Fundamentals

Ninety percent of new jobs require digital skills, yet nearly half of small businesses still operate without a website, a digital storefront in today’s marketplace. This gap isn’t just a matter of outdated technology; it signals a deeper misunderstanding of how automation truly impacts a business’s bottom line and future trajectory. For many small and medium-sized businesses (SMBs), automation feels like a distant, corporate concept, something reserved for sprawling factories or tech giants. They see robots on assembly lines or complex software streamlining multinational operations, but struggle to connect these advancements to their own Main Street enterprises.

This disconnect is understandable, yet profoundly limiting. Automation, in its essence, is about efficiency, scalability, and strategic growth ● principles that are not exclusive to large corporations. They are, in fact, vital for the survival and prosperity of SMBs in an increasingly competitive landscape. The challenge, however, lies in accurately measuring the impact of automation.

Simply implementing new software or machinery and hoping for the best is a gamble, not a strategy. To truly understand if automation efforts are paying off, SMBs need to move beyond simple observation and embrace a more rigorous approach ● causal inference.

Beyond Correlation Understanding True Impact

Imagine a local bakery deciding to automate its order-taking process with a new online system. Sales increase after implementation. Is this because of the new system? Many might assume so.

Sales figures went up, automation was introduced, therefore, automation caused the sales increase. This is correlation, and it’s a dangerous trap in business decision-making. Correlation merely indicates a relationship between two things; it doesn’t prove that one caused the other. Perhaps the sales increase was due to a highly-rated food blog featuring the bakery the same week the new system launched.

Maybe a competitor closed down, diverting customers. Or perhaps seasonal demand for baked goods simply increased. Without understanding the true cause, the bakery might incorrectly attribute success solely to automation and make further investments based on flawed assumptions. This is where causal inference Meaning ● Causal Inference, within the context of SMB growth strategies, signifies determining the real cause-and-effect relationships behind business outcomes, rather than mere correlations. steps in.

It’s not about just noticing that sales went up after automation; it’s about rigorously determining if and how much the automation caused that increase. It’s about separating the signal of automation’s impact from the noise of other factors influencing business outcomes.

Causal inference is the compass that guides SMBs through the automation maze, ensuring they invest wisely and measure impact accurately, not just assume it.

Why Simple Metrics Fall Short

Many SMBs rely on readily available metrics to assess automation impact Meaning ● Automation Impact: SMB transformation through tech, reshaping operations, competition, and work, demanding strategic, ethical, future-focused approaches. ● time saved, cost reduction, or perhaps a simple pre-and-post comparison of sales figures. These metrics are not inherently bad, but they are insufficient for truly understanding causal relationships. Consider time saved. Automation might reduce the time spent on a task, like data entry.

This seems like a clear win. However, what if the time saved is not reinvested productively? What if employees, freed from data entry, simply spend more time on social media? The potential benefit of time saved is there, but the actual impact on business performance remains unclear without examining the causal chain.

Similarly, cost reduction Meaning ● Cost Reduction, in the context of Small and Medium-sized Businesses, signifies a proactive and sustained business strategy focused on minimizing expenditures while maintaining or improving operational efficiency and profitability. metrics can be misleading. Automation might lower labor costs, but increase expenses in other areas, such as software maintenance, training, or unexpected system failures. A simple cost comparison might show a net reduction, but fail to capture hidden costs or long-term implications. Pre-and-post comparisons are perhaps the most common, and most flawed, approach.

Comparing sales or efficiency before and after automation implementation Meaning ● Strategic integration of tech to boost SMB efficiency, growth, and competitiveness. is susceptible to all the confounding factors mentioned earlier ● market changes, competitor actions, seasonal variations, and countless other external influences. These simple metrics, while easy to track, offer a superficial view. They tell a story of what happened, but not why. Causal inference, on the other hand, seeks to uncover the ‘why’, providing a deeper, more reliable understanding of automation’s true contribution.

The Core of Causal Inference ● Isolating Automation’s Effect

At its heart, causal inference is about isolating the specific effect of automation from all other factors. It’s about answering the question ● “What would have happened if we hadn’t implemented automation?”. This is a counterfactual question, something we can’t directly observe. We can only observe what did happen with automation in place.

Therefore, causal inference relies on methods to create a credible comparison group ● a ‘control’ group ● that represents what would have happened in the absence of automation. In a perfect scientific experiment, we could randomly assign some SMBs to adopt automation (the ‘treatment’ group) and others to remain without (the ‘control’ group). Random assignment helps ensure that the two groups are statistically similar in all relevant aspects, except for automation. Then, any difference in outcomes between the groups can be more confidently attributed to automation.

However, real-world business settings rarely allow for such controlled experiments. SMBs cannot be randomly assigned to automation adoption Meaning ● SMB Automation Adoption: Strategic tech integration to boost efficiency, innovation, & ethical growth. or non-adoption. They make their own decisions based on their unique circumstances. Therefore, causal inference in business relies on quasi-experimental methods ● techniques that mimic the logic of randomized experiments using observational data.

These methods attempt to create statistical ‘twins’ ● businesses that are as similar as possible, except for their automation choices ● to estimate the causal effect of automation. This might involve comparing businesses in similar industries, of similar size, and in similar markets, but with differing levels of automation adoption. The goal is to create a comparison that is as close as possible to a true experiment, allowing for more reliable conclusions about causation.

Practical Steps for SMBs ● Starting with Causal Thinking

Embracing causal inference doesn’t require SMBs to become statistical experts overnight. It starts with adopting a causal mindset. This means questioning assumptions, looking beyond simple correlations, and actively seeking to understand the ‘why’ behind business outcomes. Here are some practical first steps for SMBs:

• Define Clear Objectives ● Before implementing any automation, clearly define what you hope to achieve. Is it to increase sales, reduce costs, improve customer satisfaction, or something else? Specific, measurable objectives are crucial for later evaluating impact.
• Identify Potential Confounding Factors ● Brainstorm other factors that could influence your objectives besides automation. Think about market trends, competitor actions, seasonal effects, marketing campaigns, and internal changes within your business.
• Gather Relevant Data ● Collect data not only on your key metrics (sales, costs, etc.) but also on potential confounding factors. The more data you have, the better you can control for these factors in your analysis.
• Look for Natural Experiments ● Sometimes, real-world events create ‘natural experiments’ that can be leveraged for causal inference. For example, if you have multiple locations and roll out automation in only some, the non-automated locations can serve as a comparison group, provided they are sufficiently similar.
• Start Small and Iterate ● Don’t try to automate everything at once. Begin with a pilot project in one area of your business. This allows you to test and learn, measure impact more effectively, and refine your approach before wider implementation.

These steps are about building a foundation for causal thinking. For more rigorous analysis, SMBs might need to seek expert help, but even a basic understanding of causal inference principles can significantly improve their automation decision-making and impact measurement. By moving beyond simple correlations and embracing a causal approach, SMBs can unlock the true potential of automation to drive sustainable growth and success.

Understanding causation in automation isn’t just about data; it’s about smart business strategy, ensuring every tech investment pulls its weight and propels real growth.

Intermediate

The siren song of automation promises efficiency gains and streamlined operations, a tempting melody for SMBs navigating competitive waters. Yet, according to a recent study by McKinsey, nearly 70% of automation initiatives fail to achieve their intended outcomes. This high failure rate isn’t necessarily due to flawed technology, but often stems from a fundamental misstep ● inadequate measurement of automation’s actual impact. Many SMBs, while moving beyond rudimentary pre-post comparisons, still rely on correlational metrics, mistaking association for causation.

They observe improvements after automation implementation and assume a direct link, overlooking the complex interplay of variables that truly drive business performance. For intermediate-level analysis, SMBs must graduate to more sophisticated causal inference techniques. This transition involves understanding the limitations of correlation, appreciating the nuances of confounding variables, and employing methodologies that can disentangle automation’s specific contribution from the background noise of market dynamics and operational fluctuations.

Confounding Variables ● The Hidden Saboteurs of Accurate Measurement

Confounding variables represent a significant challenge to accurately measuring automation impact. These are factors that are related to both the adoption of automation and the outcome being measured, creating spurious correlations. Consider an SMB implementing a new CRM system alongside a revamped marketing campaign. If sales increase, is it the CRM, the marketing, or a synergistic effect?

Without careful analysis, it’s impossible to discern the individual contributions. The marketing campaign, in this case, is a confounding variable. It’s associated with both the CRM implementation (both initiatives launched concurrently) and the sales increase (marketing efforts directly aim to boost sales). Failing to account for confounding variables leads to inflated or deflated estimates of automation’s true effect.

For instance, an SMB might overestimate the CRM’s impact if the sales surge is primarily driven by the marketing campaign. Conversely, they might underestimate the CRM’s value if a simultaneous economic downturn dampens sales, masking the CRM’s positive contribution. Identifying and addressing confounding variables is paramount for robust causal inference. This requires a deep understanding of the business context, potential external influences, and the mechanisms through which automation is expected to drive change. It’s about anticipating the ‘hidden saboteurs’ that can distort the measurement of automation’s true impact.

Regression Analysis ● A Tool for Isolating Causal Effects

Regression analysis offers a powerful statistical framework for controlling for confounding variables and isolating the causal effect of automation. In its simplest form, linear regression models the relationship between an outcome variable (e.g., sales) and one or more predictor variables (e.g., automation adoption, marketing spend, seasonality). By including potential confounders as predictor variables in the regression model, we can statistically ‘hold them constant’ and estimate the independent effect of automation. For example, in the CRM and marketing campaign scenario, we can build a regression model with sales as the outcome variable and both CRM adoption and marketing spend as predictors.

The regression coefficient for CRM adoption, after controlling for marketing spend, provides a more accurate estimate of the CRM’s causal effect on sales. Regression analysis Meaning ● Regression Analysis, a statistical methodology vital for SMBs, facilitates the understanding of relationships between variables to predict outcomes. can also accommodate more complex relationships, such as non-linear effects or interactions between variables. For instance, the impact of automation might be different for SMBs of different sizes or in different industries. Interaction terms in regression models can capture these heterogeneous effects.

Furthermore, regression analysis can be extended to panel data, which tracks multiple SMBs over time. This allows for controlling for time-invariant unobserved confounders ● factors that are constant over time within each SMB but vary across SMBs and are correlated with both automation adoption and outcomes. While regression analysis is a valuable tool, it’s crucial to recognize its limitations. It relies on assumptions, such as linearity and no omitted variable bias.

Violations of these assumptions can lead to biased causal estimates. Therefore, careful model specification, diagnostic checks, and sensitivity analyses are essential for ensuring the validity of regression-based causal inference.

Regression analysis isn’t just number crunching; it’s strategic surgery, dissecting data to reveal automation’s precise impact, cutting through the noise of business complexity.

Propensity Score Matching ● Mimicking Randomization in Observational Studies

Propensity score matching (PSM) is another widely used technique for causal inference in observational studies, particularly when randomization is not feasible. PSM aims to create a ‘treatment’ group (SMBs that adopted automation) and a ‘control’ group (SMBs that did not) that are as similar as possible in terms of observed characteristics that might influence both automation adoption and outcomes. The ‘propensity score’ is the estimated probability of adopting automation, based on a set of pre-treatment covariates (confounding variables). SMBs are then matched based on their propensity scores, creating pairs or groups of SMBs that are similar in terms of their likelihood of automation adoption.

By comparing outcomes between matched treated and control SMBs, we can estimate the average treatment effect of automation. PSM is particularly useful when dealing with high-dimensional confounding ● many potential confounders. It reduces the dimensionality by summarizing all pre-treatment covariates into a single propensity score. However, PSM relies on the ‘conditional independence assumption’ ● that, conditional on the observed covariates, automation adoption is as-if random.

This assumption is untestable and can be violated if there are unobserved confounders. Furthermore, PSM only balances observed covariates; it does not address imbalances in unobserved factors. Therefore, careful selection of covariates, sensitivity analyses to unobserved confounding, and consideration of alternative methods are crucial when using PSM for causal inference. Despite these limitations, PSM offers a valuable approach for mimicking randomization in observational settings, allowing SMBs to draw more credible causal inferences about automation impact.

Difference-In-Differences ● Leveraging Panel Data for Causal Insights

Difference-in-differences (DID) is a powerful causal inference technique specifically designed for panel data, exploiting the time dimension to estimate treatment effects. DID is particularly useful when automation adoption occurs at different times across different SMBs, creating a staggered rollout. The core idea of DID is to compare the change in outcomes over time between the ‘treatment’ group (SMBs that adopted automation) and the ‘control’ group (SMBs that did not), before and after automation adoption. By differencing twice ● once across time and once across groups ● DID effectively removes the effects of time-invariant confounders (factors that are constant over time) and common time trends (factors that affect both groups similarly over time).

For example, consider a chain of retail stores implementing a new inventory management system in some stores but not others. DID can estimate the causal impact of the new system on sales by comparing the change in sales in the treated stores relative to the change in sales in the control stores, before and after implementation. DID relies on the ‘parallel trends assumption’ ● that, in the absence of automation, the treatment and control groups would have followed parallel trends in outcomes. This assumption is crucial for the validity of DID and should be carefully examined.

Visual inspection of pre-treatment trends, placebo tests, and robustness checks are common approaches for assessing the plausibility of the parallel trends assumption. DID is a widely applicable and intuitive method for causal inference in panel data settings, offering SMBs a robust tool for measuring the dynamic impact of automation over time.

Moving Towards Causal Maturity ● Strategic Data Infrastructure

Adopting intermediate-level causal inference techniques requires SMBs to invest in a more strategic data Meaning ● Strategic Data, for Small and Medium-sized Businesses (SMBs), refers to the carefully selected and managed data assets that directly inform key strategic decisions related to growth, automation, and efficient implementation of business initiatives. infrastructure. This includes not only collecting data on key performance indicators (KPIs) but also systematically gathering data on potential confounding variables, automation implementation timelines, and relevant contextual factors. Data quality is paramount. Accurate, consistent, and timely data are essential for reliable causal inference.

SMBs may need to improve their data collection processes, data storage systems, and data management practices. Furthermore, building analytical capabilities is crucial. This might involve training existing staff in data analysis techniques, hiring data analysts, or partnering with external consultants. The goal is to develop the in-house expertise to apply causal inference methods effectively and interpret the results in a business context.

Investing in data infrastructure Meaning ● Data Infrastructure, in the context of SMB growth, automation, and implementation, constitutes the foundational framework for managing and utilizing data assets, enabling informed decision-making. and analytical capabilities is not just a technical upgrade; it’s a strategic investment in decision-making. It empowers SMBs to move beyond guesswork and intuition, making data-driven decisions about automation that are grounded in robust causal evidence. This transition towards causal maturity is a journey, not a destination. It requires ongoing learning, adaptation, and a commitment to data-driven decision-making at all levels of the organization. For SMBs serious about maximizing the return on their automation investments, embracing intermediate-level causal inference is not merely an option; it’s a strategic imperative.

Strategic data infrastructure isn’t just about servers and software; it’s the foundation for causal clarity, empowering SMBs to see automation’s true ROI and steer towards data-driven success.

Advanced

Automation, heralded as a panacea for productivity woes, often presents a paradox for sophisticated SMBs. While the theoretical benefits are compelling, empirical evidence of widespread, unequivocally positive impact remains surprisingly elusive. A Harvard Business Review study highlighted that while automation can boost productivity at the task level, its broader organizational impact on profitability and sustainable growth is far from guaranteed, often contingent on complex contextual factors and strategic implementation. For advanced SMBs, the challenge transcends mere measurement of automation’s effect; it necessitates a deep, critical examination of the causal mechanisms through which automation influences business outcomes.

Moving beyond simple ‘what works’ analysis to ‘why it works, for whom, and under what conditions’ demands advanced causal inference methodologies, coupled with a nuanced understanding of organizational dynamics and market complexities. This advanced perspective acknowledges that automation is not a monolithic entity with uniform effects, but rather a heterogeneous set of technologies whose impact is shaped by intricate interactions with business strategy, organizational structure, workforce skills, and the broader competitive landscape.

Causal Mechanisms ● Unpacking the Black Box of Automation Impact

Advanced causal inference delves into the ‘black box’ of automation impact, seeking to uncover the specific causal mechanisms that mediate the relationship between automation and business outcomes. Understanding these mechanisms is crucial for optimizing automation strategies Meaning ● Automation Strategies, within the context of Small and Medium-sized Businesses (SMBs), represent a coordinated approach to integrating technology and software solutions to streamline business processes. and predicting their effects in different contexts. For example, automation might improve customer service Meaning ● Customer service, within the context of SMB growth, involves providing assistance and support to customers before, during, and after a purchase, a vital function for business survival. response times (a mediating mechanism), which in turn leads to increased customer satisfaction Meaning ● Customer Satisfaction: Ensuring customer delight by consistently meeting and exceeding expectations, fostering loyalty and advocacy. and loyalty (the ultimate outcome). However, if automation implementation is poorly managed and leads to system glitches or impersonal interactions, it might worsen customer service response times and negatively impact customer satisfaction.

Simply measuring the overall effect of automation on customer satisfaction without understanding the mediating mechanism of response times provides an incomplete and potentially misleading picture. Mechanism analysis involves identifying and empirically testing these intermediate steps in the causal chain. This can be achieved through mediation analysis techniques, which statistically decompose the total effect of automation into direct and indirect effects, the indirect effects operating through specific mediating variables. Furthermore, understanding causal mechanisms requires qualitative research methods, such as case studies and process tracing, to complement quantitative analysis.

In-depth case studies of SMBs that have successfully or unsuccessfully implemented automation can reveal rich insights into the organizational processes, managerial decisions, and contextual factors that shape automation’s impact. Process tracing, a method for systematically examining the sequence of events and decisions leading to an outcome, can help to identify the critical junctures and causal pathways through which automation exerts its influence. By combining quantitative and qualitative methods to unpack the black box of causal mechanisms, advanced SMBs can gain a more granular and actionable understanding of automation’s true impact.

Heterogeneous Treatment Effects ● Recognizing Contextual Contingencies

The assumption of uniform treatment effects ● that automation has the same impact across all SMBs ● is often unrealistic. Advanced causal inference recognizes the importance of heterogeneous treatment effects, acknowledging that automation’s impact can vary significantly depending on SMB characteristics, industry context, and implementation strategies. For example, the impact of robotic process automation (RPA) on operational efficiency might be greater for SMBs in highly regulated industries with standardized processes compared to those in creative industries with more customized workflows. Similarly, the effect of AI-powered customer service chatbots on customer satisfaction might differ depending on the complexity of customer inquiries, the quality of chatbot training data, and the level of human oversight.

Identifying and understanding these heterogeneous effects is crucial for tailoring automation strategies to specific SMB needs and maximizing their effectiveness. Subgroup analysis is a common approach for investigating heterogeneous treatment effects. This involves estimating the causal effect of automation separately for different subgroups of SMBs, defined by relevant characteristics such as size, industry, technology adoption level, or organizational culture. 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, such as causal forests, can also be used to identify complex patterns of heterogeneity and predict treatment effects for individual SMBs based on their specific characteristics.

However, subgroup analysis and machine learning approaches should be applied cautiously, avoiding data dredging and spurious findings. Pre-specification of relevant subgroups based on theoretical considerations and business domain knowledge is essential for robust and interpretable results. By embracing the concept of heterogeneous treatment effects, advanced SMBs can move beyond one-size-fits-all automation solutions and adopt a more nuanced, context-aware approach to technology implementation.

Advanced causal inference isn’t just about proving automation works; it’s about dissecting how and why it works differently for each SMB, crafting bespoke strategies for maximum impact.

Instrumental Variables ● Addressing Endogeneity and Unobserved Confounding

Endogeneity, the problem of reverse causality or correlation between the treatment variable (automation adoption) and the error term in a regression model, poses a significant challenge to causal inference. Unobserved confounding, where unmeasured factors influence both automation adoption and outcomes, further complicates the picture. Instrumental variables (IV) regression is a powerful technique for addressing endogeneity and unobserved confounding, providing more credible causal estimates. An instrumental variable is a variable that is correlated with automation adoption (the treatment) but uncorrelated with the outcome variable, except through its effect on automation.

Finding valid instrumental variables is often challenging and requires deep domain knowledge and careful consideration of potential instruments. For example, government subsidies for automation adoption might serve as an instrument if they influence SMBs’ automation decisions but do not directly affect business outcomes other than through automation. However, the validity of this instrument depends on assumptions that must be carefully scrutinized. IV regression uses the instrumental variable to isolate the exogenous variation in automation adoption ● the part of automation adoption that is driven by the instrument and not by other factors, including unobserved confounders.

By focusing on this exogenous variation, IV regression can provide a more unbiased estimate of the causal effect of automation. However, IV regression has limitations. It relies on strong assumptions about the validity of the instrument, which are often difficult to verify. Weak instruments, instruments that are only weakly correlated with automation adoption, can lead to biased and imprecise estimates.

Furthermore, IV regression typically estimates the local average treatment effect (LATE) ● the causal effect of automation for SMBs whose adoption decisions are influenced by the instrument. This LATE might not be generalizable to all SMBs. Despite these limitations, IV regression offers a valuable approach for addressing endogeneity and unobserved confounding, providing a more rigorous foundation for causal inference in complex business settings.

Causal Machine Learning ● Integrating Prediction and Explanation

Causal machine learning (CML) represents a frontier in advanced causal inference, integrating the predictive power of machine learning with the explanatory goals of causal analysis. CML methods leverage machine learning algorithms to estimate causal effects in complex, high-dimensional data settings, addressing limitations of traditional statistical methods. For example, doubly robust estimation, a CML technique, combines machine learning models for both the treatment assignment process (propensity score) and the outcome model, achieving robustness to model misspecification. Causal forests, another CML method, extend random forests to estimate heterogeneous treatment effects, providing non-parametric estimates of causal effects that can adapt to complex non-linear relationships and interactions.

CML methods are particularly well-suited for analyzing large-scale business datasets, uncovering subtle causal patterns that might be missed by traditional approaches. However, CML is not a panacea. It inherits the limitations of both machine learning and causal inference. CML models can be ‘black boxes’, making it difficult to interpret the causal mechanisms underlying the estimated effects.

Furthermore, CML methods still rely on causal assumptions, such as ignorability or instrumental validity, which must be carefully considered and justified. Therefore, CML should be viewed as a complement to, not a replacement for, traditional causal inference methods. Integrating CML with domain expertise, qualitative research, and business intuition is crucial for extracting meaningful causal insights from complex business data. Advanced SMBs that embrace CML can gain a competitive edge by leveraging the power of machine learning to understand and optimize the causal impact of automation in increasingly data-rich and complex business environments.

Ethical Considerations in Causal Inference for Automation Impact

As advanced SMBs increasingly rely on causal inference to guide their automation strategies, ethical considerations become paramount. Causal inference, while a powerful tool for understanding and optimizing automation impact, can also be misused or misapplied, leading to unintended negative consequences. For example, if causal inference reveals that automation disproportionately displaces workers in certain demographic groups, SMBs have an ethical responsibility to address these distributional effects, perhaps through retraining programs or job creation initiatives in affected areas. Similarly, if causal inference indicates that AI-powered decision-making systems perpetuate or amplify existing biases in hiring or promotion, SMBs must take steps to mitigate these biases and ensure fairness and equity.

Transparency and accountability are crucial ethical principles in the application of causal inference. SMBs should be transparent about their causal inference methods, data sources, and assumptions, allowing for external scrutiny and validation. They should also be accountable for the causal claims they make and the decisions they base on causal inference results. Furthermore, ethical considerations extend to the very definition of ‘impact’.

Focusing solely on narrow economic metrics, such as profit maximization or cost reduction, might overlook broader societal impacts of automation, such as effects on worker well-being, community development, or environmental sustainability. Advanced SMBs should adopt a more holistic and ethically informed approach to measuring automation impact, considering a wider range of stakeholder interests and societal values. This requires ongoing dialogue with employees, customers, communities, and other stakeholders to ensure that automation strategies are not only economically beneficial but also ethically sound and socially responsible. The advanced frontier of causal inference for automation impact is not just about methodological sophistication; it’s about responsible innovation, guided by ethical principles and a commitment to creating shared value for all stakeholders.

Ethical causal inference isn’t just about accurate numbers; it’s about responsible automation, ensuring SMBs leverage data power with integrity and a commitment to equitable outcomes.

References

• Angrist, Joshua D., and Jörn-Steffen Pischke. Mostly Harmless Econometrics ● An Empiricist’s Companion. Princeton University Press, 2009.
• Imbens, Guido W., and Donald B. Rubin. Causal Inference for Statistics, Social, and Biomedical Sciences ● An Introduction. Cambridge University Press, 2015.
• Pearl, Judea, Madelyn Glymour, and Nicholas P. Jewell. Causal Inference in Statistics ● A Primer. John Wiley & Sons, 2016.