Data were collected online between January 19 and February 14, 2024, using Verian’s opt-in panel in Spain. Gender and age quotas were applied to approximate the demographic distribution of the population of Castilla-La Mancha (see Supplementary Material for comparison with regional census data on key variables). The survey completion rate was 36.1%. Respondents received points redeemable for gifts as compensation for their time.

The final sample comprised 1,003 residents of Castilla-La Mancha (50.7% men), with ages ranging from 16 to 84 years (M = 45.5, SD = 14.3). The survey was administered in Spanish and included up to 57 items assessing public attitudes toward sexuality, including measures of RMA and ambivalent sexism, among other constructs. The average completion time was approximately 12 minutes (M = 12.2, mdn = 10.9, SD = 5.8). Ethical approval for the study was granted by the Social Research Ethics Committee at the University of Castilla-La Mancha (Protocol Number: CEIS-2024-23586).

Statistical analyses were conducted in R (version 4.2.3). Descriptive statistics were computed for all items in the “She Asked For It” subscale. In line with current recommendations in psychometric research, measurement invariance was evaluated at both subscale and item levels. Although MG-CFA provides a well-established framework for assessing invariance at the (sub)scale level, it may be insensitive to non-invariance affecting specific items (D’Urso et al., 2022). As a result, item-level non-invariance may exist even when (sub)scale-level indices suggest acceptable model fit. IRT-based approaches address this limitation by enabling the detection of DIF, that is, whether individual items operate differently across groups after controlling for the latent trait (Tay et al., 2014). For this reason, combining MG-CFA and IRT-based approaches provides a more comprehensive and sensitive assessment of measurement equivalence.

Subscale-level measurement invariance across gender was examined using MG-CFA implemented in the lavaan and semTools packages. Given the ordinal nature of the response scale (1–5), the Weighted Least Squares Mean and Variance adjusted (WLSMV) estimator was employed. WLSMV provides appropriate parameter estimates, standard errors, and scaled chi-square statistics for ordinal indicators.

Following recommendations for ordered-categorical indicators (Wu & Estabrook, 2016; Svetina et al., 2019), measurement invariance was tested sequentially across three levels: configural invariance (equal factor structure across groups with freely estimated thresholds and loadings), threshold invariance (thresholds constrained equal across groups), and metric invariance (thresholds and factor loadings constrained equal). Model fit was evaluated using the Comparative Fit Index (CFI), Tucker–Lewis Index (TLI), Root Mean Square Error of Approximation (RMSEA), and Standardized Root Mean Square Residual (SRMR). Nested model comparisons were conducted using scaled chi-square difference tests appropriate for WLSMV estimation, supplemented by changes in CFI and RMSEA.

Internal consistency was assessed using Cronbach’s alpha, ordinal alpha, and McDonald’s omega coefficients, calculated with the semTools package. Ordinal reliability coefficients were computed to account for the ordered-categorical nature of the data (Zumbo et al., 2007).

Item-level performance was assessed using the mirt package. A multi-group Graded Response Model (MG-GRM) was estimated to model the relationship between the latent trait (RMA) and the probability of endorsing each response category across gender groups. MG-GRMs yield item discrimination parameters (a), which indicate how strongly each item relates to the latent construct, and category difficulty parameters (β), which represent locations along the latent trait continuum at which respondents have a 50% probability of endorsing a given response category or higher (i.e., cumulative category probabilities).

Item-level measurement invariance was further evaluated using Differential Item Functioning (DIF) analyses within this multi-group IRT framework. For each item, a sequential series of likelihood ratio chi-square tests was conducted. First, an omnibus DIF test compared a constrained baseline model with a model in which both discrimination (a) and category difficulty (β) parameters were freely estimated across gender. Second, non-uniform DIF was evaluated by testing equality constraints on discrimination parameters. Third, uniform DIF was assessed by testing equality constraints on category difficulty parameters while constraining discrimination parameters to equality across groups. This sequential procedure allowed differentiation between non-uniform DIF attributable to differences in item discrimination and uniform DIF attributable to differences in category difficulty parameters. This approach complements MG-CFA by providing item-level evidence of measurement non-equivalence. Whereas MG-CFA evaluates invariance at the subscale level, DIF analyses identify specific items that may operate differently across groups, even when overall subscale-level invariance is supported.

To assess the cumulative impact of item-level DIF on total scores, DRF analyses were conducted using the mirt package (Chalmers, 2018). DRF examines whether expected total scores differ between groups at equivalent levels of the latent trait across the full continuum. To incorporate sampling uncertainty into the presentation of results, 95% confidence intervals were obtained using 500 bootstrap draws for both item-level and scale-level DRF estimates.

Overall, respondents exhibited low endorsement of rape myth statements. Table 1 presents the percentage distribution across response categories for each of the six items. Most respondents selected “totally disagree” or “quite disagree,” with pronounced floor effects for Items 1 (“If a woman is raped while she is drunk, she is responsible for what happened”) and 5 (“When women are raped, it is often because the way they said ‘no’ was unclear”).

Gender differences in response distributions appeared minimal for Item 1. In contrast, for the remaining items—particularly Items 2 (“When women go out wearing slutty clothes, they are asking for sexual advances from men”) and 6 (“If a woman starts making out, she should not be surprised if a man assumes she wants to have sex”)—women showed higher rates of disagreement and lower endorsement of agreement categories compared to men (see Table 1).

The single-factor model demonstrated excellent fit for women (scaled χ²(8) = 8.93, p = .348; CFI = 1.000; TLI = .999; RMSEA = .015; SRMR = .017), whereas model fit was comparatively weaker for men (scaled χ²(8) = 41.20, p < .001; CFI = 0.986; TLI = 0.974; RMSEA = .090; SRMR = .046). Although the chi-square test was significant for men, incremental fit indices and SRMR suggested acceptable though comparatively weaker model fit.

In the male subsample, a high modification index (MI = 28.59) was observed for the residual correlation between Items 1 (“If a woman is raped while she is drunk, she is responsible for what happened”) and 5 (“When women are raped, it is often because the way they said ‘no’ was unclear”) (Expected Parameter Change = -0.26; standardized EPC = -0.742). Because both items belong to the same subscale and address closely related victim-blaming content, this residual correlation—substantially stronger for men (r = .474) than women (r = .268)—likely reflects local item dependence not fully captured by the latent factor.

To account for this localized dependency while preserving comparability across groups, the model was re-specified to allow the residuals of these two items to correlate in both gender groups. Formal tests of measurement invariance were conducted following Wu and Estabrook’s (2016) recommendations for categorical indicators. Table 2 presents fit indices and model comparisons across the invariance sequence.

The configural model demonstrated acceptable fit (scaled χ²(16) = 55.14, p < .001; CFI = .991; TLI = .983; RMSEA = .070; SRMR = .031). Although the chi-square test was significant, incremental fit indices and SRMR indicated satisfactory model fit. Threshold invariance, which constrains item thresholds to equality across gender, showed no significant decrement in fit relative to the configural model (Δχ²(12) = 9.65, p = .647). Fit indices remained stable (CFI = .991; TLI = .990; RMSEA = .052; SRMR = .031), and changes in CFI (ΔCFI = .000) and RMSEA (ΔRMSEA = −.018) were well within commonly recommended criteria for supporting invariance.

Metric invariance, which constrains both thresholds and factor loadings to equality, similarly showed no significant worsening of fit relative to the threshold model (Δχ²(5) = 2.80, p = .731). Model fit did not deteriorate under metric constraints, providing support for metric invariance at the subscale level across gender. Changes in fit indices across models were minimal and remained within recommended thresholds (see Table 2). Taken together, these findings support metric invariance at the subscale level, indicating that the latent construct operates comparably across genders within the MG-CFA framework.

Standardized factor loadings from the metric invariance model ranged from .730 to .871, indicating strong associations between all items and the latent construct. Reliability estimates for the re-specified model were satisfactory across the full sample and within each gender subgroup. Cronbach’s alpha indicated acceptable internal consistency (α = .80 overall; α = .81 for women; α = .78 for men). As expected for ordinal data, ordinal alpha coefficients were higher (α_ord = .91 for women; α_ord = .87 for men), and McDonald’s omega values also indicated robust reliability (ω = .83 overall; ω = .84 for women; ω = .80 for men).

Item-level DIF was assessed within a multi-group IRT framework. Although subscale-level MG-CFA supported metric invariance, DIF analyses revealed non-invariance. Omnibus DIF tests indicated significant DIF in five of the six items (see Table 3), with only Item 1 demonstrating invariance across gender (χ²(5) = 6.262, p = .282). Tests of non-uniform DIF were non-significant for all items except Item 2, for which a statistically detectable difference in discrimination was observed (χ²(1) = 5.03, p = .025). In contrast, tests of uniform DIF were highly significant for Items 2, 3, 4, and 6 (all p < .001) and marginally significant for Item 5 (χ²(4) = 9.49, p = .050).

These results indicate that item-level non-invariance was driven primarily by differences in category difficulty rather than item discrimination. Across most items, men and women with equivalent levels of latent RMA differed systematically in their probability of endorsing higher response categories, while item discrimination was largely comparable across groups.

Figures 1a and 1b present expected item score functions and the corresponding expected score differences across the latent trait continuum. At equivalent trait levels, men generally exhibit higher expected item scores than women, particularly for Items 2 and 6. This pattern reflects a leftward shift in the expected score functions for men, indicating that lower levels of the latent trait are required for men to endorse higher response categories relative to women. For example, at θ = 0 (the mean of the latent trait), Item 6 yields an expected score of approximately 3 for men compared to approximately 2 for women, representing roughly a one-point difference on the 5-point response scale. In contrast, Item 1 shows overlapping expected score functions across groups, consistent with the absence of DIF.

Examination of item parameter estimates (see Table 4) further clarified the nature of the detected DIF. Discrimination parameters (a) were generally higher for women, indicating steeper item response functions and greater differentiation of the latent trait among female respondents. For example, the discrimination parameter for Item 2 was 2.78 for women compared to 1.97 for men, suggesting that responses among women more sharply distinguished between adjacent levels of RMA.

Category difficulty parameters (β₁–β₄) were consistently lower for men across Items 2, 3, 4, and 6 (see Table 4), reflecting the uniform DIF pattern identified in Table 3. Item 1 showed no meaningful differences in either discrimination or category difficulty parameters, supporting invariance at the item level.

To assess the cumulative impact of item-level DIF on total subscale scores, DRF analyses were conducted (Chalmers, 2018). DRF estimates whether expected total scores differ between groups at equivalent levels of the latent trait, thereby quantifying scale-level score differences attributable to item-level DIF.

Figure 2a displays expected total scores across the latent trait continuum for men and women, whereas Figure 2b presents the corresponding differences in expected total scores between groups, along with bootstrap confidence intervals. Conditional on equivalent levels of latent RMA, men obtained higher expected total scores than women across most trait levels. At the latent trait mean (θ = 0), the expected total score was 9.42 for men and 7.28 for women, reflecting a difference of 2.14 points despite equivalent latent trait levels. The maximum observed difference was 2.85 points at θ = 0.8 (slightly above the population mean). Scale-level differences were substantially smaller at very low (θ < −2) and very high (θ > +3) trait levels, where floor and ceiling effects reduce variability in expected scores. Notably, DRF effects were concentrated in the central region of the latent trait continuum (approximately θ = −1 to +1), where most respondents are located.

Although five of the six items exhibited statistically significant DIF, the cumulative impact at the scale level was modest in practical terms. Across the central region of the latent trait continuum, differences corresponded to approximately 2–3 points on the 30-point scale, representing a relatively small proportion of the total score range. Bootstrap confidence intervals (see Figure 2b) indicate that this scale-level difference is statistically distinguishable from zero across the central region of the latent continuum (approximately θ = −1 to +2), where the majority of respondents are located, while becoming increasingly uncertain at higher trait levels, where ceiling effects and reduced sample density limit precision. These findings indicate that statistically significant item-level DIF does not necessarily produce large distortions in total scores, and that the practical impact of DIF should be evaluated in terms of magnitude, precision, and distribution across the latent continuum rather than statistical significance of DIF tests alone.

A central contribution of this study is methodological. The combined application of MG-CFA, item-level DIF analysis, and scale-level DRF assessment illustrates the value of triangulating approaches to measurement invariance. Rather than providing a single binary judgment of equivalence, these methods offer complementary perspectives on different layers of cross-group comparability. Subscale-level invariance testing evaluates whether the latent construct is represented similarly across groups, whereas item-level analyses identify potential parameter-level differences in item functioning. DRF analyses, in turn, provide an assessment of the practical impact of such differences on expected scale scores (Millsap, 2011; Zumbo et al., 2007). Considering these levels jointly allows researchers to distinguish between statistically detectable item-level non-invariance and its substantive implications for scale interpretation.

This distinction is particularly relevant in the study of RMA, where previous cross-cultural validations of IRMA-based instruments have often relied primarily on CFA-based tests of invariance (e.g., Balezina & Zakharova, 2023; Nyúl & Kende, 2023). Although such analyses are essential for establishing structural comparability, they may not fully capture parameter-level differences in item functioning. Indeed, some studies have reported mixed patterns of invariance across items, even when overall scale-level comparability appeared adequate (Fakunmoju et al., 2019).

More broadly, methodological research has emphasized that statistically significant DIF does not necessarily translate into meaningful distortions in scale-level scores, underscoring the importance of evaluating both item-level functioning and its cumulative impact at the scale level (Zumbo, 2003). Integrating CFA-based invariance testing with IRT-based DIF and DRF analyses provides a more comprehensive framework for assessing cross-group comparability. Researchers conducting cross-group validation studies are encouraged to adopt multi-method strategies to ensure robust and substantively interpretable conclusions.

From a substantive perspective, the detected DIF was concentrated in items addressing beliefs about female sexuality, clothing, and sexual signaling cues—elements that are often central to victim-blaming narratives in rape myth discourse. In particular, Items 2 and 6 reflect cultural scripts linking women’s appearance or sexual behavior to perceived responsibility for sexual victimization, themes widely identified as core components of RMA (Burt, 1980; Suarez & Gadalla, 2010). At equivalent levels of the latent trait, men showed a higher probability than women of endorsing higher response categories on these items, primarily due to lower category difficulty parameters. This pattern was most pronounced in the moderate range of RMA (θ between −1 and +1), where the majority of respondents are located, suggesting that the impact of DIF is conditional on respondents’ position along the latent trait continuum rather than homogeneous across it.

The magnitude of these effects suggests that some previously reported gender differences in raw total scores may be partially influenced by measurement-related factors. Differences of approximately 2–3 points in observed scores may reflect, at least in part, item-level differential functioning in addition to substantive attitudinal variation. Although this does not call into question the robust evidence of gender disparities in RMA—meta-analytic research has documented moderate to large gender differences, with men consistently reporting higher levels of RMA than women (Suarez & Gadalla, 2010)—it underscores the importance of distinguishing between true differences at the latent construct level and distortions introduced at the measurement level.

Importantly, the scale-level impact of DIF, as assessed through DRF, was relatively small. These findings indicate that the translated subscale remains appropriate for research applications, particularly within latent variable modeling frameworks that explicitly model measurement parameters and group differences. The presence of item-level non-invariance does not invalidate the instrument; rather, it calls for cautious interpretation of raw total score comparisons between men and women and highlights the importance of analytic strategies that account for potential measurement non-invariance.

Several practical implications follow from these findings. When comparing gender groups using the “She Asked For It” subscale, differences in raw total scores should be interpreted with caution, particularly when observed differences are relatively small in magnitude. DRF analyses indicate that such differences may partially reflect item-level differential functioning in addition to substantive variation in the underlying construct. These findings suggest that researchers comparing men and women should avoid relying exclusively on raw total scores when drawing substantive conclusions about gender differences in RMA. Latent variable modeling approaches, which explicitly model measurement parameters and potential non-invariance, provide a more robust framework for cross-group comparisons (Putnick & Bornstein, 2016).

The findings further reinforce that DIF should not be interpreted as a purely binary phenomenon. Statistical significance alone is insufficient to determine substantive distortion; the magnitude of parameter differences and their distribution along the latent trait continuum must also be considered (Millsap, 2011; Zumbo et al., 2007). In the current analyses, DIF effects were concentrated primarily within the moderate range of the trait, indicating that potential bias varies according to respondents’ underlying level of RMA rather than affecting all individuals uniformly.

The pronounced DIF observed in items related to clothing, sexual signaling cues, and ambiguous consent may reflect gender-related differences in how such situations are interpreted within contemporary Spanish discourse. Legal reforms and public debate following the La Manada case and the 2022 “Solo Sí es Sí” law have reshaped societal understandings of consent and victim responsibility (Larrauri, 2020; Villacampa & Pujols, 2022). These developments may help explain the pronounced floor effects observed among women for more explicit victim-blaming statements, whereas more situationally or normatively framed items appear to allow greater interpretative variability. Accordingly, DIF effects were especially pronounced in items requiring evaluative judgements about women’s behavior.

That DIF effects were most pronounced in the moderate range of the trait indicates that measurement differences are not uniformly distributed but instead vary as a function of latent endorsement levels. Future research could examine potential cognitive processing differences across gender using qualitative approaches such as cognitive interviewing and think-aloud protocols. Extending invariance testing to additional sociodemographic and linguistic groups would further clarify whether observed item-level non-invariance reflects translation nuances, contextual framing effects, or broader differences in the interpretation of contested social norms (Putnick & Bornstein, 2016).

Several limitations of the current study warrant consideration. First, the sample was drawn exclusively from one region in Spain (Castilla-La Mancha), which may limit the generalizability of the findings to the broader Spanish population. Although quota sampling ensured balance by age and gender, other potentially influential variables—such as education level or political orientation—were not adjusted and may have influenced response patterns. Notably, the sample differed from the Castilla-La Mancha population in educational attainment, with a higher proportion of respondents holding a college degree (41.8%) compared to the target population (25.9%).

Second, while the IRMA-S remains one of the most widely used instruments for assessing RMA, the subscale analyzed here captures only a specific domain: victim-blaming attitudes based on perceived provocation. As public discourse and social norms evolve, new and more nuanced forms of rape myths may emerge. Future research should examine the psychometric performance and measurement invariance of these subtler forms of RMA across diverse demographic and cultural contexts.

Third, the decision to include only one of the four IRMA-S subscales was guided by theoretical relevance and practical constraints. This necessarily limits the scope of the construct assessed and restricts the generalizability of the findings. Because the remaining subscales were not administered, conclusions about the measurement properties of the full instrument cannot be drawn. Future research should validate the complete IRMA-S in Spain to explore whether similar patterns of item-level non-invariance and scale-level impact are observed across other dimensions of rape myths, such as rape denial or perpetrator exoneration.

Fourth, as this study focused solely on the Spanish adaptation of the “She Asked For It” subscale from the IRMA-S, cross-language measurement invariance could not be evaluated. Although the translation followed a structured TRAPD process and aimed to ensure conceptual equivalence, future studies should directly test language-based invariance between the English and Spanish versions to confirm that both operate equivalently across linguistic and cultural contexts.

Finally, the pronounced floor effects observed—particularly among women—reduced response variability, which may have influenced the estimation of item parameters and contributed to the detection of item-level DIF. Although DRF analyses indicated that the cumulative scale-level impact of DIF was modest, restricted variability may still affect the precision with which moderate levels of the latent trait are measured. These findings highlight the importance of continued psychometric evaluation of RMA instruments to ensure sensitivity across the full range of the construct. Future scale development should prioritize items capable of capturing subtle and implicit attitudes, especially in populations where overt endorsement of explicit rape myths has become increasingly rare.

This study highlights the importance of rigorously evaluating the Spanish adaptation of the IRMA-S “She Asked For It” subscale presented in this study before using it for gender-based comparisons in RMA. Although the subscale demonstrated strong internal consistency, a stable unidimensional structure, and metric invariance at the subscale level, item-level analyses conducted within a multi-group IRT framework revealed DIF in most items. Although the cumulative scale-level impact of DIF was modest, these findings indicate that observed gender differences in raw scores may partly reflect measurement artifacts in addition to substantive differences in RMA.

More broadly, the findings underscore the need for rigorous and multi-layered validation when adapting attitudinal instruments for comparative or policy-oriented research. Instruments designed to assess sensitive constructs such as RMA must be both culturally appropriate and psychometrically sound. In this regard, measurement invariance should not be treated as a binary property of a scale but as a multi-layered construct that requires evaluation at the item and scale levels. Careful examination of these layers is therefore not merely a technical procedure but a fundamental requirement for ensuring valid and interpretable conclusions about group differences in both academic and applied contexts.