The authors have declared that no competing interests exist.
Conceived and designed the experiments: JKA KRL. Performed the experiments: JKA KRL ØLM CHB. Analyzed the data: JKA KRL CHB. Contributed reagents/materials/analysis tools: JKA ØLM CHB. Contributed to the writing of the manuscript: JKA KRL ØLM.
Some disciplines in the social sciences rely heavily on collecting survey responses to detect empirical relationships among variables. We explored whether these relationships were
In this study, we explore how survey response patterns may be predicted using information available prior to conducting a survey. Such techniques have several interesting consequences for theory development and testing in the social sciences.
Many social science disciplines acquire data from surveys. The focus of interest is usually in how different variables relate to each other, allowing exploration of relationships such as those between leadership, motivation and work outcomes. To understand how these variables are related, researchers have hypothesised the existence of ‘latent variables’ – hidden sources of quantitative variation stemming from variables such as different types of leadership and motivation
The Achilles heel of this research is the nature of variation in survey scores. The most common input to the computational tools is the inter-item correlation matrix, or the degree to which any two items in the survey tend to co-vary in a systematic way
However, a fundamentally different explanation is possible. The main source of quantitative variation in the surveys may instead be the degree of semantic overlap among the items. We will attempt to show empirically how a
The statistical treatment of survey data in the social sciences has developed as a discipline often referred to as ‘psychometrics’, originally developed from research on intelligence
We cannot know
This semantic linkage of items is the core of what we believe to be a misunderstanding in survey-based research, demonstrable through semantic research. General psychometric theory asserts that some semantic overlap is necessary to create intra-scale consistency, usually measured by the formula called ‘Cronbach's alpha’
As we will show empirically, this assumption does not hold. The semantic relationships hold across different scales despite their apparent separation by factor analysis. The resulting inter-item correlations can be explained by their semantic relationships. This is unfortunate because it undermines the value of factor analysis in establishing scale independency and also raises fundamental questions about the empirical object of such techniques.
Our concerns are not new in research on surveys and psychometric theory. More than five decades ago, Coombs and Kao
The need of the digital community to store, search, index and extract large amount of texts has stimulated the development of techniques that are sufficiently reliable and developed to take on survey research
We have chosen two types of text algorithms for this task. One is called
LSA functions by analysing texts to create a high-dimensional ‘semantic space’ in which all terms have specific locations, represented as vectors. LSA can then ‘understand’ new texts as combinations of term vectors in this space. LSA aggregates the word contexts in which a given word does/does not appear and provides a set of constraints that determines the similarity of meanings of words and sets of words. Thus, when two terms occur in contexts of similar meaning –even in cases where they never occur in the same passage –the reduced-dimension solution represents them as similar. Similarity is indicated by the cosines of the vectors in semantic space, taking on values between −1 and 1. Some practical examples: The two sentences “doctors operate on patients” and “physicians do surgery” have no words in common, but a commonly used LSA semantic space called TASA (Touchstone Applied Science Associates) estimates their overlap in meaning at .80. Furthermore, sentences with similar words do not necessarily appear as similar. For example, the LSA cosine for the two expressions “the radius of spheres” and “a circle's diameter” is .55, but the cosine for the sentence pair “the radius of spheres” and “the music of spheres” is only .01
LSA represents a sparse matrix of documents (columns) vs. terms-in-those-documents (rows). The matrix is generally set to downweigh common words. It is sometimes normalized before using an algorithm –singular value decomposition –similar to factor analysis. LSA then yields the aforementioned semantic space. This method now has well-documented text-recognition applications
Our approach was to let LSA detect accumulated knowledge and semantic relationships within texts relevant to respondents of organisational surveys. We defined relevant texts as articles from three different domains of media: Business-press texts, general newspaper texts, and PR-related texts.
The business-press texts were excerpts from
The news excerpts were from
The PR statements were taken from
These materials allowed us to create three distinct ‘semantic spaces’, i.e. high-dimensional spaces in which all terms have a specific vector or location, allowing LSA to ‘understand’ the text of survey items. Every survey item in the study was projected into each semantic space to generate its mathematical representation (vector). These representations were, in turn, compared to each other, allowing computation of cosine angles between all the item vectors, with higher cosines indicating higher similarity between items. This procedure was repeated for all three semantic spaces.
While LSA ‘extracts’ meaning from the way words are used in texts, the MI algorithm
More detailed descriptions of these algorithms may be found in the
Particularly salient examples of our theory are found in research on constructs such as ‘leadership’, ‘motivation’ and their purported outcomes. These constructs are prevalent in the research field known as Organisational Behaviour (OB), where central research topics are different types of leadership and their relationship to psychological processes in workplace behaviours. One of the most well-researched and popular theories of leadership during the recent decades has been ‘transformational leadership’, belonging to a set of leadership theories called ‘neo-charismatic leadership’
We will argue that the analysis of Van Knippenberg and Sitkin is applicable to other theories of leadership as well, at least as long as surveys are used for measurement.
The concept of ‘motivation’ is a good case in point. An examination of the semantic network of ‘leadership’ in the lexical database WordNet shows how leadership is related not only to outcomes but also to motivation; some popular definitions of ‘leadership’ are precisely acquired through motivating people to rally around some objective
There are different ways to assure consistency in scales. Some scholars have argued that items should sample from a wide, non-synonymous domain to avoid semantically caused alpha coefficients
Our proposal is therefore that the quantitative relationships among these variables as surveyed are largely determined by their semantic properties. The psychometric validation of latent variables usually depends on the following steps of statistical analysis:
Establishing that every scale used is coherent, using Cronbach's alpha or similar;
Verifying that the scales are semantically independent of each other, using factor analytic techniques;
Establishing a quantitative model of how the variables are related to each other, usually involving structural equation models or some kind of statistical ‘mediation’ effect signalling that the findings are part of a larger nomological network; and
Using various statistical procedures to establish fit indices, used to determine the statistical significance of the whole model as such, compared to contrasting explanations or chance.
In the following sections, we will show how all these steps may be replicable in four large different samples using language algorithms applied to state-of-the art survey scales on leadership, motivation, outcomes and personality. We hope thereby to show that new technologies may be developed to illuminate this field, and also to substantiate our claim that psychometric assumptions about the origin of quantitative variation need to be revised.
The following four sections all contain their own description of methods. We declare that for all of them, data from human subjects were collected according to the ethical regulations for doing research in Norway, where the data were obtained. All respondents consented to take part voluntarily and were informed that they could withdraw from the study at any time. This research is not health related, but only asks participants to anonymously fill out survey forms with non-sensitive questions. Norway has no specific ethics committee to oversee this kind of non-clinical research, but a governmental body called the Norwegian Data Protection Authority (NDPA) rules whether such projects must be approved and registered to guarantee the legal and ethical protection of participants. We asked NDPA about the data collection (inquiry no. 28024 in 2006). The NDPA ruled that anonymous participation, i.e. submitting a completed survey questionnaire, is taken as sign of consent in three of these studies, as the procedure was seen as harmless, the questions were deemed not sensitive, and there were no way of tracing either answers or non-compliance back to individual respondents. In
Only an overview is given of the four core steps of the LSA process because of the careful treatment of LSA elsewhere, including a publication by one of the authors
LSA starts by creating a term-document matrix, A, containing a weighted count of how many times a word, i, appears inside a document, j. The weighting method employed here, log-entropy, has been found generally to outperform other LSA weighting schemes
After appropriate preparation (weighting, normalisation, etc.), this matrix is decomposed using ‘singular value decomposition’, a mathematical algorithm similar to a factor analysis, with the result being a semantic space of a given dimension represented as three matrices: U, a term-by-dimension matrix representing words; S, a singular value matrix; and V, a document-by-dimension matrix representing documents. The equation can be written as:
Given the query
To find similar items to
The MI sentence similarity measure is computed for two candidate sentences, S1 and S2, as follows.
The process begins with tokenisation and POS tagging of all the words in the survey item with their respective word classes (noun, verb, adverb, adjective and cardinal, which also plays a very important role in text understanding).
Each word in the sentence is measured against all the words from the other sentence to find the highest semantic similarity (
Once the highest semantic similarity (
Neither LSA nor MI discriminates well between negative or positive assertions, and MI does not take negative values at all. The two sentences “It is raining” and “It is not raining” are indexed as very similar with high positive semantic scores, and both of them are very different from “The cat is on the mat”.
The handling of positive and negative values is of principal importance in the following analysis, and we need to dedicate some attention to this issue. The relationship between two series of numbers depends greatly on the distribution of signs. Appropriate handling of the direction (sign) of the correlations is crucially important to estimating the true mutual variance between semantic similarity indices and observed survey correlations.
In the case of backward-scored items, to prevent biases from response sets in the respondents, these are easily corrected. However, one of the surveys we apply here–the Multifactor Leadership Questionnaire (MLQ, see further explanation below)–does not contain such scores. Within this survey, 264 (26.7%) of 990 pairs of items are negatively correlated. Theory suggests that two scales, ‘laissez-faire’ and ‘passive management by exception’, are likely to relate negatively to effective leadership. Common sense seems to indicate that pairing items from these scales with items from other scales would correlate negatively. One typical example of negative correlation is between a) an item stating that a manager is unapproachable when needed and b) another item stating that the same person uses appropriate methods of leadership. The surveyed responses to these items correlated −.42 in our sample, but the semantic identity values range between .38 and .75. There is no
We chose the Multifactor Leadership Questionnaire (MLQ) due to its widespread use and central position in leadership research over the recent decades
Cronbach's α by source of data | |||
MLQ scale | Empirically observed α | MI semantic α | α from semantically predicted correlations |
Idealised influence attributes | 0.85 | 0.82 | 0.80 |
Idealized influence behavior | 0.87 | 0.79 | 0.81 |
Inspiring motivation | 0.61 | 0.45 | 0.51 |
Intellectual stimulation | 0.87 | 0.79 | 0.79 |
Individualized consideration | 0.88 | 0.85 | 0.83 |
Conditional reward | 0.83 | 0.82 | 0.81 |
Mgmnt by exception active | 0.72 | 0.85 | 0.83 |
Mgmnt by exception passive | 0.60 | 0.83 | 0.81 |
Laissez-Faire | 0.83 | 0.83 | 0.80 |
Extra effort | 0.89 | 0.77 | 0.78 |
Effective group | 0.76 | 0.84 | 0.84 |
All outcomes | 0.92 | 0.90 | 0.91 |
We regressed the semantic similarity indices on the empirically observed correlations among the MLQ items, which yielded an R2 of .79, p<.01. Finally, we computed a General Linear Model (GLM) with: a) MLQ correlations as dependent; b) semantic indices as covariates; and c) knowledge about scale belongingness as a fixed variable (making complete use of the knowledge that is accessible prior to running a survey). This model yielded an R2 of .86, p<.01. For both regression models, we saved the predicted values and residuals.
MLQ scales with outcome variables | Average surveyed correlations | Linear regression predicted correlations | GLM predicted correlations |
Idealised influence (attrib.) with outcome | .52 | .45 | .52 |
Idealised influence (beh.) with outcome | .51 | .44 | .51 |
Inspiring motivation with outcome | .52 | .47 | .52 |
Intellectual stimulation with outcome | .50 | .43 | .50 |
Individualised consideration with outcome | .54 | .48 | .54 |
Conditional reward with outcome | .47 | .43 | .47 |
Mgmnt by exception active with outcome | .16 | .42 | .16 |
Mgmnt by exception passive with outcome | −.19 | −.25 | −.19 |
Laissez-faire with outcome | −.36 | −.25 | −.36 |
Outcome with outcome | .60 | .53 | .60 |
Random pairs of items | .18 | .19 | .18 |
GLM predicted scores in the rightmost column.
CFA of all 10 MLQ subscales in the present sample yielded a comparative fit index (CFI) of .93, a root mean square error of approximation (RMSEA) of .05, and a standardized root mean square residual (SRMR) of .07. The error terms were not correlated, and these figures are usually interpreted as indicative of an acceptable model
Again, the first 36 items describing leadership behaviours from the MLQ
The sample consisted of 255 employees at a governmental research agency, mostly scientists and engineers. Of these, 66.7% were male, and the mean age was 38 years. One quarter rated themselves as managers, and the rest termed themselves as “project team members”.
Using observed correlations, the alphas ranged from .53–.96, mean = .85. For the semantic alphas, these numbers were .58–.97, mean = .86, and for the predicted correlations, .35–.91, mean = .72. These latter values correlated (.91 and .92) with the alphas obtained empirically (p<.01).
We regressed the semantic similarity indices on the observed set of inter-item correlations, obtaining an adjusted R2 of .53 (p<.01). A second analysis applied a GLM model with information about scale belongingness, bringing the adjusted R2 to .68 (p<.01).
Saving the predicted values and residuals from the regression equations,
Main construct relationships | Scale relationship | Average observed correlations | Average correlations predicted from linear regression | GLM predicted correlations |
Transformat. leadersh.→Economic exchg. | −.10 | −.07 | −.10 | |
Transformat. leadersh.→Intrinsic motiv. | .18 | .15 | .18 | |
Transformat. leadersh.→Social exchg. | .15 | .11 | .15 | |
Transactional leadersh.→Economic exchg. | .01 | .01 | .01 | |
Transactional leadersh.→Intrinsic motiv. | .03 | .08 | .03 | |
Transactional leadersh.→Social exchg. | .05 | .06 | .05 | |
Laissez-faire→Economic exchg. | .11 | .17 | .11 | |
Laissez-faire→Intrinsic motiv. | −.11 | −.07 | −.11 | |
Laissez-faire→Social exchg. | −.07 | −.03 | −.07 | |
Intrinsic motiv.→OCB | .20 | .24 | .20 | |
Intrinsic motiv.→Turnover int. | −.22 | −.16 | −.22 | |
Intrinsic motiv.→Work effort | .26 | .24 | .26 | |
Intrinsic motiv.→Work quality | .21 | .22 | .21 | |
Social exchg.→OCB | .12 | .18 | .12 | |
Social exchg.→Turnover intent. | −.14 | −.08 | −.14 | |
Social exchg.→Work effort | .13 | .15 | .13 | |
Social exchg.→Work quality | .05 | .16 | .05 | |
Economic exchg.→OCB | −.15 | −.19 | −.15 | |
Economic exchg.→Turnover int. | .13 | .23 | .13 | |
Economic exchg.→Work effort | −.17 | −.15 | −.17 | |
Economic exchg.→Work quality | −.09 | −.14 | −.09 | |
Transformat. leadersh.→OCB | .10 | .16 | .10 | |
Transformat. leadersh.→Turnover int. | −.16 | −.07 | −.16 | |
Transformat. leadersh.→Work effort | .09 | .15 | .09 | |
Transformat. leadersh.→Work quality | .07 | .16 | .07 | |
Transactional leadersh.→Turnover int. | .05 | .08 | .05 | |
Transactional leadersh.→Turnover int. | −.07 | .02 | −.07 | |
Transactional leadersh.→Work effort | .06 | .08 | .06 | |
Transactional leadersh.→Work quality | .07 | .08 | .07 | |
Laissez-faire→OCB | −.01 | −.09 | −.01 | |
Laissez-faire→Turnover int. | .11 | .16 | .11 | |
Laissez-faire→Work effort | −.03 | −.09 | −.03 | |
Laissez-faire→Work quality | .01 | −.08 | .01 |
Observed correlations and values obtained through semantic analysis.
To explore how semantic values can explain claims of ‘mediation’ among survey variables, we ran hierarchical regression analyses with the organisational-outcome variables as dependent variables: Transformational leadership as independent in Step 1, and intrinsic motivation as independent in Step 2. Satisfying criteria for mediation, transformational leadership significantly predicted work effort, work quality, OCB, and TI in the first step (p<.01), but these relationships were rendered insignificant when intrinsic motivation was added in the analysis.
Variables | Transf. Leadership | Intrinsic motivation | Mediated by intrinsic motivation: |
Intrinsic motiv. | .32 |
||
Work effort | .17 |
.42 |
Fully |
Work quality | .13 |
.33 |
Fully |
Org. Citizen. Behav. | .19 |
.33 |
Fully |
Turnover Intention | −.30 |
−.35 |
Partly |
**Correlation is significant at the .01 level (2-tailed).
*Correlation is significant at the .05 level (2-tailed).
This dataset contains 16 individual scales. Ideally, a CFA should identify all of them with good fit. That did not happen, as a CFA for 16 factors returned a CFI of .82, a RMSEA of .05 and an SRMR of .07. If one allows the MLQ to be left out, the indices improve to a marginal fit (CFI = .89, RMSEA = .06, SRMR = .06), indicating some cross-loadings among items
This study compared semantic indices with responses to a broad range of leadership and motivation scales. Transformational leadership was measured with the 20 items from the MLQ
We used eight items of affective organisational commitment published by Meyer, Allen, and Smith
Alphas from observed correlations ranged from .86–1.00, mean = .94. For semantic alphas, these numbers were .84–.99, mean = .94 and for the predicted correlations .71–.99, mean = .89. These latter values correlated (.88 and .72) with the alphas obtained through the survey (p<.01).
Regressing the semantic similarity indices on the observed correlations, we obtained an adjusted R2 of.47 (p<.01). When scale belongingness was entered as a fixed factor in GLM, the adjusted R2 was .87 (p<.01). The predicted values and residuals were saved and are displayed in
Main construct relationships | Scale pairs | Mean observed correlations | Predicted correlations in lin. regression | Correlations predicted in GLM |
Consideration → Consideration | .55 | .29 | .55 | |
Consideration → Initiate struct. | .23 | .32 | .23 | |
Consideration → LMX | .47 | .28 | .47 | |
Consideration → Transform. lead. | .47 | .29 | .47 | |
Initiate struct. → Initiate struct. | .33 | .34 | .33 | |
Initiate struct. → LMX | .27 | .29 | .27 | |
Initiate struct.→Transform. lead. | .34 | .31 | .34 | |
LMX→LMX | .63 | .37 | .63 | |
LMX → Transform. lead. | .47 | .27 | .47 | |
Transform. lead. →Transform. lead. | .56 | .29 | .56 | |
Consideration → Affective comm. | .21 | .34 | .21 | |
Initiate struct. → Affective comm. | .13 | .35 | .13 | |
LMX- > Affective comm. | .20 | .31 | .20 | |
Transform. lead. → Affective comm. | .22 | .30 | .22 | |
Consideration → Job sat. | .36 | .31 | .36 | |
Initiate struct. → Job sat. | .19 | .34 | .19 | |
LMX → Job sat. | .33 | .32 | .33 | |
Transform. lead. → Job sat. | .32 | .30 | .32 | |
Consideration → Turnover int. | −.26 | −.17 | −.26 | |
Consideration → Work effort | .16 | .31 | .16 | |
Consideration → Work quality | .11 | .32 | .11 | |
Initiate struct. → Turnover int. | −.13 | −.20 | −.13 | |
Initiate struct. → Work effort | .12 | .35 | .12 | |
Initiate struct.→ Work quality | .10 | .34 | .10 | |
LMX → Turnover int. | −.24 | −.17 | −.24 | |
LMX → Work effort | .15 | .32 | .15 | |
LMX → Work quality | .14 | .31 | .14 | |
Transform. lead. → Turnover int. | −.23 | −.16 | −.23 | |
Transform. lead. → Work effort | .17 | .30 | .17 | |
Transform. lead.→ Work quality | .14 | .32 | .14 | |
Affective comm. → Affective comm. | .43 | .40 | .43 | |
Affective comm. → Job sat. | .40 | .34 | .40 | |
Job sat. → Job sat. | .68 | .45 | .68 | |
Affective comm. → Turnover int. | −.37 | −.20 | −.37 | |
Affective comm. → Work effort | .22 | .33 | .22 | |
Affective comm. → Work quality | .14 | .34 | .14 | |
Job sat. → Turnover int. | −.49 | −.22 | −.49 | |
Job sat. → Work effort | .31 | .38 | .31 | |
Job sat. → Work quality | .17 | .36 | .17 | |
Turnover int. → Turnover int. | .62 | .38 | .62 | |
Turnover int. → Work effort | −.15 | −.22 | −.15 | |
Turnover int. → Work quality | −.08 | −.22 | −.08 | |
Work effort → Work effort | .54 | .42 | .54 | |
Work effort → Work quality | .35 | .36 | .35 | |
Work quality → Work quality | .48 | .41 | .48 | |
The data from a CFA of this full dataset displayed the following values: CFI was .85, the RMSEA was .06 and the SRMR was .06. The considerable cross-loadings were expected due to the conceptual overlap of many of the scales and individual items included in this study.
We used an officially translated, Norwegian version of a commonly used five-factor model inventory called the NEO-FFI (the name NEO stems from the three first factors, ‘neuroticism’, ‘extraversion’ and ‘openness’)
The alphas of the NEO-FFI ranged from .94 to .96, which are considered excellent
Regressing the semantic similarity indices on the empirically obtained correlation values as the dependent variable, we found an R2 of only .004 (p<.05). The model as such reached only marginal significance. The saved and predicted values from the regression did not produce any recognisable patterns, see
Scale | Observed correlations | Correlations predicted in linear regression |
A→A | .18 | .05 |
C→A | .05 | .04 |
C→C | .28 | .05 |
E→A | .04 | .04 |
E→C | .09 | .05 |
E→E | .23 | .05 |
N→A | −.02 | .04 |
N→C | −.10 | .05 |
N→E | −.10 | .05 |
N→N | .21 | .05 |
O→A | .02 | .05 |
O→C | .03 | .05 |
O→E | .05 | .05 |
O→N | .00 | .04 |
O→O | .20 | .05 |
The scree plot from an exploratory factor analysis of the survey data indicated the usual five factors very clearly in the sample responses (see
Applying the text algorithms LSA and MI to a wide range of survey scales commonly used in management research, we were able to significantly explain the major part of variation in surveys. The correlations we predicted in multiple regression were similar to those created by human respondents. Allowing the algorithms to ‘know’ the context in GLM, we actually obtained correlations identical to those of human subjects. We were able to show that semantic relations not only predict the intra-scale coherence measured by Cronbach's alpha, but also the observed correlation relationships and proposed ‘mediating’ relationships among the variables. The factor-analytical fit indices were generally better the more semantics seemed to determine the correlation matrix. In this sense, CFA did not detect or prevent the pervasive influence of semantic commonalities. In fact, our results indicate that constructing a survey on mere semantic relationships among the items is an easy way to obtain good fit indices in CFA.
The personality test results in our study were not significantly explained by semantics. We expected this, since personality test scores are constructed to vary more freely, but still reflect the underlying construct and allow differentiated descriptions of people. This also shows that the long-proposed “lexical hypothesis”
Psychometric principles for construct validation seem, at least in their present form, as frequently applied in organisational psychology, to need revision to incorporate our findings. The semantic properties seem to pervade survey responses throughout many parts of the data analysis from the alpha coefficients to the CFA. This represents a fundamental problem to the understanding of psychometric principles in scientific research. Our study shows that the relationship between independent and dependent variables may be semantically determined a priori to conducting the survey since it follows from the wording of the items. This is in accordance with the previous theoretical analysis of van Knippenberg and Sitkin
At present, it is difficult to assess the pervasiveness of the problems we have detected here. In our study, all the commonly used measures from the field of OB were substantially affected by semantics, whereas the personality test showed very little influence. It is possible that some social scientific concepts are more abstract than others, thus being rendered more vulnerable to mere semantic relationships. It has been known for some years that common method variance usually leads to more inflated statistics in this field than in other fields
But ultimately, the only way of ruling out semantic influences as a major source of co-variation in survey data is to identify this influence in advance. Relationships among surveyed variables are commonly tested with 0-hypothesis statistics, implying the expectations that survey items are randomly related. Our findings instead suggest that all items are likely to be related though semantic commonalities. Perhaps replacing the 0-hypothesis with the semantic hypothesis is a more solid way of separating the empirical information from a merely semantic relationship in surveys.
Our findings have the following major implications:
Technologies for digital text analysis have advanced to a point of offering important and interesting usages to the social sciences. Text algorithms and similar procedures play an important role in indexing, storing and developing knowledge in the social sciences. Such knowledge is already in use for industrial purposes and we are taking it some steps further into the field of psychological research, such as OB. We believe this is now emerging as a promising field with many possible applications in our increasingly digitalised scientific society.
The fact that surveys seem predictable before the questioning of real subjects seems bewildering to many. And yet, our data show beyond reasonable doubt that this is possible. This opens up opportunities for experimental research on people's ability to understand logical propositions and our capacity to differ between logical and empirical statements. As shown in research on cognition, advanced forms of thinking in humans is an energy-consuming and partially unpleasant activity
Cross-cultural research using surveys needs to be re-examined in view of the present findings. Our findings were obtained using semantic similarity indices computed in American English, regressed on scores obtained from surveys in Norwegian. As long as the results are explained by semantics, all we can know about survey results being similar across cultures was that the survey was correctly translated. This pertains directly to the relationship among logical and empirical propositions: While the same propositions may be stated in different languages, their empirical implications in terms of behaviour dynamics may not be the same.
We are grateful to Julia I. Lane and Randy Ross for their support and for the feedback from Russel S. Cropanzano, Maw Der Foo, Kishen Iyengar and Tom Hilton. We are grateful to Thore Egeland and faculty at the Norwegian Business School for invitations to present earlier versions of this work and for the feedback that followed, especially from Tor J. Larsen and Svein Andersen.