Introduction and Validation of the Short Antinatalism Scale (S-ANS)

Generating the Initial Item Pool

The authors generated an initial pool of 22 items, the final version of which can be found in Table 1. To arrive at these items, PS and MM developed two sets of items independently based on the existing antinatalism literature. In the following discussion, we merged these two lists, removed duplicates, and added additional variations and items based on the discussion. We also added reverse-scored items based on the initial set and the discussion. Then, we sent these items to an expert in the field, the philosopher Theron Pummer. Based on his comments, we further adapted the item set. The final 22 items presented here capture antinatalism sufficiently in depth and breadth such that a large variety of different philosophical motivations for antinatalism were included (such as consent-based claims, suffering-based claims, global, as well as local versions). All items are formulated in clear language that is accessible to nonexperts.

Table 1 Factor loadings for 22-item scale and 20-item scale

Items	Model 1		Model 2
	FL	h2	FL	h2
Note. Factor loadings (FL) and communality scores (h2) for the full 22-item scale and the 20-item reduced scale.
(1) Because humans cause so much harm to the environment, it is morally wrong to have children.	.86	.75	.86	.75
(2) Bringing a person into existence is usually a considerable net benefit to that person. (R)	−.50	.25	−.49	.24
(3) In becoming a parent, one creates not only a child and a relationship, but oneself, one creates a new ideally better self-identity. (R)	−.37	.14
(4) It is better not to exist than to exist because only in existence can there be pain. If one does not exist, one cannot be in pain. As such, it is morally wrong to have children.	.84	.71	.84	.71
(5) It is usually morally permissible to bring a child into this world (R)	−.57	.33	−.58	.33
(6) Bringing someone into existence means that they will eventually suffer in one way or another, which is why it is morally wrong to have children.	.88	.77	.88	.78
(7) It is morally wrong to bring a child into this world because a child cannot consent to being born.	.86	.74	.86	.75
(8) It is morally wrong for those unable to care for their children to have children.	.34	.11
(9) Bringing a new child into this world is a morally good thing. (R)	−.60	.37	−.60	.36
(10) Irrespective of how one’s life goes, it is always better to exist than to not have existed. (R)	−.52	.27	−.52	.27
(11) Having a child in the developed world is morally wrong as it impacts the environment at least to the same extent as excessive consumption behavior.	.85	.72	.85	.72
(12) The absence of pleasure is not bad unless there is somebody for whom this absence is a deprivation. This is why not being born is not bad.	.43	.19	.43	.19
(13) People have a moral right to have children. (R)	−.40	.16	−.40	.16
(14) It is always morally wrong to bring a child into this world.	.77	.60	.77	.60
(15) People should adopt rather than have children themselves.	.54	.29	.54	.29
(16) Bringing a child into existence is always a considerable net harm to that child.	.77	.59	.77	.59
(17) Conceiving and raising children is immoral because it requires money, time, and energy that could be spent on giving to charitable organizations that save lives and reduce suffering.	.81	.65	.81	.65
(18) The absence of pain is good, even if that good is not enjoyed by anyone. This is why not being born is good.	.65	.42	.65	.42
(19) Some people’s choices to have children result in other people’s suffering by depriving them of basic rights like the right to food.	.60	.35	.60	.36
(20) No matter your economic circumstances, having a child is never bad thing. (R)	−.41	.17	−.40	.16
(21) It is better to exist than not to exist because only in existence can there be pleasure. If one does not exist, one cannot have pleasure. As such, it is morally good to have a child. (R)	−.43	.19	−.43	.18
(22) It would be good if we all stopped having children.	.78	.60	.78	.61

Table 1 Factor loadings for 22-item scale and 20-item scale

View as image HTML

Our initial item pool included a total of 13 items relating to a priori antinatalism (claims about the universal and general wrongness of procreation), five of which were reverse scored, as well as nine items relating to a posteriori antinatalism (claims about the circumstantial wrongness of procreation for reasons such as human rights and the environment), three of which were reverse scored. Furthermore, the individual items were sometimes phrased as more complex arguments, other times as moral pronouncements or statements of fact. The topics covered by the items were adoption, the relationship between procreation and global poverty, the impact of procreation on the environment, one’s relationship to one’s child, and the general moral status of procreation.

Methods

We recruited 303 participants (50.8% female) via Prolific who were between 18 and 82 years old (M = 44.76, SD = 16.08). Seven participants failed the attention check that asked people to indicate strongly agree on a 7-point Likert scale leaving the final sample at 296. This attention check was used for all three studies on Prolific. All participants had a minimum approval rate of 95% on Prolific and had participated in at least 10 studies prior. All participants were paid £0.50 for the completion of the study. They were presented with the 22 initial items. In this and all further studies, participants were asked to agree or disagree with the specific items on a 5-point Likert scale ranging from strongly disagree (1) to strongly agree (5). All items were presented in a random order to participants. At the end of the survey, participants were asked to complete a short demographic questionnaire.

We sought evidence of structure-based validity via an exploratory factor analysis based on polychoric correlations using the R (R Core Team, 2021) package psych (Revelle, 2022). We investigated the adequacy of the data and sample through the Kaiser–Meyer–Olkin criterion as well as Bartlett’s sphericity tests. We used maximum likelihood (ML) factor extraction method with Oblimin rotation. Furthermore, we drew on parallel analysis based on minimum rank factor analysis with 500 simulations for factor retention (Timmerman & Lorenzo-Seva, 2011) via the EFA.MRFA R package (Navarro-Gonzalez & Lorenzo-Seva, 2021) and FACTOR software (version factor.12.03.02; Ferrando Piera & Lorenzo-Seva, 2017) to calculate the eigenvalues and explained variance based on eigenvalues. Lastly, we used the Tucker–Lewis index (TLI) as well as root-mean-square error of approximation (RMSEA) to evaluate model fit.

Results

Our results showed robust sample and data adequacy, with Kaiser–Meyer–Olkin = .92 as well as Barlett’s sphericity χ²(231, N = 296) = 2,957.2, p < .001. Parallel analysis suggested extracting a single factor, with the explained variance for this first observed factor being 45.1% (eigenvalue = 9.93). The second factor explained 8.84% variance (eigenvalue = 1.94), while the third factor explained 5.95% of the variance (eigenvalue = 1.31). The upper bound of the 95% CI of the corresponding explained variance of the first simulated factor was 9.86%, while the second and third were 9.86% and 9.23%, respectively. This led to the following model fit indices, overall indicating poor model fit: TLI = .762, RMSEA = .118 (CI 90% .112–.126). When extracting the factors with all items, the factor loadings ranged from .40 to .88. We excluded two items that presented factor loadings of < .40. The fit indices remained below the cut-off points for good model fit. The results were, TLI = .79, RMSEA = .119 (CI 90% .111–.127), though reliability indices were rather robust: α = .91, ω = .95, and greatest lower bound of reliability, GLB = .85. Below see Table 1 with the factor loadings for all items (Model 1) and with those items that had a factor loading < .40 excluded (Model 2). To reduce the scale further and to generate hypotheses for Study 2, we conducted a non-preregistered confirmatory factor analysis (CFA). In this analysis, we found better fit statistics for the model with items whose factor loadings ≥ .60 (Model 3): χ²(54, N = 296) = 163.838, p < .001, CFI = 0.943, TLI = 0.930, RMSEA = 0.083 (CI 90% 0.069–0.098), SRMR = 0.039.

Discussion

The data suggested that the extraction of a single factor was most appropriate. While the results indicated good sample and data adequacy, the overall model fit was poor. To reduce the scale and increase model fit, we removed all items with a factor loading of .40 or smaller. However, this also did not improve model fit meaningfully. To arrive at a shortened version that researchers can more easily implement and to improve model fit going forward, we shortened the scale and included only items with a factor loading of .60 or greater in an attempt to eliminate items that loaded worst on the single factor. In non-preregistered CFA, we found improved fit which motivated us to continue working on this scale in this manner. This new 12-item scale was then used for CFA in Study 2 to test our item choice in a new data set and to provide model fit estimations in a new data set.

Methods

We recruited 399 participants (48.8% female) via Prolific who were between 18 and 84 years old (M = 34.34, SD = 11.98). Three participants failed the attention check, leaving the final sample at 396 participants. All participants had a minimum approval rate of 95% on Prolific and had participated in at least 10 studies before. They were paid £0.50 for the completion of the study. We also excluded those who had participated in Study 1. All 12 items were presented in a random order. At the end of the survey, participants were presented with a short demographic questionnaire.

In this study, we ran a CFA for the single-factor model identified in Study 1 (with factor loadings ≥ .60). This analysis was conducted in R’s lavaan package (Rosseel, 2012). We assessed model fit using ML estimator via a combination of the following statistics: chi-squared (χ²), Tucker–Lewis index (TLI), comparative fit index (CFI), root-mean-square error of approximation (RMSEA), and standardized root-mean-square residuals (SRMR). We identified the cut-off points for most of our indices via McNeish and Wolf’s (2021) dynamic fit index cut-off. This applies to SRMR, RMSEA, and CFI (using ML estimation). This dynamic cut-off has three distinct levels: Level 1, Level 2, and Level 3 (with the first level being defined as having 1/3 of items with a residual correlation of .3 that one has failed to include in one’s model, respectively, with second and third level being defined as 2/3 and 3/3 of items, respectively). We further set the cut-off for TLI at .90 in line with Brown’s (2015) recommendation. We also ran a CFA using WLSMV estimator and considering items to be ordered. When using WLSMV, we draw on Brown’s (2015) recommendation: TLI and CFI > 0.90, RMSEA < 0.08 with CI 90% not surpassing 0.10, and SRMR < 0.08.

Additionally, we conducted a Mokken scale analysis (MSA) to select the best items from this instrument (Mokken, 2011, see Franco et al., 2022 for a review). MSA is a nonparametric item response theory (IRT) model, which employs less restrictive assumptions about data that are often done by parametric statistics (in the IRT case, the relation between the probability of obtaining a particular score on a particular item of a latent variable depends on an item response function that often follows a logistic function or a normal ogive model, while in MSA only imposes order restrictions of item response functions, so they can have any shape). The montone homogeneity model (tested here) tests the three central item response theory assumptions: unidimensionality (each item is only measuring one latent trait), local independence (the item responses only depend on the latent trait being measured), and latent monotonicity (the item step response functions are nondecreasing functions of the latent trait). We implemented a genetic algorithm to conduct the Mokken scale analysis because it has been shown to perform well in recovering the correct dimensionality of scales (Straat et al., 2013). Following Straat and colleagues’ recommendations, we use a scalability coefficient of .30 as our cut-off point. We further conduct a manifest monotonicity test (Junker & Sijtsma, 2000), which involves a regression between the scores of individual items and the residual scores (of omitting selected items from the total test score). To adjust for the fact that the number of respondents at specific score levels can be small, we group respondents with adjacent residual scores until a minimum proportion of individuals per score is greater than n/5 (Sijtsma & Molenaar, 2002). We used this approach to test the assumption of nonintersection of item response functions using rest scores. Furthermore, we also assessed the reliability of the scale through the Molenaar–Sijtsma reliability statistic (MS), which is an unbiased estimator of test score reliability (Molenaar & Sijtsma, 1984; Molenaar & Sijtsma, 1988) as well as Cronbach’s α (Cronbach, 1951) and Guttman’s λ² (Guttman, 1945). We conducted the MSA and the reliability analyses in R (R Core Team, 2021) with the Mokken package (Van der Ark, 2012).

Results

Our CFA of the shortened 12-item scale again showed poor model fit using ML estimator across a variety of tests: χ²(54, N = 396) = 301.624, p < .001, CFI = 0.921, TLI = 0.904, RMSEA = 0.108 (CI 90% 0.096–0.120), SRMR = 0.043. The dynamic fit index cut-offs for Level 1 were as follows: CFI = 0.986, RMSEA = 0.044, SRMR = 0.028; Level 2: CFI = 0.972, RMSEA = 0.064, SRMR = 0.031; Level 3: CFI = 0.951, RMSEA = 0.086, SRMR = 0.035. As for the CFA using WLSMV estimator, we found poor fit for the RMSEA indicator but good fit for the remaining statistics: χ²(60, N = 396) = 341.241, p < .001, CFI = 0.973, TLI = 0.966, RMSEA = 0.116 (CI 90% 0.104–0.128), SRMR = 0.042. For factor loadings of this 12-item scale, see Table 2.

Table 2 Factor loadings for 12-item scale

Items	FL ML	FL WLSMV
Note. Factor loadings (FL) for the 12-item scale using maximum likelihood (ML) and weighted least squares mean and variance adjusted (WLSMV).
(1) Because humans cause so much harm to the environment, it is morally wrong to have children.	.807	.876
(2) It is better not to exist than to exist because only in existence can there be pain. If one does not exist, one cannot be in pain. As such, it is morally wrong to have children.	.816	.880
(3) Bringing someone into existence means that they will eventually suffer in one way or another, which is why it is morally wrong to have children.	.829	.878
(4) It is morally wrong to bring a child into this world because a child cannot consent to being born.	.813	.865
(5) Bringing a new child into this world is a morally good thing. (R)	−.475	−.537
(6) Having a child in the developed world is morally wrong as it impacts the environment at least to the same extent as excessive consumption behavior.	.782	.869
(7) It is always morally wrong to bring a child into this world.	.804	.868
(8) Bringing a child into existence is always a considerable net harm to that child.	.767	.817
(9) Conceiving and raising children is immoral because it requires money, time, and energy that could be spent on giving to charitable organizations that save lives and reduce suffering.	.764	.849
(10) The absence of pain is good, even if that good is not enjoyed by anyone. This is why not being born is good.	.682	.748
(11) Some people’s choices to have children result in other people’s suffering by depriving them of basic rights like the right to food.	.601	.677
(12) It would be good if we all stopped having children.	.738	.804

Table 2 Factor loadings for 12-item scale

View as image HTML

Because the first model showed poor model fit, we ran the same analysis on a second model where some items regarding morality had their residuals correlated (Item 6 correlated with Items 1 and 9). This was not preregistered and was as such fully exploratory. The model fit was markedly better but still remained inadequate: χ²(52, N = 396) = 189.432, p < .001, CFI = 0.956, TLI = 0.945, RMSEA = 0.082 (CI 90% 0.069–0.094), SRMR = 0.039. The dynamic fit index cut-offs for Level 1 were as follows: CFI = NONE, RMSEA = NONE, SRMR = 0.024; Level 2: CFI = 0.987, RMSEA = 0.044, SRMR = 0.028; Level 3: CFI = 0.972, RMSEA = 0.066, SRMR = 0.031. Dynamic cut-off point results of NONE indicate that this fit index is unable to differentiate between well-fitting and ill-fitting models for our specific model at this concrete level of misspecification. See OSF Appendix A, https://osf.io/rs23g/ (Maier et al., 2022), for a figure of the comparison of the fit index distributions for the true empirical model and the misspecified empirical model for the first model and the second model. As for the CFA of this model using WLSMV estimator, we also found poor fit for the RMSEA indicator but good fit for the remaining statistics: χ²(62, N = 396) = 227.378, p < .001, CFI = 0.983, TLI = 0.979, RMSEA = 0.092 (CI 90% 0.080–0.105), SRMR = 0.036.

Lastly, our Mokken scale analysis results indicated that Items 1, 5, 9, 10, 11, and 12 should be excluded from further analysis since they did not form a scale (or formed a second scale). Importantly, the excluded items included both items referring to the general claim that procreation is always wrong – or the reversed item that it is good – (Items 5 and 12) as well as more local formulations relating to environmental and human rights-based accounts (Items 1 and 11). As such, this makes it quite unlikely that these items themselves might form a distinct factor. Going forward, we decided to maintain Items 2, 3, 4, 7, and 8. In those items, we did not observe any monotonicity violations and had overall good internal consistency measures (MS = 0.916, Cronbach’s α = 0.911, Guttman’s λ² = 0.912). In further exploratory analysis, we ran a third model based on the 5-item scale drawn from the Mokken scale analysis. We observed better model fit with the ML estimator: χ²(5, N = 396) = 11.457, p = .043, CFI = 0.995, TLI = 0.990, RMSEA = 0.057 (CI 90% 0.009–0.101), SRMR = 0.014. The dynamic fit index cut-offs for Level 1 were as follows: CFI = 0.993, RMSEA = 0.068, SRMR = 0.015, Level 2: CFI = 0.98, RMSEA = 0.121, SRMR = 0.023. We also found good fit with WLSMV estimator: χ²(25, N = 396) = 9.187, p = .102, CFI = 0.999, TLI = 0.998, RMSEA = 0.046 (CI 90% 0.000–0.092), SRMR = 0.011. To test whether this better model fit simply arose because we selected for it post data collection, we ran a further preregistered study (Study 3) to test the 5-item scale specifically.

Discussion

The results of our CFA of the 12-item scale showed unsatisfactory model fit. While we could improve model fit with non-preregistered analyses by correlating some residuals, we drew on Mokken scale analysis to further reduce the scale. Based on this procedure, we removed a further seven items and constructed a short 5-item version of the antinatalism scale. To test the 5-item scale on a new sample, we preregistered and ran a second CFA (Study 3).

Methods

We recruited 300 participants (46.5% female) via Prolific who were between 18 and 74 years old (M = 34.49, SD = 12.28). Three participants failed the attention check, leaving the final sample at 297. All participants had a minimum approval rate of 95% on Prolific and had participated in at least 10 studies before. They were paid £0.50 for the completion of the study. We also excluded those who had participated in Studies 1 and 2. All five items were presented in a random order to participants. After the main scale, participants were represented with three additional scales in random order. First, they were represented with a scale measuring cognitive and affective empathy (Carré et al., 2013). This 20-item scale uses a 5-point Likert scale across all items (ranging from 1 = strongly disagree to 5 = strongly agree). The two-factor model used here consists of two components that assess empathy in young people and adults. These two components are affective empathy and cognitive empathy. (For reliability of these two factors, see Carré et al., 2013, p. 691, Appendix B). Second, they were presented with a life satisfaction scale (Margolis et al., 2019). This scale consists of six items measuring life satisfaction on a 7-point Likert scale (ranging from 1 = strongly disagree to 7 = strongly agree) that also aim to account for acquiescence bias. It has high reliability at ω = .93. Third, they were presented with a scale to measure conservatism (Everett, 2013). This scale aims to measure modern conceptualizations of conservatism on a 14-item scale and includes two subscales: social conservatism and economic conservatism. Participants rated the items on a scale from 0 to 100, where 0 was coded as feeling very negative toward an issue and 100 as feeling very positive. The scale has shown overall strong reliability at α = .88 (with the social conservatism subscale showing α = .87 and the economic conservatism subscale showing α = .70). These were added to seek evidence of validity.

First, we conducted taxometric analyses using R’s RTaxometrics package (Ruscio & Wang, 2021; for further reading, see Ruscio et al., 2013) to test whether our instrument is better described as having latent profiles or dimensions. Taxometric analysis has a core premise that not all individual differences are alike and tries to answer the question of whether the latent construct being measured falls along a continuous spectrum or if constructs form two (or more) separate groups of people. We find evidence for this via the comparison curve fit index (CCFI), where values below .40 indicate dimensions, values above .60 indicate latent profiles/classes, and values between .40 and .60 indicate ambiguity between the two (Ruscio et al., 2018). If we found evidence for latent profiles, we would conduct latent profile analysis, and if we found evidence for dimensions or if the evidence was ambiguous, we could conduct a CFA for the single-factor model. As before, the CFA was conducted in R (R Core Team, 2021), using the lavaan package (Rosseel, 2012) with the same goodness of fit statistics (χ², TLI, CFI, RMSEA, and SRMR), and using both ML and WLSMV estimators. As in previous studies, we calculate the cut-off points via McNeish and Wolf’s (2021) dynamic cut-off indices for the ML estimator and uses Brown’s (2015) recommendation for the WLSMV estimator. Second, we ran a structural equation model with a diagonally weighted least squares (DWLS) estimator using the lavaan package (Rosseel, 2012) in R. As before, we used the following statistics to determine the goodness of fit of the model: χ², TLI, CFI, RMSEA, and SRMR.

Results

Our taxometric analysis produced results that were ambiguous between latent profiles and dimensions, with the comparison curve fit index (CCFI) results being, M = 0.425, MAMBAC = 0.372, MAXEIG = 0.463, LMode = 0.460. In line with our preregistration for values between .40 and .60, we treated the 5-item antinatalism scale as a dimensional model, thus conducting a CFA.

Our CFA results showed excellent model fit across all tests for the ML estimator: χ²(5, N = 297) = 5.227, p = .389, CFI = 1.00, TLI = 0.999, RMSEA = 0.012 (CI 90% 0.000–0.082), SRMR = 0.017. The dynamic fit index cut-offs for Level 1 were as follows: NONE; Level 2: CFI = 0.986, RMSEA = 0.073, SRMR = 0.027; Level 3: CFI = 0.97, RMSEA = 0.113, SRMR = 0.037. We also found good fit with the WLSMV estimator: χ²(25, N = 297) = 7.710, p = .173, CFI = 0.999, TLI = 0.998, RMSEA = 0.043 (CI 90% 0.000–0.099), SRMR = 0.015. See Table 3 for the standardized factor loadings of this 5-item scale, the Short Antinatalism Scale (S-ANS).

Table 3 Factor loadings for the Short Antinatalism Scale (S-ANS)

Items	FL ML	FL WLSMV
Note. Factor loadings (FL) for the 5-item scale using maximum likelihood (ML) and weighted least squares mean and variance adjusted (WLSMV).
(1) It is better not to exist than to exist because only in existence can there be pain. If one does not exist, one cannot be in pain. As such, it is morally wrong to have children.	.753	.885
(2) Bringing someone into existence means that they will eventually suffer in one way or another, which is why it is morally wrong to have children.	.750	.893
(3) It is morally wrong to bring a child into this world because a child cannot consent to being born.	.760	.898
(4) It is always morally wrong to bring a child into this world.	.726	.884
(5) Bringing a child into existence is always a considerable net harm to that child.	.535	.714

Table 3 Factor loadings for the Short Antinatalism Scale (S-ANS)

View as image HTML

Furthermore, we ran a structural equation model using the DWLS estimator, correlating the antinatalism factor with empathy (both affective and cognitive), life satisfaction, and conservatism (both social and economic). We observe negative statistically significant correlations between antinatalism and affective empathy (r = −.082, p < .001), life satisfaction (r = −.198, p < .001), social conservatism (r = −.425, p < .001), and economic conservatism (r = −.360, p < .001). We did not observe a statistically significant relationship between antinatalist views and cognitive empathy (r = .025, p = .159).

Discussion

Our 5-item scale showed excellent model fit across all tests. This indicates that the items from the short scale were accurately captured by our single-factor model and was a suitable candidate to seek additional evidence of validity. We ran a structural equation model correlating antinatalism with a number of constructs that plausibly stand in a relationship to antinatalism: First, those lower on life satisfaction (Margolis et al., 2019) might be higher on antinatalism due to (over)generalization from own experiences. Furthermore, previous research has found a relationship between depression and antinatalist views (Schönegger, 2021). Furthermore, those higher on conservatism (Everett, 2013) might be lower on antinatalism as standardly conservative values like the traditional family with several children are, at least in part, inconsistent with antinatalism. Again, we observed this relationship in the predicted direction. Lastly, we included a measure of both cognitive as well as affective empathy (Carré et al., 2013). We were agnostic as to which direction the effect would run with either type of empathy. On the one hand, previous research found a positive relationship of antinatalism to dark triad traits (and as such low empathy; Schönegger, 2021), though there is also a plausible explanation of expecting antinatalists to be exceptionally high in empathy which plays out in their concern for the unborn generation. Because we failed to find a significant relationship to cognitive empathy but did find a statistically significant negative relationship to affective empathy, we take these data to speak in favor of the former possibility. However, note that while this relationship was statistically significant, its effect size was rather small at −.08, which may make it not as practically relevant as our other relationships.

These results provide support for validity in showing correlations with theoretically related constructs. Because the relationships were all in the same direction as predicted by theory (where such directional predictions based on theory existed), we take this to provide evidence for the validity of this scale and as such, its potential applicability in other contexts. However, to provide further data on validity, we also ran a further study (Study 4) with the S-ANS on a sample of self-described antinatalists to provide additional evidence of the validity of this scale.

Furthermore, the items that make up this final 5-item scale capture the central antinatalism construct by focusing on the primary a priori formulations of antinatalism. Recall that all five items directly draw on the notion of moral wrongness of bringing a child into existence in most (or all) circumstances, while varying the reason (ranging from pain-based arguments to consent-based claims). This captures the essence of the antinatalist view.

Methods

We initially recruited 351 participants via the Antinatalism sub-reddit on https://www.reddit.com. However, data collection on Reddit faces issues of repeated submissions and potential bot submissions. Therefore, we used Google’s invisible ReCaptcha bot detection software integrated in Qualtrics that estimates the probability of the participant being a human (with 1 being certainly a human and 0 certainly a bot). We excluded all participants with a bot score lower than .8 (48 participants). Furthermore, we removed participants that entered the study prior to activation of software that prevented repeated submissions by blocking participants from entering the study again (123 participants). The activation of this software was made in response to a sudden influx of low-quality data within the first hour of data collection that we had not anticipated. To ensure high data quality, we removed all entries submitted prior to the activation of this software, though all participants remained eligible for payment. This step was not preregistered, but we felt that deviating from the preregistered protocol was needed given this situation. In addition, we excluded participants that took less than or equal to 60 s to complete the survey (24 participants), that answered the attention check incorrectly (68 participants), or that did not self-identify as antinatalist (132 participants). Note that several participants may have been excluded based on more than one criterion, and as such, the sum of all these exclusions may not equal the total exclusions.

In total, we excluded 252 participants. This left us with 99 participants (40.4% female) between 18 and 59 years old (M = 26.81, SD = 7.77). Participants could enter a lottery for a £50 gift card if they wanted to. Participants were presented with the final five items before being shown a short demographic questionnaire and a question asking them if they self-identified as an antinatalist.

We tested the differential item functioning (DIF) to see if the instrument is invariant between antinatalists and nonantinatalists (Jaloto, 2021). To analyze DIF, we are following five steps: (1) item calibration and parameter estimation using the graded response model (2) scores estimation using expected a posteriori (EAP) method, (3) identification of items with DIF, (4) in case we find items with DIF, we will return to Step 1 considering that the item is not invariant and proceed with the following steps, and (5) model fit verification using Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), expecting that the AIC and BIC for the last model is lower than the first model. All DIF analyses were made using R (R Core Team, 2021) packages mirt (Chalmers, 2012) and lordif (Choi et al., 2011). For a full specification of all three models, see OSF Appendix B, https://osf.io/rs23g/ (Maier et al., 2022; Jodoin & Gierl, 2001; Zumbo, 1999).

After detecting if items show DIF or not, we estimated subjects’ EAP scores – the expected a posteriori scores, i.e., the posterior probability distribution of the latent scores of this trait – considering DIF’s results and use them to test mean differences between antinatalists and nonantinatalists (from Study 3) with a Welch t test in R (R Core Team, 2021). The effect size was calculated by Cohen’s d and Hedges’s g using the effect size (Ben-Shachar et al., 2020) R package.

Results

First, we proceed with the DIF analysis with the five items of the S-ANS. We conclude that the Short Antinatalism Scale has five items that are invariant between self-described antinatalists and non-self-described antinatalists. For a full set of results of this measurement invariance analysis, see OSF Appendix C, https://osf.io/rs23g/ (Maier et al., 2022).

Second, we proceeded with the Welch t test with EAP scores to analyze mean differences between self-described antinatalists and non-antinatalists. In Figure 1, we depict a raincloud plot visualizing the distribution of EAP scores for antinatalists and the general population.

Overall, we find that self-described antinatalists (M = 0.011, SD = 0.896) scored significantly higher on the antinatalism scale than the lay population (M = −2.708, SD = 0.994), t(184.76) = 25.435, p < .001, Cohen’s d = 2.80 [95% CI (2.50–3.10)], Hedges’s g = 2.80 [95% CI (2.50–3.09)]. The plot was created with JASP (JASP Team, 2022).

Discussion

These results provide additional evidence of the validity of our scale, as we find that the scale is measurement invariant between antinatalists and the general population. In addition, we find, as expected, that antinatalists score higher with a large effect size (d = 2.80). This further provides evidence of validity as we have shown that self-described antinatalists score considerably higher.