Propensity-score weights with the pieces needed to carry their estimation into the inference
Source:R/ps_weights.R
ps_weights.RdFits (or takes) a logistic regression for a binary treatment and returns inverse-probability
weights together with the influence of the fitted coefficients and the derivative of each weight
in those coefficients. Passed as weight_model to df_counting(), KM_diff(),
cumulative_rmst_bands() or plotKM.band_subgroups() with weight_variance = "estimated",
the object lets standard errors, intervals, bands and tests account for the fact that the
weights were estimated (see the section "Estimated weights" of df_counting()).
Arguments
- formula
A formula for the treatment model, treatment (coded 0/1) on the left-hand side. Used with
data; fitted bystats::glm(formula, family = stats::binomial(), data = data)with the default control, so that weights computed by hand from the same call agree.- data
Data frame holding the variables of
formula: exactly the rows that will be analysed.- type
"ipw"(default):1 / efor treated subjects and1 / (1 - e)for controls, withethe fitted probability of treatment."stabilized": those weights multiplied by the proportion treated (treated subjects) and the proportion not treated (controls).- model
An existing logistic
glmfit, instead offormulaanddata. Give eitherformulaanddata, ormodel, not both.- x
A
"ps_weights"object.- digits
Number of significant digits to print.
- ...
Not used.
Value
An object of class "ps_weights", a list with
- weights
The weights, one per row.
- type
"ipw"or"stabilized".- treat
The treatment (0/1), one per row.
- ps
The fitted probabilities of treatment.
- IF
Matrix, one row per subject and one column per estimated coefficient (aliased columns dropped): the influence of each subject on the coefficients.
- dw
Matrix of the same shape: the derivative of each weight in the coefficients.
- model
The
glmfit.- call
The call.
print() shows the type, the number of rows and of coefficients, and by arm the size, the range
of the weights and the effective sample size \((\sum w)^2 / \sum w^2\).
Details
With \(x_i\) the row of the model matrix of subject \(i\), \(A_i\) the treatment and
\(e_i\) the fitted probability,
$$IF_i = \Big\{\sum_j x_j x_j' e_j (1 - e_j)\Big\}^{-1} x_i (A_i - e_i)$$
is the influence of subject \(i\) on the fitted coefficients (the estimate minus its limit is
approximately \(\sum_i IF_i\)), and the derivative of the weight of subject \(i\)
in the coefficients is
\(dw_i = -\{(1 - e_i) / e_i\} x_i\) for a treated subject and
\(dw_i = \{e_i / (1 - e_i)\} x_i\) for a control. With
type = "stabilized" the weights and dw are those of "ipw" multiplied by the arm proportion,
which is treated as a fixed number; IF is the same.
The following are errors, each naming the problem: a model that is not a glm; a family other
than binomial or a link other than logit; prior weights other than 1 or an offset; a fit that did
not converge; a fitted probability within 1e-10 of 0 or 1; rows dropped for missing values; a
response that does not take both values 0 and 1.
The weights are never truncated or rescaled. The rows of weights, treat, ps, IF and dw
are in the order of the rows of the data the model was fitted on.
See also
df_counting() (section "Estimated weights"), KM_diff(), cox_rhogamma()
Other survival_analysis:
KM_diff(),
cox_rhogamma(),
df_counting(),
wt.rg.S()
Examples
library(survival)
gbsg$tte <- gbsg$rfstime / 30.4375
pw <- ps_weights(hormon ~ age + meno + size + grade + nodes + pgr + er, data = gbsg,
type = "stabilized")
pw
#> Propensity-score weights (stabilized): n = 686, 8 coefficients in the logistic model
#> arm n weight min weight max ess ess / n
#> treat = 1 246 0.4493 2.63 203.0 0.8253
#> treat = 0 440 0.7379 1.85 417.7 0.9493
gbsg$sw <- pw$weights
# weights treated as known (the default) and with their estimation carried through
known <- KM_diff(gbsg, "tte", "status", "hormon", weight.name = "sw")
est <- KM_diff(gbsg, "tte", "status", "hormon", weight.name = "sw",
weight_variance = "estimated", weight_model = pw)
identical(known$dhat, est$dhat) # estimates do not change
#> [1] TRUE
summary(sqrt(est$sig2_dhat / known$sig2_dhat)) # standard errors do
#> Min. 1st Qu. Median Mean 3rd Qu. Max. NAs
#> 0.9231 0.9330 0.9445 0.9468 0.9556 0.9979 14
dc <- df_counting(gbsg, "tte", "status", "hormon", weight.name = "sw",
weight_variance = "estimated", weight_model = pw)
dc$ess # effective sample size by arm
#> group n ess ratio
#> 1 treat 246 203.0308 0.8253283
#> 2 control 440 417.7090 0.9493386
cox_rhogamma(dc, scheme = "fh", scheme_params = list(rho = 0, gamma = 0))$hr_ci_asy
#> beta sig_bhat hr lower upper
#> 1 -0.3747545 0.1276936 0.687458 0.5352457 0.8829562