Ethical Orientation and Research Misconduct Among Business Researchers Under the Condition of Autonomy and Competition

Given the significant potential impact, researchers must refrain from research misconduct (Kornfeld, 2012; Lund, 2000). Because rules and regulations have failed to curb research misconduct effectively, researchers’ ethical orientation has become important in the debate on ethical research (Boden et al., 2009). We differentiate between the deontological and consequentialist ethical orientations, which are rooted in two distinct foundations of ethical reasoning (Reidenbach & Robin, 1990; Tanner et al., 2007).

The deontological ethical orientation in the tradition of Kant (2003, p. 1788) builds on autonomous moral reasoning that is guided by the fit of a specific action with a person’s maxims (McNair, 2000). When a deontological ethical orientation is the foundation of moral reasoning, motivations for action are based on reason alone (O’Neill, 1989). Accordingly, a deontological ethical orientation is free from external constraints and is thus autonomous (Galvin, 1999). Accordingly, the categorical imperative (Kant, 2003, p. 1788) determines what a researcher ought to do. The categorical imperative uses a thought experiment to assess the moral quality of the principles (maxims) that individuals follow when determining their intentions: If those maxims are ones that the individual might want to become a generally valid law, the intention derived from the categorical imperative is, therefore, “good” (Schneewind, 1992). In the aforementioned thought experiment, the individual may consider the needs of the whole community by claiming a fictitious general validity and by including other people’s ends in his or her deliberations (Paton, 1962). It is not the apparent alignment between the motivation and the rules that is important, but the good intention (Korsgaard, 2002; Tyler & Blader, 2005). To clarify, most researchers commit themselves to the maxim of not reporting only selected results, even when to do so would be beneficial to them; they would refrain from suppressing results because they do not want such conduct to become generally accepted practice. Actions that diverge from maxims cause unpleasant remorse (Shalvi et al., 2011) and, therefore, a deontological ethical orientation reduces research misconduct.

In practice, ethical considerations cannot be fully disentangled from the web of constraints (Schütz & Luckmann, 1973) in which researchers are embedded as members of the research community. Researchers with a consequentialist ethical orientation build their moral reasoning on the anticipated consequences of their actions, which are, in turn, determined by a web of constraints. Accordingly, their motivations are externally determined (see, the hypothetical imperative, Kant, 2003, p. 1788), for example, by legal regulations and codes of conduct that define what researchers ought to do (Luhmann, 1977) and which consequences they face if misconduct is proven. The defined consequences take effect within a community when the members accept the rule (Korsgaard, 1996) and control and sanction the misconduct of their peers (Eberl, 2004; Shalvi et al., 2011). For example, a generally accepted rule for empirical research within the research community would be that all study results should be reported, not a selection of them. Peers enforce this rule through control measures such as the peer-review process. Those researchers who break the rule face consequences, that can include the retraction of publications, the loss of their reputation, and their academic position. The anticipated negative consequences of breaches of the generally accepted rules of the game prevent researchers from acting in a manner that is contrary to the standards of the scientific community (Akaah, 1993). Accordingly, a consequentialist ethical orientation reduces research misconduct.

We argue that the relationship between researchers’ deontological ethical orientation and research misconduct is contingent on perceived autonomy. Researchers have considerable autonomy when conducting their research. However, different researchers are likely to perceive the degree of autonomy in their environment differently, leading to a variation in their behavioral responses to it.

The level of autonomy researchers perceive determines how strongly their deontological ethical orientations affect the likelihood of research misconduct. High levels of perceived autonomy permit researchers to decide what, how, when, and with whom they want to work (Rose, 2001; Sax et al., 2002), thus providing them with academic freedom (Bland et al., 2005). Zhang et al., (2017, p. 236) state that autonomy makes researchers “feel self-determined and free from external controls or constraints” (on workplace autonomy, see also, Deci et al., 1989; Fricker & Schonlau, 2002; Spreitzer, 1995). Maxims induce researchers with strong deontological postures to justify their behavior to themselves rather than to others (Douglas, 2003). The more strongly researchers feel themselves to be autonomous, the more they can draw on their internal ethical standards as they make decisions (McNair, 2000; White, 2009). Accordingly, if researchers, who root their ethical research in a deontological ethical orientation, perceive themselves to be highly autonomous, maxims will limit their research misconduct even more than their peers with less perceived autonomy in the workplace. We, therefore, formulate the following hypotheses:

While the perception of autonomy in the workplace reinforces the independence of researchers’ moral reasoning from influence by external forces (Debackere et al., 1995) and, thus, extends their freedom to follow their maxims, perceived competition affects the consequences of research misconduct. However, deontological moral reasoning builds on maxims rather than on the consequences of actions. The ethical value of the research maxims on which researchers with a deontological ethical orientation base their ethical reasoning is assessed independently from the causal constraints they face in the context they are embedded in (Kant, 2003). Accordingly, the impact of researchers’ deontological ethical orientation on research misconduct should not be contingent on the level of competition in the workplace. Therefore, we have not attempted to formulate a corresponding hypothesis.

The research environment is characterized by intense competition around publication opportunities, funding, and career opportunities (Fang & Casadevall, 2011; Hall & Martin, 2019). However, the perception of the degree of competition in the workplace is subjective, and different perceptions lead to variations in the behavioral response to the perceived competition. We argue that for researchers who base their moral reasoning on the anticipated consequences of their actions, perceived competition will influence how effectively their consequentialist ethical orientation deters research misconduct. This is because the behavior of researchers is influenced by their peers (Salancik & Pfeffer, 1978), who define expectations as the standard of peer-group comparisons (Brown et al., 1998) as well as sanctions and rewards for individual achievements (Williamson & Cable, 2003).

In intensely competitive work environments, researchers abstain from research misconduct because they anticipate that their peers are more likely to identify and report a cheating colleague, and, thus, cheats are less likely to prosper from cheating (Fink et al., 2012). While we see the merit in this whistle-blower argument, we argue that it downplays a more immediate effect of competition on the link between researchers’ consequentialist ethical orientation and research misconduct. The presence of intense competition in the first place implies that any gains obtained by cheating would be of greater relevance than in a less intense competitive environment (Shi et al., 2016). Such gains might include an extra publication, a publication in a more prestigious outlet, or extra funding from sponsors or funding agencies. Accordingly, high levels of competition reinforce the positive consequences of research misconduct anticipated by consequentialist researchers. Even if their peers monitor and report them more strictly in intensely competitive settings, the immediate competitive advantage of gains from research misconduct should outweigh the threat of negative consequences. We argue that a certain immediate gain from research misconduct is a stronger motivational factor than possible detriment owing to sanctions in the longer term (Åkerlund et al., 2016). Hence, intense competition should weaken the limiting effect of a consequentialist ethical orientation. These arguments lead us to propose:

Moreover, we assume that the extent to which anticipated consequences restrict research misconduct is not only contingent on the level of competition but also on the level of autonomy researchers perceive they have. This is because researchers who perceive they have a great deal of autonomy view their peers as having relatively little power to determine any negative consequences of their research misconduct; essentially, the chance of being sanctioned for ethical misconduct becomes less likely the greater autonomy a researcher accrues (Debackere et al., 1995). At the same time, high levels of autonomy do not reduce the positive reputational and career effects consequentialist researchers derive from the extra publications or extra funding resulting from their undetected and unsanctioned ethical misconduct. Hence, in researchers’ consequentialist ethical reasoning, ethical misconduct becomes less threatening with increasing autonomy, while the positive consequences remain intact. Because as part of their consequentialist orientation, researchers base their moral reasoning on the anticipated consequences of their action, increasing autonomy should, thus, weaken the limiting effect of their consequentialist ethical orientation on research misconduct. We therefore postulate:

The sample informing this research is drawn from the worldwide population of management scholars, from doctoral students to full professors. We used contact details submitted to an e-mail contact list of active participants of the 2016 annual meeting of the AoM, the preeminent professional association for management scholars boasting a community of 20,000 scholars from 120 countries (AoM, 2021). Sampling among participants at an international science conference offers the opportunity to address a globally distributed population of a specific profession that has a specific research focus. That sampling approach has been successfully employed in empirical research, especially in the areas of health and medicine (Pezold et al., 2016; Zozus et al., 2017). Rigorous peer-review processes affecting all contributions to the AoM conference ensured that those involved were active research scholars who authored research presented at the conference. The individuals approached could thus be guaranteed to fall into our target group.

After a notification, we sent 10,716 e-mail invitations to all authors of papers presented at the 2016 AoM annual meeting, followed by two rounds of reminders seven and 14 days after the initial mailing. A web link pointed to a landing page where the participants could choose between a classic online survey or an interactive gamified survey. The gamified survey aimed to stir curiosity and thus enhance the response rate. The gamified survey was also offered an opportunity to check for social desirability bias (Georgiou et al., 2019; Stodel, 2015). Both survey versions contained the same formulations, sequence, and scales. Of the 2533 recipients who clicked the invitation link, 1684 completed the survey. We removed responses with missing values on the dependent variable and obtained a sample of 1031 individual responses to analyze. This response represents an effective response rate of 9.6%, which is slightly below the average of similar online survey studies (Pedersen & Nielsen, 2016; Petrovčič et al., 2016; Wouters et al., 2014). However, survey response rates have been decreasing significantly since 1986 (Sheehan, 2001) and are generally lower for web-based survey instruments (Blumenberg & Barros, 2018); Manfreda et al. (2008), for example, found an average response rate of 11% and a 6–15% confidence interval for online surveys. Lower response rates were recorded by studies that either randomly addressed members of professional associations (Poynton et al., 2019), used detailed questionnaires (Sauermann & Roach, 2013), or conducted studies on sensitive topics (Trautner et al., 2020).

We addressed potential nonresponse bias at the survey design stage by carefully designing the questionnaire to maintain the respondent’s interest, keeping it to a reasonable length, and establishing the importance of the study in the introductory e-mail (Yu & Cooper, 1983). We also assessed the analytic sample for potential nonresponse bias using two techniques to target specific types of nonresponse (Rogelberg & Stanton, 2007). First, we implemented the archival approach that compares the characteristics of the sample with those of the population. This approach is especially suited for identifying passive nonresponse, which results from external factors that keep recipients from returning the questionnaire on time. Passive nonresponse typically accounts for 85% of total nonresponse (Sosdian & Sharp, 1980). For our sample, we used the respondents’ age for archival analysis, as this variable has been shown to be relevant to research misconduct (Kelley et al., 1990) and was available for the members of the AoM. The comparison identified only a minor under-sampling of older researchers. Accordingly, passive nonresponse does not seem to be a significant concern. Second, we applied wave analysis, which involves comparing the results from early and late respondents, which is especially useful for controlling active nonresponse, which refers to conscious decisions not to participate in a study (Rogelberg et al., 2003). For our sample, the wave analysis did not identify any significant differences between early (first half) and late (second half) respondents, and hence we conclude that active nonresponse bias is also unlikely to be an issue in our sample.

Another potential threat to the validity of self-reported data in surveys on sensitive topics such as ethics is social desirability bias (Chung & Monroe, 2003; Cohen & Pant, 1998; Krumpal, 2013; Randall & Gibson, 1990). Social desirability bias emerges when respondents questioned about their compliance with widely accepted behavioral norms over-report behavior that is in line with those norms and under-report deviations from them (Podsakoff et al., 2013). Because the effectiveness of social desirability scales or single items in directly measuring and correcting such bias is heavily debated, and such strategies may even introduce a systematic error (Fisher, 1993; Kam, 2013; Larson, 2019; Paulhus, 2002), we decided to implement a mix of measures in the design of the data collection that has been found to mitigate the threat of social desirability bias. Notifying all meeting participants that they would be invited to participate in a survey (DeLeeuw, 2018), the anonymity guaranteed to all respondents, and the avoidance of personal contact by opting for a self-administered web-based survey (Hunter, 2012; Moy & Murphy, 2016; Richman et al., 1999; Tourangeau & Yan, 2007) should have mitigated social desirability bias (Couper, 2017). Additionally, for those respondents who opted for the gamified version of the survey, the perception of the power relation between researcher and respondent (Denzin, 1989) and the feeling of being evaluated (Armstrong et al., 2016) should have been further ameliorated because playing a game creates a cognitive load that reduces respondents’ concern over social desirability. Accordingly, gamified surveys are less prone to faking and distortion (Georgiou et al., 2019; Stodel, 2015). Comparing the values of the dependent and independent variables collected via the traditional survey to those collected via the gamified survey revealed no significant differences. Because any social desirability bias would be stronger in the subsample collected via the traditional web survey than in the subsample collected via gamified survey, the lack of significant differences is an indication that social desirability bias is not an issue in our sample.

Before computing index scores of the multi-item scales for the subsequent regression analysis, we examined their discriminant validity using confirmatory factor analysis. Specifically, we compared a specification where all items load on their theoretically intended factors to ones where two sets of items load on a single factor, while the other two load on their own factors, and a specification where all items load on a single factor. In each case, the theoretically intended model resulted in a superior fit with the data. Indeed, the model where all items load on their intended factors shows a good fit with the data (Hu & Bentler, 1999): the comparative fit index (CFI) score is 0.976 (recommended minimum 0.95), the standardized root mean squared residual (SRMR) index value is 0.035 (recommended maximum 0.08). The root mean squared error of approximation (RMSEA) is 0.040 (recommended maximum 0.06). Furthermore, the model’s average variance extracted (AVE) scores exceed the squared correlations among the latent variables, which provides further evidence for satisfactory discriminant validity (Fornell & Larcker, 1981).

Scholars have highlighted the threat of common method bias (CMB) to empirical research relying on cross-sectional data (Lindell & Whitney, 2001), and especially to the most common self-report surveys with cognitions as dependent and independent variables (Harrison et al., 1996). Several ex-ante and ex-post measures can be applied to address the risk of CMB (Podsakoff et al., 2003). However, Spector (2006) and Richardson et al. (2009) have provided substantial evidence showing that ex-post statistical measures to adjust analyses for CMB are unreliable and often misleading. Accordingly, we employed the recommended strategies to avoid CMB ex-ante. First, we protected the respondents’ anonymity to mitigate evaluation apprehension (Podsakoff et al., 2003). Second, we used different question formats and randomized the order of scales in the questionnaire.

Finally, we checked the data for potential bias due to this study's survey strategy. A comparison on the descriptive level of the data collected with the gamified and traditional surveys shows that while the gamified survey attracted a slightly younger set of respondents and one with slightly more women, only the level of autonomy is significantly higher in the sample drawn from the traditional survey. The response rate did not differ between the two types of the survey instrument. Accordingly, bias stemming from the survey strategy does not appear to affect our study.

The ordinal nature of the dependent variable prompted us to choose ordinal logit regression as the most appropriate statistical modeling technique. We compared the Akaike and Bayesian information criteria between an ordinal and a linear regression model. The criteria favored the ordinal specification. We examined the models for multicollinearity, influential observations, and the parallel regression assumption that underlies ordinal regression models before performing the final estimations.

The mean variance inflation factor (VIF) score was 1.40, and the highest VIF, which pertained to the category Ph.D. student in the academic career stage variable, was 3.01. These are below the conventional threshold of 10 for multicollinearity.

We analyzed potentially influential observations by computing residuals, leverages, and Cook’s distance statistics, as well as examining plots of squared residuals and leverages. We identified 14 observations that had potentially influenced the model estimates and excluded them from the final analysis sample, which thus became 1031 observations.

The parallel regression assumption that underlies ordinal regression models maintains that the relationship between each pair of outcome groups is the same; that is, a single set of coefficients applies to all outcome groups. If the model violates this assumption, a generalized model where one or more coefficients vary between different outcome groups must be estimated. To test this assumption, we first compared a model specification where the parallel-lines constraint is imposed on all variables with a model where the constraint is relaxed for all variables. We found that the parallel regression assumption did not hold for our model. We then followed an iterative process of relaxing the parallel-lines constraint for one variable at a time (Williams, 2006) to determine which variables violate the assumption. These tests revealed that the coefficients of both independent variables and two control variables (promoted in the last three years and the assistant professor category in the academic career stage) need to be estimated separately for each pair of outcomes. Therefore, the final regression models use a generalized ordinal logit specification where we apply the parallel regression constraint to all variables except for the four mentioned above.

Table 1 shows the means, standard deviations, minima, and maxima for all variables used in the analysis. Table 2 displays the correlation matrix. The dependent variable and two control variables (research funding and current position) are ordinal rather than continuous, so we report Spearman’s rank-order correlations. We do not detect any unexpected correlations (Fig. 1).