paulpas Randomized Experiments
Comprehensive guide to randomized experiments in machine learning and data science workflows.
When to Use This Skill
- Solving real-world causal inference problems
- Building machine learning pipelines with randomized experiments
- Implementing best practices for randomized experiments
- Optimizing model performance using randomized experiments techniques
- Learning industry-standard approaches to randomized experiments
When NOT to Use This Skill
- When using pre-built libraries without understanding underlying concepts
- For toy problems that don't require randomized experiments rigor
- When domain expertise in specific problem requires different approach
- If your problem doesn't require the complexity this skill provides
Purpose and Key Concepts
Randomized Experiments is a critical component of the machine learning workflow. This skill covers:
- Theoretical foundations — Mathematical principles and statistical concepts
- Practical implementation — Working code examples and patterns
- Common pitfalls — Mistakes to avoid and how to recover from them
- Best practices — Industry-standard approaches and optimization techniques
Core Workflow
- Understand the problem — Clearly define what you're solving for
- Select approach — Choose the right technique for your data and constraints
- Implement solution — Write clean, tested code following best practices
- Validate results — Verify your implementation with tests and validation
- Optimize performance — Improve efficiency and accuracy incrementally
Implementation Patterns
Pattern 1: Basic Randomized Experiments
import pandas as pd
import numpy as np
from scipy import stats
from typing import Dict, Optional
def basic_rct_analysis(
data: pd.DataFrame,
treatment_col: str,
outcome_col: str,
block_col: Optional[str] = None
) -> Dict[str, float]:
"""
Perform basic two-sample analysis on randomized experiment data.
Handles optional blocking and returns statistical results.
"""
if outcome_col not in data.columns or treatment_col not in data.columns:
raise ValueError("Missing required columns in data")
treatment_data = data[data[treatment_col] == 1][outcome_col]
control_data = data[data[treatment_col] == 0][outcome_col]
if len(treatment_data) < 2 or len(control_data) < 2:
raise ValueError("Insufficient samples in treatment or control groups")
t_stat, p_value = stats.ttest_ind(treatment_data, control_data, equal_var=False)
effect_size = float(np.mean(treatment_data) - np.mean(control_data))
results: Dict[str, float] = {
"t_statistic": float(t_stat)
"p_value": float(p_value)
"effect_size": effect_size
"significant": float(1.0 if p_value < 0.05 else 0.0)
}
if block_col and block_col in data.columns:
blocked_means: Dict[str, float] = {}
for block in data[block_col].unique():
block_data = data[data[block_col] == block]
t_b, p_b = stats.ttest_ind(
block_data[block_data[treatment_col] == 1][outcome_col]
block_data[block_data[treatment_col] == 0][outcome_col]
equal_var=False
)
blocked_means[f"block_{block}_p"] = float(p_b)
results["blocked_analysis"] = blocked_means
return results
Pattern 2: Production-Ready Randomized Experiments
import logging
import pandas as pd
import numpy as np
from typing import Any, Dict, List
from scipy import stats
from statsmodels.stats.power import TTestIndPower
logger = logging.getLogger(__name__)
class RandomizedExperiment:
"""Production-grade implementation for designing and analyzing randomized experiments."""
def __init__(self, alpha: float = 0.05, power: float = 0.8, effect_size: float = 0.5) -> None:
self.alpha = alpha
self.power = power
self.effect_size = effect_size
self._analysis_results: Dict[str, Any] = {}
def calculate_sample_size(self, n_control: int = 100) -> int:
"""Calculate required treatment group size for desired power."""
analysis = TTestIndPower()
n_treatment = analysis.solve_power(
effect_size=self.effect_size,
power=self.power,
alpha=self.alpha,
nobs1=n_control,
ratio=1.0
)
logger.info(f"Calculated required treatment size: {int(np.ceil(n_treatment))}")
return int(np.ceil(n_treatment))
def assign_treatment(self, data: pd.DataFrame, block_col: str = None) -> pd.DataFrame:
"""Randomly assign treatment/control with optional blocking."""
if block_col is None:
data = data.copy()
data['treatment'] = np.random.binomial(1, 0.5, size=len(data))
else:
data = data.copy()
data['treatment'] = 0
for block in data[block_col].unique():
mask = data[block_col] == block
block_size = mask.sum()
data.loc[mask, 'treatment'] = np.random.binomial(1, 0.5, size=block_size)
return data
def analyze(self, data: pd.DataFrame, outcome_col: str) -> Dict[str, Any]:
"""Run statistical analysis on assigned experiment data."""
if 'treatment' not in data.columns:
raise ValueError("Treatment assignment missing. Run assign_treatment first.")
treatment = data[data['treatment'] == 1][outcome_col]
control = data[data['treatment'] == 0][outcome_col]
t_stat, p_value = stats.ttest_ind(treatment, control, equal_var=False)
ci = stats.t.interval(0.95, len(treatment) + len(control) - 2,
loc=np.mean(treatment) - np.mean(control)
scale=np.sqrt(np.var(treatment)/len(treatment) + np.var(control)/len(control)))
self._analysis_results = {
"t_statistic": float(t_stat)
"p_value": float(p_value)
"confidence_interval": [float(ci[0]), float(ci[1])]
"significant": bool(p_value < self.alpha)
"n_treatment": len(treatment)
"n_control": len(control)
}
return self._analysis_results
Pattern 3: Anti-Patterns and Best Practices
# BAD: Magic numbers, no error handling, violates DRY principle
def bad_experiment(data):
t = data[data['group'] == 1]['value']
c = data[data['group'] == 0]['value']
return np.mean(t) - np.mean(c) # No significance testing, hardcoded assumptions
# GOOD: Validated inputs, statistical rigor, follows DRY and SOLID principles
def good_experiment(data: pd.DataFrame, group_col: str, value_col: str) -> Dict[str, float]:
"""Analyze experiment with proper validation and statistical testing."""
if group_col not in data.columns or value_col not in data.columns:
raise ValueError("Invalid column names provided")
groups = data[group_col].unique()
if len(groups) != 2:
raise ValueError("Exactly two groups required for comparison")
t_stat, p_val = stats.ttest_ind(
data[data[group_col] == groups[0]][value_col]
data[data[group_col] == groups[1]][value_col]
equal_var=False
)
return {
"effect_size": float(np.mean(data[data[group_col] == groups[1]][value_col]) -
np.mean(data[data[group_col] == groups[0]][value_col]))
"p_value": float(p_val)
"significant": bool(p_val < 0.05)
}
Best Practices
- ✅ Always validate your implementation on test data
- ✅ Document your assumptions and methodology
- ✅ Use version control for reproducibility
- ✅ Monitor performance metrics in production
- ✅ Periodically review and update your approach
- ✅ Test with edge cases and outliers
- ✅ Log all significant operations for debugging
Common Pitfalls
| Pitfall | Problem | Solution | |
Constraints
MUST DO
- Validate all data preprocessing steps are fit-only on training data, never on validation or test sets
- Implement reproducible pipelines with fixed random seeds and deterministic operations where possible
- Report model performance with confidence intervals via bootstrapping or cross-validation across multiple runs
- Log all experiments with parameters, metrics, and artifacts using MLflow or equivalent tracking system
MUST NOT DO
- Do not evaluate a model on the same data used for training — always hold out a proper test set
- Avoid overfitting to the validation set by limiting hyperparameter search iterations
- Never use features that can only be computed at inference time (look-ahead bias)
- Do not report single-run accuracy without statistical significance testing or error bars
Live References
Authoritative documentation links for this skill's domain. The model follows markdown links at load time to resolve external references and inline content.