,
Luisa Barbanti [aut] ,
Sandra Siegfried [aut] ,
Brian Ripley [ctb],
Bill Venables [ctb],
Douglas M. Bates [ctb],
Nadja Klein [ctb]| Title: | Transformation Models |
|---|---|
| Description: | Formula-based user-interfaces to specific transformation models implemented in package 'mlt'. Available models include Cox models, some parametric survival models (Weibull, etc.), models for ordered categorical variables, normal and non-normal (Box-Cox type) linear models, and continuous outcome logistic regression (Lohse et al., 2017, <DOI:10.12688/f1000research.12934.1>). The underlying theory is described in Hothorn et al. (2018) <DOI:10.1111/sjos.12291>. An extension to transformation models for clustered data is provided (Barbanti and Hothorn, 2022, <DOI:10.1093/biostatistics/kxac048>). Multivariate conditional transformation models (Klein et al, 2022, <DOI:10.1111/sjos.12501>) and shift-scale transformation models (Siegfried et al, 2023, <DOI:10.1080/00031305.2023.2203177>) can be fitted as well. |
| Authors: | Torsten Hothorn [aut, cre] ,
Luisa Barbanti [aut] ,
Sandra Siegfried [aut] ,
Brian Ripley [ctb],
Bill Venables [ctb],
Douglas M. Bates [ctb],
Nadja Klein [ctb] |
| Maintainer: | Torsten Hothorn <[email protected]> |
| License: | GPL-2 |
| Version: | 1.1-2 |
| Built: | 2024-10-17 20:24:19 UTC |
| Source: | https://github.com/r-forge/ctm |
Aalen model with fully parameterised hazard function
Aareg(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)Aareg(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)
formula |
an object of class |
data |
an optional data frame, list or environment (or object
coercible by |
subset |
an optional vector specifying a subset of observations to be used in the fitting process. |
weights |
an optional vector of weights to be used in the fitting
process. Should be |
offset |
this can be used to specify an _a priori_ known component to
be included in the linear predictor during fitting. This
should be |
cluster |
optional factor with a cluster ID employed for computing clustered covariances. |
na.action |
a function which indicates what should happen when the data
contain |
... |
additional arguments to |
This function allows simultaneous estimation of the cumulative hazard parameterised by a Bernstein polynomial. The model is typically fitted with time-varying coefficients, all types of random censoring and trunction are allowed.
The responses is bounded (bounds = c(0, Inf)) when specified as a
Surv object. Otherwise, bounds can be specified via
....
An object of class Aareg, with corresponding coef,
vcov, logLik, estfun, summary,
print, plot and predict methods.
Torsten Hothorn, Lisa Moest, Peter Buehlmann (2018), Most Likely Transformations, Scandinavian Journal of Statistics, 45(1), 110–134, doi:10.1111/sjos.12291.
data("GBSG2", package = "TH.data") library("survival") GBSG2$time <- as.numeric(GBSG2$time) GBSG2$y <- with(GBSG2, Surv(time, cens)) ### Cox proportional hazards model m1 <- Coxph(y ~ horTh, data = GBSG2, support = c(1, 1500)) logLik(m1) ### Aalen additive hazards model with time-varying effects m2 <- Aareg(y | horTh ~ 1, data = GBSG2, support = c(1, 1500)) logLik(m2) ### compare the hazard functions nd <- data.frame(horTh = unique(GBSG2$horTh)) col <- 1:2 lty <- 1:2 plot(as.mlt(m1), newdata = nd, type = "hazard", col = col, lty = lty[1], xlab = "time") plot(as.mlt(m2), newdata = nd, type = "hazard", col = col, lty = 2, add = TRUE) legend("topright", col = rep(col, each = 2), lty = rep(1:2), bty = "n", legend = paste(rep(paste("horTh:", levels(nd$horTh)), each = 2), rep(c("Cox", "Aalen"), 2)))data("GBSG2", package = "TH.data") library("survival") GBSG2$time <- as.numeric(GBSG2$time) GBSG2$y <- with(GBSG2, Surv(time, cens)) ### Cox proportional hazards model m1 <- Coxph(y ~ horTh, data = GBSG2, support = c(1, 1500)) logLik(m1) ### Aalen additive hazards model with time-varying effects m2 <- Aareg(y | horTh ~ 1, data = GBSG2, support = c(1, 1500)) logLik(m2) ### compare the hazard functions nd <- data.frame(horTh = unique(GBSG2$horTh)) col <- 1:2 lty <- 1:2 plot(as.mlt(m1), newdata = nd, type = "hazard", col = col, lty = lty[1], xlab = "time") plot(as.mlt(m2), newdata = nd, type = "hazard", col = col, lty = 2, add = TRUE) legend("topright", col = rep(col, each = 2), lty = rep(1:2), bty = "n", legend = paste(rep(paste("horTh:", levels(nd$horTh)), each = 2), rep(c("Cox", "Aalen"), 2)))
Non-normal linear regression inspired by Box-Cox models
BoxCox(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)BoxCox(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)
formula |
an object of class |
data |
an optional data frame, list or environment (or object
coercible by |
subset |
an optional vector specifying a subset of observations to be used in the fitting process. |
weights |
an optional vector of weights to be used in the fitting
process. Should be |
offset |
this can be used to specify an _a priori_ known component to
be included in the linear predictor during fitting. This
should be |
cluster |
optional factor with a cluster ID employed for computing clustered covariances. |
na.action |
a function which indicates what should happen when the data
contain |
... |
additional arguments to |
A normal model for transformed responses, where the transformation is estimated from the data simultaneously with the regression coefficients. This is similar to a Box-Cox transformation, but the technical details differ. Examples can be found in the package vignette.
The model is defined with a negative shift term. Large values of the linear predictor correspond to large values of the conditional expectation response (but this relationship is potentially nonlinear).
An object of class BoxCox, with corresponding coef,
vcov, logLik, estfun, summary,
print, plot and predict methods.
Torsten Hothorn, Lisa Moest, Peter Buehlmann (2018), Most Likely Transformations, Scandinavian Journal of Statistics, 45(1), 110–134, doi:10.1111/sjos.12291.
data("BostonHousing2", package = "mlbench") lm(cmedv ~ crim + zn + indus + chas + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) BoxCox(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2)data("BostonHousing2", package = "mlbench") lm(cmedv ~ crim + zn + indus + chas + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) BoxCox(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2)
A proportional-odds model for continuous variables
Colr(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)Colr(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)
formula |
an object of class |
data |
an optional data frame, list or environment (or object
coercible by |
subset |
an optional vector specifying a subset of observations to be used in the fitting process. |
weights |
an optional vector of weights to be used in the fitting
process. Should be |
offset |
this can be used to specify an _a priori_ known component to
be included in the linear predictor during fitting. This
should be |
cluster |
optional factor with a cluster ID employed for computing clustered covariances. |
na.action |
a function which indicates what should happen when the data
contain |
... |
additional arguments to |
Simultaneous estimation of all possible binary logistic models obtained by dichotomisation of a continuous response. The regression coefficients can be constant allowing for an interpretation as log-odds ratios.
The model is defined with a positive shift term, thus exp(coef()) is
the multiplicative change of the odds ratio (conditional odds of treatment
or for a one unit increase in a numeric variable divided by conditional odds
of reference). Large values of the linear predictor correspond to small
values of the conditional expectation response (but this relationship is
nonlinear).
An object of class Colr, with corresponding coef,
vcov, logLik, estfun, summary,
print, plot and predict methods.
Tina Lohse, Sabine Rohrmann, David Faeh and Torsten Hothorn (2017), Continuous Outcome Logistic Regression for Analyzing Body Mass Index Distributions, F1000Research, 6(1933), doi:10.12688/f1000research.12934.1.
Torsten Hothorn, Lisa Moest, Peter Buehlmann (2018), Most Likely Transformations, Scandinavian Journal of Statistics, 45(1), 110–134, doi:10.1111/sjos.12291.
data("BostonHousing2", package = "mlbench") lm(cmedv ~ crim + zn + indus + chas + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) Colr(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2)data("BostonHousing2", package = "mlbench") lm(cmedv ~ crim + zn + indus + chas + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) Colr(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2)
An alternative approach to competing risk regression via multivariate transformation models
Compris(formula, data, subset, weights, na.action, offset, primary = c("Coxph", "Colr", "BoxCox"), competing = switch(primary, Coxph = "weibull", Colr = "loglogistic", BoxCox = "lognormal"), NPlogLik = FALSE, theta = NULL, optim = mltoptim(auglag = list(maxtry = 5)), args = list(seed = 1, type = c("MC", "ghalton"), M = 1000), scale = FALSE, ...)Compris(formula, data, subset, weights, na.action, offset, primary = c("Coxph", "Colr", "BoxCox"), competing = switch(primary, Coxph = "weibull", Colr = "loglogistic", BoxCox = "lognormal"), NPlogLik = FALSE, theta = NULL, optim = mltoptim(auglag = list(maxtry = 5)), args = list(seed = 1, type = c("MC", "ghalton"), M = 1000), scale = FALSE, ...)
formula |
an object of class |
data |
an optional data frame, list or environment (or object
coercible by |
subset |
an optional vector specifying a subset of observations to be used in the fitting process. |
weights |
an optional vector of case weights to be used in the fitting
process. Should be |
na.action |
a function which indicates what should happen when the data
contain |
offset |
this can be used to specify an _a priori_ known component to
be included in the linear predictor during fitting. This
should be |
primary |
a character defining the marginal model for the primary event of interest, that is, the first status level. |
competing |
a character defining the marginal models for the remaining competing events. |
NPlogLik |
logical, optimise nonparametric likelihood defined in terms of multivariate probabilities. |
theta |
optional starting values. |
optim |
see |
args |
a list of arguments for |
scale |
logical defining if variables in the linear predictor shall be scaled. Scaling is internally used for model estimation, rescaled coefficients are reported in model output. |
... |
addition arguments passed to primary or competing model function. |
This is a highly experimental approach to an alternative competing risk regression framework described by Czado and Van Keilegom (2023) and Deresa and Van Keilegom (2023).
An object of class Mmlt, allowing to derive marginal time-to-event
distributions for the primary event of interest and all competing events.
Claudia Czado and Ingrid Van Keilegom (2023). Dependent Censoring Based on Parametric Copulas. Biometrika, 110(3), 721–738, doi:10.1093/biomet/asac067.
Negera Wakgari Deresa and Ingrid Van Keilegom (2023). Copula Based Cox Proportional Hazards Models for Dependent Censoring. Journal of the American Statistical Association, 119(546), 1044–1054, doi:10.1080/01621459.2022.2161387.
if (require("randomForestSRC")) { library("survival") ## Competing risk data set involving follicular cell lymphoma ## (from doi:10.1002/9780470870709) data("follic", package = "randomForestSRC") ## Therapy: ### Radiotherapy alone (RT) or Chemotherapy + Radiotherapy (CMTRT) follic$ch <- factor(as.character(follic$ch), levels = c("N", "Y"), labels = c("RT", "CMTRT")) ## Clinical state follic$clinstg <- factor(follic$clinstg, levels = 2:1, labels = c("II", "I")) ## Pre-processing as in Deresa & Van Keilegom (2023) follic$time <- round(follic$time, digits = 3) follic$age <- with(follic, (age - mean(age)) / sd(age)) ## standardised follic$hgb <- with(follic, (hgb - mean(hgb)) / sd(hgb)) ## standardised ## Setup `Surv' object for fitting Compris(): ### "status" indicator with levels: ### (1) independent censoring (admin_cens) ### (2) primary event of interest (relapse) ### (3) dependent censoring (death) follic$status <- factor(follic$status, levels = 0:2, labels = c("admin_cens", "relapse", "death")) follic$y <- with(follic, Surv(time = time, event = status)) ## Fit a Gaussian Copula-based Cox Proportional Hazards Model with ## a marginal "Coxph" model for the primary event of interest and ## a Weibull "Survreg" model for dependent censoring ## Use very informative starting values to keep CRAN happy cf <- c( "Event_relapse.Event_relapse.Bs1(Event_relapse)" = -1.89058, "Event_relapse.Event_relapse.Bs2(Event_relapse)" = -1.6566, "Event_relapse.Event_relapse.Bs3(Event_relapse)" = -0.50329, "Event_relapse.Event_relapse.Bs4(Event_relapse)" = -0.50329, "Event_relapse.Event_relapse.Bs5(Event_relapse)" = -0.07402, "Event_relapse.Event_relapse.Bs6(Event_relapse)" = 0.53156, "Event_relapse.Event_relapse.Bs7(Event_relapse)" = 0.67391, "Event_relapse.Event_relapse.chCMTRT" = -0.2861, "Event_relapse.Event_relapse.age" = 0.43178, "Event_relapse.Event_relapse.hgb" = 0.02913, "Event_relapse.Event_relapse.clinstgI" = -0.55601, "Event_death.Event_death.(Intercept)" = -2.20056, "Event_death.Event_death.log(Event_death)" = 0.98102, "Event_death.Event_death.chCMTRT" = 0.25012, "Event_death.Event_death.age" = -0.64826, "Event_death.Event_death.hgb" = -0.02312, "Event_death.Event_death.clinstgI" = 0.57684, "Event_death.Event_relapse.(Intercept)" = -3.48595 ) ### gave up after multiple submissions to CRAN resulting ### in 5.02 > 5 secs m <- Compris(y ~ ch + age + hgb + clinstg, data = follic, log_first = TRUE, ### arguments below speed-up example, don't use! theta = cf, ### informativ starting values optim = mltoptim(), ### no hessian args = list(type = "ghalton", M = 80)) ### only 80 MC integration points ### log-likelihood logLik(m) ## Similar to Table 4 of Deresa & Van Keilegom (2023), ## but using a Gaussian copula instead of a Gumbel copula. ## marginal parameters coef(m, type = "marginal") ## Kendall's tau coef(m, type = "Kendall") }if (require("randomForestSRC")) { library("survival") ## Competing risk data set involving follicular cell lymphoma ## (from doi:10.1002/9780470870709) data("follic", package = "randomForestSRC") ## Therapy: ### Radiotherapy alone (RT) or Chemotherapy + Radiotherapy (CMTRT) follic$ch <- factor(as.character(follic$ch), levels = c("N", "Y"), labels = c("RT", "CMTRT")) ## Clinical state follic$clinstg <- factor(follic$clinstg, levels = 2:1, labels = c("II", "I")) ## Pre-processing as in Deresa & Van Keilegom (2023) follic$time <- round(follic$time, digits = 3) follic$age <- with(follic, (age - mean(age)) / sd(age)) ## standardised follic$hgb <- with(follic, (hgb - mean(hgb)) / sd(hgb)) ## standardised ## Setup `Surv' object for fitting Compris(): ### "status" indicator with levels: ### (1) independent censoring (admin_cens) ### (2) primary event of interest (relapse) ### (3) dependent censoring (death) follic$status <- factor(follic$status, levels = 0:2, labels = c("admin_cens", "relapse", "death")) follic$y <- with(follic, Surv(time = time, event = status)) ## Fit a Gaussian Copula-based Cox Proportional Hazards Model with ## a marginal "Coxph" model for the primary event of interest and ## a Weibull "Survreg" model for dependent censoring ## Use very informative starting values to keep CRAN happy cf <- c( "Event_relapse.Event_relapse.Bs1(Event_relapse)" = -1.89058, "Event_relapse.Event_relapse.Bs2(Event_relapse)" = -1.6566, "Event_relapse.Event_relapse.Bs3(Event_relapse)" = -0.50329, "Event_relapse.Event_relapse.Bs4(Event_relapse)" = -0.50329, "Event_relapse.Event_relapse.Bs5(Event_relapse)" = -0.07402, "Event_relapse.Event_relapse.Bs6(Event_relapse)" = 0.53156, "Event_relapse.Event_relapse.Bs7(Event_relapse)" = 0.67391, "Event_relapse.Event_relapse.chCMTRT" = -0.2861, "Event_relapse.Event_relapse.age" = 0.43178, "Event_relapse.Event_relapse.hgb" = 0.02913, "Event_relapse.Event_relapse.clinstgI" = -0.55601, "Event_death.Event_death.(Intercept)" = -2.20056, "Event_death.Event_death.log(Event_death)" = 0.98102, "Event_death.Event_death.chCMTRT" = 0.25012, "Event_death.Event_death.age" = -0.64826, "Event_death.Event_death.hgb" = -0.02312, "Event_death.Event_death.clinstgI" = 0.57684, "Event_death.Event_relapse.(Intercept)" = -3.48595 ) ### gave up after multiple submissions to CRAN resulting ### in 5.02 > 5 secs m <- Compris(y ~ ch + age + hgb + clinstg, data = follic, log_first = TRUE, ### arguments below speed-up example, don't use! theta = cf, ### informativ starting values optim = mltoptim(), ### no hessian args = list(type = "ghalton", M = 80)) ### only 80 MC integration points ### log-likelihood logLik(m) ## Similar to Table 4 of Deresa & Van Keilegom (2023), ## but using a Gaussian copula instead of a Gumbel copula. ## marginal parameters coef(m, type = "marginal") ## Kendall's tau coef(m, type = "Kendall") }
Cox model with fully parameterised baseline hazard function
Coxph(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)Coxph(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)
formula |
an object of class |
data |
an optional data frame, list or environment (or object
coercible by |
subset |
an optional vector specifying a subset of observations to be used in the fitting process. |
weights |
an optional vector of weights to be used in the fitting
process. Should be |
offset |
this can be used to specify an _a priori_ known component to
be included in the linear predictor during fitting. This
should be |
cluster |
optional factor with a cluster ID employed for computing clustered covariances. |
na.action |
a function which indicates what should happen when the data
contain |
... |
additional arguments to |
The original implementation of Cox models via the partial likelihood,
treating the baseline hazard function as a nuisance parameter, is available
in coxph. This function allows simultaneous
estimation of the log-hazard ratios and the log-cumulative baseline hazard,
the latter parameterised by a Bernstein polynomial. The model can be fitted
under stratification (time-varying coefficients), all types of random
censoring and trunction. An early reference to this parameterisation is
McLain and Ghosh (2013).
The response is bounded (bounds = c(0, Inf)) when specified as a
Surv object. Otherwise, bounds can be specified via
....
Parameters are log-hazard ratios comparing treatment (or a one unit increase in a numeric variable) with a reference.
An object of class Coxph, with corresponding coef,
vcov, logLik, estfun, summary,
print, plot and predict methods.
Alexander C. McLain and Sujit K. Ghosh (2013). Efficient Sieve Maximum Likelihood Estimation of Time-Transformation Models, Journal of Statistical Theory and Practice, 7(2), 285–303, doi:10.1080/15598608.2013.772835.
Torsten Hothorn, Lisa Moest, Peter Buehlmann (2018), Most Likely Transformations, Scandinavian Journal of Statistics, 45(1), 110–134, doi:10.1111/sjos.12291.
data("GBSG2", package = "TH.data") library("survival") (m1 <- coxph(Surv(time, cens) ~ horTh, data = GBSG2)) (m2 <- Coxph(Surv(time, cens) ~ horTh, data = GBSG2)) ### McLain & Ghosh (2013) (m3 <- Coxph(Surv(time, cens) ~ horTh, data = GBSG2, frailty = "Gamma")) ### Wald intervals confint(m1) confint(m2) ### profile likelihood interval confint(profile(m2)) ### score interval confint(score_test(m2)) ### permutation score interval; uses permutation distribution ### see coin::independence_test ## Not run: confint(perm_test(m2))data("GBSG2", package = "TH.data") library("survival") (m1 <- coxph(Surv(time, cens) ~ horTh, data = GBSG2)) (m2 <- Coxph(Surv(time, cens) ~ horTh, data = GBSG2)) ### McLain & Ghosh (2013) (m3 <- Coxph(Surv(time, cens) ~ horTh, data = GBSG2, frailty = "Gamma")) ### Wald intervals confint(m1) confint(m2) ### profile likelihood interval confint(profile(m2)) ### score interval confint(score_test(m2)) ### permutation score interval; uses permutation distribution ### see coin::independence_test ## Not run: confint(perm_test(m2))
Non-normal linear regression for Lehmann-alternatives
Lehmann(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)Lehmann(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)
formula |
an object of class |
data |
an optional data frame, list or environment (or object
coercible by |
subset |
an optional vector specifying a subset of observations to be used in the fitting process. |
weights |
an optional vector of weights to be used in the fitting
process. Should be |
offset |
this can be used to specify an _a priori_ known component to
be included in the linear predictor during fitting. This
should be |
cluster |
optional factor with a cluster ID employed for computing clustered covariances. |
na.action |
a function which indicates what should happen when the data
contain |
... |
additional arguments to |
This transformation model uses the cumulative distribution function for the standard Gumbel maximum extreme value distribution to map the shifted transformation function into probabilities. The exponential of the shift paramater can be interpreted as a Lehmann-alternative or reverse time hazard ratio.
An object of class Lehmann, with corresponding coef,
vcov, logLik, estfun, summary,
print, plot and predict methods.
Erich L. Lehmann (1953), The Power of Rank Tests, The Annals of Mathematical Statistics, 24(1), 23-43.
Torsten Hothorn, Lisa Moest, Peter Buehlmann (2018), Most Likely Transformations, Scandinavian Journal of Statistics, 45(1), 110–134, doi:10.1111/sjos.12291.
data("BostonHousing2", package = "mlbench") lm(cmedv ~ crim + zn + indus + chas + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) Lehmann(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2)data("BostonHousing2", package = "mlbench") lm(cmedv ~ crim + zn + indus + chas + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) Lehmann(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2)
Normal linear model with benefits
Lm(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)Lm(formula, data, subset, weights, offset, cluster, na.action = na.omit, ...)
formula |
an object of class |
data |
an optional data frame, list or environment (or object
coercible by |
subset |
an optional vector specifying a subset of observations to be used in the fitting process. |
weights |
an optional vector of weights to be used in the fitting
process. Should be |
offset |
this can be used to specify an _a priori_ known component to
be included in the linear predictor during fitting. This
should be |
cluster |
optional factor with a cluster ID employed for computing clustered covariances. |
na.action |
a function which indicates what should happen when the data
contain |
... |
additional arguments to |
A normal linear model with simulaneous estimation of regression coefficients and scale parameter(s). This function also allows for stratum-specific intercepts and variances as well as censoring and truncation in the response.
Note that the scale of the parameters is different from what is reported by
lm; the discrepancies are explained in the package
vignette.
The model is defined with a negative shift term. Large values of the linear predictor correspond to large values of the conditional expectation response.
An object of class Lm, with corresponding coef,
vcov, logLik, estfun, summary,
print, plot and predict methods.
Torsten Hothorn, Lisa Moest, Peter Buehlmann (2018), Most Likely Transformations, Scandinavian Journal of Statistics, 45(1), 110–134, doi:10.1111/sjos.12291.
data("BostonHousing2", package = "mlbench") lm(cmedv ~ crim + zn + indus + chas + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) Lm(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2)data("BostonHousing2", package = "mlbench") lm(cmedv ~ crim + zn + indus + chas + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) Lm(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2)
Conditional transformation models for multivariate continuous, discrete, or a mix of continuous and discrete outcomes
Mmlt(..., formula = ~ 1, data, conditional = FALSE, theta = NULL, fixed = NULL, scale = FALSE, optim = mltoptim(auglag = list(maxtry = 5)), args = list(seed = 1, type = c("MC", "ghalton"), M = 1000), fit = c("jointML", "pseudo", "ACS", "sequential", "none"), ACSiter = 2)Mmlt(..., formula = ~ 1, data, conditional = FALSE, theta = NULL, fixed = NULL, scale = FALSE, optim = mltoptim(auglag = list(maxtry = 5)), args = list(seed = 1, type = c("MC", "ghalton"), M = 1000), fit = c("jointML", "pseudo", "ACS", "sequential", "none"), ACSiter = 2)
... |
marginal transformation models, one for each response, for
|
formula |
a model formula describing a model for the dependency
structure via the lambda parameters. The default is set to |
data |
a data.frame. |
conditional |
logical; parameters are defined conditionally (only
possible when all models are probit models). This is the default as
described by Klein et al. (2022). If |
theta |
an optional vector of starting values. |
fixed |
an optional named numeric vector of predefined parameter values
or a logical (for |
scale |
a logical indicating if (internal) scaling shall be applied to the model coefficients. |
optim |
a list of optimisers as returned by |
args |
a list of arguments for |
fit |
character vector describing how to fit the model. The default
is joint likelihood estimation of all parameters, |
ACSiter |
number of iterations for |
The function implements multivariate conditional transformation models
as described by Klein et al (2020).
Below is a simple example for an unconditional bivariate distribution.
See demo("undernutrition", package = "tram") for a conditional
three-variate example.
An object of class Mmlt with coef and predict
methods.
Nadja Klein, Torsten Hothorn, Luisa Barbanti, Thomas Kneib (2022), Multivariate Conditional Transformation Models. Scandinavian Journal of Statistics, 49, 116–142, doi:10.1111/sjos.12501.
Torsten Hothorn (2024), On Nonparanormal Likelihoods. doi:10.48550/arXiv.2408.17346.
data("cars") ### fit unconditional bivariate distribution of speed and distance to stop ## fit unconditional marginal transformation models m_speed <- BoxCox(speed ~ 1, data = cars, support = ss <- c(4, 25), add = c(-5, 5)) m_dist <- BoxCox(dist ~ 1, data = cars, support = sd <- c(0, 120), add = c(-5, 5)) ## fit multivariate unconditional transformation model m_speed_dist <- Mmlt(m_speed, m_dist, formula = ~ 1, data = cars) ## log-likelihood logLik(m_speed_dist) sum(predict(m_speed_dist, newdata = cars, type = "density", log = TRUE)) ## Wald test of independence of speed and dist (the "dist.speed.(Intercept)" ## coefficient) summary(m_speed_dist) ## LR test comparing to independence model LR <- 2 * (logLik(m_speed_dist) - logLik(m_speed) - logLik(m_dist)) pchisq(LR, df = 1, lower.tail = FALSE) ## constrain lambda to zero and fit independence model ## => log-likelihood is the sum of the marginal log-likelihoods mI <- Mmlt(m_speed, m_dist, formula = ~1, data = cars, fixed = c("dist.speed.(Intercept)" = 0)) logLik(m_speed) + logLik(m_dist) logLik(mI) ## linear correlation, ie Pearson correlation of speed and dist after ## transformation to bivariate normality (r <- coef(m_speed_dist, type = "Corr")) ## Spearman's rho (rank correlation) of speed and dist on original scale (rs <- coef(m_speed_dist, type = "Spearman")) ## evaluate joint and marginal densities (needs to be more user-friendly) nd <- expand.grid(c(nd_s <- mkgrid(m_speed, 100), nd_d <- mkgrid(m_dist, 100))) nd$d <- predict(m_speed_dist, newdata = nd, type = "density") ## compute marginal densities nd_s <- as.data.frame(nd_s) nd_s$d <- predict(m_speed_dist, newdata = nd_s, margins = 1L, type = "density") nd_d <- as.data.frame(nd_d) nd_d$d <- predict(m_speed_dist, newdata = nd_d, margins = 2L, type = "density") ## plot bivariate and marginal distribution col1 <- rgb(.1, .1, .1, .9) col2 <- rgb(.1, .1, .1, .5) w <- c(.8, .2) layout(matrix(c(2, 1, 4, 3), nrow = 2), width = w, height = rev(w)) par(mai = c(1, 1, 0, 0) * par("mai")) sp <- unique(nd$speed) di <- unique(nd$dist) d <- matrix(nd$d, nrow = length(sp)) contour(sp, di, d, xlab = "Speed (in mph)", ylab = "Distance (in ft)", xlim = ss, ylim = sd, col = col1) points(cars$speed, cars$dist, pch = 19, col = col2) mai <- par("mai") par(mai = c(0, 1, 0, 1) * mai) plot(d ~ speed, data = nd_s, xlim = ss, type = "n", axes = FALSE, xlab = "", ylab = "") polygon(nd_s$speed, nd_s$d, col = col2, border = FALSE) par(mai = c(1, 0, 1, 0) * mai) plot(dist ~ d, data = nd_d, ylim = sd, type = "n", axes = FALSE, xlab = "", ylab = "") polygon(nd_d$d, nd_d$dist, col = col2, border = FALSE) ### NOTE: marginal densities are NOT normal, nor is the joint ### distribution. The non-normal shape comes from the data-driven ### transformation of both variables to joint normality in this model.data("cars") ### fit unconditional bivariate distribution of speed and distance to stop ## fit unconditional marginal transformation models m_speed <- BoxCox(speed ~ 1, data = cars, support = ss <- c(4, 25), add = c(-5, 5)) m_dist <- BoxCox(dist ~ 1, data = cars, support = sd <- c(0, 120), add = c(-5, 5)) ## fit multivariate unconditional transformation model m_speed_dist <- Mmlt(m_speed, m_dist, formula = ~ 1, data = cars) ## log-likelihood logLik(m_speed_dist) sum(predict(m_speed_dist, newdata = cars, type = "density", log = TRUE)) ## Wald test of independence of speed and dist (the "dist.speed.(Intercept)" ## coefficient) summary(m_speed_dist) ## LR test comparing to independence model LR <- 2 * (logLik(m_speed_dist) - logLik(m_speed) - logLik(m_dist)) pchisq(LR, df = 1, lower.tail = FALSE) ## constrain lambda to zero and fit independence model ## => log-likelihood is the sum of the marginal log-likelihoods mI <- Mmlt(m_speed, m_dist, formula = ~1, data = cars, fixed = c("dist.speed.(Intercept)" = 0)) logLik(m_speed) + logLik(m_dist) logLik(mI) ## linear correlation, ie Pearson correlation of speed and dist after ## transformation to bivariate normality (r <- coef(m_speed_dist, type = "Corr")) ## Spearman's rho (rank correlation) of speed and dist on original scale (rs <- coef(m_speed_dist, type = "Spearman")) ## evaluate joint and marginal densities (needs to be more user-friendly) nd <- expand.grid(c(nd_s <- mkgrid(m_speed, 100), nd_d <- mkgrid(m_dist, 100))) nd$d <- predict(m_speed_dist, newdata = nd, type = "density") ## compute marginal densities nd_s <- as.data.frame(nd_s) nd_s$d <- predict(m_speed_dist, newdata = nd_s, margins = 1L, type = "density") nd_d <- as.data.frame(nd_d) nd_d$d <- predict(m_speed_dist, newdata = nd_d, margins = 2L, type = "density") ## plot bivariate and marginal distribution col1 <- rgb(.1, .1, .1, .9) col2 <- rgb(.1, .1, .1, .5) w <- c(.8, .2) layout(matrix(c(2, 1, 4, 3), nrow = 2), width = w, height = rev(w)) par(mai = c(1, 1, 0, 0) * par("mai")) sp <- unique(nd$speed) di <- unique(nd$dist) d <- matrix(nd$d, nrow = length(sp)) contour(sp, di, d, xlab = "Speed (in mph)", ylab = "Distance (in ft)", xlim = ss, ylim = sd, col = col1) points(cars$speed, cars$dist, pch = 19, col = col2) mai <- par("mai") par(mai = c(0, 1, 0, 1) * mai) plot(d ~ speed, data = nd_s, xlim = ss, type = "n", axes = FALSE, xlab = "", ylab = "") polygon(nd_s$speed, nd_s$d, col = col2, border = FALSE) par(mai = c(1, 0, 1, 0) * mai) plot(dist ~ d, data = nd_d, ylim = sd, type = "n", axes = FALSE, xlab = "", ylab = "") polygon(nd_d$d, nd_d$dist, col = col2, border = FALSE) ### NOTE: marginal densities are NOT normal, nor is the joint ### distribution. The non-normal shape comes from the data-driven ### transformation of both variables to joint normality in this model.
Marginally interpretable transformation models for clustered data.
mtram(object, formula, data, grd = SparseGrid::createSparseGrid(type = "KPU", dimension = length(rt$cnms[[1]]), k = 10), Hessian = FALSE, tol = .Machine$double.eps, ...)mtram(object, formula, data, grd = SparseGrid::createSparseGrid(type = "KPU", dimension = length(rt$cnms[[1]]), k = 10), Hessian = FALSE, tol = .Machine$double.eps, ...)
object |
A |
formula |
A formula specifying the random effects. |
data |
A data frame. |
grd |
A sparse grid used for numerical integration to get the likelihood. |
Hessian |
A logical, if |
tol |
numerical tolerance. |
... |
Additional argument. |
A Gaussian copula with a correlation structure obtained from a random
intercept or random intercept / random slope model (that is, clustered or
longitudinal data can by modelled only) is used to capture the
correlations whereas the marginal distributions are described by a
transformation model. The methodology is described in Barbanti and Hothorn
(2022) and examples are given in the mtram package vignette.
Only coef() and logLik() methods are available at the
moment, see vignette("mtram", package = "tram") for worked
examples.
An object of class tram with coef() and logLik()
methods.
Luisa Barbanti and Torsten Hothorn (2023). A Transformation Perspective on Marginal and Conditional Models, Biostatistics, doi:10.1093/biostatistics/kxac048.
### For illustrations see ## vignette("mtram", package = "tram") ## or ## demo("mtram", package = "tram")### For illustrations see ## vignette("mtram", package = "tram") ## or ## demo("mtram", package = "tram")
P-values for a parameter in a linear transformation model and corresponding confidence intervals obtained from by the permutation principle
perm_test(object, ...) ## S3 method for class 'tram' perm_test(object, parm = names(coef(object)), statistic = c("Score", "Likelihood", "Wald"), alternative = c("two.sided", "less", "greater"), nullvalue = 0, confint = TRUE, level = .95, Taylor = FALSE, block_permutation = TRUE, maxsteps = 25, ...)perm_test(object, ...) ## S3 method for class 'tram' perm_test(object, parm = names(coef(object)), statistic = c("Score", "Likelihood", "Wald"), alternative = c("two.sided", "less", "greater"), nullvalue = 0, confint = TRUE, level = .95, Taylor = FALSE, block_permutation = TRUE, maxsteps = 25, ...)
object |
an object of class |
parm |
a vector of names of parameters to be tested.
These parameters must be present in |
statistic |
a character string specifying the statistic to be
permuted. The default |
alternative |
a character string specifying the alternative hypothesis,
must be one of |
nullvalue |
a number specifying an optional parameter used to form the null hypothesis. |
confint |
a logical indicating whether a confidence interval should be
computed. Score confidence intervals are computed by default. A
1st order Taylor approximation to the Score statistc is used with
|
level |
the confidence level. |
block_permutation |
a logical indicating wheather stratifying variables shall be interpreted as blocks defining admissible permutations. |
Taylor |
a logical requesting the use of a 1st order Taylor approximation when inverting the score statistic. |
maxsteps |
number of function evaluations when inverting the score statistic for computing confidence intervals. |
... |
additional arguments to |
Permutation test for one single parameters in the linear
predictor of object is computed. This parameters must be present
in object. This is somewhat experimental and not recommended for
serious practical use (yet!).
An object of class htest or a list thereof. See Coxph
for an example.
## Tritiated Water Diffusion Across Human Chorioamnion ## Hollander and Wolfe (1999, p. 110, Tab. 4.1) diffusion <- data.frame( pd = c(0.80, 0.83, 1.89, 1.04, 1.45, 1.38, 1.91, 1.64, 0.73, 1.46, 1.15, 0.88, 0.90, 0.74, 1.21), age = factor(rep(c("At term", "12-26 Weeks"), c(10, 5))) ) ### plot the two quantile functions boxplot(pd ~ age, data = diffusion) ### the Wilcoxon rank sum test, with a confidence interval ### for a median shift wilcox.test(pd ~ age, data = diffusion, conf.int = TRUE, exact = TRUE) ### a corresponding parametric transformation model with a log-odds ratio ### difference parameter, ie a difference on the log-odds scale md <- Colr(pd ~ age, data = diffusion) ### assess model fit by plotting estimated distribution fcts agef <- sort(unique(diffusion$age)) col <- c("black", "darkred") plot(as.mlt(md), newdata = data.frame(age = agef), type = "distribution", col = col) legend("bottomright", col = col, lty = 1, legend = levels(agef), bty = "n", pch = 19) ## compare with ECDFs: not too bad (but not good, either) npfit <- with(diffusion, tapply(pd, age, ecdf)) lines(npfit[[1]], col = col[1]) lines(npfit[[2]], col = col[2]) ### Wald confidence interval confint(md) ### Likelihood confidence interval confint(profile(md)) ### Score confidence interval confint(score_test(md)) confint(score_test(md, Taylor = TRUE)) ### exact permutation score test (pt <- perm_test(md, confint = TRUE, distribution = "exact")) (pt <- perm_test(md, confint = TRUE, distribution = "exact", Taylor = TRUE)) ### compare with probabilistic indices obtained from asht::wmwTest if (require("asht", warn.conflicts = FALSE)) { print(wt2 <- wmwTest(pd ~ I(relevel(age, "At term")), data = diffusion, method = "exact.ce")) ### as log-odds ratios print(PI(prob = wt2$conf.int)) print(PI(prob = wt2$estimate)) }## Tritiated Water Diffusion Across Human Chorioamnion ## Hollander and Wolfe (1999, p. 110, Tab. 4.1) diffusion <- data.frame( pd = c(0.80, 0.83, 1.89, 1.04, 1.45, 1.38, 1.91, 1.64, 0.73, 1.46, 1.15, 0.88, 0.90, 0.74, 1.21), age = factor(rep(c("At term", "12-26 Weeks"), c(10, 5))) ) ### plot the two quantile functions boxplot(pd ~ age, data = diffusion) ### the Wilcoxon rank sum test, with a confidence interval ### for a median shift wilcox.test(pd ~ age, data = diffusion, conf.int = TRUE, exact = TRUE) ### a corresponding parametric transformation model with a log-odds ratio ### difference parameter, ie a difference on the log-odds scale md <- Colr(pd ~ age, data = diffusion) ### assess model fit by plotting estimated distribution fcts agef <- sort(unique(diffusion$age)) col <- c("black", "darkred") plot(as.mlt(md), newdata = data.frame(age = agef), type = "distribution", col = col) legend("bottomright", col = col, lty = 1, legend = levels(agef), bty = "n", pch = 19) ## compare with ECDFs: not too bad (but not good, either) npfit <- with(diffusion, tapply(pd, age, ecdf)) lines(npfit[[1]], col = col[1]) lines(npfit[[2]], col = col[2]) ### Wald confidence interval confint(md) ### Likelihood confidence interval confint(profile(md)) ### Score confidence interval confint(score_test(md)) confint(score_test(md, Taylor = TRUE)) ### exact permutation score test (pt <- perm_test(md, confint = TRUE, distribution = "exact")) (pt <- perm_test(md, confint = TRUE, distribution = "exact", Taylor = TRUE)) ### compare with probabilistic indices obtained from asht::wmwTest if (require("asht", warn.conflicts = FALSE)) { print(wt2 <- wmwTest(pd ~ I(relevel(age, "At term")), data = diffusion, method = "exact.ce")) ### as log-odds ratios print(PI(prob = wt2$conf.int)) print(PI(prob = wt2$estimate)) }
Some regression models for ordered categorical responses
Polr(formula, data, subset, weights, offset, cluster, na.action = na.omit, method = c("logistic", "probit", "loglog", "cloglog", "cauchit"), ...)Polr(formula, data, subset, weights, offset, cluster, na.action = na.omit, method = c("logistic", "probit", "loglog", "cloglog", "cauchit"), ...)
formula |
an object of class |
data |
an optional data frame, list or environment (or object
coercible by |
subset |
an optional vector specifying a subset of observations to be used in the fitting process. |
weights |
an optional vector of weights to be used in the fitting
process. Should be |
offset |
this can be used to specify an _a priori_ known component to
be included in the linear predictor during fitting. This
should be |
cluster |
optional factor with a cluster ID employed for computing clustered covariances. |
na.action |
a function which indicates what should happen when the data
contain |
method |
a character describing the link function. |
... |
additional arguments to |
Models for ordered categorical responses reusing the interface of
polr. Allows for stratification, censoring and
trunction.
The model is defined with a negative shift term, thus exp(coef())
is the multiplicative change of the odds ratio (conditional odds for
reference divided by conditional odds of treatment or for a one unit
increase in a numeric variable). Large values of the
linear predictor correspond to large values of the conditional
expectation response (but this relationship is nonlinear).
An object of class Polr, with corresponding coef,
vcov, logLik, estfun, summary,
print, plot and predict methods.
Torsten Hothorn, Lisa Moest, Peter Buehlmann (2018), Most Likely Transformations, Scandinavian Journal of Statistics, 45(1), 110–134, doi:10.1111/sjos.12291.
data("wine", package = "ordinal") library("MASS") polr(rating ~ temp + contact, data = wine) Polr(rating ~ temp + contact, data = wine)data("wine", package = "ordinal") library("MASS") polr(rating ~ temp + contact, data = wine) Polr(rating ~ temp + contact, data = wine)
P-values and confidence intervals for parameters in linear transformation models obtained from by the score test principle
score_test(object, ...) ## S3 method for class 'tram' score_test(object, parm = names(coef(object)), alternative = c("two.sided", "less", "greater"), nullvalue = 0, confint = TRUE, level = .95, Taylor = FALSE, maxsteps = 25, ...)score_test(object, ...) ## S3 method for class 'tram' score_test(object, parm = names(coef(object)), alternative = c("two.sided", "less", "greater"), nullvalue = 0, confint = TRUE, level = .95, Taylor = FALSE, maxsteps = 25, ...)
object |
an object of class |
parm |
a vector of names of parameters to be tested.
These parameters must be present in |
alternative |
a character string specifying the alternative hypothesis,
must be one of |
nullvalue |
a number specifying an optional parameter used to form the null hypothesis. |
confint |
a logical indicating whether a confidence interval should be
computed. Score confidence intervals are computed by default. A
1st order Taylor approximation to the Score statistc is used with
|
level |
the confidence level. |
Taylor |
a logical requesting the use of a 1st order Taylor approximation when inverting the score statistic. |
maxsteps |
number of function evaluations when inverting the score statistic for computing confidence intervals. |
... |
additional arguments, currently ignored. |
Score tests and confidence intervals for the parameters in the linear
predictor of object are computed. These parameters must be present
in object.
An object of class htest or a list thereof. See Coxph
for an example. A corresponding permutation test for parameters in a
transformation models is available in
perm_test.
Weibull, log-normal, log-logistic and other parametric models (not exclusively) for survival analysis
Survreg(formula, data, subset, weights, offset, cluster, na.action = na.omit, dist = c("weibull", "logistic", "gaussian", "exponential", "rayleigh", "loggaussian", "lognormal", "loglogistic"), scale = 0, ...)Survreg(formula, data, subset, weights, offset, cluster, na.action = na.omit, dist = c("weibull", "logistic", "gaussian", "exponential", "rayleigh", "loggaussian", "lognormal", "loglogistic"), scale = 0, ...)
formula |
an object of class |
data |
an optional data frame, list or environment (or object
coercible by |
subset |
an optional vector specifying a subset of observations to be used in the fitting process. |
weights |
an optional vector of weights to be used in the fitting
process. Should be |
offset |
this can be used to specify an _a priori_ known component to
be included in the linear predictor during fitting. This
should be |
cluster |
optional factor with a cluster ID employed for computing clustered covariances. |
na.action |
a function which indicates what should happen when the data
contain |
dist |
character defining the conditional distribution of the (not necessarily positive) response, current choices include Weibull, logistic, normal, exponential, Rayleigh, log-normal (same as log-gaussian), or log-logistic. |
scale |
a fixed value for the scale parameter(s). |
... |
additional arguments to |
Parametric survival models reusing the interface of
survreg. The parameterisation is, however, a little
different, see the package vignette.
The model is defined with a negative shift term. Large values of the linear predictor correspond to large values of the conditional expectation response (but this relationship is nonlinear). Parameters are log-hazard ratios comparing a reference with treatment (or a one unit increase in a numeric variable).
An object of class Survreg, with corresponding coef,
vcov, logLik, estfun, summary,
print, plot and predict methods.
Torsten Hothorn, Lisa Moest, Peter Buehlmann (2018), Most Likely Transformations, Scandinavian Journal of Statistics, 45(1), 110–134, doi:10.1111/sjos.12291.
data("GBSG2", package = "TH.data") library("survival") survreg(Surv(time, cens) ~ horTh, data = GBSG2) Survreg(Surv(time, cens) ~ horTh, data = GBSG2)data("GBSG2", package = "TH.data") library("survival") survreg(Surv(time, cens) ~ horTh, data = GBSG2) Survreg(Surv(time, cens) ~ horTh, data = GBSG2)
Likelihood-inference for stratified linear transformation models, including linear shift-scale transformation models.
tram(formula, data, subset, weights, offset, cluster, na.action = na.omit, distribution = c("Normal", "Logistic", "MinExtrVal", "MaxExtrVal", "Exponential", "Cauchy", "Laplace"), frailty = c("None", "Gamma", "InvGauss", "PositiveStable"), transformation = c("discrete", "linear", "logarithmic", "smooth"), LRtest = TRUE, prob = c(0.1, 0.9), support = NULL, bounds = NULL, add = c(0, 0), order = 6, negative = TRUE, remove_intercept = TRUE, scale = TRUE, scale_shift = FALSE, extrapolate = FALSE, log_first = FALSE, sparse_nlevels = Inf, model_only = FALSE, constraints = NULL, ...) tram_data(formula, data, subset, weights, offset, cluster, na.action = na.omit)tram(formula, data, subset, weights, offset, cluster, na.action = na.omit, distribution = c("Normal", "Logistic", "MinExtrVal", "MaxExtrVal", "Exponential", "Cauchy", "Laplace"), frailty = c("None", "Gamma", "InvGauss", "PositiveStable"), transformation = c("discrete", "linear", "logarithmic", "smooth"), LRtest = TRUE, prob = c(0.1, 0.9), support = NULL, bounds = NULL, add = c(0, 0), order = 6, negative = TRUE, remove_intercept = TRUE, scale = TRUE, scale_shift = FALSE, extrapolate = FALSE, log_first = FALSE, sparse_nlevels = Inf, model_only = FALSE, constraints = NULL, ...) tram_data(formula, data, subset, weights, offset, cluster, na.action = na.omit)
formula |
an object of class |
data |
an optional data frame, list or environment (or object
coercible by |
subset |
an optional vector specifying a subset of observations to be used in the fitting process. |
weights |
an optional vector of case weights to be used in the fitting
process. Should be |
offset |
this can be used to specify an _a priori_ known component to
be included in the linear predictor during fitting. This
should be |
cluster |
optional factor with a cluster ID employed for computing clustered covariances. |
na.action |
a function which indicates what should happen when the data
contain |
distribution |
character specifying how the transformation function is mapped into probabilities. Available choices include the cumulative distribution functions of the standard normal, the standard logistic and the standard minimum extreme value distribution. |
frailty |
character specifying the addition of a frailty term, that is, a random component added to the linear predictor of the model, with specific distribution (Gamma, inverse Gaussian, positive stable). |
transformation |
character specifying the complexity of the response-transformation. For discrete responses, one parameter is assigned to each level (except the last one), for continuous responses linear, log-linear and smooth (parameterised as a Bernstein polynomial) function are implemented. |
LRtest |
logical specifying if a likelihood-ratio test for the null of all coefficients in the linear predictor being zero shall be performed. |
prob |
two probabilities giving quantiles of the response defining the support of a smooth
Bernstein polynomial (if |
support |
a vector of two elements; the support of a smooth
Bernstein polynomial (if |
bounds |
an interval defining the bounds of a real sample space. |
add |
add these values to the support before generating a grid via
|
order |
integer >= 1 defining the order of the Bernstein polynomial
(if |
negative |
logical defining the sign of the linear predictor. |
remove_intercept |
logical defining if the intercept shall be removed
from the linear shift predictor in favour of an (typically implicit) intercept
in the baseline transformation. If |
scale |
logical defining if variables in the linear predictor shall be scaled. Scaling is internally used for model estimation, rescaled coefficients are reported in model output. |
scale_shift |
a logical choosing between two different model types in
the presence of a |
extrapolate |
logical defining the behaviour of the Bernstein transformation
function outside |
sparse_nlevels |
integer; use a sparse model matrix if the number
of levels of an ordered factor is at least as large as
|
log_first |
logical; if |
model_only |
logical, if |
constraints |
additional constraints on regression coefficients in
the linear predictor of the form |
... |
additional arguments. |
The model formula is of the form y | s ~ x | z where y is an at
least ordered response variable, s are the variables defining strata
and x defines the linear predictor. Optionally, z defines a
scaling term (see ctm). y ~ x defines a model
without strata (but response-varying intercept function) and y | s ~
0 sets-up response-varying coefficients for all variables in s.
The two functions tram and tram_data are not intended
to be called directly by users. Instead,
functions Coxph (Cox proportional hazards models),
Survreg (parametric survival models),
Polr (models for ordered categorical responses),
Lm (normal linear models),
BoxCox (non-normal linear models) or
Colr (continuous outcome logistic regression) allow
direct access to the corresponding models.
The model class and the specific models implemented in tram are explained in the package vignette of package tram. The underlying theory of most likely transformations is presented in Hothorn et al. (2018), computational and modelling aspects in more complex situations are discussed by Hothorn (2018).
An object of class tram inheriting from mlt.
Torsten Hothorn, Lisa Moest, Peter Buehlmann (2018), Most Likely Transformations, Scandinavian Journal of Statistics, 45(1), 110–134, doi:10.1111/sjos.12291.
Torsten Hothorn (2020), Most Likely Transformations: The mlt Package, Journal of Statistical Software, 92(1), doi:10.18637/jss.v092.i01.
Sandra Siegfried, Lucas Kook, Torsten Hothorn (2023), Distribution-Free Location-Scale Regression, The American Statistician, doi:10.1080/00031305.2023.2203177.
data("BostonHousing2", package = "mlbench") ### unconstrained regression coefficients ### BoxCox calls tram internally m1 <- BoxCox(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) ### now with two constraints on regression coefficients m2 <- BoxCox(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2, constraints = c("crim >= 0", "chas1 + rm >= 1.5")) coef(m1) coef(m2) K <- matrix(0, nrow = 2, ncol = length(coef(m2))) colnames(K) <- names(coef(m2)) K[1, "crim"] <- 1 K[2, c("chas1", "rm")] <- 1 m3 <- BoxCox(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2, constraints = list(K, c(0, 1.5))) all.equal(coef(m2), coef(m3))data("BostonHousing2", package = "mlbench") ### unconstrained regression coefficients ### BoxCox calls tram internally m1 <- BoxCox(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) ### now with two constraints on regression coefficients m2 <- BoxCox(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2, constraints = c("crim >= 0", "chas1 + rm >= 1.5")) coef(m1) coef(m2) K <- matrix(0, nrow = 2, ncol = length(coef(m2))) colnames(K) <- names(coef(m2)) K[1, "crim"] <- 1 K[2, c("chas1", "rm")] <- 1 m3 <- BoxCox(cmedv ~ chas + crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2, constraints = list(K, c(0, 1.5))) all.equal(coef(m2), coef(m3))
Methods for objects inheriting from class tram
## S3 method for class 'tram' as.mlt(object) ## S3 method for class 'tram' model.frame(formula, ...) ## S3 method for class 'tram' model.matrix(object, data = object$data, with_baseline = FALSE, ...) ## S3 method for class 'stram' model.matrix(object, data = object$data, with_baseline = FALSE, what = c("shifting", "scaling"), ...) ## S3 method for class 'tram' coef(object, with_baseline = FALSE, ...) ## S3 method for class 'Lm' coef(object, as.lm = FALSE, ...) ## S3 method for class 'Survreg' coef(object, as.survreg = FALSE, ...) ## S3 method for class 'tram' vcov(object, with_baseline = FALSE, complete = FALSE, ...) ## S3 method for class 'tram' logLik(object, parm = coef(as.mlt(object), fixed = FALSE), ...) ## S3 method for class 'tram' estfun(x, parm = coef(as.mlt(x), fixed = FALSE), ...) ## S3 method for class 'tram' predict(object, newdata = model.frame(object), type = c("lp", "trafo", "distribution", "logdistribution", "survivor", "logsurvivor", "density", "logdensity", "hazard", "loghazard", "cumhazard", "logcumhazard", "odds", "logodds", "quantile"), ...) ## S3 method for class 'stram' predict(object, newdata = model.frame(object), type = c("lp", "trafo", "distribution", "logdistribution", "survivor", "logsurvivor", "density", "logdensity", "hazard", "loghazard", "cumhazard", "logcumhazard", "odds", "logodds", "quantile"), what = c("shifting", "scaling"), ...) ## S3 method for class 'tram' plot(x, newdata = model.frame(x), which = c("QQ-PIT", "baseline only", "distribution"), confidence = c("none", "interval", "band"), level = 0.95, K = 50, cheat = K, col = "black", fill = "lightgrey", lwd = 1, ...) ## S3 method for class 'tram' residuals(object, ...) ## S3 method for class 'tram' PI(object, newdata = model.frame(object), reference = 0, one2one = FALSE, ...) ## Default S3 method: PI(object, prob, link = "logistic", ...) ## S3 method for class 'tram' OVL(object, newdata = model.frame(object), reference = 0, one2one = FALSE, ...) ## Default S3 method: OVL(object, link = "logistic", ...) ## S3 method for class 'tram' TV(object, newdata = model.frame(object), reference = 0, one2one = FALSE, ...) ## Default S3 method: TV(object, link = "logistic", ...) ## S3 method for class 'tram' L1(object, newdata = model.frame(object), reference = 0, one2one = FALSE, ...) ## Default S3 method: L1(object, link = "logistic", ...) ## S3 method for class 'tram' ROC(object, newdata = model.frame(object), reference = 0, prob = 1:99 / 100, one2one = FALSE, ...) ## Default S3 method: ROC(object, prob = 1:99 / 100, link = "logistic", ...) ## S3 method for class 'ROCtram' plot(x, lty = 1:ncol(x), col = "black", fill = "lightgrey", lwd = 1, ...)## S3 method for class 'tram' as.mlt(object) ## S3 method for class 'tram' model.frame(formula, ...) ## S3 method for class 'tram' model.matrix(object, data = object$data, with_baseline = FALSE, ...) ## S3 method for class 'stram' model.matrix(object, data = object$data, with_baseline = FALSE, what = c("shifting", "scaling"), ...) ## S3 method for class 'tram' coef(object, with_baseline = FALSE, ...) ## S3 method for class 'Lm' coef(object, as.lm = FALSE, ...) ## S3 method for class 'Survreg' coef(object, as.survreg = FALSE, ...) ## S3 method for class 'tram' vcov(object, with_baseline = FALSE, complete = FALSE, ...) ## S3 method for class 'tram' logLik(object, parm = coef(as.mlt(object), fixed = FALSE), ...) ## S3 method for class 'tram' estfun(x, parm = coef(as.mlt(x), fixed = FALSE), ...) ## S3 method for class 'tram' predict(object, newdata = model.frame(object), type = c("lp", "trafo", "distribution", "logdistribution", "survivor", "logsurvivor", "density", "logdensity", "hazard", "loghazard", "cumhazard", "logcumhazard", "odds", "logodds", "quantile"), ...) ## S3 method for class 'stram' predict(object, newdata = model.frame(object), type = c("lp", "trafo", "distribution", "logdistribution", "survivor", "logsurvivor", "density", "logdensity", "hazard", "loghazard", "cumhazard", "logcumhazard", "odds", "logodds", "quantile"), what = c("shifting", "scaling"), ...) ## S3 method for class 'tram' plot(x, newdata = model.frame(x), which = c("QQ-PIT", "baseline only", "distribution"), confidence = c("none", "interval", "band"), level = 0.95, K = 50, cheat = K, col = "black", fill = "lightgrey", lwd = 1, ...) ## S3 method for class 'tram' residuals(object, ...) ## S3 method for class 'tram' PI(object, newdata = model.frame(object), reference = 0, one2one = FALSE, ...) ## Default S3 method: PI(object, prob, link = "logistic", ...) ## S3 method for class 'tram' OVL(object, newdata = model.frame(object), reference = 0, one2one = FALSE, ...) ## Default S3 method: OVL(object, link = "logistic", ...) ## S3 method for class 'tram' TV(object, newdata = model.frame(object), reference = 0, one2one = FALSE, ...) ## Default S3 method: TV(object, link = "logistic", ...) ## S3 method for class 'tram' L1(object, newdata = model.frame(object), reference = 0, one2one = FALSE, ...) ## Default S3 method: L1(object, link = "logistic", ...) ## S3 method for class 'tram' ROC(object, newdata = model.frame(object), reference = 0, prob = 1:99 / 100, one2one = FALSE, ...) ## Default S3 method: ROC(object, prob = 1:99 / 100, link = "logistic", ...) ## S3 method for class 'ROCtram' plot(x, lty = 1:ncol(x), col = "black", fill = "lightgrey", lwd = 1, ...)
object, formula, x
|
a fitted stratified linear transformation model inheriting
from class |
data |
an optional data frame. |
with_baseline |
logical, if |
as.lm |
logical, return parameters in the |
as.survreg |
logical, return parameters in the |
parm |
model parameters, including baseline parameters. |
complete |
currently ignored |
newdata |
an optional data frame of new observations. |
reference |
an optional data frame of reference observations, or a numeric vector of reference values. |
type |
type of prediction, current options include
linear predictors ( |
which |
type of plot, either a QQ plot of the probability-integral
transformed observations ( |
what |
type of model matrix / linear predictor: |
confidence |
type of uncertainty assessment. |
level |
confidence level. |
K |
number of grid points in the response, see
|
cheat |
reduced number of grid points for the computation
of confidence bands, see |
col |
line color. |
fill |
fill color. |
lwd |
line width. |
lty |
line type. |
prob |
a numeric vector of probabilities.. |
link |
a character identifying a link function. |
one2one |
logical, compute the ROC curve (and derived measures)
comparing each row in |
... |
additional arguments to the underlying methods for class
|
coef can be used to get (and set) model parameters,
logLik evaluates the log-likelihood (also for
parameters other than the maximum likelihood estimate);
vcov returns the estimated variance-covariance matrix (possibly
taking cluster into account) and
and estfun gives the score contribution by each observation.
predict and plot can be used to inspect the model on
different scales.
PI computes the probabilistic index (or concordance probability or
AUC) for all observations in newdata, relative to reference,
ie the probability
of observing a smaller value of a randomly sampled observation conditional
on compared to a randomly sampled reference observation, which
is conditional on . This is equivalent to the area under the
receiver operating curve (ROC). The probability only applies within
strata, response-varying coefficients are not allowed.
Under the same setup, OVL gives the overlap coefficient, which is
one minus the total variation and one minus half the distance
between the two conditional densities. The overlap coefficient is
identical to the Youden index and the Smirnov statistic.
PI and friends also accept an argument conf.level which
triggers computation of simultaneous
Wald confidence intervals for these measures.
Arguments in ... are forwarded to glht.
Torsten Hothorn, Lisa Moest, Peter Buehlmann (2018), Most Likely Transformations, Scandinavian Journal of Statistics, 45(1), 110–134, doi:10.1111/sjos.12291.
data("BostonHousing2", package = "mlbench") ### fit non-normal Box-Cox type linear model with two ### baseline functions (for houses near and off Charles River) BC_BH_2 <- BoxCox(cmedv | 0 + chas ~ crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) logLik(BC_BH_2) ### classical likelihood inference summary(BC_BH_2) ### coefficients of the linear predictor coef(BC_BH_2) ### plot linear predictor (mean of _transformed_ response) ### vs. observed values plot(predict(BC_BH_2, type = "lp"), BostonHousing2$cmedv) ### all coefficients coef(BC_BH_2, with_baseline = TRUE) ### compute predicted median along with 10% and 90% quantile for the first ### observations predict(BC_BH_2, newdata = BostonHousing2[1:3,], type = "quantile", prob = c(.1, .5, .9)) ### plot the predicted density for these observations plot(BC_BH_2, newdata = BostonHousing2[1:3, -1], which = "distribution", type = "density", K = 1000) ### evaluate the two baseline transformations, with confidence intervals nd <- model.frame(BC_BH_2)[1:2, -1] nd$chas <- factor(c("0", "1")) library("colorspace") col <- diverge_hcl(2, h = c(246, 40), c = 96, l = c(65, 90)) fill <- diverge_hcl(2, h = c(246, 40), c = 96, l = c(65, 90), alpha = .3) plot(BC_BH_2, which = "baseline only", newdata = nd, col = col, confidence = "interval", fill = fill, lwd = 2, xlab = "Median Value", ylab = expression(h[Y])) legend("bottomright", lty = 1, col = col, title = "Near Charles River", legend = c("no", "yes"), bty = "n") ### cars data; with quantile functions plot(dist ~ speed, data = cars) m <- Colr(dist ~ speed, data = cars) q <- predict(as.mlt(m), newdata = data.frame(speed = s <- 6:25), type = "quantile", prob = c(1, 5, 9) / 10) lines(s, q[1,]) lines(s, q[2,]) lines(s, q[3,]) nd <- data.frame(speed = s <- as.double(1:5 * 5)) # Prob(dist at speed s > dist at speed 0) # speed 0 is reference, not a good choice here PI(m, newdata = nd) # Prob(dist at speed s > dist at speed 15) lp15 <- c(predict(m, newdata = data.frame(speed = 15))) PI(m, newdata = nd, reference = lp15) PI(m, newdata = nd, reference = nd[3,,drop = FALSE]) # Prob(dist at speed s' > dist at speed s) PI(m, newdata = nd, reference = nd) # essentially: lp <- predict(m, newdata = nd) PI(object = dist(lp)) # same, with simultaneous confidence intervals PI(m, newdata = nd, reference = nd, conf.level = .95) # plot ROC curves + confidence bands # compare speed 20 and 25 to speed 15 plot(ROC(m, newdata = nd[4:5,,drop = FALSE], reference = nd[3,,drop = FALSE], conf.level = 0.95)) # Overlap of conditional densities at speed s' and s OVL(m, newdata = nd, reference = nd) ### ROC analysis (takes too long for CRAN Windows) if (require("mlbench") && .Platform$OS.type != "windows") { layout(matrix(1:4, nrow = 2)) data("PimaIndiansDiabetes2", package = "mlbench") dia <- sort(unique(PimaIndiansDiabetes2$diabetes)) nd <- data.frame(diabetes = dia, age = 29, mass = 32) ### median values ### unconditional ROC analysis: glucose tolerance test m0 <- Colr(glucose ~ diabetes, data = PimaIndiansDiabetes2) # ROC curve + confidence band plot(ROC(m0, newdata = nd[2,,drop = FALSE], conf.level = .95)) # Wald interval for AUC PI(m0, newdata = nd[2,,drop = FALSE], conf.level = .95) # score interval for AUC PI(-c(coef(m0), score_test(m0)$conf.int[2:1])) ### adjusted ROC analysis for age and mass m1 <- Colr(glucose ~ diabetes + age + mass, data = PimaIndiansDiabetes2) # ROC curve + confidence band (this is the same for all ages / # masses) plot(ROC(m1, newdata = nd[2,,drop = FALSE], reference = nd[1,,drop = FALSE], conf.level = .95)) # Wald interval for adjusted AUC PI(m1, newdata = nd[2,,drop = FALSE], reference = nd[1,,drop = FALSE], conf.level = .95) # Score interval for adjusted AUC PI(-c(coef(m1)[1], score_test(m1, names(coef(m1))[1])$conf.int[2:1])) ### conditional ROC analysis: AUC regression ~ age + mass m2 <- Colr(glucose ~ diabetes * (age + mass), data = PimaIndiansDiabetes2) # ROC curve for a person with age = 29 and mass = 32 plot(ROC(m2, newdata = nd[2,,drop = FALSE], reference = nd[1,,drop = FALSE], conf.level = .95)) # AUC for persons ages 21:81, all with mass = 32 nd1 <- data.frame(diabetes = nd[1,"diabetes"], age = 21:81, mass = 32) nd2 <- data.frame(diabetes = nd[2,"diabetes"], age = 21:81, mass = 32) auc <- PI(m2, newdata = nd2, reference = nd1, one2one = TRUE, conf.level = 0.95) plot(nd1$age, auc[, "Estimate"], xlab = "Age (in years)", ylab = "AUC", ylim = c(0, 1), type = "l") lines(nd1$age, auc[, "lwr"], lty = 3) lines(nd1$age, auc[, "upr"], lty = 3) }data("BostonHousing2", package = "mlbench") ### fit non-normal Box-Cox type linear model with two ### baseline functions (for houses near and off Charles River) BC_BH_2 <- BoxCox(cmedv | 0 + chas ~ crim + zn + indus + nox + rm + age + dis + rad + tax + ptratio + b + lstat, data = BostonHousing2) logLik(BC_BH_2) ### classical likelihood inference summary(BC_BH_2) ### coefficients of the linear predictor coef(BC_BH_2) ### plot linear predictor (mean of _transformed_ response) ### vs. observed values plot(predict(BC_BH_2, type = "lp"), BostonHousing2$cmedv) ### all coefficients coef(BC_BH_2, with_baseline = TRUE) ### compute predicted median along with 10% and 90% quantile for the first ### observations predict(BC_BH_2, newdata = BostonHousing2[1:3,], type = "quantile", prob = c(.1, .5, .9)) ### plot the predicted density for these observations plot(BC_BH_2, newdata = BostonHousing2[1:3, -1], which = "distribution", type = "density", K = 1000) ### evaluate the two baseline transformations, with confidence intervals nd <- model.frame(BC_BH_2)[1:2, -1] nd$chas <- factor(c("0", "1")) library("colorspace") col <- diverge_hcl(2, h = c(246, 40), c = 96, l = c(65, 90)) fill <- diverge_hcl(2, h = c(246, 40), c = 96, l = c(65, 90), alpha = .3) plot(BC_BH_2, which = "baseline only", newdata = nd, col = col, confidence = "interval", fill = fill, lwd = 2, xlab = "Median Value", ylab = expression(h[Y])) legend("bottomright", lty = 1, col = col, title = "Near Charles River", legend = c("no", "yes"), bty = "n") ### cars data; with quantile functions plot(dist ~ speed, data = cars) m <- Colr(dist ~ speed, data = cars) q <- predict(as.mlt(m), newdata = data.frame(speed = s <- 6:25), type = "quantile", prob = c(1, 5, 9) / 10) lines(s, q[1,]) lines(s, q[2,]) lines(s, q[3,]) nd <- data.frame(speed = s <- as.double(1:5 * 5)) # Prob(dist at speed s > dist at speed 0) # speed 0 is reference, not a good choice here PI(m, newdata = nd) # Prob(dist at speed s > dist at speed 15) lp15 <- c(predict(m, newdata = data.frame(speed = 15))) PI(m, newdata = nd, reference = lp15) PI(m, newdata = nd, reference = nd[3,,drop = FALSE]) # Prob(dist at speed s' > dist at speed s) PI(m, newdata = nd, reference = nd) # essentially: lp <- predict(m, newdata = nd) PI(object = dist(lp)) # same, with simultaneous confidence intervals PI(m, newdata = nd, reference = nd, conf.level = .95) # plot ROC curves + confidence bands # compare speed 20 and 25 to speed 15 plot(ROC(m, newdata = nd[4:5,,drop = FALSE], reference = nd[3,,drop = FALSE], conf.level = 0.95)) # Overlap of conditional densities at speed s' and s OVL(m, newdata = nd, reference = nd) ### ROC analysis (takes too long for CRAN Windows) if (require("mlbench") && .Platform$OS.type != "windows") { layout(matrix(1:4, nrow = 2)) data("PimaIndiansDiabetes2", package = "mlbench") dia <- sort(unique(PimaIndiansDiabetes2$diabetes)) nd <- data.frame(diabetes = dia, age = 29, mass = 32) ### median values ### unconditional ROC analysis: glucose tolerance test m0 <- Colr(glucose ~ diabetes, data = PimaIndiansDiabetes2) # ROC curve + confidence band plot(ROC(m0, newdata = nd[2,,drop = FALSE], conf.level = .95)) # Wald interval for AUC PI(m0, newdata = nd[2,,drop = FALSE], conf.level = .95) # score interval for AUC PI(-c(coef(m0), score_test(m0)$conf.int[2:1])) ### adjusted ROC analysis for age and mass m1 <- Colr(glucose ~ diabetes + age + mass, data = PimaIndiansDiabetes2) # ROC curve + confidence band (this is the same for all ages / # masses) plot(ROC(m1, newdata = nd[2,,drop = FALSE], reference = nd[1,,drop = FALSE], conf.level = .95)) # Wald interval for adjusted AUC PI(m1, newdata = nd[2,,drop = FALSE], reference = nd[1,,drop = FALSE], conf.level = .95) # Score interval for adjusted AUC PI(-c(coef(m1)[1], score_test(m1, names(coef(m1))[1])$conf.int[2:1])) ### conditional ROC analysis: AUC regression ~ age + mass m2 <- Colr(glucose ~ diabetes * (age + mass), data = PimaIndiansDiabetes2) # ROC curve for a person with age = 29 and mass = 32 plot(ROC(m2, newdata = nd[2,,drop = FALSE], reference = nd[1,,drop = FALSE], conf.level = .95)) # AUC for persons ages 21:81, all with mass = 32 nd1 <- data.frame(diabetes = nd[1,"diabetes"], age = 21:81, mass = 32) nd2 <- data.frame(diabetes = nd[2,"diabetes"], age = 21:81, mass = 32) auc <- PI(m2, newdata = nd2, reference = nd1, one2one = TRUE, conf.level = 0.95) plot(nd1$age, auc[, "Estimate"], xlab = "Age (in years)", ylab = "AUC", ylim = c(0, 1), type = "l") lines(nd1$age, auc[, "lwr"], lty = 3) lines(nd1$age, auc[, "upr"], lty = 3) }