After assessing balance and deciding on a matching specification, it comes time to estimate the effect of the treatment in the matched sample. How the effect is estimated and interpreted depends on the desired estimand and the type of model used (if any). In addition to estimating effects, estimating the uncertainty of the effects is critical in communicating them and assessing whether the observed effect is compatible with there being no effect in the population. This guide explains how to estimate effects after various forms of matching and with various outcome types. There may be situations that are not covered here for which additional methodological research may be required, but some of the recommended methods here can be used to guide such applications.
This guide is structured as follows: first, information on the concepts related to effect and standard error (SE) estimation is presented below. Then, instructions for how to estimate effects and SEs are described for the standard case (matching for the ATT with a continuous outcome) and some other common circumstances. Finally, recommendations for reporting results and tips to avoid making common mistakes are presented.
Before an effect is estimated, the estimand must be specified and clarified. Although some aspects of the estimand depend not only on how the effect is estimated after matching but also on the matching method itself, other aspects must be considered at the time of effect estimation and interpretation. Here, we consider three aspects of the estimand: the population the effect is meant to generalize to (the target population), the effect measure, and whether the effect is marginal or conditional.
The target population. Different matching methods
allow you to estimate effects that can generalize to different target
populations. The most common estimand in matching is the average
treatment effect in the treated (ATT), which is the average effect of
treatment for those who receive treatment. This estimand is estimable
for matching methods that do not change the treated units (i.e., by
weighting or discarding units) and is requested in
matchit()
by setting estimand = "ATT"
(which
is the default). The average treatment effect in the population (ATE) is
the average effect of treatment for the population from which the sample
is a random sample. This estimand is estimable only for methods that
allow the ATE and either do not discard units from the sample or
explicit target full sample balance, which in MatchIt
is
limited to full matching, subclassification, and profile matching when
setting estimand = "ATE"
. When treated units are discarded
(e.g., through the use of common support restrictions, calipers,
cardinality matching, or [coarsened] exact matching), the estimand
corresponds to neither the population ATT nor the population ATE, but
rather to an average treatment effect in the remaining matched sample
(ATM), which may not correspond to any specific target population. See
Greifer and Stuart (2021) for a
discussion on the substantive considerations involved when choosing the
target population of the estimand.
Marginal and conditional effects. A marginal effect is a comparison between the expected potential outcome under treatment and the expected potential outcome under control. This is the same quantity estimated in randomized trials without blocking or covariate adjustment and is particularly useful for quantifying the overall effect of a policy or population-wide intervention. A conditional effect is the comparison between the expected potential outcomes in the treatment groups within strata. This is useful for identifying the effect of a treatment for an individual patient or a subset of the population.
Effect measures. The outcome types we consider here are continuous, with the effect measured by the mean difference; binary, with the effect measured by the risk difference (RD), risk ratio (RR), or odds ratio (OR); and time-to-event (i.e., survival), with the effect measured by the hazard ratio (HR). The RR, OR, and HR are noncollapsible effect measures, which means the marginal effect on that scale is not a (possibly) weighted average of the conditional effects within strata, even if the stratum-specific effects are of the same magnitude. For these effect measures, it is critical to distinguish between marginal and conditional effects because different statistical methods target different types of effects. The mean difference and RD are collapsible effect measures, so the same methods can be used to estimate marginal and conditional effects.
Our primary focus will be on marginal effects, which are appropriate for all effect measures, easily interpretable, and require few modeling assumptions. The “Common Mistakes” section includes examples of commonly used methods that estimate conditional rather than marginal effects and should not be used when marginal effects are desired.
To estimate marginal effects, we use a method known as g-computation (Snowden, Rose, and Mortimer 2011) or regression estimation (Schafer and Kang 2008). This involves first specifying a model for the outcome as a function of the treatment and covariates. Then, for each unit, we compute their predicted values of the outcome setting their treatment status to treated, and then again for control, leaving us with two predicted outcome values for each unit, which are estimates of the potential outcomes under each treatment level. We compute the mean of each of the estimated potential outcomes across the entire sample, which leaves us with two average estimated potential outcomes. Finally, the contrast of these average estimated potential outcomes (e.g., their difference or ratio, depending on the effect measure desired) is the estimate of the treatment effect.
When doing g-computation after matching, a few additional considerations are required. First, when we take the average of the estimated potential outcomes under each treatment level, this must be a weighted average that incorporates the matching weights. Second, if we want to target the ATT or ATC, we only estimate potential outcomes for the treated or control group, respectively (though we still generate predicted values under both treatment and control).
G-computation as a framework for estimating effects after matching has a number of advantages over other approaches. It works the same regardless of the form of the outcome model or type of outcome (e.g., whether a linear model is used for a continuous outcome or a logistic model is used for a binary outcome); the only difference might be how the average expected potential outcomes are contrasted in the final step. In simple cases, the estimated effect is numerically identical to effects estimated using other methods; for example, if no covariates are included in the outcome model, the g-computation estimate is equal to the difference in means from a t-test or coefficient of the treatment in a linear model for the outcome. There are analytic approximations to the SEs of the g-computation estimate, and these SEs can incorporate pair/subclass membership (described in more detail below).
For all these reasons, we use g-computation when possible for all effect estimates, even if there are simpler methods that would yield the same estimates. Using a single workflow (with some slight modifications depending on the context; see below) facilitates implementing best practices regardless of what choices a user makes.
The goal of the outcome model is to generate good predictions for use in the g-computation procedure described above. The type and form of the outcome model should depend on the outcome type. For continuous outcomes, one can use a linear model regressing the outcome on the treatment; for binary outcomes, one can use a generalized linear model with, e.g., a logistic link; for time-to-event outcomes, one can use a Cox proportional hazards model.
An additional decision to make is whether (and how) to include covariates in the outcome model. One may ask, why use matching at all if you are going to model the outcome with covariates anyway? Matching reduces the dependence of the effect estimate on correct specification of the outcome model; this is the central thesis of Ho et al. (2007). Including covariates in the outcome model after matching has several functions: it can increase precision in the effect estimate, reduce the bias due to residual imbalance, and make the effect estimate “doubly robust”, which means it is consistent if either the matching reduces sufficient imbalance in the covariates or if the outcome model is correct. For these reasons, we recommend covariate adjustment after matching when possible. There is some evidence that covariate adjustment is most helpful for covariates with standardized mean differences greater than .1 (Nguyen et al. 2017), so these covariates and covariates thought to be highly predictive of the outcome should be prioritized in treatment effect models if not all can be included due to sample size constraints.
Although there are many possible ways to include covariates (e.g., not just main effects but interactions, smoothing terms like splines, or other nonlinear transformations), it is important not to engage in specification search (i.e., trying many outcomes models in search of the “best” one). Doing so can invalidate results and yield a conclusion that fails to replicate. For this reason, we recommend only including the same terms included in the propensity score model unless there is a strong a priori and justifiable reason to model the outcome differently.
It is important not to interpret the coefficients and tests of covariates in the outcome model. These are not causal effects and their estimates may be severely confounded. Only the treatment effect estimate can be interpreted as causal assuming the relevant assumptions about unconfoundedness are met. Inappropriately interpreting the coefficients of covariates in the outcome model is known as the Table 2 fallacy (Westreich and Greenland 2013). To avoid this, we only display the results of the g-computation procedure and do not examine or interpret the outcome models themselves.
Uncertainty estimation (i.e., of SEs, confidence intervals, and p-values) may consider the variety of sources of uncertainty present in the analysis, including (but not limited to!) estimation of the propensity score (if used), matching (i.e., because treated units might be matched to different control units if others had been sampled), and estimation of the treatment effect (i.e., because of sampling error). In general, there are no analytic solutions to all these issues, so much of the research done on uncertainty estimation after matching has relied on simulation studies. The two primary methods that have been shown to perform well in matched samples are using cluster-robust SEs and the bootstrap, described below.
To compute SEs after g-computation, a method known as the delta method is used; this is a way to compute the SEs of the derived quantities (the expected potential outcomes and their contrast) from the variance of the coefficients of the outcome models. For nonlinear models (e.g., logistic regression), the delta method is only an approximation subject to error (though in many cases this error is small and shrinks in large samples). Because the delta method relies on the variance of the coefficients from the outcome model, it is important to correctly estimate these variances, using either robust or cluster-robust methods as described below.
Robust standard errors. Also known as sandwich SEs (due to the form of the formula for computing them), heteroscedasticity-consistent SEs, or Huber-White SEs, robust SEs are an adjustment to the usual maximum likelihood or ordinary least squares SEs that are robust to violations of some of the assumptions required for usual SEs to be valid (MacKinnon and White 1985). Although there has been some debate about their utility (King and Roberts 2015), robust SEs rarely degrade inferences and often improve them. Generally, robust SEs must be used when any non-uniform weights are included in the estimation (e.g., with matching with replacement or inverse probability weighting).
Cluster-robust standard errors. A version of robust SEs known as cluster-robust SEs (Liang and Zeger 1986) can be used to account for dependence between observations within clusters (e.g., matched pairs). Abadie and Spiess (2019) demonstrate analytically that cluster-robust SEs are generally valid after matching, whereas regular robust SEs can over- or under-estimate the true sampling variability of the effect estimator depending on the specification of the outcome model (if any) and degree of effect modification. A plethora of simulation studies have further confirmed the validity of cluster-robust SEs after matching (e.g., Austin 2009, 2013a; Austin and Small 2014; Gayat et al. 2012; Wan 2019). Given this evidence favoring the use of cluster-robust SEs, we recommend them in most cases and use them judiciously in this guide1.
One problem when using robust and cluster-robust SEs along with the delta method is that the delta method is an approximation, as previously mentioned. One solution to this problem is bootstrapping, which is a technique used to simulate the sampling distribution of an estimator by repeatedly drawing samples with replacement and estimating the effect in each bootstrap sample (Efron and Tibshirani 1993). From the bootstrap distribution, SEs and confidence intervals can be computed in several ways, including using the standard deviation of the bootstrap estimates as the SE estimate or using the 2.5 and 97.5 percentiles as 95% confidence interval bounds. Bootstrapping tends to be most useful when no analytic estimator of a SE is possible or has been derived yet. Although Abadie and Imbens (2008) found analytically that the bootstrap is inappropriate for matched samples, simulation evidence has found it to be adequate in many cases (Hill and Reiter 2006; Austin and Small 2014; Austin and Stuart 2017).
Typically, bootstrapping involves performing the entire estimation process in each bootstrap sample, including propensity score estimation, matching, and effect estimation. This tends to be the most straightforward route, though intervals from this method may be conservative in some cases (i.e., they are wider than necessary to achieve nominal coverage) (Austin and Small 2014). Less conservative and more accurate intervals have been found when using different forms of the bootstrap, including the wild bootstrap develop by Bodory et al. (2020) and the matched/cluster bootstrap described by Austin and Small (2014) and Abadie and Spiess (2019). The cluster bootstrap involves sampling matched pairs/strata of units from the matched sample and performing the analysis within each sample composed of the sampled pairs. Abadie and Spiess (2019) derived analytically that the cluster bootstrap is valid for estimating SEs and confidence intervals in the same circumstances cluster robust SEs are; indeed, the cluster bootstrap SE is known to approximate the cluster-robust SE (Cameron and Miller 2015).
With bootstrapping, more bootstrap replications are always better but
can take time and increase the chances that at least one error will
occur within the bootstrap analysis (e.g., a bootstrap sample with zero
treated units or zero units with an event). In general, numbers of
replications upwards of 999 are recommended, with values one less than a
multiple of 100 preferred to avoid interpolation when using the
percentiles as confidence interval limits (MacKinnon 2006). There are several
methods of computing bootstrap confidence intervals, but the
bias-corrected accelerated (BCa) bootstrap confidence interval often
performs best (Austin
and Small 2014; Carpenter and Bithell
2000) and is easy to implement, simply by setting
type = "bca"
in the call to boot::boot.ci()
after running boot::boot()
2.
Most of this guide will consider analytic (i.e., non-bootstrapping) approaches to estimating uncertainty; the section “Using Bootstrapping to Estimate Confidence Intervals” describes broadly how to use bootstrapping. Although analytic estimates are faster to compute, in many cases bootstrap confidence intervals are more accurate.
Below, we describe effect estimation after matching. We’ll be using a
simulated toy dataset d
with several outcome types. Code to
generate the dataset is at the end of this document. The focus here is
not on evaluating the methods but simply on demonstrating them. In all
cases, the correct propensity score model is used. Below we display the
first six rows of d
:
## A X1 X2 X3 X4 X5 X6 X7 X8 X9 Y_C Y_B Y_S
## 1 0 0.1725 -1.4283 -0.4103 -2.36059 1 -1.1199 0.6398 -0.4840 -0.59385 0.07104 0 278.46
## 2 0 -1.0959 0.8463 0.2456 -0.12333 1 -2.2687 -1.4491 -0.5514 -0.31439 0.15619 0 330.63
## 3 0 0.1768 0.7905 -0.8436 0.82366 1 -0.2221 0.2971 -0.6966 -0.69516 -0.85180 1 369.94
## 4 0 -0.4595 0.1726 1.9542 -0.62661 1 -0.4019 -0.8294 -0.5384 0.20729 -2.35184 0 91.06
## 5 1 0.3563 -1.8121 0.8135 -0.67189 1 -0.8297 1.7297 -0.6439 -0.02648 0.68058 0 182.73
## 6 0 -2.4313 -1.7984 -1.2940 0.04609 1 -1.2419 -1.1252 -1.8659 -0.56513 -5.62260 0 2563.73
A
is the treatment variable, X1
through
X9
are covariates, Y_C
is a continuous
outcome, Y_B
is a binary outcome, and Y_S
is a
survival outcome.
We will need to the following packages to perform the desired analyses:
marginaleffects
provides the
avg_comparisons()
function for performing g-computation and
estimating the SEs and confidence intervals of the average estimate
potential outcomes and treatment effectssandwich
is used internally by
marginaleffects
to compute robust and cluster-robust
SEssurvival
provides coxph()
to estimate the
coefficients in a Cox-proportional hazards model for the marginal hazard
ratio, which we will use for survival outcomes.Of course, we also need MatchIt
to perform the
matching.
All effect estimates will be computed using
marginaleffects::avg_comparions()
, even when its use may be
superfluous (e.g., for performing a t-test in the matched set). As
previously mentioned, this is because it is useful to have a single
workflow that works no matter the situation, perhaps with very slight
modifications to accommodate different contexts. Using
avg_comparions()
has several advantages, even when the
alternatives are simple: it only provides the effect estimate, and not
other coefficients; it automatically incorporates robust and
cluster-robust SEs if requested; and it always produces average marginal
effects for the correct population if requested.
Other packages may be of use but are not used here. There are
alternatives to the marginaleffects
package for computing
average marginal effects, including margins
and
stdReg
. The survey
package can be used to
estimate robust SEs incorporating weights and provides functions for
survey-weighted generalized linear models and Cox-proportional hazards
models.
For almost all matching methods, whether a caliper, common support
restriction, exact matching specification, or \(k\):1 matching specification is used,
estimating the effect in the matched dataset is straightforward and
involves fitting a model for the outcome that incorporates the matching
weights3,
then estimating the treatment effect using g-computation (i.e., using
marginaleffects::avg_comparisons()
) with a cluster-robust
SE to account for pair membership. This procedure is the same for
continuous and binary outcomes with and without covariates.
There are a few adjustments that need to be made for certain scenarios, which we describe in the section “Adjustments to the Standard Case”. These adjustments include for the following cases: when matching for the ATE rather than the ATT, for matching with replacement, for matching with a method that doesn’t involve creating pairs (e.g., cardinality and profile matching and coarsened exact matching), for subclassification, for estimating effects with binary outcomes, and for estimating effects with survival outcomes. You must read the Standard Case to understand the basic procedure before reading about these special scenarios.
Here, we demonstrate the faster analytic approach to estimating confidence intervals; for the bootstrap approach, see the section “Using Bootstrapping to Estimate Confidence Intervals” below.
First, we will perform full matching on the propensity score for the ATT. Remember, all matching methods use this exact procedure or a slight variation, so this section is critical even if you are using a different matching method.
#Optimal full matching on the PS for the ATT
mF <- matchit(A ~ X1 + X2 + X3 + X4 + X5 +
X6 + X7 + X8 + X9, data = d,
method = "full", estimand = "ATT")
mF
## A `matchit` object
## - method: Optimal full matching
## - distance: Propensity score - estimated with logistic regression
## - number of obs.: 2000 (original), 2000 (matched)
## - target estimand: ATT
## - covariates: X1, X2, X3, X4, X5, X6, X7, X8, X9
## A X1 X2 X3 X4 X5 X6 X7 X8 X9 Y_C Y_B Y_S distance weights subclass
## 1 0 0.1725 -1.4283 -0.4103 -2.36059 1 -1.1199 0.6398 -0.4840 -0.59385 0.07104 0 278.46 0.08461 0.20795 150
## 2 0 -1.0959 0.8463 0.2456 -0.12333 1 -2.2687 -1.4491 -0.5514 -0.31439 0.15619 0 330.63 0.01855 0.04475 119
## 3 0 0.1768 0.7905 -0.8436 0.82366 1 -0.2221 0.2971 -0.6966 -0.69516 -0.85180 1 369.94 0.22210 0.88379 288
## 4 0 -0.4595 0.1726 1.9542 -0.62661 1 -0.4019 -0.8294 -0.5384 0.20729 -2.35184 0 91.06 0.04180 0.17676 303
## 5 1 0.3563 -1.8121 0.8135 -0.67189 1 -0.8297 1.7297 -0.6439 -0.02648 0.68058 0 182.73 0.43291 1.00000 86
## 6 0 -2.4313 -1.7984 -1.2940 0.04609 1 -1.2419 -1.1252 -1.8659 -0.56513 -5.62260 0 2563.73 0.04998 0.12626 158
Typically one would assess balance and ensure that this matching
specification works, but we will skip that step here to focus on effect
estimation. See vignette("MatchIt")
and
vignette("assessing-balance")
for more information on this
necessary step. Because we did not use a caliper, the target estimand is
the ATT.
We perform all analyses using the matched dataset, md
,
which, for matching methods that involve dropping units, contains only
the units retained in the sample.
First, we fit a model for the outcome given the treatment and
(optionally) the covariates. It’s usually a good idea to include
treatment-covariate interactions, which we do below, but this is not
always necessary, especially when excellent balance has been achieved.
You can also include the propensity score (usually labeled
distance
in the match.data()
output), which
can add some robustness, especially when modeled flexibly (e.g., with
polynomial terms or splines) (Austin
2017); see here for an
example.
#Linear model with covariates
fit1 <- lm(Y_C ~ A * (X1 + X2 + X3 + X4 + X5 +
X6 + X7 + X8 + X9),
data = md, weights = weights)
Next, we use marginaleffects::avg_comparisons()
to
estimate the ATT.
##
## Estimate Std. Error z Pr(>|z|) S 2.5 % 97.5 %
## 1.75 0.479 3.65 <0.001 11.9 0.81 2.69
##
## Term: A
## Type: response
## Comparison: mean(1) - mean(0)
## Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, predicted_lo, predicted_hi, predicted
Let’s break down the call to avg_comparisons()
: to the
first argument, we supply the model fit, fit1
; to the
variables
argument, the name of the treatment
("A"
); to the vcov
argument, a formula with
subclass membership (~subclass
) to request cluster-robust
SEs; and to the newdata
argument, a version of the matched
dataset containing only the treated units (subset(A == 1)
)
to request the ATT. Some of these arguments differ depending on the
specifics of the matching method and outcome type; see the sections
below for information.
If, in addition to the effect estimate, we want the average estimated
potential outcomes, we can use
marginaleffects::avg_predictions()
, which we demonstrate
below. Note the interpretation of the resulting estimates as the
expected potential outcomes is only valid if all covariates present in
the outcome model (if any) are interacted with the treatment.
##
## A Estimate Std. Error z Pr(>|z|) S 2.5 % 97.5 %
## 0 2.14 0.382 5.61 <0.001 25.6 1.39 2.89
## 1 3.89 0.229 16.99 <0.001 212.7 3.44 4.34
##
## Type: response
## Columns: A, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high
We can see that the difference in potential outcome means is equal to
the average treatment effect computed previously4. All of the arguments
to avg_predictions()
are the same as those to
avg_comparisons()
.
This section explains how the procedure might differ if any of the following special circumstances occur.
When matching for the ATE (including [coarsened] exact matching, full
matching, subclassification, and cardinality matching), everything is
identical to the Standard Case except that in the calls to
avg_comparisons()
and avg_predictions()
, the
newdata
argument is omitted. This is because the estimated
potential outcomes are computed for the full sample rather than just the
treated units.
When matching with replacement (i.e., nearest neighbor or genetic
matching with replace = TRUE
), effect and SE estimation
need to account for control unit multiplicity (i.e., repeated use) and
within-pair correlations (Hill and Reiter 2006; Austin and Cafri 2020). Although
Abadie and Imbens (2008) demonstrated analytically that
bootstrap SEs may be invalid for matching with replacement, simulation
work by Hill and Reiter (2006) and Bodory
et al. (2020) has found that
bootstrap SEs are adequate and generally slightly conservative. See the
section “Using Bootstrapping to Estimate Confidence Intervals” for
instructions on using the bootstrap and an example that use matching
with replacement.
Because control units do not belong to unique pairs, there is no pair
membership in the match.data()
output. One can simply
change vcov = ~subclass
to vcov = "HC3"
in the
calls to comparisons()
and predictions()
to
use robust SEs instead of cluster-robust SEs, as recommended by Hill and Reiter (2006). There is some evidence for an
alternative approach that incorporates pair membership and adjusts for
reuse of control units, though this has only been studied for survival
outcomes (Austin and
Cafri 2020). This adjustment involves using two-way
cluster-robust SEs with pair membership and unit ID as the clustering
variables. For continuous and binary outcomes, this involves the
following two changes: 1) replace match.data()
with
get_matches()
, which produces a dataset with one row per
unit per pair, meaning control units matched to multiple treated units
will appear multiple times in the dataset; 2) set
vcov = ~subclass + id
in the calls to
avg_comparisons()
and avg_predictions()
. For
survival outcomes, a special procedure must be used; see the section on
survival outcomes below.
Some matching methods do not involve creating pairs; these include
cardinality and profile matching with mahvars = NULL
(the
default), exact matching, and coarsened exact matching with
k2k = FALSE
(the default). The only change that needs to be
made to the Standard Case is that one should change
vcov = ~subclass
to vcov = "HC3"
in the calls
to avg_comparisons()
and avg_predictions()
to
use robust SEs instead of cluster-robust SEs. Remember that if matching
is done for the ATE (even if units are dropped), the
newdata
argument should be dropped.
There are two natural ways to estimate marginal effects after subclassification: the first is to estimate subclass-specific treatment effects and pool them using an average marginal effects procedure, and the second is to use the stratum weights to estimate a single average marginal effect. This latter approach is also known as marginal mean weighting through stratification (MMWS), and is described in detail by Hong (2010)5. When done properly, both methods should yield similar or identical estimates of the treatment effect.
All of the methods described above for the Standard Case also work
with MMWS because the formation of the weights is the same; the only
difference is that it is not appropriate to use cluster-robust SEs with
MMWS because of how few clusters are present, so one should change
vcov = ~subclass
to vcov = "HC3"
in the calls
to avg_comparisons()
and avg_predictions()
to
use robust SEs instead of cluster-robust SEs. The subclasses can
optionally be included in the outcome model (optionally interacting with
treatment) as an alternative to including the propensity score.
The subclass-specific approach omits the weights and uses the
subclasses directly. It is only appropriate when there are a small
number of subclasses relative to the sample size. In the outcome model,
subclass
should interact with all other predictors in the
model (including the treatment, covariates, and interactions, if any),
and the weights
argument should be omitted. As with MMWS,
one should change vcov = ~subclass
to
vcov = "HC3"
in the calls to avg_comparisons()
and avg_predictions()
. See an example below:
#Subclassification on the PS for the ATT
mS <- matchit(A ~ X1 + X2 + X3 + X4 + X5 +
X6 + X7 + X8 + X9, data = d,
method = "subclass", estimand = "ATT")
#Extract matched data
md <- match.data(mS)
fitS <- lm(Y_C ~ subclass * (A * (X1 + X2 + X3 + X4 + X5 +
X6 + X7 + X8 + X9)),
data = md)
avg_comparisons(fitS,
variables = "A",
vcov = "HC3",
newdata = subset(A == 1))
##
## Estimate Std. Error z Pr(>|z|) S 2.5 % 97.5 %
## 1.65 0.364 4.54 <0.001 17.4 0.94 2.37
##
## Term: A
## Type: response
## Comparison: mean(1) - mean(0)
## Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, predicted_lo, predicted_hi, predicted
A model with fewer terms may be required when subclasses are small;
removing covariates or their interactions with treatment may be required
and can increase precision in smaller datasets. Remember that if
subclassification is done for the ATE (even if units are dropped), the
newdata
argument should be dropped.
Estimating effects on binary outcomes is essentially the same as for
continuous outcomes. The main difference is that there are several
measures of the effect one can consider, which include the odds ratio
(OR), risk ratio/relative risk (RR), and risk difference (RD), and the
syntax to avg_comparisons()
depends on which one is
desired. The outcome model should be one appropriate for binary outcomes
(e.g., logistic regression) but is unrelated to the desired effect
measure because we can compute any of the above effect measures using
avg_comparisons()
after the logistic regression.
To fit a logistic regression model, change lm()
to
glm()
and set family = quasibinomial()
6. To
compute the marginal RD, we can use exactly the same syntax as in the
Standard Case; nothing needs to change7.
To compute the marginal RR, we need to add
comparison = "lnratioavg"
to
avg_comparisons()
; this computes the marginal log RR. To
get the marginal RR, we need to add transform = "exp"
to
avg_comparisons()
, which exponentiates the marginal log RR
and its confidence interval. The code below computes the effects and
displays the statistics of interest:
#Logistic regression model with covariates
fit2 <- glm(Y_B ~ A * (X1 + X2 + X3 + X4 + X5 +
X6 + X7 + X8 + X9),
data = md, weights = weights,
family = quasibinomial())
#Compute effects; RR and confidence interval
avg_comparisons(fit2,
variables = "A",
vcov = ~subclass,
newdata = subset(A == 1),
comparison = "lnratioavg",
transform = "exp")
##
## Estimate Pr(>|z|) S 2.5 % 97.5 %
## 1.57 <0.001 52.0 1.41 1.75
##
## Term: A
## Type: response
## Comparison: ln(mean(1) / mean(0))
## Columns: term, contrast, estimate, p.value, s.value, conf.low, conf.high, predicted_lo, predicted_hi, predicted
The output displays the marginal RR, its Z-value, the p-value for the
Z-test of the log RR against 0, and its confidence interval. (Note that
even though the Contrast
label still suggests the log RR,
the RR is actually displayed.) To view the log RR and its standard
error, omit the transform
argument.
For the marginal OR, the only thing that needs to change is that
comparison
should be set to "lnoravg"
.
There are several measures of effect size for survival outcomes. When using the Cox proportional hazards model, the quantity of interest is the hazard ratio (HR) between the treated and control groups. As with the OR, the HR is non-collapsible, which means the estimated HR will only be a valid estimate of the marginal HR when no other covariates are included in the model. Other effect measures, such as the difference in mean survival times or probability of survival after a given time, can be treated just like continuous and binary outcomes as previously described.
For the HR, we cannot compute average marginal effects and must use
the coefficient on treatment in a Cox model fit without covariates8. This
means that we cannot use the procedures from the Standard Case. Here we
describe estimating the marginal HR using coxph()
from the
survival
package. (See
help("coxph", package = "survival")
for more information on
this model.) To request cluster-robust SEs as recommended by Austin (2013b),
we need to supply pair membership (stored in the subclass
column of md
) to the cluster
argument and set
robust = TRUE
. For matching methods that don’t involve
pairing (e.g., cardinality and profile matching and [coarsened] exact
matching), we can omit the cluster
argument (but keep
robust = TRUE
)9.
library("survival")
#Cox Regression for marginal HR
coxph(Surv(Y_S) ~ A, data = md, robust = TRUE,
weights = weights, cluster = subclass)
## Call:
## coxph(formula = Surv(Y_S) ~ A, data = md, weights = weights,
## robust = TRUE, cluster = subclass)
##
## coef exp(coef) se(coef) robust se z p
## A 0.45 1.57 0.05 0.04 12 <2e-16
##
## Likelihood ratio test=63 on 1 df, p=2e-15
## n= 2000, number of events= 2000
The coef
column contains the log HR, and
exp(coef)
contains the HR. Remember to always use the
robust se
for the SE of the log HR. The displayed z-test
p-value results from using the robust SE.
For matching with replacement, a special procedure described by Austin and Cafri (2020) can be necessary for valid
inference. According to the results of their simulation studies, when
the treatment prevalence is low (<30%), a SE that does not involve
pair membership (i.e., the match.data()
approach, as
demonstrated above) is sufficient. When treatment prevalence is higher,
the SE that ignores pair membership may be too low, and the authors
recommend using a custom SE estimator that uses information about both
multiplicity and pairing.
Doing so must be done manually for survival models using
get_matches()
and several calls to coxph()
as
demonstrated in the appendix of Austin and Cafri
(2020). We demonstrate this
below:
#get_matches() after matching with replacement
gm <- get_matches(mR)
#Austin & Cafri's (2020) SE estimator
fs <- coxph(Surv(Y_S) ~ A, data = gm, robust = TRUE,
weights = weights, cluster = subclass)
Vs <- fs$var
ks <- nlevels(gm$subclass)
fi <- coxph(Surv(Y_S) ~ A, data = gm, robust = TRUE,
weights = weights, cluster = id)
Vi <- fi$var
ki <- length(unique(gm$id))
fc <- coxph(Surv(Y_S) ~ A, data = gm, robust = TRUE,
weights = weights)
Vc <- fc$var
kc <- nrow(gm)
#Compute the variance and sneak it back into the fit object
fc$var <- (ks/(ks-1))*Vs + (ki/(ki-1))*Vi - (kc/(kc-1))*Vc
fc
The robust se
column contains the computed SE, and the
reported Z-test uses this SE. The se(coef)
column should be
ignored.
The bootstrap is an alternative to the delta method for estimating
confidence intervals for estimated effects. See the section
Bootstrapping above for details. Here, we’ll demonstrate two forms of
the bootstrap: 1) the standard bootstrap, which involve resampling units
and performing matching and effect estimation within each bootstrap
sample, and 2) the cluster bootstrap, which involves resampling pairs
after matching and estimating the effect in each bootstrap sample. For
both, we will use functionality in the boot
package. It is
critical to set a seed using set.seed()
prior to performing
the bootstrap in order for results to be replicable.
For the standard bootstrap, we need a function that takes in the
original dataset and a vector of sampled unit indices and returns the
estimated quantity of interest. This function should perform the
matching on the bootstrap sample, fit the outcome model, and estimate
the treatment effect using g-computation. In this example, we’ll use
matching with replacement, since the standard bootstrap has been found
to work well with it (Bodory et al. 2020; Hill and Reiter 2006), despite some
analytic results recommending otherwise (Abadie and Imbens 2008). We’ll
implement g-computation manually rather than using
avg_comparisons()
, as this dramatically improves the speed
of the estimation since we don’t require standard errors to be estimated
in each sample (or other processing avg_comparisons()
does). We’ll consider the marginal RR ATT of A
on the
binary outcome Y_B
.
The first step is to write the estimation function, we call
boot_fun
. This function returns the marginal RR. In it, we
perform the matching, estimate the effect, and return the estimate of
interest.
boot_fun <- function(data, i) {
boot_data <- data[i,]
#Do 1:1 PS matching with replacement
m <- matchit(A ~ X1 + X2 + X3 + X4 + X5 +
X6 + X7 + X8 + X9,
data = boot_data,
replace = TRUE)
#Extract matched dataset
md <- match.data(m, data = boot_data)
#Fit outcome model
fit <- glm(Y_B ~ A * (X1 + X2 + X3 + X4 + X5 +
X6 + X7 + X8 + X9),
data = md, weights = weights,
family = quasibinomial())
## G-computation ##
#Subset to treated units for ATT; skip for ATE
md1 <- subset(md, A == 1)
#Estimated potential outcomes under treatment
p1 <- predict(fit, type = "response",
newdata = transform(md1, A = 1))
Ep1 <- mean(p1)
#Estimated potential outcomes under control
p0 <- predict(fit, type = "response",
newdata = transform(md1, A = 0))
Ep0 <- mean(p0)
#Risk ratio
return(Ep1 / Ep0)
}
Next, we call boot::boot()
with this function and the
original dataset supplied to perform the bootstrapping. We’ll request
199 bootstrap replications here, but in practice you should use many
more, upwards of 999. More is always better. Using more also allows you
to use the bias-corrected and accelerated (BCa) bootstrap confidence
intervals (which you can request by setting type = "bca"
in
the call to boot.ci()
), which are known to be the most
accurate. See ?boot.ci
for details. Here, we’ll just use a
percentile confidence interval.
##
## ORDINARY NONPARAMETRIC BOOTSTRAP
##
##
## Call:
## boot(data = d, statistic = boot_fun, R = 199)
##
##
## Bootstrap Statistics :
## original bias std. error
## t1* 1.347 0.1417 0.1937
## BOOTSTRAP CONFIDENCE INTERVAL CALCULATIONS
## Based on 199 bootstrap replicates
##
## CALL :
## boot.ci(boot.out = boot_out, type = "perc")
##
## Intervals :
## Level Percentile
## 95% ( 1.144, 1.891 )
## Calculations and Intervals on Original Scale
## Some percentile intervals may be unstable
We find a RR of 1.347 with a confidence interval of (1.144, 1.891).
If we had wanted a risk difference, we could have changed the final line
in boot_fun()
to be Ep1 - Ep0
.
For the cluster bootstrap, we need a function that takes in a vector of subclass (e.g., pairs) and a vector of sampled pair indices and returns the estimated quantity of interest. This function should fit the outcome model and estimate the treatment effect using g-computation, but the matching step occurs prior to the bootstrap. Here, we’ll use matching without replacement, since the cluster bootstrap has been found to work well with it (Austin and Small 2014; Abadie and Spiess 2019). This could be used for any method that returns pair membership, including other pair matching methods without replacement and full matching.
As before, we’ll use g-computation to estimate the marginal RR ATT,
and we’ll do so manually rather than using
avg_comparisons()
for speed. Note that the cluster
bootstrap is already much faster than the standard bootstrap because
matching does not need to occur within each bootstrap sample. First,
we’ll do a round of matching.
## A `matchit` object
## - method: 1:1 nearest neighbor matching without replacement
## - distance: Propensity score - estimated with logistic regression
## - number of obs.: 2000 (original), 882 (matched)
## - target estimand: ATT
## - covariates: X1, X2, X3, X4, X5, X6, X7, X8, X9
Next, we’ll write the function that takes in cluster membership and the sampled indices and returns an estimate.
#Unique pair IDs
pair_ids <- levels(md$subclass)
#Unit IDs, split by pair membership
split_inds <- split(seq_len(nrow(md)), md$subclass)
cluster_boot_fun <- function(pairs, i) {
#Extract units corresponding to selected pairs
ids <- unlist(split_inds[pairs[i]])
#Subset md with block bootstrapped indices
boot_md <- md[ids,]
#Fit outcome model
fit <- glm(Y_B ~ A * (X1 + X2 + X3 + X4 + X5 +
X6 + X7 + X8 + X9),
data = boot_md, weights = weights,
family = quasibinomial())
## G-computation ##
#Subset to treated units for ATT; skip for ATE
md1 <- subset(boot_md, A == 1)
#Estimated potential outcomes under treatment
p1 <- predict(fit, type = "response",
newdata = transform(md1, A = 1))
Ep1 <- mean(p1)
#Estimated potential outcomes under control
p0 <- predict(fit, type = "response",
newdata = transform(md1, A = 0))
Ep0 <- mean(p0)
#Risk ratio
return(Ep1 / Ep0)
}
Next, we call boot::boot()
with this function and the
vector of pair membership supplied to perform the bootstrapping. We’ll
request 199 bootstrap replications, but in practice you should use many
more, upwards of 999. More is always better. Using more also allows you
to use the bias-corrected and accelerated (BCa) boot strap confidence
intervals, which are known to be the most accurate. See
?boot.ci
for details. Here, we’ll just use a percentile
confidence interval.
library("boot")
set.seed(54321)
cluster_boot_out <- boot(pair_ids, cluster_boot_fun,
R = 199)
cluster_boot_out
##
## ORDINARY NONPARAMETRIC BOOTSTRAP
##
##
## Call:
## boot(data = pair_ids, statistic = cluster_boot_fun, R = 199)
##
##
## Bootstrap Statistics :
## original bias std. error
## t1* 1.588 0.001319 0.1265
## BOOTSTRAP CONFIDENCE INTERVAL CALCULATIONS
## Based on 199 bootstrap replicates
##
## CALL :
## boot.ci(boot.out = cluster_boot_out, type = "perc")
##
## Intervals :
## Level Percentile
## 95% ( 1.356, 1.857 )
## Calculations and Intervals on Original Scale
## Some percentile intervals may be unstable
We find a RR of 1.588 with a confidence interval of (1.356, 1.857).
If we had wanted a risk difference, we could have changed the final line
in cluster_boot_fun()
to be Ep1 - Ep0
.
Moderation analysis involves determining whether a treatment effect differs across levels of another variable. The use of matching with moderation analysis is described in Green and Stuart (2014). The goal is to achieve balance within each subgroup of the potential moderating variable, and there are several ways of doing so. Broadly, one can either perform matching in the full dataset, requiring exact matching on the moderator, or one can perform completely separate analyses in each subgroup. We’ll demonstrate the first approach below; see the blog post “Subgroup Analysis After Propensity Score Matching Using R” by Noah Greifer for an example of the other approach.
There are benefits to using either approach, and Green and Stuart (2014) find that either can be successful at balancing the subgroups. The first approach may be most effective with small samples, where separate propensity score models would be fit with greater uncertainty and an increased possibility of perfect prediction or failure to converge (Wang et al. 2018). The second approach may be more effective with larger samples or with matching methods that target balance in the matched sample, such as genetic matching (Kreif et al. 2012). With genetic matching, separate subgroup analyses ensure balance is optimized within each subgroup rather than just overall. The chosen approach should be that which achieves the best balance, though we don’t demonstrate assessing balance here to maintain focus on effect estimation.
The full dataset approach involves pooling information across subgroups. This could involve estimating propensity scores using a single model for both groups but exact matching on the potential moderator. The propensity score model could include moderator-by-covariate interactions to allow the propensity score model to vary across subgroups on some covariates. It is critical that exact matching is done on the moderator so that matched pairs are not split across subgroups.
We’ll consider the binary variable X5
to be the
potential moderator of the effect of A
on Y_C
.
Below, we’ll estimate a propensity score using a single propensity score
model with a few moderator-by-covariate interactions. We’ll perform
nearest neighbor matching on the propensity score and exact matching on
the moderator, X5
.
mP <- matchit(A ~ X1 + X2 + X5*X3 + X4 +
X5*X6 + X7 + X5*X8 + X9, data = d,
exact = ~X5, method = "nearest")
mP
## A `matchit` object
## - method: 1:1 nearest neighbor matching without replacement
## - distance: Propensity score - estimated with logistic regression
## - number of obs.: 2000 (original), 882 (matched)
## - target estimand: ATT
## - covariates: X1, X2, X5, X3, X4, X6, X7, X8, X9
Although it is straightforward to assess balance overall using
summary()
, it is more challenging to assess balance within
subgroups. The easiest way to check subgroup balance would be to use
cobalt::bal.tab()
, which has a cluster
argument that can be used to assess balance within subgroups, e.g., by
cobalt::bal.tab(mP, cluster = "X5")
. See the vignette
“Appendix 2: Using cobalt with Clustered, Multiply Imputed, and Other
Segmented Data” on the cobalt
website for
details.
If we are satisfied with balance, we can then model the outcome with an interaction between the treatment and the moderator.
To estimate the subgroup ATTs, we can use
avg_comparisons()
, this time specifying the by
argument to signify that we want treatment effects stratified by the
moderator.
##
## X5 Estimate Std. Error z Pr(>|z|) S 2.5 % 97.5 %
## 0 2.21 0.670 3.29 <0.001 10.0 0.894 3.52
## 1 2.18 0.569 3.83 <0.001 12.9 1.065 3.29
##
## Term: A
## Type: response
## Comparison: mean(1) - mean(0)
## Columns: term, contrast, X5, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, predicted_lo, predicted_hi, predicted
We can see that the subgroup mean differences are quite similar to
each other. Finally, we can test for moderation using another call to
avg_comparisons()
, this time using the
hypothesis
argument to signify that we want to compare
effects between subgroups:
avg_comparisons(fitP, variables = "A",
vcov = ~subclass,
newdata = subset(A == 1),
by = "X5",
hypothesis = "pairwise")
##
## Estimate Std. Error z Pr(>|z|) S 2.5 % 97.5 %
## 0.0275 0.879 0.0313 0.975 0.0 -1.7 1.75
##
## Term: 0 - 1
## Type: response
## Columns: term, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high
As expected, the difference between the subgroup treatment effects is
small and nonsignificant, so there is no evidence of moderation by
X5
.
It is important to be as thorough and complete as possible when describing the methods of estimating the treatment effect and the results of the analysis. This improves transparency and replicability of the analysis. Results should at least include the following:
glm()
in base R, avg_comparisons()
in
marginaleffects
, boot()
and
boot.ci()
in boot
)All this is in addition to information about the matching method, propensity score estimation procedure (if used), balance assessment, etc. mentioned in the other vignettes.
There are a few common mistakes that should be avoided. It is important not only to avoid these mistakes in one’s own research but also to be able to spot these mistakes in others’ analyses.
Several methods involve weights that are to be used in estimating the
treatment effect. With full matching and stratification matching (when
analyzed using MMWS), the weights do the entire work of balancing the
covariates across the treatment groups. Omitting weights essentially
ignores the entire purpose of matching. Some cases are less obvious.
When performing matching with replacement and estimating the treatment
effect using the match.data()
output, weights must be
included to ensure control units matched to multiple treated units are
weighted accordingly. Similarly, when performing k:1 matching where not
all treated units receive k matches, weights are required to account for
the differential weight of the matched control units. The only time
weights can be omitted after pair matching is when performing 1:1
matching without replacement. Including weights even in this scenario
will not affect the analysis and it can be good practice to always
include weights to prevent this error from occurring. There are some
scenarios where weights are not useful because the conditioning occurs
through some other means, such as when using the direct subclass
strategy rather than MMWS for estimating marginal effects after
stratification.
Robust SEs are required when using weights to estimate the treatment
effect. The model-based SEs resulting from weighted least squares or
maximum likelihood are inaccurate when using matching weights because
they assume weights are frequency weights rather than probability
weights. Cluster-robust SEs account for both the matching weights and
pair membership and should be used when appropriate. Sometimes,
researchers use functions in the survey
package to estimate
robust SEs, especially with inverse probability weighting; this is a
valid way to compute robust SEs and will give similar results to
sandwich::vcovHC()
.10
The distinction between marginal and conditional effects is not always clear both in methodological and applied papers. Some statistical methods are valid only for estimating conditional effects and they should not be used to estimate marginal effects (without further modification). Sometimes conditional effects are desirable, and such methods may be useful for them, but when marginal effects are the target of inference, it is critical not to inappropriately interpret estimates resulting from statistical methods aimed at estimating conditional effects as marginal effects. Although this issue is particularly salient with binary and survival outcomes due to the general noncollapsibility of the OR, RR, and HR, this can also occur with linear models for continuous outcomes or the RD.
The following methods estimate conditional effects for binary or survival outcomes (with noncollapsible effect measures) and should not be used to estimate marginal effects:
In addition, with continuous outcomes, conditional effects can be mistakenly interpreted as marginal effect estimates when treatment-covariate interactions are present in the outcome model. If the covariates are not centered at their mean in the target population (e.g., the treated group for the ATT, the full sample for the ATE, or the remaining matched sample for an ATM), the coefficient on treatment will not correspond to the marginal effect in the target population; it will correspond to the effect of treatment when the covariate values are equal to zero, which may not be meaningful or plausible. G-computation is always the safest way to estimate effects when including covariates in the outcome model, especially in the presence of treatment-covariate interactions.
#Generating data similar to Austin (2009) for demonstrating treatment effect estimation
gen_X <- function(n) {
X <- matrix(rnorm(9 * n), nrow = n, ncol = 9)
X[,5] <- as.numeric(X[,5] < .5)
X
}
#~20% treated
gen_A <- function(X) {
LP_A <- - 1.2 + log(2)*X[,1] - log(1.5)*X[,2] + log(2)*X[,4] - log(2.4)*X[,5] + log(2)*X[,7] - log(1.5)*X[,8]
P_A <- plogis(LP_A)
rbinom(nrow(X), 1, P_A)
}
# Continuous outcome
gen_Y_C <- function(A, X) {
2*A + 2*X[,1] + 2*X[,2] + 2*X[,3] + 1*X[,4] + 2*X[,5] + 1*X[,6] + rnorm(length(A), 0, 5)
}
#Conditional:
# MD: 2
#Marginal:
# MD: 2
# Binary outcome
gen_Y_B <- function(A, X) {
LP_B <- -2 + log(2.4)*A + log(2)*X[,1] + log(2)*X[,2] + log(2)*X[,3] + log(1.5)*X[,4] + log(2.4)*X[,5] + log(1.5)*X[,6]
P_B <- plogis(LP_B)
rbinom(length(A), 1, P_B)
}
#Conditional:
# OR: 2.4
# logOR: .875
#Marginal:
# RD: .144
# RR: 1.54
# logRR: .433
# OR: 1.92
# logOR .655
# Survival outcome
gen_Y_S <- function(A, X) {
LP_S <- -2 + log(2.4)*A + log(2)*X[,1] + log(2)*X[,2] + log(2)*X[,3] + log(1.5)*X[,4] + log(2.4)*X[,5] + log(1.5)*X[,6]
sqrt(-log(runif(length(A)))*2e4*exp(-LP_S))
}
#Conditional:
# HR: 2.4
# logHR: .875
#Marginal:
# HR: 1.57
# logHR: .452
set.seed(19599)
n <- 2000
X <- gen_X(n)
A <- gen_A(X)
Y_C <- gen_Y_C(A, X)
Y_B <- gen_Y_B(A, X)
Y_S <- gen_Y_S(A, X)
d <- data.frame(A, X, Y_C, Y_B, Y_S)
Because they are only appropriate with a large number of clusters, cluster-robust SEs are generally not used with subclassification methods. Regular robust SEs are valid with these methods when using the subclassification weights to estimate marginal effects.↩︎
Sometimes, an error will occur with this method, which usually means more bootstrap replications are required. The number of replicates must be greater than the original sample size when using the full bootstrap and greater than the number of pairs/strata when using the block bootstrap.↩︎
The matching weights are not necessary when performing 1:1 matching, but we include them here for generality. When weights are not necessary, including them does not affect the estimates. Because it may not always be clear when weights are required, we recommend always including them.↩︎
To verify that they are equal, supply the output of
avg_predictions()
to
hypotheses(), e.g.,
avg_predictions(…) |>
hypotheses(“revpairwise”); this explicitly compares the average potential outcomes and should yield identical estimates to the
avg_comparisons()`
call.↩︎
It is also known as fine stratification weighting, described by Desai et al. (2017).↩︎
We use quasibinomial()
instead of
binomial()
simply to avoid a spurious warning that can
occur with certain kinds of matching; the results will be identical
regardless.↩︎
Note that for low or high average expected risks
computed with avg_predictions()
, the confidence intervals
may go below 0 or above 1; this is because an approximation is used. To
avoid this problem, bootstrapping or simulation-based inference can be
used instead.↩︎
It is not immediately clear how to estimate a marginal HR when covariates are included in the outcome model; though Austin, Thomas, and Rubin (2020) describe several ways of including covariates in a model to estimate the marginal HR, they do not develop SEs and little research has been done on this method, so we will not present it here. Instead, we fit a simple Cox model with the treatment as the sole predictor.↩︎
For subclassification, only MMWS can be used; this is
done simply by including the stratification weights in the Cox model and
omitting the cluster
argument.↩︎
To use survey
to adjust for pair
membership, one can use the following code to specify the survey design
to be used with svyglm()
:
svydesign(ids = ~subclass, weights = ~weights, data = md)
where md
is the output of match.data()
. After
svyglm()
, comparisons()
can be used, and the
vcov
argument does not need to be specified.↩︎