The sample comprised 282 bachelor students from nursing, social education, social work, and police education programs, representing about 31% of the total student population in these programs at three Norwegian higher education institutions (see Table 1). This convenience sample reflects the target population for which the PSO is being validated, namely students in health, social work, and police education programs who represent future professionals expected to work with individuals convicted of sexual offences. Participants were recruited during scheduled teaching sessions and informed that participation was voluntary and anonymous. No incentives were provided. The first author was present during questionnaire administration to address any concerns, but care was taken to minimize potential power imbalances or perceptions of coercion. No students were directly supervised or assessed by the researchers in a grading capacity. The PSO was administered as a standalone instrument, followed only by a brief set of demographic questions; no additional psychological scales were included in the same session, reducing the risk that responses were primed by other attitudinal measures.

The Norwegian version of the PSO translated and used by Friestad et al. (2021) was employed in this study. The wording of all 20 items was taken directly from that version; any minor differences in how items appear in Table 2 reflect editorial shortening for publication and not changes to the administered instrument. Items are rated on a 6-point Likert scale ranging from 1 (“strongly disagree”) to 6 (“strongly agree”). Participants were instructed to respond to each item in relation to “people who have been convicted of a sexual offence”, using the standard PSO item stem (“Please indicate how much you agree or disagree with the following statements”). In addition, a free-text question asked participants which type(s) of sexual offence they had in mind when completing the questionnaire.

All ethical guidelines for voluntary and anonymous participation were followed. Participants were informed of their right to withdraw at any time. Institutional review board approval was deemed unnecessary under Norwegian regulations for anonymized survey studies with adult students. A declaration of no conflict of interest and no external funding is also included.

Confirmatory factor analysis (CFA) was used to examine whether the theoretical three-factor structure of the Perceptions of Sex Offenders Scale (PSO) (Harper & Hogue, 2014) fit the data in this sample. Because the PSO has an established structure, we did not conduct an exploratory factor analysis. Missing data was handled using full information maximum likelihood (FIML), which retains all available observations under the assumption that data are missing at random and avoids the case loss associated with listwise deletion.

Given that several PSO items showed moderate skewness, the CFA was estimated using a robust maximum likelihood estimator (MLR), which provides standard errors and model fit indices that are robust to violations of normality. Items were treated as continuous, consistent with prior validation studies using six-point Likert response formats.

Reliability was assessed using both Cronbach’s alpha and composite reliability (CR). Alpha is reported for comparability with earlier PSO studies, while CR provides a more robust estimate based on factor loadings. Convergent validity was evaluated using Average Variance Extracted (AVE), and discriminant validity was assessed using the Fornell–Larcker criterion, following established guidelines for psychometric evaluation. These indices were used to assess internal construct validity rather than relationships with external measures.

Free-text responses identifying which types of sexual offences participants had in mind when completing the scale were analysed descriptively by categorizing responses (e.g., rape, child sexual abuse, other offences) and reporting their frequencies.

The results of the confirmatory factor analysis (CFA) indicated that the three-factor PSO model showed mixed global fit in this sample. The fit indices were RMSEA = .072, SRMR = .072, CFI = .86, and TLI = .84 (Kenny et al., 2015). While RMSEA and SRMR were within commonly used thresholds for acceptable fit, the CFI and TLI values indicate that the overall model fit was suboptimal and should be interpreted with caution. Modification indices were inspected to evaluate whether theoretically defensible refinements (e.g., residual covariances or cross-loadings) could improve model fit. Although some indices suggested potential modifications, none were theoretically justified and none resulted in meaningful improvements in global fit. The original three-factor model was therefore retained.

The three factors proposed by the PSO (Harper & Hogue, 2014), Sentencing and Management, Stereotype Endorsement, and Risk Perception, showed varying standardized factor loadings (λ) across items (see Figure 1). Factor loadings differed both within and across factors, indicating that items varied in how strongly they were associated with their respective latent constructs. Sentencing and Management and Stereotype Endorsement showed generally moderate to strong loadings across several items, whereas Risk Perception displayed a weaker and more heterogeneous pattern of loadings. This variability indicates uneven measurement strength across the scale and suggests that the Risk Perception factor is less coherently operationalized than the other two dimensions.

Fit indices provide an overall assessment of model adequacy. For the three-factor structure, the CFA produced the following values: RMSEA = .072, SRMR = .072, CFI = .86, and TLI = .84. Taken together, these indices reflect a mixed pattern, with some values approaching commonly accepted thresholds and others indicating areas where the model does not fit the data optimally. Because global model fit alone does not ensure that each latent construct is well defined, we further examined convergent and discriminant validity using Average Variance Extracted (AVE), Composite Reliability (CR), and the Fornell–Larcker criterion (Fornell & Larcker, 1981; Hair et al., 2022).

For convergent validity, Sentencing and Management (AVE = .454, CR = .89) and Stereotype Endorsement (AVE = .344, CR = .71) reached acceptable levels of composite reliability (> .70), whereas Risk Perception did not (AVE = .227, CR = .58). None of the three constructs achieved the recommended AVE threshold of .50, indicating that less than half of the variance in the items was explained by their underlying factors. This points to limited convergent validity, with particularly weak evidence for Risk Perception.

For discriminant validity, the Fornell-Larcker test showed that the square roots of AVE for Sentencing and Management and for Stereotype Endorsement exceeded the squared correlations with other constructs, supporting their distinctiveness. In contrast, Risk Perception again fell short, as its AVE was low relative to its correlation with Sentencing and Management (r = .36). The inter-factor correlations were otherwise modest (r = .12 between Sentencing and Management and Stereotype Endorsement; r = -.26 between Stereotype Endorsement and Risk Perception), further supporting discriminant validity for those comparisons.

Taken together, the findings indicate that although the overall model demonstrates an acceptable global fit, the measurement quality of the constructs is uneven. In particular, the Risk Perception factor shows poor convergent validity and limited discriminant validity, suggesting that its operationalization requires refinement before stronger claims about internal construct validity can be made.

As recommended by Harper et al. (2017), a free-text item was included to identify which type of sexual offence students had in mind when completing the PSO. Of the 211 students who provided a response, 53 explicitly referred to rape, 36 referred to sexual abuse of children or paedophilia, and 104 mentioned both rape and sexual abuse of children together as a combined category. Ten students referred to specific cases they were familiar with, all involving either rape or sexual abuse of children. The remaining eight responses were more nuanced, stating that the offence type “depends” or varies. Overall, these patterns indicate that most respondents focused on a narrow subset of sexual offences, primarily rape and offences involving children, when interpreting the PSO items.

In a context where the penal system is rehabilitative but media coverage of sexual crime is intense, students’ perceptions of risk may be influenced by affectively salient associations with specific offender types, although this cannot be determined directly from the present data. Our findings align partly with Friestad et al. (2021), who reported high reliability for Sentencing and Management (α = .85–.90) and acceptable reliability for Stereotype Endorsement (α = .73–.81) among Norwegian prison officer students, but considerably weaker reliability for Risk Perception (α = .52–.64). While their study did not aim to validate the PSO, the consistency of this pattern across samples suggests that the Risk Perception factor shows limited reliability and weak indicator performance in the Norwegian context. Our findings also resonate with Togas et al. (2022), who conducted a CFA of the Greek version of the PSO in a community sample (N = 452). While Sentencing and Management (α = .90) and Stereotype Endorsement (α = .83) showed good reliability, the Risk Perception factor was not supported at all. Items originally loading on this factor either failed to reach acceptable loadings or migrated to other factors, leading the authors to retain a two-factor solution. This strengthens the concern that the Risk Perception factor may lack robust internal construct validity across different samples and settings, as indicated by consistently low reliability and unstable factor loadings reported in both the present study and previous validation efforts. Although the present study focused on evaluating the originally proposed three-factor model, emerging findings from other contexts, including the two-factor solution identified by Togas et al. (2022), suggest that alternative structures may warrant examination in future research. We did not test alternative models here because the aim was to validate the established PSO structure rather than undertake model development, and the sample size limits the stability of more extensive model comparisons. Future studies with larger and more diverse samples would be better suited for systematically comparing alternative factor structures. Together, these results call into question whether the PSO adequately captures the complexity of students’ perceptions across different samples and contexts. For future research, testing measurement invariance across professional groups will be crucial, as police, nursing, and social work students may draw on different knowledge bases when forming their attitudes.

In addition, our tests of convergent and discriminant validity provide further evidence that the PSO performs unevenly in the Norwegian context. While the global model fit was acceptable, only Sentencing and Management and Stereotype Endorsement reached satisfactory composite reliability, and none of the factors achieved the recommended AVE threshold of .50. The Fornell–Larcker test indicated adequate discriminant validity for most factor comparisons, but Risk Perception again failed to meet the criterion. These findings underscore that the Risk Perception subscale is not only less reliable but also lacks convergent validity and displays weak discriminant validity. In other words, the factor does not appear to capture a sufficiently coherent construct in this setting, which is consistent with earlier research pointing to its fragility across cultural contexts (Friestad et al., 2021; Togas et al., 2022).

A further point concerns the small negative correlation observed between Stereotype Endorsement and Risk Perception (r = -.26). Because this study aimed to examine the psychometric properties of the PSO rather than to analyse substantive attitudinal patterns, this correlation is not interpreted as a meaningful empirical finding. Instead, we consider it an artefact of the psychometric weaknesses of the Risk Perception factor, which demonstrated low internal consistency, low AVE, and limited discriminant validity. The presence of several reverse-worded items may also have introduced response noise, further reducing the reliability of the covariance estimates. In the context of validation, the negative correlation therefore does not challenge any theoretical expectation but reinforces the conclusion that the Risk Perception subscale functions inconsistently and requires refinement in future research.

One explanation may lie in how participants understand and emotionally relate to the risk dimension. In our sample, most respondents associated the term “sex offender” primarily with rape and child sexual abuse, as indicated in their free-text responses (193 out of 211). A similar pattern was observed in the Greek study by Togas et al. (2022), where participants overwhelmingly imagined either rapists or child molesters.

The narrow focus on rape and child sexual abuse may have shaped how participants interpreted the risk-related items. We cannot determine whether emotional reactions played a role, but the strong salience of these offense types may contribute to variability in how risk is understood. This is consistent with previous research showing that people often rely on morally charged associations when forming judgments about sexual offences (Imhoff, 2022). However, because we did not measure affective responses directly, any such interpretation remains speculative. While sentencing attitudes and stereotypes are likely to reflect more stable and cognitively accessible beliefs, perceptions of risk may depend more heavily on visceral reactions, media narratives, and professional familiarity with recidivism. When respondents associate sexual offences almost exclusively with child abuse and rape, affective responses can dominate and create interpretive variability. This is consistent with findings by Imhoff (2022), who showed that moral outrage and affective reactions can override more deliberative assessments when evaluating sexual offenders. This offers one possible explanation for the lower reliability observed in this dimension, but the present data do not allow causal inferences about the underlying mechanisms. Similar validity concerns have been raised in other attitudinal measures, such as the ATS, where certain subscales demonstrated instability across cultural contexts (Rawson, 2024). Narrative framing has been shown in previous studies to influence attitudes toward individuals convicted of sexual offences (Harper et al., 2018). Whether similar processes occurred in our sample cannot be concluded from the present data. Nonetheless, given the interpretive variability observed in the Risk Perception items, future development of the PSO may benefit from clearer operationalisation of this dimension.

Another factor that may have shaped item interpretation is the composition of the sample. The respondents were predominantly women and largely drawn from health and social work programmes, with only a small number of police students. These educational groups may enter the questionnaire with different forms of professional knowledge, exposure to sexual offending, and expectations about rehabilitation or risk. Such differences could contribute to variation in how items are interpreted, particularly those relating to danger, recidivism, and perceived threat. This may partly explain why the Risk Perception factor showed weaker and more heterogeneous loadings in this study.

Additionally, the broader context, where sexual crimes receive frequent media coverage in Norway, may contribute to heightened emotional responses, though we caution against speculating on media influence without more direct evidence. Nonetheless, the fact that so many students interpreted the term "sex offender" in narrow and emotionally charged terms suggests that the PSO may benefit from refinement to better account for emotional and cognitive biases in perception.

Given that the sample included students from nursing, social work, and policing programs, future research should investigate whether the PSO functions equivalently across these groups. Testing for measurement invariance, such as configural, metric, and scalar invariance, would help establish whether the instrument captures the same constructs in a comparable way across professional orientations. This would strengthen its use in interdisciplinary education and research addressing stigma toward individuals convicted of sexual offences.

Harper et al. (2017) have noted that the PSO is a relatively new instrument, still in development. Our findings support the utility of the PSO in a Norwegian context, particularly in its ability to capture attitudes related to punishment and stereotypes. However, like previous studies, our results call into question the robustness of the Risk Perception factor. Harper and Hicks (2022) similarly noted that certain PSO subscales may be particularly sensitive to cultural framing and sample composition, underscoring the need for context-specific validation. The reason for this remains unclear, and future research should explore whether this stems from cultural differences in how risk is conceptualized, from gaps in professional knowledge (e.g., risk assessment competence), or from broader emotional or cognitive framing effects.

Based on our findings and qualitative data, future research may explore whether incorporating a dimension that more explicitly captures participants’ subjective reactions could strengthen the conceptual clarity of the scale. Furthermore, future research should explore measurement invariance across educational subgroups, such as students in policing, nursing, and social work, to determine whether the scale operates consistently across different professional orientations. This would be an important step in assessing both the robustness of the instrument and how professional role identity may shape perceptions.

This study was exploratory in validating a previously translated instrument in a Norwegian educational context. The sample was non-random and drawn from three institutions, with recruitment limited to specific professional programs, which constrains the generalizability of the findings. While the overall response rate (31%) is acceptable, more detailed background information about participants would have allowed for richer interpretation. One notable limitation is the uneven distribution across educational tracks: only 10 participants were drawn from police education, compared to substantially larger numbers from nursing, social education, and social work. This imbalance may have limited the ability to assess whether the PSO functions equivalently across different professional orientations and may have skewed the psychometric results toward attitudes and interpretations more typical of health and welfare students. The predominance of female participants (223 out of 282) may also have shaped responses in ways that limit the generalizability of the findings. Lastly, the presence of the first author during questionnaire administration, though intended to provide clarification, may have introduced some degree of social desirability bias. Taken together, these factors mean that the findings should be interpreted strictly within the context of the student groups included in this study and are not intended to be generalised to broader professional or community populations.

This study contributes to the validation of the Perceptions of Sex Offenders Scale (PSO) in a Norwegian context, using a sample of students from health, social work, and policing education. The findings support the reliability of two of the scale’s three original factors but also highlight consistent weaknesses in the Risk Perception dimension, in line with previous research. These results underscore the importance of testing psychometric tools across different cultural and institutional settings before applying them in research or practice.

In addition to mixed global model fit, our analyses of convergent and discriminant validity showed that Sentencing and Management and Stereotype Endorsement displayed satisfactory composite reliability, whereas Risk Perception did not. None of the three factors reached the recommended AVE threshold, and the Fornell–Larcker test indicated that discriminant validity was only partially met. These results indicate that the weaknesses observed in the Risk Perception factor are not limited to reliability but also extend to its internal construct validity as assessed through AVE and discriminant validity indices.

Our analysis suggests that participants’ responses may be influenced by salient and affectively laden associations with specific offender types, as hinted at in their qualitative comments. However, the present data do not allow us to determine the role of emotional processes directly. Any potential emotional dimension should therefore be regarded as a topic for future research rather than a conclusion drawn from this study. Furthermore, given the differences between educational tracks, future research should examine whether the scale performs consistently across professional subgroups.