Getting started
One comparison, start to finish
mecompare can be used to calculate marginal effects within a single model and then, optionally, compare them. Or, you can fit two or more models, store them, and then call mecompare to calculate the marginal effects from each model and compare them. This page runs through one example with Stata’s nlsw88 data and then explains how to read the table.
1. Fit and store the models
The question: does the association between having a college degree and union membership change once we account for wages and work experience? Fit the model without those variables, then the model with them, storing each with estimates store.
sysuse nlsw88, clear(NLSW, 1988 extract)drop if missing(union)(368 observations deleted)
. quietly logit union i.collgrad c.age i.south i.smsa, vce(robust)
estimates store base. quietly logit union i.collgrad c.age i.south i.smsa c.wage c.ttl_exp, vce(robust)
estimates store fullMost regression commands in Stata are supported; the full list is under Supported models on the Options page. With two or more models, we recommend fitting the models with vce(robust) as suest2 is used to combine the models’ estimates and it uses robust variance estimation. This ensures results from mecompare match results from the individual models exactly. Multilevel and panel models are the exception: fit them, and any model compared with them, without vce(robust); mecompare clusters the standard errors on the highest-level group.
2. Compare the marginal effects
Name the variable whose effect you want and the stored models in models().
mecompare i.collgrad, models(base full)Predicting: Pr(union)
Marginal effects and cross-model differences (N_base=1878) (N_full=1878)
| ME # Estimate Robust SE P>|z|
---------------------------------+---------------------------------------
collgrad |
College gra - Not colleg |
base | 1 0.092 0.024 0.000
full | 2 0.057 0.025 0.024
Difference | 3 0.035 0.009 0.0003. Read the table
Each block of the table is one marginal effect. For a binary or factor variable the effect is the discrete change in the predicted probability between its levels; for a continuous variable, the default is a change for a one-unit increase (other amounts are set with amount()). Within a block:
- the first rows give the marginal effect in each model, in the order the models were listed in
models(); - the Difference row is the effect in the first model minus the effect in the second, with its standard error and p-value.
The # column numbers the rows so that you can refer to them in metest for any other contrast.
Continuous variables and the size of the change
For a continuous independent variable, amount() sets the size of the change. Here the effect of a one-standard-deviation increase in age. All changes for continuous independent variables are centered by default.
mecompare age, models(base full) amount(sd)Predicting: Pr(union)
Marginal effects and cross-model differences (N_base=1878) (N_full=1878)
| ME # Estimate Robust SE P>|z|
---------------------------------+---------------------------------------
age + SD (centered) |
base | 1 0.005 0.010 0.607
full | 2 0.004 0.010 0.676
Difference | 3 0.001 0.002 0.553Comparing effects within one model
With a single model in models(), mecompare reports the marginal effects of the variables listed, and metest can be used to compare them. Here, is the union-membership gap by college degree the same size as the gap by region?
mecompare i.collgrad i.south, models(full)Predicting: Pr(union)
Marginal effects (N_full=1878)
| ME # Estimate Robust SE P>|z|
---------------------------------+---------------------------------------
collgrad |
College gra - Not colleg |
full | 1 0.057 0.025 0.024
---------------------------------+---------------------------------------
south |
South - Not south |
full | 2 -0.114 0.020 0.000metest 1 - 2 | estimate se pvalue
---------------------------------+--------------------------------
collgrad - south | 0.171 0.034 0.000 Frequently used options
The help file (help mecompare) documents everything; these are the options that come up most often.
| Option | Purpose |
|---|---|
models(name1 [name2 ...]) |
The stored estimates to compare. One model gives marginal effects for that model alone; with three or more, every model’s effects are listed and metest can be used to test the comparisons. |
amount() |
Size of the change for a continuous variable: a number, sd or 2sd (one or two standard deviations), trimrange (5th to 95th percentile), range (minimum to maximum), p#-p# (from one percentile to another, e.g. p10-p90), or rate (the instantaneous rate of change). Several amounts for one variable go in parentheses, e.g. amount((one sd rate)). |
start(), end() |
Where the change in a continuous variable begins, start(varname=#) or start(atmeans), and, with end(varname=#), where it ends. |
cochange(varlist) |
Variables that change along with the focal variable, e.g. mediators; the table adds the effect with co-change (see generalized marginal effects). |
covariates(varname=# ...), atmeans |
Hold other variables at specified values or at their means (the default averages over the observed values). |
by(varlist) |
Marginal effects within each level of another variable, in one table, so that they can be compared with metest. |
over(varlist) |
Marginal effects within subsamples defined by another variable. |
pwcompare |
All pairwise contrasts among the levels of a nominal variable, not just contrasts with the base level. |
mcompare(method) |
Adjust the p-values of a nominal variable’s contrasts for multiple comparisons (bonferroni or sidak). |
groups |
The models were fit to separate groups (see the groups example). |
totalme, meinequality |
Add the Total Marginal Effect and/or the marginal-effect inequality summary for the compared effects. |
store(stub) |
Store the results so that they can be graphed with coefplot or tabled with esttab. |