MEPS data isn't a single spreadsheet. It's more like a library of interconnected files. Each file contains a different type of information, but they all relate back to the people surveyed. This structure keeps the data organized and manageable.

The data is organized hierarchically. At the top, you have the main person-level file. This file, often called the Full-Year Consolidated (FYC) file, contains one record for each person who participated in the survey for the full year. It includes their demographic information, health insurance coverage, and summary health care spending.

Beneath the person-level file are various event-level files. These files contain detailed information about each specific health care encounter. There are separate files for different types of events, such as doctor's office visits, hospital stays, and prescribed medicines. For example, if a person visited the doctor three times and filled five prescriptions, they would have one record in the person-level file, three records in the office-visit event file, and five records in the prescribed medicines file.

There are also condition-level files, which list the specific medical conditions reported by each person, and job-level files for employment information.

To connect the information across these different files, MEPS uses unique identifiers. The most important one is the person identifier, DUPERSID. This variable is present in almost every file and allows you to link a person's demographic data to all their medical events, conditions, and jobs.

Each MEPS file contains a host of variables. While the person-level file gives you the big picture, the event files provide the granular details. Here are some of the most commonly used public use event files:

Within these files, you'll find variables covering everything from the reason for a visit and the services provided to the total cost of care. Expenditure variables are particularly important. They detail how much was paid for a service and who paid for it—for instance, how much was paid by Medicare, private insurance, or out-of-pocket by the family.

Because MEPS is a survey of a sample of the population, not a census of every single person, your analysis must account for the survey's design. This is done using two special types of variables included in the data: survey weights and variance estimation variables.

The person-level weight variable (often PERWT##F, where ## is the year) indicates how many people in the U.S. each respondent represents. When you calculate a statistic, like the average annual health care spending, you must apply this weight. Failing to do so would be like reporting the average spending of only the few thousand people in the sample, rather than an estimate for the entire nation.

Additionally, the complex sampling design of MEPS means you can't calculate standard errors or confidence intervals using standard statistical formulas. You must use variables specifically designed for variance estimation, typically VARSTR (variance estimation stratum) and VARPSU (variance estimation primary sampling unit). These variables help statistical software like Stata account for the survey's structure to produce accurate standard errors.

Don't worry about the specific commands for now. The key takeaway is that you must always use these special variables to ensure your analysis is statistically sound and representative of the U.S. population.