An Integrated Multi-Method Framework for Gender-Based Violence Research: A Synthetic Data Demonstration Using Kenya Demographic and Health Survey Parameters

This study proposes and demonstrates a five-phase integrated analytical framework combining predictive machine learning and causal inference methods on synthetic Kenya Demographic and Health Survey data to identify key GBV predictors and test mediation pathways, serving as a proof-of-concept for future application to real-world longitudinal datasets.

The Gist

Imagine you are trying to solve a massive, complex mystery: Why does violence against women happen?

For a long time, researchers have been trying to solve this puzzle using two different tools, but they've been using them in separate rooms.

• Room A (The Statisticians): They use traditional math to find average causes. It's like looking at a map and saying, "On average, people who live near the river get wet." It's accurate, but it misses the details of how the water gets there.
• Room B (The Machine Learners): They use super-smart computers (AI) to predict who will get hurt next. It's like having a weather app that says, "It will rain here tomorrow!" It's great at prediction, but it can't explain why the rain is falling.

This paper is about building a bridge between Room A and Room B.

The author, Grold Otieno Mboya, created a new "Five-Phase Detective Framework" to solve the mystery of Gender-Based Violence (GBV) in Kenya. Since real-world data can be messy and hard to get, he didn't use real people's private stories. Instead, he built a virtual world (a computer simulation) that perfectly mimics the real demographics of Kenya. Think of it like a flight simulator: the pilot (researcher) can crash the plane a thousand times to learn how to fly, without anyone getting hurt.

Here is how the "Five-Phase Detective Framework" works, explained simply:

Phase 1: The Roll Call (Descriptive Epidemiology)

First, the detective takes a census of the virtual village. They count how many women are in the village, how old they are, how much money they have, and how many have experienced violence.

• The Result: They found that about 1 in 4 women in this virtual village experienced violence in the last year. This matches real-life statistics, proving the simulation is working correctly.

Phase 2: The Crystal Ball (Machine Learning)

Next, they handed the data to a "Crystal Ball" (a Random Forest algorithm). The goal wasn't to understand why yet, but just to guess who is most likely to be hurt.

• The Analogy: Imagine a security guard at a club trying to spot troublemakers. The Crystal Ball looked at thousands of clues and shouted, "Partner's drinking is the biggest red flag!" It also flagged childhood trauma and fighting at home.
• The Catch: The Crystal Ball was okay at guessing (about 75% accurate), but not perfect. It was only slightly better than just guessing "No violence" for everyone, because violence is a rare event compared to peace.

Phase 3: The Interrogation (Logistic Regression)

Now, the detective puts the suspects in a room and asks, "Is it really you, or is it just your friend?" This is where they use traditional math to separate the real causes from the noise.

• The Result: They confirmed the Crystal Ball's suspicions.
  • Partner's Alcohol Use: This was the "King of Risk." If a partner drinks heavily, the risk of violence jumps 6.6 times.
  • Childhood Trauma: Women who suffered as children were 2 times more likely to face violence as adults.
  • Community Norms: Living in a neighborhood where men are expected to be "bosses" (patriarchy) also increased the risk.
• The Big Win: The study proved that you can't just look at one thing (like the woman's education). You have to look at the whole ecosystem (Individual + Relationship + Community). A model that looked at all three levels fit the data much better than one that looked at just one.

Phase 4: The "Domino Effect" Test (Mediation Analysis)

The detective wanted to know: Does alcohol cause violence directly, or does it cause fighting, which then causes violence? (The Domino Effect).

• The Analogy: They set up a line of dominoes: Alcohol → Fighting → Violence. They pushed the first one.
• The Surprise: The dominoes didn't fall. In this specific simulation, the alcohol didn't need to cause a fight first; it seemed to lead to violence directly. The "middleman" (marital conflict or mental health) wasn't the main bridge.
• Why? The author admits this might be because the simulation was built to show strong direct effects. In the real world, the dominoes might fall differently.

Phase 5: The Stress Test (Sensitivity Analysis)

Finally, the detective asked, "What if we only had a tiny village? Or a huge city?" They ran the whole test again with different numbers of people.

• The Result: The main suspects (Alcohol, Trauma, Conflict) stayed the same no matter how big or small the village was. This proves the detective's method is robust and reliable.

The Final Verdict (Conclusion)

This paper is a proof-of-concept. It's like a pilot project for a new car. The author isn't saying, "We have solved the mystery of violence in Kenya." Instead, he is saying:

"We built a new, powerful tool that combines the best of AI (prediction) and traditional math (explanation). We tested it in a virtual world, and it worked perfectly. It found the same bad guys that real-world studies find, but it did it in a way that is more thorough."

What does this mean for the real world?

1. Don't just guess: We need to use both AI and traditional stats together to understand violence.
2. Focus on the big three: To stop violence, we must tackle alcohol abuse, childhood trauma, and community norms all at the same time.
3. The Next Step: Now that the "flight simulator" works, researchers need to take this framework and fly it with real data from real people to see if it can truly save lives.

In short: The author built a digital laboratory to test a new way of solving the violence puzzle. The test was a success, and the new method points clearly at alcohol, trauma, and community culture as the keys to unlocking the solution.

Technical Summary

1. Problem Statement

Current research on Gender-Based Violence (GBV) suffers from a "methodological singularity," where studies typically operate in one of two silos:

• Traditional Regression: Effective for estimating adjusted associations and controlling confounding but often assumes linearity and struggles with high-dimensional, non-linear interactions inherent in ecological systems.
• Machine Learning (ML): Offers superior predictive accuracy and handles complex interactions but is frequently criticized for lacking interpretability and causal validity.

Furthermore, while risk factors are identified, the specific causal pathways (mechanisms) through which distal factors (e.g., childhood trauma) influence proximal outcomes (e.g., Intimate Partner Violence) remain under-investigated using formal mediation analysis. There is a critical need for an integrated framework that combines predictive power with causal inference to understand who is at risk and why.

2. Methodology

The study employs a cross-sectional simulation design to develop and validate a five-phase analytical framework. Instead of using raw survey data, the authors generated a synthetic dataset ($N=3,000$) parameterized to reflect the structural properties, demographics, and risk factor distributions of the 2022 Kenya Demographic and Health Survey (KDHS).

Data Generation:

• Tool: R statistical environment (version 4.5.2) using the simstudy package.
• Structure: Hierarchical structure with 100 community clusters.
• Ground Truth: Causal relationships and effect sizes were pre-specified (e.g., setting the odds ratio for partner alcohol use at ~6.0) to allow for internal validation of the framework.
• Population: Women aged 15–49 in intimate relationships.

The Five-Phase Analytical Framework:

1. Descriptive Epidemiology: Validated simulated data against known KDHS patterns (prevalence, bivariate associations).
2. Machine Learning (Random Forest): A classification model (500 trees) was trained on 70% of the data to identify high-priority predictors without linear assumptions. Variable importance was assessed via Mean Decrease in Accuracy (MDA).
3. Multivariable Logistic Regression: Nested models were fitted to estimate adjusted odds ratios (aOR):
   • Model 1: Individual factors only.
   • Model 2: Individual + Relationship factors.
   • Model 3: Full Ecological Model (Individual + Relationship + Community).
   • Model fit was compared using Akaike Information Criterion (AIC).
4. Causal Mediation Analysis: Used the potential outcomes framework to test three hypothesized pathways:
   • Partner Alcohol $\to$ Marital Conflict $\to$ GBV
   • Childhood Trauma $\to$ Mental Health $\to$ GBV
   • Community Patriarchy $\to$ Decision Power $\to$ GBV
   • Significance was determined via 1,000 bootstrap samples.
5. Synthesis & Sensitivity Analysis: Integrated findings across methods and tested framework stability across varying sample sizes ($N=500, 1000, 5000$).

3. Key Results

Predictive Performance (Random Forest):

• Accuracy: 74.6% (AUC = 0.711).
• Baseline Comparison: This offered only a modest gain (~1.5%) over a null classifier (predicting "no GBV" for all), which achieved 73.1% accuracy due to the class imbalance (26.9% prevalence).
• Sensitivity/Specificity: Low sensitivity (37.3%) but high specificity (89.8%), indicating the model correctly identified non-cases but missed many true GBV cases.
• Top Predictors: Partner alcohol use (100% relative importance), childhood trauma (17.3%), and marital conflict (10.5%).

Causal Inference (Logistic Regression):

• Model Fit: The full ecological model (Model 3) demonstrated significantly superior fit compared to single-level models ($\Delta AIC = 67.5$ vs. Model 1), empirically validating the multi-level approach.
• Key Risk Factors (Adjusted):
  • Partner Alcohol Use: Strongest predictor ($aOR = 6.60$, 95% CI: 5.50–7.94).
  • Childhood Trauma: Nearly doubled odds ($aOR = 1.99$, 95% CI: 1.60–2.47).
  • Marital Conflict: Significant positive association ($aOR = 1.15$).
  • Community Patriarchy: Significant positive association ($aOR = 1.08$).
• Protective Factors: Higher education, stronger legal implementation, and greater decision-making power.

Mediation Analysis:

• Result: None of the three hypothesized indirect pathways reached statistical significance ($ACME \ p > 0.05$).
• Interpretation: In this specific simulation, risk factors operated primarily through direct routes rather than the specified mediators. This highlights the difficulty of detecting mediation in cross-sectional designs and suggests the data-generating process prioritized strong direct effects.

Sensitivity Analysis:

• The framework proved robust for sample sizes $N \geq 1,000$. Variable importance rankings and primary effect estimates (e.g., alcohol use $aOR$) remained stable. Smaller samples ($N=500$) resulted in wider confidence intervals and loss of significance for weaker community-level predictors.

4. Key Contributions

• Methodological Integration: Proposes a replicable, five-phase blueprint that sequentially combines ML (for variable prioritization and non-linear pattern detection) with regression (for interpretability and causal estimation) and mediation analysis.
• Empirical Validation of Ecological Theory: Provides quantitative evidence that multi-level models significantly outperform single-level models in GBV research, supporting Heise's ecological framework.
• Distinction Between Prediction and Causation: Demonstrates that a variable's predictive importance (e.g., distance to services) does not always equate to its causal significance after adjustment, and vice versa (e.g., decision-making power was a significant protective factor despite lower predictive rank).
• Synthetic Data Benchmarking: Establishes a proof-of-concept using KDHS parameters to test analytical pipelines before applying them to complex real-world data.

5. Significance and Limitations

Significance:
The study offers a "proof-of-concept" that GBV research can move beyond siloed methods. It suggests that holistic, multi-level interventions are necessary, as risk is driven by the interaction of individual history, relationship dynamics, and community norms. The framework provides a rigorous path for future research to move from identifying risk factors to understanding mechanisms.

Limitations:

• Synthetic Nature: Results are based on pre-specified effect sizes; they validate the framework's operability rather than providing new empirical evidence on real-world GBV determinants.
• Cross-Sectional Design: The lack of temporal ordering limits definitive causal and mediation inference.
• Simplification: Complex constructs (e.g., patriarchy) were reduced to single scores.
• Mediation Null Results: The lack of significant mediation was likely due to the simulation design (strong direct effects) rather than the framework's inability to detect mediation.
• Clustering: Standard errors were not cluster-robust, potentially inflating precision.

Future Directions:
The authors recommend applying this framework to real-world longitudinal data (e.g., KDHS panel surveys) to establish temporal precedence, incorporating Directed Acyclic Graphs (DAGs) for confounding control, and expanding mediation pathways to include economic and social variables.