Causal Economic Machine Learning (CEML): “Human AI”

Abstract

This paper proposes causal economic machine learning (CEML) as a research agenda that utilizes causal machine learning (CML), built on causal economics (CE) decision theory. Causal economics is better suited for use in machine learning optimization than expected utility theory (EUT) and behavioral economics (BE) based on its central feature of causal coupling (CC), which models decisions as requiring upfront costs, some certain and some uncertain, in anticipation of future uncertain benefits that are linked by causation. This multi-period causal process, incorporating certainty and uncertainty, replaces the single-period lottery outcomes augmented with intertemporal discounting used in EUT and BE, providing a more realistic framework for AI machine learning modeling and real-world application. It is mathematically demonstrated that EUT and BE are constrained versions of CE. With the growing interest in natural experiments in statistics and causal machine learning (CML) across many fields, such as healthcare, economics, and business, there is a large potential opportunity to run AI models on CE foundations and compare results to models based on traditional decision-making models that focus only on rationality, bounded to various degrees. To be most effective, machine learning must mirror human reasoning as closely as possible, an alignment established through CEML, which represents an evolution to truly “human AI”. This paper maps out how the non-linear optimization required for the CEML structural response functions can be accomplished through Sequential Least Squares Programming (SLSQP) and applied to data sets through the S-Learner CML meta-algorithm. Upon this foundation, the next phase of research is to apply CEML to appropriate data sets in various areas of practice where causality and accurate modeling of human behavior are vital, such as precision healthcare, economic policy, and marketing.

1. Introduction

Artificial intelligence (AI) involves the creation of machines that are able to think and act like humans. An important goal is to ensure that these are safe for humans and are consistent with human values [1]. As a result, the AI alignment problem [2] is a central area of research in the science of AI. Training machines to think like human brains requires modeling machine learning in line with the way human brains make decisions in the real world [3]. The highly formalized human decision-making models of economics based on utility functions lend well to this purpose. In the science of AI, the counterparts to economic utility functions are known by various terms, including value functions, objective functions, loss functions, reward functions, and preference orderings [4].

Explicitly grounding AI in economic theory has been shown to yield improved results [5]. Many AI scientists strive to make their AI systems fully rational by utilizing rational choice theory from economics [6], specifically in the form of expected utility theory [4]. However, there is much evidence that human decision-makers are not 100% rational and calculating in practice [7]. It has also been well-documented that expected utility is not effective as an applied model of individual choice [8]. Comprehensive reviews of experimental research have demonstrated violations in each of the axioms of expected utility [9]. Behavioral economics [10] emerged as a powerful approach to decision-making heavily grounded in psychology, incorporating the evidence that humans are restricted to bounded rationality in practice [11]. The reality of decision-making reflects many challenges, including computation, information, and biases. Behavioral economics has been most robustly structured within the framework of cumulative prospect theory [12], and this has been a big leap forward in modeling. Today, much of AI modeling falls back on some form of bounded rationality [13].

These are significant enhancements in theoretical foundations, yet when it comes to real-world application in decision-making under uncertainty, economics and AI both face an environment with multiple individuals, human-agent collectives [4], and decisions that span multiple, sequential, causally linked time periods where information is costly [4]. AI is a time-based process, with causality from percepts to outcomes [14]. Given this context, the theoretical frameworks of EUT and BE typically have to be put aside in applied economics in favor of dynamic stochastic programming [15] based on a Markov decision process [16]. In a similar fashion, much of applied AI is focused on reinforcement learning [17], which is also sequential, based on a Markov decision process. The sequential nature of applied decision-making has also led to economists and AI scientists each modeling intelligent agents with Bayesian probability theory to inform their beliefs (priors), and utility theory to inform their preferences [4].

This complex environment renders expected utility and behavioral economics unable to serve as core models in real-world applications, primarily due to their reliance on single-period lotteries with intertemporal choice augmentations such as exponential discounting [18] and quasi-hyperbolic discounting [19]. These approaches do not explicitly define the causal relationship that uncertain benefits follow upfront certain and uncertain costs over time on significant decisions [20]. In contrast, the structure of causal economics does align with this real-world environment, as this paper will explore. Causal economics provides a consistent theoretical underpinning to applied modeling in economics and AI that utilizes dynamic stochastic programming and AI machine learning algorithms [4].

The theoretical foundations of causal economics align strongly with the current surge of interest in causal inference throughout econometrics, statistics, and computer science, represented in research on natural experiments [21] based on the potential outcomes approach [22], the Rubin causal model [23], and the recent development of causal machine learning [24] with its expanding set of literature [25]. This invigorated and cross-disciplinary interest in causality presents an opportunity for causal economic theory (CE), causal machine learning (CML), and the natural experiments approach, to be used together as a research agenda in causal economic machine learning (CEML).

This paper’s contribution is that it lays out the theoretical foundations of CEML, casting it as preferred to alternatives that do not utilize causal economics as a model of human behavior in conjunction with causal machine learning, and then delineating the relevant implementation tools for optimization, causal inference, and machine learning applications. The next phase of this research agenda is the identification of data sets that are capable of reflecting the structure of CEML and collaborating with innovative researchers to put it to the test against more traditional approaches, such as non-causal machine learning and the human decision models of expected utility or behavioral economics.

This paper begins by recapping causal economic decision theory, as many AI practitioners will not yet be well versed in it. The mathematical framework is then presented with a practical illustrative example. It is next shown that the prominent human economic decision models in use—expected utility and behavioral economics—are mathematically constrained versions of causal economics. Supporting research from neuroscience and psychology is presented to reinforce the relevance of the causal economic decision model in humans, which is then connected directly to AI and machine learning. Causal inference and causal machine learning are then introduced and connected with causal economics, to arrive at the intersection, causal economic machine learning (CEML). The remainder of the paper guides the reader through the approach and required tools to apply CEML in areas such as the private sector, public sector, healthcare, and tax policy in particular.

2. Theory

2.1. Causal Economics Theory

There have been many significant advances in decision theory across academic disciplines, but more work is needed to provide a consistent, unified foundational framework at the theoretical and empirical levels [26]. Causal economics attempts to address this gap by providing a robust framework well-suited to testing a wide range of algorithms that address issues in AI research. The theory is grounded in behavioral economics, psychology, biology, and neuroscience [20]. In the following section, it will be demonstrated that expected utility [27] and cumulative prospect theory [12] also represent constrained mathematical reductions of causal economics that limit applications. The richness of causal economics is possible due to the definition of outcomes. Outcomes in causal economics must always contain both costs and benefits, each containing certain (deliberate) and uncertain elements. Causation must also run in at least one direction of the relationship [4]. The approach to preferences used in causal economics is built upon the structural model of cumulative prospect theory [28]. With this foundation in place, preferences are expanded to include cost broadly (incorporating certain and uncertain components), including risk in the form of the uncertain component. Psychological trade-off constraints are then defined as the primary umbrella constraint for optimization, which includes budget constraints. The following summary briefly introduces the equations of the causal economic framework for those not familiar with it [4], beginning with key definitions:

i = an outcome where P incrementally results from Q and/or Q incrementally results from P (Q is required for P and/or visa versa)

t = time period

+, − = positive overall outcome and negative overall outcome, respectively

(+), (−) = personal total benefit and personal total cost outcome values, respectively

w = probability weighting function

v = value function

π = capacity function of individual events

Ψ = rank-dependent cumulative weighting of overall weighted event values

Ω = rank-dependent cumulative probability weighting

P_i = perceived magnitude of positive element in outcome i

Q_i = perceived magnitude of negative element in outcome i

p_i = perceived probability of positive element in outcome i

q_i = perceived probability of negative element in outcome i

P_Ci = acceptable relative magnitude of positive element in outcome i

Q_Ci = acceptable relative magnitude of negative element in outcome i

p_Ci = acceptable relative probability of positive element in outcome i

q_Ci = acceptable relative probability of negative element in outcome i

Ψ_C = acceptable rank-dependent cumulative weighting of overall weighted event values

Z_U = perceived magnitude of outcome i (used when positive and negative outcomes are not specified)

z_U = perceived probability of outcome i (used when positive and negative outcomes are not specified)

Optimization in causal economics involves maximization of the non-linear value function, v, shown in Equation (1):

$$\begin{array}{l}Max v\left(f\right)= {{\displaystyle \int }}_{t=0}^M\{{{\displaystyle \int }}_{i=0}^N [w^{+(+)}(p_{U i t}) v^{+(+)} (P_{U i t})+ v^{+(+)}\left(P_{A i t }\right)- w^{+(-)}(q_{U i t }) v^{+(-)}\left(Q_{U i t}\right)\\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}- v^{+(-)}(Q_{A i t })] {\phi }_i^{+}+ {{\displaystyle \int }}_{i=-k}^{-1}[w^{-(+)}(p_{U i t}) v^{-(+)} \left(P_{U i t}\right) + v^{-{(}_{+})}\left(P_{A i t}\right)\\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}- w^{-(-)}\left(q_{U i t}\right) v^{-(-)}\left(Q_{U i t}\right)- v^{-(-)} \left(Q_{A i t}\right)] {\phi }_i^{-} \}\end{array}$$

(1)

This maximization is subject to a pair of personal psychological trade-off constraints, defined by (2) and (3):

$$c_i={\displaystyle \frac{\left[ w^{(+)}\left(p_{U i t}^C \right)v^{(+)} \left(P_{U i t}^C\right)+v^{+}\left(P_{A i t}^C\right) \right]}{\left[ w^{(-)} (q_{U i t}^C\right) v^{(-)}(Q_{U i t}^C)+v^{(-)}\left(Q_{A i t}^C\right)]}} \times {\phi }_i^C$$

(2)

$$L=Max \left[ \left\{w^{(-)}\left(q_{U i t}^C\right) v^{(-)}\left(Q_{U i t}^C\right)+v^{(-)}\left(Q_{A i t}^C\right)\right\}, {\phi }_i^C\right]$$

(3)

Optimization in causal economics differs fundamentally from the approach utilized in traditional economics, mainly through the addition of these two non-linear psychological trade-off constraints and a causal coupling constraint introduced later in Equation (8). Decision-makers optimize by making trade-off decisions in light of their personal, internal psychological constraint of acceptable cost relative to causally coupled benefit [4]. External constraints, such as the decision-maker’s financial budget, impact utility indirectly through the personal interpretation of the decision-maker, based on that agent’s personal psychological constraint, which is a matter of personal choice based on their individual experiences [29].

In causal economic optimization, the value function, v, and the psychological trade-off constraint, c, play separate but complementary roles. The value function captures a decision-maker’s personal judgments regarding the particular set of prospects they believe they may face when making a decision and the specific costs and benefits they associate as realistically possible with each event. It incorporates preferences over both risk and time, but also represents perceived realities that the decision-maker feels they must take as given. This is contrasted with the psychological trade-off constraint, c, that directly conveys a decision-maker’s preferences across costs (uncertain and uncertain) as well as risk (uncertain). It conveys the marginal amount of benefit required for a decision-maker to take on one additional unit of cost, or similarly the marginal amount of cost they will bear in order to obtain an additional unit of benefit.

The resulting optimization dynamic is a process in which anticipated prospects are evaluated, via v, and compared to acceptable scenarios, via c, subject to L as a limit (at least in the short-term), as well as the causal coupling constraint that costs precede benefits, as shown subsequently in Equation (8). The latter enforces a causal structure on the model, reflecting not just preferences but also the reality of the environment in which they are applied.

2.2. A Real World Example

A powerful way to illustrate how causal economics is able to effectively model the complex decisions agents face in the real world, where traditional economic models cannot, is through a common example that most people can relate to—the pursuit of weight loss. A goal to lose weight requires the decision-maker to weigh known costs and risks in advance of anticipated resulting future benefits. It is certainly more complex than going to the gym and eating a healthy diet for a day and then immediately losing a few pounds. Losing weight actually requires motivating goals [30] as well as deliberate and sustained commitment to incurring challenging personal costs in the form of dietary sacrifices and physical exertion that both extend over months [4]. Neither the effort required nor the results obtained are primarily random lotteries, as would be utilized in traditional economic optimization [4].

Causal economics is able to effectively model this weight loss scenario because its framework builds in a multi-period structure, containing both certain (deliberate) and uncertain costs and benefits, that also requires costs to be causally matched to resulting benefits, both contemporaneously and in future time periods. This causal coupling means that costs (some certain and some uncertain) are explicitly and directly tied to the expected resulting benefits that occur over the entire decision time horizon. Costs are generally greater in earlier periods with relatively higher certainty as contrasted with potential benefits that can occur in later periods with relatively less certainty. In the weight loss example, costs that can be controlled to a large degree include exercise and diet. The associated causal gains involve weight loss, improved health, and positive self-esteem.

The complexity of modeling an individual’s decision to achieve a weight loss goal does not end with this matching of costs and benefits over the decision time frame. Decision-makers are all individuals with their own unique subjective preferences and subjective probability assessments, as captured by the subjective expected utility theory [31]. For example, an individual may pursue an exercise program in order to obtain their own perceptions of personal benefits, such as improved health and attractiveness, with full knowledge that exercising and eating well will allow them to achieve their goal with certainty, but only if they commit to the effort over time. There is uncertainty concerning the actual outcome quantities of individual parameters, but the decision-maker’s ability to deliberately decide and control the overall outcomes of their decisions (versus a lottery) is a large gap in expected utility and behavioral economics models [4]. In the real world, decision-makers do not face a series of random lotteries. They impact and are impacted by outcomes. Many popular expressions in the English language capture this observed reality, such as ‘getting out what you put in’, ‘no pain, no gain’, and ‘bearing the fruit of one’s labor’.

2.3. Reducing Causal Economics to Expected Utility Theory

Expected utility [32] can be directly derived from the model of causal economics with a number of restrictions. The first element of constraining causal economics to expected utility is restricting outcomes to single-value, net positive, or negative lottery outcomes. This is accomplished by applying the following conditions to Equations (1)–(3):

w⁺⁽⁺⁾, w⁺⁽⁻⁾, w⁻⁽⁺⁾, w⁻⁽⁻⁾ = 0

v⁺⁽⁺⁾, v⁺⁽⁻⁾, v⁻⁽⁺⁾, v⁻⁽⁻⁾ = 0

p_A, q_A, P_A, Q_A = 0

p_U, q_U, P_U, Q_U = 0

w, v, Z, z ≠ 0

Ψ, Ψ^C = 1

Given that outcomes in causal economics typically span a significant number of time periods, the second element of reducing causal economics to expected utility is the requirement that

t = t₀

Based further on the assumption of linear independence as required for expected utility [33], the expected utility value function is defined by the familiar Equation (4):

$$\begin{array}{c} f={\displaystyle {\displaystyle \sum }_{i=0}^N}\left[v\left(Z_{U i}\right) \left(z_{U i}\right)\right]\end{array}$$

(4)

In expected utility theory, the value function is optimized in the context of standard consumer choice theory, based on indifference curves and budget constraints [34]. In order to reduce causal economic optimization to expected utility optimization, the internal psychological trade-off constraint function (2) must also be limited to a budget constraint and the personal total cost threshold constraint (3) must be allowed to take on an infinite value.

Time preferences are often added to the risk preferences of expected utility through the delta model [19], which incorporates exponential discounting [18]. Exponential discounting is time-separable, also known as time-consistent or dynamically consistent, which implies that a decision-maker’s trade-offs between utility today and delayed utility in the future are independent of when that delay occurs. In the delta model, each subsequent period utility value is multiplied by δ^t as shown in Equation (5):

$$f={\displaystyle \sum }_{t=1}^M{\delta }^t{\displaystyle \sum }_{i=1}^N\left[U(Z_{U i}) (z_{U i})\right]$$

(5)

This approach to intertemporal time discounting applies the delta discount factor to the overall utility function, which is a reduction from the framework of causal economics, where time discounting can be captured within any or all of the weighting functions, w⁺⁽⁺⁾, w⁺⁽⁻⁾, w⁻⁽⁺⁾, w⁻⁽⁻⁾, v⁺⁽⁺⁾, v⁺⁽⁻⁾, v⁻⁽⁺⁾, and v⁻⁽⁻⁾, with a functional form that fits the data appropriately.

2.4. Reducing Causal Economics to Behavioral Economics

Causal economics can also be mathematically reduced to cumulative prospect theory, which is the most robust formalized framework within behavioral economics. The first step in doing so is to restrict outcomes to single-value, net positive, or negative lottery outcomes, as was done with expected utility. This is accomplished by applying the following conditions to Equations (1)–(3):

w⁺⁽⁺⁾, w⁺⁽⁻⁾, w⁻⁽⁺⁾, w⁻⁽⁻⁾ = 0

v⁺⁽⁺⁾, v⁺⁽⁻⁾, v⁻⁽⁺⁾, v⁻⁽⁻⁾ = 0

p_A, q_A, P_A, Q_A = 0

p_U, q_U, P_U, Q_U = 0

w, v, Z, z ≠ 0

Ψ, Ψ^C = 1

Outcomes in causal economics typically span a significant number of time periods, which results in the second requirement for reducing causal economics to cumulative prospect theory:

t = t₀

In cumulative prospect theory, Z and z are relative to the status quo (instead of zero), Ω is a rank-dependent cumulative probability weighting, and sign comonotonic trade-off consistency holds [28]. Under all of these conditions, the causal economic value function (1) reduces to the utility value function for cumulative prospect theory represented by Equation (6):

$$\begin{array}{c}V\left(f\right)=\{{{\displaystyle \int }}_{i=0}^N\left[\left[v\left(Z_{U i}\right) w\left(z_{U i}\right)\right]\right]{\mathsf{\Omega }}_i+\\ {{\displaystyle \int }}_{i=-k}^{-1}\left[\left[v\left(Z_{U i}\right) w\left(z_{U i}\right)\right] \right]{\mathsf{\Omega }}_i \}\end{array}$$

(6)

In cumulative prospect theory, the value function is optimized in the context of standard consumer choice theory, based on indifference curves and budget constraints [34]. As a result, in order to reduce causal economic optimization to cumulative prospect theory optimization, the internal psychological trade-off constraint function (2) must also be limited to a budget constraint, and the personal total cost threshold constraint (3) must be allowed to take on an infinite value.

This reduction to cumulative prospect theory addresses the risk preference element of behavioral economics. However, intertemporal choice is an important component of behavioral economics [10] and is typically incorporated through quasi-hyperbolic discounting via the beta-delta model [19]. Quasi-hyperbolic discounting captures time-inconsistent preferences, such as present bias and changing preferences as time elapses. These are important additions, as extensive research demonstrates that decision-makers generally do not have time-consistent preferences [35]. They tend to be more patient for long-term gains and relatively impatient for short-term gains, which means that they often plan to do something in the future, but subsequently change their mind. These characteristics align closely with the concept of causal coupling, especially as illustrated in the real-world example in Section 2.2. Causal economics weighting functions often employ quasi-hyperbolic discounting.

In the beta-delta model, each subsequent period utility value is multiplied by β and δ^t as shown in Equation (7).

$$\begin{array}{c}V\left(f\right)=\beta \{{{\displaystyle \int }}_{i=0}^N{\delta }^t \left[\left[v\left(Z_i\right) w\left(p_i\right)\right]\right]{\mathsf{\Omega }}_i+\\ {{\displaystyle \int }}_{i=-k}^{-1}{\delta }^t \left[\left[Z\left(Q_i\right) w\left(q_i\right)\right] \right]{\mathsf{\Omega }}_i \}\end{array}$$

(7)

This approach to intertemporal time discounting applies the beta and delta discount factors overall to the utility function, which is a reduction from the framework of causal economics, where time discounting can be captured within any or all of the weighting functions, w⁺⁽⁺⁾, w⁺⁽⁻⁾, w⁻⁽⁺⁾, w⁻⁽⁻⁾, v⁺⁽⁺⁾, v⁺⁽⁻⁾, v⁻⁽⁺⁾, and v⁻⁽⁻⁾, with a functional form that fits the data appropriately.

2.5. Causal Economics and Artificial Intelligence

2.5.1. The Principle of Causal Coupling

Microeconomics

The previous sections have mapped out causal economics and demonstrated that it can be used as a single unified decision model that captures the breadth of scenarios applicable in the real world, including those tackled by AI. This provides increased clarity for AI modeling at the individual level. AI is a broad discipline that attacks many problems with a range of techniques [36]. However, the richness of causal economics creates complexity and there is a need to be aware of the risk of combinatorial explosion slowing down computation and impacting solvability [14]. Causal economics is most relevant to AI problems in the areas of planning, decision-making, and learning. These areas deal with intelligent reasoning in the face of uncertainty and incomplete information over time. As a result, they incorporate probabilities, feedback, and learning, utilizing various tools from probability and economics [14]. Research in adaptive economics has mapped out the importance of this feedback and learning in a complex and evolving environment [37]. The fundamental anchor in causal economics that aligns it closely with this context is the concept of causal coupling (CC). Causal coupling builds in cost and benefits that reflect subjectivity, causality, uncertainty, and information availability over time. Causal coupling is formally defined in Equation (8):

$$\begin{array}{l}when {\displaystyle {\displaystyle \sum }_{t=0}^M}[ v^{(+)}(P_{Ait}) +v^{(+)}(P_{Uit})]>0 it holds that w^{(+)}\left(p_{Uit}\right)\\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}>0 and {\displaystyle {\displaystyle \sum }_{t=0}^M}[ v^{(-)}(Q_{Ait}) + v^{(-)}(Q_{Uit}) ]\\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}>0 and w^{(-)}\left(q_{Uit}\right)\\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}>0 and {\displaystyle {\displaystyle \sum }_{t=0}^M}t over values where \{v^{(+)}(P_{Ait})\\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}+v^{(+)}(P_{Uit})>0 \}\\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}> {\displaystyle {\displaystyle \sum }_{t=0}^M}t over values where \{v^{(-)}(Q_{Ait}) +v^{(-)}(Q_{Uit})>0 \}\end{array}$$

(8)

Equation (8) requires that each outcome must contain both benefits and costs (contemporaneous and subsequent causation). It reflects the notion that decision-makers in general have more immediate control over and face more immediate impact on their costs in terms of upfront effort than they do over the benefits that typically follow with some time delay. Equations (1)–(3) and (8) directly build causal coupling into the definitions of percepts, outcomes, preferences, and constraints. This mathematical grounding provides a robust foundation for AI algorithm development and application in individual agent decision-making.

Macroeconomics

Whereas causal coupling connects cost and benefit for individual decision-makers as a causal relationship over time, it can also be applied at a macro policy level in pursuit of societal optimization. This possibility centers on a bold conclusion of causal economics—that economic social coordination failures (outcomes that are not Pareto optimal) within societies are at their root the result of persistent causal decoupling of benefits and costs within populations of involved and/or impacted agents, and that effective solutions are the result of ensuring that causal coupling is achieved [20]. In essence, social coordination failure is a generalization of the well-entrenched concept of market failure.

The pursuit of self-interest through the power of the free market, captured by Adam Smith’s invisible hand metaphor, has been demonstrated in practice [38], but there are significant market failures in society that reduce welfare [39]. The application of causal coupling at a macro level attempts to close this gap through policies that couple costs and benefits for individuals, building in the notion of freedom with accountability [20]. A loose metaphor for causal coupling at the macro level is perhaps the generalizing of Adam Smith’s ‘invisible hand’ into ‘invisible hands’. One hand represents the freedom to pursue self-interest for personal benefit, while the other represents the obligation to contribute a prorated share to societal costs as determined through democracy. This perspective allows a fresh opportunity for AI researchers to apply causal coupling as a macro policy condition when testing for policy optimality in populations.

Market failure occurs when the voluntary actions of agents fail to take into account broader externalities to the current transaction [40]. Social coordination failure typically results when an agent is able to obtain personal enrichment without bearing the causally preceding costs, essentially gaining benefit through means other than a voluntary interaction/exchange, and/or avoiding their prorated share of the cost of public programs. It is essentially a causal decoupling of cost and benefit over the intermediate to long term across individuals and society at large. In causal economic theory, market failure is not due to exogenous externalities, but is instead the result of endogenous causal decoupling between costs and benefits across individuals and the population [20].

In practice, the largest source of causal decoupling in societies is when intermediaries receive/distribute compensation/benefits that are not tied to performance. Social policies are Pareto optimal when they couple benefits and costs for agents and do not allow intermediaries to accrue extra benefits from and/or offload extra costs to others who have no choice in the matter, taking into account realistic preferences, usage, and risk [20]. When involuntary allocations are applied through taxation to fund democratically determined public goods and services and risk-sharing programs, optimal allocations of benefit and cost are achieved by prorating costs across individuals either through a flat tax or user fee, with relief for those facing significant challenges to paying a proportionate share, subsidized by taxpayers with a higher ability to pay. Causal economics asserts the optimality of taxation based on voter-approved specific usage with committed metrics (use-based taxes), as opposed to source-based taxation that is automatically collected on activities associated with economic value creation, such as income and consumption, and placed into general government coffers. This does not necessitate higher or lower taxes on individuals or overall. It just places the focus on a causally coupled, efficient means of raising funds for directed purposes desired by a democratic majority of voters. This is a highly idealized discussion, as many entrenched stakeholders in the real world would vigorously resist the implementation of such a model.

2.6. The Neuroscience Underpinning Causal Economics

The concepts and structural setup of causal economics are closely aligned with findings in neuroscience [41]. Alongside psychology, neuroscience adds a fundamental perspective to our understanding of human decision-making. This convergence in research provides a strong grounding for causal economics and its use as a foundation in AI. Studies in neuroscience have established the integration of reward and pain during choice [42]. It has been demonstrated that the cost required in action is a vital determinant of choice behavior [43]. Multiple human and animal experiments have shown how the effort one must exert impacts choice and how the brain calculates and integrates effort into action value [43,44,45,46]. Total effort is a combination of upfront and ongoing efforts and can be expended as either mental work [47] or physical work [41]. Additional research in neuroscience reinforces the discounting of prospects where outcomes contain potential pain or loss [42,48] and are separated over time [49,50,51]. Overall, these observations on effort, costs, and outcome value over time from research in neuroscience ground the upfront and ongoing cost component of causal coupling at the center of causal economics.

2.7. The Psychology Underpinning Causal Economics

The science of large and complex choices that take upfront and continued effort over time is comprehensively covered in the field of psychology [52]. The causal economics framework essentially consolidates and provides a common formalized structure to many proven concepts in psychology [20]. From a theoretical perspective, the structural design of causal economics relies heavily on many principles of psychology, as does that of behavioral economics [53]. This is an important foundation for a robust framework that can be most useful when used in AI. However, this alignment to psychology does not end with theory. In applied practice, psychologists actively work to assist in the management of a decision-maker’s perceived costs and benefits to obtain desired outcomes. These therapies can focus on thoughts and behaviors to various degrees, through methods such as cognitive behavioral therapy, behavioral therapy, or behavioral choice theory [54].

2.8. Decision-Making in Artificial Intelligence

AI seeks to address a broad range of challenges, spanning reasoning, planning, learning, and perception, with applications that span areas such as software, robotics, and the Internet of Things [55]. AI researchers have devised many tools to tackle these problems, and since AI agents generally operate with incomplete or uncertain information in the real world, for the most part, methodologies from probability theory and economics are leveraged [14]. Modeling of decision-making agents in AI typically parallels the approach of microeconomic decision-making theory, based on bounded rational utility maximization in the face of constraints and perceptions of the environment [4]. Causal economics and its use of causal coupling provide the best structural fit as an underlying model of choice in the context of AI research because it more closely models actual human decision-making.

AI takes the bold step of attempting to automate the decision process for efficiency and results that would parallel a human decision-maker. Automated planning [56] involves agents pursuing a specific goal, in the face of complexity, where solutions need to be discovered and optimized in multidimensional space [56]. Decision-makers choose an action with some known costs and make probabilistic guesses about future additional costs and benefits. In multi-agent situations, the decision-maker’s environment and even preferences can be dynamically unknown, requiring ongoing iterative revision to the model and associated strategies and policies [14]. Agents are often forced to reassess their situation once results are experienced [14]. As the field of artificial intelligence is fundamentally concerned with solving applied problems, not just theory, a range of computational techniques are deployed, including dynamic programming, reinforcement learning, and combinatorial optimization [14].

Like applied economics, AI research typically utilizes dynamic stochastic programming [15], based on a Markov decision process [57]. In this approach, a transition model is defined that describes the probability that a specific action will change the resulting state in a certain way as well as a reward function that defines the utility benefit value of each state and the cost of each action. Within this context, a policy is calculated or learned and associates a decision with each possible state of the world [14].

There has been increasing momentum among AI researchers for the use of subjective probability, in order to model utility functions based on actual consequences and the decision taken, a priori probability distributions, and other probability distributions impacting the decision [58]. Currently, AI researchers often model intelligent agents using Bayesian probability theory to inform beliefs (priors), and utility theory to inform preferences [4]. A broad range of practical problems are addressed with Bayesian tools, such as Bayesian networks [59]. Particular algorithms are often applied to particular classes of problems, such as the Bayesian inference algorithm for reasoning, the expectation-maximization algorithm for learning [14], decision networks for planning using decision networks [14], and dynamic Bayesian networks for perception [60].

The power of AI’s impact on society is demonstrated in many areas. AI is helping to drive higher revenues and greater profitability for companies through optimized marketing [61]. Environmental sustainability is a clear and urgent societal priority, and AI is being leveraged to analyze possible solutions in this area [62]. It also cannot be taken for granted that AI and machine learning are only possible given the stability and performance of the computing infrastructure that underpins their use, which makes cyber security of paramount importance. AI and machine learning are also playing a pivotal role in maximizing the security of this infrastructure and related applications using deep neural networks [63].

It is clear that the environment in which AI research is conducted is one of complex decisions involving causally linked cost and benefit trade-offs over extended periods of time, in the face of uncertainty, and across many agents. This environment is naturally modeled with causal economics and causal coupling, and does not correspond highly to the underlying frameworks of expected utility or behavioral economics. The examples above show the sizable impact AI can have in important areas of our society, so it is very important that our approach to AI mirrors human reasoning as closely as possible, in order to produce truly intelligent decisions. For this reason, AI research must not rely on unrealistic models of human behavior in their optimization, such as expected utility and even behavioral economics. Causal economics provides a framework that does not rely on 100% rationality and allows the modeling of a full range of cost and benefit interactions, both certain and uncertain, connected by the reality that costs/effort must almost always causally precede future expected benefits [20]. For example, in order to build a company, initial investments and effort are required, and in order to achieve target health outcomes, up-front and sustained changes to behavior are required.

2.9. Machine Learning

The previous section illustrated the underlying context in which AI agents must make decisions: one of subjective, causally linked cost–benefit trade-offs over time, in the face of uncertainty, with multiple interacting individuals and feedback from actions taken. However, stopping at this point would forego the opportunity to incorporate learning insights that reinforce effective strategies while simultaneously avoiding ineffective ones [64]. In real-world decision-making situations, humans iteratively discover information without a priori knowledge of eventual results. They regularly experience outcomes that are demonstrated to be suboptimal posteriori. However, these situations provide learning opportunities, and as a result, decision-makers are able to modify future decisions in pursuit of improved future results [65].

The fundamental driver of the study of machine learning is to understand and apply programs that are able to improve their own performance on a particular task automatically [14]. The notion of computers being able to learn from their own decisions and outcomes has been fundamental to AI from the outset [66]. Today, machine learning has become central to the resurgence of interest that is occurring in AI [67]. Fields such as healthcare are making very strong use of machine learning and seeing positive results [68]. The study of machine learning can fall into one of two major categories: unsupervised and supervised. In the unsupervised machine learning approach, algorithms analyze data in order to find patterns and make predictions from that analysis automatically without any human guidance. In contrast, the supervised machine learning approach involves a human, with the role of labeling data inputs and categorizing them [14].

Reinforcement learning is one of the most utilized forms of machine learning in practice and the approach aligns very closely with the principle of causal coupling. The approach of reinforcement learning applies rewards to the decision-making agent for all actions that the model defines as good, and applies penalties to all actions that the model defines as bad [69]. Researchers have gone further, deriving the additional concept of inverse reinforcement learning. In such models, decision-making agents are able to seek out new information and then apply this experience-based knowledge to improve their preference functions. Inverse reinforcement learning does not utilize an explicit specification of a reward function because the reward function is actually inferred based on the observed behavior of the decision-maker [70]. The field of machine learning continues to evolve, bringing with it powerful new approaches that can improve the predictive power of models. An example is the inclusion of the concept of transfer learning, whereby knowledge obtained from a particular problem can be intelligently applied to a new problem to improve outcomes [14]. Building upon the previous discussion regarding the importance of grounding theory in multiple related disciplines, the area of deep learning is another core branch of machine learning that deserves attention. It leverages artificial neural networks that are inspired by biological systems with the objective of closely resembling the biological processes of human decision-making [71].

Optimization in causal economics occurs within this context, through maximization of the utility function, V, defined in Equation (1), subject to three constraint functions: specifically, the equality C, as defined in Equation (2), the inequality L, as defined in Equation (3), and the set of inequalities in Equation (8). In these equations, the weighting function components, v, w, and Ψ, will take shape based on the data set, incorporating any learning dynamics involved. These parameters allow real human personalities to be reflected in data sets. Because all of these functions are non-linear, the optimization procedure involves constrained non-linear programming. The problem is then solved using Sequential Least Squares Programming (SLSQP) [72]. In practice, this can be solved with machine learning, using a tool such as SciPy within Python. This practical approach directly aligns the time-based, information-constrained, and uncertain process of decision, action, feedback, and learning captured in causal economics with machine learning. Utility maximization within causal economics differs significantly from alternative approaches to utility maximization due to the use of these non-linear structural functions.

However, one component of causal economics that is missing in traditional machine learning methodology is causation, which is at the heart of causal machine learning and causal economics. Causal inference is addressed in the following section as a result.

2.10. Causal Inference

The previous section has sought to demonstrate that applied economics can certainly benefit from the use of machine learning in practice and that machine learning can benefit from the explicit use of causal economics in its modeling. However, the opportunity for these disciplines to work in tandem for improved results has become greater than ever with the surge in research interest in two areas fundamentally focused on causality: natural experiments in econometrics and causal machine learning in computer science [25,73]. Causality is at the very heart of causal economic theory through the concept of causal coupling [20], so it is particularly exciting that causality is seeing a surge of interest in the field of economics. This momentum has had an impact, as it is now considered best practice in empirical work to identify all assumptions necessary to estimate causal effects [20]. Causal inference is vital in modern scientific research because without it researchers only have correlations to rely on. In today’s highly connected world, data almost always contains variables that impact each other, producing observations that are spurious and appear to have causal relationships that do not exist. It is clear that reliance on correlation cannot be relied upon for intervention decisions in areas such as healthcare [74].

2.10.1. Natural Experiments

A recent revolution in empirical research has brought the importance of causality front and center. The current prominence of causality in empirical work is highlighted clearly in the econometrics and statistics literature through the work of 2021 Nobel laureates David Card, Joshua Angrist, and Guido Imbens on natural experiments [75] and independently in the biostatics literature [76]. These insights build upon the potential outcomes framework [77] first proposed by Jerzy Neyman and the subsequent Neyman–Rubin causal model [23] devised by Donald Rubin. This invigorated and cross-disciplinary interest in causality presents an opportunity for causal economic theory, causal machine learning, and the natural experiments approach to be used together in future research.

Natural experiments are so powerful because they allow researchers to extract causal relationships from the real world, which is much more challenging to work with than data from a controlled clinical trial environment. In the latter, exposure to test and control groups can be controlled to ensure randomization. In natural experiments, individuals and/or groups are exposed to both the test and control conditions which are observed from nature rather than controlled by the experimenter. In nature, there is often a confounding variable, one that is related to both the input variable and the output variable, which can distort the causal relationship of interest and even give rise to an apparent causal relationship that is actually spurious. Therefore, natural experiments attempt to approximate a random assignment. If a researcher suspects that changes in variable A cause changes in variable B, they compare the level of B across systems that vary in their level of A. So instead of manipulating A directly, the researcher analyzes the variations in the level of A observed in nature, not in a lab.

It is much more difficult to interpret cause and effect in natural experiments because individuals themselves have chosen whether they participate in the program/policy of interest. Even in this challenging context, robust conclusions can be drawn from natural experiments in which individuals cannot be forced or forbidden to participate in a policy/program [78], such that changes in outcomes may be plausibly attributed to exposure to the intervention [79]. Since most randomized experiments in practice do not allow complete control over who actually participates in a policy intervention, these findings are broadly relevant and widely adopted by researchers in practice. Natural experiments can shed the most insight when there exists a clearly defined exposure to a policy/program intervention for a well-defined subpopulation and the absence of such exposure in a similar subpopulation. Though extremely powerful, the methods described can only provide an estimate of the effect of a policy intervention on the people who actually changed their behavior as a result of the natural experiment.

The Rubin causal model (RCM) measures the causal effect of a policy by comparing the two alternative potential outcomes that an individual could experience. Since only one can actually be observed as an outcome, one is always missing, an issue known as the fundamental problem of causal inference [80]. The fundamental problem of causal inference means that causal effects cannot be directly measured at an individual level. A researcher must instead turn to population-level estimates of causal effects drawn from randomized experiments [81]. To the extent that an experiment can be randomized, individuals can be identified within one of two groups and then a difference in the means of the policy variable can be observed to ascertain the causal effect [78], known as the average treatment effect (ATE).

In natural experiments, randomization is not always possible, as individuals may select particular outcomes based on other factors. In these cases, techniques such as propensity score matching [82] are utilized in an attempt to reduce treatment assignment bias, and mimic randomization. This is accomplished by creating a sample of individuals that received the treatment which is comparable on all observed covariates to a sample of units that did not receive the treatment. More precise estimates can be achieved by focusing only on individuals who are compliant [83], which leads to the use of measures such as the local average treatment effect (LATE), also known as the complier average causal effect (CACE) or the conditional average treatment effect (CATE). The LATE is typically calculated as either the ratio of the estimated intent-to-treat effect relative to the estimated proportion of compliers, or through an instrumental variable estimator. However, this increase in precision comes with the trade-off of ignoring the effect of non-compliance to a policy that is typical in real-world environments.

An alternative model of causal inference to the RCM is a structural causal model (SCM) [84], which introduces a highly graphical approach that can be very helpful in modeling assumptions or identifying whether an intervention is even possible or not. The potential outcomes literature is more focused on quantifying the impact of policy/program interventions. In addition, there is ongoing work to potentially unify these methods, such as single-world intervention graphs [74]. Continuing in the spirit of unification, there is further opportunity at present to explore whether predictive power in experiments can be increased by building the causal definitions of outcomes, preferences, and constraints of causal economic theory into research models across broad areas such as computer science, economics, healthcare, and others.

These innovative developments in natural experiments have propelled causality to the forefront of statistics, econometrics, and machine learning. Causal machine learning will be discussed in the following section, but here it is important to demonstrate how causal inference is formally defined and quantified. This is a foundation of causal machine learning and as we will illustrate subsequently, it is also foundational to CEML as a result.

Each individual in a population of interest can be denoted by i, where each is either a member of the control group W_i = 0 or the treatment group W_i = 1. In general, the response metric of interest under study will be the utility function, v. When a particular individual is within the control group, the response metric is V_i (0) (V_i (1)_resp. Only one outcome can be observed in practice, which is denoted as V_i^obs = V_i (W_i). In this formulation, each individual in the population will possess a vector of pre-treatment covariates, denoted by X_i. With these definitions in place, causality can be extracted from the data through the conditional average treatment effect (CATE), which is specifically defined in Equation (9):

$$\tau \left(x\right):=(\mathbb{E} [Y\left(1\right)-Y\left(0\right) | X=x])$$

(9)

This can then be written as the difference in means of the response function of the treatment and control groups, as defined in Equation (10), where µ represents the mean:

$$\tau \left(x\right)={\mu }_1\left(x\right)-{\mu }_0\left(x\right)$$

(10)

2.10.2. Causal Machine Learning (CML)

Causal machine learning [24] is currently an emerging area of computer science with an expanding literature [25]. The increased focus on causality in empirical work spans many disciplines, such as computer science, econometrics, epidemiology, marketing, and statistics, as well as business [73]. Some of the most significant work underlying this area is the Neyman–Rubin causal model [85] based on the potential outcome approach of statistics [23] discussed in the previous section. Causal machine learning builds upon this underlying work and others in causal inference [86] and applies machine learning methods [73] to answer well-identified causal questions using large and informative data [87]. The continual pursuit of better treatment/policy outcomes by practitioners provides reason to believe that causal inference and machine learning [88] will come to play a dominant role in empirical analysis over time.

Pioneering economists are driving the increased use of causal machine learning in econometrics [89] and its application to economic policy evaluation [73]. Prominent causal machine learning methods have been developed by economists directly, an example being causal forests as proposed by Susan Athey and Stefan Wager [90]. The causal forest approach identifies neighborhoods in the covariate space through recursive partitioning. Each causal tree within the causal forest learns a low-dimensional representation of the heterogeneity of the treatment effect. The causal forest is an average of a large number of individual causal trees, where each tree differs based on subsampling [89]. Studies find a value-added contribution from the use of causal machine learning in welfare economics applications [91]. With the continued expansion of the big data trend, there will most likely be further increased interest in causal machine learning to deal with data sets that have a large number of covariates relative to the number of observations, posing a challenge to non-causal techniques.

Causal machine learning has experienced increasing popularity in healthcare [74] in tandem with a continued focus on real-world evidence (RWE) and improvement of intervention outcomes [92]. Obtaining valid statistical inference when using machine learning in causal research is currently a critical topic in public health and beyond [93]. Medical researchers and practitioners are making some of the most powerful advances in the use of causal machine learning, asserting that machine learning promises to revolutionize clinical decision-making and diagnosis [94]. For example, a 2020 study conducted by Richens et al. showed that a causal, counterfactual algorithm delivered predictions that ranked within the top 25% of doctors, achieving expert clinical accuracy, compared to a standard associative algorithm that only placed in the 48% of doctors. These findings demonstrate that causal reasoning is a vital missing ingredient when applying machine learning to medical diagnosis situations [95]. Recent research on the past expansion of health insurance programs using causal machine learning has provided tangible and robust insights into potential future program expansions [94].

A range of powerful meta-algorithms has been developed for causal machine learning, including the S-learner, the T-learner, the X-learner, the R-learner, and the doubly robust (DR) learner. Each of these uses slightly different ways of estimating the average output and defining the conditional average treatment effect [96]. For example, in epidemiology, expert researchers suggest prioritizing the use of doubly robust estimators, the super learner, and sample splitting to reduce bias and improve inference [97]. Work is ongoing to further categorize and apply causal machine learning [98].

The best way to bring these advances in causal machine learning into CEML is through the S-learner meta-learner algorithm because it can accommodate non-linear policy functions, constraints that are both equalities and inequalities as well as both binary and continuous treatments. In this approach, the outcome is predicted using the intervention, confounder variables, and additional covariates as features of the model [72]. Once set up, the model is utilized to estimate the difference between the means of the potential outcomes under different treatment conditions, which quantifies the treatment effect of interest.

The S-learner is powerful, but practitioners must keep in mind the risk of regularization bias [72]. It is common for machine learning algorithms to utilize regularization in order to reduce the challenge of overfitting. Unfortunately, this approach also creates issues for the measurement of causality. More information on this can be found in the research of Künzel [96]. This risk is in addition to the typical challenges that must be managed when using causal machine learning, such as issues of data quality, model complexity, and potential for bias when invoking causal inference methods [99].

The S-learner can be used to test various possible treatment variables and their effects. The S-learner is a single model, so the treatment variable W_i (which can be any of the input parameters) is treated just like any other covariate in the model, such as any that are contained in the vector X_i. Equation (11) formally illustrates this setup:

$$\mu \left(x,w\right):=\mathbb{E} \left[Y^{obs} | X=x, W=w\right]$$

(11)

Implementation of the S-learner is carried out in two steps [96]. First, all of the observations in the data set are used to estimate the response function of interest $\widehat{\mu }$ (x, w), and second, the CATE is estimated via Equation (12):

$$\widehat{{\tau }_S}\left(x\right)=\widehat{\mu } \left(x, 1\right)-\widehat{\mu } \left(x, 0\right)$$

(12)

The increased interest in causal machine learning is driven by the desire to use knowledge of a causal relationship to better predict outcomes of policies and treatments and to support confident and informed decision-making in important situations where reliance on spurious correlations could be dangerous or costly. Causal machine learning is where the theoretical and statistical approaches discussed previously see power in action in a world of big data. It is where the principles of causal economics connect directly with artificial intelligence, where theoretical and applied principles come together, unified in theory and practice.

2.10.3. Causal Economic Machine Learning (CEML)

CEML as an Evolution of Human AI

Causal economic machine learning (CEML) represents the intersection of causal machine learning and causal economics, resulting in a more ‘human AI’. It essentially involves the replacement of traditional economic decision modeling with causal economic decision modeling in the practice of machine learning, particularly with respect to reinforcement learning and deep learning contexts. This primarily means observing outcomes that contain both certain and uncertain costs upfront that cause future uncertain benefits and costs when modeling individual decision-makers. It can also mean the application of causal coupling at the macroeconomic policy level to gauge the effectiveness of policies that do a better or worse job of coupling cost and benefit across individuals.

Table 1. Four phases of evolution in AI leading to CEML/human AI.

Generation	1st Generation (Rational AI)	2nd Generation (Behavioral AI)	3rd Generation (Casual AI)	4th Generation (Human AI)
Comparative Synopsis	Relies on 100% human rationality, which is proven to be unrealistic.	Lacks real predictive power based on associations not causation and relies on slightly improved models of human behavior that incorporate biases.	Benefits from powerful causal inference statistical methods, but relies on unrealistic models of human behavior.	Benefits from powerful causal inference statistical methods and relies on realistic models of human behavior.
Model of Human Behavior	Expected Utility.	Behavioral Economics.	Behavioral Economics.	Causal Economics.
Model of Artificial Behavior	Artificial Intelligence.	Machine Learning.	Natural Experiments and Causal Machine Learning.	Causal Economic Machine Learning.
Methods and Tools	Linear optimization.	Non-linear optimization using methods such as SLSQP, deployed through standard (non-causal) machine learning tools such as SciPy (1.7.0) in Python.	Non-linear optimization using methods such as SLSQP, deployed via causal machine learning meta-learners (such as S-learner, 0.15.1) in tools such as EconML in Python.	Non-linear objective and constraint functions enforcing cost to benefit causation. Deployed via non-linear SLSQP (1.14.1) (Ex. SciPy) and S-Learner (Ex. EconML).

illustrates the way in which these trends have interacted, to bring us to the fourth generation of AI, which can be thought of as “human AI”. As indicated, CEML is human AI because it is the first model to combine realistic causal decision-making in humans with realistic causality in machine learning.

Implementation of CEML

To implement CEML a researcher is likely to follow three major steps:

• Label the data in line with the CE framework, capturing current and future costs and benefits, both certain and uncertain.
• Utilize Sequential Least Squares Programming (SLSQP) for constrained non-linear optimization of Equations (1)–(3) and (8).
• Apply the S-Learner machine learning algorithm.

Because causal economics involves both a utility function and constraint functions that are non-linear, where those constraints include both equalities and inequalities, the optimization involves constrained non-linear programming. As is common in economics, a Lagrangian is often employed to uncover solutions consistent with the set of equations that must hold in order to preserve the framework of causal economics. This problem is often solved by using Sequential Least Squares Programming (SLSQP) [72], using software such as the SciPy tool within Python.

The response functions of this optimization model can then be deployed through the S-learner meta-algorithm, known as a meta-learner, which estimates the CATE. The S-learner is able to accommodate non-linearity and the existence of both equalities and inequalities. The S-learner is frequently deployed through EconML within Python [100].

There is currently a tremendous amount of potential for a research agenda in CEML, as a causal machine learning framework that employs the model of causal economics—in particular, for studies that employ big data and the increasingly powerful software in use today. CEML essentially sits at the convergence of three major trends:

• Expected utility theory and behavioral economics theory have led to causal economic theory.
• Causal inference (natural experiments, structural causal models) and machine learning have led to causal machine learning.
• Causal economic theory and causal machine learning have led to causal economic machine learning.

Challenges to the Implementation of CEML

The value of causal economic machine learning is directly related to its robust specification of input variables and the relationships between them specified by the utility and constraint functions. This complexity can add more powerful, human-like modeling, but it also poses a number of additional challenges. These include:

• Limited data sets. This challenge exists due to the large number of parameters involved in the equations of CEML and in particular the inclusion of a number of psychological inputs. CEML studies can require extensive up-front work in preparing and labeling data sets.
• Risk of combinatorial explosion. The complexity of the CEML optimization presents a risk of slowing down computation and impacting solvability [14].
• Implications. The concept of causal coupling that is built into the functional forms of CEML can produce implications in applications that represent significant upheaval of the status quo in areas such as tax policy. The following section lays out potential applications that largely represent ideal scenarios, so policymakers evaluating these possibilities must appreciate the difficult reality of driving major change in the real world against vested interests.

3. Applications and Discussion

3.1. Decision Making

The real world presents humans and machines with complex decision scenarios, involving causally linked cost and benefit trade-offs over extended periods of time, in the face of uncertainty, while interacting with other competing and cooperating decision-making agents. This context renders expected utility theory and behavioral economics theory unable to serve as core models in machine learning applications, primarily as a result of their reliance on single-period lotteries with intertemporal choice augmentations such as exponential discounting [18] and quasi-hyperbolic discounting [19], that do not explicitly define the causal relationship that uncertain benefits follow upfront certain and uncertain costs over time on significant decisions [20].

In contrast, the framework of causal economics does provide an effective theoretical underpinning for use in AI and machine learning because it is structurally built to capture complex decision scenarios that involve causally linked cost and benefit trade-offs over extended periods of time, in the face of uncertainty. It does not however explicitly build in multi-agent interaction dynamics, so additional methods such as game theory can be employed when defining an underlying model for an AI research program. Research has demonstrated the power of modeling inference with a combination of insight from game theory and machine learning [101].

The major contribution of this paper is to demonstrate that CEML serves as a preferred model for machine learning research because it combines the causal advances in machine learning with the causal advances in human decision modeling. This alignment of human and machine behavior provides a model of ‘human AI’, and this paper has sought to lay out the tools to implement the framework in practice. Below are examples of application areas, and the next phase of this research program will be to apply CEML to data sets in these fields. A major challenge at this point in demonstrating CEML in practice relative to alternative approaches is finding a data set with observations across the larger number of variables contained in the model. The complexity presented by this situation will not stop progress, however, because AI and ML have always taken on the seemingly impossible and put structure on it.

3.2. Macroeconomic/Social Outcomes

There is a growing body of literature on the use of machine learning in macroeconomics [83], but it is hard to find much that utilizes causal machine learning in particular [102]. Even without utilizing causal machine learning, the use of machine learning is producing improved results over full reliance on traditional econometric methods [103]. Comprehensive research from the International Monetary Fund applies the elastic net, recurrent neural network (RNN), and super learner machine learning algorithms to broad economic analysis across the world, and finds that overall they outperform traditional econometric techniques [104].

Causal economics and its core concept of causal coupling serve as a consistent tool for analyzing macroeconomic environments. It asserts broadly that structural institutions (ex. a free market, government body, central bank, etc.) and intervention policies (ex. a tax policy, etc.) will produce Pareto optimality when all decision-makers and impacted stakeholders are able to enjoy the specific benefits that result from the costs they incur and bear the costs of the benefits they receive. A number of characteristics reflect an economy with high levels of causal coupling in institutions and policies. A non-exhaustive list of a few prominent examples includes the following [20] and provides for some novel experimental research:

• Free interaction and exchange (free markets) with corrections for externalities
• Significant levels of direct democracy
• Government by legislation prioritized over bureaucracy
• Widespread compensation for results in the private and public sector
• Use-based taxation driven by flat and use-based taxes with relief for those in need
• Social risk/cost sharing programs for exposures/projects that would be catastrophic/prohibitive to individuals

The application of causal coupling in macro policy would assert the optimality of conditions such as an engaged democratic majority, free competitive markets with corrections for externalities, a fair legal framework in place of government bureaucracies, minimal monetary and fiscal policy, and taxes collected for specific voter-approved spending programs instead of being automatically collected and fed into general accounts based on income, consumption, property, and wealth. Few of these conditions prevail together on a widespread scale in real economies, generally the result of vested interests and power, with a natural, biological motivation for those in power to pursue outcomes in their favor [105].

3.2.1. Private Sector Application

The economic literature has a long history of demonstrating, through precise models, the power of competitive free markets to deliver Pareto optimal efficient allocations of goods and services through supply and demand [38]. Restrictions on competition, such as monopoly, duopoly, oligopoly, and monopsony, and government intervention through price and quantity controls, are generally known to reduce this Pareto economic efficiency. Causal economics provides a simple and consolidated fresh new way to look at this. Competitive markets are optimal precisely because they maximize the causal coupling of costs and benefits for those involved in the transaction [20]. Buyers receive the perceived benefit of their purchase and part with the total personal cost (financial and/or psychological) of the item. Sellers part with the cost of the product/service they provide and obtain the benefit of funds that can be used for things they value more. Buyers and sellers make a voluntary decision on the trade-off between cost and benefit, which couples cost to benefit, and achieves Pareto optimality. It is clear that competitive free markets build causal coupling in directly at the transaction level.

However, transactions in the competitive free market are not overall optimal to the extent that they create involuntary externalities of costs and/or benefits to others. In these cases, government policies can correct the causal decoupling through tax/spend policies, for example through ‘carbon taxes’ to address environmental damage as directly as possible. The application of a tax/credit is essentially an effort to recouple cost and benefit. An example could involve a tax charged on producers of high-pollution fuels that profit from the activity and the redistribution of those funds as a credit to companies and individuals that develop, deploy, and utilize more environmentally sound, viable fuels.

Free and competitive markets are critical to freedom and economic prosperity, but they are not perfect. Private, profit-driven companies will understandably focus on segments of the market that are very willing and able to pay, often limiting the availability and affordability of essential products and services, such as food and housing, to lower-income individuals. Government intervention is not always at odds with the desires of taxpayers. There is much evidence that the majority of economic citizens support some degree of government-driven support to help those in serious need. Social safety nets exist in most advanced economies, demonstrating that even those that do not benefit directly from a program, but instead finance it, can still see value in its importance in society.

None of this understanding of markets and government intervention delivers new implications to existing economic theory; the purpose of this discussion is to illustrate how causal coupling provides a simple, consistent, and universal way to look at economic and distributive societal issues that everyone can understand. At the level of the buyer and the firm, there is also an increased interest in CML in digital marketing and ecommerce [106], where personalization is key to targeting profitable customers and particular customer engagement actions are optimal at each stage of the time-based, causal buying process. CEML provides an even greater application fit in these areas than CML on its own, by incorporating the buyer’s personal psychological trade-off preferences.

3.2.2. Public Sector Applications

The public sector plays an important role in most economics and has a large impact on the economy through taxation and government spending. As a result, it has received a lot of attention in analysis. The use of deep reinforcement learning machine learning in the analysis of taxation is providing quantifiable insights even without the explicit use of causality. An applied example is the empirical work completed via Microsoft’s The AI Economist project [107]. This study utilizes a two-level deep reinforcement learning approach in economic simulations to generate learning with respect to dynamic tax policies. In the model, both individuals and governments are able to learn and adapt [107]. The study demonstrates that AI-driven tax policies can improve the trade-off between equality and productivity by 16% over baseline policies.

Causal economic theory implies that taxation is optimal only when it is use-based, because this approach preserves the criterion of causal coupling. Use-based taxation requires that governments obtain advance approval from citizens prior to collection and are then subjected to compensation that is tied to performance goals. If a particular spending program is approved by voters then taxes are applied directly to cover the costs. This approach is in stark contrast to source-based taxation, which automatically collects taxes on economic variables such as income, consumption, and wealth. Use-based taxation more closely parallels causally coupled private sector funding, where entities must convince funding parties in advance that they will deliver results, based on their track record and plan.

Source-based taxation is non-optimal in causal economic theory because it directly decouples costs and benefits. The optimal alternatives that do achieve causal coupling are transparent flat taxes and user fees that are tied to results. Under such as system, tax relief is typically provided in democracies to those that are significantly challenged in contributing their prorated share. User fees are a very powerful causal-coupling mechanism because they directly couple the cost of the fee to the benefit the user obtains as directly as possible. Where specific users of particular government products/services can be identified, user fees are ideal. Where this is not the case, and it can be reasonably assumed that all citizens benefit from a particular government product/service, flat taxes are the ideal method of collection. Flat taxes spread the democratically approved cost of a particular government program across all members of society, regardless of whether or not they personally benefit directly. In summary, causal coupling is maximized when government programs are funded through use-based tax policies, collected via user fees and/or flat taxes [20]. Furthermore, the democratic majority can direct that tax relief and support be provided to individuals/groups that are unable to bear their prorated share of program costs.

There is an interesting opportunity to use CEML to test the effectiveness of changes in tax and/or spending programs as they represent clear policy interventions. Because source-based and use-based policies are so significantly contrasted, and because source-based taxes are extremely entrenched in economies across the world, there is exciting potential to use CEML to test the effectiveness of incremental tax/spending programs with varying degrees of causal coupling.

3.2.3. Healthcare Applications

Healthcare often finds a home in the public and/or private sphere and is such a big impact area that it is considered here within its own category. Real-world evidence (RWE) has long been a prominent principle driving healthcare research [108], and it remains a top priority due to its approach to evaluating patient interventions and outcomes based on data from actual clinical practice. By placing a central focus on data-driven insight, RWE provides a natural point of extension into the application of machine learning in public health contexts [109,110]. The fit makes sense, because there has been a commensurate increase in the popularity of CML in healthcare research [74]. The most urgent challenges facing healthcare today require the comparison of the costs and benefits of alternative approaches, requiring that algorithms used can lay out the consequences of a doctor’s particular intervention treatment [111]. Promising new research continues to develop innovative CML methods that provide estimates of heterogeneous treatment effects in order to improve the individual-level, clinical decision-making of healthcare professionals in practice [112].

Traditional machine learning is very effective in identifying risks and predicting outcomes, but it cannot be relied upon during clinical interventions, because in these settings, identification of causation is vital [113]. Researchers have identified causal inference and reasoning as a critical yet often missing component in the application of machine learning to medical diagnosis [94]. However, incorporating causation is not always easy in practice, since observational studies very often include confounding variables that impact each of the variables under consideration. The challenge of confounding biases in individualized medical interventions has been demonstrated through research that applies causal machine learning analysis to RWE data contained in electronic health records (EHR) [114].

Additional research has shown the promising application of causal machine learning to evaluate understanding of the effectiveness of hospital discharges. Specifically, a study conducted at Kaiser Permanente Northern California (KPNC) healthcare demonstrated significant heterogeneous results in the responses of individual patients to intervention. In this study, causal machine learning was utilized in an attempt to identify preventable hospital readmissions so they could be avoided. The results of this research illustrated that focusing on predicted treatment effects instead of predicted risk could potentially reduce readmissions by 30% annually [115].

In each of the scenarios discussed, the introduction of causality into traditional correlation-based machine learning has been shown to provide a much more realistic modeling of human decision-making and behaviors through CML. CEML has extensive potential to further this progress by even more robustly modeling real human decision-making and behaviors, including all of the benefits of CML and adding an additional layer of representation of the trade-offs between costs and benefits into the definitions of outcomes, preferences, and constraints.

4. Conclusions

This paper has sought to demonstrate the unique and timely opportunity that currently exists to align the surge of interest in natural experiments and artificial intelligence, specifically causal machine learning, and tie them together with the framework of causal economics. There is a strong case for a specific research focus on causal economic machine learning (CEML). Beyond strengthened micro-foundations, the implications of causal coupling provide an interesting approach to macro policy analysis and testing.

Researchers in AI are making rapid progress and the field is highly relevant and visible in society today. AI researchers focus on applied results but are simultaneously interested in ensuring that their models are grounded in strong microeconomic decision modeling. Economic theory has made tremendous strides in many areas, but the dominant decision models of expected utility and behavioral economics have not provided a structure that works for AI and the real-world focus of machine learning because they rely predominantly on lottery-based, non-causal outcomes. Causal economics fills this gap through a framework that builds causality directly into the definitions of outcomes, preferences, and trade-off constraints, incorporating multi-period decisions in the face of uncertainty, constrained information, and limited computational capabilities. Additionally, causal economics leverages contributions to decision theory from many disciplines, including economics, psychology, neuroscience, statistics, and computer science. To the extent that a single powerful model can be utilized, the focus should be on solving real-world problems rather than dividing attention on building inconsistent models [116].

Causal economic machine learning (CEML) sits at a convergence of three trajectories:

• Expected utility theory and behavioral economics theory have led to causal economic theory.
• Causal inference (natural experiments and structural causal models) and machine learning have led to causal machine learning.
• Causal economic theory and causal machine learning have led to causal economic machine learning.

There is exciting potential to adapt and develop algorithms that utilize the definitions of outcomes, preferences, and trade-off constraints of causal economics which build in causality at the core. On this foundation, there is also an opportunity to apply causal coupling as a constraint at the macro level to evaluate economic and social policy with causal economic machine learning.

This paper’s contribution has been to lay out the theoretical foundations of CEML, casting it as preferred to alternatives based on it being a more realistic model of human behavior, and then delineate the relevant implementation tools for optimization, causal inference, and machine learning applications. The proposed steps and tools are as follows:

• Labelling the data in line with the structure of the causal economic framework (certain and uncertain costs, upfront and in the future, as well as current and future benefits).
• Optimizing the non-linear response functions for utility, Equation (1), and the two constraints. Equations (2) and (3) using SLSQP (Sequential Least Squares Programming).
• Applying the S-Learner meta-learner machine learning algorithm to fit the model.

The next phase of this research agenda is the identification of data sets that are capable of reflecting the structure of CEML and collaborating with innovative researchers to put it to the test against more traditional approaches, such as non-causal machine learning and human decision models of expected utility or behavioral economics. This is no small task, as the CE response functions contain numerous parameters and weighting functions, but AI and machine learning have never shied away from challenges that seem impossible. The entire field thrives on the mission of finding structure in the face of overwhelming information.