A burgeoning body of research on attitudes toward homosexuality emerged in the 1970s (e.g., Davison and Wilson, 1973; Morris, 1973; Thompson and Fishburn, 1977; Garfinkle and Morin, 1978; Dressler, 1979). These studies were not explicitly informed by theoretical frameworks but theorized homosexuality as a psychopathology. A single item (e.g., “Most homosexual persons are mentally ill”) was used instead of a theory-based scale based on responses to a category variable (e.g., “yes,” “no,” or “cannot say”).

Subsequently, three theories of attitudes toward homosexuality were developed: the essentialist orientation of homophobia, the social constructivist orientation of the old-fashioned prejudice, and modern prejudice against homosexuality (the second and third are also known as old-fashioned and modern homonegativity; Morrison et al., 2009).

Homophobia, old-fashioned homonegativity, and modern homonegativity focus on negative attitudes toward lesbians and gay men. These theories take such attitudes as unidimensional structures and suppose that individuals’ attitudes are presented through cognition, feelings, and behavioral tendencies. As a result, existing measurements using the above theories as frameworks use item composite scores with a unidimensional bipolar scale (extreme positive to extreme negative) to interpret scores. Positive and negative attitudes are considered mutually exclusive, which implies that attitudes toward lesbians and gay men cannot be both positive and negative. As such, previous research generally categorized attitudes toward lesbians and gay men into positive and negative classes (Larsen et al., 1980; Kite and Deaux, 1986; Herek, 1988; Morrison et al., 1999; Morrison and Morrison, 2002).

However, positive and negative attitudes toward lesbians and gay men could be independent. For example, Jackman (2020) found that the majority of a heterosexual sample held conflicting attitudes toward anti-gay laws in Barbados and sexual activity between people of the same sex. Clausell and Fiske (2005) also found that college students simultaneously had positive and negative stereotypes of gay men. Through qualitative research methods, Wang (2011) found that some lesbians and gay men also had contradictory attitudes toward themselves.

Additionally, the types of attitudes toward lesbians and gay men could fit into more than two classes (i.e., positive and negative). Guo (2023) identified that the dichotomy of positive and negative limits researchers’ ability to distinguish between purely positive and discriminatorily positive attitudes, which differ from one another. Compared to the purely positive class, the discriminatorily positive class entails conditional, selective, and incomplete positivity based on a heteronormative culture. In line with the manifestation of tepid homophobia and veiled homophobia in a novel framework analyzing homophobia (Lyonga, 2021), people holding discriminatorily positive attitudes may not oppose the practice of homosexuality but may be against granting lesbian women and gay men the same rights as heterosexual people.

Lesbians and gay men may also suffer from subtle forms of prejudice masked by a guise in daily scenarios. For example, some individuals could accept others’ homosexuality but not their children’s (Tan et al., 2012; Liang, 2015), or they might have positive attitudes only toward “good” lesbians and gay men who meet their expectations based on heteronormative norms (Bai et al., 2013). Common views of discriminatorily positive attitudes include “homosexuality is normal, but I hope they will not be [sic] conspicuous” and “if homosexuals could become heterosexuals, they would be happier.” The above “positive” attitudes may be well-intentioned to a certain extent, but they reinforce power and status discrepancies between heterosexuals and non-heterosexuals.

Inspired by postmodern feminist psychology, Guo (2023) deconstructed and reconstructed the theory of attitudes toward homosexuality from a sexual-orientation-diverse perspective rather than a single perspective (i.e., prejudice against homosexuality). In this vein, Guo (2023) developed a two-factor and three-class model of attitudes toward homosexuality and the Two-factor Attitudes toward Homosexuality Scale (TAHS).

Theories on prejudice against homosexuality have highlighted the importance of the binary opposition between homosexuality and heterosexuality. Preference for heterosexuality is established based on an individual’s conscious or unconscious admiration for and reinforcement of heteronormativity. As such, when prejudice against homosexuality (Factor 1) is non-existent or low, attitudes should be identified as positive; when the level of Factor 1 is high, attitudes should be negative. Likewise, when there is no or a low preference for heterosexuality (Factor 2), the corresponding attitude is considered positive, while a high level of Factor 2 is considered negative. An inconsistency in attitudes toward Factor 1 and Factor 2 reflects an ambivalent attitude. This hypothesis of a two-factor structure of attitudes toward homosexuality can reconcile the problems posed by the coexistence of positive and negative attitudes.

Additionally, the combination of prejudice against homosexuality and preference for heterosexuality theoretically results in three classes of attitudes toward homosexuality: purely positive, discriminatorily positive, and negative. This hypothesis of a three-class model of attitudes toward homosexuality can help researchers distinguish between purely positive and discriminatorily positive classes. Examples of the three combinations are as follows:

Positive in Factor 1 + positive in Factor 2 = purely positive attitude

Note that the two-factor and three-class model focuses on attitudes toward homosexuality instead of the specific sub-groups of lesbians and gay men. The term “homosexuality” used in measures might cause measurement bias. Such a use of “homosexuality” led to two main issues in previous studies.

Firstly, the results from such a measurement hardly explain whether the measurement represents attitudes toward homosexuality or attitudes toward gay men. There are several reasons for this. For one, when presented with the term “homosexuals,” some participants might consider gay men the target group (Black and Stevenson, 1984; Herek, 1984). Similarly, qualitative research has indicated that lesbians are easily neglected or replaced by gay men in some homosexual rights movements for historical and social reasons (Cheng, 2018). Additionally, descriptions of gay men are more common than those of lesbians in literary works, artistic works, and historical records (Zhang, 2007), which may cause the concept of lesbians to be absent from some people’s concept of ‘homosexuality (Ruan and Tsai, 1987; Cheng, 2018).

Secondly, even if the term “homosexual” is considered to represent both lesbians and gay men, individuals’ attitudes toward lesbians and gay men are different—normally, attitudes toward lesbians are more positive than those toward gay men (Poteat and Espelage, 2005; Kosciw et al., 2009; Zhu et al., 2012). As such, if a participant has positive attitudes toward lesbians and negative attitudes toward gay men, the use of the term “homosexuality” limits participants’ ability to reflect on whether their attitudes toward lesbians differ from their attitudes toward gay men.

We conducted two studies using variable-centered and person-centered approaches to validate whether the two-factor and three-class model of attitudes toward homosexuality is suitable for assessing attitudes toward lesbians and gay men. Additionally, we developed two new scales—the Two-factor Attitudes toward Lesbians and Gay Men Scales (TAHS-L and TAHS-G)—to evaluate the applicability of this model and the interpretability of the results of people’s attitudes.

We utilized EFA to extract common factors of the TAHS-L/G and assessed the TAHS-L/G’s construct validity through CFA. The TAHS-L/G’s convergent validity was tested based on the correlation coefficient between the TAHS-L/G and existing scales. The reliability tests included the Pearson correlation coefficients between item scores and total scores, Cronbach’s α, and split-half reliability. We ran multi-group CFAs and DIFFTEST to assess the measurement invariance of sex. All analyses were conducted in SPSS 22.0 and Mplus 8.

The mean, standard deviation, skewness, and kurtosis coefficients of each item are shown in Table 6. The absolute skewness coefficient values of all items ranged from 0.481 to 3.848, and the absolute kurtosis coefficient values ranged from 0.276 to 15.925.

We conducted a CFA on the two-factor model, which contains two latent variables and 12 measurement indicators with 78 pieces of information (p*(p + 1)/2). There are 25 parameters to estimate, including 12 factor loadings, one covariance between the two factors, and the error variance of 12 items (78–25 = 53 (df)). Furthermore, there are more than three items for each factor without relevant error. Therefore, the model conforms to the rules of CFA model identification.

The WLSMV estimation yielded the following results: (1) TAHS-L: χ² = 874.316，df = 53，χ²/df = 16.497， RMSEA (90% CI) = 0.075 [0.071, 0.080]， CFI = 0.980，TLI = 0.975， SRMR = 0.035; (2) TAHS-G: χ² = 1103.453, df = 53, χ²/df = 20.820, RMSEA (90% CI) = 0.085 [0.081, 0.090], CFI = 0.979, TLI = 0.974, SRMR = 0.037. These results show that the two-factor models of TAHS-L and TAHS-G fit well. Figures 1, 2 present the path diagrams of the TAHS-L and TAHS-G.

TAHS-L: The Pearson correlation coefficients between each item score and the total score were between 0.560 and 0.799. The coefficients between total scores on Factor 1 and its six constituent items ranged between 0.722 and 0.847, and those between total scores on Factor 2 and its six constituent items ranged between 0.643 and 0.789. The Cronbach’s α for the full scale was 0.885 (0.877 for Factor 1 and 0.810 for Factor 2). The split-half reliability of the scale was 0.833 (0.742 for Factor 1 and 0.707 for Factor 2).

We found that the TAHS-L has a significant positive correlation with the AHS-L, r (361) = 0.806, p < 0.001. The scores of the prejudice against homosexuality factor of the TAHS-L had a significant positive correlation with the AHS-L, r (361) = 0.786, p < 0.001, and the scores of the preference for heterosexuality factor of the TAHS-L had a significant positive correlation with the AHS-L, r (361) = 0.849, p < 0.001. Likewise, the TAHS-G had a significant positive correlation with the AHS-G, r (361) = 0.940, p < 0.001. The scores of the prejudice against homosexuality factor of the TAHS-G had a significant positive correlation with the AHS-G, r (361) = 0.886, p < 0.001, and the scores of the preference for heterosexuality factor of the TAHS-G had a significant positive correlation with the AHS-G, r (361) = 0.875, p < 0.001.

Because the chi-square test is easily affected by sample size, and since we used a 4-point Likert scale and, thus, the variable measured using this scale is considered an ordinal variable, we used the WLSMV method for parameter estimation, as well as several fitting indexes, including the Tucker-Lewis Index (TLI), comparative fit index (CFI), root mean square error of approximation (RMSEA), and standardized root mean square residual (SRMR). When comparing nested models, we used the DIFFTEST, which can handle ordinal variables (Li and Liu, 2011). To compare multi-group CFAs and nested models, we used four models: the Configural Invariance Model, Metric Invariance Model, Threshold Invariance Model, and Residual Variance Invariance Model.

CFA in a Single Group. Table 7 presents the CFA results for each group. The results suggest that each model fits well and that it is acceptable to use these models as baseline models for assessing measurement invariance in subsequent analyses.

The equality of the unstandardized item thresholds across groups was then examined in a Threshold Invariance Model (M3). The DIFFTEST shows M3 fitted significantly worse than M2B (p < 0.01), and the modification indices suggest that Items 8 and 9 were the largest sources of the misfit. After the thresholds of Items 8 and 9 were freed, the partial Threshold Invariance Model (M3B) did not fit significantly worse than M2B (p = 0.23). Finally, the equality of the unstandardized residual variances across groups was examined in a Residual Variance Invariance Model (M4). The model comparison at this step proceeded backward. M4 had good fit, and it did not fit significantly worse than M3B (p = 0.32).

Measurement Invariance of the TAHS-G. We considered the male group as the reference group. As shown in Table 9, the Configural Invariance Model (M1) had good fit. The equality of the unstandardized item factor loadings between groups was then examined in M1. The DIFFTEST shows that the Metric Invariance Model (M2) fitted significantly worse than M1 (p < 0.01). The modification indices suggest a localized misfit for the constrained loading of Item 8. After this item’s loading between groups was freed, the Partial Metric Invariance Model (M2B) did not fit significantly worse than M1 (p = 0.15).

The equality of the unstandardized item thresholds across groups was then examined in a Threshold Invariance Model (M3). The DIFFTEST shows M3 fitted significantly worse than M2B (p < 0.01), and the modification indices suggest that Items 7, 10, and 12 were the largest sources of misfit. After the thresholds of Items 7, 10, and 12 were freed, the Partial Threshold Invariance Model (M3B) did not fit significantly worse than M2B (p = 0.08). Finally, the equality of the unstandardized residual variances across groups was examined in a Residual Variance Invariance Model (M4). The model comparison at this step proceeded backward. M4 had good fit, and it did not fit significantly worse than M3B (p = 0.09).

In conclusion, the above analyses show that the partial measurement invariance of both the TAHS-L and TAHS-G was obtained across males and females. That is, the relationships of the items with the latent factor of independent living were equivalent between males and females.

We conducted latent class analysis in Mplus 8—a person-centered approach—to test the validity of the three-class model. We used the maximum likelihood method to estimate the parameters, and we used the maximum expectation method in the iterative process. The evaluation indexes of the model included the Akaike information criterion, the Bayesian information criterion, and the adjusted Bayesian information criterion. Generally, smaller values for these indexes indicate better model fit. We evaluated the classification accuracy with the entropy index, and we used the Lo–Mendell–Rubin test and the bootstrapped likelihood ratio test to compare fit differences among the latent class models. If the value of p of an index indicated statistical significance, we concluded that the k-class model fitted significantly better than the k-1 class model (Wang, 2014).

Taking the 12 scale items as indicators, we converted the 2,720 responses provided on the 4-point scale into values between 0 and 1. “Relatively disagree” and “strongly disagree” responses were recorded as 0 (which represents a negative tendency), and “relatively agree” and “strongly agree” responses were recorded as 1 (which represents a positive tendency). We performed latent class analyses, with each analysis considering between one and four classes.

Tables 10, 11 display the fit statistics of the TAHS-L and TAHS-G, respectively, which are demonstrated through comparisons of the different models’ fit statistics. Critically, the Akaike information criterion, Bayesian information criterion, and adjusted Bayesian information criterion values of the three-class model decreased the most significantly. These results support the use of the three-class model for both the TAHS-L and TAHS-G.

Figures 3A–C display the conditional probability diagrams of two to four classes of attitudes toward lesbians.

Conditional probability diagrams of two to four classes. The abscissa represents the item, and the ordinate represents the response probability (i.e., the probability of each participant’s positive tendency on each item).

In the two-class model (Figure 3A), the scale could divide participants into two classes, namely a positive attitude group (red line) and a negative attitude group (blue line). In the three-class model (Figure 3B), the conditional probabilities of the first class (red line) and second class (green line) for the prejudice toward homosexuality factor (the items on the abscissa were 1–6) were very similar; the conditional probabilities of the second class (green line) and the third class (blue line) for the preference for heterosexuality factor (the items on the abscissa were 7–12) were also very similar. In the four-class model (Figure 3C), the fourth class (black line) appeared in the conditional probabilities, whereas the third class was more explanatory than the other classes.

Therefore, we selected the three-class model, combined with the various fit indexes, the model diagrams, and the interpretability of each class as the most suitable model. According to the responses for all items of the two factors in the three-class model, the first class’s scores on both factors were very high, which we designated as the purely positive class (2,079 individuals) and which accounted for 76.4% of the data. The second class had high scores on Factor 1 and low scores on Factor 2, which we designated as the discriminatorily positive class (554 individuals) and which accounted for 20.4% of the data. The third class scored low on both factors, which we designated as the negative class (87 individuals) and which accounted for 3.2% of the data.

Figures 4A–C display the conditional probability diagrams of two to four classes of attitudes toward gay men. Similar to the findings regarding lesbians, the third class was more explanatory than the other classes. The purely positive class accounted for 74.9% of the data (2,036 individuals). The discriminatorily positive class accounted for 19.8% of the data (540 individuals), and the negative class accounted for 5.3% of the data (144 individuals).