A Pesquisa Nacional Por Amostra De Domicílios

Getting data from PNAD isn't as simple as it sounds

A pesquisa nacional por amostra de domicílios is IBGE's flagship household survey. It's been running since 1967, and if you're working with Brazilian socioeconomic data, you will eventually need it. The common mistake people make is assuming it's just one dataset. It's not. There's the original PNAD Continuada (monthly), the older annual PNAD samples that go back decades, and then there's the microdados you actually download. They don't line up perfectly on variable names, geography codes, or weighting schemes. I've spent years cleaning and merging these files, and the thing that catches everyone off guard is the territorial reclassification. IBGE updates municipal codes almost every year. If you're pulling data from 2015 and 2022 and trying to track a specific municipality across both, you'll hit rows that vanish because the code changed or the municipality was split. The workaround is to always join on the codmun_ibge variable rather than the municipality name string, and then cross-reference against the official IBGE temporal geography table, which they publish separately. It adds maybe ten minutes to your pipeline but saves you from spending three days debugging missing observations.

Accessing a pesquisa nacional por amostra de domicílios microdados

The data lives at the PNAD Contínua section of the IBGE sidra platform. You can download the full microdataset directly — no registration required for basic access. For the older annual PNAD samples, you go through the same portal but under the "Pesquisas" menu, then "Domiciliares." The files are released as compressed SPSS, STATA, or SAS datasets depending on what you need. If you're doing anything heavier than a quick cross-tabulation, grab the STATA version. The variable labels are cleaner and the data structure is more consistent across years. The weighting variables are where most people stall out. Every PNAD file comes with peso_final (the final household weight) and sometimes a person-level weight depending on the module. If you run any analysis without applying these weights, your estimates will be biased, often significantly. IBGE uses a complex sampling design with stratification by macroregion and urban/rural classification, so unweighted results won't match official IBGE publications. The difference isn't trivial — in my experience, unweighted poverty rate estimates can deviate by two to four percentage points from the weighted figures, and that gap widens in states with smaller sample sizes like those in the North region.

The variables you actually need vs. what's in the file

The raw files contain dozens of variables, but the useful ones cluster into a few categories. Geography is handled through the uf, regiao, and area (urban/rural) fields. Demographics use standard codes for age brackets, sex, and education levels, though the education variable has undergone multiple reclassifications across survey years. The labor module, which is what most researchers actually use, tracks employment status, income brackets, and formal/informal classification. The tricky part is that the labor questions changed structure between 2012 and 2016 when IBGE switched from the old PNAD to the continuous PNAD model. Before 2016, the employment questions were structured differently and some variables don't map 1-to-1. You'll need the official recoding tables IBGE publishes for each transition year. One counter-intuitive thing about PNAD: the sample is designed to be representative at the state level, not at the municipal level for small cities. If you're working with a municipality under 100,000 inhabitants, your effective sample size drops considerably and confidence intervals widen substantially. I learned this the hard way when I published a paper on informal employment in a small interior city in Minas Gerais. The point estimates looked fine until I calculated the design effect and realized the confidence interval spanned nearly eighteen percentage points. You can mitigate this by aggregating across multiple survey periods — the continuous PNAD gives you twelve months of data per year — but even pooling three years of data for a small municipality rarely gets you below a five-point margin of error at the 95 percent level.

Common pitfalls I've seen repeat myself

People frequently forget that PNAD excludes institutions. If you're studying prison populations, military barracks, or hospital long-stay residents, that entire segment is invisible in the data. It's a minor issue for most labor market analyses, but it matters if you're researching unemployment among vulnerable populations. Another frequent error is double-counting household members when merging across modules. The PNAD structure uses a head-of-household file linked to individual respondent files, and if you join incorrectly, you'll inflate your N. Always verify your household ID variable matches exactly across files before merging. The sampling weight also changes meaning depending on which module you're using. The employment module weight is not identical to the general demographic weight. IBGE provides separate weight variables for each thematic module, and using the wrong one introduces bias. I recommend always checking the technical notes that accompany each release — they're brief but they flag exactly which weight applies to which set of variables. Without them, you're guessing, and guessing with survey data is how you get numbers that look plausible but are technically wrong. PNAD has real limitations. The income questions rely on self-reporting, which means underreporting is systematic and varies by income level. High earners underreport significantly more than low earners. The survey also doesn't capture the informal economy well — street vendors, day laborers without formal arrangements, and people working in household production often fall through the gaps. If your research question is about the full scope of informal work, PNAD will give you a floor, not a ceiling. For that, you'd need to supplement with the Pesquisa de Orçamentos Familiares or specialized studies like the PNAD Special Module on Informality, which IBGE occasionally releases.

Practical tips that come from actually using this data

Download the recoding manuals before you start any analysis. They'rePDF files on the same page as the data, and they save hours of guessing what a variable code means. The geographic codes change, but the manual explains each reclassification. Also, always keep a copy of the raw data untouched. I've seen too many people recode variables in place and then lose the ability to go back and verify their logic. Store the original file separately, create a new cleaned version, and document every transformation in a separate notes file. If you're working in R or Python rather than STATA, be aware that the label metadata from IBGE's proprietary formats doesn't always carry over cleanly. You'll need to rebuild your variable dictionaries from the documentation. It's tedious but necessary — relying on unlabeled numeric codes is how you accidentally analyze age instead of education level. The time investment upfront, roughly thirty to forty-five minutes per dataset, prevents days of errors later. And if you're combining multiple years, do it year by year and validate each merge before proceeding to the next one. Cross-year consistency checks catch more problems than any single comprehensive merge ever would. The data is freely available and one of the best sources for Brazilian socioeconomic analysis, but it demands respect for its structure. Treat it like a technical instrument rather than a simple spreadsheet, and it will serve you well. Handle it carelessly and you'll get results that look authoritative and aren't.