The construct of moral vitalism provides an important avenue through which to examine the role of spiritual beliefs in moral reasoning and judgement. It goes beyond prior work focused on how a belief in God may shape thinking and behavior (e.g., [6]) or work focusing on concerns over sanctity or purity (e.g., [7]), by examining a general lay theory about spiritual forces that is not limited to specific religious commitments or political beliefs. Moral vitalism provides people with a convenient explanation for why good and bad things happen, as well as what makes people good or bad. Like other lay theories, moral vitalism may often be largely implicit and poorly articulated.

By focusing on how people explain their moral worlds, moral vitalism provides an account of moral cognition which diverges from the view that morality arises from the need to protect persons, groups, or norms (such as moral foundations theory, see [7]). While it shares some similarities with accounts that emphasize the role of sense making in the context of harm (such as the notion of dyadic completion: [8]), it goes beyond such accounts by suggesting that in efforts to understand their worlds, people often rely on beliefs that have explanatory power. Beyond completing a moral dyad of victim and perpetrator, moral vitalism serves to explain why there are victims and perpetrators in the first place. Yet, in contrast to theories focusing on the moral character of agents (e.g., [9]), moral vitalists see moral action as in part determined by forces that are independent of people, but which can possess and influence them (see [1]). Supportive of this, moral vitalism appears to be relatively distinct from a similar construct focusing on pure good and evil published by [10], but which focused on purely good or evil people, as opposed to purely good or evil forces (see [1]).

In short, moral vitalism provides a filler explanation, or a placeholder concept, for why morally relevant events occur. As an approach to understanding moral cognition, it suggests that morality may be as much characterized by a set of beliefs about the nature of the world as it is by the basis on which people make moral judgements.

As a formal theory, moral vitalism reflects a basic form of cognition (belief) that is likely universal and probably arose as an explanation for life threatening events, such as disease in contexts where other (i.e., more scientific) explanations were not available (see [11]). From this perspective, moral vitalism is perhaps most closely aligned with the purity dimension of moral foundations theory, given a similar emphasis on concerns regarding purity and contagion, and links to concerns over biological disease ([7, 12]). As such, moral vitalism is likely to be evident across a range of cultures, and yet it is also likely to be reinforced within particular cultural contexts.

The theory and measurement of moral vitalism provides exciting and novel avenues for cross-cultural research. For example, [11] showed in a cross-national study that historically higher levels of pathogen prevalence were positively associated with the endorsement of moral vitalistic beliefs. Other social ecological factors could also be explored, such as the prevalence and frequency of natural disasters, because they may equally be related to the development of moral vitalism as a cultural meaning-making belief system. Given that moral vitalism takes account of the role of spiritual belief within the moral domain by focusing on core underlying assumptions, rather than specific tenets of religious, cultural or political beliefs, it opens up new avenues for comparison across culturally diverse populations, and across varying ideological and religious commitments. A cross-nationally reliable scale on moral vitalism is able to provide novel insights into the factors that stimulated the development and preservation of such beliefs.

The current version of the MVS features five items assessing the belief in real, agentic forces of good and evil (e.g., “There are underlying forces of good and evil in this world”) on a 6-point Likert scale ranging from 1 = strongly disagree to 6 = strongly agree [1]. The item wordings are presented in Table 1. The scale was translated from English into the respective languages of countries (see Table 2) by bilinguals. Accuracy of the translation was verified through back-translations or a committee approach. S1 File include translations of the MVS into these languages.

A total of 3,202 undergraduate university students residing in 28 countries participated in this study. Ethical approval for this study was obtained by the last author from the University of Melbourne’s Behavioural and Social Sciences Ethical Review Committee, project no. 2009001486. Informed consent was obtained in line with the requirements of ethical approval. All other samples in this study were collected in line with relevant ethical protocols and informed consent procedures for each country.

The participants and cultures were sampled on a convenience basis. The central team reached researchers through professional networks and asked them to collect a sample of 80–100 respondents. Participants responded to a larger questionnaire that included the MVS. Respondents who took part in the study received non-monetary incentive such as course credits.

The average age of the total sample was 22.6 years (SD = 6.3), ANOVA test showed significant differences in age across samples. In Spain, the average age was substantially higher as the participants were students of an open university, which attracts more mature attendance. Gender composition was also significantly different across samples as indicated by a significant χ² test. Overall 65% of all participants were female (see Table 2). The data contained less than one percent of cases with missing values, these were treated with full information maximum likelihood in frequentist models, whereas Bayesian models implicitly incorporated missing values.

The analysis followed three major steps. At first, we identified a pool of the items that showed similar factor structure across cultural groups. Second, we ran a series of full and partial exact invariance tests. Third, an approximate invariance approach was applied to adjust the model to the population more closely. At all stages we applied confirmatory factor analysis models using R package “lavaan” [13] and Mplus version 7.3 software [14]. All the codes with details of the analyses and the replication data can be found in the S1 File.

A conventional way to assess measurement invariance is to run a series of multiple groups confirmatory analyses [15]. First, a configural model is fitted to the data. Configural model does not constrain factor loadings or item intercepts. Second, a metric invariance model is fit to the data, which is similar to the configural model, but the factor loadings are constrained to be equal across groups. And finally, a scalar invariance model is fitted, which constrains both factor loadings and item intercepts. After fitting these three models to the data, they can be compared using likelihood ratio χ² test. However, χ² difference test was shown to be overly conservative with larger sample sizes [16]. Thus, another set of criteria was suggested: differences in comparative fit index (CFI) larger than .01 and difference in RMSEA larger than .015 as an indication of substantial differences between models [17]. If a configural model has a substantially better fit than a metric invariance model, it is preferred over the metric invariance model. Likewise, if a metric invariance model has a substantially better fit than a scalar invariance model, the former is preferred over the latter.

An important prerequisite to this sequence of model testing is a well-fitting configural model. As a criterion of an acceptable fit for a factor model we used CFI/TLI > .90, RMSEA < .08; SRMR < .08 [18]; χ² was ignored in this analysis due to a large (over 3000) sample. The model was identified using a marker variable approach (see [19]). After a preliminary analysis, we decided to use item “existence” as a marker, because it showed highly invariant parameters across groups when using different model specifications. Switching of a marker variable to the other items did not affect substantive results.

At the second stage of analysis we tested several partial invariance models. Byrne, Shavelson, and Muthén [20] proposed the idea of partial invariance, which claims that some of the loadings/intercepts are allowed to vary without creating a substantial bias. They suggested that it is sufficient to have two invariant loadings for a partial metric invariance and two invariant intercepts for a partial scalar invariance. However, the consecutive tests of partial invariance might lead to an inductive solution which brings a danger of overfitting, that is, a possibility that the final model may not replicate with different data. For this reason, we were especially careful in handling model modifications which were not expected by the theory.

The third stage of analysis was based on Bayesian statistics with informative priors. The prior between-group difference in factor loadings and/or item intercepts was set to zero (no differences) and its variance was set to a conventionally low value. The between-group variance in factor loadings and intercepts of the size .01 was considered negligibly small for most substantive conclusions (it defines the 95% confidence interval of absolute differences in parameters of .20, see also [21]). Between-group variance in parameters of .10 was considered large and deviating from the invariance (corresponds to ±.63 difference on a standardized scale).

First, we tested an approximate configural model, in which the factor loadings and item intercepts were allowed to vary with a prior between-group variance of .10. Given a well-fitting configural model, we then set the prior between-group variance to .01 for loadings (approximate metric invariance) and then for both loadings and intercepts (approximate scalar invariance). In order to evaluate the model fit to the data, two key statistics were examined: 95% confidence interval for the difference between the observed and the replicated χ² values (if it includes zero, it indicates acceptable fit), and related standardized index of posterior predictive p-value (PPP, which should be higher than .05 with a perfect fit indicated by value of .50). As long as the sample sizes were relatively small, we expected convergence issues. To ensure convergence the models were run using 4 chains and a minimum of 30,000 iterations. All the models revealed stable solutions indicated by potential scale reduction factor (R-hat), which was smaller than 1.1 for every parameter. Scanning of trace plots did not reveal issues with convergence as well. Visual examination of autocorrelation plots detected no substantial problems with the parameter estimation.

The MVS initially included eight items: three reverse-coded items and five non-reversed items. Non-reversed items are listed in Table 1; reversed items were following: “Good and evil are human constructions”, “Things happen and sometimes they have good or evil consequences, but there is nothing that is truly good or truly evil”, “There is nothing that is really good or really evil in this world, it’s all a matter of perspective”. A preliminary analysis on the pooled sample demonstrated acceptable fit of a two-factor model: one factor loaded on all the reverse-coded items and the other factor loaded on the rest of items [CFI = .964, RMSEA = .069, SRMR = .045, χ²(df = 19) = 299]. However, applying this model in the multiple group settings was challenging, as in many groups it resulted in a non-positive definite matrix, which signaled potential problems with the model and the data (such as high multicollinearity combined with small samples). Moreover, running the two-factor model separately in each group resulted in various model problems across groups.

We concluded that configural invariance across groups could not be detected with the eight items. Therefore, we dropped the reverse-coded items and examined measurement invariance for a one-factor model with the five straight-coded items only.

A simple multiple-group factor model with a single factor at the pooled sample showed unacceptable fit as indicated by a high value of RMSEA (.098; the other fit measures were within the recommended range, CFI = .971, TLI = .942). Modification indices suggested adding a covariance between items “awareness” and “natural” which improved the model and yielded a marginally acceptable value of .078 for RMSEA. At the next step, the modification indices suggested to add a covariance between residuals of the items “natural” and “existence” which makes sense, because both items claim the existence of good and evil forces in a natural world. The resulting model (Fig 1) had appropriate fit in a multiple-group setting (see model 1 in Table 3), as well as it fitted to the data from each country separately–p-values of χ² tests of model fit were above .01 in every group. Therefore, this model was further used in the invariance testing as a configural model. The alternative model with two factors, one including “existence”, “natural”, and “responsible”, and the other including “motivates” and “awareness” showed acceptable fit on the pooled sample but did not converge in multiple-group analysis (configural model).

Table 3 lists fit indices from a series of the exact measurement invariance tests with multiple group confirmatory factor analyses. Constraining factor loadings across groups to equality, that is, testing the metric invariance, decreased CFI by .045 and increased RMSEA by .034, both values are much greater than the recommended ones. Therefore, we had to select configural invariance model over metric invariance and concluded that there is no full metric invariance.

Next, we tested partially invariant models. We examined the factor loadings in each group from the configural model in which they were estimated independently in every group. This information as well as modification indices indicated that factor loadings of items “existence” and “awareness” had the largest differences across groups, so we relaxed these equality constraints. Comparing the model fit of the partial metric invariance model (model 4) to the fit of the configural model revealed a decrease in CFI of .011 and increase in RMSEA of .013, which almost fell in the recommended range. So, we were able to conditionally accept this partial metric invariance model. However, when we tested the partial scalar invariance model, the decrease in CFI and increase in RMSEA were very large (model 5), thus, it was rejected. We also tried a minimal partial metric invariance model, releasing equality constraints on loadings of “existence”, “nature”, and “awareness” items, which expectedly showed an acceptable decrease in the model fit (model 6) compared to configural invariance model. This model was then used to test for the partial scalar invariance (model 7), a decrease in CFI was .014, close to the suggested threshold, but increase in RMSEA was .022, substantially larger than the recommended threshold of .015, and the absolute value of RMSEA of .096 pointed to the model’s overall poor fit to the data.

The resulting partial metric invariance model is prone to criticism, because, first, the RMSEA statistic, although technically being under the recommended threshold of .08, had an upper confidence bound which was beyond this threshold. Moreover, these models had a very few degrees of freedom, which may imply that the models could have been overfitted, that is, be applicable only to the current sample. Given the nature of the convenience sample, the problem of overfitting is especially important issue, so the minimally fitting model might not be the population model. Due to these concerns, we switched to the approximate measurement invariance approach [21–22], which in a way avoids the problem of overfitting by allowing some across-group variance in item intercepts and factor loadings.

Table 4 lists the model fit statistics of approximate invariance testing. The approximate configural model fits the data very well, showing PPP of .164 and a confidence interval of χ² including zero. The metric model showed a similarly good fit with PPP = .060. The scalar invariance model showed a deterioration in fit with a PPP dropping below the recommended threshold of .05 to essentially zero, and BIC increasing compared to the metric and configural models. Increasing prior variance of the intercepts of the items “natural”, “awareness”, and “responsible” resulted in a partial approximate scalar invariance model that fitted the data sufficiently well: PPP of .061 was in the recommended range, χ² confidence interval included zero and the increase in BIC was relatively small. We consider this model final and conclude that the measurement model of moral vitalism with five items and two residual covariances (of the items “natural” with “awareness” and “existence”) is approximately and partially invariant at the scalar level, which implies that the means and regression/correlation coefficients can be compared across all 28 cultural groups.

The latent means estimated by the partial approximate scalar invariance model are listed in Table 2. The latent mean had to be fixed to zero in one of the groups, so we randomly chose Japan. The other means represent differences in moral vitalism compared to Japan. Expectedly, Western European countries occupied the lower ranks, while Asian cultures as well as USA and Cyprus scored the highest on moral vitalism. Previous analyses have shown that country-level the moral vitalism score correlated robustly with pathogen prevalence, therefore, underscoring the cross-cultural validity of the scale [11]. Interestingly, the lower levels of moral vitalism tended to coincide with the larger confidence interval of the estimate.

It is noteworthy that the latent means estimated by different models were highly similar. For example, the latent means from the exact scalar invariance model (model 3 in Table 3 which showed poor fit) correlated with the means estimated by the approximate partial scalar invariance model (model 4 in Table 4) at .99.