Accessing causal-model objects via get_ methods e.g. get_nodal_types(), get_parameters is no longer supported. Objects may now be accessed via a unified syntax through the grab() function (see New Functionality). The following functions are no longer exported:
get_causal_types()get_nodal_types()get_all_data_types()get_event_probabilities()get_ambiguities_matrix()get_parameters()get_parameter_names()get_parmap()get_parameter_matrix()get_priors()get_param_dist()get_type_prob_multiple()grab()causal-model objects can now be accessed via grab() like so:
grab(model, "parameters_df")See documentation for an exhaustive list of accessible objects. causal-model objects now additionally come with dedicated print methods returning short informative summaries of the given object.
A summary of parameter values and convergence information produced by the update_model() Stan model can now be accessed via:
grab(model, "stan_summary")Advanced model diagnostics on raw Stan output via external packages is possible by saving the stan_fit object when updating. This is facilitated via the keep_fit option in update_model():
model <- make_model("X -> Y") |>
update_model(data, keep_fit = TRUE)
model |> grab("stan_fit")nodal_types to make_model() now implements correct error handlingPreviously this make_model("X -> Y" , nodal_types = list(Y = c("0", "1"))) was permissible leading to setting nodal_types:
$X
NULL
$Y
[1] "0" "1"This led to undefined behavior and unhelpful downstream error messages. When passing nodal_types to make_model() users are now forced to specify a set of nodal_types on each node.
query_distribution() are no longer overwrites type distribution internallymake_model()Previously hyphenated names would not throw an error and be corrupted silently through the conversion of model definition strings into dagitty objects.
make_model("institutions -> political-inequality")
Statement:
[1] "institutions -> political-inequality"
DAG:
parent children
1 institutions politicalChecks for correct variable naming are now reinstated.
Calls to sapply() have ben replaced with vapply() wherever possible to enforce type safety.
Looping via index has been replaced by range based looping wherever possible to guard against 0 length exceptions.
goodpractice::gp()goodpractice code improvements have been implemented.
query_distribution() now supports the use of multiple queries in one function call and thus returns a DataFrame of distribution draws instead of a single numeric vector.
query_distribution(): now supports the specification of multiple queries and givens to be evaluated on a single model in one function call.
model <- make_model("X -> Y")
query_distribution(model,
query = list("(Y[X=1] > Y[X=0])", "(Y[X=1] < Y[X=0])"),
given = list("Y==1", "(Y[X=1] <= Y[X=0])"),
using = "priors")|>
head()query_model(): now supports the specification of multiple models to evaluate a set of queries on in one function call.
models <- list(
M1 = make_model("X -> Y"),
M2 = make_model("X -> Y") |> set_restrictions("Y[X=1] < Y[X=0]")
)
query_model(
models,
query = list(ATE = "Y[X=1] - Y[X=0]", Share_positive = "Y[X=1] > Y[X=0]"),
given = c(TRUE, "Y==1 & X==1"),
using = c("parameters", "priors"),
expand_grid = FALSE)
query_model(
models,
query = list(ATE = "Y[X=1] - Y[X=0]", Share_positive = "Y[X=1] > Y[X=0]"),
given = c(TRUE, "Y==1 & X==1"),
using = c("parameters", "priors"),
expand_grid = TRUE)This eliminates the need for redundant function calls when querying models and substantially improves computation time as computationally expensive function calls to produce data structures required for querying are now reduced to a minimum via redundancy elimination and caching.
realise_outcomes(): specifying the node option now produces a DataFrame detailing how the specified node responds to its parents in the presence or absence of do operations. This produces a reduced form of the usual realise_outcomes() output detailing all causal-types; and aids in the interpretation of both nodal- and causal-types. This update resolves previous bugs and errors relating to specification of nodes with multiple parents in the node option.
model <- make_model("X1 -> M -> Y -> Z; X2 -> Y") |>
realise_outcomes(dos = list(M = 1), node = "Y") Previously set_parameters() and set_priors() would default applying changes in the order in which parameters appeared in the parameters_df DataFrame; regardless of the order in which changes were specified in the aforementioned functions. Calling:
model <- make_model("X -> Y")
set_priors(model, alphas = c(3,4), nodal_type = c("10",00))would results in the following parameters_df.
param_names node gen param_set nodal_type given param_value priors
<chr> <chr> <int> <chr> <chr> <chr> <dbl> <dbl>
1 X.0 X 1 X 0 "" 0.5 1
2 X.1 X 1 X 1 "" 0.5 1
3 Y.00 Y 2 Y 00 "" 0.25 3
4 Y.10 Y 2 Y 10 "" 0.25 4
5 Y.01 Y 2 Y 01 "" 0.25 1
6 Y.11 Y 2 Y 11 "" 0.25 1Now changes to parameters values get applied in the order specified in the function call; resulting in the following parameters_df for the above example:
param_names node gen param_set nodal_type given param_value priors
<chr> <chr> <int> <chr> <chr> <chr> <dbl> <dbl>
1 X.0 X 1 X 0 "" 0.5 1
2 X.1 X 1 X 1 "" 0.5 1
3 Y.00 Y 2 Y 00 "" 0.25 4
4 Y.10 Y 2 Y 10 "" 0.25 3
5 Y.01 Y 2 Y 01 "" 0.25 1
6 Y.11 Y 2 Y 11 "" 0.25 1Additionally we have implemented helpful warnings for when instructions identifying parameters to be updated are under specified. This is particularly useful when setting priors or parameters on models with confounding as changes may inadvertently be applied across param_sets.
Previously updating models with censored types would fail as 0s in the w vector induced by censoring would evaluate to -Inf as the Stan MCMC algorithm began sampling from the posterior of the multinational distribution. We resolved this issue by pruning the w vector when the multinomial is run. This preserves the true w vector (event probabilities without censoring) while still updating with the censored data-
Previously wildcards in set_restrictions() were erroneously interpreted as valid nodal types, leading to errors and undefined behavior. Proper unpacking and mapping of wildcards to existing nodal types has been restored.
Previously misspecifications in queries like Y[X==1]=1 would lead to undefined behavior when mapping queries to nodal or causal types. We now correct misspecified queries internally and warn about the misspecification. For example; running:
model <- CausalQueries::make_model("X -> Y")
get_query_types(model, "Y[X=1]=1")now produces
Causal types satisfying query's condition(s)
query = Y[X=1]==1
X0.Y01 X1.Y01
X0.Y11 X1.Y11
Number of causal types that meet condition(s) = 4
Total number of causal types in model = 8
Warning message:
In check_query(query) :
statements to the effect that the realization of a node should equal some value should be specified with `==` not `=`.
The query has been changed accordingly: Y[X=1]==1
Previously a parameter matrix P that was attached to a causal_model object could not be overwritten. Overwrites are now possible.
realise_outcomes()We achieved a ~100 fold speed gain in the realise_outcomes() functionality. Nodal types on a given node are generated as the Cartesian product of parent realizations. Consider the meaning of nodal types on a node \(Y\) with 3 parents \([X1,X2,X3]\):
| X1 | X2 | X3 |
|---|---|---|
| 0 | 0 | 0 |
| 1 | 0 | 0 |
| 0 | 1 | 0 |
| 1 | 1 | 0 |
| 0 | 0 | 1 |
| 1 | 0 | 1 |
| 0 | 1 | 1 |
| 1 | 1 | 1 |
Each row in the above DataFrame corresponds to a digit in Y's nodal types. The first digit of each nodal type of \(Y\) (see first row above), corresponds to the realization of \(Y\) when \(X1 = 0, X2 = 0, X3 = 0\). The fourth digit of each nodal type of \(Y\) (see fourth row above), corresponds to the realization of \(Y\) when \(X1 = 1, X2 = 1, X3 = 0\). Finding the position of the realization value of \(Y\) in a nodal type given parent realizations is equivalent to finding the row number in the Cartesian product DataFrame. By definition of the Cartesian product, the number of consecutive 0 or 1 elements in a given column is \(2^{columnindex}\), when indexing columns from 0. Given a set of parent realizations \(R\) indexed from 0, the corresponding row in a number in a DataFrame indexed from 0 can thus be computed via: \[row = (\sum_{i = 0}^{|R| - 1} (2^{i} \times R_i))\]. We implement a fast C++ version of this computing powers of 2 via bit shifting.
Stan updateWe updated to the new array syntax introduced in Stan v2.33.0