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

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Overall, this a well-written manuscript. This article was solid with mathematical foundations and innovative integration of causal inference into economic and machine learning frameworks. The integration of causal inference into economic models and AI represents a novel approach that aligns with current trends in interdisciplinary research. This novelty can make the paper an attractive candidate for journals focusing on econometrics, AI, and applied economics. This article also presents a high level of academic rigor, particularly in how it handles the math models and the foundational theories. Please see my concerns in the following:

 

1. the paper could be more benefit if the authors could test the assumptions for each model;

2. the paper could benefit from a more detailed comparison with existing methods, such as empirical validations or simulations;

 

Author Response

REVIEWER INSIGHT #1

Overall, this a well-written manuscript. This article was solid with mathematical foundations and innovative integration of causal inference into economic and machine learning frameworks. The integration of causal inference into economic models and AI represents a novel approach that aligns with current trends in interdisciplinary research. This novelty can make the paper an attractive candidate for journals focusing on econometrics, AI, and applied economics. This article also presents a high level of academic rigor, particularly in how it handles the math models and the foundational theories. Please see my concerns in the following:

AUTHOR RESPSONSE #1

This is very clear and actionable insight. It is much appreciated and very helpful. I have endeavored to incorporate it throughout the revision o the best of my ability within the context of this research.

REVIEWER INSIGHT #2

The paper could benefit from a more detailed comparison with existing methods, such as empirical validations or simulations.

AUTHOR RESPONSE #2

To address this feedback I have added a synopsis table to the CEML section (Section 2.10.3) that illustrates the assumptions of each model and the very specific tools necessary to deploy the model in comparison to alternative approaches. This particular paper sets up the theory and application approach and tools and establishes as a next phase the testing of this model on real data sets. This approach has been taken because the model is a significant change in approach and the intent is to set that out for clarity and map our possibilities for experts in various fields, such as healthcare, economic policy etc. with relevant data sets. The intent is to deploy and test the model in collaboration with experts in running ML models in the various fields. The author is currently engaged in connecting with other researchers to implement as directed. I have provided a thorough comparison and delineated how the non-linear optimization of the value and constraint functions is solved via Sequential Least Squares Programming (SLSQP) using SciPy in Python, and how causal machine learning is incorporated via an S-Learner that can handle this non-linear problem, with deployment via EconML in Python. I have been much more clear on what this paper contributes as a framework and prescriptive action plan and where the next phase of research lays.

REVIEWER RESPONSE #3

The paper could benefit from a more detailed comparison with existing methods, such as empirical validations or simulations.

AUTHOR RESPONSE #3

This feedback has been very helpful. As noted above, I have now added the ‘more detailed comparison’ of the proposed model and existing alternative models in the form of a summary table in the CEML section. This table shows the structural differences enforced in the model response functions, the optimization methods and specific machine learning approaches and software that can be utilized. I would have liked to have this article include an applied test of this model on real data sets, but I thought setting out this unique approach as a research agenda was a significant task to layout on its own, and I felt that I needed access to data sets and some ML deployment expertise beyond my experience. For that reason I felt it was best to provide a theoretical paper with very specific direction on the precise tools needed for optimization and testing of the model on data sets.

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors

 

The paper titled “Causal Economic Machine Learning (CEML)” addresses important issues however it needs improvements.

 1.      The abstract has some drawbacks that could be addressed for better clarity and effectiveness:

The terms, such as "causal coupling" and "multi-period, causal-linked process," is not clearly explained. Additionally, the abstract does not provide specific examples or applications of how CEML can be used.

 

2.      Issues need to be considered for the introduction.

 

The connection between AI, economics, and machine learning is complex, but the introduction does not do enough to break down these connections in a simple and accessible way. Additionally, the introduction jumps between different ideas, such as AI alignment, decision-making models, and causal economics, without a clear, cohesive narrative. It does not sufficiently explain how the proposed causal economics model addresses the issues.

3.      This subsection on "Decision Making in Artificial Intelligence" has some issues that impact its clarity and accessibility. The text jumps between different ideas, such as decision-making, reinforcement learning, and Bayesian theory, without a clear or logical flow. Additionally, while it emphasizes the use of causal economics in AI, it does not adequately explain why this approach is superior or how it practically applies to AI decision-making.

 

To improve the clarity and readability of the subsection, it would be beneficial to incorporate practical examples from relevant studies. For instance, the "Analysis of environmental factors using AI and ML methods" can illustrate AI's application in environmental analysis, while "Smotednn: a novel model for air pollution forecasting and AQI classification" can showcase AI's role in predicting air quality. Additionally, " DBoTPM: a deep neural network-based botnet prediction model” serves as an example of AI in cybersecurity.

 

4.      Section CML: There is a lack of practical examples that could help bridge the gap between theoretical concepts and real-world applications. The focus on advanced statistical methods and the prominence of certain studies, while valuable, may overshadow the need to explain fundamental concepts in simpler terms. The discussion is somewhat fragmented, jumping between various applications and methodologies without clearly linking them back to the core principles of CML. The section could benefit from a more balanced discussion of the limitations and challenges of CML, such as issues with data quality, the complexity of model interpretation, and the potential for bias in causal inference, which are briefly mentioned but not explored in depth.

5.      Section CEML: This section highlights an interesting intersection between causal machine learning and economics, but it needs some improvements. The text is somewhat vague and lacks detailed explanations, making it difficult to understand how CEML actually works or differs from traditional methods. The section also overemphasizes the potential of CEML without adequately addressing the challenges and limitations. Additionally, the brief mention of big data and powerful software is not sufficiently expanded upon.

6.      The "Results and Discussion" section has some issues that can affect its clarity and practical application. The text introduces concepts like expected utility theory and causal economics but does not explain them in simple terms. The section also criticizes existing economic models but does not provide concrete examples or clear explanations of how causal economics and machine learning would practically improve decision-making processes. In discussing macroeconomic outcomes, the text suggests that causal machine learning is underutilized but does not provide sufficient evidence or detailed case studies to support this claim. Additionally, while the text discusses the importance of causal coupling in economic policies, it fails to address the real-world challenges and resistance that such changes would face, particularly from entrenched systems. The discussion on public sector applications is interesting, but it spins over the complexity of implementing use-based taxation and the potential unintended consequences.

7.      Limitations and the future scope should be added with more clarity.

 

8.      The conclusion section needs flow and transition. 

Author Response

REVIEWER INSIGHT #1

The paper titled “Causal Economic Machine Learning (CEML)” addresses important issues however it needs improvements. The abstract has some drawbacks that could be addressed for better clarity and effectiveness: The terms, such as "causal coupling" and "multi-period, causal-linked process," is not clearly explained. Additionally, the abstract does not provide specific examples or applications of how CEML can be used.

AUTHOR RESPONSE #1

Thank you for this very helpful feedback. I have made modifications to this section to more clearly define these concepts briefly, and I have added specific examples of how CEML can be used in certain fields, using certain, very specific methods. I have tried to add this further clarity within the constraint of keeping things short for the abstract.

REVIEWER INSIGHT #2

Issues need to be considered for the introduction. The connection between AI, economics, and machine learning is complex, but the introduction does not do enough to break down these connections in a simple and accessible way. Additionally, the introduction jumps between different ideas, such as AI alignment, decision-making models, and causal economics, without a clear, cohesive narrative. It does not sufficiently explain how the proposed causal economics model addresses the issues.

AUTHOR RESPONSE #2

I have added a paragraph to the end of the introduction which I hope does a better job of framing up the contribution of this paper. I wasn't completely sure on how to flow the paper better in line with your thought process. It still goes through the same steps, but I have now also added a paragraph at the end that illustrates how CEML as it relates to alternatives and where it has impact. In the CEML section later I have added Table 1, which showcases how each model compares, both theoretically and practically in terms of tools utilized to implement. I believe you will find this table helpful. Perhaps you would like it to go into the end of the introduction section as well? I am hoping these elements work. I find that I am writing from two perspectives--and economist, and an AI researcher, and sometimes this contributes to some choppiness in the flow I suspect. I thank you for your insight and guidance and hope I have been able to address the excellent points you note, in the revised manuscript. 

REVIEWER INSIGHT #3

This subsection on "Decision Making in Artificial Intelligence" has some issues that impact its clarity and accessibility. The text jumps between different ideas, such as decision-making, reinforcement learning, and Bayesian theory, without a clear or logical flow. Additionally, while it emphasizes the use of causal economics in AI, it does not adequately explain why this approach is superior or how it practically applies to AI decision-making. To improve the clarity and readability of the subsection, it would be beneficial to incorporate practical examples from relevant studies. For instance, the "Analysis of environmental factors using AI and ML methods" can illustrate AI's application in environmental analysis, while "Smotednn: a novel model for air pollution forecasting and AQI classification" can showcase AI's role in predicting air quality. Additionally, " DBoTPM: a deep neural network-based botnet prediction model” serves as an example of AI in cybersecurity.

AUTHOR RESPONSE #3

Thank you for this guidance. I have added a paragraph highlighting the powerful range of applications of AI as a foundation, which makes great sense to establish before moving into CML and CEML. I have also updated this in an attempt to illustrate more directly in this section how causal economics is preferred and practical for use in AI. I used your examples as guidance of the excellent areas where applications are having an impact.

REVIEWER INSIGHT #4

Section CML: There is a lack of practical examples that could help bridge the gap between theoretical concepts and real-world applications. The focus on advanced statistical methods and the prominence of certain studies, while valuable, may overshadow the need to explain fundamental concepts in simpler terms. The discussion is somewhat fragmented, jumping between various applications and methodologies without clearly linking them back to the core principles of CML. The section could benefit from a more balanced discussion of the limitations and challenges of CML, such as issues with data quality, the complexity of model interpretation, and the potential for bias in causal inference, which are briefly mentioned but not explored in depth.

AUTHOR RESPONSE #4

This really helps. I have added the examples which does help to add more direct relevance to the section. I have also directly included the challenges you noted with respect to causal machine learning. I have tried to ground this section with a formal and thorough build out of the primary causal machine learning meta learner, the s-learner that applies to the causal economic and hence CEML framework—which comprises a non-linear objective function and non-linear constraints.  I have also included the suggested element of noting the limitations and challenges of CML.

REVIEWER INSIGHT #5

Section CEML: This section highlights an interesting intersection between causal machine learning and economics, but it needs some improvements. The text is somewhat vague and lacks detailed explanations, making it difficult to understand how CEML actually works or differs from traditional methods. The section also overemphasizes the potential of CEML without adequately addressing the challenges and limitations. Additionally, the brief mention of big data and powerful software is not sufficiently expanded upon.

AUTHOR RESPONSE #5

This is great guidance. I have revised this section to build out how CEML actually works, in three defined stages, starting with the supervised labelling of data, building out the nature of the non-linear optimization, and then deployment in machine learning through an S-Learner that is able to accommodate non-linear optimization with both equalities and inequalities. I have also made mention of very specific software packages and tools that can be employed by researchers with an interest in testing this model in practice on data sets For example SLSQP, S-Learning, Python, EconML and SciPy. I have identified that the stopped at this point of setting the foundation as the next step of uncovering appropriate data sets is a big and separate phase that will take time.

REVIEWER INSIGHT #6

The "Results and Discussion" section has some issues that can affect its clarity and practical application. The text introduces concepts like expected utility theory and causal economics but does not explain them in simple terms. The section also criticizes existing economic models but does not provide concrete examples or clear explanations of how causal economics and machine learning would practically improve decision-making processes. In discussing macroeconomic outcomes, the text suggests that causal machine learning is underutilized but does not provide sufficient evidence or detailed case studies to support this claim. Additionally, while the text discusses the importance of causal coupling in economic policies, it fails to address the real-world challenges and resistance that such changes would face, particularly from entrenched systems. The discussion on public sector applications is interesting, but it spins over the complexity of implementing use-based taxation and the potential unintended consequences.

AUTHOR RESPONSE #6

In the revised draft, I have endeavored to layout the differences between traditional economic optimization and causal economic optimization. My goal has been to convey that expected utility and behavior economics are not yet complete and realistic models of human decision making. The first requires 100% rationality and the second incorporates biases, but not the full dynamic between costs that cause benefits over time--where some are certain and some are not. The difference in these models is massive. The comparative table in the CEML section lines them up side by side. I didn't want to turn this into an economics paper on causal economics vs, behavioral economics (BE) vs expected utility (EUT), but can add even more in this regard if you see fit. I wanted to provide the clarity, mathematically, that BE and EUT are constrained versions of CE, and that the additional parameters possible in CE allow for more robust application in machine learning research.

I have also added some discussion around the 'ideal', theory implications of the model and the stark reality that would be faced in implementation. I have received similar advice from globally renowned tax economists... that the implications would be powerful, but that implementation in the real world of vested interests would be daunting. For the purposes of this paper, I hope I have highlighted that reality further, as you advised.  

REVIEWER INSIGHT #7

Limitations and the future scope should be added with more clarity and the conclusion section needs flow and transition. 

AUTHOR RESPONSE #7

This was a significant gap and has now been added through paragraphs in the introduction and in the conclusion, which layout that this paper is presenting a model and specifics of how to implement with certain tools, and clarifies that the next phase is identification of appropriate data sets and collaborating with others as needed to deploy the model on real data sets. This clarity you were seeking on where the current research makes a contribution and where the next phase would clearly start is extremely helpful direction that I believe I have addressed. Thank you so much. I believe the flow and transition works better now with these changes. I am currently working with industry ML experts at top tech corporations to collaborate on specific data sets and am seeing great traction for putting this into application with their proprietary data.

Reviewer 3 Report

Comments and Suggestions for Authors

The paper reviews causal economics theory and concept, based on  causal coupling (CC), which models decisions as requiring upfront costs, in anticipation of future, uncertain causally-linked benefits.

The paper promotes in an interesting way causal economics and proposes potential applications. However, the relation of the causal economics to Machine learning is not presented in depth, but only as an idea.

Lines 17-20: "With the growing interest in natural experiments in statistics and 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." This is a nice idea but is not proved in the paper.

So, it would be beneficial to add in the last part of introduction, a paragraph on what is the paper's contribution and what is left for future work. Also, add future work section at the end of the paper, explaining what it really needs to be developed in order to connect causal economics with machine learning and build related applications (run AI models on CE foundations).

There is no compare and contrast section between the proposed theory and related theories: compare results to models based on traditional decision making models that focus only on rationality), as mentioned by the author.

3. Results and Discussion, should be better named Applications and Discussion, since there are no results.

Some refs are missing, eg. 29, 37.

As a final conclusion, if the author can add more depth to the machine learning part and provide comparison with other approaches, then the paper would be publishable and provide interest to the readers.

Author Response

REVIEWER INSIGHT #1

The paper reviews causal economics theory and concept, based on  causal coupling (CC), which models decisions as requiring upfront costs, in anticipation of future, uncertain causally-linked benefits.

The paper promotes in an interesting way causal economics and proposes potential applications. However, the relation of the causal economics to Machine learning is not presented in depth, but only as an idea.

Lines 17-20: "With the growing interest in natural experiments in statistics and 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." This is a nice idea but is not proved in the paper.

So, it would be beneficial to add in the last part of introduction, a paragraph on what is the paper's contribution and what is left for future work. Also, add future work section at the end of the paper, explaining what it really needs to be developed in order to connect causal economics with machine learning and build related applications (run AI models on CE foundations).

There is no compare and contrast section between the proposed theory and related theories: compare results to models based on traditional decision making models that focus only on rationality), as mentioned by the author.

AUTHOR RESPONSE #1

This is the most helpful advice! I have added the suggested changes to the end of the introduction, laying out specifically the contributions of this paper as well as the next phase of implementing on data sets. I have been very specific in the revised version about the the methods and specific tools. For example clarifying the optimization method for the non-linear value function and constraints (Sequential Least Squares Programming) using tools such as SciPy in Python and clarifying the use of an S-Learner meta learner to incorporate the causal machine learning (ex. using EconML in Python). In the CEML section, I have added a summary table of how CEML differs from alternate methods, both at theoretical levels and in terms of the practical tools of implementation. I believe this helps to clarify how this evolution better aligns the convergence of developments of causal machine learning and causal economics to produce a more 'human AI'.

REVIEWER INSIGHT #2

Results and Discussion, should be better named Applications and Discussion, since there are no results. Some refs are missing, eg. 29, 37.

AUTHOR RESPONSE #2

I have made these changes. Thank you for this feedback.

REVIEWER INSIGHT #3

As a final conclusion, if the author can add more depth to the machine learning part and provide comparison with other approaches, then the paper would be publishable and provide interest to the readers.

AUTHOR RESPONSE #3

This guidance is very helpful. As noted above, I have added additional and prescriptive depth to the machine learning part of the paper and included a summary table that shows CEML relative to other models—comparing both the structural differences and the inventory of tools applicable to putting the model to the test on data sets. I am hoping that the table makes clear the aligning evolution that has been taking place and inspires research that puts this very directly to the test.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

All the comments have been addressed satisfactorily.

 

Author Response

Thank your for you insight and guidance.

Reviewer 3 Report

Comments and Suggestions for Authors

At the last paragraph of the Introduction, the structure of the paper should be added.

Table 1 lacks caption, and has some grammatical errors.

The challenges related to Causal Economic Machine Learning, should be put in a separate subsection and analyzed more thoroughly.

There is problem with refs. Please avoid repetion. Also in ref 187, the paper and not the blog article should be put as a ref.

Author Response

REVIEWER INSIGHT #1

At the last paragraph of the Introduction, the structure of the paper should be added.

AUTHOR RESPONSE #1

This greatly improves readability. Thank you. I have added this per the guidance.

REVIEWER INSIGHT #2

Table 1 lacks caption, and has some grammatical errors.

AUTHOR RESPONSE #2

I have added a caption at the bottom of the table and fixed the grammatical errors.

REVIEWER INSIGHT #3

The challenges related to Causal Economic Machine Learning, should be put in a separate subsection and analyzed more thoroughly.

AUTHOR RESPONSE #3

I have broken the CEML section out into sub-sections to more clearly present it and to address the feedback. These sub-sections now introduce CEML and its proposed place in the evolution of AI modelling, followed by a discussion of implementation, and then through a new sub-section, discusses the challenges to implementation more appropriately as guided.

REVIEWER INSIGHT #4

There is problem with refs. Please avoid repetition.

AUTHOR RESPONSE #4

I have removed all duplication. Sorry about this. I normally use APA and didn't convert properly. It is now addressed.

REVIEWER INSIGHT #5

Also in ref 187, the paper and not the blog article should be put as a ref.

AUTHOR RESPONSE #5

I have carefully reviewed this, and unfortunately the blog article is what actually contains the point about using Python EconML. There is no paper by this author to cite on this point, or I certainly would have selected that instead. I hope it is still okay to cite an online article such as this, as I certainly don't make a habit of it.

 

Round 3

Reviewer 3 Report

Comments and Suggestions for Authors

The paper has been improved. I have enjoyed reading it.