survival object can now be constructed manually with
survival() and simulated state probabilities can be computed from survival curves with
sim_stateprobs.survival(). These features are useful for partitioned survival analyses when a user would like to use a survival model not supported by
hesim to predict survival curves (#49).
tpmatrix() is more flexible:
There is now a
complement argument where users can define transitions that are “complements”. This is particularly helpful for creating large transition probability matrices since there is no longer a need to manually enter
"C" in the
... portion of
tpmatrix(). In other words, users can pass a single object to
tpmatrix() and use the
complement argument to ensure probabilities sum to 1 in each row. One reasonable workflow with
define_tparams() would be to (i) create a single matrix of initial values (say zeros), (ii) modify the transitions that differ from the initial values, and (iii) pass the resulting object to
tpmatrix() while using the
sep arguments can be used to name the columns (i.e., the transitions) and the states. The named states are, in turn, passed to the
$value element of a
tparams_transprobs object with
eval_rng() object can now be passed directly to
define_model(), meaning that users can sample parameters prior to defining a model. Previously users could only pass an
rng_def object to
define_model(), which meant that sampling could only occur while evaluating a
sim_qalys() are now exported functions that give users additional flexibility in their modeling pipelines and provide improved documentation for computation of expected values in cohort models.
sim_ev() is particularly useful for computing outcomes that depend on time in state other than costs or quality-adjusted life-years (QALYs).
Multiple absorbing states (or none at all) are possible in
hesim::IndivCtstm models (previously the final health state was always assumed to be a death state). In cohort models, the absorbing states can be set manually using the
absorbing field in the
hesim::CohortDtstmTrans class; if not, they are set automatically based on the transition probabilities. The number of health states in state value models (class
hesim::StateVals) must equal the number of health states in the transition models less the number of absorbing states.
create_CohortDtstmTrans.params_mlogit_list() method allows the transition component of a cohort discrete time state transition model (cDTSTM) to be created directly from multinomial logistic regression parameter objects.
The coefficient elements of parameter objects can be constructed from any object (e.g., data frame) than can be passed to
as.matrix() (rather than only from matrices as in previous versions). See, for instance,
A transition intensity matrix can be created from a multi-state model fit using
msm::msm() with the
qmatrix.msm() method. Similarly, a
CohortDtstmTrans object can be created with
... argument to
create_PsmCurves() now passes arguments to
object is of class
flexsurvreg_list. This allows more control over bootstrapping (i.e., use of the
plot_evpi() plot the cost-effectiveness plane, cost-effectiveness acceptability curve (CEAC), cost-effectiveness acceptability frontier (CEAF), and expected value of perfect information (EVPI), respectively.
The first column of each matrix listed in the
coef element returned by
create_params.flexsurvreg() is now named “(Intercept)” instead of the name of the corresponding parameter.
create_params() methods now use the argument
uncertainty to draw parameters and the old arguments
bootstrap are deprecated. This also affects
trans in the data table returned by the
$cumhazard() methods from the
hesim::CtstmTrans class has been renamed
hesim::Psmnow contains the
by_grpargument so that subgroup analyses can be performed.
There are new functions to construct (and debug the construction of) the multiple transition probability matrices stored in
tparams_transprobs() objects and used for cDTSTMs. These can either stored as 3-dimensional arrays or as 2-dimensional tabular objects (i.e.,
qmatrix() lets you store transition intensity matrices which can be used to construct transition probability matrices with the matrix exponential via
expmat(). The latter is a simple wrapper around
msm::MatrixExp() that computes the matrix exponential for all matrices in an array rather than just a single matrix.
apply_rr() applies relative risks (stored in a 2-dimensional tabular object) to (potentially multiple elements) of transition probability matrices stored in an array. This function is vectorized so it can be performed very quickly even for large arrays.
as.data.table.tparams_transprobs() converts the array of transition probability matrices stored in a
tparams_transprobs object into a
data.table which can be helpful for debugging to ensure that the right transition probability matrices correspond to the right observations (i.e., treatment strategies, patients, etc.).
2-dimensional tabular objects (in addition to vectors) can now be passed to
tpmatrix(). See the new examples.
A new dataset
hesim::onc3 was added as an example multi-state dataset for an oncology use case with 3 health states (Stable, Progression, Death) and 3 possible transitions (Stable -> Progression, Stable -> Death, and Progression -> Death). This is similar to
hesim::mstate3_exdata but does not allow for reversible transitions and does not contain cost or utility data.
as_pfs_os() can convert a multi-state dataset in the same format as
hesim::onc3 into a dataset with one row per patients containing time to event information for progression free survival (PFS) and overall survival (OS).
cycle_length field in
CohortDtstmTrans was fixed so that it corresponds to a model cycle in terms of years (e.g., a value of 2 means a model cycle is 2 years long and that state probabilities are computed every 2 years with
The simulated dataset
multinom3_exdata was fixed by removing a bug where some patients were simulated to have died more than once.
Fixed bug where the
$sim_costs() method of
IndivCtstm was erroneously returning a life-years column in addition to the costs column.
Modification to creation of input matrices from a
flexsurvreg object to properly capture levels of factor variables.
Minor updates to the documentation and fixes to small problems in the
C++ code identified with the CRAN package checks.
IndivCtstmTrans objects can be constructed from a
Survival models can randomly sample from piecewise exponential and proportional hazards Weibull distributions. A
fixed distribution has also been added so that survival times can be set to a single constant value. Random number generation from truncated versions of these distributions is also supported. Note that functionality beyond random number generation (e.g., hazard functions, cumulative hazard functions, cumulative density functions) is not yet complete. See
A new vignette incorporates the two bullets above and shows how a time-inhomogeneous Markov model can be simulated using individual patient simulation.
Disease progression (i.e., a trajectory through a multi-state model) can be simulated using the
sim_disease() method of the
A more computationally efficient approach to simulation of time-inhomogeneous Markov cohort models has been added to the corresponding vignette. This was aided by the new
The “Articles” on the package website have been reorganized so that they align more closely with the different types of economic models.
hesim now supports cDTSTM via
hesim::CohortDtstm objects. Users can build a model by either fitting multinomial logistic regressions with
nnet::multinom() or with a mathematical expression using
$summarize() methods now have a
by_grp option to facilitate subgroup analyses.
hesim::CohortDtstm class simulates cDTSTMs. State transitions in a cDTSTM are simulated using the
hesim::CohortDtstmTrans class, which can be constructed from a
multinom_list() object or using
$summarize() methods now have a
by_grp = "TRUE" option to facilitate subgroup analyses. If
by_grp = "FALSE", then estimates are aggregated across groups. A new
patient_wt argument in the
patients table of
hesim_data() can be used to weight groups during the aggregation.
General cumulative hazard functions can now be simulated using a discrete time approximation where the probability of an event during each time period is simulated from a Bernoulli distribution. This is more efficient than the previous method based on a
C++ version of the
sample() function. See the
random_method = "discrete" option in
rdirichlet_mat() has a new argument
output so that multiple object types can be returned.
The auxiliary argument
random_method = "sample" in
params_surv() is deprecated and
random_method = "discrete" should be used instead.
stateval_tbl now contains a
grp_var column used to assign state values to “groups” of patients. This is distinct from
hesim_data(), which is used to define groups for subgroup analyses.
The public field
input_mats has been renamed
R6 classes for disease progression and state values. This is a more generic name and will allow for potential feature enhancements in which
input_data is a data frame rather than a matrix.
rdirichlet_mat() has been modified to better facilitate sampling from transition matrices within the context of
tparams_transprobs(). One implication is that the number of rows in
alpha must now be less than or equal to the number of columns and that the number of columns can be greater than the number of rows.
Remove a documented
... that was not used in
input_mats class now contains an element
TRUE, then time intervals reset each time a patient enters a new health state. In other words, state values can depend on time since entering a health state.
To illustrate, consider an oncology application with three health states (stable disease, progressed disease, and death). In these models it is common to assume that patients begin second line treatment after disease progression. Suppose the second line treatment is a chemotherapy that patients take for 12 cycles (or approximately 1 year). Then drug costs would accrue for the first year but not afterwards.
State values like this can be specified by setting
time_reset = TRUE in
hesim_dat <- hesim_data(strategies = data.frame(strategy_id = c(1, 2)), patients = data.frame(patient_id = seq(1, 3)), states = data.frame(state_id = c(1, 2))) drugcosts <- stateval_tbl(tbl = data.frame(state_id = rep(c(1, 2), each = 2), time_start = c(0, 1, 0, 1), est = c(10000, 0, 12500, 0)), dist = "fixed", hesim_data = hesim_dat) drugcostsmod <- create_StateVals(drugcosts, time_reset = TRUE)
hesim now provides a general framework for integrating statistical models with economic evaluation. Users build a decision model by specifying a model structure, which consists of a set of statistical models for disease progression, utilities, and costs. Each statistical model is used to simulate outcomes as a function of estimated parameters and input data. N-state partitioned survival models (PSMs) and individual-level continuous time state transition models (iCTSTMs) are now supported.
Economic models—which combine the disease, utility, and cost models—are constructed with the following classes:
Disease models are constructed using the classes:
hesim::PsmCurvesto simulate survival curves for each endpoint of interest
hesim::IndivCtstmTransto simulate health state transitions with a iCTSTM
Utility and cost models are constructed with the
The economic models are used to simulate disease progression (
$sim_stateprobs()), quality-adjusted life-years (QALYs) (
$sim_qalys()), and costs (
$sim_costs()). Parameter uncertainty is propagated to model outcomes using probabilistic sensitivity analysis. Summaries of the simulated costs and QALYs are used to perform model-based cost-effectiveness analyses (CEAs) and represent decision uncertainty with
The initial CRAN submission containing support for CEA but not for model development. Decision uncertainty is represented using cost-effectiveness planes, cost-effectiveness acceptability curves, cost-effectiveness acceptability frontiers, and the expected value of perfect information. CEAs by subgroup (i.e., individualized CEAs) are performed with