This page explains the details of estimating weights from SuperLearner-based propensity scores by setting method = "super"
in the call to weightit()
or weightitMSM()
. This method can be used with binary, multi-category, and continuous treatments.
In general, this method relies on estimating propensity scores using the SuperLearner algorithm for stacking predictions and then converting those propensity scores into weights using a formula that depends on the desired estimand. For binary and multi-category treatments, one or more binary classification algorithms are used to estimate the propensity scores as the predicted probability of being in each treatment given the covariates. For continuous treatments, regression algorithms are used to estimate generalized propensity scores as the conditional density of treatment given the covariates. This method relies on SuperLearner::SuperLearner()
from the SuperLearner package.
Binary Treatments
For binary treatments, this method estimates the propensity scores using SuperLearner::SuperLearner()
. The following estimands are allowed: ATE, ATT, ATC, ATO, ATM, and ATOS. Weights can also be computed using marginal mean weighting through stratification for the ATE, ATT, and ATC. See get_w_from_ps()
for details.
Multi-Category Treatments
For multi-category treatments, the propensity scores are estimated using several calls to SuperLearner::SuperLearner()
, one for each treatment group; the treatment probabilities are not normalized to sum to 1. The following estimands are allowed: ATE, ATT, ATC, ATO, and ATM. The weights for each estimand are computed using the standard formulas or those mentioned above. Weights can also be computed using marginal mean weighting through stratification for the ATE, ATT, and ATC. See get_w_from_ps()
for details.
Continuous Treatments
For continuous treatments, the generalized propensity score is estimated using SuperLearner::SuperLearner()
. In addition, kernel density estimation can be used instead of assuming a normal density for the numerator and denominator of the generalized propensity score by setting use.kernel = TRUE
. Other arguments to density()
can be specified to refine the density estimation parameters. plot = TRUE
can be specified to plot the density for the numerator and denominator, which can be helpful in diagnosing extreme weights.
Longitudinal Treatments
For longitudinal treatments, the weights are the product of the weights estimated at each time point.
Missing Data
In the presence of missing data, the following value(s) for missing
are allowed:
"ind"
(default)First, for each variable with missingness, a new missingness indicator variable is created which takes the value 1 if the original covariate is
NA
and 0 otherwise. The missingness indicators are added to the model formula as main effects. The missing values in the covariates are then replaced with the covariate medians (this value is arbitrary and does not affect estimation). The weight estimation then proceeds with this new formula and set of covariates. The covariates output in the resultingweightit
object will be the original covariates with theNA
s.
Details
SuperLearner works by fitting several machine learning models to the treatment and covariates and then taking a weighted combination of the generated predicted values to use as the propensity scores, which are then used to construct weights. The machine learning models used are supplied using the SL.library
argument; the more models are supplied, the higher the chance of correctly modeling the propensity score. The predicted values are combined using the method supplied in the SL.method
argument (which is nonnegative least squares by default). A benefit of SuperLearner is that, asymptotically, it is guaranteed to perform as well as or better than the best-performing method included in the library. Using Balance SuperLearner by setting SL.method = "method.balance"
works by selecting the combination of predicted values that minimizes an imbalance measure.
Note
Some methods formerly available in SuperLearner are now in SuperLearnerExtra, which can be found on GitHub at https://github.com/ecpolley/SuperLearnerExtra.
The criterion
argument used to be called stop.method
, which is its name in twang. stop.method
still works for backward compatibility. Additionally, the criteria formerly named as es.mean
, es.max
, and es.rms
have been renamed to smd.mean
, smd.max
, and smd.rms
. The former are used in twang and will still work with weightit()
for backward compatibility.
Additional Arguments
discrete
if
TRUE
, uses discrete SuperLearner, which simply selects the best performing method. DefaultFALSE
, which finds the optimal combination of predictions for the libraries usingSL.method
.
An argument to SL.library
must be supplied. To see a list of available entries, use SuperLearner::listWrappers()
.
All arguments to SuperLearner::SuperLearner()
can be passed through weightit()
or weightitMSM()
, with the following exceptions:
obsWeights
is ignored because sampling weights are passed usings.weights
.method
inSuperLearner()
is replaced with the argumentSL.method
inweightit()
.
For continuous treatments only, the following arguments may be supplied:
density
A function corresponding to the conditional density of the treatment. The standardized residuals of the treatment model will be fed through this function to produce the numerator and denominator of the generalized propensity score weights. If blank,
dnorm()
is used as recommended by Robins et al. (2000). This can also be supplied as a string containing the name of the function to be called. If the string contains underscores, the call will be split by the underscores and the latter splits will be supplied as arguments to the second argument and beyond. For example, ifdensity = "dt_2"
is specified, the density used will be that of a t-distribution with 2 degrees of freedom. Using a t-distribution can be useful when extreme outcome values are observed (Naimi et al., 2014). Ignored ifuse.kernel = TRUE
(described below).use.kernel
If
TRUE
, uses kernel density estimation through thedensity()
function to estimate the numerator and denominator densities for the weights. IfFALSE
, the argument to thedensity
parameter is used instead.bw
,adjust
,kernel
,n
If
use.kernel = TRUE
, the arguments to thedensity()
function. The defaults are the same as those indensity
except thatn
is 10 times the number of units in the sample.plot
If
use.kernel = TRUE
, whether to plot the estimated density.
Balance SuperLearner
In addition to the methods allowed by SuperLearner()
, one can specify SL.method = "method.balance"
to use "Balance SuperLearner" as described by Pirracchio and Carone (2018), wherein covariate balance is used to choose the optimal combination of the predictions from the methods specified with SL.library
. Coefficients are chosen (one for each prediction method) so that the weights generated from the weighted combination of the predictions optimize a balance criterion, which must be set with the criterion
argument, described below.
criterion
A string describing the balance criterion used to select the best weights. See
cobalt::bal.compute()
for allowable options for each treatment type. For binary and multi-category treatments, the default is"smd.mean"
, which minimizes the average absolute standard mean difference among the covariates between treatment groups. For continuous treatments, the default is"p.mean"
, which minimizes the average absolute Pearson correlation between the treatment and covariates.
Note that this implementation differs from that of Pirracchio and Carone (2018) in that here, balance is measured only on the terms included in the model formula (i.e., and not their interactions unless specifically included), and balance results from a sample weighted using the estimated predicted values as propensity scores, not a sample matched using propensity score matching on the predicted values. Binary and continuous treatments are supported, but currently multi-category treatments are not.
Additional Outputs
info
For binary and continuous treatments, a list with two entries,
coef
andcvRisk
. For multi-category treatments, a list of lists with these two entries, one for each treatment level.coef
The coefficients in the linear combination of the predictions from each method in
SL.library
. Higher values indicate that the corresponding method plays a larger role in determining the resulting predicted value, and values close to zero indicate that the method plays little role in determining the predicted value. Whendiscrete = TRUE
, these correspond to the coefficients that would have been estimated haddiscrete
beenFALSE
.cvRisk
The cross-validation risk for each method in
SL.library
. Higher values indicate that the method has worse cross-validation accuracy. WhenSL.method = "method.balance"
, the sample weighted balance statistic requested withcriterion
. Higher values indicate worse balance.
obj
When
include.obj = TRUE
, the SuperLearner fit(s) used to generate the predicted values. For binary and continuous treatments, the output of the call toSuperLearner::SuperLearner()
. For multi-category treatments, a list of outputs to calls toSuperLearner::SuperLearner()
.
References
Binary treatments
Pirracchio, R., Petersen, M. L., & van der Laan, M. (2015). Improving Propensity Score Estimators’ Robustness to Model Misspecification Using Super Learner. American Journal of Epidemiology, 181(2), 108–119. doi:10.1093/aje/kwu253
Multi-Category Treatments
Imai, K., & Ratkovic, M. (2014). Covariate balancing propensity score. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 76(1), 243–263.
Continuous treatments
Kreif, N., Grieve, R., Díaz, I., & Harrison, D. (2015). Evaluation of the Effect of a Continuous Treatment: A Machine Learning Approach with an Application to Treatment for Traumatic Brain Injury. Health Economics, 24(9), 1213–1228. doi:10.1002/hec.3189
Balance SuperLearner (SL.method = "method.balance"
)
Pirracchio, R., & Carone, M. (2018). The Balance Super Learner: A robust adaptation of the Super Learner to improve estimation of the average treatment effect in the treated based on propensity score matching. Statistical Methods in Medical Research, 27(8), 2504–2518. doi:10.1177/0962280216682055
See method_glm
for additional references.
Examples
library("cobalt")
data("lalonde", package = "cobalt")
#Balancing covariates between treatment groups (binary)
(W1 <- weightit(treat ~ age + educ + married +
nodegree + re74, data = lalonde,
method = "super", estimand = "ATT",
SL.library = c("SL.glm", "SL.stepAIC",
"SL.glm.interaction")))
#> Loading required package: nnls
#> A weightit object
#> - method: "super" (propensity score weighting with SuperLearner)
#> - number of obs.: 614
#> - sampling weights: none
#> - treatment: 2-category
#> - estimand: ATT (focal: 1)
#> - covariates: age, educ, married, nodegree, re74
summary(W1)
#> Summary of weights
#>
#> - Weight ranges:
#>
#> Min Max
#> treated 1.0000 || 1.0000
#> control 0.0059 |---------------------------| 5.8129
#>
#> - Units with the 5 most extreme weights by group:
#>
#> 5 4 3 2 1
#> treated 1 1 1 1 1
#> 411 589 269 409 296
#> control 2.4388 2.528 2.7673 3.3921 5.8129
#>
#> - Weight statistics:
#>
#> Coef of Var MAD Entropy # Zeros
#> treated 0.00 0.00 0.000 0
#> control 1.13 0.74 0.446 0
#>
#> - Effective Sample Sizes:
#>
#> Control Treated
#> Unweighted 429. 185
#> Weighted 188.67 185
bal.tab(W1)
#> Balance Measures
#> Type Diff.Adj
#> prop.score Distance 0.0535
#> age Contin. -0.1475
#> educ Contin. 0.0161
#> married Binary -0.0049
#> nodegree Binary 0.0115
#> re74 Contin. -0.0247
#>
#> Effective sample sizes
#> Control Treated
#> Unadjusted 429. 185
#> Adjusted 188.67 185
# \donttest{
#Balancing covariates with respect to race (multi-category)
(W2 <- weightit(race ~ age + educ + married +
nodegree + re74, data = lalonde,
method = "super", estimand = "ATE",
SL.library = c("SL.glm", "SL.stepAIC",
"SL.glm.interaction")))
#> A weightit object
#> - method: "super" (propensity score weighting with SuperLearner)
#> - number of obs.: 614
#> - sampling weights: none
#> - treatment: 3-category (black, hispan, white)
#> - estimand: ATE
#> - covariates: age, educ, married, nodegree, re74
summary(W2)
#> Summary of weights
#>
#> - Weight ranges:
#>
#> Min Max
#> black 1.3125 |-------------------| 14.0815
#> hispan 1.7509 |--------------------------| 19.0788
#> white 1.0815 |----| 4.9344
#>
#> - Units with the 5 most extreme weights by group:
#>
#> 184 244 485 182 181
#> black 7.5917 7.7461 10.1133 11.6337 14.0815
#> 346 392 371 269 345
#> hispan 16.6957 17.1862 17.2636 17.8517 19.0788
#> 457 23 409 589 296
#> white 4.0945 4.1846 4.3505 4.6072 4.9344
#>
#> - Weight statistics:
#>
#> Coef of Var MAD Entropy # Zeros
#> black 0.640 0.391 0.138 0
#> hispan 0.476 0.372 0.110 0
#> white 0.388 0.317 0.069 0
#>
#> - Effective Sample Sizes:
#>
#> black hispan white
#> Unweighted 243. 72. 299.
#> Weighted 172.57 58.85 260.06
bal.tab(W2)
#> Balance summary across all treatment pairs
#> Type Max.Diff.Adj
#> age Contin. 0.0968
#> educ Contin. 0.0794
#> married Binary 0.0450
#> nodegree Binary 0.0246
#> re74 Contin. 0.0328
#>
#> Effective sample sizes
#> black hispan white
#> Unadjusted 243. 72. 299.
#> Adjusted 172.57 58.85 260.06
#Balancing covariates with respect to re75 (continuous)
#assuming t(8) conditional density for treatment
(W3 <- weightit(re75 ~ age + educ + married +
nodegree + re74, data = lalonde,
method = "super", density = "dt_8",
SL.library = c("SL.glm", "SL.ridge",
"SL.glm.interaction")))
#> A weightit object
#> - method: "super" (propensity score weighting with SuperLearner)
#> - number of obs.: 614
#> - sampling weights: none
#> - treatment: continuous
#> - covariates: age, educ, married, nodegree, re74
summary(W3)
#> Summary of weights
#>
#> - Weight ranges:
#>
#> Min Max
#> all 0.0439 |---------------------------| 21.0528
#>
#> - Units with the 5 most extreme weights:
#>
#> 431 483 484 485 354
#> all 9.4162 16.527 18.1028 19.0692 21.0528
#>
#> - Weight statistics:
#>
#> Coef of Var MAD Entropy # Zeros
#> all 1.334 0.504 0.341 0
#>
#> - Effective Sample Sizes:
#>
#> Total
#> Unweighted 614.
#> Weighted 221.04
bal.tab(W3)
#> Balance Measures
#> Type Corr.Adj
#> age Contin. 0.0311
#> educ Contin. 0.0348
#> married Binary 0.0613
#> nodegree Binary -0.0599
#> re74 Contin. 0.0405
#>
#> Effective sample sizes
#> Total
#> Unadjusted 614.
#> Adjusted 221.04
# }
#Balancing covariates between treatment groups (binary)
# using balance SuperLearner to minimize the maximum
# KS statistic
(W4 <- weightit(treat ~ age + educ + married +
nodegree + re74, data = lalonde,
method = "super", estimand = "ATT",
SL.library = c("SL.glm", "SL.stepAIC",
"SL.lda"),
SL.method = "method.balance",
criterion = "ks.max"))
#> A weightit object
#> - method: "super" (propensity score weighting with SuperLearner)
#> - number of obs.: 614
#> - sampling weights: none
#> - treatment: 2-category
#> - estimand: ATT (focal: 1)
#> - covariates: age, educ, married, nodegree, re74
summary(W4)
#> Summary of weights
#>
#> - Weight ranges:
#>
#> Min Max
#> treated 1.0000 || 1.0000
#> control 0.0222 |---------------------------| 2.0438
#>
#> - Units with the 5 most extreme weights by group:
#>
#> 5 4 3 2 1
#> treated 1 1 1 1 1
#> 411 595 269 409 296
#> control 1.3303 1.4365 1.5005 1.6369 2.0438
#>
#> - Weight statistics:
#>
#> Coef of Var MAD Entropy # Zeros
#> treated 0.000 0.000 0.00 0
#> control 0.823 0.701 0.33 0
#>
#> - Effective Sample Sizes:
#>
#> Control Treated
#> Unweighted 429. 185
#> Weighted 255.99 185
bal.tab(W4, stats = c("m", "ks"))
#> Balance Measures
#> Type Diff.Adj KS.Adj
#> prop.score Distance 0.0199 0.0944
#> age Contin. 0.0459 0.2764
#> educ Contin. -0.0360 0.0601
#> married Binary 0.0044 0.0044
#> nodegree Binary 0.0080 0.0080
#> re74 Contin. -0.0275 0.2839
#>
#> Effective sample sizes
#> Control Treated
#> Unadjusted 429. 185
#> Adjusted 255.99 185