Causal models, causal inference, and why this is trending

Why Is Causal Inference a Trend in 2025?

Traditional AI (LLMs) Can Predict, Not Explain:
- Before: AI tools like GPT can say, "The man is likely to cheat based on patterns (correlation)."
- Problem: They don’t tell why cheating happens. Is it because the wife gained weight or because there's a sexier mistress? Predictive models miss the cause.
- Need for Causal Inference: In 2025, users demand smarter AI that not only predicts what will happen but also explains why it happens and answers what-if scenarios.
Agentic AI Requires Decision-Making:
- Trend: Agentic AI needs to make autonomous decisions based on causes, not just patterns. For example:
  - Should the AI recommend couples therapy for the wife, or block access to the mistress?
  - Without causal reasoning, the AI might choose interventions that don’t work.
Why It Wasn’t Used Before:
- Lack of accessible tools: Probabilistic frameworks like Pyro and DoWhy have only recently made causal inference approachable.
- Computational challenges: Running causal inference requires dealing with counterfactuals (what-if worlds), which are more complex than simple predictions.

Causal Inference with Wife and Mistress Example

Scenario: Why do men cheat?

Variables:

X1 (Wife gains weight)
X2 (Mistress is sexier)
C (Confounder: Marriage dissatisfaction)
Y (Cheating)

Step 1: Correlation vs. Causation

Observation: Men whose wives gain weight and have sexier mistresses are more likely to cheat.
Problem: Correlation does not mean causation.
- Is the man cheating because the wife gained weight?
- Or is he already dissatisfied (confounder) and the mistress is an opportunity?

Step 2: Causal Graphs

We can represent this as a causal model (graph):

    X1 (Wife gains weight) → Y (Cheating)
    X2 (Sexy mistress) → Y (Cheating)
    C (Marriage dissatisfaction) → X1, X2, Y

Explanation:
- Dissatisfaction (C) causes the wife to gain weight (X1) and leads to cheating (Y).
- A sexier mistress (X2) increases the likelihood of cheating, but only if dissatisfaction exists.

Step 3: Counterfactuals (What-Ifs)

Causal inference uses counterfactuals to ask:

What if the wife didn’t gain weight (X1)?
- Does cheating still happen because of the mistress?
What if there was no sexy mistress (X2)?
- Would dissatisfaction (C) still lead to cheating?

Step 4: Tools Like Pyro or DoWhy

We can now use tools like Pyro and DoWhy to test these causal relationships and counterfactuals programmatically.

Code Example with DoWhy

import dowhy
from dowhy import CausalModel
import pandas as pd
import numpy as np

# Generate synthetic data
data = pd.DataFrame({
    "wife_weight_gain": np.random.choice([0, 1], size=1000),  # 0 = No, 1 = Yes
    "sexy_mistress": np.random.choice([0, 1], size=1000),  # 0 = No, 1 = Yes
    "dissatisfaction": np.random.choice([0, 1], size=1000),  # 0 = No, 1 = Yes
    "cheating": np.random.choice([0, 1], size=1000),  # 0 = No, 1 = Yes
})

# Define the causal model
model = CausalModel(
    data=data,
    treatment="wife_weight_gain",
    outcome="cheating",
    common_causes=["dissatisfaction", "sexy_mistress"]
)

# Identify causal effect
identified_estimand = model.identify_effect()
print(identified_estimand)

# Estimate the causal effect
estimate = model.estimate_effect(
    identified_estimand,
    method_name="backdoor.propensity_score_matching"
)
print("Causal Estimate:", estimate.value)

# Check counterfactual
refute = model.refute_estimate(identified_estimand, estimate, method_name="placebo")
print(refute)

What Happens Here?

Input:
- wife_weight_gain (X1), sexy_mistress (X2), dissatisfaction (C), and cheating (Y).
Output:
- The causal effect of wife_weight_gain on cheating is calculated, adjusting for dissatisfaction and sexy_mistress.
- Counterfactuals: The tool tests whether cheating would still happen if X1 (wife weight gain) didn’t occur.

Sample Output

Causal Estimate: 0.15
Refutation: No significant change in estimate after placebo test

Interpretation:
- Weight gain has a small causal effect on cheating.
- Most of the cheating is driven by dissatisfaction (C) and the presence of a sexy mistress (X2).

Why Is Causal Inference a Trend in 2025?
Traditional AI (LLMs) Can Predict, Not Explain:
- Before: AI tools like GPT can say, "The man is likely to cheat based on patterns (correlation)."
- Problem: They don’t tell why cheating happens. Is it because the wife gained weight or because there's a sexier mistress? Predictive models miss the cause.
- Need for Causal Inference: In 2025, users demand smarter AI that not only predicts what will happen but also explains why it happens and answers what-if scenarios.
Agentic AI Requires Decision-Making:
- Trend: Agentic AI needs to make autonomous decisions based on causes, not just patterns. For example:
  - Should the AI recommend couples therapy for the wife, or block access to the mistress?
  - Without causal reasoning, the AI might choose interventions that don’t work.
Why It Wasn’t Used Before:
- Lack of accessible tools: Probabilistic frameworks like Pyro and DoWhy have only recently made causal inference approachable.
- Computational challenges: Running causal inference requires dealing with counterfactuals (what-if worlds), which are more complex than simple predictions.

Causal Inference with Wife and Mistress Example

Scenario: Why do men cheat?

Variables:

X1 (Wife gains weight)
X2 (Mistress is sexier)
C (Confounder: Marriage dissatisfaction)
Y (Cheating)

Step 1: Correlation vs. Causation

Observation: Men whose wives gain weight and have sexier mistresses are more likely to cheat.
Problem: Correlation does not mean causation.
- Is the man cheating because the wife gained weight?
- Or is he already dissatisfied (confounder) and the mistress is an opportunity?

Step 2: Causal Graphs

We can represent this as a causal model (graph):

    X1 (Wife gains weight) → Y (Cheating)
    X2 (Sexy mistress) → Y (Cheating)
    C (Marriage dissatisfaction) → X1, X2, Y

Explanation:
- Dissatisfaction (C) causes the wife to gain weight (X1) and leads to cheating (Y).
- A sexier mistress (X2) increases the likelihood of cheating, but only if dissatisfaction exists.

Step 3: Counterfactuals (What-Ifs)

Causal inference uses counterfactuals to ask:

What if the wife didn’t gain weight (X1)?
- Does cheating still happen because of the mistress?
What if there was no sexy mistress (X2)?
- Would dissatisfaction (C) still lead to cheating?

Step 4: Tools Like Pyro or DoWhy

We can now use tools like Pyro and DoWhy to test these causal relationships and counterfactuals programmatically.

Code Example with DoWhy

import dowhy
from dowhy import CausalModel
import pandas as pd
import numpy as np

# Generate synthetic data
data = pd.DataFrame({
    "wife_weight_gain": np.random.choice([0, 1], size=1000),  # 0 = No, 1 = Yes
    "sexy_mistress": np.random.choice([0, 1], size=1000),  # 0 = No, 1 = Yes
    "dissatisfaction": np.random.choice([0, 1], size=1000),  # 0 = No, 1 = Yes
    "cheating": np.random.choice([0, 1], size=1000),  # 0 = No, 1 = Yes
})

# Define the causal model
model = CausalModel(
    data=data,
    treatment="wife_weight_gain",
    outcome="cheating",
    common_causes=["dissatisfaction", "sexy_mistress"]
)

# Identify causal effect
identified_estimand = model.identify_effect()
print(identified_estimand)

# Estimate the causal effect
estimate = model.estimate_effect(
    identified_estimand,
    method_name="backdoor.propensity_score_matching"
)
print("Causal Estimate:", estimate.value)

# Check counterfactual
refute = model.refute_estimate(identified_estimand, estimate, method_name="placebo")
print(refute)

What Happens Here?

Input:
- wife_weight_gain (X1), sexy_mistress (X2), dissatisfaction (C), and cheating (Y).
Output:
- The causal effect of wife_weight_gain on cheating is calculated, adjusting for dissatisfaction and sexy_mistress.
- Counterfactuals: The tool tests whether cheating would still happen if X1 (wife weight gain) didn’t occur.

Sample Output

Causal Estimate: 0.15
Refutation: No significant change in estimate after placebo test

Interpretation:
- Weight gain has a small causal effect on cheating.
- Most of the cheating is driven by dissatisfaction (C) and the presence of a sexy mistress (X2).

Why This Matters for Agentic AI in 2025

Better Decision-Making:
Agentic AI can now:
- Explain actions: "The man cheats because of dissatisfaction, not the wife’s weight."
- Recommend interventions: Focus on fixing dissatisfaction, not superficial factors like weight.
Transparent AI:
Users trust AI more when it provides causal reasoning instead of black-box predictions.
What’s New?
- Tools like Pyro and DoWhy bring causality into AI, making it possible to simulate real-world "what-if" scenarios.

Causal models, causal inference, and why this is trending (Why men cheat analogy)

Why Is Causal Inference a Trend in 2025?

Causal Inference with Wife and Mistress Example

Scenario: Why do men cheat?

Step 1: Correlation vs. Causation

Step 2: Causal Graphs

Step 3: Counterfactuals (What-Ifs)

Step 4: Tools Like Pyro or DoWhy

Code Example with DoWhy

What Happens Here?

Sample Output

Why Is Causal Inference a Trend in 2025?

Causal Inference with Wife and Mistress Example

Scenario: Why do men cheat?

Step 1: Correlation vs. Causation

Step 2: Causal Graphs

Step 3: Counterfactuals (What-Ifs)

Step 4: Tools Like Pyro or DoWhy

Code Example with DoWhy

What Happens Here?

Sample Output

Why This Matters for Agentic AI in 2025

Subscribe to my newsletter

Anix Lynch

Anix Lynch