Fundamentals

Consider this ● a local bakery, a cornerstone of its community, decides to automate its online ordering system using an algorithm to predict demand and manage inventory. Initially, efficiency skyrockets, waste plummets, and profits rise. However, subtle biases creep into the algorithm, perhaps unintentionally favoring orders from certain zip codes or based on past purchasing patterns that inadvertently exclude new customers or those from less affluent areas. This scenario, seemingly benign, highlights a growing challenge for small and medium-sized businesses (SMBs) ● algorithmic fairness.

Understanding Algorithmic Fairness For Small Businesses

Algorithms, at their core, are sets of instructions. They guide computers to make decisions or predictions based on data. Think of them as digital recipes.

In business, these recipes are increasingly used for everything from targeted advertising and customer service chatbots to loan applications and hiring processes. For SMBs, algorithms offer the promise of efficiency, automation, and data-driven decision-making, tools once only accessible to large corporations.

Fairness, in this context, is about ensuring these algorithmic recipes do not unfairly discriminate against certain groups of people. Discrimination can occur along various lines ● race, gender, age, location, or any other protected characteristic. The challenge arises because algorithms learn from data, and if that data reflects existing societal biases, the algorithm can amplify and perpetuate those biases, sometimes in unexpected ways. For an SMB, unintentional algorithmic bias Meaning ● Algorithmic bias in SMBs: unfair outcomes from automated systems due to flawed data or design. can lead to reputational damage, legal issues, and, most importantly, unfair treatment of customers or employees.

Algorithmic fairness for SMBs Meaning ● SMBs are dynamic businesses, vital to economies, characterized by agility, customer focus, and innovation. is about ensuring automated systems treat all individuals equitably, preventing unintentional bias from undermining business values and customer trust.

Why Should SMBs Care About Algorithmic Fairness?

Some might argue that algorithmic fairness Meaning ● Ensuring impartial automated decisions in SMBs to foster trust and equitable business growth. is a concern only for tech giants or heavily regulated industries. This perspective misses a crucial point. SMBs operate within communities. Their reputations are built on trust and personal relationships.

Unfair algorithmic practices can erode this trust quickly. Imagine the bakery scenario again. If word gets out that the online ordering system is inadvertently discriminating against certain neighborhoods, customer loyalty could plummet. In today’s hyper-connected world, news travels fast, and negative perceptions can be difficult to reverse.

Beyond reputation, there are practical business reasons to care. Algorithms that perpetuate bias can lead to missed market opportunities. If a loan application algorithm unfairly rejects applicants from certain demographics, the SMB lender loses potential customers and limits its growth potential.

Similarly, biased hiring algorithms can exclude qualified candidates, hindering a company’s ability to build a diverse and talented workforce. In the long run, fairness is not just an ethical consideration; it is a smart business strategy.

Practical Business Metrics for SMBs

Measuring algorithmic fairness might seem daunting, especially for SMBs with limited resources. The good news is that it does not require complex statistical models or expensive software. Simple, practical business metrics can provide valuable insights. The key is to focus on outcomes and impact, rather than getting bogged down in technical details.

Here are some actionable ways SMBs can begin to measure algorithmic fairness:

Outcome Audits

This involves examining the results of algorithmic decisions for different groups. For example, if an SMB uses an algorithm to screen job applications, an outcome audit would compare the selection rates for different demographic groups (e.g., men vs. women, different racial groups). Significant disparities in selection rates could indicate potential bias.

Data Disaggregation

Break down your data to understand how different groups are represented. Are certain customer segments underrepresented in your marketing data? Is your customer service chatbot performing equally well for all users, regardless of their background? Disaggregating data can reveal hidden patterns and potential areas of concern.

Customer Feedback Analysis

Actively solicit and analyze customer feedback, paying attention to any complaints or concerns related to fairness or discrimination. Are customers from certain groups reporting negative experiences with your automated systems? Are there patterns in the feedback that suggest algorithmic bias?

Proxy Variable Identification

Algorithms often rely on proxy variables ● data points that are correlated with protected characteristics, even if they do not directly include them. For example, zip code can be a proxy for race or socioeconomic status. SMBs should be aware of potential proxy variables in their data and consider their impact on fairness.

Human Oversight and Review

Algorithms should not operate in a vacuum. Implement human oversight and review processes for critical algorithmic decisions, especially those that impact individuals directly. This could involve manual review of loan applications flagged by an algorithm or human intervention in customer service interactions handled by chatbots.

These metrics are not exhaustive, but they provide a starting point for SMBs to begin measuring and addressing algorithmic fairness. The goal is not to achieve perfect fairness ● a complex and perhaps unattainable ideal ● but to identify and mitigate potential biases, ensuring that automated systems align with business values and ethical principles.

Consider a small online retailer using an algorithm to personalize product recommendations. By tracking click-through rates and purchase rates for different demographic groups, they might discover that the algorithm is less effective at recommending products to older customers. This could be due to biases in the training data or the algorithm itself. By identifying this disparity, the retailer can take steps to adjust the algorithm or diversify its recommendation strategies, ensuring a fairer and more effective customer experience for all.

Simple Tools and Techniques for SMBs

SMBs do not need to invest in expensive AI fairness software to get started. Many readily available tools and techniques can be used to measure and improve algorithmic fairness:

• Spreadsheet Software ● Tools like Microsoft Excel or Google Sheets can be used for basic data analysis, outcome audits, and tracking fairness metrics.
• Data Visualization Tools ● Free or low-cost data visualization tools can help SMBs identify patterns and disparities in their data, making potential biases more visible.
• A/B Testing ● Experiment with different algorithmic approaches or parameters using A/B testing to compare outcomes for different groups and identify fairer alternatives.
• Regular Monitoring ● Establish a routine for regularly monitoring fairness metrics Meaning ● Fairness Metrics, within the SMB framework of expansion and automation, represent the quantifiable measures utilized to assess and mitigate biases inherent in automated systems, particularly algorithms used in decision-making processes. and reviewing algorithmic outcomes. This ongoing process is crucial for detecting and addressing bias over time.

The process of measuring algorithmic fairness for SMBs is iterative. It starts with awareness, moves to measurement, and then to action. By taking these initial steps, SMBs can begin to navigate the complexities of algorithmic fairness and build more equitable and sustainable businesses.

For a local coffee shop using an algorithm to manage employee scheduling, they could use spreadsheet software to track employee satisfaction and work-life balance across different demographic groups. If they notice that younger employees are consistently assigned less desirable shifts, they can investigate whether the scheduling algorithm is unintentionally biased and make adjustments to ensure fairer shift assignments.

In conclusion, algorithmic fairness is not a niche concern for SMBs. It is a fundamental business imperative in an increasingly automated world. By adopting practical measurement strategies and utilizing readily available tools, SMBs can proactively address algorithmic bias, build stronger customer relationships, and foster a more equitable business environment. It begins with recognizing that even simple algorithms can have fairness implications and taking concrete steps to understand and mitigate those implications.

Intermediate

The initial allure of algorithms for SMBs often centers on efficiency gains and cost reduction. However, as SMBs integrate algorithms deeper into their operations, a more critical understanding of algorithmic fairness becomes essential. The seemingly neutral code, designed to optimize processes, can inadvertently amplify existing societal inequalities, leading to unintended business consequences.

Consider a growing e-commerce SMB utilizing algorithmic marketing to personalize customer experiences. If the algorithm, trained on historical sales data, disproportionately targets affluent demographics while neglecting lower-income segments, the SMB inadvertently reinforces economic disparities and limits its market reach.

Moving Beyond Basic Metrics ● Deeper Fairness Analysis

While outcome audits and data disaggregation provide a foundational understanding, intermediate-level algorithmic fairness measurement Meaning ● Ensuring unbiased, equitable outcomes from automated systems in SMBs. for SMBs requires a more nuanced approach. This involves exploring different dimensions of fairness and employing metrics that capture these complexities. The concept of “fairness” itself is not monolithic; various definitions exist, each with its own implications for business practice.

Here are some key fairness concepts and associated metrics relevant for SMBs at an intermediate stage:

Statistical Parity

This concept aims for equal representation in outcomes across different groups. For example, in a loan application scenario, statistical parity would mean that the approval rate for loan applications should be roughly the same for all demographic groups. The metric used to measure this is often the Disparate Impact Ratio, which compares the selection rate for a disadvantaged group to the selection rate for an advantaged group. A ratio significantly below 1 (e.g., below 0.8) is often considered indicative of disparate impact.

Equal Opportunity

This focuses on ensuring that qualified individuals from all groups have an equal chance of achieving a positive outcome. In a hiring context, equal opportunity would mean that among candidates who are equally qualified for a job, the selection rate should be similar across different demographic groups. Metrics for equal opportunity often involve comparing True Positive Rates (the rate at which qualified candidates are correctly identified) across groups.

Predictive Parity

This concept emphasizes the accuracy of predictions across different groups. It aims to ensure that the algorithm is equally reliable in its predictions for all segments of the population. Metrics for predictive parity include comparing Positive Predictive Values (the proportion of positive predictions that are actually correct) and Negative Predictive Values (the proportion of negative predictions that are actually correct) across groups.

Selecting the appropriate fairness metric depends on the specific business context and the potential harms associated with algorithmic bias. Statistical parity might be relevant in situations where equal representation is a primary concern, while equal opportunity might be more appropriate when focusing on ensuring fair access to opportunities for qualified individuals. Predictive parity is crucial when algorithmic predictions directly impact individuals’ lives, such as in risk assessment or fraud detection.

Intermediate algorithmic fairness measurement for SMBs involves selecting and applying relevant fairness metrics, moving beyond basic outcome analysis to capture the nuances of equitable algorithmic performance.

Practical Tools and Frameworks for Intermediate Analysis

As SMBs delve deeper into algorithmic fairness, they may benefit from utilizing more sophisticated tools and frameworks. While complex AI fairness toolkits might be overkill, several accessible options can enhance their measurement capabilities:

Fairness Libraries in Programming Languages

For SMBs with in-house technical expertise, fairness libraries available in programming languages like Python (e.g., AIF360, Fairlearn) offer a more structured approach to fairness measurement. These libraries provide implementations of various fairness metrics, bias detection algorithms, and fairness-aware machine learning Meaning ● Machine Learning (ML), in the context of Small and Medium-sized Businesses (SMBs), represents a suite of algorithms that enable computer systems to learn from data without explicit programming, driving automation and enhancing decision-making. techniques. While requiring some technical proficiency, they offer a powerful way to analyze algorithmic fairness in a more rigorous manner.

Cloud-Based AI Fairness Services

Cloud platforms like Google Cloud AI Platform and Azure Machine Learning offer services that include fairness assessment tools. These services often provide user-friendly interfaces and pre-built fairness metrics, making them more accessible to SMBs without extensive data science expertise. They can help automate fairness evaluation and provide reports on potential biases in algorithmic models.

Algorithmic Auditing Frameworks

Several frameworks and guidelines for algorithmic auditing have emerged, providing structured approaches to assess fairness. These frameworks often involve a multi-stage process, including defining fairness goals, identifying potential biases, measuring fairness metrics, and implementing mitigation strategies. While some frameworks are designed for large organizations, SMBs can adapt and simplify them to suit their needs and resources. For example, they can adopt a simplified auditing process that involves regular reviews of algorithmic outcomes, data quality checks, and stakeholder consultations.

These tools and frameworks empower SMBs to move beyond basic fairness assessments and conduct more in-depth analyses. They facilitate the calculation of relevant fairness metrics, the identification of bias sources, and the evaluation of mitigation strategies. However, it is crucial to remember that tools are only enablers. The real value lies in the thoughtful application of these tools and the commitment to embedding fairness into the algorithmic design Meaning ● Algorithmic Design for SMBs is strategically using automation and data to transform operations, create value, and gain a competitive edge. and deployment process.

Consider a small fintech SMB using an algorithm to assess creditworthiness for loan applications. Using a fairness library, they can calculate disparate impact ratios for different demographic groups and identify if their algorithm is disproportionately denying loans to certain communities. They can then explore techniques to mitigate this bias, such as re-weighting training data or adjusting algorithmic parameters, and re-evaluate fairness metrics to ensure improvements.

Integrating Fairness into the Algorithmic Lifecycle

Measuring algorithmic fairness should not be a one-off exercise. It needs to be integrated into the entire algorithmic lifecycle, from design and development to deployment and monitoring. This proactive approach ensures that fairness considerations are embedded from the outset, rather than being treated as an afterthought.

Key stages in integrating fairness into the algorithmic lifecycle for SMBs include:

1. Fairness Goal Definition ● Clearly define what fairness means in the specific business context. What are the potential harms of algorithmic bias? Which fairness concepts are most relevant? What are the business values and ethical principles that should guide algorithmic design?
2. Data Assessment and Preprocessing ● Evaluate the training data for potential biases. Are there imbalances in representation? Are there proxy variables that could lead to discrimination? Implement data preprocessing techniques to mitigate biases, such as re-sampling or data augmentation.
3. Algorithmic Design and Development ● Consider fairness-aware algorithmic design techniques. Are there algorithmic choices that could exacerbate bias? Explore algorithms that are inherently more robust to bias or that incorporate fairness constraints during training.
4. Fairness Evaluation and Measurement ● Regularly measure fairness metrics throughout the development process. Use appropriate tools and frameworks to assess fairness at different stages. Identify and address any fairness issues that arise.
5. Deployment and Monitoring ● Continuously monitor algorithmic performance and fairness metrics in real-world deployment. Establish feedback loops to detect and address any unexpected biases or fairness drifts over time. Regularly re-evaluate fairness goals and metrics as business needs and societal context evolve.

By embedding fairness into each stage of the algorithmic lifecycle, SMBs can build more responsible and equitable algorithmic systems. This proactive approach not only mitigates fairness risks but also fosters trust, enhances reputation, and contributes to long-term business sustainability.

For a small HR tech SMB developing an algorithmic resume screening tool, integrating fairness into the lifecycle would involve defining fairness goals (e.g., equal opportunity for qualified candidates from all backgrounds), assessing resume data for potential biases (e.g., gendered language patterns), choosing algorithms less prone to bias, regularly evaluating fairness metrics during development, and continuously monitoring the tool’s performance in real-world hiring processes.

In conclusion, intermediate-level algorithmic fairness measurement for SMBs requires a deeper understanding of fairness concepts, the use of more sophisticated tools and frameworks, and the integration of fairness considerations into the entire algorithmic lifecycle. This transition from basic awareness to proactive fairness management is crucial for SMBs seeking to harness the power of algorithms responsibly and ethically, ensuring that automation contributes to a more equitable and inclusive business environment.

Moving to intermediate algorithmic fairness involves proactive integration into the lifecycle, ensuring fairness is a core design principle, not just a post-deployment check.

Advanced

The progression from basic awareness to intermediate engagement with algorithmic fairness marks a significant step for SMBs. However, truly embedding fairness into the fabric of algorithmic operations necessitates an advanced perspective, one that transcends metric-driven assessments and delves into the systemic and strategic implications of algorithmic decision-making. At this level, SMBs must confront the inherent complexities of fairness itself, acknowledging that it is not a static concept but a dynamic and context-dependent construct.

Consider a rapidly scaling SaaS SMB employing AI-powered customer support chatbots. While initial fairness assessments might focus on demographic parity in response times, a more advanced analysis must consider intersectional fairness, recognizing that biases can compound for individuals belonging to multiple marginalized groups (e.g., elderly women from minority ethnic backgrounds) and that seemingly fair average response times might mask significant disparities in service quality for these specific segments.

Navigating the Complexities of Fairness Definitions

Advanced algorithmic fairness measurement acknowledges that no single definition of fairness universally applies. The choice of fairness metric is inherently value-laden and depends on the specific ethical and societal context. SMBs operating at this level must engage in critical reflection on different fairness definitions and their trade-offs, recognizing that optimizing for one type of fairness might inadvertently compromise another. This necessitates a shift from a purely technical approach to a more socio-technical perspective, one that integrates ethical considerations, stakeholder values, and domain-specific knowledge into the fairness assessment process.

Key advanced fairness concepts and considerations for SMBs include:

Intersectional Fairness

This recognizes that individuals belong to multiple social groups and that biases can be compounded at the intersection of these identities. Traditional fairness metrics, which often focus on single demographic attributes (e.g., race or gender in isolation), can fail to capture these intersectional biases. Advanced fairness analysis requires metrics and techniques that explicitly account for intersectionality, such as Group-Wise Fairness Metrics calculated for all combinations of protected attributes or Causal Inference Methods to disentangle the complex interplay of factors contributing to unfair outcomes.

Counterfactual Fairness

This definition of fairness focuses on causal reasoning and asks ● “Would the outcome be different if the individual belonged to a different protected group, holding all other factors constant?” Counterfactual fairness aims to eliminate discrimination based on protected attributes by ensuring that algorithmic decisions are not causally influenced by these attributes. Measuring counterfactual fairness often involves Causal Modeling Techniques and Sensitivity Analysis to assess the extent to which protected attributes causally impact algorithmic outcomes.

Procedural Fairness

This shifts the focus from outcomes to the processes and procedures used to design and deploy algorithms. Procedural fairness emphasizes transparency, accountability, and explainability in algorithmic decision-making. It involves establishing clear guidelines and protocols for algorithmic development, ensuring human oversight and review mechanisms, and providing avenues for redress and appeals for individuals affected by algorithmic decisions. Measuring procedural fairness is less about quantitative metrics and more about qualitative assessments of organizational processes and governance structures.

Dynamic and Contextual Fairness

Fairness is not a static property of an algorithm but a dynamic concept that evolves over time and varies across contexts. Advanced fairness measurement recognizes this dynamism and context-dependence. It involves continuous monitoring of fairness metrics in real-world deployment, adapting fairness goals and metrics as societal norms and business contexts change, and considering the specific domain and application when evaluating fairness. This requires a flexible and adaptive approach to fairness assessment, rather than relying on fixed metrics and thresholds.

Advanced algorithmic fairness for SMBs involves grappling with the inherent complexities of fairness definitions, moving beyond single metrics to encompass intersectional, counterfactual, procedural, and dynamic perspectives.

Strategic Business Integration of Algorithmic Fairness

At an advanced level, algorithmic fairness is not merely a risk mitigation exercise but a strategic business imperative. SMBs that proactively embrace algorithmic fairness can gain a competitive advantage, enhance their brand reputation, and foster long-term customer loyalty. Integrating fairness strategically requires embedding it into the core business strategy, aligning algorithmic design with organizational values, and leveraging fairness as a differentiator in the marketplace.

Key strategic integration points for algorithmic fairness in SMBs include:

Fairness-Driven Innovation

Instead of viewing fairness as a constraint, SMBs can leverage it as a driver of innovation. Designing algorithms with fairness in mind from the outset can lead to more robust, reliable, and ethically sound systems. This can also open up new market opportunities and attract customers who value ethical and responsible AI. For example, an SMB developing AI-powered financial products could differentiate itself by explicitly designing algorithms that promote financial inclusion and reduce bias in lending decisions.

Transparency and Explainability as Competitive Advantages

In an increasingly algorithm-driven world, transparency Meaning ● Operating openly and honestly to build trust and drive sustainable SMB growth. and explainability are becoming crucial for building trust with customers and stakeholders. SMBs that prioritize algorithmic transparency and explainability can gain a competitive edge by demonstrating their commitment to responsible AI. This could involve providing clear explanations of how algorithms work, disclosing fairness metrics, and offering avenues for customers to understand and challenge algorithmic decisions. This transparency can be particularly valuable in sectors where trust is paramount, such as healthcare, finance, and education.

Fairness as a Brand Value and Reputation Enhancer

In today’s socially conscious marketplace, consumers are increasingly demanding that businesses operate ethically and responsibly. SMBs that actively promote algorithmic fairness can enhance their brand reputation and attract customers who align with these values. Communicating a commitment to fairness through marketing materials, public statements, and corporate social responsibility initiatives can strengthen brand loyalty and attract socially conscious investors and partners. This can be particularly impactful for SMBs that target younger, more ethically minded demographics.

Long-Term Sustainability and Risk Mitigation

While addressing algorithmic fairness might require upfront investment, it can contribute to long-term business sustainability and risk mitigation. Fairer algorithms are less likely to generate negative publicity, legal challenges, or customer backlash. By proactively addressing fairness issues, SMBs can avoid costly reputational damage, regulatory fines, and erosion of customer trust. This long-term perspective is crucial for SMBs seeking to build resilient and sustainable businesses in the age of AI.

For a small e-learning SMB using AI to personalize learning paths, strategically integrating fairness could involve designing algorithms that actively promote equitable learning outcomes for students from diverse backgrounds, transparently communicating fairness metrics to students and educators, and positioning fairness as a core brand value to attract ethically conscious customers and educational institutions.

Advanced Tools and Methodologies for Deep Analysis

Advanced algorithmic fairness measurement often requires employing more sophisticated tools and methodologies, drawing upon research from fields like causal inference, algorithmic auditing, and social sciences. While SMBs may not have the resources to conduct cutting-edge research, they can leverage existing tools and methodologies to deepen their fairness analysis:

Causal Inference Techniques

Techniques like Do-Calculus and Instrumental Variables can be used to estimate causal effects and measure counterfactual fairness. These methods can help SMBs understand the extent to which protected attributes causally influence algorithmic outcomes and identify potential sources of discrimination. While requiring statistical expertise, readily available software packages and online resources can make these techniques more accessible.

Algorithmic Auditing Methodologies

Advanced auditing methodologies go beyond simple metric calculation and involve more in-depth investigations of algorithmic behavior. This can include Sensitivity Analysis to assess how algorithmic outcomes change under different input conditions, Adversarial Testing to identify vulnerabilities and biases, and Qualitative Audits involving stakeholder interviews and expert reviews to understand the social and ethical implications of algorithms. SMBs can adapt these methodologies to their context, focusing on areas of highest risk and impact.

Fairness-Aware Machine Learning Algorithms

Beyond measuring fairness, advanced approaches also involve developing and deploying algorithms that are explicitly designed to be fair. This includes techniques like Adversarial Debiasing, Fairness Constraints, and Representation Learning, which aim to mitigate bias during the algorithm training process. While these techniques are technically complex, pre-trained fairness-aware models and cloud-based AI fairness services are becoming increasingly available, making them more accessible to SMBs.

Ethical Frameworks and Stakeholder Engagement

Advanced fairness analysis is not solely a technical endeavor but also an ethical and social one. SMBs should adopt ethical frameworks for algorithmic development and engage with stakeholders (including customers, employees, and community groups) to understand their fairness concerns and values. This participatory approach can help ensure that fairness goals and metrics are aligned with societal expectations and that algorithmic systems are developed and deployed in a responsible and ethical manner. This engagement can take various forms, from customer surveys and focus groups to advisory boards and public consultations.

By employing these advanced tools and methodologies, SMBs can move beyond superficial fairness assessments and conduct deeper, more nuanced analyses of their algorithmic systems. This deeper understanding is crucial for navigating the complexities of fairness, mitigating potential harms, and leveraging fairness as a strategic business asset.

For a small healthcare tech SMB using AI to diagnose medical conditions, advanced fairness analysis could involve using causal inference to assess whether diagnostic accuracy differs across demographic groups due to factors beyond underlying health conditions, conducting adversarial testing to identify potential biases in the diagnostic algorithm, and engaging with patient advocacy groups to understand their fairness concerns and incorporate ethical considerations into the algorithm design process.

In conclusion, advanced algorithmic fairness measurement for SMBs represents a paradigm shift from reactive risk mitigation to proactive strategic integration. It requires grappling with the inherent complexities of fairness definitions, embedding fairness into the core business strategy, and employing sophisticated tools and methodologies for deep analysis. SMBs that embrace this advanced perspective can not only navigate the ethical challenges of algorithmic decision-making but also unlock new opportunities for innovation, competitive advantage, and long-term sustainable growth in an increasingly AI-driven world. It is about recognizing that fairness is not a technical problem to be solved but a fundamental business value to be cultivated and championed.

Advanced algorithmic fairness transcends metrics, becoming a strategic business value, driving innovation, enhancing reputation, and ensuring long-term sustainability in the AI era.

References

• Barocas, Solon, and Andrew D. Selbst. “Big data’s disparate impact.” California Law Review, vol. 104, no. 3, 2016, pp. 671-732.
• Holstein, Klaus, et al. “Improving fairness in machine learning.” Advances in Neural Information Processing Systems, vol. 32, 2019, pp. 11591-11601.
• Mehrabi, Ninareh, et al. “A survey on bias and fairness in machine learning.” ACM Computing Surveys (CSUR), vol. 54, no. 6, 2021, pp. 1-35.
• Mitchell, Margaret, et al. “Model cards for model reporting.” Proceedings of the Conference on Fairness, Accountability, and Transparency, 2019, pp. 220-229.
• Selbst, Andrew D., et al. “Fairness and abstraction in sociotechnical systems.” Proceedings of the Conference on Fairness, Accountability, and Transparency, 2019, pp. 59-68.