Running Statistical Analysis

Configure and run reproducible statistical analyses with full traceability.

Analysis Types

MedTWIN supports several analysis types, all deterministic—same input always produces same output.

Logistic Regression

Use when: Binary outcome (yes/no, 0/1)

Output: - Odds ratios with 95% CI - P-values for each predictor - Model fit statistics (AUC, Hosmer-Lemeshow) - Classification metrics (sensitivity, specificity)

Example: Predicting 30-day mortality based on age, comorbidities, and lab values.

Cox Proportional Hazards

Use when: Time-to-event with censoring

Output: - Hazard ratios with 95% CI - Kaplan-Meier curves - Log-rank test - Concordance index

Example: Survival analysis comparing treatment groups.

Linear Regression

Use when: Continuous outcome

Output: - Coefficients with 95% CI - R² and adjusted R² - Residual diagnostics - Predicted vs actual plots

Example: Predicting length of stay based on patient characteristics.

Descriptive Statistics

Use when: Summarizing baseline characteristics

Output: - Table 1 (demographics by group) - Mean/SD for continuous variables - N (%) for categorical variables - Standardized differences

Configuration

Step 1: Select Analysis Type

┌─────────────────────────────────────────┐
│  Select Analysis Type                   │
├─────────────────────────────────────────┤
│  ○ Logistic Regression                  │
│  ● Cox Proportional Hazards  ← selected │
│  ○ Linear Regression                    │
│  ○ Descriptive Statistics               │
└─────────────────────────────────────────┘

Step 2: Define Variables

For Cox Regression:

Field	Description	Example
Time Variable	Time to event or censoring	`followup_days`
Event Indicator	1 = event occurred, 0 = censored	`mortality`
Primary Predictor	Main exposure of interest	`treatment_group`
Covariates	Confounders to adjust for	`age`, `sex`, `comorbidity_score`

Step 3: Configure Options

Analysis Configuration:

  # Variable selection
  feature_selection: true
  selection_method: "backward"  # forward, backward, stepwise
  selection_criterion: "AIC"    # AIC, BIC, p-value

  # Validation
  cross_validation: true
  cv_folds: 5

  # Missing data
  missing_strategy: "complete_case"  # complete_case, imputation

  # Output
  confidence_level: 0.95
  decimal_places: 3

Running the Analysis

Start Analysis

Click Run Analysis. You'll see:

┌─────────────────────────────────────────┐
│  Analysis Running...                    │
│                                         │
│  ████████████░░░░░░░░  60%             │
│                                         │
│  ✓ Data validation                      │
│  ✓ Missing data handling                │
│  ◉ Fitting model...                     │
│  ○ Generating results                   │
│  ○ Creating visualizations              │
└─────────────────────────────────────────┘

Typical Runtime

Data Size	Expected Time
< 1,000 rows	5-15 seconds
1,000 - 10,000 rows	15-45 seconds
10,000 - 100,000 rows	1-3 minutes
> 100,000 rows	3-10 minutes

Results

Results Table

┌────────────────────────────────────────────────────────────┐
│  Cox Regression Results                                    │
│  Outcome: mortality | N = 1,247 | Events = 312            │
├──────────────────┬─────────┬───────────────┬──────────────┤
│  Variable        │   HR    │    95% CI     │   P-value    │
├──────────────────┼─────────┼───────────────┼──────────────┤
│  age (per year)  │   1.04  │  1.02 - 1.06  │   <0.001 **  │
│  sex (male)      │   1.23  │  0.98 - 1.54  │    0.072     │
│  hba1c (per %)   │   1.42  │  1.18 - 1.71  │   <0.001 **  │
│  egfr (per 10)   │   0.91  │  0.85 - 0.97  │    0.004 *   │
│  treatment       │   0.67  │  0.52 - 0.86  │    0.002 *   │
└──────────────────┴─────────┴───────────────┴──────────────┘

Model Diagnostics

Concordance Index: 0.72 (95% CI: 0.68 - 0.76)
Proportional Hazards Test: Global p = 0.34 (assumption met)
Sample Size: 1,247 (312 events, 935 censored)

Visualizations

Generated automatically:

Kaplan-Meier Curves - Survival by group
Forest Plot - Hazard ratios with CI
Schoenfeld Residuals - PH assumption check
Predicted Risk Distribution - Model calibration

Traceability

Every result links back to its source.

Click Any Statistic

HR = 1.42 (95% CI: 1.18 - 1.71)
│
├── Variable: hba1c_baseline
├── Analysis: Cox Regression
├── Run ID: run_20240115_143022
├── N: 1,247 (312 events)
│
├── Data Source:
│   └── master_dataset v2
│   └── Mapping: mapping_v2_20240110
│   └── Original file: diabetes_cohort.xlsx
│
├── Computation:
│   └── Python: lifelines.CoxPHFitter
│   └── [View full code]
│
└── Audit:
    └── Created: 2024-01-15 14:30:22
    └── User: dr.smith@hospital.org
    └── [View audit log]

Chain of Custody

Original Data → Mapping → Master Dataset → Analysis → Result
     │              │            │             │         │
     ▼              ▼            ▼             ▼         ▼
  SHA-256       Version 2    Snapshot      Run #3    Locked
  verified      committed    frozen        complete  immutable

Re-running Analysis

When Data Changes

If you update your mapping:

Previous analysis results are preserved
Click Re-run with Current Data
New results are saved as a new run
Compare results side-by-side

Sensitivity Analysis

Run variations to test robustness:

Run	Configuration	Key Result
Run 1	Complete case	HR = 1.42
Run 2	Multiple imputation	HR = 1.38
Run 3	Excluding outliers	HR = 1.45

Exporting Results

Export Options

Format	Contents
Tables (Word/Excel)	Results tables, formatted
Figures (PNG/SVG)	High-res visualizations
Code (Python/R)	Reproducible analysis script
Full Bundle	Data + code + results + audit log

For Journal Submission

Click Export for Publication:

Tables formatted per journal guidelines
Figures at 300 DPI
Methods text describing analysis
Supplementary materials

Best Practices

Before Running

✅ Verify variable mappings are correct
✅ Check for sufficient sample size
✅ Review missing data patterns
✅ Pre-specify your analysis plan

Interpreting Results

✅ Check model diagnostics before interpreting coefficients
✅ Consider clinical significance, not just statistical
✅ Review outliers and influential observations
✅ Run sensitivity analyses

For Reproducibility

✅ Lock your dataset version before analysis
✅ Document any exclusion criteria
✅ Export the analysis code
✅ Keep the audit trail

Troubleshooting

"Model failed to converge"

Causes: - Perfect separation (outcome perfectly predicted) - Multicollinearity (highly correlated predictors) - Too many predictors for sample size

Solutions: 1. Remove highly correlated variables 2. Reduce number of predictors 3. Use regularization (coming soon)

"Not enough events"

Cox regression needs sufficient events per variable. Rule of thumb: ≥10 events per predictor.

Predictors	Minimum Events Needed
3	30
5	50
10	100

"Proportional hazards violated"

The PH assumption may not hold. Options:

Stratify by the violating variable
Use time-varying coefficients
Consider parametric survival models