Interpreting Model Estimates: Marginal Effects

Marcelo Coca Perraillon

[email protected]

University of Colorado
Anschutz Medical Campus
—

Nova School of Business and Economics

Lisbon, Portugal
July, 2019
Updated lectures are here: https://clas.ucdenver.edu/marcelo-perraillon/code-
and-topics/marginal-effects
© Marcelo Coca Perraillon, 2021. See link above for entire copyright notice.
1
These notes are part of a forthcoming book to be published by
Cambridge University Press. For more information and additional
datasets, examples, and sample chapters, go to

https://www.perraillon.com/PLH/

Please note copyright notice here and on website.

Outline

Why do we need marginal effects?

Derivatives to interpret more complex models
Derivatives definition, two ways
A collection of terms
1 Average Marginal Effects (AME)
2 Marginal Effect at the Mean (MEM)
3 Marginal Effects at Representative values (MER)
Interactions
Marginal effects for two-part models

2
Why do we need marginal effects?

In a simple linear model, say, y = β0 + β1 age + β2 male, we can easily

interpret the coefficients
It is less straightforward when there are non-linear terms, for example:
y = β0 + β1 age + β2 age 2 + β3 male
Using the analytical derivative makes interpretation a lot easier:
∂E [y |age,male]
∂age = β1 + 2β2 age
No single effect of age, the effect depends on age; that is, one effect
at age 20, another at 50, etc. We just plug in numbers for age in the
above expression to get the effect at different ages
Note that effects are not the same as predictions

3
Why do we need marginal effects?

With interactions, it’s more complicated:

y = β0 + β1 age + β2 male + β3 male ∗ age
But similar in the sense that the effect of age now depends on sex; or
the other way around, the effect of sex depends on age
With interactions, taking the derivative still helps with interpretation:
∂E [y |age,male]
∂age = β1 + β3 ∗ male
Centering also helps with parameter interpretation:
y = β0 + β1 (age − m) + β2 male + β3 male ∗ (age − m)
If m is average age, then β2 is E [Y ] for males versus females of
average age. What is β1 ?
In non-linear models interpretation is often more difficult

4
Logit/probit model reminder

There are several ways of deriving the logit model. We can assume a
latent outcome or assume the observed outcome 1/0 distributes either
Binomial or Bernoulli. The latent approach is convenient because it
can be used to derive both logit and probit models
We assume that there is a latent (unobserved) variable y ∗ that is
continuous. Think about it as a measure of illness
If y ∗ crosses a threshold, then the person dies. We only observe if the
person died but we can’t, by definition, observe the latent variable y ∗
What is the probability of dying? We can write this problem as:
P(y = 1|X ) = P(β0 + β1 X1 + · · · + βp Xp + u > 0) = P(−u <
β0 + β1 X1 + · · · + βp Xp ) = F (β0 + β1 X1 + · · · + βp Xp )
F() is the cdf of -u. If we assume logistic distribution, we get logistic
regression, if we assume cumulative normal, we get a probit model

5
Logit/probit model reminder

Remember that the cumulative distribution function (cdf) gives

you P(X < a). Remember too that to getR the probability you need to
a
integrate the density f(t) from −∞ to a: −∞ f (t)dt
If we assume standard normal cdf, our model then becomes
R β0 +β1 x 1 (− t 2 )
P(y = 1|x) = −∞ 2π e
2 dt

And that’s the probit model. Note that because we use the cdf, the
probability will obviously be constrained between 0 and 1 because,
well, it’s a cdf
If we assume that u distributes standard logistic then our model
e β0 +β1 x
becomes P(y = 1|x) = 1+e β0 +β1 x

Remember that there are two different concepts: logistic response

function and logistic distribution. The standard logistic cdf
happens to have the above formula (the pdf is different)

6
Logit/probit model reminder

For now, the most important part to remember is that the scale of
estimation is not the same as the scale of interest (more on this in
30 seconds)
This is because we use transformations to constraint the probability
between 0 and 1
Keep in mind that the logistic model can be derived in different ways.
This tends to confuse students. All ways lead to same likelihood
function and therefore the same parameters
Back to why we need marginal effects...

7
Why do we need marginal effects?

p
We can write the logistic model as: log ( 1−p ) = β0 + β1 age + β2 male
The estimated parameters are in the log-odds scale, which, other
than the sign, don’t have any useful interpretation
In the above equation, β1 is the effect of age on the log-odds of the
outcome, not on the probability, which is often what were care
about
As an alternative, economists prefer to estimate Probit models for
binary outcomes:
P(y = 1|male, age) = Φ(γ0 + γ1 age + γ2 male)
But still similar problem. In the estimation scale, γ1 is interpreted as
shifts in the standard cumulative normal, which, again, is of little help

8
Why do we need marginal effects?
With the logit model we could present odds ratios (e β1 and e β2 ) but
odds-ratios are often misinterpreted as if they were relative
risks/probabilities (nonetheless presenting odds-ratios is standard
practice in the medical literature)
A simple example with no covariates: Say that the probability of
death in a control group is 0.40. The probability of death in the
treatment group is 0.20
0.2
1−0.2
The odds-ratio is: 0.4 = 0.375. The treatment reduces the odds of
1−0.4
death by a factor of 0.375. Or in reverse, the odds of death are 2.67
1
higher in the control group ( 0.375 )
But that’s not the relative risk, even though most people, including
journalists, would interpret the odds ratio as a relative risk. The
relative risk is 0.2
0.4 = 0.5. The probability of death is reduced by half
in the treatment group
9
Why do we need marginal effects?
Note something else. With odds ratios and relative risks, we don’t
have a sense of the magnitude. Same example but now the
probability of death in the control group is 0.0004 and 0.0002 in the
treatment group
The odds ratio is still 0.375 and the relative risk is still 0.5. The
magnitudes are of course quite different
A journalist could still say that, for example, eating broccoli sprouts
daily reduces the probability of dying of cancer by half. By half!!!
But if you learned that the reduction is (0.0004-0.0002) 0.0002 or
0.02 percent points, you probably are not going to run to Miosotis or
Celeiro or Pingo Doce to get a $4 serving of broccoli sprouts every day
On the other hand, a difference of 20 percent point looks quite
impressive
As we will see, marginal effects is a way of presenting results as
differences in probabilities, which is more informative than odds
ratios and relative risks 10
Why do we need marginal effects?: Recap
Ideally, we want to understand what the model saying in the
probability scale and not in the odds scale, much less in the
estimation scale, the log-odds.
In the probability scale, all effects are non-linear because,
conditional on covariate values, the probability must be bounded
between 0 and 1
Here is when numerical methods come to the rescue
We call them marginal effects in econometrics but they come in
many other names and there are different types
Big picture: marginal effects use model PREDICTION for
INTERPRETATION. We are using the estimated model to make
predictions so we can better interpret the model in the scale that
makes more sense (but we are not trying to evaluate how good is the
model at predicting...)
11
Big picture: not just for logit/probit models

We are going to use the logistic model to introduce marginal effects

But marginal effects are applicable to any other model
We will also use them to interpret linear models with more difficult
functional forms
Marginal effects can be use with Poisson models, GLM, two-part
models. In fact, most parametric models

12
It’s about derivatives

We are going to estimate a model and we are going to use

predictions to help us interpret the model
Our goal will be to take numerical derivatives of functions for which
analytical derivatives are more complicated
As we saw, in the model yi = β0 + β1 Xi + β2 Xi2 + i we can take the
[yi |Xi ]
derivative with respect to X : dE dX i
= β1 + 2β2 Xi
In the logistic model, our estimation scale is the log-odds but we
would like to interpret our model in the probability scale
1
In the probability scale, the model is p = 1+e − (β0 +β1 X1 +···+βp Xp )
Reminder: To go from log-odds to probabilities, solve for p in
p e β0 +β1 X 1
log ( 1−p ) = β0 + β1 X . You get p = 1+e β0 +β1 X = 1+e − (β +β X )
0 1

13
It’s about derivatives

We know that for a function y = e X the derivative with respect to X

dy
is dX = e X so the above expression is not going to simplify to a
simpler expression
In fact, the effect of X1 depends on the value of X1 and the values of
all other covariates:
∂p β1 e β0 +β1 X1 +···+βp Xp
∂X1 = (1+e −(β0 +β1 X1 +···+βp Xp ) )2
We would need to choose values for all the covariates in the model
Instead, we will approximate the analytical derivative numerically
using the definition of derivative and we will compute the average
effect of X1 on p

14
Derivative, review
The analytical derivative is a limit:
limh→0 f (x+h)−f
h
(x)

All the formulas for the derivative can be derived using the
definition and taking the limit. For example, an easy one for
f (x) = X 2
2 2
limh→0 (x+h)h −x = x +2xh+h
2 2 2
−x 2
x = 2xh+h
h = 2x + h = 2x
Numerically, that is, without finding the analytical formula, we could
use the definition plugging in a number for h that is small enough. In
that case:
limh→0 f (x+h)−f
h
(x)
≈ f (x+h)−f
h
(x)

Computationally, it’s not trivial to come up with a number h that is

small but “large enough” so that computations are numerically
accurate
For more on choosing h see Gould, Pitblado, and Poi (2010), Chapter
1 and Greene (2017) Appendix E
15
Two-sided derivatives

The above approximation to the analytical derivative

f 0 (x) ≈ f (x+h)−f
h
(x)
is not the only way to define the derivative
A two-sided derivative calculates a positive and negative small change
h:
f (x+h)−f (x−h)
f20 (x) ≈ 2h
The same issues about how to choose a small h remain
We will calculate marginal effects “by hand” and then we will use the
margins command
We will use both definitions of the derivative but Stata uses the
two-sided version
Our calculations will be very close but Stata uses an additional
iterative procedure that changes the value of h to achieve numerical
accuracy. This is what slows down marginal effects calculations

16
Terminology
As usual, language that originates in one discipline doesn’t translate
well to others. The term “marginal affects” is common in economics
and is the language of Stata
Gelman and Hill (2007) use the term “average predicted probability”
to refer to the same concept as marginal effects (in the logit model)
SAS and R have some procedures that can get marginal effects and
are also called marginal effects
One confusion is that when you tell your statistician friend about
marginal effects, your friend imagines an integral because of marginal
probability density functions (in a table of joint probabilities, the
probabilities “at the margin” are the marginal probabilities)
In economics, marginal means “additional” or “incremental,” which is
a derivative
Career advice: When you use marginal effects in a
presentation/paper for non-economists, make sure that you explain
what you mean when you show marginal effects
17
Digression: Is it a unit change?

In a model like y = β0 + β1 x +  we have have said that β1 is the

change in E [y ] for a one unit change in x
And that’s fine when the change is constant (linear)
In the model y = β0 + β1 x + β2 x 2 + , the marginal effect/change
is no longer for a 1 unit change even though most people would
interpret it that way when using marginal effects. Do it by hand:
Start with x = x 0 . Then change by one unit to x 0 + 1 and compare
the two predictions for y . (For simplicity, dropping the “hat” from the
betas but the betas are the same)
y 1 − y 0 = β0 + β1 (x 0 + 1) + β2 (x 0 + 1)2 − β0 − β1 x 0 − β2 (x 0 )2
You will find: y 1 − y 0 = β1 + 2β2 x 0 + β2 , which is not the same as
dy /dx = β1 + 2β2 x evaluated at x = x 0 . There is the extra β2 term
lingering there

18
Digression: Is it a unit change?

Easy to understand why if you use the definition of the derivative:

β0 +β1 (x+h)+β2 (x+h)2 −β0 −β1 x−β2 x
limh→0 h = β1 + 2β2 x + β2 h
So when h tends to zero both ways will be the same (increasing by
one or taking the derivative). The marginal effect applies to a
very small change in x, not to an one unit change in x
When are they going to be the same? The above formula provides
some intuition: β2 is the coefficient for x 2 , the curvature. The smaller
the β2 the smaller the curvature and closer to a straight line
So they are close when the curve looks more like a line
Bottom line: marginal effects for continuous variables apply to
a small change in x when effects are non-linear. They are not
changes by 1 unit, strictly speaking

19
Data

We will use birth weight data from Wooldridge (bcuse bwght)

I created an indicator for low birth weight. Very low birth weight is
about 50 ounces (1,500 grams) but here I used 100 ounces
About 15% of children (mean of lw)
bwght birth weight, ounces
faminc 1988 family income, $1000s
motheduc mother’s yrs of educ
cigs cigs smked per day while preg

gen lw = 0
replace lw = 1 if bwght < 100 & bwght ~= .

sum lw faminc motheduc cigs

Variable | Obs Mean Std. Dev. Min Max
-------------+---------------------------------------------------------
lw | 1,388 .1491354 .3563503 0 1
faminc | 1,388 29.02666 18.73928 .5 65
motheduc | 1,387 12.93583 2.376728 2 18
cigs | 1,388 2.087176 5.972688 0 50

20
Model

We will model low birth weight as a function of cigarettes, mother’s

education, and family income
lwi
log ( 1−lw i
) = β0 + β1 cigsi + β2 faminci + β3 motheduci
logit lw cigs faminc motheduc, nolog

Logistic regression Number of obs = 1,387

LR chi2(3) = 24.63
Prob > chi2 = 0.0000
Log likelihood = -572.15891 Pseudo R2 = 0.0211

------------------------------------------------------------------------------
lw | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs | .0449006 .0104436 4.30 0.000 .0244316 .0653696
faminc | -.0080855 .004801 -1.68 0.092 -.0174953 .0013243
motheduc | .0031552 .037153 0.08 0.932 -.0696634 .0759738
_cons | -1.678173 .4497551 -3.73 0.000 -2.559676 -.7966687
------------------------------------------------------------------------------

21
Model

We can show odds ratios: An additional cigarette increases the odds

of low birth weight by 4.6%
Of course, we now know that it does NOT mean that an additional
cigarette increases the relative probability of low birth weight by 4.6%
or that it’s even a difference in probability, which is what we want
logit, or

Logistic regression Number of obs = 1,387

LR chi2(3) = 24.63
Prob > chi2 = 0.0000
Log likelihood = -572.15891 Pseudo R2 = 0.0211

------------------------------------------------------------------------------
lw | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs | 1.045924 .0109232 4.30 0.000 1.024733 1.067554
faminc | .9919471 .0047623 -1.68 0.092 .9826569 1.001325
motheduc | 1.00316 .0372704 0.08 0.932 .9327077 1.078934
_cons | .1867149 .083976 -3.73 0.000 .0773298 .4508283
------------------------------------------------------------------------------

22
Model
We can also run our trusty linear model with the caveat that SEs are
likely not right (but probably close) and that since low birth
probability is (relatively) low we should be extra careful
Now, in the probability scale, an extra cigarette increases the
probability of low birth weight by 0.7%. With 10 cigarettes, 7% (but
that assumes a linear effect)
reg lw cigs faminc motheduc, robust

Linear regression Number of obs = 1,387

F(3, 1383) = 6.51
Prob > F = 0.0002
R-squared = 0.0212
Root MSE = .35304

------------------------------------------------------------------------------
| Robust
lw | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs | .007757 .0020677 3.75 0.000 .0037009 .0118131
faminc | -.0009345 .0005785 -1.62 0.106 -.0020693 .0002004
motheduc | .0005403 .0042972 0.13 0.900 -.0078895 .00897
_cons | .1531912 .0532648 2.88 0.004 .0487027 .2576797
------------------------------------------------------------------------------

23
A plot is always helpful
A plot will help you understand the shape of the relationship of
interest but remember that other variables may change the shape
lowess lw cigs, gen(lw_c)
scatter lw cigs, jitter(3) msize(small) || ///
line lw_c cigs, color(blue) sort legend(off) saving(l.gph, replace)
graph export l.png, replace

24
Average Marginal Effect (AME)

Sketch: we will compute the numerical derivative of a variable Xj for

each observation using the other covariates Xk as they were
observed. We will change the value of a predictor by “a litlle bit”
and then we compute the average change across observations
In other words, the numerical derivative. Think of marginal effects as
∂p
getting an average derivative ∂X j
that starts by computing a small
change for each observation
When using the margins command, make sure that 1) you use the
option dydx(varname) and 2) make sure you use factor syntax so
Stata knows that variables are continuous or dummy. Also, Stata will
figure out if there are interactions
If you coded interactions by hand (say: gen agemale = age*male),
you will get wrong results using margins without factor syntax

25
Average Marginal Effect (AME)
Let’s calculate AME for the cigarette variable using the typical
formula for the analytical derivative
* Get the "small change"
qui sum cigs
scalar h = (abs(r(mean))+.0001)*.0001
di h
*.00020873
preserve
qui logit lw cigs faminc motheduc, nolog
* as is
predict double lw_0 if e(sample)
* Change cigs by a bit
replace cigs = cigs + scalar(h)
predict lw_1 if e(sample)
* For each obs
gen double dydx = (lw_1-lw_0)/scalar(h)
* Average
sum dydx
restore
Variable | Obs Mean Std. Dev. Min Max
-------------+---------------------------------------------------------
dydx | 1,387 .0055768 .0012444 .0040507 .0113006

An additional cigarette increases the probability of low birth weight by

0.56 percent points
26
Average Marginal Effect (AME)

Replicate using margins command

. margins, dydx(cigs)

Average marginal effects Number of obs = 1,387

Model VCE : OIM

Expression : Pr(lw), predict()

dy/dx w.r.t. : cigs

------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs | .0055782 .0012814 4.35 0.000 .0030666 .0080898
------------------------------------------------------------------------------

Very close but Stata uses the two-sided derivative

The formula h = (|x̄| + .0001) × .0001 is how Stata calculates the
initial small change. In this case, .00020873

27
Average Marginal Effect (AME) the (almost) Stata way
* Two-sided derivative
preserve
qui logit lw cigs faminc motheduc
* Define small change for cigs
qui sum cigs
scalar h = (abs(r(mean))+0.0001)*0.0001

* Duplicte variable
clonevar cigs_c = cigs

* Small negative change

replace cigs = cigs_c - scalar(h)
predict double lw_0 if e(sample)
* Small positive change change
replace cigs = cigs_c + scalar(h)
predict double lw_1 if e(sample)

gen double dydx = (lw_1-lw_0)/(2*scalar(h))

sum dydx
restore
* .0055771

Close. Remember, Stata has an iterative procedure to ensure

numerical accuracy. This is one reason the margins command for
some complex models can be very slow. Stata also calculates the
variance using the delta method
28
Average Marginal Effect (AME)
One advantage of doing it “by hand” rather than using margins is
that you can change the “small” change to any change or just
calculated the a difference ∆Y
∆X . What about 10 extra cigarettes?
preserve
qui logit lw cigs faminc motheduc
predict double lw_0 if e(sample)

replace cigs = cigs + 10

predict double lw_1 if e(sample)

gen double dydx = (lw_1-lw_0)/10

* Could be
* gen dy = (lw_1-lw_0)
sum dydx
restore
Variable | Obs Mean Std. Dev. Min Max
-------------+---------------------------------------------------------
dydx | 1,387 .0064608 .0012196 .0048265 .0111532

In some cases we may not be interested in an approximation for a

small change. You have more flexibility doing it by hand. On the
other hand, we could just change the scale of cigs so changes are for
10 units 29
AME for indicator variables

Understanding marginal effects is easier with dummy variables; that’s

why I have focused on continuous variables
With dummy variables we don’t have to do a “small” change. We
change from 0 to 1
preserve
qui logit lw smoked faminc motheduc
* Nobody smoked
replace smoked = 0
predict double lw_0 if e(sample)

* Everybody smokes
replace smoked = 1
predict double lw_1 if e(sample)

gen double dydx = (lw_1-lw_0)

sum dydx
restore

Variable | Obs Mean Std. Dev. Min Max

-------------+---------------------------------------------------------
dydx | 1,387 .118284 .0105063 .0935378 .1363816

30
AME for indicator variables
We can of course also use the margins command with caution (!)
qui logit lw smoked faminc motheduc, nolog
margins, dydx(smoked)
------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
smoked | .0988076 .0230959 4.28 0.000 .0535405 .1440748
------------------------------------------------------------------------------
qui logit lw i.smoked faminc motheduc, nolog
margins, dydx(smoked)
------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
1.smoked | .118284 .0322576 3.67 0.000 .0550602 .1815078
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.

Even though same margins statement, different results. The first one
is not what we wanted. We did not use the factor syntax in the first
model so Stata didn’t go from 0 to 1; instead it used a “small” change
Smoking increases the probability of low birth weight by almost 12%
points (yikes)
31
The margins command must be treated with caution

Stata’s margins command is worth the price of Stata. It’s truly

awesome
But it’s very easy to get an answer that is different from what you
wanted
A small change in syntax produces very different results. Always
read the output and be very, very careful
“adjust” and “mfx” used to be two separate commands. Now
margins does the same (plus more things) but the syntax became
more complex
Stata is so careful with version control and continuity that you can
still use those commands
Bottom line: be careful

32
AME for indicator variables
With indicator variables, we can also get what Stata calls predictive
margins (not marginal effects). Marginal effects are their difference
We can also use the results to go from margins to relative risk and to
odds ratios
qui logit lw i.smoked faminc motheduc, nolog
margins smoked
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
smoked |
0 | .1305183 .0099014 13.18 0.000 .1111118 .1499248
1 | .2488023 .0304311 8.18 0.000 .1891584 .3084461
------------------------------------------------------------------------------

* marginal effects from predictive margins

. di .2488023 - .1305183
.118284
. di .2488023 / .1305183
1.9062637
. di exp(_b[1.smoked])
2.2115807
* Same as
. di (.2488023/(1-.2488023))/((0.1305183)/(1-0.1305183))
2.206425

Odds ratio: 2.21, relative probability: 1.90. Difference: 0.118.

Remember, the rarer the event the closer they are
33
AME for indicator variables

No predictive margins for continous variables

* no margins for continuous variables (marginal effects instead)
qui logit lw cigs faminc motheduc, nolog
margins cigs
factor ’cigs’ not found in list of covariates
r(322);

No margins for continuous variables because they take too

many values

34
Marginal Effect at the Mean (MEM)
We have left the values of the covariates as they were observed
rather than holding them fixed at a certain value
We can also calculate marginal effects at the mean (of each
covariate)
There is some discussion about which way is better (see Williams,
2012)
For example, does it make sense to hold male at 0.6 male? In a sense,
yes. We are giving males the value of the proportion in the sample,
0.6. In another sense, it seems odd because male is a dummy variable
Don’t waste too much time thinking about this. When we
calculate marginal effects (not margins), it doesn’t really matter at
which value we hold the other covariates constant because we are
taking differences in effects. There could some differences in small
samples
In general, the difference will be so small that it is better to spend
mental resources somewhere else 35
Marginal Effect at the Mean (MEM)
Keep covariates at mean values instead
preserve
qui sum cigs
scalar h = (abs(r(mean))+0.0001)*0.0001
qui logit lw cigs faminc motheduc, nolog
clonevar cigs_c = cigs

* At mean
replace faminc = 29.02666
replace motheduc = 12.93583

* Small negative change

replace cigs = cigs_c - scalar(h)
predict double lw_0 if e(sample)
* Small positive change change
replace cigs = cigs_c + scalar(h)
predict double lw_1 if e(sample)
gen double dydx = (lw_1-lw_0)/(2*scalar(h))
sum dydx
restore

Variable | Obs Mean Std. Dev. Min Max

-------------+---------------------------------------------------------
dydx | 1,387 .0055624 .0010396 .0051876 .011267

MEM not that different from AME 36

Marginal Effect at the Mean (MEM)

Using the margins command

margins, dydx(cigs) at((mean) faminc motheduc)

Average marginal effects Number of obs = 1,387

Model VCE : OIM

Expression : Pr(lw), predict()

dy/dx w.r.t. : cigs
at : faminc = 29.04218 (mean)
motheduc = 12.93583 (mean)

------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs | .005563 .0012843 4.33 0.000 .0030458 .0080801
------------------------------------------------------------------------------

Always read Stata’s output; remember, be afraid of the margins

command!

37
Marginal Effect at the Mean (MEM)
Not the same as using the atmeans option
margins, dydx(cigs) atmeans

Conditional marginal effects Number of obs = 1,387

Model VCE : OIM

Expression : Pr(lw), predict()

dy/dx w.r.t. : cigs
at : cigs = 2.088681 (mean)
faminc = 29.04218 (mean)
motheduc = 12.93583 (mean)

------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs | .0055506 .0012879 4.31 0.000 .0030264 .0080749
------------------------------------------------------------------------------

In this one, cigarettes were held at its mean, 2.088. Not a big deal in
this example because in this example the effect of cigs is relatively
linear (see lowess plot above) but you could have gotten a very
different answer
One more time: please be careful with the margins command
38
Marginal effects at representative values (MER)

We can hold values at observed values (AME) or at mean values

(MEM)
We could also choose representative values; values that are of interest
For example, what is the marginal effect of an additional cigarette on
the probability of low birth weight at different levels of income, say
10K, 20K, 30K and 40K?
Leave other covariates as observed

39
Marginal effects at representative values (MER)
We will do it “by hand” for low income (10K) and higher income
(40K) using the one-sided version to make the code shorter
preserve
qui logit lw cigs faminc motheduc, nolog
* income 10k
replace faminc = 10
predict double lw_0_10 if e(sample)
replace cigs = cigs + .00597269
predict double lw_1_10 if e(sample)
gen double dydx10 = (lw_1_10-lw_0_10)/.00597269

* income 40k
replace faminc = 40
predict double lw_0_40 if e(sample)
replace cigs = cigs + .00597269
predict double lw_1_40 if e(sample)
gen double dydx40 = (lw_1_40-lw_0_40)/.00597269

sum dydx*
restore
Variable | Obs Mean Std. Dev. Min Max
-------------+---------------------------------------------------------
dydx10 | 1,387 .0061672 .0010198 .005653 .0112164
dydx40 | 1,387 .0052304 .001039 .0047327 .0111981

40
Marginal effects at representative values (MER)
Below, income reduces the effect of smoking. Better access to health
care? So income is a modifier of the effect?
qui logit lw cigs faminc motheduc, nolog
margins, dydx(cigs) at(faminc=(10 20 30 40)) vsquish

Average marginal effects Number of obs = 1,387

Model VCE : OIM
Expression : Pr(lw), predict()
dy/dx w.r.t. : cigs
1._at : faminc = 10
2._at : faminc = 20
3._at : faminc = 30
4._at : faminc = 40
------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs |
_at |
1 | .0061667 .0014205 4.34 0.000 .0033825 .0089508
2 | .0058463 .001337 4.37 0.000 .0032258 .0084668
3 | .0055334 .0012803 4.32 0.000 .0030239 .0080428
4 | .0052289 .0012501 4.18 0.000 .0027788 .0076789
------------------------------------------------------------------------------

41
Marginal effects at representative values (MER)

But wait. Why different effects by income? We did not add

interactions between income and cigs and income in the model
Because in the probability scale the effect cigs depends on the values
of income and all other covariates!!!!
This won’t happen in the linear probability model
If we think that the effect of cigs depend on income, we should add
interactions to make this explicit

42
Same but with LPM
Since there are no interactions, the marginal effect doesn’t depend on
the value of income
qui reg lw cigs faminc motheduc
margins, dydx(cigs) at(faminc=(10 20 30 40)) vsquish

Average marginal effects Number of obs = 1,387

Model VCE : OLS

Expression : Linear prediction, predict()

dy/dx w.r.t. : cigs
1._at : faminc = 10
2._at : faminc = 20
3._at : faminc = 30
4._at : faminc = 40

------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs |
_at |
1 | .007757 .001631 4.76 0.000 .0045574 .0109566
2 | .007757 .001631 4.76 0.000 .0045574 .0109566
3 | .007757 .001631 4.76 0.000 .0045574 .0109566
4 | .007757 .001631 4.76 0.000 .0045574 .0109566
------------------------------------------------------------------------------

43
With interactions the effect should be more noticiable
Now adding interactions between cigarettes and income. This is the
right way of making the effect of cigs depend on income
Note how the conclusion is different. However, the interaction is not
statistically significant
qui logit lw c.cigs##c.faminc motheduc, nolog
margins, dydx(cigs) at(faminc=(10 20 30 40)) vsquish
Average marginal effects Number of obs = 1,387
Model VCE : OIM
Expression : Pr(lw), predict()
dy/dx w.r.t. : cigs
1._at : faminc = 10
2._at : faminc = 20
3._at : faminc = 30
4._at : faminc = 40

------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs |
_at |
1 | .0054953 .0016702 3.29 0.001 .0022217 .0087689
2 | .0059858 .0013512 4.43 0.000 .0033374 .0086342
3 | .0064006 .0016609 3.85 0.000 .0031453 .0096558
4 | .0067452 .0022472 3.00 0.003 .0023408 .0111497
------------------------------------------------------------------------------

44
Marginsplot
You can visualize changes using marginsplot. This is a way to get
adjusted plots using the margins command
marginsplot, saving(mp.gph, replace)
graph export mp.png, replace

45
Interactions

We have estimated the model

lwi
log ( 1−lw i
) = β0 + β1 cigsi + β2 faminci + β3 motheduci
We didn’t use interactions between cigarettes and income so we have
assumed the same effect regardless of income in the log-odds scale
In other words, same slope and and same intercept
If we add interactions, the margins command is a life saver because
effects are hard to interpret otherwise
In particular, the magnitude of effects is hard to make sense when a
logistic model has interactions and we use odds ratios for
interpretations

46
Interactions

As usual, interactions are easier to understand with two indicator

variables. Let’s say more than high school (hsp) and male as in
p
previous examples: log ( 1−p ) = β0 + β1 hsp + β2 male + β3 hsp ∗ male
In the log-odds scale, we interpret them exactly in the same way as in
the linear model and we can use derivatives and the other tricks we
learned (What’s β1 ? β2 ?)
pm
For males: log ( 1−p m
) = β0 + β1 hsp + β2 + β3 hsp
pf
For females: log ( 1−p f
) = β0 + β1 hsp
pm pf
Difference males - females: log ( 1−p m
) − log ( 1−p f
) = β2 + β3 hsp
So the effect depends on the level of education
Pm
Note that we can write it in terms of odds: 1−Pm
Pf = e β2 +β3 hsp
1−Pf

47
Interactions

pm pf
Difference males - females: log ( 1−p m
) − log ( 1−p f
) = β2 + β3 hsp
Difference male - female for educated:
pme pfe
log ( 1−p me
) − log ( 1−p fe
) = β2 + β3
Difference male - female for uneducated:
pmu pfu
log ( 1−p mu
) − log ( 1−p fu
) = β2
Difference in difference:
pme pfe pmu pfu
log ( 1−p me
) − log ( 1−p fe
) − [log ( 1−p mu
) − log ( 1−p fu
)] = β3
So same as with linear model. In the log-odds scale, it is a
difference-in-difference

48
Interactions in the odds scale
pme pfe pmu pfu
log ( 1−p me
) − log ( 1−p fe
) − [log ( 1−p mu
) − log ( 1−p fu
)] = β3
We can apply the rules of logs and take e () on both sides:
Pme Pmu
1−Pme
Pfe / 1−P
Pfu
mu
= e β3
1−Pfe 1−Pfu

In the odds scale, an interaction is not a difference-in-difference but

rather a ratio of odds ratios
The interpretation is... well, good luck with that...
The interpretation is not that hard, actually. If greater than one, the
odds of outcome for educated are higher than for the uneducated (see
the numerator–it’s the odds ratio for males females for those
“educated” or hsp=1. The denominator is same but for those with
hsp=0)
Interpreting the magnitude in the probability scale is a lost
cause. A lot easier with margins or stratifying the models
49
Interactions and stratification
Remember, a fully interacted model is the same as a stratified model;
it helps to think about it when verifying that interactions in the odds
scale is a ratio of odds ratios. Back low birth babies
Make sure you understand this (below, you can replicate the
0.947913–the interaction–running two stratified models and using
odds ratios)
gen inc = 0 if faminc ~=.
replace inc = 1 if faminc > 40 & faminc ~=.
logit lw c.cigs##i.inc, nolog or
------------------------------------------------------------------------------
lw | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs | 1.050702 .0109996 4.72 0.000 1.029363 1.072483
1.inc | .9045562 .1740226 -0.52 0.602 .6204083 1.318844
|
inc#c.cigs |
1 | .947913 .0553484 -0.92 0.360 .8454093 1.062845
|
_cons | .1583417 .0151181 -19.30 0.000 .1313182 .1909263
------------------------------------------------------------------------------
qui logit lw cigs if inc == 0, nolog
. di exp(_b[cigs])
1.0507018
qui logit lw cigs if inc == 1, nolog
. di exp(_b[cigs])
.99597388
. di .99597388/1.0507018 50
Interactions and marginal effects

What about if we use marginal effects instead? The model is:

The model is logit(lw ) = β0 + β1 cigs + β2 inc + β3 cigs ∗ inc
inc is a dummy and equal 1 if income is higher than 40k
margins, dydx(*)
Average marginal effects Number of obs = 1,388
Model VCE : OIM
Expression : Pr(lw), predict()
dy/dx w.r.t. : cigs 1.inc
------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs | .0046941 .0018482 2.54 0.011 .0010717 .0083165
1.inc | -.0288422 .0225222 -1.28 0.200 -.0729849 .0153006
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.

Wait, two effects? The model has three coefficients. Where is the
interaction?
51
Interactions and marginal effects

This may seem confusing but it’s not when you remember how Stata
calculates marginal effects
For cigs, a continuous variable, it’s using the two-sided derivative
increasing cigs by a little bit and calculating predictions. It’s
increasing cigs in both the main effect and the interaction
Then it takes an average so the marginal effect of cigs is the
numerical derivative for both inc=1 and inc=0 combined
For the marginal effect of inc, it’s doing the same going from 0 to 1,
averaging over the values of cigs
To get what we need, which in this case is the marginal effect of cigs
separately for inc=1 and inc=0, we have to be more specific

52
Interactions and marginal effects

With interactions the effect of cigs depends on income

margins, dydx(cigs) at(inc=(0 1)) vsquish
Average marginal effects Number of obs = 1,388
Model VCE : OIM
Expression : Pr(lw), predict()
dy/dx w.r.t. : cigs
1._at : inc = 0
2._at : inc = 1

------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cigs |
_at |
1 | .0062867 .0012881 4.88 0.000 .0037621 .0088113
2 | -.0004394 .0062301 -0.07 0.944 -.0126501 .0117713
------------------------------------------------------------------------------

You need to remember this. Be careful using the margins

command when you have interactions

53
Interactions and marginal effects
Of course, interactions go both ways. So the effect of income
depends on the number of cigs. But cigs is continuous; we have to
choose some values
margins, dydx(inc) at(cigs=(0 10 20 40)) vsquish
Conditional marginal effects Number of obs = 1,388
Model VCE : OIM
Expression : Pr(lw), predict()
dy/dx w.r.t. : 1.inc
1._at : cigs = 0
2._at : cigs = 10
3._at : cigs = 20
4._at : cigs = 40
------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
0.inc | (base outcome)
-------------+----------------------------------------------------------------
1.inc |
_at |
1 | -.0114123 .0214935 -0.53 0.595 -.0535388 .0307143
2 | -.0851994 .0622866 -1.37 0.171 -.2072788 .03688
3 | -.1819228 .1224851 -1.49 0.137 -.4219893 .0581436
4 | -.4251388 .2397838 -1.77 0.076 -.8951064 .0448287
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.

54
Digression

We can of course estimate marginal effects for linear models without

nonlinear terms but they will be boring
sysuse auto, clear
reg price mpg turn
------------------------------------------------------------------------------
price | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mpg | -259.6967 76.84886 -3.38 0.001 -412.929 -106.4645
turn | -38.03857 101.0624 -0.38 0.708 -239.5513 163.4742
_cons | 13204.27 5316.186 2.48 0.015 2604.1 23804.45
------------------------------------------------------------------------------
margins, dydx(*)
------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mpg | -259.6967 76.84886 -3.38 0.001 -412.929 -106.4645
turn | -38.03857 101.0624 -0.38 0.708 -239.5513 163.4742
------------------------------------------------------------------------------

55
Margins are predictions

The essence of margins and marginal effects is that they are

predictions
We are using our estimated model to make predictions when we
change a continuous variable by a small amount or when we change
an indicator variable from 0 to 1
They are extremely useful because they allow us to interpret our
models
They are truly indispensable when the scale of estimation is not the
same as the scale of interest (logit, Poisson, etc) or when we have
non-linear terms

56
A very brief summary of margins and the margins
command
Most common uses: estimate “effects” in the scale of interest.
That is a 1) a numerical derivative for continuous covariates or 2)
incremental effects for dummy variables. Syntax is “margins,
dydx(varname)”
Another possibility is to use margins to obtain “predictive margins”
of dummy variables–and if you fix a continuous covariate at some
values, for continuous variable as well. Syntax is “margins varname”
or “margins varname, at(...)”
With the previous syntax you can use margins to obtain predictions,
Just specify values for all covariates: “margins, at(var1=10 var2=20
var3=...)”
You can also use margins to obtain “adjusted predictions,” which is
essentially the same ideas as previous point. You need to fix covariates
at some values: margins, at(cigs=(0(1)50) faminc=50 motheduc=13)
57
A very brief summary of margins and the margins
command
Confusion alert: Make sure you understand the difference between
marginal effects and predictive margins. I guarantee you are going to
get confused
Marginal effects (dydx) is about effects; the other is about calculating
predictions but not effects. Yet, part of the confusion is that in order
to calculate effects you also use predictions BUT changing values by a
“small” amount or from 0 to 1
We haven’t discussed other features but you can use the margins
command to express effects as elasticities, for example
The marginsplot has many options. It’s specially helpful to display
interactions and understand the model
You can produce adjusted plots as in the example in Excel using a
reference population...
58
Predictions in logistic models
We saw that we can easily make predictions in the probability scale
logit lw smoked faminc, nolog
Logistic regression Number of obs = 1,388
LR chi2(2) = 24.30
Prob > chi2 = 0.0000
Log likelihood = -572.48309 Pseudo R2 = 0.0208

------------------------------------------------------------------------------
lw | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
smoked | .7898437 .1830874 4.31 0.000 .4309989 1.148688
faminc | -.0076573 .0043414 -1.76 0.078 -.0161662 .0008516
_cons | -1.681833 .151239 -11.12 0.000 -1.978256 -1.38541
------------------------------------------------------------------------------
di exp(_b[_cons] + _b[smoked] + _b[faminc]*30) / (1+exp(_b[_cons] + _b[smoked] + _b[faminc]*30))
.24569462

A mother who smoked during pregnancy with 30K in family income

has 25% probability of having a child weighting less than 100 ounces
at birth
Pseudo R 2 is low but LRT test tell us that the two explanatory
variables are better than a model with no explanatory variables (never
a surprise)
59
What about other models?

Once you understand the logic of numerical derivatives is not difficult

to understand what the margins command does when applied to other
models
By default, Stata calculates marginal effects in the same units as
the post-estimation command predict
For example, if you use the poisson command for counts, the default
is to obtain predictions as counts, not rates
So the margins command by default will provide effects in the counts
scale

60
More custom: Two-part models

Two-part models are often used to estimate cost data with a large
proportion of costs that are zero
The idea is to estimate the model in two parts:
1) Estimate the probability that the cost is greater than zero
conditional on covariates: P(yi > 0|Xi ) (using logit, probit,
complementary log-log models)
2) For those observations with non-zero costs, estimate the expected
costs conditional on covariates: E (yi |yi > 0, Xi ) (using Poisson, linear
models, Gamma, log (y ), Box-Cox, etc)
Predictions are obtaining combining both parts (multiplication)

61
Two-part models marginal effects
If you know how to get predictions, you now how to calculate
marginal effects
First, using the twopm user-written command and the margins
command
* Get data from example in twopm command
webuse womenwk
replace wage = 0 if wage==.

* First part logit, second part GLM with Gamma family

qui twopm wage i.married children, firstpart(logit) secondpart(glm, family(gamma) link(log))
margins, dydx(married)

Average marginal effects Number of obs = 2,000

Expression : twopm combined expected values, predict()
dy/dx w.r.t. : 1.married
------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
1.married | 6.615493 .5323105 12.43 0.000 5.572183 7.658802
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.

62
Two-part models marginal effects “by hand”
gen nozeroc = 0
replace nozeroc = 1 if wage >0
preserve
clonevar marr = married
* not married
qui logit nozeroc i.married children
replace married = 0
predict double fp0
replace married = marr
qui glm wage i.married children if wage > 0, f(gamma) l(log)
replace married = 0
predict double c0
gen chat0 = fp0*c0

* married
replace married = marr
qui logit nozeroc i.married children
replace married = 1
predict double fp1
replace married = marr
qui glm wage i.married children if wage > 0, f(gamma) l(log)
replace married = 1
predict double c1
gen chat1 = fp1*c1

gen dif = chat1 - chat0

sum dif
restore
Variable | Obs Mean Std. Dev. Min Max
-------------+---------------------------------------------------------
dif | 2,000 6.615493 1.660084 2.485561 8.001396
63
Summary

Marginal effects come to the rescue when we need to interpret

non-linear models
Different terms for these types of effects
What we did today was about PREDICTION, 100 percent. We use
predictions to interpret models using numerical derivatives
The margins command has unified some of the terminology
But if you talk to your friendly statistician, you need to explain
what you mean by marginal effects

64