Package {oalasso}

oalasso: Outcome-Adaptive Lasso Propensity Scores

Description

Estimates propensity scores by the outcome-adaptive lasso (OAL) of Shortreed and Ertefaie (2017) and the generalized outcome-adaptive lasso (GOAL) of Balde, Yang and Lefebvre (2023), using glmnet with an exact penalty-scale correction so that the published objectives and tuning grids are reproduced, and tuning by the papers' weighted absolute mean difference (wAMD) balance criterion. The resulting score is designed to be supplied directly to MatchIt::matchit() as a distance measure, to WeightIt::weightit() as a propensity score, or to psAve::psave() as an appended candidate.

Details

The single estimation function is oal(). Its result hands off to the existing ecosystem: oal_match() / oal_weight() (or the equivalent explicit MatchIt::matchit() / WeightIt::weightit() calls) and cobalt::bal.tab() (which has a method for oal objects). oal_wamd() exposes the selection criterion for methods research and testing.

Design notes

glmnet sits in Imports – the one principled departure from the suite's engines-in-Suggests rule – because it is the sole solver of the estimator's objective, and the exact-scale correction depends on its penalty-factor rescaling, its infinite-penalty-to-exclusion conversion, and its coef(..., exact = TRUE) re-supply semantics (floor ⁠>= 4.1-2⁠). The archived lqa package of the original reference implementation is not a dependency. cobalt powers the display balance table and bal.tab.oal(); the wAMD criterion itself is native code.

Note the name collision: the GOAL implemented here (method = "goal") is the generalized outcome-adaptive lasso of Balde, Yang and Lefebvre (2023) for binary treatments, not the unrelated "GOAL" generalized propensity score software for continuous exposures by Gao and colleagues.

Author(s)

Maintainer: Daijiro Kabata daijiro.kabata@port.kobe-u.ac.jp [copyright holder]

Authors:

• Daijiro Kabata daijiro.kabata@port.kobe-u.ac.jp [copyright holder]

References

Shortreed SM, Ertefaie A (2017). Outcome-adaptive lasso: variable selection for causal inference. Biometrics, 73(4), 1111-1122. doi:10.1111/biom.12679

Balde I, Yang Y, Lefebvre G (2023). Reader reaction to "Outcome-adaptive lasso: variable selection for causal inference" by Shortreed and Ertefaie (2017). Biometrics, 79(1), 514-520. doi:10.1111/biom.13683

Jones B, Ertefaie A, Shortreed SM (2023). Rejoinder to reader reaction "On the use of the outcome-adaptive lasso for propensity score estimation". Biometrics, 79(1), 521-525. doi:10.1111/biom.13681

Balde I (2025). Oracle properties of the generalized outcome-adaptive lasso. Statistics & Probability Letters. doi:10.1016/j.spl.2025.110379

Zou H (2006). The adaptive lasso and its oracle properties. Journal of the American Statistical Association, 101(476), 1418-1429. doi:10.1198/016214506000000735

Brookhart MA, Schneeweiss S, Rothman KJ, Glynn RJ, Avorn J, Sturmer T (2006). Variable selection for propensity score models. American Journal of Epidemiology, 163(12), 1149-1156. doi:10.1093/aje/kwj149

Myers JA, Rassen JA, Gagne JJ, Huybrechts KF, Schneeweiss S, Rothman KJ, Joffe MM, Glynn RJ (2011). Effects of adjusting for instrumental variables on bias and precision of effect estimates. American Journal of Epidemiology, 174(11), 1213-1222. doi:10.1093/aje/kwr364

See Also

Useful links:

• https://github.com/kabajiro/oalasso

• Report bugs at https://github.com/kabajiro/oalasso/issues

Balance tables for oal objects

Description

A method for cobalt::bal.tab(): assesses balance on the (original-scale) covariates of an oal() fit under the inverse-probability weights implied by the outcome-adaptive lasso propensity score, which is supplied as a distance measure.

Usage

## S3 method for class 'oal'
bal.tab(x, ...)

Arguments

Details

The call delegates to the default cobalt machinery as ⁠cobalt::bal.tab(<covariates>, treat = x$treat, weights = x$weights, s.d.denom = <by estimand>, distance = data.frame(ps = x$ps), ...)⁠, so all the usual cobalt arguments (un, stats, thresholds, ...) are available and display conventions are cobalt's own. The selection criterion inside oal() is the papers' wAMD instead, computed natively on the standardized covariates; see oal_wamd().

Value

A bal.tab object; see cobalt::bal.tab().

References

Shortreed SM, Ertefaie A (2017). Outcome-adaptive lasso: variable selection for causal inference. Biometrics, 73(4), 1111-1122. doi:10.1111/biom.12679

See Also

oal(), cobalt::bal.tab(), plot.oal()

Examples

data("lalonde", package = "MatchIt")
fit <- oal(treat ~ age + educ + married + re74, data = lalonde,
           outcome = ~ re78)
cobalt::bal.tab(fit, un = TRUE)

Extract propensity score model coefficients

Description

Returns the selected outcome-adaptive lasso propensity score coefficients (intercept first), either on the original covariate scale (default; the scale used by predict.oal()) or on the standardized scale that the penalized objective was actually optimized on (the scale of the adaptive weights and of the wAMD – methods-paper material).

Usage

## S3 method for class 'oal'
coef(object, scale = c("original", "standardized"), ...)

Arguments

Value

A named numeric vector of length p + 1.

See Also

oal(), predict.oal()

Outcome-adaptive lasso propensity scores

Description

oal() estimates a propensity score by the outcome-adaptive lasso (OAL) of Shortreed and Ertefaie (2017) or its generalized elastic-net form (GOAL) of Balde, Yang and Lefebvre (2023): an adaptive lasso on the treatment (propensity score) log-likelihood whose penalty weights come from an outcome regression, so that covariates unrelated to the outcome – instruments and noise variables – are excluded from the propensity score model. Tuning parameters are selected by the papers' weighted absolute mean difference (wAMD) balance criterion. The result is deliberately modest: a numeric score vector designed to be handed to MatchIt::matchit() as distance, to WeightIt::weightit() as ps, or to psAve::psave() as an appended candidate.

Usage

oal(
  formula,
  data,
  outcome,
  method = c("oal", "goal"),
  estimand = c("ATE", "ATT"),
  family = gaussian(),
  lambda = c(-10, -5, -2, -1, -0.75, -0.5, -0.25, 0.25, 0.49),
  gamma = NULL,
  gamma.factor = 2,
  lambda2 = c(0, 10^c(-2, -1.5, -1, -0.75, -0.5, -0.25, 0, 0.25, 0.5, 1)),
  outcome.coef = NULL,
  refit = FALSE,
  clip = c(0.01, 0.99),
  keep.path = TRUE,
  keep.fits = FALSE,
  verbose = FALSE,
  ...
)

Arguments

Details

Objective and exact glmnet evaluation

OAL solves, on the TOTAL log-likelihood scale,

\hat\alpha = \arg\min_\alpha\; -\ell(\alpha; A, X_s) + \lambda_n \sum_j |b_j|^{-\gamma} |\alpha_j|,

where \ell is the binomial log-likelihood of treatment on the standardized covariates and b_j are the outcome-model coefficients. glmnet minimizes the per-observation loss and internally rescales penalty factors to sum to the number of variables, so each grid point is evaluated exactly at

s = \overline{\mathrm{pen}} \cdot \lambda_n / n, \qquad \mathrm{pen}_j = |b_j|^{-\gamma},

via coef(fit, s = s, exact = TRUE, x = , y = , penalty.factor = ), which re-solves at that penalty (the KKT-verified recipe of the Shortreed/Ertefaie rejoinder code). When outcome.coef contains zeros, some \mathrm{pen}_j = \infty: glmnet converts those to exclusions and internally resets their factor to 1 before the sum-to-nvars rescale, so the constant generalizes to s = (\sum_{j: \mathrm{finite}} \mathrm{pen}_j + \#\{\mathrm{pen}_j = \infty\})/p \cdot \lambda_n / n – identical to \overline{\mathrm{pen}}\,\lambda_n/n when all factors are finite. This correction is always on; it is what makes the n^\delta grid carry its published meaning.

Standardization protocol (fixed, not an argument)

The model matrix X <- model.matrix(formula, data)[, -1] (default treatment contrasts) is standardized ONCE, Xs <- scale(X, TRUE, TRUE), all columns including dummies (the rejoinder convention). The outcome model AND glmnet are both fit on this SAME Xs with glmnet(standardize = FALSE). Per-subset scaling is structurally impossible. Coefficients are back-transformed for output: \alpha_j = \alpha^{std}_j / s_j and \alpha_0 = \alpha^{std}_0 - \sum_j \alpha^{std}_j c_j / s_j, with centers c and scales s stored in info.

Weight (outcome) model and the degenerate-beta policy

By default b comes from lm(y ~ A + Xs) (or glm(y ~ A + Xs, binomial)) on the FULL sample, both arms pooled – the Shortreed-Ertefaie/rejoinder convention (not control-arm-only). Only the covariate coefficients are used; the treatment coefficient is discarded and the treatment is never penalized anywhere. Any NA or exactly zero internal coefficient is an ERROR naming the columns: OLS betas are almost surely nonzero, so an exact zero or NA signals rank deficiency or aliasing that would silently corrupt the penalty-scale correction. Remove the collinear columns or supply outcome.coef.

Tuning grid and the \gamma pairing

lambda supplies exponents \delta with \lambda_n = n^{\delta}, n = nrow(model.matrix) always. With gamma = NULL each \delta is paired with \gamma = 2(\mathrm{gcf} - \delta + 1) (default \mathrm{gcf} = 2), the formula derivable from \lambda_n\, n^{\gamma/2 - 1} = n^{\mathrm{gcf}}; there are no \lambda \times \gamma cross-products in paired mode. A scalar gamma is crossed with all \delta (Schnitzer's \gamma = 2.5 is reachable this way; convert their raw \lambda grid via \delta = \log \lambda_n / \log n).

GOAL

For \lambda_2 > 0 the elastic-net term is solved by the Zou-Hastie augmentation: X_{aug} = \mathrm{rbind}(X_s, \sqrt{\lambda_2} I_p) with p pseudo-responses 0, followed by rescaling ALL coefficients INCLUDING the intercept by (1 + \lambda_2) (Balde 2025 supplement: PS = \mathrm{expit}(\mathrm{cbind}(1, X)\,(1 + \lambda_2)\,\hat\alpha)); the propensity score uses the original n rows only. On augmented grid points the raw penalty constant is \lambda_n = (n + q)^{\delta} with q = p augmentation rows (Balde's adaptive.lasso(lambda = n.q^(il)), n.q = n + q), and the penalty-scale constant uses 1/(n + q) in place of 1/n, i.e. s = \overline{\mathrm{pen}} (n + q)^{\delta} / (n + q). \lambda_2 = 0 grid points run the plain-OAL code path (no augmentation, \lambda_n = n^{\delta}), so method = "goal" with lambda2 = 0 nests method = "oal" exactly. (Balde's script literally augments with q zero rows and the (n+q)^\delta base even at \lambda_2 = 0; the zero rows still shift the intercept, so exact nesting is deliberately preferred here.) \lambda_2 is selected jointly with (\delta, \gamma) by the wAMD: the flat first-minimum over the \lambda_2-outer \times \delta-inner grid, which is equivalent to Balde's nested rule (per-\lambda_2 minima over (\delta, \gamma), then the first minimum over \lambda_2).

wAMD criterion and selection

Every grid point is scored by oal_wamd(): propensity scores clipped to clip first (note that Balde's reference implementation is unclipped; see the clip argument), IPW weights at estimand, per-column weighted absolute mean differences on the standardized matrix summed with weights |b_j| over ALL columns (including excluded ones). The selected grid point is the FIRST minimum (relative tolerance 1e-9) in the grid order \lambda_2 ascending (0 first) outer \times \delta ascending inner – ties therefore prefer plain OAL and then the smallest penalty. The wAMD is a balance heuristic: no published proof guarantees it lands \lambda_n inside the window required by the theory (\lambda_n/\sqrt{n} \to 0, \lambda_n n^{\gamma/2 - 1} \to \infty, \gamma > 1), and it inherits the outcome model's misspecification vulnerability.

Why exclude instruments?

Covariates that predict treatment but not outcome should be excluded from a propensity score model: they increase the variance of the effect estimate without decreasing bias (Brookhart et al. 2006) and amplify the bias from any unmeasured confounding (Myers et al. 2011). OAL's penalty implements exactly this exclusion.

Provenance labels

print() and summary() always print a provenance line first: "OAL (Shortreed & Ertefaie 2017, doi:10.1111/biom.12679)" or "GOAL (Balde, Yang & Lefebvre 2023, doi:10.1111/biom.13683)" with a clause stating whether the lambda2 grid is the author's published grid (Balde 2025 supplement) or user-specified. Supplying outcome.coef appends "user-supplied outcome coefficients — screening use only; no oracle property"; family = binomial() and refit = TRUE append their own flags.

Handoff contract

⁠$ps⁠ is a numeric vector of length n, named by rownames(data), strictly inside (0, 1) after clipping – it satisfies psAve::psave(ps.append = ), MatchIt::matchit(distance = ), and WeightIt::weightit(ps = ) by construction.

There is no seed argument: the whole pipeline is deterministic.

Value

An object of class "oal": a list with components

ps

numeric(n), named by rownames(data): the propensity score, clipped to clip, strictly in (0, 1) – the deliverable.

weights

numeric(n): the IPW weights at estimand implied by ps.

coefficients

named numeric(p+1): selected propensity score coefficients on the ORIGINAL covariate scale (intercept first).

coefficients.std

the same on the standardized scale actually optimized.

selected

data frame, one row per model-matrix column: term, outcome.coef (signed b_j, standardized scale), penalty.factor, coef (original scale), selected (logical), and role ("retained"/"excluded").

lambda

named numeric: delta, lambda.n (= n^\delta; = (n + p)^\delta on GOAL grid points with \lambda_2 > 0), gamma, lambda2 (NA for method = "oal") at the selected point.

criterion, criterion.value

"wamd" and its value at the selection.

path

data frame (or NULL without keep.path), one row per grid point in grid order: lambda2, delta, lambda.n, gamma, s (the exact glmnet s used), wamd, n.selected, refit.converged (NA unless refit), selected.

sets

list (or NULL): the selected variable set at every grid point.

coef.path

(p+1) x G matrix (or NULL): propensity score coefficients across the grid, original scale.

outcome.coef

named numeric(p): the b_j actually used (internal or user), standardized scale.

outcome.model

compact list: coefficient table, family, and a label ("lm(y ~ A + X)", "glm(y ~ A + X, binomial)", or "user-supplied"); the full fit object only under keep.fits.

balance

data frame per ORIGINAL covariate column: smd.un, smd.wt, ks.un, ks.wt via cobalt – identical layout to psAve::psave()$balance.

provenance

the fixed provenance label, printed first by print()/summary().

treat

integer(n) 0/1 treatment as coerced.

covs

n x p numeric original-scale model matrix with attr(, "bin.vars").

estimand, method, refit, clip

as resolved.

formula, data, terms, xlevels

stored to power oal_match(), oal_weight() and predict.oal().

fits

list (or NULL): glmnet (one path per (\lambda_2, \gamma)), outcome, refit – iff keep.fits.

info

list: n, n.treated, p, grid.size, family, gamma.mode ("paired"/"fixed"/"user-coef"), center, scale, contrasts, glmnet.version, oalasso.version.

call

the matched call.

References

Shortreed SM, Ertefaie A (2017). Outcome-adaptive lasso: variable selection for causal inference. Biometrics, 73(4), 1111-1122. doi:10.1111/biom.12679

Balde I, Yang Y, Lefebvre G (2023). Reader reaction to "Outcome-adaptive lasso: variable selection for causal inference" by Shortreed and Ertefaie (2017). Biometrics, 79(1), 514-520. doi:10.1111/biom.13683

Jones B, Ertefaie A, Shortreed SM (2023). Rejoinder to reader reaction "On the use of the outcome-adaptive lasso for propensity score estimation". Biometrics, 79(1), 521-525. doi:10.1111/biom.13681

Balde I (2025). Oracle properties of the generalized outcome-adaptive lasso. Statistics & Probability Letters. doi:10.1016/j.spl.2025.110379

Zou H (2006). The adaptive lasso and its oracle properties. Journal of the American Statistical Association, 101(476), 1418-1429. doi:10.1198/016214506000000735

Brookhart MA, Schneeweiss S, Rothman KJ, Glynn RJ, Avorn J, Sturmer T (2006). Variable selection for propensity score models. American Journal of Epidemiology, 163(12), 1149-1156. doi:10.1093/aje/kwj149

Myers JA, Rassen JA, Gagne JJ, Huybrechts KF, Schneeweiss S, Rothman KJ, Joffe MM, Glynn RJ (2011). Effects of adjusting for instrumental variables on bias and precision of effect estimates. American Journal of Epidemiology, 174(11), 1213-1222. doi:10.1093/aje/kwr364

Schnitzer ME, Talbot D, Liu Y, Berger C, Wang G, O'Loughlin J, Sylvestre M-P, Ertefaie A (2025). Outcome-adaptive LASSO for longitudinal data. Statistics in Medicine (arXiv:2410.08283).

See Also

oal_match(), oal_weight(), oal_wamd(), predict.oal(), plot.oal(), bal.tab.oal()

Examples

data("lalonde", package = "MatchIt")

## 1) Matching
fit <- oal(treat ~ age + educ + race + married + nodegree + re74 + re75,
           data = lalonde, outcome = ~ re78)
fit
m <- oal_match(fit)   # = MatchIt::matchit(f, lalonde, distance = fit$ps)

## 2) Weighting (requires WeightIt)
# w <- oal_weight(fit)  # = WeightIt::weightit(f, lalonde, ps = fit$ps, "ATE")

## 3) psAve composition (requires psAve); cbind() labels the candidate
##    while keeping the row names aligned with rownames(data)
# ma <- psAve::psave(treat ~ age + educ + race + married + nodegree +
#                      re74 + re75,
#                    data = lalonde, outcome = ~ re78,
#                    ps.append = cbind(oal = fit$ps))

Match on an outcome-adaptive lasso propensity score

Description

Convenience pass-through to MatchIt::matchit(): matches on the propensity score of an oal() fit, reusing the formula and data stored in the object. Equivalent to the canonical explicit call

MatchIt::matchit(<formula>, data = <data>, distance = fit$ps, ...)

but with no opportunity for row misalignment between the two steps. All ... arguments are forwarded verbatim; the return value is an ordinary matchit object, so the full MatchIt/cobalt toolkit applies.

Usage

oal_match(object, ...)

Arguments

Value

A matchit object; see MatchIt::matchit().

See Also

oal(), oal_weight(), MatchIt::matchit()

Examples

data("lalonde", package = "MatchIt")
fit <- oal(treat ~ age + educ + married + re74, data = lalonde,
           outcome = ~ re78)
m <- oal_match(fit, method = "nearest", caliper = 0.2)

Weighted absolute mean difference (wAMD) of Shortreed and Ertefaie (2017)

Description

Scores an arbitrary candidate propensity score on the exact wAMD balance criterion used by oal() to select its tuning parameters. This function is the package's single source of truth for the criterion: the internal grid search calls it verbatim, and exporting it lets any candidate score – from any package – be scored on the same yardstick (the mirror of psAve::psave_criteria()).

Usage

oal_wamd(
  ps,
  treat,
  covs,
  coef,
  estimand = c("ATE", "ATT"),
  clip = c(0.01, 0.99)
)

Arguments

Details

The propensity scores are first clipped to clip; the inverse-probability weights at estimand are then (with e_i the clipped score)

ATE:\; w_i = A_i/e_i + (1 - A_i)/(1 - e_i), \qquad ATT:\; w_i = A_i + (1 - A_i)\, e_i/(1 - e_i).

For each covariate column j the weighted absolute mean difference is

d_j = \left| \frac{\sum_i w_i X_{ij} A_i}{\sum_i w_i A_i} - \frac{\sum_i w_i X_{ij} (1 - A_i)}{\sum_i w_i (1 - A_i)} \right|,

and the criterion is \mathrm{wAMD} = \sum_j |\beta_j|\, d_j, summed over all columns of covs (including any the propensity score model excluded), with \beta_j the outcome-model coefficients supplied in coef (their absolute values are taken internally).

Inside oal(), covs is the standardized model matrix and coef the outcome coefficients on that same standardized scale (the wAMD, like the adaptive weights, is scale-dependent). The formula is implemented natively – never delegated to cobalt – so that it reproduces the published criterion exactly.

Value

A list with elements

total

the wAMD, \sum_j |\beta_j| d_j;

by.covariate

named numeric vector of the per-covariate contributions |\beta_j| d_j (they sum to total).

References

Shortreed SM, Ertefaie A (2017). Outcome-adaptive lasso: variable selection for causal inference. Biometrics, 73(4), 1111-1122. doi:10.1111/biom.12679

See Also

oal(), weights.oal()

Examples

set.seed(7)
X <- matrix(rnorm(200), 100, 2, dimnames = list(NULL, c("x1", "x2")))
A <- rbinom(100, 1, plogis(X[, 1]))
ps <- plogis(0.8 * X[, 1])
oal_wamd(ps, A, X, coef = c(x1 = 0.5, x2 = 0.1), estimand = "ATE")

Weight by an outcome-adaptive lasso propensity score

Description

Convenience pass-through to WeightIt::weightit(): constructs balancing weights from the propensity score of an oal() fit, reusing the stored formula and data. Equivalent to the canonical explicit call

WeightIt::weightit(<formula>, data = <data>, ps = fit$ps, estimand = <estimand>, ...)

All ... arguments are forwarded verbatim; the return value is an ordinary weightit object.

Usage

oal_weight(object, estimand = object$estimand, ...)

Arguments

Value

A weightit object; see WeightIt::weightit().

See Also

oal(), oal_match(), WeightIt::weightit(), WeightIt::get_w_from_ps()

Examples

data("lalonde", package = "MatchIt")
fit <- oal(treat ~ age + educ + married + re74, data = lalonde,
           outcome = ~ re78)
w <- oal_weight(fit)

Plot an oal object

Description

Three diagnostic displays for an oal() fit:

"wamd"

the wAMD selection criterion against the exponent \delta (lambda), one curve per lambda2 value for method = "goal"; the selected point is marked. Requires keep.path = TRUE.

"coef"

the original-scale propensity score coefficient paths against \delta at the selected lambda2, one line per covariate; covariates driven to zero (instruments, noise) are visible as lines absorbed into the axis. Requires keep.path = TRUE.

"balance"

a Love plot of covariate balance before/after weighting, via cobalt::love.plot() (dispatched through bal.tab.oal(); cobalt is an Import, so always available).

Usage

## S3 method for class 'oal'
plot(x, type = c("wamd", "coef", "balance"), ...)

Arguments

Value

For "balance", the ggplot object from cobalt::love.plot() (invisibly, after printing); otherwise x, invisibly.

See Also

oal(), bal.tab.oal(), cobalt::love.plot()

Predict outcome-adaptive lasso propensity scores for new data

Description

Computes propensity scores (or the linear predictor) for new observations from the stored terms, factor levels, and ORIGINAL-scale coefficients of an oal() fit. Works without keep.fits = TRUE and for both refit modes; the training standardization is never re-estimated (the original-scale coefficients already absorb it). Missing values in newdata are an error.

Usage

## S3 method for class 'oal'
predict(object, newdata = NULL, type = c("ps", "link"), ...)

Arguments

Value

A numeric vector with one value per row, named by rownames. For newdata = NULL and type = "ps" this is exactly object$ps.

See Also

oal(), coef.oal()

Print an oal object

Description

Prints a one-screen teaching summary of a fitted oal() object: the provenance label first (which method, which tuning provenance), the sample, the selected tuning point and its wAMD, the retained (outcome-related) versus excluded (outcome-unrelated) covariates with their outcome-coefficient magnitudes, the fixed instrument-exclusion rationale, and then the literal next call – echoing the formula and data name from your own oal() call – that hands the score to MatchIt::matchit() or WeightIt::weightit(). A near-positivity diagnostic is appended when more than 5\ clipping bound.

Usage

## S3 method for class 'oal'
print(x, digits = 3, ...)

Arguments

Value

x, invisibly.

See Also

oal(), summary.oal()

Summarize an oal object

Description

Produces (a) the full per-covariate selection table (selected), (b) a path summary (the best grid row per lambda2 value, when the path was kept), (c) the outcome (weight) model coefficient table, (d) the full balance table (unweighted vs. weighted SMD and KS, with a * marker at weighted SMD > 0.1), and (e) the near-positivity clip diagnostic.

Usage

## S3 method for class 'oal'
summary(object, ...)

## S3 method for class 'summary.oal'
print(x, digits = 3, ...)

Arguments

Value

For summary.oal(), an object of class "summary.oal": a list with elements call, provenance, method, estimand, refit, lambda, criterion, criterion.value, selected, path.summary, outcome.model, balance, clip, clip.share, and nn. print.summary.oal() returns x invisibly.

See Also

oal(), print.oal(), bal.tab.oal()

Extract the inverse-probability weights of an oal fit

Description

Returns object$weights: the IPW weights at the fitted estimand implied by the clipped propensity score (ATE: A/e + (1-A)/(1-e); ATT: A + (1-A)\,e/(1-e)). For other estimands use WeightIt::get_w_from_ps(object$ps, object$treat, estimand = ...) or oal_weight().

Usage

## S3 method for class 'oal'
weights(object, ...)

Arguments

Value

The numeric vector object$weights, named by rownames(data).

See Also

oal(), oal_weight(), oal_wamd()