📊 statistics

Comparison of generalised additive models and neural networks in applications: A systematic review

This systematic review of 143 studies comparing Generalised Additive Models (GAMs) and neural networks on tabular data finds no consistent superiority for either approach, suggesting they are complementary tools where GAMs offer a competitive performance advantage in smaller datasets while preserving interpretability.

Original authors: Jessica Doohan, Lucas Kook, Kevin Burke

Published 2026-06-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jessica Doohan, Lucas Kook, Kevin Burke

Original paper licensed under CC BY 4.0 (http://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 predict the future based on a list of facts (like weather data, house prices, or medical records). You have two main tools in your toolbox to do this:

The "Generalised Additive Model" (GAM): Think of this as a careful, transparent architect. It builds a prediction by looking at each piece of information one by one and drawing a smooth, understandable curve for how that specific piece affects the result. It's like saying, "If the temperature goes up, the ice cream sales go up like this." You can see exactly how it works.
The "Neural Network" (NN): Think of this as a super-fast, super-complex black box. It's a digital brain that connects all the dots at once, finding hidden patterns and wild relationships that a human might miss. It's incredibly powerful, but if you ask it why it made a prediction, it often can't explain itself clearly. It just gives you the answer.

For years, people have argued: "Which one is better?" This paper is a massive systematic review (a giant study of other studies) that looked at 143 different research papers involving 430 real-world datasets to settle the score.

Here is what they found, broken down simply:

1. The "Who Wins?" Question

The researchers didn't just ask the authors of those papers who they thought won; they took the actual numbers (the scores) and crunched them together.

The Result: There is no consistent winner.
The Analogy: It's like a boxing match where sometimes the careful architect wins, sometimes the black box wins, and often they end in a tie. Neither tool is universally superior. If you look at the most common ways to measure success (how close the prediction is to reality), they are generally on the same level.

2. When Does the "Black Box" Shine?

The study found that the Neural Network (the black box) tends to have a slight edge in two specific situations:

When the data is huge: If you have a massive dataset (like millions of rows), the black box can sometimes find patterns the careful architect misses.
When there are many variables: If you are juggling hundreds of different factors, the black box handles the complexity well.

However, the paper notes that this advantage is shrinking. As time goes on, the careful architect (GAM) is catching up, and the gap is getting smaller.

3. When Does the "Architect" Shine?

The GAM (the transparent model) remains a very strong competitor, especially in smaller datasets.

The Big Plus: The main reason to choose the GAM isn't just about being slightly more accurate; it's about interpretability. Because it's built like a clear blueprint, you can understand why it made a decision. In fields like science or medicine, knowing why is often just as important as being right.

4. The "Missing Manual" Problem

One of the most interesting findings wasn't about the models, but about how the studies were written. The researchers found that many papers were missing crucial details.

The Analogy: Imagine reading a recipe for a cake, but the author forgot to list the amount of sugar or the oven temperature. You can't really tell if the cake turned out good because of the recipe or just luck.
The Reality: In many of the 143 papers reviewed, the authors didn't report how big their dataset was, how many variables they used, or how complex their neural network was. This makes it hard to trust the results or repeat the experiments. The paper calls for scientists to start writing down these details more clearly.

The Final Verdict

The authors conclude that these two tools shouldn't be seen as enemies fighting for the top spot. Instead, they are complementary tools.

If you need to squeeze out every tiny bit of extra accuracy from a massive, complex dataset, the Neural Network might be your go-to.
If you need a model that is accurate and explainable (so you can tell a doctor, a judge, or a scientist exactly how it works), the GAM is often the better choice.

The paper suggests that for most everyday "tabular" data (rows and columns of numbers), the performance difference is so small that you should choose the model based on whether you care more about raw power or clear understanding.

Technical Summary: Comparison of Generalised Additive Models and Neural Networks in Applications

Problem Statement
Neural networks (NNs) have become dominant in predictive modelling, particularly for tasks involving image recognition and natural language processing, often prioritizing predictive accuracy over interpretability. Conversely, Generalised Additive Models (GAMs) offer a flexible, non-linear statistical framework that retains interpretability through an additive structure. While both are state-of-the-art, there is a lack of systematic understanding regarding their relative performance on real-world tabular data. Previous systematic reviews have largely focused on specific domains (e.g., transportation, heart health) or compared NNs against simpler linear models (e.g., linear or logistic regression), failing to account for the non-linear capabilities of modern statistical models like GAMs. This paper addresses the gap by conducting a systematic review and meta-analysis comparing GAMs and NNs across a broad spectrum of application areas.

Methodology
The study adheres to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.

Data Collection: A search of the Scopus database yielded 445 documents. After screening for eligibility (peer-reviewed, English, full-text, empirical comparison of GAMs and NNs on the same dataset, and availability of performance metrics), 143 papers were selected. These papers contained 430 distinct datasets.
Data Extraction: Attributes were extracted at both the paper level (domain, journal, year) and dataset level (sample size, number of predictors, problem type, neural network architecture, complexity, and performance metrics).
Statistical Analysis: To address Research Question 3 (factors associated with performance differences), the authors employed linear mixed-effects modelling. This approach was chosen to account for the hierarchical structure of the data, where multiple datasets are nested within single papers, introducing intra-paper dependence.
- Response Variables: The analysis focused on the three most frequently reported metrics: Root Mean Square Error (RMSE), $R^2$ , and Area Under the Curve (AUC). To ensure scale invariance across diverse datasets, the authors computed log-ratios of performance metrics (e.g., $\log(\text{RMSE}_{\text{GAM}}/\text{RMSE}_{\text{NN}})$ ), converting $R^2$ and AUC to loss measures ( $1-R^2$ , $1-\text{AUC}$ ) to align with RMSE.
- Predictors: Fixed effects included publication year, domain area, sample size, number of predictors, and neural network complexity. Pairwise interactions with the year of publication were included to assess temporal trends. A random intercept for "Paper ID" was included to capture paper-specific heterogeneity.
- Handling Missingness: Given that sample size, predictor count, and complexity were frequently unreported, missing values were converted into categorical "Unknown" levels rather than being discarded, allowing for the assessment of whether missingness itself correlated with outcomes.

Key Results

Overall Performance: Across the 430 datasets, there was no consistent evidence of superiority for either GAMs or neural networks when considering the most common metrics (RMSE, $R^2$ , AUC). The intercepts of the mixed-effects models were not statistically significantly different from zero, indicating that, on average, neither model class systematically outperforms the other.
Neural Network Complexity: For RMSE, neural network complexity and its interaction with time were significant. In 2011, medium and large complexity networks performed worse relative to GAMs compared to smaller networks. However, this negative effect diminished over time; by 2024, complexity was no longer associated with a performance advantage or disadvantage relative to GAMs.
Dataset Dimensions (Sample Size and Predictors): For AUC, larger sample sizes and higher numbers of predictors were historically associated with better neural network performance relative to GAMs. However, the interaction with time showed that this advantage has narrowed significantly. In recent years, no statistically significant association exists between dataset dimensions and the relative performance of the two models.
Reporting Quality: A significant finding was the high rate of missing information in the literature. Only 21% of datasets reported all three key attributes (sample size, predictors, and complexity). Unreported neural network complexity accounted for the majority of missingness (reported in only 35% of datasets).
Intra-Paper Correlation: The Intraclass Correlation Coefficient (ICC) was high (ranging from 0.76 to 0.97), indicating substantial variability between papers that could not be explained by the observed dataset-level variables. This suggests that unreported methodological details (e.g., preprocessing, hyperparameter tuning) likely drive performance differences more than the model class itself.

Significance and Conclusions
The paper claims that GAMs and neural networks should be viewed as complementary approaches rather than competitors for tabular data. The study concludes that:

Performance Trade-offs: The performance trade-off between the two methods is modest. Neither model class holds a consistent advantage across diverse applications.
Interpretability: Given the comparable performance, GAMs remain a robust choice, particularly in scientific research where interpretability and transparency are paramount.
Reporting Standards: The review highlights a critical lack of standardised reporting in machine learning literature, specifically regarding dataset characteristics and model complexity. This opacity limits reproducibility and the ability to draw definitive conclusions about model superiority.
Future Directions: The authors suggest that future research should focus on standardised reporting protocols (e.g., "model cards") and potentially conducting independent re-analyses of datasets from this review using consistent protocols to isolate the effects of modelling choices from reporting inconsistencies.

The paper emphasizes that while neural networks excel in complex, non-tabular domains, they do not offer a decisive performance advantage over GAMs for tabular data, making the choice between them dependent on the specific value placed on interpretability versus predictive gains.

1. The "Who Wins?" Question

2. When Does the "Black Box" Shine?

3. When Does the "Architect" Shine?

4. The "Missing Manual" Problem

The Final Verdict

More like this