Stata Migration Guide

Complete translation reference for economists migrating from Stata's margins command

Basic Command Translation

Core Margins Commands

Variable Selection

Grouping and Stratification

Basic Grouping

Nested Analysis

Scenario Analysis (at() Specification)

Basic Scenarios

Population-Level Counterfactuals

Key Difference: Stata's at() creates evaluation points, while Margins.jl's scenarios creates population counterfactuals.

Profile vs Population Interpretation

# Stata: margins, at(treatment=(0 1))
# → Effects at two evaluation points

# Margins.jl Profile Equivalent
profile_results = profile_margins(model, data, 
    cartesian_grid(treatment=[0, 1]);
    type=:effects)

# Margins.jl Population Alternative (often more relevant)  
population_results = population_margins(model, data;
    scenarios=(treatment=[0, 1]),
    type=:effects)

Skip Rule: dydx(x) with over(x) (and scenarios)

Unlike Stata, population_margins intentionally skips computing the effect of a variable when that same variable appears in groups (Stata over()) or in scenarios (Stata at()). This avoids the contradiction of “compute the effect of x while holding x fixed” or “using x both as an effect variable and a grouping key.”

Recommended translations:

# 1) Stata: margins, dydx(x) over(x)
# → Profile-style alternative: evaluate derivatives at specific x values
mem_like = profile_margins(model, data,
    cartesian_grid(x=[-2.0, 0.0, 2.0]);
    type=:effects,
    vars=[:x])

# 2) Population stratification by x without contradiction:
#    Create a derived bin variable and group by it, not by :x directly
df.x_bin = cut(df.x, 4)  # quartiles via user code; or use groups=(:x, 4)
by_xbins = population_margins(model, df;
    type=:effects,
    vars=[:x],
    groups=:x_bin)  # allowed since groups variable ≠ :x

# 3) Effects of other variables within x strata (population approach)
effects_in_xbins = population_margins(model, data;
    type=:effects,
    vars=[:z, :w],
    groups=(:x, 4))

# 4) Counterfactual predictions as x changes (not effects of x)
preds_under_x = population_margins(model, data;
    type=:predictions,
    scenarios=(x=[-2.0, 0.0, 2.0]))

See also: “Skip Rule” note in the Population Grouping docs for rationale and guidance.

Short example: grouping by x_bin to compute dydx(x) across strata

using Random
using DataFrames, CategoricalArrays
using Statistics  # for quantile
using GLM
using Margins

Random.seed!(42)
n = 500
df = DataFrame(
    y = rand(Bool, n),
    x = randn(n),
    z = randn(n)
)

# Fit a simple model
m = glm(@formula(y ~ x + z), df, Binomial(), LogitLink())

# Create quartile bins for x as a separate column "x_bin"
edges = quantile(df.x, 0:0.25:1.0)
labels = ["Q1", "Q2", "Q3", "Q4"]
df.x_bin = cut(df.x, edges; labels=labels, extend=true)

# Now compute population AME of x within x_bin strata (no contradiction)
res = population_margins(m, df;
    type=:effects,
    vars=[:x],
    groups=:x_bin)

DataFrame(res)  # Shows dydx(x) by Q1..Q4

Fixed Effects Models (reghdfe / ivregress)

Basic Fixed Effects

Predictions with Fixed Effects

Predictions require save=:fe when fitting the model:

model = reg(df, @formula(y ~ x1 + x2 + fe(state) + fe(year)); save=:fe)

Instrumental Variables

See Fixed Effects Models for full documentation.

Combined Grouping and Scenarios

Complex Analysis Patterns

Advanced Patterns Beyond Stata

Margins.jl extends far beyond Stata's capabilities with features unavailable in Stata:

Continuous Variable Binning

* Stata approach (manual and cumbersome):
gen income_q = .
_pctile income, nq(4)
replace income_q = 1 if income <= r(r1)
replace income_q = 2 if income > r(r1) & income <= r(r2)
replace income_q = 3 if income > r(r2) & income <= r(r3)
replace income_q = 4 if income > r(r3)
margins income_q

# Julia approach (automatic):
population_margins(model, data; groups=(:income, 4))  # Automatic Q1-Q4 quartiles

Custom Policy Thresholds

* Stata approach:
gen income_bracket = .
replace income_bracket = 1 if income < 25000
replace income_bracket = 2 if income >= 25000 & income < 50000
replace income_bracket = 3 if income >= 50000 & income < 75000
replace income_bracket = 4 if income >= 75000
margins income_bracket

# Julia approach:
population_margins(model, data; groups=(:income, [25000, 50000, 75000]))

Hierarchical Grouping

* Stata approach (requires multiple commands or complex by groups):
by region: margins education
* No native support for deep nesting

# Julia approach (native hierarchical support):
population_margins(model, data; groups=:region => :education)
population_margins(model, data; groups=:country => (:region => :education))  # Deep nesting

Multi-Variable Scenarios

* Stata approach (requires multiple separate commands):
margins, at(treatment=0 policy=0)
margins, at(treatment=0 policy=1) 
margins, at(treatment=1 policy=0)
margins, at(treatment=1 policy=1)

# Julia approach (automatic Cartesian product):
population_margins(model, data; scenarios=(treatment=[0, 1], policy=[0, 1]))

Note: scenarios in Julia are population‑level counterfactuals (everyone receives each setting in turn). For Stata’s point‑evaluation semantics of at(), use profile_margins(model, data, reference_grid) with a grid builder (e.g., means_grid, cartesian_grid) or an explicit DataFrame.

Complete Workflow Examples

Example 1: Education Policy Analysis

Stata Workflow

* Fit model
logit outcome education income female urban policy_treatment

* Basic effects
margins, dydx(*)

* Effects by education
margins education, dydx(income)

* Policy scenarios (multiple commands required)
margins education, at(policy_treatment=0)
margins education, at(policy_treatment=1)

* Manual difference calculation needed for treatment effects

Julia Workflow

# Fit model  
model = glm(@formula(outcome ~ education + income + female + urban + policy_treatment),
            data, Binomial(), LogitLink())

# Basic effects
basic_effects = population_margins(model, data; type=:effects)

# Effects by education
education_effects = population_margins(model, data; 
                                     type=:effects, 
                                     vars=[:income],
                                     groups=:education)

# Policy scenarios (automatic treatment effect calculation)
policy_analysis = population_margins(model, data;
                                   type=:effects,
                                   groups=:education,
                                   scenarios=(:policy_treatment => [0, 1]))

# All results readily available as DataFrames
DataFrame(policy_analysis)

Example 2: Complex Demographic Analysis

Stata Approach (Cumbersome)

* Multiple manual commands needed:
margins education, over(region)
margins gender, over(region)  
margins education#gender, over(region)

* Income quartiles require manual creation:
xtile income_q4 = income, nq(4)
margins education, over(income_q4)

* No native support for hierarchical analysis

Julia Approach (Comprehensive)

# Single comprehensive analysis
comprehensive_results = population_margins(model, data;
    type=:effects,
    groups=:region => [:education, :gender, (:income, 4)]
)

# Results: Region × (Education + Gender + Income-Quartiles) automatically computed
# Professional Q1-Q4 labeling included
DataFrame(comprehensive_results)

Performance Comparisons

Computational Advantages

Stata Command Count Reduction

# This single Julia command:
result = population_margins(model, data;
    groups=:region => [:education, (:income, 4)],
    scenarios=(treatment=[0, 1], policy=["old", "new"])
)

# Replaces ~30 individual Stata margins commands:
# 4 regions × 3 education × 4 income × 2 treatment × 2 policy = 192 combinations
# Plus manual variable creation, looping, and results compilation

Migration Best Practices

Common Translation Patterns

# Pattern 1: Simple margins → population_margins
# margins → population_margins(model, data; type=:predictions)

# Pattern 2: Effects by groups → groups parameter  
# margins education, dydx(*) → population_margins(model, data; type=:effects, groups=:education)

# Pattern 3: Multiple at() values → scenarios or profile grids
# margins, at(x=(0 1 2)) → profile_margins(model, data, cartesian_grid(x=[0, 1, 2]))
# OR population_margins(model, data; scenarios=(x=[0, 1, 2]))  # for counterfactuals

# Pattern 4: Complex manual analysis → comprehensive single call
# Multiple Stata commands → single population_margins with groups + scenarios