| Title: | Diagnostic and Plotting Functions to Supplement 'bartCause' |
|---|---|
| Description: | Functions to assist in diagnostics and plotting during the causal inference modeling process. Supplements the 'bartCause' package. |
| Authors: | Joseph Marlo [aut, cre], George Perrett [aut] |
| Maintainer: | Joseph Marlo <[email protected]> |
| License: | MIT + file LICENSE |
| Version: | 0.1.30 |
| Built: | 2025-02-19 05:15:43 UTC |
| Source: | https://github.com/priism-center/plotbart |
Lalonde dataset
lalondelalonde
An object of class data.frame with 445 rows and 12 columns.
https://CRAN.R-project.org/package=arm
Visualize balance of variables between treatment and control groups. Balance plot reflects balance in standardized units.
plot_balance( .data, treatment, confounders, compare = c("means", "variance", "covariance"), estimand = c("ATE", "ATT", "ATC"), limit_continuous = NULL, limit_catagorical = NULL )plot_balance( .data, treatment, confounders, compare = c("means", "variance", "covariance"), estimand = c("ATE", "ATT", "ATC"), limit_continuous = NULL, limit_catagorical = NULL )
.data |
dataframe |
treatment |
the column denoted treatment. Must be binary. |
confounders |
character list of column names denoting the X columns of interest |
compare |
character of either means or variance denotes what to compare balance on |
estimand |
character of either ATE, ATT or ATC the causal estimand you are making inferences about |
limit_continuous |
integer that can be used to limit the plot to only show the limit_continuous most imbalanced variables |
limit_catagorical |
integer that can be used to limit the plot to only show the limit_categorical most imbalanced variables |
ggplot object
George Perrett & Joseph Marlo
data(lalonde) plot_balance(lalonde, 'treat', c('re78', 'age', 'educ'), compare = 'means', estimand = 'ATE') + labs(title = 'My new title')data(lalonde) plot_balance(lalonde, 'treat', c('re78', 'age', 'educ'), compare = 'means', estimand = 'ATE') + labs(title = 'My new title')
Plot the conditional average treatment effect (CATE) of a 'bartCause' model. The conditional average treatment effect is derived from taking the difference between predictions for each individual under the control condition and under the treatment condition averaged over the population. Means of the CATE distribution will resemble SATE and PATE but the CATE distribution accounts for more uncertainty than SATE and less uncertainty than PATE.
plot_CATE( .model, type = c("histogram", "density"), ci_80 = FALSE, ci_95 = FALSE, reference = NULL, .mean = FALSE, .median = FALSE )plot_CATE( .model, type = c("histogram", "density"), ci_80 = FALSE, ci_95 = FALSE, reference = NULL, .mean = FALSE, .median = FALSE )
.model |
a model produced by 'bartCause::bartc()' |
type |
histogram or density |
ci_80 |
TRUE/FALSE. Show the 80% credible interval? |
ci_95 |
TRUE/FALSE. Show the 95% credible interval? |
reference |
numeric. Show a vertical reference line at this value |
.mean |
TRUE/FALSE. Show the mean reference line |
.median |
TRUE/FALSE. Show the median reference line |
ggplot object
George Perrett, Joseph Marlo
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_CATE(model_results)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_CATE(model_results)
Plot common support based on the standard deviation rule, chi squared rule, or both.
plot_common_support( .model, .x = NULL, .y = NULL, rule = c("both", "sd", "chi") )plot_common_support( .model, .x = NULL, .y = NULL, rule = c("both", "sd", "chi") )
.model |
a model produced by 'bartCause::bartc()' |
.x |
a character string denoting which covariate to use a the x axis default is to use a regression tree to predict the variable with least common support. |
.y |
a character string denoting which covariate to use a the y axis default is the outcome variable y |
rule |
one of c('both', 'sd', 'chi') denoting which rule to use to identify lack of support |
Sufficient overlap/common support is an assumption of causal inference. BART models use the uncertainty of counter factual uncertainty. When the posterior distribution of an individual's counterfactual prediction extends beyond a specified cut-point, that point likely has insufficient common support. 'bartCause' model offer the option to automatically remove points without common support from analyses, however, this must be specified during model fitting. Cut-points are determined through one of two rules: the standard deviation (sd) or chi-squared (chi). Under the standard deviation rule, a point has weak common support if its posterior distribution of the counterfactual deviation is greater than the maximum posterior of the observed predictions with 1 standard deviation of the distribution of standard deviations for each individual's predicted outcome under the observed assignment. Under the chi-squared rule, a point is discarded if the variance between its counterfactual prediction over observed prediction are statistically different under a chi-squared distribution with 1 degree of freedom. For more details on discard rules see Hill and Su 2013.
When called this plot will show how many points would have been removed under the standard deviation and chi-squared rules. This plot should be used as a diagnostic for 'bartCause' models fit without a common-support rule.
ggplot object
George Perrett, Joseph Marlo
Hill, J., & Su, Y. S. (2013). Assessing lack of common support in causal inference using Bayesian nonparametrics: Implications for evaluating the effect of breastfeeding on children's cognitive outcomes. The Annals of Applied Statistics, 1386-1420.
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_common_support(model_results) plot_common_support(model_results)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_common_support(model_results) plot_common_support(model_results)
Visualize balance of the covariance of variables between treatment and control groups. Balance plot reflects balance in standardized units.
plot_covariance(.data, treatment, confounders)plot_covariance(.data, treatment, confounders)
.data |
dataframe |
treatment |
the column denoted treatment. Must be binary. |
confounders |
character list of column names denoting the X columns of interest |
ggplot object
George Perrett
data(lalonde) plot_covariance(lalonde, 'treat', c('re75','re74' , 'age', 'educ')) + labs(title = 'My new title')data(lalonde) plot_covariance(lalonde, 'treat', c('re75','re74' , 'age', 'educ')) + labs(title = 'My new title')
Plots a histogram of Individual Conditional Average Treatment effects (ICATE). ICATEs are the difference in each individual's predicted outcome under the treatment and predicted outcome under the control averaged over the individual. Plots of ICATEs are useful to identify potential heterogeneous treatment effects between different individuals. ICATE plots can be grouped by discrete variables.
plot_ICATE(.model, .group_by = NULL, n_bins = 30, .alpha = 0.7)plot_ICATE(.model, .group_by = NULL, n_bins = 30, .alpha = 0.7)
.model |
a model produced by 'bartCause::bartc()' |
.group_by |
a grouping variable as a vector |
n_bins |
number of bins |
.alpha |
transparency of histograms |
ggplot object
George Perrett
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_ICATE(model_results, lalonde$married)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_ICATE(model_results, lalonde$married)
Use a regression tree to optimally bin a continous variable
plot_moderator_c_bin( .model, moderator, type = c("density", "histogram", "errorbar"), .alpha = 0.7, facet = FALSE, .ncol = 1, .name = "bin" )plot_moderator_c_bin( .model, moderator, type = c("density", "histogram", "errorbar"), .alpha = 0.7, facet = FALSE, .ncol = 1, .name = "bin" )
.model |
a model produced by 'bartCause::bartc()' |
moderator |
the moderator as a vector |
type |
string to specify if you would like to plot a histogram, density or error bar plot |
.alpha |
transparency value [0, 1] |
facet |
TRUE/FALSE. Create panel plots of each moderator level? |
.ncol |
number of columns to use when faceting |
.name |
sting representing the name of the moderating variable |
ggplot object
George Perrett
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_moderator_c_bin(model_results, lalonde$age, .name = 'age')data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_moderator_c_bin(model_results, lalonde$age, .name = 'age')
Plot the LOESS prediction of ICATEs by a continuous covariate. This is an alternative to partial dependency plots to assess treatment effect heterogeneity by a continuous covariate. See Carnegie, Dorie and Hill 2019.
plot_moderator_c_loess(.model, moderator, line_color = "blue")plot_moderator_c_loess(.model, moderator, line_color = "blue")
.model |
a model produced by 'bartCause::bartc()' |
moderator |
the moderator as a vector |
line_color |
the color of the loess line |
ggplot object
George Perrett, Joseph Marlo
Carnegie, N., Dorie, V., & Hill, J. L. (2019). Examining treatment effect heterogeneity using BART. Observational Studies, 5(2), 52-70.
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_moderator_c_loess(model_results, lalonde$age)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_moderator_c_loess(model_results, lalonde$age)
Plot a partial dependency plot with a continuous covariate from a 'bartCause' model. Identify treatment effect variation predicted across levels of a continuous variable.
plot_moderator_c_pd(.model, moderator, n_bins = NULL)plot_moderator_c_pd(.model, moderator, n_bins = NULL)
.model |
a model produced by 'bartCause::bartc()' |
moderator |
the moderator as a vector |
n_bins |
number of bins to cut the moderator with. Defaults to the lesser of 15 and number of distinct levels of the moderator |
Partial dependency plots are one way to evaluate heterogeneous treatment effects that vary by values of a continuous covariate. For more information on partial dependency plots from BART causal inference models see Green and Kern 2012.
ggplot object
George Perrett, Joseph Marlo
Green, D. P., & Kern, H. L. (2012). Modeling heterogeneous treatment effects in survey experiments with Bayesian additive regression trees. Public opinion quarterly, 76(3), 491-511.
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none', keepTrees = TRUE ) plot_moderator_c_pd(model_results, lalonde$age)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none', keepTrees = TRUE ) plot_moderator_c_pd(model_results, lalonde$age)
Plot the Conditional Average Treatment Effect split by a discrete moderating variable. This plot will provide a visual test of moderation by discrete variables.
plot_moderator_d( .model, moderator, type = c("density", "histogram", "errorbar"), .alpha = 0.7, facet = FALSE, .ncol = 1 )plot_moderator_d( .model, moderator, type = c("density", "histogram", "errorbar"), .alpha = 0.7, facet = FALSE, .ncol = 1 )
.model |
a model produced by 'bartCause::bartc()' |
moderator |
the moderator as a vector |
type |
string to specify if you would like to plot a histogram, density or error bar plot |
.alpha |
transparency value [0, 1] |
facet |
TRUE/FALSE. Create panel plots of each moderator level? |
.ncol |
number of columns to use when faceting |
ggplot object
George Perrett
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_moderator_d(model_results, lalonde$educ)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_moderator_d(model_results, lalonde$educ)
Plots the range of the Conditional Average Treatment Effect grouped by a discrete variable. This is analogous to plot_moderator_d_density but is preferable for moderators with many categories. Rather than plotting the full density, the posterior range is shown.
plot_moderator_d_linerange(.model, moderator, .alpha = 0.7, horizontal = FALSE)plot_moderator_d_linerange(.model, moderator, .alpha = 0.7, horizontal = FALSE)
.model |
a model produced by 'bartCause::bartc()' |
moderator |
the moderator as a vector |
.alpha |
transparency value [0, 1] |
horizontal |
flip the plot horizontal? |
ggplot object
George Perrett, Joseph Marlo
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_moderator_d_linerange(model_results, lalonde$educ)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_moderator_d_linerange(model_results, lalonde$educ)
Plot a single regression tree for exploratory heterogeneous effects. Fit single regression tree on bartc() ICATEs to produce variable importance plot. This plot is useful for identifying potential moderating variables. Tree depth may be set to depths 1, 2 or 3. Terminal nodes signal the Conditional Average Treatment effect within levels of moderation variables. Trees with different values across terminal nodes suggest strong treatment effect moderation.
plot_moderator_search(.model, max_depth = c(2, 1, 3))plot_moderator_search(.model, max_depth = c(2, 1, 3))
.model |
a model produced by 'bartCause::bartc()' |
max_depth |
one of c(1, 2, 3). Maximum number of node levels within the tree. 2 is recommended |
ggplot object
George Perrett, Joseph Marlo
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_moderator_search(model_results)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_moderator_search(model_results)
Plot histograms showing the overlap between propensity scores by treatment status.
plot_overlap_pScores( .data, treatment, confounders, plot_type = c("histogram", "density"), trim = TRUE, min_x = NULL, max_x = NULL, pscores = NULL, ... )plot_overlap_pScores( .data, treatment, confounders, plot_type = c("histogram", "density"), trim = TRUE, min_x = NULL, max_x = NULL, pscores = NULL, ... )
.data |
dataframe |
treatment |
character. Name of the treatment column within .data |
confounders |
character list of column names denoting confounders within .data |
plot_type |
the plot type, one of c('Histogram', 'Density') |
trim |
a logical if set to true y axis will be trimmed to better visualize areas of overlap |
min_x |
numeric value specifying the minimum propensity score value to be shown on the x axis |
max_x |
numeric value specifying the maximum propensity score value to be shown on the x axis |
pscores |
propensity scores. If not provided, then propensity scores will be calculated using BART |
... |
additional arguments passed to 'dbarts::bart2' propensity score calculation |
ggplot object
George Perrett, Joseph Marlo
data(lalonde) plot_overlap_pScores( .data = lalonde, treatment = 'treat', confounders = c('age', 'educ'), plot_type = 'histogram', pscores = NULL, seed = 44 )data(lalonde) plot_overlap_pScores( .data = lalonde, treatment = 'treat', confounders = c('age', 'educ'), plot_type = 'histogram', pscores = NULL, seed = 44 )
Plot histograms showing the overlap between variables by treatment status.
plot_overlap_vars( .data, treatment, confounders, plot_type = c("histogram", "density"), min_x = NULL, max_x = NULL )plot_overlap_vars( .data, treatment, confounders, plot_type = c("histogram", "density"), min_x = NULL, max_x = NULL )
.data |
dataframe |
treatment |
character. Name of the treatment column within .data |
confounders |
character list of column names denoting confounders within .data |
plot_type |
the plot type, one of c('histogram', 'density'). Defaults to 'histogram' |
min_x |
numeric value specifying the minimum value to be shown on the x axis |
max_x |
numeric value specifying the maximum value to be shown on the x axis |
ggplot object
George Perrett, Joseph Marlo
data(lalonde) plot_overlap_vars( .data = lalonde, treatment = 'treat', confounders = c('age', 'educ'), plot_type = 'Histogram' )data(lalonde) plot_overlap_vars( .data = lalonde, treatment = 'treat', confounders = c('age', 'educ'), plot_type = 'Histogram' )
Plot shows the Population Average Treatment Effect which is derived from the posterior predictive distribution of the difference between and .
Mean of PATE will resemble CATE and SATE but PATE will account for more uncertainty and is recommended for informing inferences on the average treatment effect.
plot_PATE( .model, type = c("histogram", "density"), ci_80 = FALSE, ci_95 = FALSE, reference = NULL, .mean = FALSE, .median = FALSE )plot_PATE( .model, type = c("histogram", "density"), ci_80 = FALSE, ci_95 = FALSE, reference = NULL, .mean = FALSE, .median = FALSE )
.model |
a model produced by 'bartCause::bartc()' |
type |
histogram or density |
ci_80 |
TRUE/FALSE. Show the 80% credible interval? |
ci_95 |
TRUE/FALSE. Show the 95% credible interval? |
reference |
numeric. Show a vertical reference line at this value |
.mean |
TRUE/FALSE. Show the mean reference line |
.median |
TRUE/FALSE. Show the median reference line |
ggplot object
George Perrett, Joseph Marlo
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_PATE(model_results)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_PATE(model_results)
Identify variables that predict lack of overlap
plot_predicted_common_support( .model, max_depth = 3, rule = c("both", "sd", "chi") )plot_predicted_common_support( .model, max_depth = 3, rule = c("both", "sd", "chi") )
.model |
a model produced by 'bartCause::bartc()' |
max_depth |
a number indicatin the max depth of the tree. Higher numbers are more prone to overfitting. |
rule |
one of c('both', 'sd', 'chi') denoting which rule to use to identify lack of support |
ggplot object
George Perrett, Joseph Marlo
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_predicted_common_support (model_results) plot_predicted_common_support (model_results, max_depth = 2, rule = 'chi')data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_predicted_common_support (model_results) plot_predicted_common_support (model_results, max_depth = 2, rule = 'chi')
Visualize balance of variables between treatment and control groups. Balance plot reflects balance in standardized units.
plot_residual_density(.model)plot_residual_density(.model)
.model |
a model produced by 'bartCause::bartc()' |
ggplot object
George Perrett & Joseph Marlo
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_residual_density(model_results)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_residual_density(model_results)
Plot a histogram or density of the Sample Average Treatment Effect (SATE). The Sample Average Treatment Effect is derived from taking the difference of each individual's observed outcome and a predicted counterfactual outcome from a BART model averaged over the population. The mean of SATE will resemble means of CATE and PATE but will account for the least uncertainty.
plot_SATE( .model, type = c("histogram", "density"), ci_80 = FALSE, ci_95 = FALSE, reference = NULL, .mean = FALSE, .median = FALSE, check_overlap = FALSE, overlap_rule = c("none", "sd", "chisq") )plot_SATE( .model, type = c("histogram", "density"), ci_80 = FALSE, ci_95 = FALSE, reference = NULL, .mean = FALSE, .median = FALSE, check_overlap = FALSE, overlap_rule = c("none", "sd", "chisq") )
.model |
a model produced by 'bartCause::bartc()' |
type |
histogram or density |
ci_80 |
TRUE/FALSE. Show the 80% credible interval? |
ci_95 |
TRUE/FALSE. Show the 95% credible interval? |
reference |
numeric. Show a vertical reference line at this x-axis value |
.mean |
TRUE/FALSE. Show the mean reference line |
.median |
TRUE/FALSE. Show the median reference line |
check_overlap |
TRUE/FALSE. Check if any overlap rules are applicable |
overlap_rule |
enter overlap rules to view how different bartCause removal rules would have influenced results. Only applicable if check_overlap is TRUE. |
ggplot object
George Perrett, Joseph Marlo
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_SATE(model_results)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_SATE(model_results)
Returns a ggplot of the estimated effect over each iteration of the model fit. This is used to visually assess the convergence of Markov chain Monte Carlo (MCMC) sampling. Chains should be well mixed such that no single color is notably separate from others.
plot_trace(.model, type = c("cate", "sate", "pate", "sigma"))plot_trace(.model, type = c("cate", "sate", "pate", "sigma"))
.model |
a model produced by 'bartCause::bartc()' |
type |
parameter to plot options are average treatment effects: 'cate', 'sate' and 'pate' as well as posterior predicitve uncertainty 'sigma' |
ggplot object
Joseph Marlo, George Perrett
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_trace(.model = model_results)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_trace(.model = model_results)
Trace plots may occlude convegence issues within Markov Chains. Trace rank plots present an alternartive convergence diagnostic of MCMC convergence. Trank plots are described in detail in Vehtari et al. (2021).
plot_trank(.model, type = c("cate", "sate", "pate", "sigma"))plot_trank(.model, type = c("cate", "sate", "pate", "sigma"))
.model |
a model produced by 'bartCause::bartc()' |
type |
parameter to plot options are average treatment effects: 'cate', 'sate' and 'pate' as well as posterior predicitve uncertainty 'sigma' |
ggplot object
George Perrett
Vehtari, A., Gelman, A., Simpson, D., Carpenter, B., & Bürkner, P. C. (2021). Rank-normalization, folding, and localization: An improved R ̂ for assessing convergence of MCMC (with discussion). Bayesian analysis, 16(2), 667-718.
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_trank(.model = model_results)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSup.rule = 'none' ) plot_trank(.model = model_results)
Plots the point and posterior intervals of each individual's ICATE ordered by the ICATE or a continuous variable. Points can be colored by a discrete variable. Waterfall plots are a useful visual diagnostic of possible treatment effect heterogeneity. A flat line implies little treatment effect heterogeneity while a steeper curve implies that the treatment effect varies across individuals in the sample. Ordering points by a continuous variable or coloring points by a discrete variable can be helpful to identify potential moderators of the treatment effect.
plot_waterfall( .model, descending = TRUE, .order = NULL, .color = NULL, .alpha = 0.5 )plot_waterfall( .model, descending = TRUE, .order = NULL, .color = NULL, .alpha = 0.5 )
.model |
a model produced by 'bartCause::bartc()' |
descending |
order the ICATEs by value? |
.order |
a vector representing a custom order |
.color |
a vector representing colors |
.alpha |
transparency value [0, 1] |
ggplot object
George Perrett
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_waterfall(model_results)data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) plot_waterfall(model_results)
See balance statisitics of variables between treatment and control groups.
print_balance(.data, treatment, confounders, estimand = c("ATE", "ATT", "ATC"))print_balance(.data, treatment, confounders, estimand = c("ATE", "ATT", "ATC"))
.data |
dataframe |
treatment |
the column denoted treatment. Must be binary. |
confounders |
character list of column names denoting the X columns of interest |
estimand |
character of either ATE, ATT or ATC the causal estimand you are making inferences about |
tibble
George Perrett
data(lalonde) print_balance(lalonde, 'treat', confounders = c('re78', 'age', 'educ'), estimand = 'ATE')data(lalonde) print_balance(lalonde, 'treat', confounders = c('re78', 'age', 'educ'), estimand = 'ATE')
See balance statistics of covariance for specified variables between treatment and control groups.
print_covariance( .data, treatment, confounders, estimand = c("ATE", "ATT", "ATC") )print_covariance( .data, treatment, confounders, estimand = c("ATE", "ATT", "ATC") )
.data |
dataframe |
treatment |
the column denoted treatment. Must be binary. |
confounders |
character list of column names denoting the X columns of interest |
estimand |
character of either ATE, ATT or ATC the causal estimand you are making inferences about |
tibble
George Perrett
data(lalonde) print_covariance(lalonde, 'treat', confounders = c('re78', 'age', 'educ'), estimand = 'ATE')data(lalonde) print_covariance(lalonde, 'treat', confounders = c('re78', 'age', 'educ'), estimand = 'ATE')
Use a regression tree to optimally bin a continuous variable, this function will print out a table with estimates and 95
table_moderator_c_bin(.model, moderator, .name = "bin")table_moderator_c_bin(.model, moderator, .name = "bin")
.model |
a model produced by 'bartCause::bartc()' |
moderator |
the moderator as a vector |
.name |
sting representing the name of the moderating variable |
a data.frame object
George Perrett
data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) table_moderator_c_bin(model_results, lalonde$age, .name = 'age')data(lalonde) confounders <- c('age', 'educ', 'black', 'hisp', 'married', 'nodegr') model_results <- bartCause::bartc( response = lalonde[['re78']], treatment = lalonde[['treat']], confounders = as.matrix(lalonde[, confounders]), estimand = 'ate', commonSuprule = 'none' ) table_moderator_c_bin(model_results, lalonde$age, .name = 'age')