tabscore: Build, simplify, validate and present a clinical/statistical...

In R4VN: Health Data Analysis and Publication-Ready Reporting

Build, simplify, validate and present a clinical/statistical scorecard

Description

tabscore() converts a multivariable prediction model into a complete, publication-ready scorecard. The function is designed for the full workflow, not merely for rounding regression coefficients. In one call it can:

• develop or accept a final prediction model;

• optionally select predictors from a candidate set;

• create score-ready categories for continuous predictors;

• derive an exact/model score and a simpler clinical integer score;

• produce a complete Predictor–Category–Point table and theoretical total score range;

• map every possible total score to predicted risk (or expected count/rate);

• select clinically/statistically useful score cutoffs;

• report diagnostic/prognostic properties at the selected cutoff;

• compare the original model, the model score and the clinical score;

• perform bootstrap internal validation of the entire development pipeline;

• optionally evaluate an external validation data set; and

• retain plot-ready data and prediction methods for deployment in R4VN Studio.

Version 1 supports logistic regression, Cox proportional hazards regression, and Poisson regression. Logistic and Cox models receive the most complete discrimination/cutoff workflow. A Poisson model may be used for count/rate scores; when its outcome is binary, ROC/cutoff summaries are also available.

Usage

tabscore(
  outcome,
  predictors = NULL,
  data = NULL,
  family = c("auto", "logistic", "cox", "poisson"),
  event = NULL,
  time = NULL,
  times = NULL,
  cutoff_time = NULL,
  select = c("none", "full", "backward", "forward", "purposeful", "lasso"),
  force = NULL,
  exclude = NULL,
  entry = 0.25,
  stay = 0.1,
  confound = 0.15,
  cuts = NULL,
  continuous = c("easy", "auto", "quantile", "keep"),
  bins = 4L,
  points = "auto",
  pdo = 20,
  maxscore = "auto",
  simplify = TRUE,
  tolerance = 0.01,
  riskonly = TRUE,
  scoreref = "lowest",
  cutoff = c("iu", "youden", "risk", "prevalence", "sens", "spec", "cost", "manual",
    "refprob", "none"),
  riskcut = NULL,
  cutoff_value = NULL,
  sens = NULL,
  spec = NULL,
  cost_fp = 1,
  cost_fn = 1,
  refprob = NULL,
  refcut = NULL,
  validate = c("bootstrap", "none"),
  bootstrap = 500L,
  validation = NULL,
  risktable = TRUE,
  compare = TRUE,
  calibration = TRUE,
  decision = TRUE,
  plot = TRUE,
  show = TRUE,
  console = FALSE,
  seed = NULL,
  ai = FALSE
)

Arguments

outcome	Outcome variable. It can be an unquoted variable name, a character variable name, an already fitted glm/coxph model, or an R4VN regression result containing raw$model (for example from logistic() or poisson()). For Cox analysis this is the event/status indicator; supply follow-up time in ⁠time=⁠. Binary outcomes may be numeric/logical/factor/character.
predictors	Candidate predictors. Accepts a character vector, unquoted variables inside c(), or an R4VN vars() expression. Predictors are treated as model variables rather than individual dummy coefficients. When vars() is used, its R4VN declarations are honored: no prefix and b2., b3., ... declare categorical predictors and select the corresponding factor reference; c., q. and f. declare numeric predictors as continuous for model fitting. Numeric variables carrying complete named value labels are also treated as categorical scorecard variables. Omit this argument when outcome is an already fitted model.
data	Data frame. When omitted, tabscore() attempts to use the active R4VN data set created by usedf(). When outcome is an already fitted model, the stored model frame is the authoritative development sample; data is not used to refit or silently change that model.
family	Model family: auto, logistic, cox, or poisson. With auto, the presence of ⁠time=⁠ selects Cox; otherwise a binary outcome selects logistic and a non-negative integer count selects Poisson.
event	Event level for a binary outcome/status. For 0/1 outcomes the default is 1. For a two-level factor the default is its second factor level; for character outcomes it is the second observed non-missing value. Set this explicitly whenever the event direction matters, e.g. event="Co".
time	Cox follow-up-time variable, supplied as an unquoted or character variable name. This is the observed time, not the prediction horizon.
times	Prediction horizons for Cox score-to-risk tables, e.g. times=c(1,3,5). When omitted, useful event-time quantiles are selected.
cutoff_time	Time horizon at which a Cox cutoff is evaluated. Defaults to the largest value in times, which is often the main clinical horizon.
select	Predictor-selection strategy. none and full keep all supplied predictors. backward/forward use AIC stepwise selection. purposeful uses univariable screening, multivariable removal, confounding assessment and re-entry. lasso uses cross-validated glmnet at lambda.1se (falling back to lambda.min if necessary), then refits a standard model.
force	Predictors that must remain in model-selection procedures. Accepts character names, c(...), or vars(...), using the same naming conventions as predictors.
exclude	Candidate predictors to remove before model development. Accepts character names, c(...), or vars(...).
entry	Univariable screening p-value for purposeful selection; default 0.25.
stay	Multivariable retention/re-entry p-value for purposeful selection; default 0.10.
confound	Relative coefficient-change threshold used to retain a variable as a confounder during purposeful selection; default 0.15 (15 percent).
cuts	Optional named list of user-defined cut points for continuous predictors, e.g. list(age=c(40,50,60), bmi=c(23,25,30)). These are used in the score-ready model and therefore in the scorecard. As a convenience, a single character value "easy", "auto", "quantile", or "keep" is accepted as an alias for ⁠continuous=⁠. The original final model remains available for comparison.
continuous	Handling of continuous predictors when the clinical scorecard is built. easy (default) uses quantile-informed cut points snapped to easy-to-use numbers; auto is an alias; quantile uses unsnapped empirical quantiles; keep leaves continuous predictors continuous in the score-ready model. A complete bedside Predictor–Category–Point table nevertheless needs explicit categories, so use ⁠cuts=⁠ when keep is requested.
bins	Desired number of categories when automatic continuous-variable categorization is used. Default 4.
points	Point-construction method. auto (or clinical/integer) searches for a small integer score that preserves score-ready model discrimination within tolerance. model or pdo uses model/PDO scaling. A named list or a data frame with columns Predictor, Category and Point may be supplied for a completely user-defined integer score. With the default riskonly=TRUE, manual points are shifted within each predictor so its minimum becomes zero; therefore even a supplied negative protective point is converted to an equivalent add-only risk score. With riskonly=FALSE, negative manual points are retained. tabscore() still calibrates, tests, compares and validates the manual score. When manual points refer to categories of a continuous predictor, also supply explicit ⁠cuts=⁠ so the category definitions remain fixed during validation. Named point vectors are matched to displayed category labels after normalizing common typographic equivalents, so ASCII input such as 40-49 and ⁠>=60⁠ also matches publication labels such as 40-49 and ⁠>=60⁠. Unnamed vectors are matched in displayed category order.
pdo	Points to double the effect on the model's log scale. For logistic regression this is Points to Double the Odds: factor = pdo/log(2). For Cox the same number of points doubles hazard; for Poisson it doubles modeled rate. Default 20. The PDO/model score is retained separately from the clinical score.
maxscore	Maximum preferred theoretical clinical score. auto searches compact totals (approximately 5–30 points). A numeric value constrains the theoretical range. This range is based on all possible scorecard categories, not merely the observed sample minimum/maximum.
simplify	Logical. With points="auto", TRUE (default) searches for a compact clinical integer score; FALSE uses rounded PDO/model points instead, which is useful when preserving model-scale resolution is more important than bedside compactness.
tolerance	Maximum tolerated decrease in the primary discrimination metric when simplifying to the clinical score. For binary models this is AUC; for Cox it is C-index. The same threshold is also used to flag excessive loss caused by automatic predictor categorization. Default 0.01. If no compact score meets the tolerance, the best-performing candidate is retained and a warning is stored.
riskonly	Logical. Default TRUE. Build the bedside score as a pure add-only risk score: each predictor is re-referenced to its lowest modeled risk category, all scoring effect ratios are at least 1, and all clinical points are non-negative. For example, if Male is the regression reference and Female has OR=0.50, the score representation becomes Female=0 points (score reference) and Male has risk-oriented OR=2.00 with positive points. This reparameterization does not change the fitted model, subject ranking, or predicted risks. Set FALSE only when a signed score with negative protective points is specifically desired.
scoreref	Scoring reference. Default "lowest" chooses the lowest-risk category independently within every categorical predictor. "model" uses each fitted model reference and is mainly useful with riskonly=FALSE. A named list/vector can set explicit scoring references, for example list(sex="Female", exercise="Yes"). With riskonly=TRUE, an explicitly requested reference must be one of the predictor's lowest-risk categories; otherwise tabscore() stops because satisfying that reference would require negative risk points.
cutoff	Selected cutoff principle: iu, youden, risk, prevalence, sens, spec, cost, manual, refprob, or none. All available methods are shown; this argument chooses the threshold used for final classification.
riskcut	One or more clinically meaningful probability thresholds. With cutoff="risk", the first value is converted to the nearest integer score. Multiple values create Low/Intermediate/High/etc. groups in the risk table.
cutoff_value	User-specified integer score threshold for cutoff="manual".
sens	Minimum desired sensitivity for cutoff="sens", e.g. 0.90.
spec	Minimum desired specificity for cutoff="spec", e.g. 0.90.
cost_fp	Relative cost assigned to a false positive. Default 1.
cost_fn	Relative cost assigned to a false negative. Example: cost_fn=5 makes a false negative five times as costly as a false positive when cost_fp=1.
refprob	Optional reference predicted probability for binary-outcome workflows: a numeric vector or a probability-variable name in data. tabscore() reports correlation, MAE, RMSE and mean difference between reference and score-derived probabilities. It is not used as a surrogate outcome and is not called a gold standard unless it truly is one.
refcut	Probability threshold applied to refprob. With cutoff="refprob", the score threshold that best reproduces this reference classification is selected, then evaluated against the actual outcome.
validate	Internal validation: bootstrap (default) or none. Bootstrap validation reruns the whole development process inside each resample, including selection, automatic cuts and point simplification.
bootstrap	Number of bootstrap resamples. Default 500. Use 1000 or more for a final analysis when feasible; small values are for code testing only.
validation	Optional external validation data frame. The frozen final scorecard is applied without re-estimating cuts, points, calibration or the chosen score cutoff. Binary validation reports AUC, Brier, calibration and fixed-cutoff performance; Cox validation reports C-index, time-specific IPCW Brier and fixed-cutoff IPCW performance; count Poisson reports prediction error.
risktable	Logical; create score-to-risk/score-to-expected-value table.
compare	Logical; compare original model, model score and clinical score.
calibration	Logical; prepare calibration data for plots.
decision	Logical; prepare decision-curve net-benefit data for binary outcomes.
plot	Logical. Default TRUE. Prepare all graphics supported by the chosen model family, embed every available graph directly in the HTML Viewer, and draw them into the interactive R/RStudio Plot history. Set FALSE when only tables are wanted. Plotting uses base R and does not add a required package.
show	Logical. If TRUE, write a self-contained publication-oriented HTML result and open it in the RStudio Viewer (or the default browser).
console	Logical. If TRUE, print a concise console summary.
seed	Optional random seed used for LASSO, bootstrap fallback and validation. The default NULL does not set a seed.
ai	Logical or endpoint name. If R4VN aiask() is available, request an optional AI interpretation after the statistical object is complete.

Details

Missing data and development sample

When tabscore() develops a model from candidate predictors, it uses one complete-case development sample across the outcome/status, Cox time (when applicable), and all candidate predictors supplied before model selection. This keeps candidate models comparable but can reduce sample size when many predictors have missing values. Perform the intended imputation or missing-data strategy before tabscore() when complete-case analysis is inappropriate.

R4VN vars() declarations and reference categories

tabscore() understands the R4VN variable culture rather than merely stripping prefixes. For example, vars(c.age, b2.sex, smoking) fits age continuously, treats sex as categorical with its second factor level as model reference, and treats smoking as categorical with its first factor level as reference. This affects the fitted model table and model-selection calculations. Point assignment itself is then shifted within each predictor so the lowest-risk category receives zero automatic points; consequently the zero-point category does not have to be the regression reference category. The final prediction-model table preserves the original statistical reference and may therefore legitimately show OR/HR/RR below 1. The separate risk_orientation table shows the scoring contrast after re-referencing; with riskonly=TRUE, every displayed scoring ratio is >=1 and the clinical score contains only zero or positive points. With ordinary c(age, sex) syntax, data type is inferred from the columns instead of imposing R4VN categorical declarations.

Original model, model score and clinical score

tabscore() distinguishes three objects. The original model is the selected or fixed model using the original predictor representation. The model score is a monotone point transformation of the score-ready model linear predictor. With pdo=20, a 20-point increase doubles odds (logistic), hazard (Cox), or modeled rate (Poisson). The clinical score uses small integer points and is the score shown in the clean Predictor–Category–Point table.

Intercepts are never artificially divided among predictors; they remain in the risk mapping. Within each predictor, the lowest modeled contribution is shifted to zero before automatic non-negative points are assigned. Thus a zero-point category need not be the regression reference if another category has lower risk.

Protective factors and add-only risk scoring

With riskonly=TRUE, categorical contributions are transformed predictor by predictor as beta_score = beta_category - min(beta_categories). Therefore exp(beta_score) >= 1. For a binary predictor with an original protective contrast OR=0.50, reversing the scoring contrast gives 1/0.50=2.00 for the higher-risk category. For multi-level predictors the same minimum-risk rebasing is used; the procedure is not abs(beta), which can distort category ordering. Cox HR and Poisson RR/IRR are handled identically on their log-effect scales. For a continuous coefficient retained without categories, a negative coefficient is described technically as risk per unit decrease; a complete bedside integer score still requires explicit/automatic categories. The transformation changes only the score origin/reference: the original fitted model and its absolute-risk predictions remain untouched.

Starting from an already fitted model

A base glm/coxph model or an R4VN regression result containing raw$model may be passed as outcome. In that workflow the fitted predictor set is fixed, so select, force and exclude are not used. This version deliberately rejects interactions, transformed/spline terms, no-intercept models, non-unit analysis weights, and non-zero offsets/exposures rather than silently changing the fitted model during score simplification. Represent required transformed predictors as explicit columns and refit before calling tabscore(), or provide a manual point system.

Automatic and manual cut points

Automatic scorecard categorization is a simplification step rather than part of the original continuous model. continuous="easy" uses empirical quantiles and snaps thresholds toward simple numbers. Prefer clinically established ⁠cuts=⁠ when available. Because automatic cuts are data-driven, bootstrap validation repeats the cut-selection step inside each resample.

Theoretical total score

The displayed range is calculated from all category combinations implied by the scorecard, not from the smallest/largest observed subject score. This keeps the bedside score stable in new data.

Score-to-risk conversion

For logistic models, the clinical score is recalibrated by logit(P)=a+b*Score; every possible total score receives a predicted probability and 95 percent CI. This recalibration is important because categorization and integer rounding mean the clinical score is no longer exactly the original LP. Cox scorecards use a one-predictor Cox calibration model to provide risk at each ⁠times=⁠ horizon. Poisson count scores provide expected count/rate and CI.

Cutoff methods

• youden: maximize sensitivity + specificity - 1.

• iu: minimize abs(sensitivity-AUC)+abs(specificity-AUC).

• risk: map a clinical probability in riskcut to an integer score.

• prevalence: use development event prevalence as probability threshold, reproducing a common legacy workflow; it is not automatically clinically best.

• sens: among thresholds meeting sens, maximize specificity.

• spec: among thresholds meeting spec, maximize sensitivity.

• cost: minimize cost_fn*FN + cost_fp*FP.

• manual: use cutoff_value.

• refprob: use refprob and refcut to reproduce a reference probability rule.

Cox support in this version is for standard right-censored proportional-hazards models. Start-stop/time-dependent, multi-state, competing-risk and other complex survival structures are not silently reduced to a simple integer score. Cox ROC/cutoff calculations at cutoff_time use cumulative/dynamic IPCW sensitivity and specificity with the censoring distribution estimated by Kaplan-Meier. IPCW cutoff sensitivity, specificity, PPV, NPV, accuracy and likelihood ratios are point estimates in the apparent cutoff table; unlike the ordinary binary-outcome table, simple binomial confidence intervals are not reported because censoring weights make them inappropriate. Full-pipeline bootstrap validation supplies optimism-corrected cutoff performance and the empirical stability interval of the selected score threshold. If an intervention probability is known, riskcut is generally easier to interpret clinically than a purely statistical cutoff.

Model versus score comparison

A score is not accepted merely because an AUC difference is non-significant. Binary comparisons include AUC with 95 percent CI, Brier score, calibration intercept/slope, and paired AUC difference. Cox comparisons include C-index and time-specific IPCW Brier scores. Decision-curve data compare net benefit across probability thresholds.

Bootstrap internal validation

Each bootstrap resample repeats predictor selection, automatic categorization, point derivation, and the requested cutoff-selection rule. Apparent performance is compared with performance when the bootstrap-derived score and cutoff are applied to the original sample. For binary outcomes the validation table includes optimism-corrected AUC, Brier score, calibration intercept/slope, sensitivity, specificity, PPV, NPV, accuracy, and bootstrap cutoff stability. For Cox models the validation table includes optimism-corrected C-index and, when a cutoff is requested, time-dependent IPCW sensitivity, specificity, PPV, NPV and accuracy at cutoff_time plus bootstrap cutoff stability. Mean optimism is subtracted from the apparent final performance. This is more rigorous than bootstrapping a fixed already-developed score.

External validation and deployment

Supply ⁠validation=⁠ to apply the frozen developed score to a separate data set. Predictor cut points, integer points, risk mapping, and the selected score threshold are not re-optimized in the external data. This avoids turning an external validation into a second development exercise. After development, predict(score_object, newdata, type="all") returns bedside score, predicted risk/value and risk group where applicable. New factor values that were absent from the development scorecard cannot be scored and therefore yield missing score/risk rather than being silently assigned zero points.

Viewer, plots and package requirements

With show=TRUE, the Viewer contains the publication tables and, when plot=TRUE, every graph that is actually available for the fitted family. Logistic/binary scorecards can show score-to-risk, ROC, calibration, decision curve and score-distribution plots. Cox scorecards show score-to-risk across requested horizons, time-dependent ROC at cutoff_time, and score distribution. Poisson count scorecards show score-to-expected-value and the observed-count distribution. The Viewer plots are rendered with base R, using a system sans-serif font and an embedded raster image when possible; this keeps the report self-contained and avoids a ggplot2/htmlwidgets dependency. Use plot(result) to draw all available graphs in the RStudio Plots pane, or plot(result, which="roc"), for example, to draw one.

The core logistic and Poisson workflows use base/recommended R only. survival is needed only for Cox scorecards, glmnet only for select="lasso", and pROC is optional because R4VN has a base-R fallback for AUC calculations and paired AUC comparison. Thus users do not need to install a large collection of packages for ordinary tabscore() analyses.

Modeling cautions

Prediction modeling is not equivalent to retaining only p<0.05 predictors. Use subject-matter knowledge and ⁠force=⁠ for essential variables. Data-driven cuts and cutoffs can overfit and should be validated. Interactions, spline bases, time-varying Cox effects, competing risks and machine-learning distillation are not silently converted into a bedside integer score in this version.

Value

An object of class r4vn_tabscore with fitted models, score rules, development scores/predictions, selected cutoff, theoretical range, publication tables, technical tables, validation results and plot-ready data. Important elements include models, scores, tables, publication_tables, selected_predictors, selected_cutoff, score_range, riskonly, effect_measure, plots, plot_titles and plot_data. tables$risk_orientation explicitly compares the score-ready model-reference effect ratio with the risk-oriented scoring ratio. The publication_tables list contains only ready-to-export non-NULL tables and can be passed directly to tabexport(). predict() can then score new patients without re-estimating the scorecard.

See Also

predict.r4vn_tabscore, plot.r4vn_tabscore

Other R4VN tables: tab(), tabexport(), tabforest(), tablong(), tabmeta(), tabmulti(), tabscale(), tabsurvey(), vars()

Examples

set.seed(2026)
n <- 220
d <- data.frame(
  age = round(rnorm(n, 52, 12)),
  bmi = round(rnorm(n, 24, 4), 1),
  hypertension = factor(rbinom(n, 1, .30), 0:1, c("No", "Yes")),
  smoking = factor(rbinom(n, 1, .25), 0:1, c("No", "Yes")),
  alcohol = factor(rbinom(n, 1, .20), 0:1, c("No", "Yes"))
)
lp <- -4.2 + .04*d$age + .06*(d$bmi - 24) +
  .8*(d$hypertension == "Yes") + .6*(d$smoking == "Yes")
d$event <- rbinom(n, 1, plogis(lp))

# 1. Simplest publication-ready logistic scorecard.
# show=TRUE and plot=TRUE are the user-facing defaults.
s1 <- tabscore(
  event, c(age, bmi, hypertension, smoking), data=d,
  validate="none", show=FALSE, plot=FALSE
)
s1$tables$scorecard
s1$tables$risk
s1$tables$comparison

# 2. R4VN variable declarations: continuous variables and chosen references.
s2 <- tabscore(
  event, vars(c.age, c.bmi, b2.hypertension, b2.smoking), data=d,
  validate="none", show=FALSE, plot=FALSE
)
s2$tables$model
s2$tables$risk_orientation

# 3. Clinically prespecified cut points.
s3 <- tabscore(
  event, c(age, bmi, hypertension, smoking), data=d,
  cuts=list(age=c(40,50,60), bmi=c(23,25,30)),
  validate="none", show=FALSE, plot=FALSE
)

# 4. Apply the frozen scorecard to new patients.
newp <- data.frame(
  age=c(45,68), bmi=c(24,29),
  hypertension=factor(c("No","Yes"), levels=c("No","Yes")),
  smoking=factor(c("Yes","No"), levels=c("No","Yes"))
)
predict(s3, newp, type="all")


# 5. Purposeful selection; force variables that must remain clinically.
s5 <- tabscore(
  event, c(age,bmi,hypertension,smoking,alcohol), data=d,
  select="purposeful", force=c("age","hypertension"),
  validate="none", show=FALSE, plot=FALSE
)

# 6. Backward or forward AIC selection.
s6a <- tabscore(event, c(age,bmi,hypertension,smoking,alcohol), data=d,
                select="backward", validate="none", show=FALSE, plot=FALSE)
s6b <- tabscore(event, c(age,bmi,hypertension,smoking,alcohol), data=d,
                select="forward", validate="none", show=FALSE, plot=FALSE)

# 7. LASSO is optional and only needs glmnet for this selection method.
if (requireNamespace("glmnet", quietly=TRUE)) {
  s7 <- tabscore(event, c(age,bmi,hypertension,smoking,alcohol), data=d,
                 select="lasso", validate="none", show=FALSE, plot=FALSE)
}

# 8. Compact score versus PDO/model-scale points.
s8a <- tabscore(event, c(age,hypertension,smoking), data=d,
                cuts=list(age=c(40,50,60)), maxscore=10,
                validate="none", show=FALSE, plot=FALSE)
s8b <- tabscore(event, c(age,hypertension,smoking), data=d,
                cuts=list(age=c(40,50,60)), points="pdo", pdo=20,
                validate="none", show=FALSE, plot=FALSE)

# 9. Completely manual bedside points; R4VN still calibrates and validates it.
s9 <- tabscore(
  event, c(age,hypertension,smoking), data=d,
  cuts=list(age=c(40,50,60)),
  points=list(
    age=c("<40"=0, "40-49"=1, "50-59"=2, ">=60"=3),
    hypertension=c("No"=0,"Yes"=2),
    smoking=c("No"=0,"Yes"=1)
  ), validate="none", show=FALSE, plot=FALSE
)

# 10. Common cutoff rules.
s10_iu <- tabscore(event, c(age,hypertension,smoking), data=d,
                   cutoff="iu", validate="none", show=FALSE, plot=FALSE)
s10_youden <- tabscore(event, c(age,hypertension,smoking), data=d,
                       cutoff="youden", validate="none", show=FALSE, plot=FALSE)
s10_sens <- tabscore(event, c(age,hypertension,smoking), data=d,
                     cutoff="sens", sens=.90, validate="none", show=FALSE, plot=FALSE)
s10_cost <- tabscore(event, c(age,hypertension,smoking), data=d,
                     cutoff="cost", cost_fn=5, cost_fp=1,
                     validate="none", show=FALSE, plot=FALSE)

# 11. Clinically meaningful probability threshold and multiple risk groups.
s11 <- tabscore(event, c(age,hypertension,smoking), data=d,
                riskcut=c(.05,.10,.20), cutoff="risk",
                validate="none", show=FALSE, plot=FALSE)
s11$tables$risk

# 12. Compare the score with an existing/reference probability.
d$reference_risk <- plogis(-4 + .04*d$age + .7*(d$hypertension == "Yes"))
s12 <- tabscore(event, c(age,hypertension,smoking), data=d,
                refprob=reference_risk, refcut=.10, cutoff="refprob",
                validate="none", show=FALSE, plot=FALSE)
s12$tables$reference_probability

# 13. Convert an already fitted logistic model.
m13 <- glm(event ~ age + hypertension + smoking, data=d, family=binomial())
s13 <- tabscore(m13, validate="none", show=FALSE, plot=FALSE)

# 14. External validation with a frozen scorecard.
dev <- d[1:150, ]
val <- d[151:nrow(d), ]
s14 <- tabscore(event, c(age,hypertension,smoking), data=dev,
                cuts=list(age=c(40,50,60)), validation=val,
                validate="none", show=FALSE, plot=FALSE)
s14$tables$external_validation

# 15. Full-pipeline bootstrap validation. B=20 is only a quick code check;
# use bootstrap=500 or more for the final report.
s15 <- tabscore(event, c(age,hypertension,smoking), data=d,
                cuts=list(age=c(40,50,60)),
                validate="bootstrap", bootstrap=20,
                show=FALSE, plot=FALSE)
s15$tables$validation

# 16. Cox scorecard: survival is the only package required for this family.
if (requireNamespace("survival", quietly=TRUE)) {
  ds <- d
  true_t <- rexp(nrow(ds), rate=exp(-3 + .02*ds$age +
                  .6*(ds$hypertension == "Yes")))
  censor_t <- rexp(nrow(ds), rate=.08)
  ds$status <- as.integer(true_t <= censor_t)
  ds$ftime <- pmin(true_t, censor_t)
  sc <- tabscore(status, c(age,hypertension,smoking), data=ds,
                 family="cox", time=ftime, times=c(1,3,5), cutoff_time=5,
                 cuts=list(age=c(40,50,60)),
                 validate="none", show=FALSE, plot=FALSE)
  sc$tables$risk
  sc$tables$time_brier
}

# 17. Poisson count scorecard.
dp <- d
dp$count <- rpois(nrow(dp), exp(-1 + .015*dp$age + .35*(dp$smoking == "Yes")))
sp <- tabscore(count, c(age,smoking), data=dp, family="poisson",
               cuts=list(age=c(40,50,60)), validate="none",
               show=FALSE, plot=FALSE)
sp$tables$risk

# 18. Every available plot; Viewer includes the same figures when show=TRUE.
sv <- tabscore(event, c(age,hypertension,smoking), data=d,
               cuts=list(age=c(40,50,60)), validate="none",
               show=FALSE, plot=TRUE)
plot(sv, which="risk")
plot(sv, which="roc")
plot(sv, which="calibration")
plot(sv, which="decision")
plot(sv, which="distribution")
plot(sv)  # all available plots in Plot history

# 19. Protective predictors: default risk-only coding versus a signed score.
dr <- d
dr$exercise <- factor(rbinom(nrow(dr), 1, .55), 0:1, c("No", "Yes"))
dr$event2 <- rbinom(nrow(dr), 1,
                    plogis(-2.5 + .04*dr$age - .8*(dr$exercise == "Yes")))
srisk <- tabscore(event2, c(age,exercise), data=dr,
                  cuts=list(age=c(40,50,60)), riskonly=TRUE,
                  validate="none", show=FALSE, plot=FALSE)
ssigned <- tabscore(event2, c(exercise), data=dr,
                    points=list(exercise=c("No"=0,"Yes"=-2)),
                    riskonly=FALSE, scoreref="model",
                    validate="none", show=FALSE, plot=FALSE)

# 20. Additional cutoff strategies.
s20_prev <- tabscore(event, c(age,hypertension,smoking), data=d,
                     cutoff="prevalence", validate="none",
                     show=FALSE, plot=FALSE)
s20_spec <- tabscore(event, c(age,hypertension,smoking), data=d,
                     cutoff="spec", spec=.90, validate="none",
                     show=FALSE, plot=FALSE)
s20_manual <- tabscore(event, c(age,hypertension,smoking), data=d,
                       cutoff="manual", cutoff_value=3, validate="none",
                       show=FALSE, plot=FALSE)
s20_none <- tabscore(event, c(age,hypertension,smoking), data=d,
                     cutoff="none", validate="none",
                     show=FALSE, plot=FALSE)

# 21. Quantile-based automatic categorization.
s21 <- tabscore(event, c(age,bmi,hypertension,smoking), data=d,
                continuous="quantile", bins=4,
                validate="none", show=FALSE, plot=FALSE)

# 22. An R4VN logistic() result can be converted directly as well.
rfit <- logistic(event, vars=vars(c.age, hypertension, smoking),
                 data=d, show=FALSE)
s22 <- tabscore(rfit, validate="none", show=FALSE, plot=FALSE)

# 23. Prediction outputs after the scorecard is frozen.
predict(s3, newp, type="score")
predict(s3, newp, type="risk")
predict(s3, newp, type="group")
predict(s3, newp, type="model")

# 24. Export all publication-ready tables without another tabscore-specific
# dependency. Word/Excel writers are only needed when those formats are chosen.
# tabexport(sv$publication_tables, export=c("html","docx","xlsx"),
#           file="tabscore_report", open=FALSE)

R4VN documentation built on Sept. 30, 2026, 5:13 p.m.