sample contained “mostly white” or “mostly heterosexual” participants). These reporting practices limited the moderators included in our analysis and particularly limited what could be learned about the experiences of marginalized communities. Increasing the amount of sample- and methods- related detail in publications will open additional avenues of exploration in future meta-analyses.

The current study responds to calls for more empirical research on how study methods impact obtained rates of sexual victimization (Krebs, 2014; Muehlenhard et al., 2017). The results presented identify multiple variables that predict variation, including the mean age of participants, the source of a study’s sample, the perpetration tactics included in the study’s operationalization of rape, and the interaction between sample source and perpetration tactics. Study-to-study variation is to be expected in the research literature, and clearly situating research findings in the context of a study’s sample and methods can help ensure that this variation does not detract from the fundamental conclusion that sexual violence is experienced at an alarming rate.

Appendix A

Authors (publication year)	N	Sample source	Perp. tactics	Recruitment method	Mean age	Age cutoff	Prev. rate	95 % CI
Abbey et al. (1996)	1160	College	FR/IR	Group	*25.9	14	23.0	[20.7, 25.5]
Amacker and Littleton (2013)	167	College	FR/IR	Signup	21.3	14	19.8	[14.4, 26.5]
Barker and Galliher (2017)	131	College	FR/IR	Signup	21.9	14	4.6	[2.1, 9.8]
Barnett et al. (1987)	262	College	FR	Signup	*19.0	14	10.3	[7.2, 14.6]
Benson et al. (2007)	350	College	FR/IR	Signup	19.3	14	13.0	[9.9, 16.9]
Botta and Pingree (1997)	623	College	FR/IR	Mailing	19.0	14	20.0	[17.0, 23.3]
Breitenbecher (1999)	406	College	FR/IR	Signup	*19.0	14	24.0	[18.9, 30.0]
Breitenbecher (2008)	224	College	FR/IR	Mailing	*20.6	14	26.0	[21.8, 30.7]
Breitenbecher and Gidycz (1998)	84	College	FR/IR	Signup	*20.9	14	22.0	[18.2, 26.3]
Breitenbecher and Scarce (1999)	377	College	FR/IR	Signup	21.4	14	27.0	[21.6, 33.2]
Breitenbecher and Scarce (2001)	224	College	FR/IR	Signup	*19.7	14	26.0	[17.7, 36.4]
Brown et al. (2005)	329	College	FR/IR	Signup	19.2	17	12.2	[9.1, 16.2]
Campbell et al. (2008)	268	Military	FR	Group	40.6	18	39.0	[33.3, 45.0]
Carey et al. (2015)	483	College	FR	Mailing	*18.1	14	6.0	[4.2, 8.5]
Cleere and Lynn (2013)	302	College	FR/IR	Signup	19.0	14	15.2	[11.6, 19.7]
Clements and Ogle (2009)	328	College	FR	Signup	19.0	14	10.1	[7.3, 13.9]
Cook and Messman-Moore (2018)	1293	College	FR/IR	Signup	18.8	14	21.0	[18.9, 23.3]
Copenhaver and Grauerholz (1991)	140	College	FR/IR	Mailing	*20.1	14	17.0	[11.6, 24.1]
Donde (2017)	1115	College	FR/IR	Signup	20.0	14	16.7	[14.6, 19.0]
Edwards et al. (2009)	1056	College	FR/IR	Signup	18.7	14	7.0	[5.6, 8.7]
Franklin (2010)	185	College	FR	Signup	*20.1	14	11.4	[7.6, 16.8]
Franklin (2013)	257	College	FR	Signup	20.7	14	8.6	[5.7, 12.7]
Georgia et al. (2018)	701	Comm.	FR	Signup	34.4	18	12.4	[10.2, 15.1]
Gidycz et al. (1993)	857	College	FR/IR	Signup	*19.7	14	13.8	[11.6, 16.3]
Gidycz et al. (2008)	540	College	FR/IR	Signup	*18.8	14	9.3	[7.1, 12.1]
Hahn et al. (2016)	141	College	FR/IR	Signup	19.5	14	18.0	[12.5, 25.2]
Haikalis et al. (2017)	673	College	FR	Signup	18.9	14	7.1	[5.4, 9.3]
Hollander (2014)	286	College	FR/IR	Signup	21.1	14	25.0	[20.3, 30.3]
Humphrey and White (2000)	1569	College	FR	Group	18.3	14	13.0	[11.4, 14.8]
Jenkins and Dambrot (1987)	323	College	FR	Group	*19.0	14	13.0	[9.8, 17.1]
Jordan et al. (2014)	750	College	FR	Mailing	*18.5	14	11.1	[9.0, 13.6]
Kalof (2000)	383	College	FR/IR	Mailing	21.0	18	22.1	[18.2, 26.5]
Kelley and Gidycz (2017)	501	College	FR/IR	Signup	18.9	14	20.8	[17.5, 24.6]
Kelley and Gidycz (2019)	420	College	FR/IR	Signup	19.3	14	11.4	[8.7, 14.8]
Koss et al. (1987)	3187	College	FR/IR	Group	21.4	14	15.4	[14.2, 16.7]
Koss et al. (1991)	2291	Comm.	FR/IR	Mailing	36.5	14	13.9	[12.5, 15.4]
Lawyer et al. (2010)	314	College	FR/IR	Signup	20.1	14	13.1	[9.8, 17.3]
Layman et al. (1996)	591	College	FR/IR	Signup	19.2	14	14.0	[11.4, 17.0]
Littleton and Breitkopf (2006)	1253	College	FR/IR	Signup	–	14	20.4	[18.3, 22.7]
Littleton et al. (2008)	72	Comm.	FR/IR	Group	27.0	14	16.2	[14.1, 18.6]
Littleton, Axsom, and Grills-Taquechel (2009)	99	College	FR/IR	Signup	21.7	14	20.2	[18.4, 22.2]
Littleton, Tabernik, et al. (2009)	1744	College	FR/IR	Signup	22.6	14	21.0	[14.4, 29.7]
Littleton et al. (2014)	1616	College	FR/IR	Signup	19.3	14	18.9	[17.1, 20.9]
Littleton and Dodd (2016) (Sample 1)	646	College	FR/IR	Signup	18.7	14	22.2	[15.1, 31.4]
Littleton and Dodd (2016) Sample 2	826	College	FR/IR	Signup	19.1	14	12.5	[6.6, 22.3]
Littleton et al. (2018)	1033	Comm.	FR/IR	Group	28.8	14	16.1	[13.5, 19.1]
Littleton et al. (2019)	109	College	FR/IR	Signup	18.8	14	22.4	[19.7, 25.4]
Marx et al. (2000)	176	College	FR	Signup	19.5	14	9.0	[5.6, 14.2]
Merrill et al. (1999)	1140	Military	FR	Group	20.6	14	34.7	[32.0, 37.5]
Messman-Moore and Long (2002)	300	Comm.	FR/IR	Signup	37.4	17	25.7	[21.1, 30.9]
Messman-Moore et al. (2010)	752	College	FR/IR	Signup	18.8	14	17.8	[15.2, 20.7]
Messman-Moore et al. (2013)	369	College	FR/IR	Signup	18.8	14	19.2	[15.5, 23.5]
Moore and Waterman (1999)	87	College	FR	Group	20.3	14	20.0	[12.9, 29.7]
Nelson and Lepore (2013)	1494	Comm.	FR	Group	23.4	16	24.6	[22.5, 26.8]
Newins et al. (2018)	964	College	FR/IR	Signup	19.5	14	18.8	[16.5, 21.4]
Orchowski and Gidycz (2012)	374	College	FR/IR	Signup	*18.5	14	8.3	[5.9, 11.6]
Osman (2016) (Sample 1)	267	College	FR/IR	Signup	19.1	14	15.0	[11.2, 19.8]
Osman (2016) (Sample 2)	358	College	FR/IR	Signup	19.2	14	28.0	[23.6, 32.9]
Pihlgren et al. (1993)	439	College	FR/IR	Signup	*18.6	14	18.0	[14.7, 21.9]
Reilly et al. (1992)	534	College	FR	Group	19.9	14	8.6	[6.5, 11.3]
Schry and White (2013)	672	College	FR/IR	Signup	19.4	14	18.8	[16.0, 21.9]
Schry and White (2016)	135	College	FR/IR	Signup	–	14	19.4	[13.6, 26.9]
Schultz et al. (2006)	124	Military	FR	Mailing	45.3	14	48.9	[40.2, 57.6]
Schumm et al. (2006)	777	Comm.	FR	Group	21.7	16	23.0	[20.2, 26.1]
Segal (2009)	138	College	FR/IR	Other	24.4	14	22.0	[15.9, 29.7]
Smith and Frieze (2003) (Sample 1)	127	College	FR/IR	–	*19.4	14	11.8	[7.2, 18.7]
Smith and Frieze (2003) (Sample 2)	89	College	FR/IR	–	–	14	23.6	[15.9, 33.5]
Stander et al. (2007)	2431	Military	FR/IR	Group	19.7	14	25.6	[23.9, 27.4]
Stepakoff (1998)	393	College	FR/IR	Signup	20.0	17	8.7	[6.3, 11.9]
Testa et al. (2003)	1014	Comm.	FR/IR	Other	23.8	14	17.2	[15.0, 19.6]
Tirabassi et al. (2017)	435	College	FR/IR	Signup	19.3	18	13.3	[10.4, 16.8]
Tromp et al. (1995) (Sample 1)	1037	Comm.	FR/IR	Mailing	36.6	14	30.0	[27.3, 32.9]
Tromp et al. (1995) (Sample 2)	2142	Comm.	FR/IR	Mailing	40.5	14	29.0	[27.1, 31.0]
Turchik and Hassija (2014)	309	College	FR/IR	Signup	18.9	16	27.5	[22.8, 32.7]
Turchik et al. (2007)	520	College	FR/IR	Signup	18.9	14	12.4	[9.8, 15.5]
Untied et al. (2013)	191	College	FR/IR	Signup	*19.2	14	8.4	[5.2, 13.3]
Vanzile-Tamsen et al. (2005)	318	Comm.	FR/IR	Mailing	24.0	14	18.2	[14.3, 22.8]
Walch and Broadhead (1992) Sample 1	147	Comm.	FR/IR	Group	35.0	18	7.5	[4.2, 13.0]
Walch and Broadhead (1992) Sample 2	258	College	FR/IR	Group	23.1	18	15.5	[11.6, 20.4]
Walker and Messman-Moore (2011)	501	College	FR/IR	Signup	18.7	14	16.8	[13.8, 20.3]
Walsh, DiLillo, et al. (2013)	714	College	FR/IR	Signup	19.7	18	22.4	[19.5, 25.6]
Walsh, Messman-Moore, et al. (2013)	546	College	FR/IR	Signup	18.7	14	22.9	[19.6, 26.6]
Walsh et al. (2016)	3001	Comm.	FR/IR	Other	33.2	14	14.8	[13.6, 16.1]
Wilson and Scarpa (2017)	1192	College	FR/IR	Signup	19.9	18	14.9	[13.0, 17.0]
Winslett and Gross (2008)	88	College	FR/IR	Signup	19.8	14	15.1	[9.0, 24.2]
Yeater et al. (2010)	334	College	FR/IR	Signup	19.5	14	17.0	[12.0, 23.5]
Yeater et al. (2016)	481	College	FR/IR	Signup	19.9	14	20.4	[16.4, 25.1]
Yeater et al. (2019)	168	College	FR/IR	Signup	18.2	14	16.0	[13.0, 19.6]
Yuan et al. (2006)	793	Comm.	FR	Other	41.0	18	14.0	[11.8, 16.6]

Note. Across moderators, – indicates missing data. For mean age moderator, * indicates that mean age was not provided in the article but a value was estimated based on included data (e.g., modal age, breakdown of years in college). For recruitment method, “other” and “missing” were collapsed for analysis in order to minimize the impact of missing data.

Data availability

Data will be made available on request.