Title stata.

com
mediate — Causal mediation analysis

Description Quick start Menu Syntax

Options Remarks and examples Stored results Methods and formulas
Acknowledgments References Also see

Description
mediate fits causal mediation models and estimates effects of a treatment on an outcome. The
treatment effect can occur both directly and indirectly through another variable, a mediator. The
outcome and mediator variables may be continuous, binary, or count. The treatment may be binary,
multivalued, or continuous. The estimated direct, indirect, and total effects have a causal interpretation
provided that assumptions pertaining to causal mediation models are met.

Quick start
Fit the mediation model with continuous outcome y1, continuous mediator m1, and categorical
treatment t1, and estimate the total effect, natural direct effect, and natural indirect effect
mediate (y1) (m1) (t1)
Same as above, but with covariates in both the outcome and the mediator equations
mediate (y1 x1 x2) (m1 x1 x3) (t1)
Same as above, but with probit model for binary outcome y2 and Poisson model for count mediator
m2
mediate (y2 x1 x2, probit) (m2 x1 x3, poisson) (t1)
Same as above, but estimate only the natural indirect effect (NIE)
mediate (y2 x1 x2, probit) (m2 x1 x3, poisson) (t1), nie
Same as above, but also estimate potential-outcome means
mediate (y2 x1 x2, probit) (m2 x1 x3, poisson) (t1), nie pomeans
Fit the mediation model with continuous treatment t2, and evaluate at values 0 and 4 of the treatment
with 0 as the control
mediate (y2 x1 x2, probit) (m2 x1 x3, poisson) (t2, continuous(0 4))

Menu
Statistics > Causal inference/treatment effects > Continuous outcomes > Causal mediation

Statistics > Causal inference/treatment effects > Binary outcomes > Causal mediation

Statistics > Causal inference/treatment effects > Count outcomes > Causal mediation

Statistics > Causal inference/treatment effects > Nonnegative outcomes > Causal mediation

1
 2 mediate — Causal mediation analysis

Syntax
 
mediate (ovar omvarlist, omodel noconstant )
 
(mvar mmvarlist, mmodel noconstant )
         
(tvar , continuous(numlist) ) if in weight , stat options

ovar is a continuous, binary, or count outcome of interest.

omvarlist specifies the covariates in the outcome model.
mvar is the mediator variable and may be continuous, binary, or count.
mmvarlist specifies the covariates in the mediator model.
tvar is the treatment variable and may be binary, multivalued, or continuous.

omodel Description
Model
linear linear model; the default
expmean exponential-mean model
logit logistic regression model
probit probit regression model
poisson Poisson model
omodel specifies the model for the outcome variable.

mmodel Description
Model
linear linear model; the default
expmean exponential-mean model
logit logistic regression model
probit probit regression model
poisson Poisson model
mmodel specifies the model for the mediator variable.
The logit outcome model may not be combined with the linear or expmean mediator model; probit rather than
logit may be used in these cases.
 mediate — Causal mediation analysis 3

stat Description
Stat
Pearl’s labeling of effects
nie natural indirect effect
nde natural direct effect
te total effect
pnie pure natural indirect effect
tnde total natural direct effect
ATE labeling of effects
aite average indirect treatment effect; synonym for nie
adte average direct treatment effect; synonym for nde
ate total average treatment effect; synonym for te
aitec average indirect treatment effect with respect to controls; synonym for pnie
adtet average direct treatment effect with respect to the treated; synonym for tnde
pomeans potential-outcome means
all all effects and potential-outcome means
Multiple effects may be specified; default is nie nde te.

options Description
Model
nointeraction exclude interaction of mediator and treatment
control(# | label) specify the level of tvar that is the control; default is first treatment level
SE/Robust
vce(vcetype) vcetype may be robust, cluster clustvar, bootstrap, or jackknife
nose do not estimate standard errors
Reporting
level(#) set confidence level; default is level(95)
ateterms use ATE terminology to label effects
aequations display auxiliary-equation results
nolegend suppress table legend
display options control columns and column formats, row spacing, line width,
display of omitted variables and base and empty cells, and
factor-variable labeling
Optimization
optimization options control the optimization process; seldom used
Advanced
force force estimation when the number of treatment groups exceeds 10
coeflegend display legend instead of statistics
 4 mediate — Causal mediation analysis

omvarlist and mmvarlist may contain factor variables; see [U] 11.4.3 Factor variables.
bootstrap, by, collect, jackknife, and statsby are allowed; see [U] 11.1.10 Prefix commands.
Weights are not allowed with the bootstrap prefix; see [R] bootstrap.
pweights, fweights, and iweights are allowed; see [U] 11.1.6 weight.
coeflegend does not appear in the dialog box.
See [U] 20 Estimation and postestimation commands for more capabilities of estimation commands.

Options

 Model
noconstant; see [R] Estimation options.
continuous(numlist) specifies that the treatment variable is continuous; numlist specifies the values
at which the potential-outcome means are to be evaluated, where the first value in the list is taken
as the control.
nointeraction excludes the interaction between the treatment and the mediator; by default, the
model includes the treatment–mediator interaction.
control(# | label) specifies the level of tvar that is the control. The default is the first treatment
level. You may specify the numeric level # (a nonnegative integer) or the label associated with
the numeric level. control() may not be specified with continuous treatments.

 Stat
stat specifies the statistics to be estimated. You may select from among five effects, each of which can
be labeled according to terminology used by Pearl and others or by ATE terminology. In addition
to effects, you may request that potential-outcome means be reported. The default is nie nde te.
stat may be one or more of the following:
stat Definition
nie natural indirect effect
nde natural direct effect
te total effect
pnie pure natural indirect effect
tnde total natural direct effect
aite average indirect treatment effect; synonym for nie
adte average direct treatment effect; synonym for nde
ate average treatment effect; synonym for te
aitec average indirect treatment effect with respect to controls; synonym for pnie
adtet average direct treatment effect with respect to the treated; synonym for tnde
pomeans potential-outcome means

all specifies that all effects and potential-outcome means be estimated; specifying all is
equivalent to specifying nie nde te pnie tnde pomeans. When option ateterms is
specified, all is equivalent to specifying aite adte ate aitec adtet pomeans.
 mediate — Causal mediation analysis 5


 SE/Robust
vce(vcetype) specifies the type of standard error reported, which includes types that are robust to
some kinds of misspecification (robust), that allow for intragroup correlation (cluster clustvar),
and that use bootstrap or jackknife methods (bootstrap, jackknife); see [R] vce option.
nose suppresses calculation of the variance–covariance matrix and standard errors.

 Reporting
level(#); see [R] Estimation options.
ateterms specifies that ATE terminology be used to label effects. ateterms is strictly a labeling
option. This option may not be specified on replay.
aequations specifies that the estimation results for the outcome model and the mediator model be
displayed. By default, they are not displayed.
nolegend suppresses the display of the table legend.
display options: noci, nopvalues, noomitted, vsquish, noemptycells, baselevels,
allbaselevels, nofvlabel, fvwrap(#), fvwrapon(style), cformat(% fmt), pformat(% fmt),
sformat(% fmt), and nolstretch; see [R] Estimation options.

 Optimization
optimization
  options: conv maxiter(), conv ptol(), conv vtol(), tracelevel(), and
no log. See [M-5] optimize( ).
conv maxiter(#) specifies the maximum number of iterations. The default is the number set
using set maxiter, which by default is 300.
conv ptol(#) specifies the convergence criteria for the parameters. The default is
conv ptol(1e-6).
conv vtol(#) specifies the convergence criteria for the gradient. The default is
conv vtol(1e-7).
tracelevel(tracelevel) allows you to display additional information about the iterative process
in the iteration log. tracelevel may be none, value, tolerance, step, params, or gradient.
See tracelevel in [M-5] optimize( ) for details.
log and nolog specify whether to display the iteration log. The iteration log is displayed by
default unless you used set iterlog off to suppress it; see set iterlog in [R] set iter.

 Advanced
force forces estimation when the number of treatment groups exceeds 10. By default, only 10 groups
are allowed for multivalued treatments. Do not use the force option if the treatment is continuous;
instead, use the continuous() option.

The following option is available with mediate but is not shown in the dialog box:
coeflegend; see [R] Estimation options.
 6 mediate — Causal mediation analysis

Remarks and examples stata.com

Remarks are presented under the following headings:
Introduction
Approaches to mediation analysis
Workflow for causal mediation
Forming research questions
Potential outcomes and effect decompositions
Evaluating assumptions for causal inference
Estimation of effects
Technical overview of causal mediation
Mediation analysis in the potential-outcomes framework
Total, direct, and indirect effects
Comparison of potential outcomes and classical mediation analysis
Accounting for treatment–mediator interaction
Assumptions for causal identification
Examples
Example 1: A simple causal mediation model
Example 2: Including covariates and relaxing the no-interaction assumption
Example 3: Referring to treatment effects using an alternative naming scheme
Example 4: Causal mediation model with a binary mediator
Example 5: Causal mediation model with a binary outcome
Example 6: Causal mediation model with a binary mediator and binary outcome
Example 7: Causal mediation model with a count mediator
Example 8: Causal mediation model with an exponential-mean outcome
Example 9: Causal mediation model with multivalued treatment
Example 10: Causal mediation model with continuous treatment
Example 11: Estimating controlled direct effects
Example 12: Estimating treatment effects on different scales

Introduction
Causal inference is an essential goal in many research areas and aims at identifying and quantifying
causal effects. For example, we might wish to find out whether physical exercise leads to an
improvement in self-perceived well-being, and if so, to what extent. Causality in this context typically
means that there is some cause T that has an effect on some outcome Y . We could visualize this
relation with a simple causal diagram:
T Y
Figure 1
If T is a measure of exercise and Y is well-being, then under certain assumptions, we could
use the above causal model to identify the total effect of exercise on well-being (by means of a
randomized controlled trial, for instance). However, a question that we cannot answer empirically
with our simple causal model is why exercise may increase well-being. Perhaps exercising causes
an increase in certain chemicals or hormones in the human body, which in turn affects perceptions
of well-being. To assess such intermediary effects, we need to expand our simple causal model by
adding variables that lie on the causal pathway between T and Y :
T M Y
Figure 2
Suppose that, in our exercise example, the variable M represents the production of a certain
chemical in the human body. With this new model, we now hypothesize that exercising leads to the
production of this chemical, which in turn leads to an increase in well-being. However, it might be
unrealistic to assume that the effect of exercise on well-being hinges exclusively on the production
 mediate — Causal mediation analysis 7

of that chemical. Perhaps we would like to allow for the possibility that exercise has an effect on
well-being beyond its path through the mediating variable, and so a better model might be
M

T Y
Figure 3
Here, we include a direct path from T to Y in addition to the indirect path of T to Y via M .
In other words, we assume that exercise produces a particular chemical that affects well-being, but
we also allow for the possibility of a direct effect of exercise on well-being that is not related to the
chemical. This is the classical mediation model that decomposes the total effect into a direct and an
indirect effect. Causal mediation analysis aims to identify these direct and indirect effects and give
them a causal interpretation.

Approaches to mediation analysis

Mediation analysis can be performed in a variety of ways. The classical approach of Baron and
Kenny (1986) fits two linear regression models, one for M and one for Y , and estimates direct, indirect,
and total effects as functions of the coefficients. Estimation can be simplified by fitting the models
for M and Y simultaneously via structural equation modeling as discussed in [SEM] Example 42g.
In Stata, you can use sem to fit linear models for the outcome and mediator, and you can then use
estat teffects to obtain a decomposition of direct and indirect effects based on the results from
sem. Note that this classical approach relies on the specification of a particular model at the outset
of the process.
Another approach to mediation analysis is based on the potential-outcomes framework. The potential
outcomes are values of the outcome that would be obtained under different conditions, such as when the
treatment occurs. Differences in potential outcomes yield direct, indirect, and total effects of interest.
This is the approach typically referred to as causal mediation analysis and is the one implemented in
mediate.
The causal mediation framework allows much flexibility. In this framework, it is common to allow
the mediator and the treatment to interact; thus, we do not assume that the effect of the mediator
on the outcome is the same for the treated and untreated groups. The total effect of the treatment
on the outcome can be decomposed into direct and indirect effects in two ways, and the researcher
can select the decomposition that matches his or her research question. The effects are defined in a
model-free manner, so the researcher can select an estimation method that is appropriate for his or
her data and then compute estimates of the effects of interest.
In the situation where both the outcome and the mediator are modeled using linear regression
and there is no treatment–mediator interaction, the classical approach and causal mediation via the
potential-outcomes framework will lead to the same results.

Workflow for causal mediation

The general workflow for researchers performing causal mediation analysis is as follows:
1. Specify your research question.
2. Identify the treatment, mediator, and outcome variables to be analyzed.
3. Determine which effect decomposition can be used to answer your research question.
 8 mediate — Causal mediation analysis

4. Evaluate whether assumptions for causal interpretation are appropriate.

5. Select a method for estimating the causal effects of interest.
6. Interpret the results.
In our introductory discussion, we provided an example of step 2 by using exercise, chemical
production, and well-being as the treatment, mediator, and outcome variables, respectively. Below,
we will provide a conceptual introduction to steps 1, 3, 4, and 5. In Examples, we will interpret the
results in different scenarios.

Forming research questions

Before performing causal mediation analysis, we must decide which research questions motivated
the desire to perform the analysis. Here are some types of research questions that may arise:
1. A scenario in which the primary interest is to determine whether there is an indirect effect and,
if so, to quantify it. In our example above, we might assume there will be some direct impact
of exercise on well-being, but we also wonder if and to what extent there is an indirect effect,
such as exercise increasing production of a chemical which in turn increases well-being. In this
case, we would be interested in decomposing the effects according to ATE decomposition 1.
2. A scenario in which the primary interest is to determine whether there is a direct effect and,
if so, to quantify it. Continuing with our example, perhaps we expect an indirect effect, but
we also wish to determine if there are any other ways in which exercise causes changes in
well-being. In this case, we would be interested in decomposing the effects according to ATE
decomposition 2.
3. A scenario in which the primary interest is to determine how the total effect can be decomposed
into direct and indirect effects, with focus remaining on all effects and not just direct or
just indirect effects. In our example, we simply want to explore the breakdown of the total
effect of exercise on well-being into all possible direct and indirect effects. In this case, we
would likely be interested in looking at both decompositions, ATE decomposition 1 and ATE
decomposition 2.
4. A scenario in which we want to determine the effect of the treatment on the outcome when
the mediator is set to a specific value. In our example, we might want to know the effect of
exercise on well-being for individuals whose level of this particular chemical is 10, which is the
mean value in the population. In this case, we would be interested in controlled direct effects.

Potential outcomes and effect decompositions

Below, we introduce statistics that may be of interest when performing causal mediation analysis.
Many of these statistics have a variety of names in the causal mediation literature. See, for instance,
Robins and Greenland (1992), VanderWeele (2015), and Pearl and MacKenzie (2018) for some of
the various terminology.
1. Potential-outcome means. These estimate the population-average value of the outcome that
would be expected if everyone was given the treatment (denoted here as Y [1, M (1)]) or if
everyone was given the control (denoted Y [0, M (0)]). In our example, Y [1, M (1)] is the
expected average well-being if everyone exercises, and Y [0, M (0)] is the expected average
well-being if no one exercises.
In addition, there are two cross-world potential outcomes. These are a bit less intuitive because
they correspond to situations that do not exist for any individual in the population. The first is
 mediate — Causal mediation analysis 9

the expected value of the outcome when everyone is treated but counterfactually experiences
the value of the mediator associated with being untreated (denoted Y [1, M (0)]). The second is
the expected value of the outcome when everyone is untreated but counterfactually experiences
the value of the mediator associated with being treated (denoted Y [0, M (1)]). In our example,
Y [1, M (0)] is the expected well-being if everyone was treated but experiencing the chemical
level as if untreated. Y [0, M (1)] is the expected value if everyone was untreated but experiencing
the chemical level as if treated.
The mediate command reports these potential-outcome means when the pomeans option is
specified.
2. Total effect (TE). This estimates the average difference in outcomes that we expect when
everyone receives the treatment versus when no one receives the treatment. In our case, it
estimates the improvement in well-being that we would expect if everyone exercises versus if
no one exercises.
The total effect is also referred to as the average treatment effect (ATE), the total average
treatment effect, or the marginal total effect.
The mediate command reports this statistic when the te option or its synonym ate is specified.
The total effect can be decomposed into direct and indirect effects in two ways when we allow
for a treatment–mediator interaction.
Decomposition 1. This decomposition separates the direct effect under the untreated mediator
condition from the total indirect effect. Nguyen, Schmid, and Stuart (2021) recommend using
this decomposition when a direct effect is assumed and the researcher is questioning whether a
mediation effect also exists. In our example, we would be interested in this decomposition if we
expect that exercise has a direct effect on well-being but want to determine whether a portion of
the total effect can be attributed to the increase in the chemical (and if so, how much of the total
effect is due to this mediation effect).
3. Natural direct effect (NDE). This estimates the average direct effect of the treatment on
the outcome when the mediator is held at its value associated with being untreated. It is the
difference Y [1, M (0)] − Y [0, M (0)].
This effect is sometimes referred to as the pure natural direct effect or the average direct
treatment effect (ADTE). All remaining effects of the treatment on the outcome are included in
the natural indirect effect.
The mediate command reports this statistic when the nde option or its synonym adte is
specified.
4. Natural indirect effect (NIE). This estimates the average indirect effect through a mediator. It
is the difference Y [1, M (1)] − Y [1, M (0)].
This effect is sometimes referred to as the total natural indirect effect, causal mediation effect,
or average indirect treatment effect (AITE).
The mediate command reports this statistic when the nie option or its synonym aite is
specified.
Decomposition 2. This decomposition separates the indirect effect under the untreated condition
from the total direct effect. Nguyen, Schmid, and Stuart (2021) recommend using this decomposition
when an indirect effect is assumed and the researcher is questioning whether a direct effect also
exists. In our example, we would be interested in this decomposition if we believe that exercise
increases production of the chemical which in turn increases well-being but want to determine if
there is also some change in well-being that is not caused by this mediation effect (and if so, how
much of the total effect is not due to the mediation effect).
 10 mediate — Causal mediation analysis

5. Pure natural indirect effect (PNIE). This estimates the average indirect effect of a mediator
under the untreated/control condition. It is the difference Y [0, M (1)] − Y [0, M (0)].
This is sometimes referred to as the average indirect treatment effect with respect to controls
(AITEC). All remaining effects of the treatment on the outcome are included in the total natural
direct effect.
The mediate command reports this statistic when the pnie option or its synonym aitec is
specified.
6. Total natural direct effect (TNDE). This estimates the average direct treatment effect when
the mediator is held at its value associated with being treated. It is the difference Y [1, M (1)] −
Y [0, M (1)].
This effect is sometimes referred to as the average direct treatment effect with respect to the
treated (ADTET).
The mediate command reports this statistic when the tnde option or its synonym adtet is
specified.
When no prior assumptions are made about the existence of direct and indirect effects, Nguyen,
Schmid, and Stuart (2021) recommend reporting both Decomposition 1 and Decomposition 2.
7. Controlled direct effects. These are the direct effects when the mediator is controlled by setting
it to a specific value. After fitting your model with mediate, you can estimate the average
controlled direct effect with the mediator set to your selected value by using estat cde; see
Example 11: Estimating controlled direct effects. In our well-being example, controlled direct
effects provide the direct effect of exercise on well-being when the chemical is assumed to be
a specific value.

Evaluating assumptions for causal inference

Before proceeding to estimation and interpretation of the effects of interest, we need to verify that
it is reasonable to give them a causal interpretation in our particular research context.
General assumptions for causal inference are discussed in [CAUSAL] Intro, and more precise
definitions in the context of mediation are provided in Assumptions for causal identification below.
To evaluate whether assumptions of causality are met for our mediation model, we must first
consider all potential variables, both observed and unobserved, that could affect the relationships
among our treatment, mediator, and outcome. If we anticipate that there are confounders (variables
that affect both an outcome and a predictor), we must determine whether these confounders will lead
to biased results in the estimation of effects from our mediation analysis. In particular, we want to
assume that
1. There is no unobserved confounding in the treatment–outcome relationship, and observed
confounders are included as covariates in the outcome model.
2. There is no unobserved confounding in the mediator–outcome relationship, and observed
confounders are included as covariates in the outcome model.
3. There is no unmeasured confounding in the treatment–mediator relationship, and observed
confounders are included as covariates in the mediator model.
4. There are no confounders in the mediator–outcome relationship that are caused by the treatment.
No variable exists that affects both the mediator and the outcome and that itself is caused by
the treatment.
 mediate — Causal mediation analysis 11

Estimation of effects

When assumptions are met, the mediate command can be used to estimate the causal parameters
of interest.
While the effects derived under the potential-outcomes framework required no particular model,
we now need to decide how to model our data to obtain estimates.
We first select models for the outcome and the mediator. Outcomes can be continuous, binary,
or counts and can be modeled using a linear, exponential-mean, logistic, probit, or Poisson model.
Mediators can also be continuous, binary, or counts and can also be modeled using a linear, exponential-
mean, logistic, probit, or Poisson model. Covariates can be included in the outcome and mediator
models. The treatment may be binary, categorical (multivalued), or continuous.
As a simple example of the mediate command, say that we have a binary outcome y, a continuous
mediator m, and a binary treatment t. We can fit a mediation model by typing
. mediate (y, probit) (m, linear) (t)

The first set of parentheses specifies a model for the outcome. The second set of parentheses specifies
the model for the mediator. The third set of parentheses defines the treatment. By default, the TE and
its decomposition into NDE and NIE are reported in the output.
If we would instead like to see the second type of decomposition, we can obtain the TE, PNIE,
and TNDE by typing
. mediate (y, probit) (m, linear) (t), te pnie tnde

Many combinations of models and effects can be obtained. See Examples below for additional
syntax examples as well as interpretation of the results.

Technical overview of causal mediation

Above, we provided a conceptual introduction to the concepts in causal mediation analysis. Here
we more formally define the potential-outcomes framework; explain total, direct, and indirect effects;
and introduce the assumptions necessary for causal inference. If you are familiar with the technical
aspects of causal mediation and are ready to see mediate demonstrated, go directly to Examples.

Mediation analysis in the potential-outcomes framework

The potential-outcomes framework is commonly employed for identifying causal effects. If we go

back to the model associated with figure 1 and assume T is a binary treatment, we can identify two
sets of potential outcomes, Yi (1) and Yi (0). Yi (t) is the outcome that would be realized if the ith
individual were exposed to treatment level t.
Consider a randomized experiment where the experimental group exercises while the control group
spends the same amount of time in a resting state. The outcome is subjective well-being that is
measured after exercising/resting. If it were possible to observe an individual in both states at the
same time, we would observe one outcome value under treatment, Yi (1), and one value under the
control condition, Yi (0). Then the treatment effect would be the difference τi = Yi (1) − Yi (0).
In other words, there is a potential outcome for each treatment level that could be administered.
Averaging the difference over all individuals in the sample would yield an estimate of the ATE
τ = E[Yi (1) − Yi (0)] = E[Yi (1)] − E[Yi (0)].
 12 mediate — Causal mediation analysis

However, it is not possible to observe the same individual under both conditions at the same
time; we can only observe one of these while the other is missing. If an individual is treated, we
observe Yi (1), and if not, we observe Yi (0). This has been coined the “fundamental problem of
causal inference” (Holland 1986). Much of the treatment effects and causal inference literature deals
with the question of how to estimate an ATE in the presence of this problem.
In a simple experiment where treatment is randomly assigned, the potential outcomes are independent
of treatment assignment and the missing potential outcomes are missing completely at random. In this
case, the average of the treatment group outcomes are a valid estimate of E[Yi (1)], and the average of
the control group outcomes are a valid estimate of E[Yi (0)]. Then τb = E[Yb i (1)] − E[Y
b i (0)] where
PNt
E[Yi (t)] = 1/Nt i=1 1(Ti = t)Yi is a valid estimator of the ATE. This estimation strategy follows
b
from the identification result that E[Yi (t)] = E(Yi |Ti = t) such that τ = E[Yi (1)] − E[Yi (0)] =
E[Yi |Ti = 1] − E[Yi |Ti = 0].
With observational rather than experimental data, however, the potential outcomes are not indepen-
dent of the treatment assignment process, and the causal effect is not identifiable without imposing
further assumptions such as conditional independence. Stata’s teffects suite of commands provides
a variety of estimators from this class of treatment-effects estimators.
For further information about identification and estimation in the context of causal models as well
as an overview of estimators implemented in Stata, see [CAUSAL] Intro. Here we focus on causal
inference and potential outcomes specifically for mediation analysis. In this situation, we have another
set of potential outcomes, Mi (1) and Mi (0), because M is also affected by the treatment. That is,
we can only observe Mi (1) for the group of individuals who were treated, and we can only observe
Mi (0) for the controls. If we let t denote the treatment level with respect to the outcome and let t0
be the treatment level with respect to the mediator, then the potential outcomes become Yi [t, Mi (t0 )].
Similar to the nonmediation case above, we can define a treatment effect as a difference between
potential outcomes. The treatment effect is identified if

E[Yi (t, Mi (t0 ))] = EMi |Ti =t0 E[Yi |Mi , Ti = t]

where EMi |Ti =t0 is the expectation of the mediator conditional on the treatment taking on the value
t0 and where E[Yi |Mi , Ti = t] is the expectation of the outcome conditional on the mediator and
treatment taking on the value t.

Total, direct, and indirect effects

In mediation analysis, we are interested not only in the total treatment effect but also in its
decomposition into direct effects and indirect effects.
Notice that if t = t0 for a given potential outcome, the resulting potential outcome is equivalent
to Yi (t). Assuming again a binary treatment, we have that

τ = E[Yi (1)] − E[Yi (0)] = E[Yi (1, Mi (1))] − E[Yi (0, Mi (0))]
In the context of mediation analysis, this treatment effect is also referred to as the total effect.
The total effect can be decomposed further into direct and indirect effects using contrasts between
potential-outcome means. The contrasts yielding direct and indirect effects use potential outcomes for
which t 6= t0 , which means we set the treatment level to t and set the mediator to its potential value
under treatment level t0 .
The natural indirect effect is then defined as

δ(t) ≡ E[Yi (t, Mi (1))] − E[Yi (t, Mi (0))], t ∈ {0, 1}

 mediate — Causal mediation analysis 13

Notice that here we “switch” the treatment from on to off in its effect on the mediator but keep the
treatment fixed at value t in its effect on the outcome. This natural indirect effect is also sometimes
referred to as the causal mediation effect (Imai, Keele, and Tingley 2010).
Likewise, the natural direct effect can be defined as

ζ(t) ≡ E[Yi (1, Mi (t))] − E[Yi (0, Mi (t))], t ∈ {0, 1}

Comparison of potential outcomes and classical mediation analysis

For those familiar with classical mediation analysis for linear models, it may be helpful to see
how the calculation of total, direct, and indirect effects in the potential-outcomes framework relates
to the classical product-of-coefficients approach.
We first write our mediation model corresponding with figure 3 as

Yi = β0 + β1 Mi + β2 Ti + i

Mi = α0 + α1 Ti + νi

where i and νi are uncorrelated error terms with means 0 and variances σ2 and σν2 , respectively.
Let’s consider the indirect effect δ(1). To calculate δ(1) in the potential-outcomes framework, we
need estimates for the potential-outcome means E[Yi (1, Mi (1))] and E[Yi (1, Mi (0))]. Intuitively,
what we want is a world where everyone in the population is exposed to the treatment, that is, Yi (1),
but where we can switch the treatment on and off in regard to the effect of the treatment on the
mediator, that is, Mi (1) and Mi (0). The difference when going from the treatment switched on to
the treatment switched off will inform us about the effect of the treatment on the outcome that goes
through the mediator. First, we write the above model in reduced form:

E[Yi |Mi , Ti ] = β0 + β1 (α0 + α1 Ti ) + β2 Ti

= β0 + β1 α0 + β1 α1 Ti + β2 Ti

This yields the conditional expectation E[Yi |Mi , Ti ] that we can observe from the data.
To obtain the potential-outcome means, we can modify the reduced-form model by replacing Mi
with the expectation of Mi that we would observe if Ti had taken on the value t0 for every unit in
the population. That is,

E[Yi (t, Mi (t0 ))] = β0 + β1 E[Mi (t0 )] + β2 t, t ∈ {0, 1}

Thus, to compute the potential-outcome mean E[Yi (1, Mi (1))], we must set the treatment Ti to 1 in
both the outcome and the mediator equations. In other words, we fix both t and t0 at 1:

E[Yi (1, Mi (1))] = β0 + β1 E[Mi (t0 )] + β2 t, t = t0 = 1

= β0 + β1 α0 + β1 α1 × 1 + β2 × 1
= β0 + β1 α0 + β1 α1 + β2
 14 mediate — Causal mediation analysis

However, to compute E[Yi (1, Mi (0))], we need to set treatment Ti to 1 in the outcome equation and
need to set it to 0 in the mediator equation. Specifically, we fix t0 = 0 and t = 1:

E[Yi (1, Mi (0))] = β0 + β1 E[Mi (t0 )] + β2 t, t = 1; t0 = 0

= β0 + β1 α0 + β1 α1 × 0 + β2 × 1
= β0 + β1 α0 + β2

Calculating the difference between these two potential-outcome means yields the indirect treatment
effect
δ(1) = (β0 + β1 α0 + β1 α1 + β2 ) − (β0 + β1 α0 + β2 )
= β0 + β1 α0 + β1 α1 + β2 − β0 − β1 α0 − β2
= β1 α1

In this case of a linear model, the indirect treatment effect is the product of the treatment coefficient
from the mediator equation and the mediator coefficient from the outcome equation. This is congruent
with the indirect effect definition in the product-of-coefficients method for mediation as proposed by
the classical mediation literature; see Baron and Kenny (1986).

Accounting for treatment–mediator interaction

Notice that the indirect effect we estimated above would be the same if we had estimated δ(0)
instead. Thus far, we assumed that the effect of the mediator on the outcome is the same for both
treatment groups. Presumably, a more realistic assumption would be to allow the mediator effects to
vary by treatment. This can be achieved by including a treatment–mediator interaction term.
When we allow an interaction, δ(0) 6= δ(1). Now we have two indirect effects, one with respect
to treatment [δ(1)] and one with respect to controls [δ(0)]. In the following, we will refer to δ(1)
as the NIE and to δ(0) as the PNIE.
To illustrate computation of the NIE under inclusion of a treatment–mediator interaction, we write
a new model
Yi = β0 + β1 Mi + β2 Ti + β3 Mi Ti + i

Mi = α0 + α1 Ti + νi
Here, NIE ≡ E[Yi (1, Mi (1)) − Yi (1, Mi (0))], whereas PNIE ≡ E[Yi (0, Mi (1)) − Yi (0, Mi (0))].
As before, to calculate NIE, we need potential-outcome means E[Yi (1, Mi (1))] and E[Yi (1, Mi (0))].
Writing the model in reduced form, we get

E[Yi |Mi , Ti ] = β0 + β2 Ti + β1 (α0 + α1 Ti ) + β3 Ti (α0 + α1 Ti )

= β0 + β2 Ti + (β1 + β3 Ti )(α0 + α1 Ti )
 mediate — Causal mediation analysis 15

Fixing the values for the treatment in both equations accordingly, we have potential-outcome means
E[Yi (1, Mi (1))] = β0 + β2 × 1 + (β1 + β3 × 1)(α0 + α1 × 1)
= β0 + β2 + (β1 + β3 )(α0 + α1 )
and
E[Yi (1, Mi (0))] = β0 + β2 × 1 + (β1 + β3 × 1)(α0 + α1 × 0)
= β0 + β2 + (β1 + β3 )α0

Taking the difference yields the NIE

E[Yi (1, Mi (1))] − E[Yi (1, Mi (0))] = (β1 + β3 )α1

We could proceed similarly for the other direct and indirect treatment effects. In this case
with treatment–mediator interaction, we also have two direct treatment effects. We have NDE ≡
E[Yi (1, Mi (0)) − Yi [0, Mi (0))] and TNDE ≡ E[Yi (1, Mi (1)) − Yi (0, Mi (1))].
Notice that both treatment-effect decompositions—NIE and NDE as well as PNIE and TNDE—sum
to the total treatment effect (or, as we will call it, the TE).
TE ≡ E[Yi (1, Mi (1)) − Yi (0, Mi (0))]
For further details and discussion on the different direct and indirect effects, as well as a discussion on
the differences between causal inference and traditional mediation approaches, see Nguyen, Schmid,
and Stuart (2021).

Assumptions for causal identification

The above discussion shows that the estimands of interest are the result of contrasts between potential-
outcome means, which are conditional expectations of the outcome with respect to counterfactuals in
both the outcome and the mediator equations. In other words, once we can estimate E[Yi (t, Mi (t0 ))],
we can estimate all direct and indirect treatment effects of interest.
The general form for causal mediation potential-outcome means, which includes covariates Xi , can
be written as the integral of the conditional expectation of the outcome with respect to the conditional
distribution of the mediator (see Imai, Keele, and Tingley [2010]):
Z
E[Yi (t, Mi (t ))|Xi = x] = E[Yi |Mi = m, Ti = t, Xi = x] dF [m|Ti = t0 , Xi = x]
0

This is a general, nonparametric solution that applies regardless of the underlying outcome and
mediator models. Stata’s mediate command uses analytical solutions of this integral for a variety
of parametric outcome and mediator model combinations. Also, while so far we assumed a binary
treatment for simplicity purposes, this approach generalizes straightforwardly to multivalued as well
as continuous treatments.
As is the case with nonmediation causal inference, there are assumptions to be met for the estimated
effects to be given a causal interpretation. Most notably, a crucial assumption in the nonmediation
case is the conditional independence assumption, also known as conditional ignorability assumption,
unconfoundedness, or selection on observables. This assumption states that potential outcomes are
independent of treatment assignment after conditioning on a set of observed covariates that affect both
the outcome and the selection into treatment (see Imbens [2004]). Intuitively, we have a model that
resembles an experiment once we account for observable characteristics. More formally, we have that
Yi (t) ⊥ Ti |Xi
 16 mediate — Causal mediation analysis

In the mediation case, however, we have an additional selection process because “selection” into
the mediator is also typically not based on random assignment. This leads to the following two
conditional independence assumptions:
{Yi [t, m], Mi (t0 )} ⊥ Ti |Xi = x
Yi [t, m] ⊥ Mi (t0 )|Ti = t0 , Xi = x
The first assumption states that treatment assignment is independent of potential outcomes and
potential mediators after conditioning on observed (pretreatment) covariates, or confounders. The
second assumption states that potential mediators are independent of the potential outcomes given
the observed treatment and observed (pretreatment) covariates. Because these assumptions are being
made sequentially, this has also been coined the sequential ignorability assumption (Imai, Keele, and
Tingley 2010).
Similarly, there is an additional overlap assumption with causal mediation models. In the nonmedi-
ation case, the overlap assumption states that each individual has a positive probability of receiving
each treatment:
0 < Pr(Ti = t|Xi = x), t ∈ {0, 1}
In the mediation case, the same principle applies to the mediator:
0 < p(Mi (t) = m|Ti = t, Xi = x), t ∈ {0, 1}

Finally, as is the case with nonmediation treatment-effects models, causal mediation models rely
on the stable unit treatment-value assumption, which states that potential outcomes do not depend
on treatments assigned to other individuals. For a detailed overview of effect identification and
assumptions for causal mediation analysis, see Nguyen et al. (2022).

Examples

Example 1: A simple causal mediation model

Suppose we wish to find out whether exercise affects perceptions of well-being among some
population of individuals. To the extent that there is such a causal relationship, we also wish to find
out why exercise affects well-being.
We have fictional data from a randomized controlled trial with individuals randomized into two
groups—one group performs physical exercise and the other group spends the same amount of time in
a resting state. Subjective well-being is measured before and after treatment sessions. In addition, the
level of the (fictional) hormone bonotonin is measured. The researchers wish to determine whether
exercise leads to an increase in bonotonin levels, which in turn has a positive effect on subjective
well-being. Here is an excerpt from our dataset:
. use https://www.stata-press.com/data/r18/wellbeing
(Fictional well-being data)
. list wellbeing bonotonin exercise age gender in 1/5, abbreviate(12)

wellbeing bonotonin exercise age gender

1. 71.73816 196.5467 Control 58 Male

2. 68.66573 195.8572 Exercise 38 Female
3. 71.05155 228.6035 Exercise 53 Female
4. 69.44469 206.6651 Exercise 44 Female
5. 75.62035 261.6855 Exercise 28 Female
 mediate — Causal mediation analysis 17

To estimate the treatment effects with mediate, we specify wellbeing as the outcome variable
in the first set of parentheses, bonotonin as the mediator variable in the second set of parentheses,
and exercise as the binary treatment variable in the third set of parentheses. Although inclusion of
a treatment–mediator interaction is commonly recommended, we specify the nointeraction option
here to omit the interaction and fit the simplest model possible.
. mediate (wellbeing) (bonotonin) (exercise), nointeraction
Iteration 0: EE criterion = 1.627e-25
Iteration 1: EE criterion = 3.061e-28
Causal mediation analysis Number of obs = 2,000
Outcome model: Linear
Mediator model: Linear
Mediator variable: bonotonin
Treatment type: Binary

Robust
wellbeing Coefficient std. err. z P>|z| [95% conf. interval]

NIE
exercise
(Exercise
vs
Control) 9.694617 .377312 25.69 0.000 8.955099 10.43413

NDE
exercise
(Exercise
vs
Control) 2.996658 .2109357 14.21 0.000 2.583231 3.410084

TE
exercise
(Exercise
vs
Control) 12.69127 .4005769 31.68 0.000 11.90616 13.47639

Note: Outcome equation does not include treatment--mediator interaction.

In the header of the output, we see that mediate fit linear models (the default) for both the
outcome and the mediator. Three treatment-effect estimates are reported in the table. TE is the total
effect of exercise on well-being and is estimated to be 12.7. The interpretation is the same as for the
ATE in the nonmediation case: if everyone in the population exercised, their well-being would be, on
average, 12.7 points higher than their well-being would be if no one exercised. The decomposition of
the TE into direct and indirect effects is of primary interest. The NIE is estimated to be 9.7, whereas
the NDE is estimated to be 3.0. These sum to the total effect of 12.7. The indirect effect is much
larger than the direct effect, indicating that the effect of exercise on well-being is largely due to
exercise affecting bonotonin levels, which in turn affect well-being. The direct effect of 3.0 is the
effect of exercise on well-being beyond the effect through bonotonin.
Instead of comparing estimates of the direct and indirect effects, we might ask what proportion of
the total effect is due to mediation. We can answer this question by using estat proportion.
 18 mediate — Causal mediation analysis

. estat proportion
Proportion mediated Number of obs = 2,000

Robust
wellbeing Proportion std. err. z P>|z| [95% conf. interval]

exercise
(exercise
vs
control) .7638805 .0154928 49.31 0.000 .7335151 .7942459

The indirect effect via bonotonin accounts for 76% of the effect of physical activity on well-being,
and the remaining 24% is due to other mechanisms.

Example 2: Including covariates and relaxing the no-interaction assumption

The previous example was somewhat unrealistic. For causal inference, we must evaluate the
potential of confounding. With causal mediation models, there are three types of confounders we
should consider: treatment–outcome confounders, treatment–mediator confounders, and mediator–
outcome confounders. A treatment–outcome confounder, for example, is a variable that affects both
the selection into treatment and the outcome. If confounders exist and we observe them in our data,
we can add them as covariates to the model to prevent biased results.
Above, we noted that the wellbeing data come from a randomized controlled trial. In this case,
we do not have to worry about treatment–outcome and treatment–mediator confounders because
treatment assignment is random. We do, however, need to consider variables such as age, gender,
and hstatus (a person’s health status) that affect both the mediator and the outcome. We include
these variables as covariates in the model for well-being. We also make our model a bit more realistic
by including baseline well-being in the outcome equation and baseline bonotonin level in the mediator
equation. In addition, we omit the nointeraction option to allow the bonotonin coefficients to vary
across treatment groups.
 mediate — Causal mediation analysis 19

. mediate (wellbeing age gender i.hstatus basewell)

> (bonotonin basebono)
> (exercise)
Iteration 0: EE criterion = 1.664e-25
Iteration 1: EE criterion = 1.192e-28
Causal mediation analysis Number of obs = 2,000
Outcome model: Linear
Mediator model: Linear
Mediator variable: bonotonin
Treatment type: Binary

Robust
wellbeing Coefficient std. err. z P>|z| [95% conf. interval]

NIE
exercise
(Exercise
vs
Control) 9.941404 .2307909 43.08 0.000 9.489062 10.39375

NDE
exercise
(Exercise
vs
Control) 3.08372 .1684778 18.30 0.000 2.753509 3.41393

TE
exercise
(Exercise
vs
Control) 13.02512 .2356989 55.26 0.000 12.56316 13.48709

Note: Outcome equation includes treatment--mediator interaction.

The interpretation of the treatment effects is the same as before. The total effect of exercise on
well-being is 13.0. Of this effect, 3.1 is attributed to the direct effect, while the remaining 9.9 is due
to the indirect path via bonotonin. These results are similar to our simpler model above.
20 mediate — Causal mediation analysis

We find that the expected effect of exercise on well-being is 13.0, but what is the expected well-
being when everyone exercises? When no one exercises? We can estimate four such potential-outcome
means by specifying the pomeans option:
. mediate (wellbeing age gender i.hstatus basewell)
> (bonotonin basebono)
> (exercise), pomeans
Iteration 0: EE criterion = 1.660e-25
Iteration 1: EE criterion = 1.473e-28
Causal mediation analysis Number of obs = 2,000
Outcome model: Linear
Mediator model: Linear
Mediator variable: bonotonin
Treatment type: Binary

Robust
wellbeing Coefficient std. err. z P>|z| [95% conf. interval]

POmeans
Y0M0 56.94195 .2300492 247.52 0.000 56.49107 57.39284
Y1M0 60.02567 .2571311 233.44 0.000 59.52171 60.52964
Y0M1 66.78952 .2642177 252.78 0.000 66.27167 67.30738
Y1M1 69.96708 .232508 300.92 0.000 69.51137 70.42278

Note: Outcome equation includes treatment--mediator interaction.

Y1M1 is an estimate of the potential-outcome mean E[Yi (1, Mi (1))]. If everyone in the popu-
lation exercised, we would expect the average of well-being to be around 70. The values labeled
Y1M0 and Y0M1 are estimates of the “cross-world” potential-outcome means E[Yi (1, Mi (0))] and
E[Yi (0, Mi (1))]. For these, we set different counterfactuals in the outcome and mediator equations.
In this case, the Y1M0 estimate tells us the expected average well-being if, for the outcome equation,
we assume that everyone in the population exercised, but we assume that no one exercised in regard
to the effect of treatment on the mediator. If we compare the Y1M1 and Y1M0 estimates, we imagine
a world where everyone received the treatment, except that the treatment is switched on and off in
its effect on the mediator. The difference between these is 69.96708 − 60.02567 = 9.94141, which
is our NIE reported above.
By default, the TE, NIE, and NDE are computed, but we can request specific effects. For example,
we could estimate only the NIE by typing
. mediate (wellbeing age gender i.hstatus basewell)
> (bonotonin basebono)
> (exercise), nie
 mediate — Causal mediation analysis 21

Alternatively, we could estimate all available effects and potential-outcome means at once by
specifying the all option:
. mediate (wellbeing age gender i.hstatus basewell)
> (bonotonin basebono)
> (exercise), all
Iteration 0: EE criterion = 1.668e-25
Iteration 1: EE criterion = 1.532e-28
Causal mediation analysis Number of obs = 2,000
Outcome model: Linear
Mediator model: Linear
Mediator variable: bonotonin
Treatment type: Binary

Robust
wellbeing Coefficient std. err. z P>|z| [95% conf. interval]

POmeans
Y0M0 56.94195 .2300492 247.52 0.000 56.49107 57.39284
Y1M0 60.02567 .2571311 233.44 0.000 59.52171 60.52964
Y0M1 66.78952 .2642177 252.78 0.000 66.27167 67.30738
Y1M1 69.96708 .232508 300.92 0.000 69.51137 70.42278

NIE
exercise
(Exercise
vs
Control) 9.941404 .2307909 43.08 0.000 9.489062 10.39375

NDE
exercise
(Exercise
vs
Control) 3.08372 .1684778 18.30 0.000 2.753509 3.41393

PNIE
exercise
(Exercise
vs
Control) 9.84757 .2318329 42.48 0.000 9.393186 10.30195

TNDE
exercise
(Exercise
vs
Control) 3.177554 .1800896 17.64 0.000 2.824585 3.530523

TE
exercise
(Exercise
vs
Control) 13.02512 .2356989 55.26 0.000 12.56316 13.48709

Note: Outcome equation includes treatment--mediator interaction.

Here we obtain estimates for two additional effects, PNIE and TNDE, which provide a different
decomposition of the TE into direct and indirect effects. In this case, PNIE and TNDE are similar to
NIE and NDE, respectively, because the coefficient on the treatment–mediator interaction term is quite
small in the model for well-being. We can see the results for the underlying models, including this
small coefficient of 0.002, if we add the aequations option to our mediate command.
 22 mediate — Causal mediation analysis

Example 3: Referring to treatment effects using an alternative naming scheme

The effects we have discussed so far are sometimes referred to by different names. The default
naming conventions originate in the works of Pearl and others. However, we can instead use terminology
more closely tied to ATEs if we specify the ateterms option:
. mediate (wellbeing age gender i.hstatus basewell)
> (bonotonin basebono)
> (exercise), all ateterms
Iteration 0: EE criterion = 1.668e-25
Iteration 1: EE criterion = 1.532e-28
Causal mediation analysis Number of obs = 2,000
Outcome model: Linear
Mediator model: Linear
Mediator variable: bonotonin
Treatment type: Binary

Robust
wellbeing Coefficient std. err. z P>|z| [95% conf. interval]

POmeans
Y0M0 56.94195 .2300492 247.52 0.000 56.49107 57.39284
Y1M0 60.02567 .2571311 233.44 0.000 59.52171 60.52964
Y0M1 66.78952 .2642177 252.78 0.000 66.27167 67.30738
Y1M1 69.96708 .232508 300.92 0.000 69.51137 70.42278

AITE
exercise
(Exercise
vs
Control) 9.941404 .2307909 43.08 0.000 9.489062 10.39375

ADTE
exercise
(Exercise
vs
Control) 3.08372 .1684778 18.30 0.000 2.753509 3.41393

AITEC
exercise
(Exercise
vs
Control) 9.84757 .2318329 42.48 0.000 9.393186 10.30195

ADTET
exercise
(Exercise
vs
Control) 3.177554 .1800896 17.64 0.000 2.824585 3.530523

ATE
exercise
(Exercise
vs
Control) 13.02512 .2356989 55.26 0.000 12.56316 13.48709

Note: Outcome equation includes treatment--mediator interaction.

Using this notation, ATE can be decomposed into AITE and ADTE or into AITEC and ADTET. Notice
that the estimates are the same as in the previous example; they now just have different names.
 mediate — Causal mediation analysis 23

Example 4: Causal mediation model with a binary mediator

In the previous examples, both outcome and mediator variables were continuous. We now look at
the case where the mediator variable is binary. To this end, we use the binary variable bbonotonin,
an indicator of higher bonotonin levels after exercise, where improvement is defined as an increase
of at least 10%. We could use a probit or a logit model for this mediator; we choose a logit model:
. mediate (wellbeing age gender i.hstatus basewell)
> (bbonotonin, logit)
> (exercise)
Iteration 0: EE criterion = 8.253e-18
Iteration 1: EE criterion = 8.223e-18 (backed up)
Causal mediation analysis Number of obs = 2,000
Outcome model: Linear
Mediator model: Logit
Mediator variable: bbonotonin
Treatment type: Binary

Robust
wellbeing Coefficient std. err. z P>|z| [95% conf. interval]

NIE
exercise
(Exercise
vs
Control) 4.41435 .4666635 9.46 0.000 3.499706 5.328994

NDE
exercise
(Exercise
vs
Control) 8.429238 .5696256 14.80 0.000 7.312792 9.545683

TE
exercise
(Exercise
vs
Control) 12.84359 .3712965 34.59 0.000 12.11586 13.57132

Note: Outcome equation includes treatment--mediator interaction.

Direct and indirect effect estimates differ from previous results because we used a different bonotonin
measure as our mediator variable. However, because we still have the continuous well-being outcome,
the interpretation of the effects is the same as before. Here we estimate a total effect of 12.8 with
direct and indirect effects of 8.4 and 4.4, respectively. That is, we expect an increase of 12.8 in
well-being due to treatment, of which 4.4 is due to an increase in bonotonin levels whereas the
remaining 8.4 is due to other mechanisms.
 24 mediate — Causal mediation analysis

Example 5: Causal mediation model with a binary outcome

Interpretation of effects did not change with a binary mediator, but interpretation does change
when we specify a different type of outcome.
To demonstrate, we return to the continuous mediator but use a binary outcome variable. The
outcome bwellbeing indicates higher well-being and is defined as an increase in well-being of at
least 10% compared with the baseline measurement. Using bwellbeing as the outcome variable and
specifying a probit outcome model, we get
. mediate (bwellbeing age gender i.hstatus, probit)
> (bonotonin basebono, linear)
> (exercise)
Iteration 0: EE criterion = 2.177e-25
Iteration 1: EE criterion = 9.730e-29
Causal mediation analysis Number of obs = 2,000
Outcome model: Probit
Mediator model: Linear
Mediator variable: bonotonin
Treatment type: Binary

Robust
bwellbeing Coefficient std. err. z P>|z| [95% conf. interval]

NIE
exercise
(Exercise
vs
Control) .2346259 .0145763 16.10 0.000 .2060568 .263195

NDE
exercise
(Exercise
vs
Control) .033732 .0237585 1.42 0.156 -.0128338 .0802978

TE
exercise
(Exercise
vs
Control) .2683579 .0200872 13.36 0.000 .2289877 .3077281

Note: Outcome equation includes treatment--mediator interaction.

We interpret the effects as expected differences measured on the probability scale, sometimes
referred to as risk differences. The TE of 0.27 indicates that if everyone in the population exercised,
we would expect the probability of increased well-being to be 0.27 higher than the probability of
increased well-being if no one exercised. In other words, the chance of experiencing an increase in
well-being goes up by 27 percentage points when exposed to the exercise treatment. We can see
that about 23 points are due to the indirect path via bonotonin, and about 3 points are due to other
mechanisms.
 mediate — Causal mediation analysis 25

Example 6: Causal mediation model with a binary mediator and binary outcome

We could also have the case where both the outcome and the mediator are binary. Here we use a
logit model for both:
. mediate (bwellbeing age gender i.hstatus, logit)
> (bbonotonin, logit)
> (exercise)
Iteration 0: EE criterion = 4.223e-16
Iteration 1: EE criterion = 2.107e-30
Causal mediation analysis Number of obs = 2,000
Outcome model: Logit
Mediator model: Logit
Mediator variable: bbonotonin
Treatment type: Binary

Robust
bwellbeing Coefficient std. err. z P>|z| [95% conf. interval]

NIE
exercise
(Exercise
vs
Control) .0959618 .0288699 3.32 0.001 .0393778 .1525457

NDE
exercise
(Exercise
vs
Control) .1676141 .0358902 4.67 0.000 .0972706 .2379577

TE
exercise
(Exercise
vs
Control) .2635759 .0212488 12.40 0.000 .221929 .3052228

Note: Outcome equation includes treatment--mediator interaction.

The interpretation is again in terms of differences in probabilities. We observe a TE of around

0.26, which is partially due to the indirect effect via bbonotonin (0.10) and partially due to other
mechanisms (0.17).

Example 7: Causal mediation model with a count mediator

We use a fictional dataset on birthweights and demonstrate how to perform causal mediation
analysis when using a Poisson model for a count mediator.
We now pretend to have observational data instead of experimental data. The sample includes
women who gave birth to a child. We wish to find out whether socioeconomic status and education of
the mother affects the child’s health. The outcome variable is the birthweight of the baby (bweight),
and the treatment variable is whether or not the mother has a college degree (college). The mediator
variable is the number of cigarettes smoked per day during pregnancy (ncigs). The hypothesis is
that women with a higher educational degree are likely to smoke fewer cigarettes and that smoking
during pregnancy has negative effects on birthweight. Here is an excerpt from the dataset:
26 mediate — Causal mediation analysis

. use https://www.stata-press.com/data/r18/birthweight, clear

(Fictional birthweight data)
. list bweight ncigs college ses sespar age in 1/5

bweight ncigs college ses sespar age

1. 3621 1 No 5.3581 3.308523 29

2. 3278 0 Yes 9.556957 4.376035 38
3. 3073 1 No 3.980829 6.580275 39
4. 3306 0 Yes 11.17643 12.12075 30
5. 4517 0 Yes 9.026146 4.738766 28

We fit a linear model for the outcome bweight and a Poisson model for the mediator ncigs,
and we specify college as the binary treatment variable. Because we have fully observational data,
where selection into treatment is no longer completely random, we have to be concerned about all
confounder types as mentioned in Example 2: Including covariates and relaxing the no-interaction
assumption. We specify several potential confounders as covariates in both equations.
We do not assume that the adverse effects of smoking are different between women with a college
degree and women without a college degree. Therefore, we use the nointeraction option.
. mediate (bweight sespar c.age##c.age)
> (ncigs sespar c.age##c.age, poisson)
> (college), nointeract
Iteration 0: EE criterion = 1.939e-21
Iteration 1: EE criterion = 1.937e-21 (backed up)
Causal mediation analysis Number of obs = 2,000
Outcome model: Linear
Mediator model: Poisson
Mediator variable: ncigs
Treatment type: Binary

Robust
bweight Coefficient std. err. z P>|z| [95% conf. interval]

NIE
college
(Yes vs No) 167.3075 21.36134 7.83 0.000 125.4401 209.175

NDE
college
(Yes vs No) 347.3375 34.44561 10.08 0.000 279.8253 414.8496

TE
college
(Yes vs No) 514.645 28.65043 17.96 0.000 458.4912 570.7988

Note: Outcome equation does not include treatment--mediator interaction.

As before, the type of model we use for the mediator does not affect the interpretation of the
estimated treatment effects. Effects are expected differences on the scale of the outcome variable. The
TE indicates that if all women had a college degree, the average birthweight of newborn babies would
be almost 515 grams higher than the average birthweight if no woman had a college degree. Of this
weight increase, around 167 grams are due to women with higher educational degrees smoking less,
while 347 grams are due to other mechanisms.
 mediate — Causal mediation analysis 27

Example 8: Causal mediation model with an exponential-mean outcome

Here we use an exponential-mean model for the outcome bweight.

. mediate (bweight sespar c.age##c.age, expmean)
> (ncigs sespar c.age##c.age, poisson)
> (college), nointeract
Iteration 0: EE criterion = 3.250e-13
Iteration 1: EE criterion = 1.159e-17
Causal mediation analysis Number of obs = 2,000
Outcome model: Exponential mean
Mediator model: Poisson
Mediator variable: ncigs
Treatment type: Binary

Robust
bweight Coefficient std. err. z P>|z| [95% conf. interval]

NIE
college
(Yes vs No) 198.978 23.53279 8.46 0.000 152.8546 245.1014

NDE
college
(Yes vs No) 320.3318 34.47792 9.29 0.000 252.7563 387.9072

TE
college
(Yes vs No) 519.3098 28.70435 18.09 0.000 463.0503 575.5693

Note: Outcome equation does not include treatment--mediator interaction.

Because we are still modeling a continuous outcome, the interpretation does not change. The TE is
about 519 grams, of which 199 grams are due to women with a college degree smoking less.

Example 9: Causal mediation model with multivalued treatment

So far we have only dealt with treatments that are binary. However, experiments often have more
than two treatment arms, or an observational treatment could consist of multiple categories. Then we
would refer to the treatment as multivalued.
To demonstrate, we return to our well-being data and use treatment variable mexercise, which
captures three treatment groups: a control group, a group where individuals exercised for 45 minutes,
and a group where individuals exercised for 90 minutes. Such a design would allow the researcher
to find out whether and how the duration of exercise affects bonotonin levels and thereby well-being.
Here we use a linear model for both the outcome and the mediator, and we include the multivalued
treatment mexercise as our treatment variable:
 28 mediate — Causal mediation analysis

. use https://www.stata-press.com/data/r18/wellbeing, clear

(Fictional well-being data)
. mediate (wellbeing age gender i.hstatus basewell)
> (bonotonin basebono)
> (mexercise)
Iteration 0: EE criterion = 1.697e-25
Iteration 1: EE criterion = 2.577e-26
Causal mediation analysis Number of obs = 2,000
Outcome model: Linear
Mediator model: Linear
Mediator variable: bonotonin
Treatment type: Multivalued

Robust
wellbeing Coefficient std. err. z P>|z| [95% conf. interval]

NIE
mexercise
(45 minutes
vs
Control) 5.128899 .3505171 14.63 0.000 4.441898 5.815899
(90 minutes
vs
Control) 9.780537 .2880877 33.95 0.000 9.215895 10.34518

NDE
mexercise
(45 minutes
vs
Control) 1.197498 .1750038 6.84 0.000 .8544965 1.540499
(90 minutes
vs
Control) 3.051084 .2071236 14.73 0.000 2.645129 3.457039

TE
mexercise
(45 minutes
vs
Control) 6.326396 .3894269 16.25 0.000 5.563134 7.089659
(90 minutes
vs
Control) 12.83162 .2967962 43.23 0.000 12.24991 13.41333

Note: Outcome equation includes treatment--mediator interaction.

We now have two effects per estimand because we compare the two treated groups to the control
group. Starting with the TE, we expect nearly a 13-point increase in well-being if everyone in the
population exercised for 90 minutes. Of these 13 points, around 10 points are due to the increase in
bonotonin levels and 3 points are due to other mechanisms. The results for the 45-minute treatment
arm, though expectedly smaller in magnitude, are interpreted similarly.

Example 10: Causal mediation model with continuous treatment

Instead of a binary or multivalued treatment, we could have a continuous treatment variable. With
continuous treatments, we have to specify at least two values, one to be the treatment and another to be
the control. We return to our birthweight data and use socioeconomic status (ses) as our continuous
treatment variable. Here are some summary statistics for ses:
 mediate — Causal mediation analysis 29

. use https://www.stata-press.com/data/r18/birthweight
(Fictional birthweight data)
. summarize ses
Variable Obs Mean Std. dev. Min Max

ses 2,000 7.804412 2.287496 1.304026 16.27844

We can see that ses ranges from around 1 to 16 and has a mean of about 8. These values, however,
do not tell us much because the variable is measured on an arbitrary scale. Therefore, we standardize
it so that the resulting variable has a mean of 0 and a standard deviation of 1:
. generate std_ses = (ses-r(mean))/r(sd)

We will use the new variable, std ses, as our treatment variable. We include the continuous()
option within the third set of parentheses where we define the treatment. This option tells mediate
to treat the variable as continuous and to use the values specified within the option as the control
and treatment points. The first value is the control, and the remaining values are treatments that are
compared with the control. Here we will specify one standard deviation below the mean as our control
value and one standard deviation above the mean as our treatment value:
. mediate (bweight sespar c.age##c.age, expmean)
> (ncigs sespar c.age##c.age, poisson)
> (std_ses, continuous(-1 1)), nointeract
Iteration 0: EE criterion = 1.470e-12
Iteration 1: EE criterion = 1.980e-17
Causal mediation analysis Number of obs = 2,000
Outcome model: Exponential mean
Mediator model: Poisson
Mediator variable: ncigs
Treatment type: Continuous
Continuous treatment levels:
0: std_ses = -1 (control)
1: std_ses = 1

Robust
bweight Coefficient std. err. z P>|z| [95% conf. interval]

NIE
std_ses
(1 vs 0) 171.3015 14.68778 11.66 0.000 142.514 200.089

NDE
std_ses
(1 vs 0) 170.0598 32.14841 5.29 0.000 107.05 233.0695

TE
std_ses
(1 vs 0) 341.3613 31.73741 10.76 0.000 279.1571 403.5655

Note: Outcome equation does not include treatment--mediator interaction.

Even though we used a continuous treatment variable, we interpret the results as before: if everyone
in the population had a socioeconomic status one standard deviation above the mean, the birthweight
of newborn children would be about 341 grams higher than the birthweight if everyone’s status value
is one standard deviation below the mean. Of these 341 grams, roughly half is due to women with a
higher status smoking less, and the other half is due to other mechanisms.
30 mediate — Causal mediation analysis

We could also evaluate the treatment effects at more than two values. Here we use the mean (0)
of the standardized variable as the base, and we evaluate the treatment effects at −2, −1, 1, and 2:
. mediate (bweight sespar c.age##c.age, expmean)
> (ncigs sespar c.age##c.age, poisson)
> (std_ses, continuous(0 -2 -1 1 2)), nointeract
Iteration 0: EE criterion = 1.470e-12
Iteration 1: EE criterion = 2.773e-17
Causal mediation analysis Number of obs = 2,000
Outcome model: Exponential mean
Mediator model: Poisson
Mediator variable: ncigs
Treatment type: Continuous
Continuous treatment levels:
0: std_ses = 0 (control)
1: std_ses = -2
2: std_ses = -1
3: std_ses = 1
4: std_ses = 2

Robust
bweight Coefficient std. err. z P>|z| [95% conf. interval]

NIE
std_ses
(1 vs 0) -276.2757 27.69004 -9.98 0.000 -330.5471 -222.0042
(2 vs 0) -100.1155 9.170566 -10.92 0.000 -118.0894 -82.14148
(3 vs 0) 65.84585 5.423096 12.14 0.000 55.21678 76.47493
(4 vs 0) 110.1346 8.724232 12.62 0.000 93.03538 127.2337

NDE
std_ses
(1 vs 0) -170.9012 31.33649 -5.45 0.000 -232.3196 -109.4828
(2 vs 0) -86.56069 16.08129 -5.38 0.000 -118.0794 -55.04193
(3 vs 0) 88.83929 16.94031 5.24 0.000 55.6369 122.0417
(4 vs 0) 180.0172 34.77372 5.18 0.000 111.8619 248.1724

TE
std_ses
(1 vs 0) -447.1769 35.41401 -12.63 0.000 -516.5871 -377.7667
(2 vs 0) -186.6761 15.73291 -11.87 0.000 -217.5121 -155.8402
(3 vs 0) 154.6851 16.31969 9.48 0.000 122.6991 186.6712
(4 vs 0) 290.1517 33.85571 8.57 0.000 223.7958 356.5077

Note: Outcome equation does not include treatment--mediator interaction.

 mediate — Causal mediation analysis 31

We now get four effect estimates for each treatment effect, which capture the expected differences
in the outcome with respect to the control point. With multiple effect estimates, it can be convenient
to plot the results. We use the postestimation command estat effectsplot to do so:
. estat effectsplot

Effects plot
400

200

0
NIE
Effect

NDE
TE
-200

-400

-600
-2 -1 0 1 2
std_ses

For more information about estat effectsplot, see [CAUSAL] mediate postestimation.

Example 11: Estimating controlled direct effects

Controlled direct effects (CDEs) are different from the other estimands we have dealt with so far.
Here, rather than having potential outcomes of the form Yi (t, Mi (t0 )), we have potential outcomes of
the form Yi (t|Mi = m). That is, we have potential outcomes for each treatment level that are evaluated
at set values of the mediator. Thus, CDEs only use the results of the outcome equation. Assuming a
binary treatment, the CDE for value m of the mediator is CDE(m) = Yi (1|Mi = m) − Yi (0|Mi = m).
CDEs can be estimated using the postestimation command estat cde.
32 mediate — Causal mediation analysis

To demonstrate, we begin by fitting a mediation model using the well-being data:

. use https://www.stata-press.com/data/r18/wellbeing, clear
(Fictional well-being data)
. mediate (bwellbeing age gender i.hstatus, probit)
> (bbonotonin, probit)
> (exercise)
Iteration 0: EE criterion = 4.326e-19
Iteration 1: EE criterion = 5.424e-32
Causal mediation analysis Number of obs = 2,000
Outcome model: Probit
Mediator model: Probit
Mediator variable: bbonotonin
Treatment type: Binary

Robust
bwellbeing Coefficient std. err. z P>|z| [95% conf. interval]

NIE
exercise
(Exercise
vs
Control) .0962832 .0288905 3.33 0.001 .0396588 .1529076

NDE
exercise
(Exercise
vs
Control) .1672304 .0358936 4.66 0.000 .0968803 .2375805

TE
exercise
(Exercise
vs
Control) .2635137 .0212346 12.41 0.000 .2218946 .3051327

Note: Outcome equation includes treatment--mediator interaction.

We fit probit models for the outcome and the mediator, but the type of model is not important here;
we can use estat cde after any model fit with mediate.
What is the TE if everyone in the population has a mediator value of 0 (no improvement in
bonotonin levels)? To find out, we estimate CDE(0) by specifying the mvalue(0) option with estat
cde:
. estat cde, mvalue(0)
Controlled direct effect Number of obs = 2,000
Mediator variable: bbonotonin
Mediator value = 0

Delta-method
CDE std. err. z P>|z| [95% conf. interval]

exercise
(Exercise
vs
Control) .1605355 .039731 4.04 0.000 .0826641 .2384068
 mediate — Causal mediation analysis 33

This CDE is around 0.16. The probability of increased well-being when everyone exercises is 0.16
higher than the probability of increased well-being when no one exercises, provided that no one in
the population had at least a 10% increase in bonotonin levels.
We could perform the same analysis specifying multiple values for the mediator. Here we wish to
estimate both CDE(0) and CDE(1):
. estat cde, mvalue(0 1)
Controlled direct effect Number of obs = 2,000
Mediator variable: bbonotonin
Mediator values:
1._at: bbonotonin = 0
2._at: bbonotonin = 1

Delta-method
CDE std. err. z P>|z| [95% conf. interval]

exercise@_at
(Exercise
vs
Control)
1 .1605355 .039731 4.04 0.000 .0826641 .2384068
(Exercise
vs
Control)
2 .2224479 .0493025 4.51 0.000 .1258166 .3190791

If we “switch on” the mediator, the CDE is higher by around 0.06 points. We could also estimate this
difference directly by using the contrast option:
. estat cde, mvalue(0 1) contrast
Controlled direct effect Number of obs = 2,000
Mediator variable: bbonotonin
Mediator values:
1._at: bbonotonin = 0
2._at: bbonotonin = 1

Delta-method
CDE std. err. z P>|z| [95% conf. interval]

_at#exercise
(2 vs 1)
(Exercise
vs
Control) .0619124 .0630241 0.98 0.326 -.0616126 .1854373

See [CAUSAL] mediate postestimation for further information about estat cde.

Example 12: Estimating treatment effects on different scales

The mediate command estimates treatment effects on the natural scale of the outcome variable.
However, some researchers may want to present their estimated effects on a different scale such as
on the odds-ratio or risk-ratio scale if the outcome variable is binary or on the incidence-rate–ratio
scale if the outcome variable is a count. The postestimation commands estat rr, estat or, and
estat irr transform estimated treatment effects onto these different scales.
34 mediate — Causal mediation analysis

To see how this works, we first fit the following model with a binary outcome variable. We could
use a probit or logit model for the outcome variable; here we are using probit:
. mediate (bwellbeing age gender i.hstatus, probit)
> (bbonotonin, probit)
> (exercise), all
Iteration 0: EE criterion = 4.326e-19
Iteration 1: EE criterion = 5.886e-32
Causal mediation analysis Number of obs = 2,000
Outcome model: Probit
Mediator model: Probit
Mediator variable: bbonotonin
Treatment type: Binary

Robust
bwellbeing Coefficient std. err. z P>|z| [95% conf. interval]

POmeans
Y0M0 .3014153 .0146737 20.54 0.000 .2726554 .3301752
Y1M0 .4686457 .0328309 14.27 0.000 .4042984 .5329931
Y0M1 .3536121 .0381972 9.26 0.000 .2787469 .4284773
Y1M1 .5649289 .0154435 36.58 0.000 .5346603 .5951976

NIE
exercise
(Exercise
vs
Control) .0962832 .0288905 3.33 0.001 .0396588 .1529076

NDE
exercise
(Exercise
vs
Control) .1672304 .0358936 4.66 0.000 .0968803 .2375805

PNIE
exercise
(Exercise
vs
Control) .0521968 .0348642 1.50 0.134 -.0161357 .1205293

TNDE
exercise
(Exercise
vs
Control) .2113169 .041136 5.14 0.000 .1306917 .291942

TE
exercise
(Exercise
vs
Control) .2635137 .0212346 12.41 0.000 .2218946 .3051327

Note: Outcome equation includes treatment--mediator interaction.

 mediate — Causal mediation analysis 35

Given that our outcome variable is binary and our outcome model is probit, the potential-outcome
means are averaged probabilities, and because the treatment effects are differences between potential-
outcome means, the estimated effects can be interpreted as risk differences. For example, the natural
indirect effect is the difference between potential-outcome means Y1M1 and Y1M0:
. display _b[POmeans:Y1M1]-_b[POmeans:Y1M0]
.09628322

If, instead of interpreting the effect on the risk-difference scale, we wanted to interpret it on the
risk-ratio scale, we could simply compute the ratio of the potential-outcome means:
. display _b[POmeans:Y1M1]/_b[POmeans:Y1M0]
1.2054499

This is what estat rr is doing:

. estat rr
Transformed treatment effects Number of obs = 2,000

Robust
bwellbeing Risk ratio std. err. z P>|z| [95% conf. interval]

NIE
exercise
(Exercise
vs
Control) 1.20545 .0746555 3.02 0.003 1.06766 1.361023

NDE
exercise
(Exercise
vs
Control) 1.554817 .132328 5.19 0.000 1.315937 1.837062

PNIE
exercise
(Exercise
vs
Control) 1.173172 .1157431 1.62 0.105 .966905 1.423442

TNDE
exercise
(Exercise
vs
Control) 1.597595 .1778207 4.21 0.000 1.284469 1.987055

TE
exercise
(Exercise
vs
Control) 1.874254 .1043578 11.28 0.000 1.680482 2.09037

The total treatment effect is now decomposed into multiplicative components NIE and NDE as well as
PNIE and TNDE. That is, taking their product, rather than their sum, will yield the total effect.
36 mediate — Causal mediation analysis

Similarly, we can express all effects on the odds-ratio scale by using estat or:
. estat or
Transformed treatment effects Number of obs = 2,000

Robust
bwellbeing Odds ratio std. err. z P>|z| [95% conf. interval]

NIE
exercise
(Exercise
vs
Control) 1.472222 .1708025 3.33 0.001 1.172788 1.848106

NDE
exercise
(Exercise
vs
Control) 2.044157 .3042049 4.80 0.000 1.527008 2.736449

PNIE
exercise
(Exercise
vs
Control) 1.267908 .1933291 1.56 0.120 .9403671 1.709534

TNDE
exercise
(Exercise
vs
Control) 2.373558 .4231302 4.85 0.000 1.673622 3.366217

TE
exercise
(Exercise
vs
Control) 3.009452 .281479 11.78 0.000 2.505377 3.614945

Here the total average treatment effect is 3 on the odds-ratio scale and is composed of odds ratios
1.47 and 2.04 in regard to NIE and NDE, respectively, and of odds ratios 1.27 and 2.37 in regard to PNIE
and TNDE. Typically, the treatment effects can be interpreted more intuitively on the risk-difference
scale, but there may be applications where transforming them to the risk-ratio or odds-ratio scale is
desirable.
Notice that estat rr, estat or, and estat irr require estimation of potential-outcome means
with mediate. If the fitted model does not contain potential-outcome mean estimates, these estat
commands will refit the model. The reestimation does not affect the results, but computation takes
longer. See [CAUSAL] mediate postestimation for further information about estat rr, estat or,
and estat irr.
 mediate — Causal mediation analysis 37

Stored results
mediate stores the following in e():
Scalars
e(N) number of observations
e(N clust) number of clusters
e(k eq) number of equations in e(b)
e(k levels) number of levels in treatment variable
e(rank) rank of e(V)
e(interact) 1 if treatment–mediator interaction included, 0 otherwise
e(converged) 1 if converged, 0 otherwise
Macros
e(cmd) mediate
e(cmdline) command as typed
e(depvar) name of outcome variable
e(mvar) name of mediator variable
e(tvar) name of treatment variable
e(omodel) linear, logit, probit, poisson, or expmean
e(mmodel) linear, logit, probit, poisson, or expmean
e(wtype) weight type
e(wexp) weight expression
e(title) title in estimation output
e(clustvar) name of cluster variable
e(tlevels) levels of treatment variable
e(tvartype) binary, multivalued, or continuous
e(control) control level
e(vce) vcetype specified in vce()
e(vcetype) title used to label Std. err.
e(properties) b V
e(estat cmd) program used to implement estat
e(predict) program used to implement predict
e(marginsnotok) predictions disallowed by margins
Matrices
e(b) coefficient vector
e(V) variance–covariance matrix of the estimators
Functions
e(sample) marks estimation sample

In addition to the above, the following is stored in r():

Matrices
r(table) matrix containing the coefficients with their standard errors, test statistics, p-values,
and confidence intervals

Note that results stored in r() are updated when the command is replayed and will be replaced when
any r-class command is run after the estimation command.
 38 mediate — Causal mediation analysis

Methods and formulas

mediate fits causal mediation models and estimates direct, indirect, and total treatment effects.
Using the potential-outcomes framework, the estimated treatment effects are the result of contrasts
between potential-outcome means. Without loss of generality, let Ti be a binary treatment, t ∈ {0, 1},
for observations i = 1, . . . , N , and let Yi be the outcome and Mi be the mediator variable. The
potential-outcome means are
0
POMt,t0 ≡ E[Yi (t, Mi (t ))]

The treatment effects are then defined as follows:

NIE ≡ E[Yi (1, Mi (1)) − Yi (1, Mi (0))]

NDE ≡ E[Yi (1, Mi (0)) − Yi (0, Mi (0))]
PNIE ≡ E[Yi (0, Mi (1)) − Yi (0, Mi (0))]
TNDE ≡ E[Yi (1, Mi (1)) − Yi (0, Mi (1))]
TE ≡ E[Yi (1, Mi (1)) − Yi (0, Mi (0))]

Synonyms for NIE, NDE, PNIE, TNDE, and TE are AITE, ADTE, AITEC, ADTET, and ATE, respectively.
The potential-outcome means are the result of an integral of the conditional expectation of the
outcome with respect to the conditional distribution of the mediator (Imai, Keele, and Tingley 2010):
Z
E[Yi (t, Mi (t0 ))|Xi = x] = E[Yi |Mi = m, Ti = t, Xi = x] dF [m|Ti = t0 , Xi = x] (1)

They are estimated as the sample average

N
1 X
POM
d t,t0 = Yi (t, Mi (t0 ))|Xi = x
N i=1

The estimated treatment effects are then the result of differences between estimated potential-outcome
means.
mediate uses analytical solutions for the integral in (1) for a variety of parametric outcome and
mediator model combinations. Let Xi = {Wi , Zi }, the index function of the outcome model is

ηiY = β0 + β1 Ti + β2 Mi + β3 Ti Mi + Wi γ (2)

and that of the mediator model is

ηiM = α0 + α1 Ti + Zi ζ (3)

where Wi and Zi are potentially overlapping sets of covariates. If the nointeraction option is
used, ηiY reduces to the simpler function where β3 = 0. Depending on which model is specified, the
expected values of the outcome and mediator follow these functional forms:

Model Link function

linear ηi
exponential mean eηi
Poisson eηi
logit Π(ηi )
probit Φ(ηi )
 mediate — Causal mediation analysis 39

Π and Φ are the cumulative logistic and cumulative normal distribution functions, respectively.
Between the outcome and mediator models, all combinations of the above functional forms are
allowed with the exception of logit outcome models in combination with linear or exponential-mean
mediator models.
mediate uses the estimated coefficients from (2) and (3) to estimate POM d t,t0 . Calculation of
d t,t0 depends on the combination of functional forms of the outcome and mediator models. We
POM
define the following terms, where t represents counterfactual values for the treatment with respect to
the outcome equation and t0 represents treatment counterfactuals in regard to the mediator equation:

ν t = β0 + β1 t + W i γ
ξt0 = α0 + α1 t0 + Zi ζ
κt = β2 + β3 t

If the outcome model is linear, we have

E[Yi (t, Mi (t0 ))] = νt + Θm (ξt0 )κt

where Θm (·) denotes the identity function if the mediator model is linear, the cumulative normal
distribution function if probit, the cumulative logistic distribution function if logit, and the exponential
function if exponential mean or Poisson. In this case, exponential mean and Poisson are synonyms
when specifying the mediator model; notice, though, that this is not necessarily the case for other
model combinations.
If the outcome model is probit and the mediator model is linear or exponential mean, we have
" #
0 νt + Γm (ξt0 )κt
E[Yi (t, Mi (t ))] = Φ p
1 + κ2t σm
2

where Γm (·) denotes the identity function if the mediator model is linear and denotes the exponential
2
function if it is exponential mean, and σm is the error variance pertaining to the mediator model.
For probit and logit outcome models in combination with probit and logit mediator models, the
potential outcomes are

E[Yi (t, Mi (t0 ))] = Λy (νt + κt )Λm (ξt0 ) + Λy (νt ){1 − Λm (ξt0 )}

where Λy (·) is the cumulative normal distribution function if the outcome model is probit and the
cumulative logistic distribution function if the outcome model is logit. Λm (·) denotes the cumulative
normal distribution function if the mediator model is probit and the cumulative logistic distribution
function if the mediator model is logit.
Regarding the outcome model, notice that, unlike the case of the mediator model, Poisson and
exponential mean always refer to the same model. Thus, we can use the terms Poisson and exponential
mean interchangeably in regard to the outcome model. The potential outcomes in the case of the
exponential-mean outcome model and the linear or exponential-mean mediator model are
2 2
E[Yi (t, Mi (t0 ))] = eνt +κt Γm (ξt0 )+(κt σm )/2

where Γm (·) is the identity function if the mediator model is linear and is the exponential function if
the mediator model is exponential mean. For probit and logit mediator models, the potential outcomes
are
E[Yi (t, Mi (t0 ))] = Λm (ξt0 )eνt +κt + {1 − Λm (ξt0 )}eνt
40 mediate — Causal mediation analysis

If the outcome model is exponential mean, probit, or logit, and if the mediator model is Poisson,
the potential outcomes are
K ξ 0
X ejξt0 e−e t

E[Yi (t, Mi (t0 ))] = Ψy (νt + κt j)

j=0
j!
where j indexes the counts of the mediator variable, and Ψy (·) denotes the exponential function if
the outcome model is exponential mean, denotes the cumulative normal distribution function if the
outcome model is probit, and denotes the cumulative logistic distribution function if the outcome
model is logit.
mediate uses a method of moments estimator, also known as an estimating equations estimator,
to estimate all auxiliary and effect parameters as well as their variance–covariance matrix. For more
information about the underlying gmm command, see [R] gmm.
The postestimation commands estat or, estat rr, and estat irr calculate treatment effects on
different scales. If the outcome variable is binary and the model for the outcome variable is logit
or probit, estat or computes marginal treatment effects on the odds-ratio scale, whereas estat
rr computes marginal treatment effects on the risk-ratio scale. If the model for the outcome variable
is poisson or expmean, estat irr computes marginal treatment effects on the incidence-rate–
ratio scale. Let YtMt0 be a shorthand for E[Yi (t, Mi (t0 ))]; the treatment effects on risk-ratio and
incidence-rate–ratio scales are ratios of potential-outcome means:
RR
NIE ≡ Y1M1 /Y1M0
RR
NDE ≡ Y1M0 /Y0M0
RR
PNIE ≡ Y0M1 /Y0M0
RR
TNDE ≡ Y1M1 /Y0M1
RR
TE ≡ Y1M1 /Y0M0
For logit and probit outcome models, YtMt0 are probabilities, and so the treatment effects on odds-ratio
scale are
OR
NIE ≡ Y1M1 /(1 − Y1M1 )/{Y1M0 /(1 − Y1M0 )}
OR
NDE ≡ Y1M0 /(1 − Y1M0 )/{Y0M0 /(1 − Y0M0 )}
OR
PNIE ≡ Y0M1 /(1 − Y0M1 )/{Y0M0 /(1 − Y0M0 )}
OR
TNDE ≡ Y1M1 /(1 − Y1M1 )/{Y0M1 /(1 − Y0M1 )}
OR
TE ≡ Y1M1 /(1 − Y1M1 )/{Y0M0 /(1 − Y0M0 )}
Notice that, with all three of these scales, the total effect is the product of direct and indirect effects,
rather than their sum.
CDEs use only the results of the outcome equation. Rather than having potential outcomes of the
form Yi (t, Mi (t0 )), here we have potential outcomes Yi (t|Mi = m). That is, we have potential
outcomes for each treatment level t that are evaluated at value m of the mediator. CDE(m) then is
the average of the differences between potential outcomes. For binary treatment, CDE(m) is defined
as Yi (1|Mi = m) − Yi (0|Mi = m). Letting Ytm be a shorthand for Yi (t|Mi = m), we have that
CDE(m) ≡ Y1m − Y0m
RR
CDE(m) ≡ Y1m /Y0m
IRR
CDE(m) ≡ Y1m /Y0m
OR
CDE(m) ≡ Y1m /(1 − Y1m )/{Y0m /(1 − Y0m )}
 mediate — Causal mediation analysis 41

Acknowledgments
Stata has an active research community adding features to the area of causal mediation. We would
like to acknowledge their previous and ongoing contributions to the area: paramed by Hanhua Liu,
Richard Emsley, Graham Dunn, Tyler VanderWeele, and Linda Valeri; medeff by Raymond Hicks
and Dustin Tingley; rwrmed by Ariel Linden, Chuck Huber, and Geoffrey T. Wodtke; and many
more. Type search casual mediation to see Stata’s official and community-contributed features
for causal mediation.

References
Baron, R. M., and D. A. Kenny. 1986. The moderator–mediator variable distinction in social psychological research:
Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology 51: 1173–1182.
https://doi.org/10.1037//0022-3514.51.6.1173.
Holland, P. W. 1986. Statistics and causal inference. Journal of the American Statistical Association 81: 945–960.
https://doi.org/10.2307/2289064.
Imai, K., L. Keele, and D. Tingley. 2010. A general approach to causal mediation analysis. Psychological Methods
15: 309–334. https://psycnet.apa.org/doi/10.1037/a0020761.
Imbens, G. W. 2004. Nonparametric estimation of average treatment effects under exogeneity: A review. Review of
Economics and Statistics 86: 4–29. https://doi.org/10.1162/003465304323023651.
Nguyen, T. Q., I. Schmid, E. L. Ogburn, and E. A. Stuart. 2022. Clarifying causal mediation analysis: Effect
identification via three assumptions and five potential outcomes. Journal of Causal Inference 10: 246–279.
https://doi.org/10.1515/jci-2021-0049.
Nguyen, T. Q., I. Schmid, and E. A. Stuart. 2021. Clarifying causal mediation analysis for the applied researcher:
Defining effects based on what we want to learn. Psychological Methods 26: 255–271.
https://psycnet.apa.org/doi/10.1037/met0000299.
Pearl, J., and D. MacKenzie. 2018. The Book of Why: The New Science of Cause and Effect. New York: Basic
Books.
Robins, J. M., and S. Greenland. 1992. Identifiability and exchangeability for direct and indirect effects. Epidemiology
3: 143–155. https://doi.org/10.1097/00001648-199203000-00013.
VanderWeele, T. J. 2015. Explanation in Causal Inference: Methods for Mediation and Interaction. New York: Oxford
University Press.

Also see
[CAUSAL] mediate postestimation — Postestimation tools for mediate
[CAUSAL] teffects — Treatment-effects estimation for observational data
[CAUSAL] teffects ra — Regression adjustment
[SEM] sem — Structural equation model estimation command
[U] 20 Estimation and postestimation commands

Stata, Stata Press, and Mata are registered trademarks of StataCorp LLC. Stata and
®
Stata Press are registered trademarks with the World Intellectual Property Organization
of the United Nations. Other brand and product names are registered trademarks or
trademarks of their respective companies. Copyright c 1985–2023 StataCorp LLC,
College Station, TX, USA. All rights reserved.