Guide to Randomization

Step-by-step instructions for implementing stratified randomization in Stata, including reproducible seed setting, balance checking, and troubleshooting common issues.

Learning Objectives

Randomization is a fundamental technique in impact evaluation that ensures treatment and control groups are comparable, allowing for unbiased estimates of program effects. This guide will help you:

Understand the principles of randomization and its importance in impact evaluation
Implement stratified randomization in Stata to maintain balance across key characteristics
Set up reproducible randomization with proper seed management
Verify treatment assignment and check for balance across groups

Problem: Assign Participants to Treatment Groups

You need to randomly assign participants to treatment and control groups while maintaining balance across key characteristics like grade level, school, or other strata.

Prerequisites:

Stata installed with randtreat package
Dataset with participant identifiers and stratification variables
Basic familiarity with Stata commands

What you’ll accomplish:

Set up reproducible randomization with proper seed management
Implement stratified randomization to achieve balance
Verify treatment assignment worked correctly

Step 1: Set Up Your Environment

Set Stata version and seed for reproducibility:

Use a different seed for each project
Choose seeds through random methods—such as from random.org or dice rolls
Set only one seed per do-file
Always set the Stata version at the beginning since algorithms change between versions

Step 2: Prepare Your Data

Example scenario: School supply program evaluation

500 students across 10 schools
Budget to treat 250 students
Need to ensure balance across grades

Create your dataset with participant identifiers and strata:

%%stata
clear
set obs 500
gen student_id = _n
gen school_id = ceil(_n/50)         // 10 schools, 50 students each
gen grade = mod(_n-1,5) + 1         // Grades one to five, evenly distributed


. clear

. set obs 500
Number of observations (_N) was 0, now 500.

. gen student_id = _n

. gen school_id = ceil(_n/50)         // 10 schools, 50 students each

. gen grade = mod(_n-1,5) + 1         // Grades one to five, evenly distributed

.

Step 3: Set Random Seed

Set a random seed for replicability:

%%stata
set seed 12345

Step 4: Verify Your Strata

Check that your stratification variable covers all participants:

%%stata
tab grade


      grade |      Freq.     Percent        Cum.
------------+-----------------------------------
          1 |        100       20.00       20.00
          2 |        100       20.00       40.00
          3 |        100       20.00       60.00
          4 |        100       20.00       80.00
          5 |        100       20.00      100.00
------------+-----------------------------------
      Total |        500      100.00

Step 5: Implement Stratified Randomization

Assign half of students to treatment within each grade—total treated equals 250. Generate a treatment variable where 1 equals treated and 0 equals control:

%%stata

* Install randtreat if not already installed
cap ssc install randtreat

* Use randtreat for stratified randomization by grade
randtreat, strata(grade) unequal(1/2 1/2) generate(treatment)


. 
. * Install randtreat if not already installed
. cap ssc install randtreat

. 
. * Use randtreat for stratified randomization by grade
. randtreat, strata(grade) unequal(1/2 1/2) generate(treatment)
assignment produces 0 misfits

.

Step 6: Verify Balance

Tabulate treatment by grade and school to verify balance:

%%stata
tab treatment grade
tab treatment school_id


. tab treatment grade

           |                         grade
 treatment |         1          2          3          4          5 |     Total
-----------+-------------------------------------------------------+----------
         0 |        50         50         50         50         50 |       250 
         1 |        50         50         50         50         50 |       250 
-----------+-------------------------------------------------------+----------
     Total |       100        100        100        100        100 |       500 

. tab treatment school_id

           |                       school_id
 treatment |         1          2          3          4          5 |     Total
-----------+-------------------------------------------------------+----------
         0 |        26         27         25         24         18 |       250 
         1 |        24         23         25         26         32 |       250 
-----------+-------------------------------------------------------+----------
     Total |        50         50         50         50         50 |       500 


           |                       school_id
 treatment |         6          7          8          9         10 |     Total
-----------+-------------------------------------------------------+----------
         0 |        25         28         25         26         26 |       250 
         1 |        25         22         25         24         24 |       250 
-----------+-------------------------------------------------------+----------
     Total |        50         50         50         50         50 |       500 

.

Troubleshooting Common Issues

Not setting a random seed: Always use set seed in Stata to ensure your results are replicable.

Failing to check balance: After randomization, verify that treatment and control groups have balance on key variables.

Confusing unit identifiers: Double-check IDs—such as village, school, or participant names—to avoid misassignment.

Contamination: Monitor to prevent control group members from receiving the treatment.

Poor documentation: Keep detailed records of your randomization procedure for transparency and reproducibility.

Sorting Considerations

Ensuring Reproducible Sorting

The sort command can produce non-reproducible results if the sorting variables don’t uniquely identify observations. Always include a unique ID as the last sorting variable:

* Check if ID is unique
isid unique_id

* Sort by unique ID before generating random numbers
sort unique_id
gen rand = runiform()

* When sorting by strata and random number, include unique ID last
sort region rand unique_id

This prevents Stata from breaking ties inconsistently, ensuring your randomization is reproducible.

For more detailed guidance on randomization and power calculations in Stata, see the Stata power and sample size reference manual Release 18 ¹.

Additional Resources:

Footnotes

StataCorp. 2023. Stata power and sample size reference manual: Release 18. https://www.stata.com/manuals/power.pdf↩︎

Reuse

CC BY 4.0