📊 statistics

Analytical Validation of Wistats v3.0 Through Comparison With SPSS and scikit-learn

This study validates the analytical accuracy of Wistats v3.0 by demonstrating its complete agreement with established statistical software (SPSS) and machine learning libraries (scikit-learn) across diverse analytical scenarios, confirming its reliability as a user-centered platform for automated, publication-ready scientific reporting.

Original authors: Levent Korkmaz

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Levent Korkmaz

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 bake a complex cake, but you aren't sure if you should use a gas oven or an electric one, or whether you need to preheat it. You have a trusted, old-school cookbook (let's call it SPSS) that gives you the exact recipe and baking times, but the instructions are written in a very technical, confusing language. You also have a brilliant, high-tech kitchen robot (scikit-learn) that can bake the cake perfectly if you program it with code, but you need to know how to code to use it.

Now, imagine a new, friendly kitchen assistant named Wistats. Its job isn't to invent new recipes or replace the old cookbook or the robot. Instead, it acts as a translator and a guide. It looks at your ingredients (your data), decides which oven and recipe to use automatically, bakes the cake, and then writes the instructions in plain English so you can understand exactly what happened.

This paper is essentially a "taste test" to see if Wistats is telling the truth. The author, Dr. Levent Korkmaz (who also built Wistats), wanted to prove that Wistats doesn't make mistakes when it translates or calculates things.

Here is how they tested it, using simple analogies:

1. The Setup: A "Practice Kitchen"

Since they didn't want to risk ruining real patient data, they created a synthetic dataset. Think of this as a "practice dough" made up of fake numbers representing things like age, blood pressure, and hospital stay length. They made sure this dough had all kinds of textures: some parts were smooth (normal distribution), some were lumpy (non-normal), and some were distinct chunks (categorical data like "Male/Female").

2. The Test: The "Three-Way Mirror"

The researchers ran the same "practice dough" through three different machines:

SPSS: The gold-standard, old-school cookbook (the reference for statistics).
Manual Python Code: A human expert writing code from scratch to use the robot (the reference for machine learning).
Wistats: The new assistant.

They asked a simple question: "Do all three machines give the exact same answer?"

3. The Results: Perfect Mirrors

The study found that Wistats was 100% identical to the other two in every single test they tried. Here is what they checked:

The "Smell Test" (Distribution): Before baking, you need to know if your dough is ready. Wistats checked if the data was "normal" or "weird" just like SPSS did. They agreed on every single ingredient.
The "Comparison Test" (Statistics): They asked, "Is Group A different from Group B?" Whether they used a t-test, a chi-square test, or an ANOVA, Wistats gave the exact same "Yes/No" answer and the exact same numbers as the SPSS cookbook.
The "Prediction Test" (Machine Learning): They tried to predict things like "Will this patient respond to treatment?" or "How long will they stay in the hospital?" Wistats used the same algorithms (like Support Vector Machines or Random Forests) as the human coder. The accuracy scores, error rates, and predictions were identical.

4. The Special Sauce: No "Magic" or "Guessing"

One of the most important points the paper makes is about how Wistats decides what to do.

Some new AI tools use "Large Language Models" (like a very smart but sometimes hallucinating chatbot) to guess what analysis you need.
Wistats does not do this. It uses a rulebook. Think of it like a traffic light system: "If the data is red (non-normal), stop and use the Mann-Whitney test. If it's green (normal), go and use the t-test."
Because it follows strict rules and not "guesses," if you put the same data in twice, you get the exact same result twice. This makes it reproducible and transparent.

5. The Goal: Making the Cake Accessible

The paper concludes that Wistats isn't trying to replace the expert chefs (statisticians) or the powerful robots (programmers). Instead, it's trying to help the home cooks (researchers who aren't experts in math or coding).

It takes the scary, technical numbers and turns them into narrative stories (e.g., "The treatment group stayed in the hospital significantly shorter than the control group").
It automatically creates tables that look ready to be pasted into a scientific paper.

The Bottom Line

The paper claims that Wistats v3.0 is a reliable, honest translator. It takes complex data, runs it through established, trusted mathematical engines (the same ones SPSS and Python use), and presents the results in a way that is easy to read and write up. The "taste test" proved that the new assistant doesn't change the flavor of the cake; it just serves it on a nicer plate with a clearer description.

A Note on Limits: The paper admits this test was done on a "practice dough" (synthetic data) and only tested specific types of recipes. It didn't test every possible statistical method in the world, but for the ones they tried, the results were perfect matches.

Technical Summary: Analytical Validation of Wistats v3.0

Problem Statement
Contemporary scientific research increasingly relies on heterogeneous data structures and complex analytical workflows, yet the selection of appropriate statistical methods and the accurate interpretation of outputs remain significant challenges, particularly for researchers lacking formal training in biostatistics or programming. Conventional software packages (e.g., SPSS, R, Stata) and machine learning libraries (e.g., scikit-learn) offer robust capabilities but often require manual test selection, deep methodological expertise, and extensive effort to transform technical outputs into publication-ready formats. This manual process introduces risks of methodological errors, inappropriate test selection, and inconsistencies in reporting, which can compromise reproducibility and the reliability of published findings. Furthermore, while generative artificial intelligence has emerged as a tool for interpretation, its probabilistic nature raises concerns regarding hallucination and reproducibility in scientific contexts. There is a need for an analytical environment that integrates automated test selection, deterministic interpretation, and reproducible machine learning workflows without sacrificing methodological rigor.

Methodology
This study evaluated the analytical validity of Wistats v3.0, a Python-based platform developed by WisdomEra, by comparing its outputs against established reference environments. The validation framework utilized a synthetic dataset intentionally designed to encompass a broad spectrum of analytical scenarios, including normally and non-normally distributed continuous variables, binary and multiclass categorical variables, and predictive modeling applications.

The validation process was divided into five domains:

Distributional Assessment: Normality was evaluated using Shapiro–Wilk statistics, skewness, and kurtosis. Wistats outputs were compared against IBM SPSS Statistics version 32.0.
Comparative Statistical Analyses: The study validated independent-samples t-tests, Mann–Whitney U tests, Pearson and Spearman correlations, one-way ANOVA, Kruskal–Wallis tests, and Pearson chi-square analyses against SPSS.
Multivariable Logistic Regression: Regression coefficients, odds ratios, confidence intervals, and significance levels were compared against SPSS.
Machine Learning Classification: Models for predicting treatment response (using Support Vector Machines, Random Forest, and Logistic Regression) were validated against manually implemented scikit-learn (version 1.7.2) workflows in Python. Metrics included accuracy, cross-validation accuracy, and F1-scores.
Machine Learning Regression: Models predicting hospital length of stay were validated against manual scikit-learn implementations using $R^2$ , Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE).

Wistats v3.0 operates as an interface layer utilizing established scientific computing libraries (SciPy, statsmodels, scikit-learn). Crucially, its automated test selection and interpretation engines rely on deterministic rule-based algorithms rather than large language models or generative AI, ensuring that identical inputs consistently produce identical outputs. Agreement was defined as identical analytical outputs after numerical rounding.

Key Contributions

Automated Test Selection: The study demonstrates a rule-based framework that automatically selects appropriate statistical tests (parametric vs. non-parametric) based on variable types, distributional assumptions (Shapiro–Wilk, skewness, kurtosis), and group structures, reducing the burden of manual decision-making.
Deterministic Interpretation and Reporting: Unlike generative AI systems, Wistats employs template-based narrative generation and structured table creation. This approach ensures transparency and reproducibility, transforming technical outputs into publication-ready summaries without probabilistic variability.
Integrated Validation: The paper provides a comprehensive validation of both traditional statistical procedures and machine learning workflows within a single platform, confirming that the integration layer does not introduce computational discrepancies when compared to industry-standard reference implementations.

Results
The study reported 100% agreement between Wistats v3.0 and the reference environments across all evaluated scenarios:

Distributional Analysis: Normality assessments (Shapiro–Wilk p-values, skewness, and kurtosis) were identical between Wistats and SPSS for all variables.
Comparative Statistics: All comparative procedures (t-tests, Mann–Whitney U, chi-square, ANOVA, Kruskal–Wallis, and correlations) yielded identical p-values between Wistats and SPSS.
Logistic Regression: Regression coefficients, odds ratios, confidence intervals, and significance levels matched SPSS outputs exactly.
Machine Learning Classification: Predictive performance metrics (accuracy, cross-validation accuracy, F1-score) for treatment response prediction models ranged from 0.66 to 0.70 in accuracy. All metrics were identical between Wistats and manual scikit-learn implementations.
Machine Learning Regression: Models for predicting hospital length of stay showed $R^2$ values between −0.01 and 0.01, with MAE ranging from 2.98 to 3.28 and RMSE from 4.00 to 4.06. All performance indicators were identical between Wistats and scikit-learn.

Significance and Claims
The authors conclude that Wistats v3.0 represents a reliable, reproducible, and user-centered analytical environment. The primary significance of the platform lies not in the development of novel statistical methodologies, but in its ability to facilitate interpretation and scientific reporting for users with varying levels of expertise. By automating test selection and generating publication-ready narratives and tables through deterministic rules, Wistats aims to reduce technical complexity and methodological errors while maintaining strict consistency with established statistical and machine learning libraries.

The paper modestly claims that Wistats is designed to complement, rather than replace, comprehensive statistical software or advanced programming environments. It is positioned as a tool to improve accessibility and reproducibility in scientific communication, particularly for researchers who require robust analytical workflows but may lack extensive programming skills. The study emphasizes that while the platform automates individual analytical findings and reporting, the synthesis of evidence, hypothesis generation, and broader scientific interpretation remain the responsibility of the investigator. Future validation using large-scale real-world datasets and diverse analytical environments is suggested to further confirm the platform's generalizability.