📈 economics

Machine Learning Feature Importance vs. Fixed-Effects Regression: Which Variables Actually Drive CSR–Firm Value Relationships in Africa?

This study demonstrates that combining machine learning feature importance with fixed-effects regression reveals critical nonlinearities and interaction effects—such as thresholds in board independence and context-dependent CEO duality—that traditional linear models overlook, thereby offering a superior two-step framework for understanding the drivers of CSR-firm value relationships in African financial markets.

Original authors: Mutala Zubeiru, Professor Uwuigbe Olubukunola Ranti, Osman Issah

Published 2026-07-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mutala Zubeiru, Professor Uwuigbe Olubukunola Ranti, Osman Issah

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out what makes a specific type of car (let's call it an "African Business Car") valuable. You have a massive garage with 80 of these cars, and you've tracked their performance over 10 years. You have a list of 83 possible things that might make them valuable: how clean they are, how big the engine is, who drives them, the quality of the road they are on, and even the weather.

For a long time, researchers have used a standard, old-school tool (called Fixed-Effects Regression) to figure out which factors matter. Think of this tool like a straight ruler. It's great at measuring things that go up in a straight line (e.g., "more engine size = more speed"). But it struggles if the relationship is curved, or if something only works after you reach a certain point.

This paper introduces a new, high-tech tool called Machine Learning (specifically Random Forests and Gradient Boosting). Think of this tool like a smart, flexible 3D scanner. It doesn't just look for straight lines; it can detect complex curves, hidden thresholds, and how different factors twist together.

Here is what the authors found when they compared the "Straight Ruler" against the "Smart 3D Scanner":

1. The Agreement: The "Big Four"

Both tools agreed on the top four most important things that drive value. Whether you use the old ruler or the new scanner, these four factors are the undisputed champions:

CSR Score: How well the company talks about its social and environmental responsibility.
ROA (Return on Assets): How efficiently the company uses its money to make a profit.
Institutional Quality: How good the country's laws and government are.
Firm Size: How big the company is.

The paper says this is good news because it proves that the old studies were right about the most obvious factors.

2. The Disagreement: Where the "Ruler" Failed

This is where the paper gets interesting. The "Smart 3D Scanner" found that the "Straight Ruler" was missing some crucial details because those details don't follow a straight line.

A. The "Switch" Effect (Thresholds)
The ruler told the researchers that Board Independence (having outside directors on the board) didn't really matter. But the scanner showed a "light switch" effect.

The Analogy: Imagine a team of people trying to stop a bad decision. If you have 30% independent people, they are too few to stop anything (the light is off). But once you cross 40%, they suddenly have enough power to make a real difference (the light turns on).
The Finding: The old ruler couldn't see this "switch" and thought the factor was unimportant. The scanner saw that once you hit that 40% mark, the company value jumps.

B. The "Tail Risk" Effect
The ruler said Political Risk (unstable governments) wasn't a big deal. The scanner saw that risk is like a cliff.

The Analogy: Driving on a road is fine until you hit a cliff. Below a certain level of risk, nothing happens. But once the risk gets too high (the cliff edge), the value of the car crashes. The ruler just saw the flat road and missed the cliff.

C. The "Hidden Ingredient" Effect
The researchers used to look at a single "CSR Score" (a big bucket of everything). The ruler said this bucket was important.

The Finding: The scanner looked inside the bucket and found that Environmental disclosure was the real star, doing 2.3 times more work than Social disclosure.
The Analogy: It's like judging a smoothie by its total weight. The ruler says "heavy smoothie = good." The scanner says, "Wait, the weight is mostly from the spinach (Environment), not the strawberries (Social). If you want value, focus on the spinach." The old method was hiding this difference by mixing them together.

D. The "Fake Correlation" Trap
The ruler gave high importance to things like Sales Growth and Leverage (debt).

The Finding: The scanner said, "These are just mechanical numbers that happen to move with value, but they aren't actually causing the value."
The Analogy: It's like seeing a rooster crow and the sun rise. The ruler thinks the rooster causes the sun to rise. The scanner knows the rooster is just a side effect and isn't the real driver.

3. The New Recipe: "Discover, Then Infer"

The authors propose a new way to do research, like a two-step cooking process:

Step 1 (The Scanner): Use Machine Learning first to scan all 83 ingredients and find out which ones actually have flavor (predictive power) and which ones are just garnish.
Step 2 (The Ruler): Take only the important ingredients found in Step 1 and use the traditional statistical ruler to prove why they work and how they interact.

Summary

The paper concludes that while the old methods were right about the "big picture" (CSR, Profit, Size, and Country Laws matter), they were blind to the complex rules that govern the details. They missed the "tipping points" (like the 40% board rule) and the "hidden drivers" (like environmental vs. social disclosure).

By using the "Smart 3D Scanner" (Machine Learning) to guide the "Straight Ruler" (Regression), researchers can finally see the full, complex picture of what drives business value in Africa, rather than just seeing a flat, simplified version.

Technical Summary: Machine Learning Feature Importance vs. Fixed-Effects Regression in African CSR Research

Problem Statement
Empirical research on the relationship between Corporate Social Responsibility (CSR) disclosure and firm value in African capital markets has historically relied on fixed-effects (FE) panel regressions. While this approach has established a general consensus on the positive direction of the CSR–value link, it suffers from two critical methodological weaknesses that obscure the true drivers of this relationship:

Variable Selection Bias: Pre-specifying a limited set of controls risks omitting crucial, under-theorized moderating variables.
Linearity Assumption: Standard FE regressions impose linear additive effects, failing to detect non-linearities, threshold effects, or heterogeneous treatments that are theoretically plausible in CSR contexts.

Consequently, there is ambiguity regarding which variables robustly drive firm value, as traditional methods may miss complex interactions or misattribute significance to mechanically correlated variables.

Methodology
This study employs a two-step comparative framework using a balanced panel of 800 firm-year observations (80 firms across South Africa, Nigeria, Kenya, and Ghana) from 2015 to 2024. The dataset includes 83 predictors spanning CSR disclosure (GRI G3.1), financial performance, corporate governance, and institutional quality.

Machine Learning (ML) Discovery: The authors utilize Random Forest (RF) and Gradient Boosting (GB) algorithms to generate data-driven feature importance rankings. To ensure interpretability and comparability with econometric results, Shapley Additive Explanations (SHAP) are applied to decompose model predictions into individual feature contributions. The models are trained on data from 2015–2022 and tested on 2023–2024 to prevent temporal leakage.
Econometric Benchmark: A two-way fixed-effects regression (entity and year fixed effects) serves as the causal inference benchmark, utilizing the same 83 predictors.
Concordance Framework: The study compares the ranking of variable importance across RF, GB, SHAP, and standardized FE regression coefficients. Variables are classified into three categories:
- Concordant: Ranks within 5 positions across all methods.
- FE-Underweighted: ML ranks the variable significantly higher (≥8 positions) than FE, suggesting FE misses non-linear or threshold effects.
- FE-Overweighted: FE ranks the variable significantly higher than ML, suggesting the FE significance stems from mechanical within-firm correlations rather than genuine predictive power.

Key Results
The analysis reveals a high degree of concordance (Spearman $\rho = 0.91$ ) for the top-tier predictors, validating the selection of key variables in prior literature. The top four drivers identified by all methods are: CSR Score, Return on Assets (ROA), Institutional Quality, and Firm Size.

However, significant discordance emerges in the mid-tier variables, revealing specific limitations of the FE approach:

FE-Underweighted Variables (Governance & Institutional): ML identifies Board Independence, Environmental Disclosure, CEO Duality, and Political Risk as substantially more important than FE regression suggests. SHAP analysis uncovers the reasons for this discrepancy:
- Threshold Effects: Board independence only generates value when it exceeds a 40% threshold; below this level, independent directors lack the critical mass to constrain management.
- Non-linearity: Environmental disclosure drives firm value 2.3 times more than social disclosure, a nuance masked by aggregate CSR scores.
- Contextual Interactions: The negative impact of CEO duality is contingent on institutional quality and is significant only in Ghana and Nigeria, not South Africa.
FE-Overweighted Variables (Financial Structure): Variables such as Sales Growth, Leverage, and Current Ratio appear statistically significant in FE models but possess low predictive importance in ML models. The study attributes this to mechanical correlations with Tobin's Q that do not reflect genuine causal drivers of value.
Dependent Variable Robustness: When using Standardized Economic Value Added (EVA) as the dependent variable, the CSR score drops in importance, confirming that CSR operates primarily through market perception channels (Tobin's Q) rather than direct operating performance (EVA).

Key Contributions

Methodological Innovation: The paper introduces a replicable "ML to discover, regression to infer" workflow. It proposes using ML feature importance as a diagnostic tool to identify non-linearities and thresholds before specifying regression models, thereby reducing specification error.
Substantive Insights for African Markets:
- It demonstrates that environmental disclosure is the primary driver of the CSR–value relationship in Africa, overshadowing social dimensions.
- It identifies a 40% board independence threshold, suggesting that governance reforms must achieve a critical mass to be effective, rather than merely aiming for average representation.
- It highlights the country-contingent nature of governance mechanisms (e.g., CEO duality), arguing against one-size-fits-all governance policies across African markets.
Taxonomy of Specification Error: The study provides a classification system for variables that are either underweighted (due to non-linearity) or overweighted (due to mechanical correlation) in standard panel regressions.

Significance and Claims
The authors claim that this study provides the first systematic evaluation of the agreement between ML and FE regression in the context of African CSR research. The findings suggest that while traditional FE models are valid for top-tier linear drivers, they systematically fail to capture the complexity of governance and institutional variables.

The paper does not claim that ML replaces causal inference; rather, it argues that ML is essential for variable discovery and specification diagnosis. By revealing hidden thresholds and interactions, ML allows researchers to construct more accurate regression models that account for non-linear realities. The proposed concordance framework is presented as a necessary pre-specification step for future panel data research in emerging markets to avoid omitted variable bias and misinterpretation of statistical significance.

1. The Agreement: The "Big Four"

2. The Disagreement: Where the "Ruler" Failed

3. The New Recipe: "Discover, Then Infer"

Summary

Technical Summary: Machine Learning Feature Importance vs. Fixed-Effects Regression in African CSR Research

More like this