Are employee surveys biased? CHAS Journal club, Oct 13, 2021

Impression management as a response bias in workplace safety constructs

In October,, 2021 the CHAS Journal club reviewed the 2019 paper by Keiser & Payne examining the impact of “impression management” on the way workers in different sectors responded to safety climate surveys. The authors were able to attend to discuss their work with the group on October 13. Below is their presentation file as well as the comments from the table read the week before.

Our thanks to Drs. Keiser and Payne for their work and their willingness to talk with us about it!

10/06 Table Read for The Art & State of Safety Journal Club

Excerpts from “Are employee surveys biased? Impression management as a response bias in workplace safety constructs”

Full paper can be found here: https://www.sciencedirect.com/science/article/abs/pii/S0925753518315340?casa_token=oOShJnb3arMAAAAA:c4AcnB3fwnlDYlol3o2bcizGF_AlpgKLdEC0FPjkKg8h3CBg0YaAETq8mfCY0y-kn7YcLmOWFA

Meeting Plan

  • (5 minutes) Sarah to open meeting
  • (15 minutes) All participants read complete document
  • (10 minutes) All participants use “Comments” function to share thoughts
  • (10 minutes) All participants read others’ Comments & respond
  • (10 minutes) All participants return to their own Comments & respond
  • (5 minutes) Sarah announces next week’s plans & closes meeting

Introduction

The ultimate goal of workplace safety research is to reduce injuries and fatalities on the job.[a] Safety surveys that measure various safety-related constructs,including safety climate (Zohar, 1980), safety motivation and knowledge (Griffin and Neal, 2000), safety participation and compliance (Griffin and Neal, 2000), and outcome indices (e.g., injuries, incidents, and near misses) are the primary way that researchers gather relevant safety data. They are also used extensively in industry. It is quite common to administer self-report measures of both safety predictors and outcomes in the same survey, which introduces the possibility that method biases prevalent in self-report measures contaminate relationships among safety constructs (Podsakoff et al., 2012).

The impetus for the current investigation is the continued reliance by safety researchers and practitioners on self-report workplace safety surveys. Despite evidence that employees frequently underreport in-juries (Probst, 2015; Probst and Estrada, 2010), researchers have not directly examined the possibility that employees portray the workplace as safer than it really is on safety surveys[b]. Correspondingly, the current investigation strives to answer the following question: Are employee safety surveys biased? In this study,we focus on one potential biasing variable, impression management, defined as conscious attempts at exaggerating positive attributes and ignoring negative attributes (Connelly and Chang, 2016; Paulhus, 1984).The purpose of this study is to estimate the prevalence of impression management as a method bias in safety surveys based on the extent to which impression management contaminates self-reports of various workplace safety constructs and relationships among them.[c][d][e]

Study 1

Method

This study was part of a larger assessment of safety climate at a public research university in the United States using a sample of research laboratory personnel. The recruitment e-mail was concurrently sent to people who completed laboratory safety training in the previous two years (1841) and principal investigators (1897). Seven hundred forty-six laboratory personnel responded to the survey… To incentivize participation, respondents were given the option to provide their name and email address after they completed the survey in a separate survey link, in order to be included in a raffle for one of five $100 gift cards.

Measures:

  • Safety climate
  • Safety knowledge, compliance, and participation
  • Perceived job risk and safety outcomes
  • Impression management

Study 2

a second study was conducted to

  1. Further examine impression management as a method bias in self-reports of safety while
  2. Accounting for personality trait variance in impression management scales.

A personality measure was administered to respondents and controlled to more accurately estimate the degree to which self-report measures of safety constructs are susceptible to impression management as a response bias.

Method

A similar survey was distributed to all laboratory personnel at a different university located in Qatar. A recruitment email was sent to all faculty, staff, and students at the university (532 people), which included a link to an online laboratory safety survey. No incentive was provided for participating and no personally identifying information was collected from participants. A total of 123 laboratory personnel responded.[f]

Measures:

  • Same constructs as Study 1, plus
  • Personality

Study 3

Two limitations inherent in Study 1 and Study 2 were addressed in a third study, specifically, score reliability and generalizability.

Method

A safety survey was distributed to personnel at an oil and gas company in Qatar, as part of a larger collaboration to examine the effectiveness of a safety communication workshop. All employees (∼370) were invited to participate in the survey and 107 responded (29% response rate). Respondents were asked to report their employee identification numbers at the start of the survey, which was used to identify those who participated in the workshop. A majority of employees provided their identifying information (96, 90%).

Measures:

  • Same constructs used in Study 1, plus
  • Risk propensity
  • Safety communication
  • Safety motivation
  • Unlikely virtues

Conclusion[g][h][i][j][k][l]

Safety researchers have provided few direct estimates of method bias [m][n][o][p]in self-report measures of safety constructs. This oversight is especially problematic considering they rely heavily on self-reports to measure safety predictors and criteria.

The results from all three studies, but especially the first two, suggest that self-reports of safety are susceptible to dishonesty[q][r][s][t][u] aimed at presenting an overly positive representation of safety.[v][w][x][y][z][aa] In Study 1, self reports of safety knowledge, climate, and behavior appeared to be more susceptible to impression management compared to self-reports of perceived job risk and safety outcomes. Study 2 provided additional support for impression management as a method bias in self-reports of both safety predictors and outcomes. Further, relationships between impression management and safety constructs remained significant even when controlling for Alpha personality trait variance (conscientiousness, agreeableness, emotional stability). Findings from Study 3 provided less support for the biasing effect of impression management on self-report measures of safety constructs (average VRR=11%). However, the unlikely virtues measure [this is a measure of the tendency to claim uncommon positive traits] did reflect more reliable scores as those observed in Study 1 and Study 2 and it was significantly related to safety knowledge, motivation, and compliance. Controlling for the unlikely virtues measure led to the largest reductions in relationships with safety knowledge. Further exploratory comparison of identified vs. anonymous respondents observed that mean scores on the unlikely virtues measure were not significantly different for the identified subsample compared to the anonymous subsample; however, unlikely virtues had a larger impact on relationships among safety constructs for the anonymous subsample.

The argument for impression management as a biasing variable in self-reports of safety relied on the salient social consequences to responding and other costs to providing a less desirable response, including for instance negative reactions from management, remedial training, or overtime work[ab][ac]. Findings suggest that the influence of impression management on self-report measures of safety constructs depends on various factors[ad] (e.g., distinct safety constructs, the identifying approach, industry and/or safety salience) rather than the ubiquitous claim that impression management serves as a pervasive method bias.

The results of Study 1 and Study 3 suggest that impression management was most influential as a method bias in self-report measures of safety climate, knowledge, and behavior, compared to perceived risk and safety outcomes. These results might reflect the more concrete nature of these constructs based on actual experience with hazards and outcomes. Moreover, these findings are in line with Christian et al.’s (2009) conclusion that measurement biases are less of an issue for safety outcomes compared to safety behavior. These findings in combination with theoretical rationale suggest that the social consequences of responding are more strongly elicited by self-report measures of safety climate, knowledge, and behavior, compared to self-reports of perceived job risk and safety outcomes. Items in safety perception and behavior measures fittingly tend to be more personally (e.g., safety compliance – “I carry out my work in a safe manner.”) and socially relevant (e.g., safety climate – “My coworkers always follow safety procedures.”).

The results from Study 2, compared to findings from Study 1 and Study 3, suggest that assessments of job risk and outcomes are also susceptible to impression management. The Alpha personality factor generally accounted for a smaller portion of the variance in the relationships between impression management and perceived risk and safety outcomes. The largest effects of impression management on the relationships among safety constructs were for relationships with perceived risk and safety outcomes. These results align with research on injury underreporting (Probst et al., 2013; Probst and Estrada, 2010) and suggest that employees may have been reluctant to report safety outcomes even when they were administered on an anonymous survey used for research purposes.

We used three samples in part to determine if the effect of impression management generalizes. However, results from Study 3 were inconsistent with the observed effect of impression management in Studies 1 and 2. One possible explanation is that these findings are due to industry differences and specifically the salience of safety. There are clear risks associated with research laboratories as exemplified by notable incidents; [ae]however, the risks of bodily harm and death in the oil and gas industry tend to be much more salient (National Academies of Sciences, Engineering, and Medicine, 2018). Given these differences, employees from the oil and gas industry as reflected in this investigation might have been more motivated to provide a candid and honest response to self-report measures of safety.[af][ag][ah][ai][aj] This explanation, however, is in need of more rigorous assessment.

These results in combination apply more broadly to method bias [ak][al][am]in workplace safety research. The results of these studies highlight the need for safety researchers to acknowledge the potential influence of method bias and to assess the extent to which measurement conditions elicit particular biases.

It is also noteworthy that impression management suppressed relationships in some cases; thus, accounting for impression management might strengthen theoretically important relationships. These results also have meaningful implications for organizations because positively biased responding on safety surveys can contribute to the incorrect assumption that an organization is safer than it really is[an][ao][ap][aq][ar][as][at].

The results of Study 2 are particularly concerning and practically relevant as they suggest that employees in certain cases are likely to underreport the number of safety outcomes that they experience even when their survey responses are anonymous. However, these findings were not reflected in results from Study 1 and Study 3. Thus, it appears that impression management serves as a method bias among self-reports of safety outcomes only in particular situations. Further research[au][av][aw] is needed to explicate the conditions under which employees are more/less likely to provide honest responses to self-report measures of safety outcomes.

———————————————————————————————————————

BONUS MATERIAL FOR YOUR REFERENCE:

For reference only, not for reading during the table read

Respondents and Measures

  • Study 1

Respondents:

graduate students (229,37%),

undergraduate students (183, 30%),

research scientists and associates (123,20%),

post-doctoral researchers (28,5%),

laboratory managers (25, 4%),

principal investigators (23, 4%)

329 [53%] female;

287 [47%] male

377 [64%] White;

16 [3%] Black;

126 [21%] Asian;

72 [12%] Hispanic

Age (M=31,SD=13.24)

Respondents worked in various types of laboratories, including:

biological (219,29%),

Animal biological (212,28%),

human subjects/computer (126,17%),

Chemical (124,17%),

mechanical/electrical (65,9%)

Measures:

  • Safety Climate

Nine items from Beus et al. (2019) 30-item safety climate measure were used in the current study. The nine-item measure included one item from each of five safety climate dimensions (safety communication, co-worker safety practices, safety training, safety involvement, safety rewards) and two items from the management commitment and safety equipment and  housekeeping dimensions. The nine items were identified based on factor loadings from Beus et al. (2019). Items were responded to on a five-point agreement scale (1=strongly disagree, 5=strongly agree).

  • Safety knowledge, compliance, and participation

Respondents completed slightly modified versions of Griffin and Neal’s (2000) four-item measures of safety knowledge (e.g., “I know how to perform my job in the lab in a safe manner.”), compliance (e.g., “I carry out my work in the lab in a safe manner.”), and participation (e.g., “I promote safety within the laboratory.”). Items were completed using a five-point agreement scale (1=strongly disagree, 5=strongly agree).

  • Perceived job risk and safety outcomes

Respondents completed a three-item measure of perceived job risk (e.g., “I encounter personally hazardous situations while in the laboratory;” 1=almost always untrue, 5=almost always true; Jermier et al., 1989). Respondents also provided safety incident data regarding the number of injuries, incidents, and near misses that they experienced in the last 12 months.

  • Impression Management

Four items were selected from Paulhus’s (1991) 20-item Balanced Inventory of Desirable Responding. These items were selected based on a review of Paulhus’s (1991) full measure and an assessment of those items that were most relevant and best representative of the full measure (Table 1). Items were completed using a five-point accuracy scale (1=very inaccurate, 5=very accurate). Ideally this survey would have included Paulhus’s (1991) full 20-item measure. However, as is often the case in survey research, we had to balance construct validity with survey length and concerns about respondent fatigue and for these reasons only a subset of Paulhus’s (1991) measure was included.

  • Study 2

Respondents:

research scientists or post-doctoral researchers (43; 39%)

principal investigators (12; 11%)

laboratory managers and coordinators (12; 11%)

graduate students (3; 3%)

Faculty teaching in a laboratory (3; 3%)

 one administrator (1%)

Respondents primarily worked in:

chemical (55; 45%)

mechanical/electrical (39; 32%)

Uncategorized laboratory (29; 24%)

Measures:

  • Safety Constructs

Respondents completed the same six self-report measures of safety constructs that were used in Study 1: safety climate, safety knowledge, safety compliance, safety participation, perceived job risk, and injuries, incidents, and near misses in the previous 12 months.

  • Impression Management

Respondents completed a five-item measure of impression management from the Bidimensional Impression Management Index (Table 1; Blasberg et al., 2014). Five items from the Communal Management subscale were selected based on an assessment of their quality and degree to which they represent the 10-item scale.5 A subset of Blasberg et al.’s (2014) full measure was used because of concerns from management about survey length. Items were responded to on a five-point agreement scale (1=strongly disagree, 5=strongly agree).

  • Personality

Conscientiousness, agreeableness, and emotional stability were assessed using six items from Gosling et al. (2003) 10-item personality measure. Four items from the 10-item measure assessing openness to experience and extraversion were not included in this study. Respondents were asked to indicate the degree to which adjectives were representative of them (i.e., Conscientiousness – “dependable, self-disciplined;”  agreeableness – “sympathetic, warm;” Emotional stability – “calm, emotionally stable”; 1=strongly disagree, 7=strongly agree) and combined to represent the Alpha personality factor. One conscientiousness item was dropped because it had a negative item-total correlation (“disorganized, careless” [reverse coded]). This was not surprising as it was the only reverse-scored personality item administered.

  • Study 3

Respondents:

The typical respondent was male (101, 94%) and had no supervisory responsibility (72, 67%); however, some women (6, 6%), supervisors (17, 16%), and managers/senior managers (16, 15%) also completed the survey.7 The sample was diverse in national origin with most respondents from India (44, 42%) and Pakistan (25, 24%).

Measures:

  • Safety Constructs

Respondents completed five of the same self-report measures of

safety constructs used in Study 1 and Study 2, including safety climate (Beus et al., 2019), safety knowledge (Griffin and Neal, 2000), safety compliance (Griffin and Neal, 2000), safety participation (Griffin and Neal, 2000), and injuries, incidents, and near misses in the previous 6 months. Respondents completed a similar measure of perceived job risk (Jermier et al., 1989) that included three additional items assessing the degree to which physical, administrative, and personal controls

  • Unlikely Virtues

Five items were selected from Weekley’s (2006) 10-item unlikely

virtues measure (see also Levashina et al., 2014; Table 1) and were responded to on a 5-point agreement scale (1=strongly disagree; 5=strongly agree). Akin to the previous studies, an abbreviated version of the measure was used because of constraints with survey length and the need to balance research and organizational objectives.

[a]In my mind, this is a negative way to start a safety research project. The ultimate goal of the organization is to complete its mission and injuries and fatalities are not part of the mission. So this puts the safety researcher immediately at odds with the organization.

[b]I wonder if this happens beyond surveys—do employees more generally portray a false sense of safety to co-workers, visitors, employers, trainees, etc? Is that made worse by surveying, or do surveys pick up on bias that exists more generally in the work culture?

[c]Employees always portray things in a better light on surveys because who really knows if its confidential

[d]Not just with regard to safety; most employees, I suspect, want to portray their businesses in a positive light. Good marketing…

[e]I think that this depends on the quality of the survey. If someone is pencil whipping a questionnaire, they are probably giving answers that will draw the least attention. However, if the questions are framed in an interesting way, I believe it is possible to have a survey be both a data collection tool and a discussion starter. Surveys are easy to generate, but hard to do well.

[f]In my experience, these are pretty high response rates for the lab surveys (around 20%).

[g]A concern that was raised by a reviewer on this paper was that it leads to a conclusion of blaming the workers. We certainly didn't set out to do that, but I can understand that perspective. I'm curious if others had that reaction.

[h]I had the same reaction and I can see how it could lead to a rosier estimate of safety conditions.

[i]There is an interesting note below where you mention the possible outcomes of surveys that "go poorly" if you will. If the result is that the workers are expected to spend more of their time and energy "fixing" the problem, it is probably no surprise that they will just say that there is no problem.

[j]I am always thinking about this type of thing—how results are framed and who the finger is being pointed at. I can see how this work can be interpreted that way, but I also see it from an even bigger picture—if people are feeling that they have to manage impressions (for financial safety, interpersonal safety, etc) then to me it stinks of a bigger cultural, systemic problem. Not really an individual one.

[k]Well – the "consequences" of the survey are really in the hands of the company or institution. A researcher can go in with the best of intentions, but a company can (and often does) respond in a way that discourages others from being forthright.

[l]Oh for sure! I didn't mean to shoulder the bigger problem on researchers or the way that research is conducted—rather, that there are other external pressures that are making individuals feel like managing people's impressions of them is somehow more vital than reporting safety issues, mistakes, needs, etc. Whether that's at the company, institution, or greater cultural level (probably everywhere), I don't think it's at the individual level.

[m]My first thought on bias in safety surveys had to do more with the survey model rather than method bias.  Most all safety surveys I have taken are based on the same template and questions generally approach safety from the same angle.  I haven't seen a survey that asks the same question several ways in the course of the survey or seen any control questions to attempt to determine validity of answers.  Perhaps some of the bias comes from the general survey format itself….

[n]I agree. In reviewing multiple types of surveys trying to target safety, there are many confounding variables. Trying to find a really good survey is tough – and I'm not entirely sure that it is possible to create something that can be applied by all. It is one of the reasons I was so intrigued by the BMS approach.

[o]Though—a lot of that work (asking questions multiple ways, asking control questions, determining validity and reliability, etc) is done in the original work that initially develops the survey metric. Just because it's not in a survey that one is taking or administering, doesn't necessarily mean that work isn't there

[p]Agreed – There are a lot of possible method biases in safety surveys. Maybe impression management isn't the most impactful. There just hasn't been much research in this area as it relates to safety measures, but certainly there is a lot out there on method biases more broadly. Stephanie and I had a follow up study (conference paper) looking at blatant extreme responding (using only the extreme endpoints on safety survey items). Ultimately, that too appears to be an issue

[q]In looking back over the summary, I was drawn to the use of the word 'dishonesty."  That implies intent.  I'm wondering whether it is equally likely that people are lousy at estimating risk and generally overestimate their own capabilities (Dunning Kruger anyone?).  So it is not so much about dishonesty but more about incompetency.

[r]They are more likely scared of retribution.

[s]This is an interesting point and I do think there is a part of the underestimation that has to do with an unintentional miscalibration. But, I think the work in this paper does go to show that some of the underestimation is related to people's proclivity to attempt to control how people perceive them and their performance.

[t]Even so, that proclivity is not necessarily outright dishonesty.

[u]I agree. I doubt that the respondents set out with an intent to be fraudulent or dishonest. Perhaps a milder or softer term would be more accurate?

[v]I wonder how strong this effect is for, say, graduate students who are in research labs under a PI who doesn't value safety

[w]I thinks its huge. I know I see difference in speaking with people in private versus our surveys

[x]Within my department, I know I became very cynical about surveys that were administered by the department or faculty members. Nothing ever seemed to change, so it didn't really matter what you said on them.

[y]I also think it is very significant. We are currently dealing with an issue where the students would not report safety violations to our Safety Concerns and Near Misses database because they were afraid of faculty reprisal. The lab is not especially safe, but if no one reports it, the conclusion might be drawn that no problems exist.

[z]And to bring it back to another point that was made earlier: when you're not sure if reporting will even trigger any helpful benefits, is the perceived risk of retribution worth some unknown maybe-benefit?

[aa]I heard a lot of the same concerns when we tried doing a "Near Miss" project. Even when anonymity was included, I had several people tell me that the details of the Near Miss would give away who they were, so they didn't want to share it.

[ab]Interesting point. It would seem here that folks fear if they say something is amiss with safety in the workplace, it will be treated as something wrong with themselves that must be fixed.

[ac]Yeah I feel like this kind of plays in to our discussion from last week, when we were talking about people feeling like they're personally in trouble if there is an incident

[ad]A related finding has been cited in other writings on surveys – if you give a survey, and nothing changes after the survey, then people catch on that the survey is essentially meaningless and they either don't take surveys anymore or just give positive answers because it isn't worth explaining negative answers.

[ae]There are risks associated with research labs, but I don't know if I would call them "clear". My sense is that "notable incidents" is a catchphrase people are using about academic lab safety to avoid quantitating the risks any more specifically.

[af]This is interesting to think about. One the one hand, if one works in a higher hazard environment maybe they just NOTICE hazardous situations more and think of them as more important. On the other hand, there is a lot of discussion around the normalization of hazards in an environment that would seem to suggest that they would not report on the hazards because they are normal.

[ag]Maybe they receive more training as well which helps them identify hazards easier. Oil & Gas industry Chemical engineers certainly get more training from my experience.

[ah]Oil and gas workers were also far more likely to participate in the study than the academic groups.  I think private industry has internalized safety differently (not necessarily better or worse) than academia.  And high hazard industries like oil and gas have a good feel for the cost of safety-related incidents.  That definitely gets passed on to the workforce

[ai]How does normalization take culture into effect? Industries have a much longer history of self-reporting and reporting of accidents in general than do academic institutions.

[aj]Some industries have histories of self-reporting in some periods of time. For example, oil and gas did a lot of soul searching after the Deepwater explosion (which occurred the day of a celebration of 3 years with no injury reports), but this trend can fade with time. Alcoa in the 1990s and 2000s is a good example of this. For example, I've looked into Paul H. O'Neill's history with Alcoa. He was safety champion whose work faded soon after he left.

[ak]I wonder if this can be used as a way to normalize the surveys somehow

[al]Hmm, yeah I think you could, but you would also have to take a measure of impression management so that you could remove the variance caused by that from your model.

Erg, but then long surveys…. the eternal dilemma.

[am]I bet there are overlapping biases too that have opposite effects, maybe all you could do is determine to what extent of un-reliability your survey has

[an]In the BMS paper we covered last semester, it was noted that after they started to do the managerial lab visits, the committee actually received MORE information about hazardous situations. They attributed this to the fact that the committee was being very serious about doing something about each issue that was discovered. Once people realized that their complaints would actually be heard & addressed, they were more willing to report.

[ao]and the visits allowed for personal interactions which can be kept confidential as opposed to a paper trail of a complaint

[ap]I imagine that it was also just vindicating to have another human listen to you about your concerns like you are also a human. I do find there is something inherently dehumanizing about surveys (and I say this as someone who relies on them for multiple things!). When it comes to safety in my own workplace, I would think having a human make time for me to discuss my concerns would draw out very different answers.

[aq]Prudent point

[ar]The Hawthorne Effect?

[as]I thought that had to do with simply being "studied" and how it impacts behavior. With the BMS study, they found that people were reporting more BECAUSE their problems were actually getting solved. So now it was actually "worth it" to report issues.

[at]It would be interesting to ask the same question of upper management in terms of whether their safety attitudes are "true" or not. I don't know of any organizations that don't talk the safety talk. Even Amazon includes a worker safety portion to its advertising campaign despite its pretty poor record in that regard.

[au]I wish they would have expanded on this more, I'm really curious to see what methods to do this are out there or what impact it would have, besides providing more support that self-reporting surveys shouldn't be used

[av]That is an excellent point and again something that the reviewers pushed for. We added some text to the discussion about alternative approaches to measure these constructs. Ultimately, what can we do if we buy into the premise that self-report surveys of safety are biased? Certainly one option is to use another referent (e.g., managers) instead of the workers themselves. But that also introduces its own set of bias. Additionally, there are some constructs that would be odd to measure based on anything other than self-report (e.g., safety climate). So I think it's still somewhat of an open question, but a very good one. I'm sure Stephanie will have thoughts on this too for our discussion next week. 🙂 But to me that is the crux of the issue: what do we do with self-reports that tend to be biased?

[aw]Love this, I will have to go read the full paper. Especially your point about safety climate, it will be interesting to see what solutions the field comes up with because everyone in academia uses surveys for this. Maybe it will end up being the same as incident reports, where they aren't a reliable indicator for the culture.

Leave a Reply

Your email address will not be published. Required fields are marked *