DIRE Warnings, Part 5: Failure Modes
How Diversity, Inclusion, Representation, and Equity are supposed to work and where they go wrong.
This article is Part 5 in a series. Part 1 summarized what the series will demonstrate. Part 2 addressed the definitions of equity, diversity, and inclusion including misconceptions. Part 3 addressed relevant (Canadian) legislation and psychology. Part 4 addressed representation proper application of concepts.
The prior parts of DIRE Warnings identified that equity, diversity, and inclusion (EDI) are about internal processes, treatment of individuals, and treating people as multi-dimensional rather than one-dimensional, especially with respect to immutable traits. EDI is not about head-counts by social categories or where they are in org charts. Representation is about one dimension, primarily about immutable traits, and includes head-counts and percentages. But, representation is a regulated feedback mechanism for checking an individual organization, or collection of organizations, against the availability pool to help identify any clogs in the hiring and retention system within the organizations.
Diversity, inclusion, representation, and equity (DIRE) together make for a happier, better performing workforce free from unnecessary and unfair biases, with a flow free from clogs and a feedback monitoring mechanism to detect when clogs start. But, they have a common failure mechanism. This failure mechanism is the attempt to directly control outcome numbers based on single-dimension social categories, treating DIRE as if it is a quota system. The failure can occur whether it is specifically designed as a mechanism to increase current levels (“more of category X”), or attempting to hit a specific target quota.
When it comes to EDI specifically, the incorrect definition of diversity is particularly prone toward this failure mechanism. From Part 2 of DIRE Warnings, the Misconceived EDI definition of diversity includes elements of “differences in race, colour, place of origin, religion, immigrant and newcomer status, ethnic origin, ability, sex, sexual orientation, gender identity, gender expression and age”. This definition is typically tied to increasing head-count numbers or relative percentages of people with the various trait-based social categories listed, often with an outcome target.
Setting specific target numbers, even the identified representation rates, can also drive a DIRE system to a failure mode. From Part 4 of the series, representation in the Canadian Employment Equity Act (EEA) is achieved by identifying the source of shortfalls and fixing them, not by setting target numbers.
There are many reasons that target quota systems lead to failure. In this article, I will discuss four mechanisms of failure.
(1) Disrespect for Diversity
The whole point of diversity (Part 2, Figure 2) is that people have differences in their values, choices, and habits, including cultural differences. The value of diversity, even the Misconceived EDI definition of diversity, is that different ideas and approaches bring a wider range of solutions to problems or options to examine, which produce different outcomes. If they didn’t produce any difference in outcomes, there’d be no value in that diversity.
Any person setting target head-counts implicitly declares that what they value, or what the “over-represented” groups value, is what everybody should value. The reason there are low numbers of Amish in the military, police, STEM fields, and business leadership is not because of a systemic bias against the Amish. It is because Amish cultural beliefs are in direct contradiction to the fundamentals of those professions. Any person deciding to target a specific outcome number of Amish in these positions is inserting their own personal values of what the Amish should want to do instead of letting Amish individuals decide what they do want to do based on their own values.
Properly designed DIRE systems makes an organization flexible to adjust to what people want to do by their own values (when it comes to their own choices), not the other way around. If the available pool for physicists is 0% Amish, despite a 10% Amish population in the region, then 0% is the correct number. A society with 10% Amish by population that has 10% Amish in military, policy, STEM, and business is necessarily oppressive. It is consistent with a society that integrates or assimilates to a singular monoculture. DIRE is based on respecting multicultural diversity which is consistent with near 0% Amish in these fields.
To be clear, that doesn’t mean there should be an effort to keep it at 0% either, as that also would be a target outcome. Representation has no “should” component; it is not foundationally a moralistic nor political concept. Representation is a statistical measurement tool. Representation is a measure relative to what is available, not what ought to be available. Conceptually what ought to be available is whatever count there is by self-selection under conditions with no biases or barriers, which is not easily predictable. It is fundamentally based on what people choose when there are no bias or barriers, which can only be measured by applying good EDI policy and seeing what that outcome is. It is not a pre-defined outcome and can’t be known beforehand.
While the Amish example is an extreme version to make the point, different cultures and different experiences in the world (diversity!) result in different statistical desires and choices (Blustein & Ellis, 2000) (Mau W.-C. , 2000) (Mau W.-C. J., 2004) (Dickson, 2010). From (Mau W.-C. , 2000),
“Results suggested that career decision-making styles have differential impacts on career decision-making self-efficacy, depending on the cultural background of the individuals. Results also showed significant differences in career decision-making style and career decision-making self-efficacy as a function of nationality and gender.”
This should not be surprising. After all, differences in outcome is the whole point of valuing diversity. From the research literature and CCDI training in Part 2, its value comes from the fact that doing things different ways results in different outcomes. Grouping people statistically by their common diversity characteristics and noting that they have different statistical outcomes is exactly what is expected with a healthy EDI social environment. Some cultures focus on arts, others focus on science and academia. Some focus on community, others focus on personal growth. People high in the Openness personality trait may be great at starting businesses, and people high in Conscientiousness may be better at keeping them organized and running. Extroverts may prefer jobs working with other people and excel at them; introverts may prefer jobs working alone and be more happy and productive that way.
If there we no differences in outcomes, that means either diversity is trivial with no value, or that the diversity is suppressed in the systemic processes, which is exactly what EDI is supposed to fix. When applied to a monocultural system that suppresses diversity, good EDI policy should result in increased diversity of outcomes, not homogeneity.
The reality is more complex once you consider the opposing effects of assimilative monoculture pushing uniformity of outcome on the one hand, and biases and barriers pushing disparity of outcome on the other hand. Since EDI, applied properly, serves to reduce and remove both sets of biases whether the demographic differences in careers increase or decrease, and by how much, depends on the pre-existing relative level of bias in either direction. In principle, EDI reducing both factors in equal amounts could have a zero net effect on demographic statistical outcomes.
It is important to note, however, that even in such cases of no demographic changes in outcomes, EDI is still successful at accomplishing its objectives; EDI’s objectives are not about changing demographic outcomes; they are about removing barriers and creating a higher performing and happy workforce.
Sometimes these results seem unintuitive, but they follow directly from reason and evidence. For example, it appears that the more we empower women and men to pursue whatever they want to do for fields of study, and remove the sociological and financial pressures, the outcomes tends toward bigger differences in what they chose to do, not smaller. This is known as the Gender Equality Paradox (Stoet, Bailey, Moore, & Geary, 2016) (Stoet & Geary, 2018) (Khazan, 2018).
The implication is that the dominant bias factor for gender-based career choices is to drive women and men to choose the same type of careers more than they would by their own wishes, and hence freeing people to make desired choices results in greater differences, not smaller. Note that this does not mean that gender-based bias causing greater disparity is zero; there is plenty of evidence that it is non-zero. Rather, that bias does not exist in a vacuum; when taken as a whole, the Gender Equality Paradox merely suggests that the bias toward making identical choices may be larger.
EDI policies reduce both biases, hence the end resulting outcome statistics cannot easily be predicted.
(2) Over-constrained System Design
Actively trying to control the outcome statistics creates an over-constrained system that can’t work. It attempts to control inputs, outputs, and the transfer function. If you create a fair system (“transfer function”) that is inclusive, equitable, and values diversity, and the input side has a diverse population across the dimensions of diversity, you cannot also control the output side. The only way to control the output side to produce a desired statistical outcomes is to either force the input side to be a homogeneous monoculture or to create an unfair system that is not inclusive or equitable and eliminates diversity within individuals to force the desired outcome (e.g., forcing the Amish to become physicists against their will).
Put simply, for y = f(x), you cannot control all three components of input x, output y, and transfer function f(). A multicultural society has a wide range of characteristics in people on the input side (x). EDI targets creating an employment system (f()) such that everybody can maximally contribute to the success of the organization without unnecessary barriers or biases, and attract the widest range of input populations (x). The outcomes (y) are the dependent variables in this case, and include a multitude of possible metrics that may be traded off, such as income vs work hours, responsibilities vs work-life balance, risk vs reward, or stress levels vs health. The distribution of outcomes (y) by demographic groups will depend on how their input diversity affects each outcome variable.
Over-constrained systems must fail at some point in the system. If the parts and constraints are both strong, the system will simply grind to a halt and stop functioning. In the case of using target outcomes, this might look like trying to solve the puzzle of where to move or hire people across the complete employment system to get the desired combinations of characteristics, only to realize that there is no configuration that produced the desired outcome, so nobody can move through it at all. The more factors required (e.g., race, gender, bilingualism, disability), the more unrealistic it is that there is a combination that can possibly work.
If parts are strong but some constraints are weak, the system will fail along a weak constraint such as sliding or loosening. That might look like attempting to hit one target by letting another constraint fail, such as trying to hit visible minority targets but letting bilingualism, gender, or disability target outcomes fail. Or, the legislative enforcement might be weak and so covert illegal acts take place to try to hit the targets with a low chance of getting caught. If constraints are strong but parts are weak, the system will fail by finding new degrees of freedom by breaking the parts such as bending, fatiguing, or bursting. This may look like departments or job functions that are understaffed or overstaffed in order to hit targets and cause the functions of various parts of the organization to fail.
(3) Undercutting Representation
There is a hidden failure mode in this case with respect to the DIRE system. The Employment Equity Act and Privacy Act federally in Canada, and PIPEDA and related provincial privacy legislation, all require that identification of people as visible minority, indigenous, or disabled must be via voluntary because these categorizations are considered personal information. Privacy legislation limits the conditions under which an employer is allowed to collect and disseminate such information and limit the conditions of maintaining confidentiality. Gender is not considered personal information, though this may change as gender identity and gender expression constraints have emerged in recent years, as it could force the outing of transgendered individuals whom do not want to be outed but also do not want to lie.
Under the EEA, employers can only count self-identified individuals toward representation. This adds a confounding variable in terms of the rate of self-identity for each category. When representation is applied as intended as a system check, this confounding variable should not be a significant issue as the rates of self-identity across organizations should not differ much, so comparisons should be valid.
But, this new variable adds an additional source to over-constrain the system as well as provide potential failure modes. The over-constraint condition occurs when the availability pool denominator is intentionally manipulated in an attempt to pressure increasing numbers. For the EEA representation check to work properly, the availability baseline number needs to be reasonable accurate to the (self-identified) statistical average.
Suppose the reality is that 25% of all software engineers are Klingon, but only 80% of them self-identify to their employer. Then all employers will see 20% (0.8 x 25%) of software engineers as self-identified Klingons. The real number being 25% is irrelevant to the representation check; if an organization sees it has 20% self-identified Klingons and so does everybody else, the numbers still indicate there is nothing especially wrong with the organization. If instead it sees 10%, then that organization needs to investigate, understand, and plan how to fix or justify the difference. It could be a low number of employed, or it could be potentially that Klingons at this one organization refuse to self-identify, which is itself an indication of a problem at the organization.
Suppose instead somebody were to replace the statistical baseline of 20% self-identified Klingons with, say, 40%, perhaps with the intent to make organizations appear lower than they really were and to pressure hiring more Klingons as a result. Now all of the organizations at 20% would report internally that they were at half of the availability. Following the EEA, they investigate and try to find out why. But, since they actually aren’t low in the first place, they will never find out why. Or worse, they will develop ad hoc explanations and make unnecessary changes that do nothing to help, because they didn’t really need help.
A potential example of this might be found in the Accessibility Strategy for the Public Service of Canada (Treasury Board of Canada Secretariat, 2021), with respect to plans for addressing low representation of persons with disabilities by replacing the workforce availability with a “work potential” concept which is a projected target based on a hypothetical number of persons with disabilities employed under ideal conditions. Unfortunately, in doing so, it will likely remove the value of the representation check as feedback tool and instead train organizations to ignore the reported value.
Another failure mode that could follow from this is the organizations do everything they can to try to hit targets, including violating the Canadian Human Rights Act (or provincial equivalent) and depleting the actual availability pool. In the fictional example of Klingon software engineers, trying to get numbers up to 40% when the real availability pool is only 20% means the rapid hiring will drive the availability pool from 20% Klingon toward 0%, and so future hiring will have a trough of very low percentage, and in the long run will re-settle to 20% since that is the accurate number, resulting in legal risks and no changes.
Another failure mode is poaching from other organizations which is a zero-sum effort in which the increases at one organization result in decreases at another, and turn organizations against each other as Klingons essentially become a currency. At no point does any of this effort actually help Klingons or increase the number of Klingon software engineers, as the average will always end up to be 20%.
Another failure mode is that, if it becomes apparent that the baseline availability number has been manipulated and trust is lost in it, organizations that actually have internal issues will no longer be able to identify they have issues. If the true value is 20% but nobody knows that because it was manipulated to be 40%, and everybody knows it, then an organization that is at 10% doesn’t know it is below average and may have an internal problem. If the baseline were properly reported as 20%, the organization could do statistical analysis along the hiring and retention chain to see where the missing numbers show up. But, with the incorrect availability number, the correct numbers to be looking for are unknown. (Hence, the case for the Accessibility Strategy may make it more difficult to find source issues and more prone toward ignoring the number.)
Finally, there is a failure mode from the self-identity variable regardless of whether the availability baseline is correct or not; any effort to hit a target can be addressed by the employer organization coercing employees to self-identify. The targets might be hit without any changes to employees or organization. In this failure mode nothing has changed, nobody has been helped, and no EDI-based improvements were made. In fact, the individuals who were intended to be helped are potentially slightly worse off because they have provided their employer with some personal information that they had not originally intended to share. Their privacy has been somewhat harmed and there is now a risk that their private information could possibly be inadvertently shared or leaked. This failure mode is an easy out for organizations that don’t want to undertake the hard work. Target numbers can be hit simply by changing how the employees identify. Note that in proper use of EDI this failure is not possible. EDI as intended has nothing to do with head-counts, org charts, or one-dimensional statistics so this approach can’t address EDI at all when it is used properly.
This failure mode only exists by confusing EDI with representation and using it to achieve some target head-count numbers instead of as a bias-checking mechanism.
(3) Shortcutting Representation
When given target outcomes, the easiest thing to do is to hire directly to achieve the target head-counts by demographic categories. Doing this both tends to run afoul of human rights legislation, such as the Canadian Human Rights Act, but also avoids the hard work of determining why the headcounts were low in the first place, which may be EDI issues such as policies, practices, or internal culture that are not inclusive and create conditions that cause people to leave or turn down positions. Even with direct hiring for numbers, without fixing the underlying causes the numbers will likely return to where they were before. This is not a problem when EDI targets are not based on numerical target outcomes, but instead about measuring the flexibility, accommodation, and inclusiveness of the work environment, and using the representation check to investigate where along the hiring and retention process that the numbers change.
This failure mode only gets worse when the target head-counts are artificially larger than the availability pool, because of the flowrate mathematics. The more the organization hires above the true availability rate, the more the pool becomes depleted of that demographic, making it even harder to attract more candidates in that category. As noted above, this also does not help individuals in the targeted demographic category, does not change their average rate of participation, and very likely violates legislation.
None of the identified failure mechanisms or modes result when DIRE policies are applied properly. Target numbers can make good motivational tools, such as in charity drives, but care must be taken not to use them as measurements of success or failure, or as incentives, for EDI policies and initiatives.
In the next article, Part 6 will summarize an conclude the DIRE Warnings series.