Quick look at the margins command

Jean-Philippe Gauvin
[email protected]
October 9, 2012

Abstract
In this document I try to give an overview of the margins command in Stata. I start with
a basic explanation of the difference between a predictive margin and a marginal effect. I then
explain how to use the marginsplot command to visualise results. Finally, I explain how to
run multiple models with margins and overlay the results on one graph.

Predictive Margins

In Stata 12, the margins command can do multiple things, one of them being a predictive margin,
or more generally an adjusted prediction. This means that the margins command can, after running
a model, predict the expected value of an estimate, while holding every other variable either at
their means or their observed value.
For example:
regress y x1 x2 x3 x4 x5
margins, at(x2=3) atmeans
In this example, I predict y (i.e. the outcome of y) when x2 is 3, while holding every other
variables constant at their mean. If I used the asobserved argument instead of atmeans, Stata
would run a simulation for each observation and then average the effect.
The adjustment is not limited to one variable; it can also be used to run profile simulations.
For example:
regress income female i.region i.ideology i.education
margins, at(female=1 region=3 ideology=3 education=1)
In this example, the answer given by margins will be the linear prediction of income when the
respondent is a female living in region 3 while having a right-leaning ideology with low education.
The interesting thing with margins is that you can repeat the process and compute that adjusted
prediction for different values of x. Also, it is not limited to linear models.
For example:
logit GW_xp01 c.hierarchy##i.RISK ideology female
margins RISK, at(hierarchy=(1 2 3 4 5 6)) atmeans

In this example, GW_xp01 is a dichotomous variable where 0 means that the expert is not found
credible and 1 means that the expert is found to be credible. hierarchy is a continuous variable
going from 1 to 6. RISK is a dichotomous variable, where 0 means that the respondent was in
the control group and 1 means that he was in the experiment group. We use the ## argument
to tell Stata to use the whole multiplication term in the interaction, as it should. In sum, we
are estimating a predicted probability: that of finding the expert credible depending on the group
(control or experiment), at different values of hierarchy, while maintaining every other covariate at
its mean.
The marginsplot command then provides the graphical representation of those adjusted predictions.

Marginal Effect

The marginal effect (also called partial change or partial derivative) is the difference in prediction
of the presence and absence of a variable. The idea is simple to grasp when thinking of a dummy
variable in a logit model. Since the y outcome is dichotomous, the prediction is the probability
of getting a 1 rather than a 0. When taking the marginal effect of a dummy variable, we try to
get the effect of that variable when it is one, minus its effect when it is zero. In other words, the
marginal effect of x1 is the difference in the probability of y being 1 given that x is 1, minus the
probability that y = 1 given that x is equal to 0 as shown in equation 1:
me (x1 ) = P r(y = 1|x = 1) P r(y = 1|x = 0)

(1)

But the marginal effect might not always be appropriate in the case of dependent categorical
variable. To understand this specifically, let
P r(y = 1|x) = F (x)

(2)

where F is any cumulative density function (as for the normal distribution in the case of probit
or for the logistic distribution in the case of logit). Hence, the marginal effect is also called the
partial change in the probability since it is computed by simple calculus, ie. by taking the partial
derivative of equation 2 with respect to xk :
P r(y = 1|x)
F (x)
dF (x) x
=
=
= f (x)k
xk
xk
dx xk

(3)

As we can see from this transformation, we start with the cdf (F ) and end up with the probability
density function or pdf (f ). In Stata, we would use the option dydx(*) after the margins command,
leaving the star in to provide marginal effects for all covariates, or placing one or several covariates
between the brackets.
As showed in Figure 1, the derivative method means that we must be careful, since we are
estimating the effect from the tangent and not the slope itself. In other words, there is no problem
in linear relationships, but there might be some discrepancies in non-linear relationships like in
logits and probits. The following graph provides an exaggerated example. Long (1997) suggests
alternatives to the marginal effect, like the discrete change. Indeed, using discrete change will
actually touch the slope, compared to the orange line in the following figure who only touches
the derived tangent.
2

Figure 1: Measuring the marginal effect

ME =.25

x1

2.1

X1 + 1

Marginal Effects and Linear Models

The last section took the first derivative of a logistic function with the margins command. But
is it possible to use it with a linear model? The short answer is yes, but the results would not
be very useful. Indeed, the marginal effect of a linear relationship is the beta coefficient. This is
because, as explained in the last section, a marginal effect or partial change in the function is
computed from the derivative of the function at a certain point on the curve. In other words, the
first derivative measures the slope of the tangent. But if the relationship is linear, the result will
be the slope of the regression line, hence the beta coefficient.
But if this result does not seem useful, the marginsplot command provide a neat way to
present regression results. However, using it directly after taking the marginal effects of the linear
prediction slope might not produce the best graphic, as can be seen from the left portion of figure 2.
Indeed, the connected line makes it uninterpretable, while the x-axis is impossible to read. But the
graphic can be greatly improve by flipping it horizontally, recasting the graphic as a dot plot and
setting a line at 0 on the x-axis. This can be done by specifying the following options:
marginsplot, recast(dot) horizontal xline(0, lcolor(red))
The recast command reshapes the graphic in a different format in this case, a dot plot. The
horizontal option flips the graph on its side. Finally, the xline option creates a vertical line
on the x-axis at value 0 and formats it in red. If no color is necessary, xline(0) can be used
on its own. The result can be seen in the right portion of figure 2. Notice that the betas are
now standardized, which facilitates reading. With their confidence interval, the variables that are
significantly different from zero are easy to identify. The next step would be to reorder the variables
by effect size.

Figure 2: Marginsplot with Regression Results

Average Marginal Effects with 95% CIs

.5

fem
5.party

3.party

Effects with Respect to

Effects on Linear Prediction

4.party

-.5

2.party

gender
4.equal
3.equal
2.equal

-1

harper

harper 2.equal
3.equal
4.equal
gender 2.party
3.party
4.party
5.party
fem

-1

Effects with Respect to

-.5

.5

Effects on Linear Prediction

One Limit of the Margins Command

While margins can be very useful at isolating a factor that explains the dependant variable, it
can only simulate one model at a time. This can be very unpractical when trying to estimate
multinomial logit regression models, since it will only predict one outcome at a time. The same
thing applies to ordered logits.
Say we want to know how the perception of a specific leader influences the probability of voting
for different parties. We could run the following model in Stata:
mlogit party gender##c.fem rightscale Cons_leader, rr
Where Cons_leader is a thermometer scale (from 0 to 100) of the perception of the Conservative
leader in this election. In a model where there are four parties, we could use a loop to create and
combine four different graphics:
forvalues i=1/4 {
margins, at(Cons_leader=(0(10)100)) predict(outcome(i))
marginsplot
graph save marginsi.gph
}
graph combine margins1.gph margins2.gph margins3.gph margins4.gph, ///
ycommon ti(Influence of Conservative Leader Feelings on Vote Choice)
This would give us a graphic similar to the one in Figure 3. While it is not an absolutely
ugly graph except for the axis labels most people would want to simplify interpretation by
overlaying the different slopes on one and only graph. Fortunately, it is possible to create those
4

Figure 3: Example graph

Pr(Party==Conservatives)

Inuence of Conservative Leader Feelings on Vote Choice

Predictive Margins with 95% CIs

10 20 30 40 50 60 70 80 90 100

Predictive Margins with 95% CIs

10 20 30 40 50 60 70 80 90 100

Pr(Party==Bloc_Quebecois)

RECODE of CPS11_23 (Feelings about Stephen Harper) RECODE of CPS11_23 (Feelings about Stephen Harper)

Pr(Party==Ndp)

Predictive Margins with 95% CIs

10 20 30 40 50 60 70 80 90 100

Predictive Margins with 95% CIs

10 20 30 40 50 60 70 80 90 100

RECODE of CPS11_23 (Feelings about Stephen Harper) RECODE of CPS11_23 (Feelings about Stephen Harper)

slopes manually by extracting the data from Statas inner memory. But this requires knowledge
of how Stata handles its data. In fact, when posting results, it actually stores data in different
matrices. We can then run the following syntax to gather the relevant data for the slopes. I will
go into a line by line explanation in the next paragraph.
set more off
mlogit party gender##c.fem rightscale Cons_leader, rr
forvalues j=1/4 {
margins, at(Cons_leader=(0(10)100)) predict(outcome(j)) atmeans vsquish post
mat t=J(11,3,.)
mat a = (0\10\20\30\40\50\60\70\80\90\100)
forvalues i=1/11
mat t[i,1] =
mat t[i,2] =
mat t[i,3] =
}

{
_b[i._at]
_b[i._at] - 1.96*_se[i._at]
_b[i._at] + 1.96*_se[i._at]

/* get the "at" values

/* get probability estimates */

mat t=t,a
mat colnames t = prob_partyj lci_partyj uci_partyj at_partyj
svmat t, names(col)
}
5

*/

In order to save some place, I use a loop within a loop to repeat the process multiple times.
The first step is to run the model we want to estimate, hence the mlogit command. The results
include four different outcomes for four different political parties. I then use a loop to repeat the
extraction procedure four times, one for each party. The j argument will be repeatedly replace
by 1, 2, 3 and then 4.
Inside the first loop, we run the margins command, setting our predicted probability at 11
values of Cons_leader, being from 0 to 100, by increments of 10. I also use the vsquish argument
to synthesise the output and the post argument to save the results in matrices.
Afterwards, we need three informations in order to create the slopes. The first is in the first
column of the results, being the predicted probability. These estimates will become our datapoints
in the final graph. We also need two other informations: the lower and upper limits of the confidence
interval so to create slopes refering to those intervals.
We start by typing mat t=J(11,3,.) to create a new t matrix composed of eleven rows and
three columns, all being empty. We then create an a vector that consists of all the _at values from
the first column. We then use a loop to extract everything automatically in eleven steps one for
each _at value and store it in our t matrix. The first line extracts the first beta specified by
the _b argument and stores it in first column of the matrix. The second and third step computes
the lower and upper limit of the confidence interval by subtracting and adding to the beta 1.96
times its standard error specified by the _se argument. This is the classic formula to compute
confidence interval at a 95% level of certainty.
Finally, we reset the t matrix by appending the a vector to it. We then change the column
names so they are more easily interpretable. And as a last step, we save the columns as variables,
using the column names. We now have twelve variables four variables for each party each
consisting of eleven rows, for the values from 0 to 100 by increments of 10. We can finally graph
those slopes with the following syntax.
twoway (line prob_party1 at_party1, lcolor(red) lpattern(solid)) /// party 1 set
(line lci_party1 at_party1, lcolor(red) lpattern(dash_3dot) lwidth(thin)) ///
(line uci_party1 at_party1, lcolor(red) lpattern(dash_3dot) lwidth(thin)) ///
///
(line prob_party2 at_party2, lcolor(dknavy) lpattern(solid) ) /// party 2 set
(line lci_party2 at_party2, lcolor(dknavy) lpattern(longdash) lwidth(thin)) ///
(line uci_party2 at_party2, lcolor(dknavy) lpattern(longdash) lwidth(thin)) ///
///
(line prob_party3 at_party3, lcolor(orange) lpattern(solid) ) /// party 3 set
(line lci_party3 at_party3, lcolor(orange) lpattern(dash_3dot) lwidth(thin)) ///
(line uci_party3 at_party3, lcolor(orange) lpattern(dash_3dot) lwidth(thin)) ///
///
(line prob_party4 at_party4, lcolor(eltblue) lpattern(solid) ) /// party 4 set
(line lci_party4 at_party4, lcolor(eltblue) lpattern(dash_3dot) lwidth(thin)) ///
(line uci_party4 at_party4, lcolor(eltblue) lpattern(dash_3dot) lwidth(thin)) ///
, legend(off) ///
xtitle(Feeling about Conservative Leader) ytitle("pr(Vote=Party x)") ///
ti(Estimate of vote choice)

Figure 4: Better Visual Representation

pr(Vote=Party x)

.8

.6

.4

.2

0
0

20

40

60

80

100

Feeling about Conservative Leader

This syntax might seem, yet again, complex. In fact, it is pretty straightforward. It is composed
of only one command: scatter and multiple slopes. Hence, this syntax is one huge line of command,
separated by breaks identified by three slashes (///) to make it more easy to read. As explained
earlier, we are not creating four slopes but twelve: four slopes for the different parties and 8 slopes
for the confidence intervals. Hence, I create the line with the line argument. Each line and its
options are contained in brackets. I start by plotting the slope of the estimate, using lcolor to
set the color and lpattern to choose a special design. For the confidence intervals, I also add the
lwidth option so the graph is more easily read. I then add axis titles and a graph title.
The result can be seen in Figure 4. While it is not the most wonderful figure in the world, it is
better than figure 3. But more importantly, the technique can be used with any models where you
wish to overlay different slopes with margins results, such as multinomial logits, ordered logits or
simply different logits.