Supervisor Interpersonal Justice and Worker Turnover Intention in Global Supply Chains: Evidence from Bangladesh Garment Factories

Supervisor Interpersonal Justice and Worker Turnover Intention

Interpersonal justice refers to the perception that authority figures such as supervisors treat employees with respect and propriety (Bies and Moag 1986; Colquitt 2001). It typically includes treating subordinates in a polite manner, with dignity, respect, and refraining from improper remarks and comments (Colquitt 2001). Although individuals’ interpersonal justice orientation may vary to a certain degree, justice is a fundamental human motive and “one of the most sacred and pervasive themes in social behavior” (Lerner 1975: 1; Lerner and Clayton 2011). As a fundamental and sometimes subconscious motive, people may use and respond to justice in a social context with automatic reactions as well as with thoughtful controlled reactions (Lerner and Clayton 2011). Focusing on the more automatic reaction to justice akin to Kahneman’s (2011)“System 1” fast, intuitive cognition, we propose a direct effect from SIJ to worker turnover intention based on the assumption that workers will have an intuitive attraction to supervisor justice. Workers may also evaluate SIJ in the light of the factory context, in line with Kahneman’s (2011)“System 2” slow, complex thinking. Such reasoned responses to SIJ via workers’ positive affect and how labor rights institutions shape workers’ attributions and emotional reactions are elaborated in Hypotheses (2)–(4).

Consistent with our expectation that workers may be intuitively attracted by SIJ and less inclined to leave the factory, Özkan’s (2022) meta-analysis showed a negative relationship (r = −0.361) between interpersonal justice—which is often enacted by supervisors—and turnover intention. We maintain that this negative relationship will also hold in the GSC context because workers at export factories also react strongly to disrespect or abuse from supervisors (Pike and Godfrey 2012: 12; Ahmed and Uddin 2022). Supervisor abuse may influence their job choices, as shared by one Bangladeshi garment worker (Akhter et al. 2019: 6): “I changed two other factories because the supervisors of the factories do not behave with the workers nicely. I heard that this factory is good . . . so I joined.” Relatedly, Seo and Chung (2019) found that abusive supervision was linked with higher worker turnover intention among Chinese factory workers. Therefore, we expect justice to invert this relationship and lower such intentions.

We do acknowledge that SIJ, especially in a GSC context, does not impact workers in a vacuum. Many factors may affect SIJ and worker turnover. These include stringent MSIs and collective worker voice mechanisms that may improve compliance with labor standards in some cases (Oka 2016; Kabeer et al. 2020; Pike 2020), which in turn relate to worker turnover rates (Li and Kuruvilla 2023). However, such regulatory institutions may not determine SIJ for two reasons. Conceptually, although labor rights institutions may include standards or procedures to prevent adverse supervisor behaviors, as transaction cost theory (Williamson 1981) suggests, everyday control of specific supervisor behaviors is difficult to achieve. Supervisor enactment of interpersonal justice in their daily work—which also entails broader normative/moral principles such as respect—varies and may have independent effects on workers beyond institutional or contractual prescriptions. Empirically, most research suggests that various private regulatory institutions did not reduce abuse of workers (Barrientos and Smith 2007; Kabeer et al. 2020: 1387; Ahmed and Uddin 2022; for an exception see Pike 2020). Anner (2019) even suggested that the global brands’ conflicting demands of social compliance and predatory purchasing practices contribute to more abuse of workers.

Overall, owing to its interpersonal and processual nature we posit that SIJ has a unique impact on turnover intentions not captured by other institutions (i.e., even in a well-regulated environment, supervisor conduct varies in ways that affect workers). Indeed, in some cases these institutional factors are not necessarily the top priority in workers’ turnover decisions (Li and Kuruvilla 2023). We thus hypothesize that:

Indirect Effect of Supervisor Interpersonal Justice through Positive Affect

One major pathway ¹ through which SIJ may influence workers’ intentions and behaviors is “affect” (Colquitt et al. 2013). Interpersonal justice often leads to positive feelings such as positive affect or pleasantness in recipients (Colquitt et al. 2013), which in turn influence worker intentions and behaviors. SIJ may meet some important human needs such as self-esteem and self-determination (Cropanzano, Byrne, Bobocel, and Rupp 2001; Decker and Van Quaquebeke 2015), thus leading to higher positive affect as shown in Colquitt et al.’s (2013: 212) meta-analysis. This finding is also congruent with research from a resources lens showing how supervisors provide subordinates with psychological resources or help them cope with work-related stressors, resulting in higher positive affect (e.g., Jasiński and Derbis 2023).

We propose that SIJ may improve factory workers’ positive affect at work, including feelings such as enthusiasm and pride. Emotional reactions to interpersonal (in)justice also feature in workers’ accounts in the GSC context. For instance, workers often describe negative feelings toward supervisor verbal abuse such as anger (Kabeer et al. 2020: 1385) or feelings of being “humiliated, helpless, unworthy and unaccepted at work” (Ahmed and Uddin 2022: 546). Describing positive/negative feelings toward justice/injustice from supervisors, another worker shared, “If they [supervisors] talk and behave with us well . . . we become happy after working hard. When they are rude and harsh, we become upset and feel threatened” (Akhter et al. 2019: 6). We thus posit that:

Building on this, positive affect may in turn influence workers’ turnover intentions. The broader turnover literature suggests that affect can explain employee turnover through a hedonistic approach of seeking desirable or avoiding undesirable situations (Maertz and Griffeth 2004). In the GSC context, Seo and Chung’s (2019) study on Chinese factory workers found that psychological emotional states mediate the relationship between abusive supervision and turnover intention. We thus hypothesize the following:

Synergy between Supervisor Interpersonal Justice and Workplace Collective Voice Mechanism

The organizational context may shape the effect of the overall level of SIJ by influencing how supervisors enact justice and how workers attribute the reasons for supervisor justice. Specifically, when justice such as fair treatment is delivered in a way sensitive to the personal situation of the recipient (e.g., by mentioning the latter’s name), this personal sensitivity can accentuate positive reactions from the recipient (e.g., Ployhart, Ryan, and Bennett 1999; Margolis and Molinsky 2008). In this regard, supervisors can benefit from feedback on workers’ concerns and needs collected from formal collective voice mechanisms at the workplace. As many worker concerns relate to supervisors (Pike 2020), workers may feel freer and safer to communicate their issues and needs via collective voice platforms such as workers’ committees or unions rather than through individual confrontation with supervisors (Freeman and Medoff 1984). Supervisors can then draw on such worker voice and feedback to attend to the particular needs of workers under their supervision. Indeed, Babbitt and Brown (2018) showed that supervisors at garment factories who are exposed to workers’ views exhibit a higher level of perspective-taking and humanization of workers, which reduces worker abuse. Absent collective voice, supervisors might act out their personal (often negative) perceptions of workers (Akhter et al. 2019). Informed supervisor treatment is likely to create more positive feelings among workers, thereby increasing retention.

Further, workers may make more positive attributions of informed supervisor treatment in lieu of collective voice mechanisms. Workers may think that their supervisors responded to their collective voice and thus respect them in the way they desired. Here, the effects on workers are a combination of SIJ per se—as in a context without unions/committees—plus the value of workers’ perception that their voice was heard and implemented. Indeed, voice is an important element of procedural justice (Thibaut and Walker 1975), and procedural justice has been found to strengthen the effect of interpersonal justice (Huang and Huang 2016). In the GSC context, Adhvaryu, Molina, and Nyshadham’s (2019) study found that the Indian garment workers who had the opportunity to express voice via a survey showed lower turnover rates and absenteeism. Arguably, workers’ perception of their voice being responded to by supervisors could have stronger effects on worker affect and behavior.

We note that sensitive SIJ and favorable worker attributions are more likely to occur when collective voice mechanisms function well. Such synergy may diminish with constrained voice mechanisms (e.g., Anner 2018). Although the actual synergy may vary across contexts, we propose that, on average, a stronger SIJ-positive relationship is more likely to materialize under collective voice mechanisms.

Synergy between Supervisor Interpersonal Justice and External Stringent Monitoring Programs

Although the organizational justice literature has not yet examined how institutions outside the organization may shape the effects of organizational justice (Colquitt et al. 2023), we suggest that monitoring programs in the environment may influence justice enactment and worker attributions as well. Pertaining to GSCs, we propose that stringent labor rights monitoring programs may amplify the effects of SIJ on workers by facilitating more stable enactment of fair treatment. Justice variability—for example, being polite one day but rude the next day—creates uncertainty and thus increases employee stress (Matta et al. 2017) and can negate the benefits of an overall positive level of interpersonal justice (Matta, Scott, Guo, and Matusik 2020).

External monitoring programs—whether led by individual brands or MSIs (Ashwin et al. 2020)—differ in their stringency (Rodríguez-Garavito 2005; Fransen and Burgoon 2012). The more stringent ones (e.g., FairWear and Better Work) have the following three attributes: comprehensive labor standards, strong enforcement mechanisms that include worker complaint channels, and a degree of control by societal actors such as involving unions at the global and/or factory level (Fransen and Burgoon 2012). By contrast, permissive MSIs including the amfori Business Social Compliance Initiative (BSCI) and the Worldwide Responsible Apparel Production Certification Program (WRAP) lack one or all of the three stringent elements. The effects of stringent MSIs often emanate from both the standards and enforcement of the MSI per se as well as the private regulation programs of individual reputation-conscious buyers who joined the MSI (Oka 2010).

External stringent monitoring programs may create a more conducive environment for supervisors to show a consistent level of propriety and respect to workers. First, such programs may prompt factory management to address some factory-wide issues such as health and safety standards and minimum wages (e.g., Barrientos and Smith 2007; Kabeer et al. 2020). Such improvements may save supervisors’ mental and psychological energy, enabling them to enact more consistent fair treatment of workers. Managers are less capable of enacting fair treatment when they are themselves depleted (Whiteside and Barclay 2018) or have heavy workloads (Sherf, Venkataramani, and Gajendran 2019).

External stringent monitoring programs may also contribute to workers’perception that they have some power via these external institutions to influence supervisors’ behavior. This potential empowerment may emanate from stringent programs’ complaint mechanisms (e.g., worker hotlines) through which workers can file their grievances to internal voice mechanisms or external parties (Fransen and Burgoon 2012). For example, the Accord on Fire and Building Safety in Bangladesh and the Alliance for Bangladesh Workers’ Safety (hereafter Accord/Alliance) both established worker complaint channels. Although Accord/Alliance focused on safety issues, their complaint mechanisms received worker reports on many other issues (e.g., 79% of worker calls to Alliance’s helpline were “non-safety” issues; Donaghey and Reinecke 2018: 32), perhaps because Bangladeshi workers perceived global buyers and institutions “as the primary guardian of workers’ welfare and rights” (Kabeer et al. 2020: 1387).

Workers’ perceived power may also come from the involvement of union or worker representatives in stringent MSIs. For instance, both the Accord and Alliance gave union or worker representatives the right to be present during safety inspections, which signaled symbolic power to workers who often “enjoy little respect” but saw union leaders’ co-presence with inspectors as “a powerful demonstration that workers were eligible to speak on the same level with outside authorities and managers” (Donaghey and Reinecke 2018: 29). As grassroots workers may have very limited knowledge of specific standards in the codes of MSI and related brands (Graz, Piazza, and Walter 2022), they may perceive stringent programs and associated brand buyers, regardless of single issue-focused programs (e.g., Accord/Alliance for safety) or comprehensive ones (e.g., Better Work), as potential external allies against various grievances, including supervisor behaviors.

Such a personal sense of power—a perceived ability to influence other people in specific contexts (Anderson, John, and Keltner 2012: 316)—is related with more positive emotions (Anderson and Berdahl 2002) and can amplify the effects of justice especially among workers in low hierarchical positions (van Dijke, De Cremer, Langendijk, and Anderson 2018). That is, the effects of SIJ in factories subject to external stringent monitoring programs may comprise SIJ per se and associated worker perceptions of increased power (from external source) to influence SIJ.

Admittedly, the abovementioned more stable justice enactment and positive worker attribution are contingent on functioning and visible monitoring programs in the environment. Workers may be unaware of some codes or MSIs (Graz et al. 2022) and may sometimes collude with managers to hoodwink auditors (Hoang and Jones 2012). On average, and especially for stringent and visible monitoring programs, we hypothesize a synergy:

Overall, we have hypothesized a direct relationship between SIJ and turnover intention and an indirect effect via positive affect. Workplace voice mechanisms and external stringent monitoring programs may amplify such effects. Figure 1 graphically summarizes our hypotheses.

OPEN IN VIEWER

Figure 1.

Proposed Model

Data

We drew on a survey of 1,500 garment workers in Bangladesh in 2017 collected by Naila Kabeer and colleagues (2020). The Bangladeshi garment industry provides a suitable case to examine our hypotheses for two reasons. First, several studies documented severe abuse of Bangladeshi garment workers (Akhter et al. 2019; Kabeer et al. 2020; Ahmed and Uddin 2022) but the take-up of supervisor training is low (Macchiavello et al. 2015). Demonstrating the benefits of SIJ may thus provide support for positive intervention/supervisor training or selection to address a prevalent issue in this context. Second, Bangladesh hosts many visible labor rights institutions providing an opportunity to examine their interaction with supervisor behavior. As regards workplace collective voice mechanisms, the Bangladeshi 2006 Labor Act required Worker Participation Committees (WPCs) of managers and workers—for social dialogue—in factories with 50 or more workers, and the 2013 amended Act requires open election of worker representatives to WPCs (Kabeer et al. 2020: 1381–82). Further, after the Rana Plaza tragedy that killed 1,134 garment workers in 2013, two stringent monitoring schemes were established: the Accord on Fire and Building Safety in Bangladesh, and the Alliance for Bangladesh Workers’ Safety. Unlike other schemes such as Better Work, Accord/Alliance does not mandate WPCs or promote supervisor training, making supervisor treatment relatively independent of these two institutions and an appropriate case to explore their interactions.

The workers in our study were sampled from five industrial locations within the Dhaka area for their high concentration of garment workers: 200 women and 100 men were randomly sampled from each neighborhood, reflecting the higher percentage of women in the garment sector (Kabeer et al. 2020: 1367). Local researchers visited workers in their residences and filled in workers’ responses to ensure the quality of the answers. Our final sample comprises 1,377 workers from 107 factories after deleting observations not suitable for our analysis (explained below).

Measures and Items

Supervisor interpersonal justice was measured by three items. Two items overlap with seminal statements on interpersonal justice capturing respect/politeness and proper language/no improper remarks (Bies and Moag 1986; Colquitt 2001). Those items are: “My supervisor respects me, is polite with me” (reflecting “respect”); “My supervisor does not use bad language with me or make indecent comments” (reflecting “propriety”). As justice rules may evolve with time and/or context and because an important reason for supervisor verbal abuse of garment workers is mistakes or falling behind targets (Ahmed and Uddin 2022), we also added a third item based on later research that found supervisor support (with tasks or defending the employee against others in the firm) is an important justice rule in how employees assess supervisor justice (Hollensbe, Khazanchi, and Masterson 2008; Fortin et al. 2020: 1640). Specifically, our third item is: “If I unintentionally make a mistake, my supervisor will side with me.” These three items were measured on a 5-point scale ranging from strongly disagree (1) to strongly agree (5). Cronbach’s alpha for the three items is 0.85.

Positive affect was measured by three items from Zevon and Tellegen’s (1982) mood checklist of 60 adjectives: enthusiastic/eager, proud, and contented. The question asks the workers, “How do you usually feel in the workplace?” (3-point Likert scale). “Enthusiastic” and “proud” are among Watson, Clark, and Tellegen’s (1988) classic 10-item scale of positive affect; “contented” is among Zevon and Tellegen’s (1982) longer list, which is relevant in the wake of “constant pressure” among Bangladeshi garment workers (Akhter et al. 2019: 5). Although the item on enthusiasm shows lower correlation with the other two items and has lower factor loading, we include all three items ² to maximize use of information. Cronbach’s alpha for this scale is 0.69.

Turnover intention was measured by a single item used by Hom, Griffeth, and Sellaro (1984: 148): “What are the chances of you leaving this factory within the next 12 months?” The choices coded from 1 to 5: will not go (1), strong change of staying, not sure, strong change of leaving, and going to leave (5).

The workplace collective voice mechanism was measured in two stages. Analytically we needed to determine if the factory had a WPC, based on workers’ perceptions. To that end, in our first stage we relied on a survey question that asked workers, “Is there a worker participation committee (WPC) in the factory?” Among the 107 factories with two or more workers sampled, the majority of workers answered “yes” (n = 1,121, 81.41%), followed by “I don’t know” (n = 146, 10.6%), and “no” (n = 110, 7.99%). We looked at the workers’ background information and their responses regarding WPC within each factory. Workers from 28 factories (10.7% of workers) unanimously said “yes” (coded as WPC = 1) and 5 factories (0.7%) unanimously “no” (coded as WPC = 0). We also observed that 72 factories (88% of workers) had mixed responses. Among these 72 factories, 33 (46.55%) were easy to score because the disagreeing respondent(s) said, “I don’t know” and we followed the consensus answer.

For the remaining 39 factories with inconsistent explicit worker perceptions of WPC presence, a second stage of coding was involved in which, methodologically, we used distinct individual perceptions to assign a factory-level characteristic. We followed the “key informants” methodology used by inter-organizational scholars. Specifically, we draw on the work of Kumar, Stern, and Anderson (1993) who acknowledged the necessity of relying on human informants in the absence of reliable archival data, and the challenges associated with multiple raters who diverge from each other. In such cases, the authors mentioned three approaches based on the literature. Two of the options involve some statistical approach, such as aggregation (e.g., using structural equation modeling, SEM), or a latent trait modeling (LTM) approach. The third is the consensus method, which uses the agreement of competent, key informants. We have preferred the latter option above the others because our analysis ideally requires a definite answer in terms of the presence (or absence) of WPC. To determine who the key informants are in each factory we needed to assess their competence. To that end, we examined 1) the workers’ demographic information, including their job designations, seniority, and years of schooling, along with 2) their knowledge of matters that a competent informant would know. Specifically, we paid attention to whether any workers affirmatively responded to knowing about buyer codes of conduct, local labor laws, and whether they voted in the WPC elections.

Following the key informants approach, it was relatively straightforward to operationalize our WPC variable for most of the 39 factories where respondents had conflicting views. ³ For such cases, Kumar et al. (1993) proposed a hybrid approach in their article that combined elements of aggregation, LTM, and consensus methods. But given that our variable is binary (yes/no) and is a measure of the presence or absence of the WPCs, we primarily relied on the consensus approach. However, we also paid attention to whether our key informants were the majority in a group—which they were in 23 out of the 39 factories (includes 91% workers in this subgroup). We identify approximately 16 factories (9% of workers within this subgroup) where the group with the key informant consensus had the same number of people as the opposite consensus (yes vs. no), mostly attributable to small sample sizes (n≤ 11). Regardless, in most cases within the 39 factories the consensus answer was very clear. For example, among 33 factories, the few disagreeing workers had lower job positions such as helper and/or a shorter tenure in the factory, whereas the majority group typically included some senior workers such as quality inspectors who often had more years in the factory and/or garments industry. We thus used the response of the majority workers with key informants. We found only six boundary cases (n = 2 to 4 workers, though one had n = 6) with equal or close number of “yes” or “no” answers on WPC from two or three workers at similar jobs and with similar seniority. We selected the answer of the worker(s) who indicated better knowledge of codes, laws, and whether they had voted in the WPC elections.

Finally, in two factories (n = 3 and 4 workers) everyone responded, “I don’t know” to the WPC question. We decided to extrapolate an answer rather than abandon them because these factories were still valid for measurement of other variables in our analysis. These two factories are relatively small (with respondents indicating 200 to 600 total workers) and some workers are near the legal minimum working age of 14 to 17. Such small factories were less likely to have a WPC. And one of these factories had an operator with 12 years in the garment industry and a quality inspector in our sample; even they responded “don’t know” regarding WPC presence, signaling its absence. We thus report the main results coding WPC = 0 for the two factories in which all workers did not know whether their factory had a WPC. Even if these two factories were incorrectly assigned a score (along with the boundary cases mentioned above), such errors are more likely to increase noise in the data and thus make our moderation analyses conservative.

On the whole, we acknowledge two important limitations. First, there is still a risk of making a mistake as even competent informants may not always agree with each other (Kumar et al. 1993). Second, although we are treating this variable as an objective measure at the factory level, it is still a proxy variable (albeit a highly sophisticated one) rooted in workers’ perceptions. However, such perceptions could be an important pathway that relates to how the factory-level WPC may shape workers’ attribution and reactions to SIJ, as we explained in H3.

External stringent monitoring program was captured by asking workers the name of their current factory. Kabeer et al. (2020) independently checked whether the factory’s name was on supplier factory lists disclosed online by Accord and Alliance. Although differences exist between Accord and Alliance—for example, legally binding vs. voluntary, and industrial democracy vs. corporate social responsibility—they both improved some safety issues with their own methods (Donaghey and Reinecke 2018). Our data set makes it difficult to tease out the effect of each because most factories covered by the Alliance are also in the Accord: 31 of the 107 factories (29%) are covered by both Accord and Alliance, 39 (36.4%) factories are covered by just Accord, and only 2 (1.9%) by just Alliance. Therefore, we aggregate the affiliation to either initiative in a single Accord/Alliance variable: 72 (67%) factories, or 87.7% of the workers, are covered by them whereas 35 other factories in our sample are outside their remit. This measure primarily captures the coverage of the Accord, which covers 70 factories jointly or alone.

Although the Accord and the Alliance also promoted safety committees, these differ from WPCs mandated by the law that cover more general issues. Of the 72 factories covered by Accord/Alliance, 63 also have WPCs, perhaps because these are large factories for which the Bangladeshi law mandated WPCs. Among the 35 factories outside Accord/Alliance, 13 have WPCs based on majority confirmation and the key informants rule. And among the 32 factories without WPCs, 9 were covered by Accord/Alliance. These results allow us to examine the moderation effect of each institution in the absence of the other in subsample analysis (see the Robustness Analysis section).

We also checked factory names against the Better Work Bangladesh (BWB) factory list in 2017 and the WRAP certified factory list in 2025 (the only publicly available lists at the time of writing). None of the factories not covered by Accord/Alliance were covered by these two programs. Among those covered by Accord/Alliance, 13 factories (18% workers) are also in BWB, and 12 factories (18% workers) were covered by WRAP. This suggests that factories that score 0 on our variable are poorly monitored, if not completely outside external private regulation. Therefore, scoring a 1 on our variable suggests being covered by some form(s) of stringent monitoring, especially the Accord.

Control variables include gender (coded 1 for woman), age (years), tenure in garments industry (years), job satisfaction, and monthly pay for their potential influence on turnover or turnover intention found in meta-analysis (Rubenstein et al. 2018). Job satisfaction was measured by a single item on a 3-point scale: “Considering everything, how satisfied are you with your job?” We also controlled for monthly pay (basic wage plus overtime pay and log transformed) because of the importance of income in influencing worker turnover (Li and Kuruvilla 2023).

We had two approaches controlling for the potental confounding effect of factory management systems that may influence both SIJ and worker turnover intention. First, we included variables on perceived management style. The survey asked workers to pick which one of the following statements best describe the relationship between management and workers in their factory: “Workers are treated like family members”; “The management regularly consults workers or their representatives to find out about workers’ problems”; “The factory has rules that all workers must be aware of and follow”; and “We are lucky to get the job; therefore, we must obey orders.” We labeled the first item (treated as family members, 14.3% workers) as a perceived paternalistic management style to indicate worker perception of management being predominantly characterized by benevolence or provision of decent work conditions, stemming from top management values or business strategy (Pellegrini and Scandura 2008; Perry, Wood, and Fernie 2015). We code the second item (regular consultation, 8.6%) as perceived consultative management style, which also allows us to control for factory-level procedural justice—allowing voice—that may trickle down to supervisor justice and worker attitudes (Colquit et al. 2013). As the last two items both emphasize rule/order obedience with similar negative implications for our variables, we combined them under the label perceived authoritarian management style, which emphasizes worker perceptions of authority, discipline, and control (Pellegrini and Scandura 2008); we use this as the baseline group in our analysis. Second, we use factory fixed-effects by including dummy variables for each (n– 1) factory to control for other unmeasured factory-wide practices. This approach excludes 128 observervations with only one worker in the factory.

Finally, we included a fixed effect for job types as exposure to supervisors may vary across jobs and the labor market or alternative jobs for distinct jobs/skills may differ and influence turnover intention. In the subset in which factories with only one worker were omitted, there were approximately 20 types of jobs, with the largest groups being operator (49.39%) and helper (23.43%). ⁴ After deleting jobs with only one worker for job fixed effect, our final sample comprises 1,377 workers at 13 types of jobs across 107 factories.

Analytical Strategy

We conducted all the analyses using an ordinary least squares (OLS) regression framework, including the mediation analyses. For the mediation test in H2b, the basic intuition popularized by Baron and Kenny (1986) is to look at three regression models: the relationships between the independent variable (X) and the mediator (M), between the mediator (M) and dependent variable (Y), and whether the relationship between X and Y is significantly reduced with the inclusion of M. Since then, mediation tests have developed further. For instance, in the following models a×X→M and X+b×M→Y, the product of coefficients a×b (the mediation index) is used as the size of the mediation or indirect effect of X to Y through M (e.g., Hayes 2018). ⁵

Multiple ways are available to test the statistical significance of this mediation index. ⁶ We rely on the bootstrapping method, which balances risk and power better than other methods (Preacher and Hayes 2008). We performed a bootstrap with 10,000 iterations and reported confidence intervals at 95% for our mediation (H2b) and moderated mediation (H3b and H4b) tests. We also triangulated these significance tests with Monte Carlo simulations as a robustness check.

Descriptive Statistics

Our final sample comprises mostly permanent workers (96.8%) in the sewing/making sections (67.9%) or the finishing section (18.88%). Most workers were under monthly wage contracts (96.01%), with approximately 3.78% on piece-rate contracts. Only 60 workers (4.36%) reported having a union in the workplace, while the majority did not know (61.29%).

Table 1 reports descriptive statistics of our final sample. SIJ was rated an average of 3.4—the median being 3.7— on a 5-point scale, echoing prior findings of verbal abuse of workers (e.g., Ahmed and Uddin 2022) as well as fair treatment for many workers. Workers reported low turnover intention, with the majority choosing 1 (“will not go,” 67.6%). Relatedly, workers scored high on positive affect as well: 2.4 on a 3-point scale.

OPEN IN VIEWER

Table 1.

Variable Overview

Variables	Mean	Std. dev.	Min	Max
Turnover intention	1.76	1.25	1	5
Supervisor interpersonal justice	3.40	0.92	1	5
Positive affect	2.48	0.45	1	3
Accord/Alliance	0.88	0.33	0	1
Worker Participation Committee	0.90	0.30	0	1
Job satisfaction	2.50	0.63	1	3
Monthly pay (log)	9.08	0.24	8.01	10.02
Age (years)	26.61	5.94	14	51
Gender (woman)	0.68	0.47	0	1
Garment tenure (years)	5.60	3.89	1	29
Perceived paternalistic management style	0.15	0.35	0	1
Perceived consultative management style	0.09	0.29	0	1
Perceived authoritarian management style	0.76	0.43	0	1

Notes: Accord/Alliance = the Accord on Fire and Building Safety in Bangladesh, and the Alliance for Bangladesh Workers’ Safety.

Table 2 reports bivariate correlations among the variables. As we expected, SIJ shows a significant negative relationship with turnover intention and is associated with higher positive affect. Further, SIJ did not vary significantly with WPC (r = 0.02) or Accord/Alliance (r = 0.01), suggesting the relative independence of these variables and the safety focus of Accord/Alliance.

OPEN IN VIEWER

Table 2.

Bivariate Correlations among Studied Variables

	Variable	1	2	3	4	5	6	7	8	9	10	11	12
1	Turnover intention
2	Supervisor interpersonal justice	–0.15***
3	Positive affect	–0.22***	0.36***
4	Accord/Alliance	–0.18***	0.01	0.16***
5	Worker Participation Committee	–0.23***	0.02	0.17***	0.64***
6	Job satisfaction	–0.18***	0.34***	0.50***	0.11***	0.11***
7	Monthly pay (log)	–0.17***	0.08**	0.16***	0.23***	0.22***	0.07**
8	Age	–0.06*	0.02	–0.02	0.06*	0.09***	–0.01	0.16***
9	Women	–0.02	–0.09***	–0.03	–0.01	0	0.01	–0.36***	–0.08**
10	Garment tenure	–0.02	0.04	–0.01	0.02	–0.05	–0.02	0.38***	0.45***	–0.16***
11	Paternalistic management style	–0.12***	0.30***	0.10***	0.02	0.02	0.16***	0.02	–0.01	–0.03	0
12	Consultative management style	–0.05	0.10***	0.19***	0.09***	0.06*	0.17***	0.09***	0	–0.05	0.02	–0.13***
13	Authoritarian management style	0.13***	–0.32***	–0.21***	–0.08**	–0.05*	–0.25***	–0.08**	0.01	0.05*	–0.02	–0.74***	0.56***

Notes: Paternalistic, consultative, and authoritarian management style(s) are worker perceived management styles. Accord/Alliance = the Accord on Fire and Building Safety in Bangladesh, and the Alliance for Bangladesh Workers’ Safety.

*

p <.05; **p <.01; ***p <.001 two-tailed test.

Analysis Results

As shown in Table 3, model 1’s OLS regression results, H1—the relationship between SIJ and turnover intention—was supported (β = −0.106; p = 0.008). The coefficient suggests that, a 1 unit increase in SIJ is associated with a 0.106 unit decrease in turnover intention (both were on a 5-point scale), controlling for other variables among workers in the same factory. As the average turnover intention was 1.76 in our sample, a 0.106 decrease represents a 6% drop. ⁷

OPEN IN VIEWER

Table 3.

Direct and Indirect Effects of Supervisor Interpersonal Justice

	Model 1	Model 2	Model 3
	DV = Turnover intention		DV = Positive affect
Supervisor interpersonal justice	–0.106**(0.040)	–0.081*(0.041)	0.112***(0.012)
Positive affect		–0.221*(0.090)
Age	–0.009(0.007)	–0.009(0.007)	–0.002(0.002)
Woman	–0.256**(0.087)	–0.248**(0.087)	0.034(0.027)
Garment tenure	0.017(0.011)	0.016(0.011)	–0.005(0.0035)
Job satisfaction	–0.185**(0.056)	–0.125*(0.061)	0.271***(0.018)
Monthly pay (log)	–0.760***(0.22)	–0.719***(0.22)	0.185**(0.068)
Perceived paternalistic management style (vs. authoritarian style)	–0.318**(0.098)	–0.323***(0.097)	–0.0227(0.031)
Perceived consultative management style	–0.044(0.12)	–0.018(0.12)	0.118**(0.036)
Job fixed effect (13 jobs)	Included	Included	Included
Factory fixed effect (107 factories)	Included	Included	Included
Constant	9.796***(1.94)	9.749***(1.94)	–0.213(0.61)
R-squared	0.246	0.249	0.418

Notes: Ordinary least squares (OLS) regression coefficients with standard errors in parentheses; N = 1,377. DV = dependent variable.

*

p <.05; **p <.01; ***p <.001 two-tailed test.

H2a suggested a postive relationship between SIJ and positive affect. The results in model 3 of Table 3 support this hypothesis: β = 0.112 (p < 0.001). This result suggests that a 1 unit increase in SIJ is related with a 0.112 unit (on a 3-point scale, or 4.5%) increase in worker positive affect, accounting for other variables within each factory. H2b proposed an indirect effect from SIJ to turnover intention via positive affect. Model 2 in Table 3 supported a signficant negative relationship from positive affect to turnover intention: β = −0.221, p = 0.014. And the coefficient for SIJ on turnover intention with the inclusion of the mediator/positive affect in model 2 decreased to β = −0.081 (p = 0.046) from −0.106 in model 1. The indirect effect from SIJ to turnover intention via positive affect is thus −0.025 (i.e., 0.112 ×–0.221). Based on bootstrap (10,000 repeated sampling), the mediation index was significant at 95% confidence intervals (CI): [–0.0473, −0.0047], which excludes zero. In other words, the mediation index that signifies the indirect effect is unlikely to be zero in the population. H2b is thus also supported.

Regarding control variables, turnover intention was lower for female workers and workers with higher job satisfaction or monthly pay, consistent with prior findings on turnover rates (Li and Kuruvilla 2023). Perceived paternalistic management style (compared to authoritarian style) is significantly related to lower turnover intention, but not with positive affect. Perceived consultative management style is related with significantly higher positive affect but not turnover intention.

H3 proposed synergy between SIJ and workplace voice mechanisms such as WPC. We examine this moderation by adding an interaction term between SIJ and WPC in the regression for positive affect. As shown in Table 4, model 4, the coefficient for the interaction term is significant: β = 0.136, p = 0.002. As we included factory fixed effects, this essentially suggests that the relationship between SIJ and positive affect is stronger (by 0.136 units) among factories with WPCs compared to those without. As the positive sign indicates, Figure 2 plotting the moderation effect shows a steeper line between SIJ and positive affect when the factory has a WPC. Simple slope tests show that the relationship between SIJ and positive affect is b = −0.014 (p = 0.729) in the absence of a WPC; this relationship becomes positive and significant when a WPC exists: b = 0.122 (p < 0.001). As the main effect of SIJ on positive affect is 0.112 (model 3 in Table 3), having a WPC is associated with 8.9% increase in the effect of SIJ. H3a is supported.

OPEN IN VIEWER

Table 4.

Synergy between Supervisor Interpersonal Justice and Labor Rights Institutions

	DV = Positive affect
	Model 4	Model 5
	H3a moderator: WPC	H4a moderator: Accord/Alliance
Supervisor interpersonal justice (SIJ)	–0.014(0.040)	0.033(0.035)
SIJ × Worker Participation Committee (WPC)	0.136**(0.042)
SIJ × Accord/Alliance		0.088*(0.037)
Age	–0.002(0.002)	–0.002(0.002)
Woman	0.038(0.027)	0.036(0.027)
Garments tenure	–0.005(0.003)	–0.005(0.004)
Job satisfaction	0.272***(0.018)	0.271***(0.018)
Monthly pay (log)	0.195**(0.068)	0.190**(0.068)
Perceived paternalistic management style (vs. authoritarian style)	–0.020(0.031)	–0.022(0.031)
Perceived consultative management style	0.118**(0.036)	0.115**(0.036)
Job fixed effect (13 jobs)	Included	Included
Factory fixed effect (107 factories)	Included	Included
Intercept	–0.039(0.67)	–0.183(0.62)
R-squared	0.423	0.420

Notes: Ordinary least squares (OLS) regression coefficients with standard errors in parentheses; N = 1,377. Accord/Alliance = the Accord on Fire and Building Safety in Bangladesh, and the Alliance for Bangladesh Workers’ Safety; DV = dependent variable.

*

p <.05; **p <.01; ***p <.001 two-tailed test.

OPEN IN VIEWER

Figure 2.

Interaction with Worker Participation Committee (WPC)

For H3b on moderated mediation, we conducted a bootstrap of 10,000 iterations that effectively multiplies the coefficient for the interaction term—SIJ × WPC (model 1 in Table 4)—and the coefficient for positive affect→turnover intention (model 2 in Table 3). ⁸ The indirect relationship between SIJ and turnover intention was moderated by WPC at 95% CI: [–0.0693, −0.003], which excludes zero. H3b is supported.

Turning to synergy with Accord/Alliance, the interaction between SIJ and Accord/Alliance is positive and significant for positive affect: β = 0.088, p = 0.016 (model 5 in Table 4). As with the discussion on H3b above, our fixed effect model implies factories under Accord/Alliance coverage benefit from a stronger effect of SIJ on positive affect (by 0.088). Figure 3 plots this interaction effect, showing a steeper line among workers covered by Accord/Alliance. Simple slope tests show that the relationship between SIJ and positive affect was not significant among workers not covered by Accord/Alliance: b = 0.033, p = 0.342; but this relationship becomes significant with the coverage of Accord/Alliance: b = 0.121, p < 0.001. Accord/Alliance coverage is related with an 8% increase from the main effect of SIJ on positive affect. H4a is supported. For H4b on moderated mediation by Accord/Alliance, the index is also significant at 95% bootstrap (10,000 iterations) confidence interval: [–0.0497, −0.0003] albeit the upper bound is very close to zero. H4b is thus supported.

OPEN IN VIEWER

Figure 3.

Interaction with Accord/Alliance

Robustness Analysis

We conducted the following robustness analyses. First, we note a strong overlap between WPC and Accord/Alliance in our data. Among 72 factories covered by Accord/Alliance, 63 (87.5% factories) also have WPC. We think it is still meaningful to have these two moderators in our analysis because, besides their conceptual difference, Accord/Alliance is an objective measure and WPC presence is based on the majority of workers’ perceptions or those of competent respondents (which leaves some room for error). To probe their separate moderation effects, we analyze WPC’s moderation on the relationship between SIJ and positive affect (H3a) among the 169 workers from 35 factories outside Accord/Alliance. The moderation term in the full model (H3a) with the 169 workers was not significant (β = 0.108, p = 0.173) due to limited observations and a large number of control variables (53) and thus low statistical power to find significant results. We then drop job fixed effect (10-plus dummy variables), and the interaction term was marginally significant ⁹ : β = 0.138, p = 0.066. We also analyze the moderation effect of Accord/Alliance among 32 factories (133 workers) without WPC. The interaction term for the full model (H4a) was not significant (β = 0.048, p = 0.713), and remained so even when we removed the dummy variables for job designation (β = 0.065, p = 0.606). This might result from the small sample size. These separate analyses suggest that WPC may moderate the relationship between SIJ and positive affect outside Accord/Alliance, but the moderation effect of Accord/Alliance outside WPC deserves more research (we elaborate in the Discussion section).

Second, we analyze potential endogeneity of supervisor justice; in particular, is SIJ predicted by WPC or Accord/Alliance? In our sample, they differ conceptually (none of the schemes mandates the others) and are statistically uncorrelated. We run a random effects model using WPC and Accord/Alliance, together with age, gender, garment tenure, monthly pay, jobs, and perceived management styles, to predict SIJ. The coefficients for WPC and Accord/Alliance are not significant (p > 0.163) while perceived paternalistic and consultative management styles both relate with better SIJ (p < 0.001).

We analyze the variance of SIJ and find that 6% of the difference is between factories whereas 94% of the variance arises from differences across individuals within the same factory. Although the between-variance is relatively small, there is a chance that some factory characteristics systematically influence SIJ. We allow for this possibility by accounting for potential clustering of SIJ in factories (Abadie, Athey, Imbens, and Woolridge 2023). More specifically, we use the clustered bootstrap approach to resample at the group/factory level to test all our hypotheses. The advantage of this method is that it is non-parametric and does not make any assumptions about the model (Field and Welsh 2007), and it outperforms the OLS-based or cluster robust standard error approach (Harden 2011). It is also an intuitive extension of the regular bootstrap approach and could accommodate our sensitive moderated mediation tests. In 10,000 iterations, we resampled at the cluster/factory level and constructed a 95% confidence interval of the coefficient for all the analyses (including the main effects and moderations). We found the same pattern as our OLS/regular bootstrap findings, except the moderated mediation of Accord/Alliance, which became insignificant. Table 5 summarizes confidence intervals for the mediation (H2b) and moderated mediations (H3b and H4b) using regular bootstrapping, Monte Carlo simulations (MCS), and clustered bootstrapping. Overall, results suggest that the moderation effect of Accord/Alliance is less robust or it may shape other factory practices that together directly influence SIJ. We return to this in the Discussion section.

OPEN IN VIEWER

Table 5.

Alternative Significance Analysis Results of Mediation and Moderated Mediation

	Regular bootstrap		Clustered bootstrap		Monte Carlo method
95% CI	Lower	Upper	Lower	Upper	Lower	Upper
Mediation	–0.0473	–0.0047	–0.0485	–0.001	–0.0461	–0.0048
Moderated mediation by WPC	–0.0693	–0.003	–0.0699	–0.0007	–0.0651	–0.0045
Moderated mediation by Accord/Alliance	–0.0497	–0.0003	–0.0496	0.0006	–0.0459	–0.0012

Notes: Bootstrap based on 10,000 iterations. Accord/Alliance = the Accord on Fire and Building Safety in Bangladesh, and the Alliance for Bangladesh Workers’ Safety; CI = confidence interval; WPC = Worker Participation Committee.

Third, we further examine the sensitivity of our main results to potential unmeasured omitted variables that could influence both SIJ and outcome variables. We followed the partial R² method of Cinelli and Hazlett (2020) and summarize the steps and detailed test results in the Online Appendix. The SIJ–turnover intention relationship (H1) can become non-significant (p > 0.05) in the presence of a confounder that explains 1.99% of the residual variances of SIJ and turnover intention; a confounder as strong as perceived paternalistic management style—a benchmark variable in our data—can do just that. So, this relationship can be explained away. The SIJ–positive affect relationship (H2a) can become non-significant by a confounder that explains 17.87% of the residual variances of both SIJ and positive affect. In our data, a confounder as strong as job satisfaction would not make it non-significant, though a confounder two times as strong as job satisfaction would. As job satisfaction already explains 10.9% (partial R² = 0.109) of the residual variance of positive affect, it is less likely to have confounders stronger than this. The relationship is relatively robust, though an extreme confounder might exist. The positive affect–turnover intention relationship (H2b) can also be explained away by a confounder accounting for 1.41% of the residual variances of both, and a confounder as strong as job satisfaction can do that. We extend this approach to simulate a mediation index from SIJ to positive affect to turnover intention and it is not robust to confounder(s). These results suggest that our findings on SIJ’s direct and indirect effect on turnover are fragile while the relationship between SIJ and positive affect is relatively robust. We discuss implications in the next section.