LASSO Cox Regression for Variable Selection in Survival Analysis

Note: The lassocox() function is designed for use within jamovi’s GUI. The code examples below show the R syntax for reference. To run interactively, use devtools::load_all() and call the R6 class directly: lassocoxClass$new(options = lassocoxOptions$new(...), data = mydata).

LASSO Cox Regression

Overview

The Lasso-Cox Regression module (lassocox) performs automatic variable selection in survival analysis using L1-penalized Cox proportional hazards regression. It identifies the most important predictors while preventing overfitting — ideal for clinicopathological research with many candidate variables.

Key features:

• Data suitability assessment (traffic-light system)
• Cross-validated lambda selection (lambda.min vs lambda.1se)
• Variable selection with coefficient shrinkage
• Risk score calculation and patient stratification
• Comprehensive visualizations (CV plot, coefficient plot, survival curves)
• Clinical interpretation guidance

---

What is LASSO Cox Regression?

LASSO (Least Absolute Shrinkage and Selection Operator) Cox regression combines the Cox proportional hazards model with L1 regularization for automatic variable selection.

Standard Cox model:

$$h(t) = h_0(t) \exp(\beta_1 x_1 + \beta_2 x_2 + \cdots + \beta_p x_p)$$

LASSO Cox adds a penalty:

$$\text{Maximize:} \quad \ell(\beta) - \lambda \sum_{j=1}^{p} |\beta_j|$$

where $\ell(\beta)$ is the partial log-likelihood and $\lambda$ controls penalty strength.

The effect: small coefficients shrink toward zero, irrelevant variables become exactly zero, and only important predictors remain in the model.

LASSO vs Ridge vs Elastic Net

Method	Penalty	Variable Selection	Best For
Ridge	$\lambda \sum \beta_j^2$ (L2)	No — shrinks but keeps all	Correlated predictors, all must stay
LASSO	$\lambda \sum |\beta_j|$ (L1)	Yes — sets some to zero	Sparse models, feature selection
Elastic Net	$\alpha \lambda \sum |\beta_j| + (1-\alpha) \lambda \sum \beta_j^2$	Yes — grouped selection	Correlated groups should stay/go together

In ClinicoPath: lassocox uses alpha=1 (pure LASSO). For Elastic Net or Ridge, use the Penalized Cox Regression module.

---

When to Use LASSO Cox Regression

Scenario	Recommendation
Many predictors relative to sample (p >= n/3)	LASSO strongly indicated
Radiomics / genomics feature selection	LASSO ideal
Building a prognostic risk score	LASSO recommended
Few predictors (p < 5) with ample events	Standard Cox may suffice
Correlated predictors that should all be retained	Consider Elastic Net
Inference / p-values are the primary goal	Use standard Cox regression

---

Datasets Used in This Guide

Dataset	N	Events	Predictors	Description
Synthetic “standard”	200	80	3 numeric	Clean scenario (all green suitability)
Synthetic “small”	60	25	5 numeric	Yellow suitability scenario
Synthetic “collinear”	100	45	4 (2 correlated)	Multicollinearity detection
Synthetic “categorical”	100	40	3 factors	All categorical predictors
histopathology	250	~100	8 mixed	Real clinical pathology data
lassocox_genomic	80	–	50 gene features	High-dimensional genomic scenario
lassocox_multicollinear	180	–	12 correlated	Multicollinearity stress test
lassocox_breast_cancer	250	–	20 clinical	Breast cancer clinicopathological

---

1. Basic Analysis (Green Suitability)

A standard scenario with adequate sample size, events, and predictors.

# Create synthetic data: 200 patients, 80 events, 3 numeric predictors
set.seed(42)
n <- 200
data_green <- data.frame(
  time = abs(rnorm(n, 20, 10)) + 0.1,
  status = factor(c(rep("event", 80), rep("censored", 120)),
                  levels = c("censored", "event")),
  var1 = rnorm(n),
  var2 = rnorm(n),
  var3 = rnorm(n)
)

lassocox(
  data = data_green,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("var1", "var2", "var3"),
  suitabilityCheck = FALSE,
  lambda = "lambda.1se",
  nfolds = 5,
  standardize = TRUE,
  cv_plot = TRUE,
  coef_plot = TRUE,
  survival_plot = TRUE
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

The suitability report should show all green indicators:

• EPV: 80/3 = 26.7 (excellent)
• Sample size: n=200 (adequate)
• Event rate: 40% (balanced)
• Multicollinearity: uncorrelated predictors

---

2. Small Sample (Yellow Suitability)

When sample size is limited, the suitability assessment flags potential issues.

set.seed(42)
data_small <- data.frame(
  time = abs(rnorm(60, 15, 8)) + 0.1,
  status = factor(c(rep("event", 25), rep("censored", 35)),
                  levels = c("censored", "event")),
  var1 = rnorm(60), var2 = rnorm(60), var3 = rnorm(60),
  var4 = rnorm(60), var5 = rnorm(60)
)

lassocox(
  data = data_small,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("var1", "var2", "var3", "var4", "var5"),
  suitabilityCheck = FALSE,
  lambda = "lambda.1se",
  nfolds = 5,
  standardize = TRUE,
  cv_plot = TRUE,
  coef_plot = TRUE,
  survival_plot = TRUE
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

Expected suitability flags:

• EPV: 25/5 = 5.0 (yellow — adequate for LASSO but interpret with caution)
• Sample size: n=60 (yellow — small sample)
• Regularization: yellow — standard Cox may be an alternative

---

3. Multicollinearity Detection

When predictors are highly correlated, the suitability check identifies the pairs.

set.seed(42)
base_var <- rnorm(100)
data_collinear <- data.frame(
  time = abs(rnorm(100, 20, 10)) + 0.1,
  status = factor(c(rep("event", 45), rep("censored", 55)),
                  levels = c("censored", "event")),
  var1 = base_var,
  var2 = base_var + rnorm(100, 0, 0.1),  # r > 0.95 with var1
  var3 = rnorm(100),
  var4 = rnorm(100)
)

lassocox(
  data = data_collinear,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("var1", "var2", "var3", "var4"),
  suitabilityCheck = FALSE,
  lambda = "lambda.1se",
  nfolds = 5,
  standardize = TRUE,
  cv_plot = TRUE,
  coef_plot = TRUE,
  survival_plot = TRUE
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

The suitability report should flag:

• Multicollinearity: yellow/red with max |r| > 0.95
• Top correlated pair: var1 & var2
• Recommendation: Consider Elastic Net (Penalized Cox in ClinicoPath)

---

4. Categorical Predictors

LASSO handles factor variables via dummy coding. The collinearity check correctly excludes within-factor dummy correlations.

set.seed(42)
data_cat <- data.frame(
  time = abs(rnorm(100, 20, 10)) + 0.1,
  status = factor(c(rep("event", 40), rep("censored", 60)),
                  levels = c("censored", "event")),
  grade = factor(sample(c("Low", "Med", "High"), 100, replace = TRUE)),
  stage = factor(sample(c("I", "II", "III"), 100, replace = TRUE)),
  treatment = factor(sample(c("A", "B"), 100, replace = TRUE))
)

lassocox(
  data = data_cat,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("grade", "stage", "treatment"),
  suitabilityCheck = FALSE,
  lambda = "lambda.1se",
  nfolds = 5,
  standardize = TRUE,
  cv_plot = TRUE,
  coef_plot = TRUE,
  survival_plot = TRUE
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

---

5. Lambda Selection: lambda.min vs lambda.1se

The choice of lambda controls the trade-off between model complexity and parsimony.

# lambda.min: minimum CV error — more variables, better in-sample fit
lassocox(
  data = data_green,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("var1", "var2", "var3"),
  suitabilityCheck = FALSE,
  lambda = "lambda.min",
  nfolds = 5,
  standardize = TRUE,
  cv_plot = TRUE,
  coef_plot = TRUE,
  survival_plot = TRUE
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

# lambda.1se: most regularized within 1 SE — fewer variables, better generalization
lassocox(
  data = data_green,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("var1", "var2", "var3"),
  suitabilityCheck = FALSE,
  lambda = "lambda.1se",
  nfolds = 5,
  standardize = TRUE,
  cv_plot = TRUE,
  coef_plot = TRUE,
  survival_plot = TRUE
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

Interpretation:

• lambda.min: Selects more variables. Better in-sample performance but higher overfitting risk.
• lambda.1se (default): More parsimonious. Recommended for clinical models that need to generalize.

---

6. Suitability Check Toggle

The suitability assessment is advisory — it never blocks the analysis.

# Disable suitability assessment
lassocox(
  data = data_green,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("var1", "var2", "var3"),
  suitabilityCheck = FALSE,  # No suitability report shown
  lambda = "lambda.1se",
  nfolds = 5,
  standardize = TRUE,
  cv_plot = TRUE,
  coef_plot = TRUE,
  survival_plot = TRUE
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

---

7. Explanatory Output Options

Method explanations

lassocox(
  data = data_green,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("var1", "var2", "var3"),
  suitabilityCheck = FALSE,
  lambda = "lambda.1se",
  nfolds = 5,
  standardize = TRUE,
  cv_plot = TRUE,
  coef_plot = TRUE,
  survival_plot = TRUE,
  showSummary = TRUE,              # Natural-language results paragraph for manuscripts
  showExplanations = TRUE,         # LASSO methodology + plot explanations
  showMethodologyNotes = TRUE,     # Technical details
  includeClinicalGuidance = TRUE   # Clinical interpretation guide
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

Variable importance and model comparison

lassocox(
  data = data_green,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("var1", "var2", "var3"),
  suitabilityCheck = FALSE,
  lambda = "lambda.min",  # More likely to select variables
  nfolds = 5,
  standardize = TRUE,
  cv_plot = FALSE,
  coef_plot = FALSE,
  survival_plot = FALSE,
  showVariableImportance = TRUE,   # Importance scores + stability ranks
  showModelComparison = TRUE       # LASSO vs Standard Cox comparison
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

---

8. Risk Score Output

The computed risk score can be saved as a new variable in the dataset.

result <- lassocox(
  data = data_green,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("var1", "var2", "var3"),
  suitabilityCheck = FALSE,
  lambda = "lambda.min",
  nfolds = 5,
  standardize = TRUE,
  cv_plot = FALSE,
  coef_plot = FALSE,
  survival_plot = FALSE
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

The risk score is computed as:

$$\text{Risk Score} = \beta_1 X_1 + \beta_2 X_2 + \cdots + \beta_k X_k$$

where only LASSO-selected variables have non-zero coefficients.

---

Interpreting Results

Model Summary Table

Statistic	Meaning
Total Variables	Number of candidate predictors (after dummy coding)
Selected Variables	Variables with non-zero coefficients
Selection Proportion	Percentage of variables retained
Optimal Lambda	The regularization strength used
Sample Size	Complete cases used
Number of Events	Events observed
Censoring Rate	Percentage of censored observations

Selected Variables Table

Column	Meaning
Variable	Predictor name
Coefficient	Log hazard ratio (positive = risk, negative = protective)
Hazard Ratio	exp(coefficient) — multiplicative effect on hazard
95% CI (Lower/Upper)	Confidence interval from post-selection unpenalized Cox refit
p	P-value from post-selection Cox refit (interpret with caution)
Importance	Absolute coefficient value

Performance Table

Metric	Meaning
C-index	Discrimination (0.5 = random, 1.0 = perfect)
Log-rank p-value	Statistical significance of risk group separation
Hazard Ratio	Risk of high vs low group (with 95% CI)

C-index Interpretation

C-index	Discrimination	Clinical Utility
0.50–0.60	Poor	Limited
0.60–0.70	Fair	May inform decisions
0.70–0.80	Good	Useful for stratification
0.80–0.90	Excellent	Strong utility
> 0.90	Outstanding	May indicate overfitting

---

Common Pitfalls and Best Practices

Pitfalls to Avoid

1. Using LASSO with too few events: EPV < 2 produces unreliable results. Even with regularization, LASSO needs sufficient events — a common problem in small radiomics cohorts (e.g., 80 patients with 300 features).

2. Not standardizing variables: LASSO is sensitive to scale. A variable measured in 0–20,000 (e.g., tumor area) will dominate one measured in 0–100 (e.g., age) unless standardized. Use standardize = TRUE (the default).

3. Interpreting coefficients as p-values: LASSO performs selection, not testing. Incorrect: “LASSO identified Ki67 as statistically significant.” Correct: “LASSO selected Ki67 as an important predictor.”

4. Feature selection on full data (data leakage): A very common mistake. If you perform LASSO feature selection on all patients, then build and test the model on the same data, information leaks from the test set into the selection step. Feature selection must be done within the training set only.

5. No external validation: Internal cross-validation produces optimistically biased estimates. Journals increasingly require external validation cohorts. The validation hierarchy:

   • Internal cross-validation (minimum)
   • Internal train/test split
   • External independent dataset (preferred)

6. Ignoring feature stability: Radiomic or WSI features may change with scanner, staining protocol, segmentation method, or resolution. Before LASSO, features should be tested for robustness using intraclass correlation coefficient (ICC) or test–retest analysis.

7. Reporting only selected variables: Good reporting requires: selected features, lambda value used, cross-validation method, coefficient values, and performance metrics. Follow the TRIPOD statement for transparent reporting.

8. Ignoring clinical variables: Some radiomics papers include only image features. Combined models (clinical + pathological + radiomic + molecular) usually perform better.

9. Not assessing model performance properly: Survival models should report C-index and calibration curves. High discrimination (C-index) alone is insufficient — the model must also be well-calibrated. Consider decision curve analysis for clinical utility.

10. No calibration analysis: A model with high AUC/C-index but poor calibration makes systematically wrong predictions. Always check calibration plots and consider the Hosmer–Lemeshow test or calibration slope.

Best Practices

1. Check suitability first: Use the built-in assessment before interpreting results
2. Prefer lambda.1se: More parsimonious and generalizable than lambda.min
3. Report transparently: Lambda, CV method, all selected variables, coefficients
4. Validate externally: Test in independent cohorts when possible
5. Combine with clinical variables: Don’t use only image/genomic features
6. Follow TRIPOD guidelines: For prognostic model reporting
7. Test feature stability: Use ICC before entering features into LASSO
8. Perform selection within training set: Never select features on the full dataset

---

Radiomics and Digital Pathology Pipeline

LASSO Cox regression is central to many radiomics and digital pathology prognostic studies. A robust analysis pipeline follows these steps:

WSI / Imaging Data
       ↓
Feature Extraction (e.g., texture, morphology, deep features)
       ↓
Feature Stability Testing (ICC, test–retest)
       ↓
Feature Reduction (remove unstable/redundant features)
       ↓
LASSO Cox Selection (cross-validated, training set only)
       ↓
Model Construction (risk score from selected features)
       ↓
Internal Validation (bootstrap or cross-validation)
       ↓
External Validation (independent cohort)
       ↓
Calibration + Decision Curve Analysis

Application Areas

Domain	Example Use
WSI radiomics	Texture and morphology features from whole-slide images
Deep feature selection	CNN-extracted features reduced via LASSO for interpretability
Tumor microenvironment signatures	Spatial features (immune cell density, stroma ratio)
Gene expression survival models	Selecting prognostic genes from expression panels

Typical Wording in Papers

When reporting LASSO Cox results, consider these example phrasings:

“Feature selection was performed using LASSO penalized Cox regression with 10-fold cross-validation. The selected features were used to construct a radiomic risk score.”

“Variables were selected using LASSO regression with 10-fold cross-validation, and selected predictors were entered into the multivariable Cox model.”

---

9. Using Real Test Datasets

The package includes several realistic test datasets for LASSO Cox analysis.

Breast cancer clinical data

data(lassocox_breast_cancer)

lassocox(
  data = lassocox_breast_cancer,
  elapsedtime = "survival_months",
  outcome = "death",
  outcomeLevel = "Dead",
  explanatory = c("age", "tumor_size_cm", "grade", "lymph_nodes_positive",
                   "ki67_percent", "er_status", "her2_status", "lvi"),
  suitabilityCheck = FALSE,
  lambda = "lambda.1se",
  nfolds = 10,
  standardize = TRUE,
  cv_plot = TRUE,
  coef_plot = TRUE,
  survival_plot = TRUE,
  showSummary = TRUE
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

The showSummary = TRUE option produces a natural-language results paragraph suitable for copying into manuscripts or reports.

---

10. Reproducibility with Random Seeds

The random_seed option ensures reproducible cross-validation fold assignments. Using the same seed produces identical results every time.

# Fixed seed for reproducible results
lassocox(
  data = data_green,
  elapsedtime = "time",
  outcome = "status",
  outcomeLevel = "event",
  explanatory = c("var1", "var2", "var3"),
  suitabilityCheck = FALSE,
  random_seed = 42,
  lambda = "lambda.1se",
  nfolds = 5,
  standardize = TRUE,
  cv_plot = FALSE,
  coef_plot = FALSE,
  survival_plot = FALSE
)
#> Error in `lassocox()`:
#> ! argument "censorLevel" is missing, with no default

Different seeds may produce different variable selections, especially when predictors have similar importance. For publications, always report the seed used.

---

Related ClinicoPath Functions

Function	Use When
Survival Analysis	Standard Kaplan-Meier with log-rank test (few predictors)
Multivariable Survival	Standard Cox with multiple predictors (no regularization needed)
Penalized Cox Regression	Elastic Net or Ridge (retain correlated predictors)
Adaptive LASSO	Two-stage LASSO with oracle properties
Group LASSO	Select/exclude grouped variables together

---

References

• Tibshirani R. The lasso method for variable selection in the Cox model. Stat Med. 1997;16(4):385-395.
• Simon N, et al. Regularization paths for Cox’s proportional hazards model via coordinate descent. J Stat Softw. 2011;39(5):1-13.
• Collins GS, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD). BMJ. 2015;350:g7594.