Where are Filipinos moving?

I found the PSA's internal migration dataset to be hard to work with, so I wrote another one. It uses public government data and the residual method, which is a fairly basic method to estimate demographics. This work was inspired by Tobi's work on OpenHalalan (stay posted; Tobi, Quintin, Ignacio, and I have some upcoming work on estimating the economic effects of flood control projects). My website is here, and my Google Scholar is here.

Download the data

1,626towns & cities with migration estimates

8censuses per town, 1960–2020

42,041barangays with 2020 population

19years of birth & death records

2020–2024

Who gained and who lost people since 2020

The census counted everyone in May 2020 and again in July 2024. Where a town grew faster than its own births and deaths can explain, people moved in. Where it grew slower, or shrank, people moved out. Blue = people moved in, red = people moved out. Hover over any town for its numbers.

Jump to:

Boundaries: PSA/NAMRIA via philippines-json-maps (2023) drag to pan · ctrl-scroll to zoom

Gray means no estimate could be published, mostly because a boundary changed between the two censuses. Rates are people gained or lost per 1,000 residents per year.

1960–2024

The same map since 1960

Town-level birth and death records only begin in 2017, so the migration estimate above cannot be computed for earlier decades. Census counts, however, go back to 1960, and this answers a simple question: which places grew faster than the country as a whole, and which fell behind? However, this confounds migration with local differences in birth rates, so it's a rougher estimate than our data above. But large effects are still pretty clear, such as growth rates in Mindanao in the 60s and 70s, and the long pull towards Manila and its suburbs. The small maps below show all eight windows at once; click one to explore it in detail, or press play.

Boundaries: PSA/NAMRIA via philippines-json-maps (2023)

1960–1970

Four maps of provinces: growth relative to the national rate for 1990-1995, 1995-2000, 2000-2007, and 2007-2010 — The censuses of 1995 and 2007 were published at town level only on paper, so the map above steps over them, from 1990 to 2000 and from 2000 to 2010. Province totals for those censuses survive, and they split the two decades into the four shorter windows shown here. Short windows amplify differences in how well each census counted, so read these as broad patterns. Leyte's 1995 and 2007 counts and Palawan's 1990 count carry footnotes in the source and are left gray.

Findings

What the data shows

1 · People move toward the capital, but around it rather than into it

This is the longest-running pattern in the data. Cavite, Laguna, Rizal, and Bulacan, the four provinces around Metro Manila, more than doubled their share of the national population over sixty years, and between 2015 and 2020 they overtook the capital itself. The flows behind that are large: in 2015–2020 Cavite alone gained almost half a million people net (24.6 per 1,000 residents per year) while core cities like Quezon City lost people.

Line chart: Metro Manila's and its four surrounding provinces' share of the national population, 1960 to 2020 — Metro Manila's share of the national population peaked around 1990 and has drifted down since; the four provinces around it passed it between 2015 and 2020.

2 · After 2020, the map rearranged

The newest census window changes the ranking. The capital region flipped from losing about one person per 1,000 per year to gaining 3.3. Ten of its sixteen cities gained people, up from seven in 2015–2020. Meanwhile several regions that had been gaining (Mimaropa, Soccsksargen, Central Visayas) flipped to losing, and the Bicol region lost more people relative to its size than any other.

Two bar charts ranking regions by net migration rate, 2015-2020 and 2020-2024 — Regions ranked by their 2020–2024 rate, same order in both panels. Blue bars gain people, red bars lose them. In 2015–2020 almost every region shows a gain because the estimates carry a positive counting error that period; the ranking within each panel is what to read. BARMM is grayed out because its numbers mostly reflect the civil-registration gap described under limits.

Dumbbell chart of Metro Manila cities: net migration rate in 2015-2020 versus 2020-2024 — Each row is one Metro Manila city; the coral dot is its 2015–2020 rate and the navy dot its 2020–2024 rate. Most navy dots sit to the right of the coral ones. Quezon City and Marikina flipped from clear losses to gains; Valenzuela moved the other way, from the biggest gain in the region to a small loss.

3 · Disasters push people out late, not immediately

Typhoon Haiyan's landfall corridor grew at a normal rate in the census window that contained the storm, then fell 1.1 percentage points a year behind comparable towns in 2015–2020. The population cost arrived after the emergency ended. Mount Pinatubo shows the same signature stretched over the 1990s. The Bohol earthquake, after which reconstruction was fast, left no visible mark at all.

Chart comparing six events: Pinatubo, Bohol earthquake, Haiyan twice, Marawi, Odette — Each line is one disaster, measured as the growth of affected towns minus the growth of similar towns that were not affected, with a 95% confidence interval. Marawi seemed to gain during the siege (†), which I find confusing.

Method

How the estimates are made

A town's population can only really change for three reasons: people are born, people die, or people move. We have some information available to us which can rule out the first two: a census gives us the population at two dates, and the civil registry gives us births and deaths in between. So whatever part of the change is left over after births and deaths must be migration. This is called the residual method, and it is a standard demographic tool: the same logic is used to estimate populations that no survey can count directly, like undocumented immigrants in the US census (Warren and Passel 1987; Demography 2021).

For example: a town counts 10,000 people in 2020 and 11,000 in 2024, and its registry shows 1,500 births and 400 deaths. Births and deaths explain a gain of 1,100, but the town only gained 1,000. The difference, 100 people, is the number who moved away on net.

Most of the work here is bookkeeping, since I found the PSA census data hard to work with. Files are scattered across sites and formats, and it's kind of mind-boggling to keep track of all the changes in name, category, etc. In this dataset we do our best to track every such change since 1977 so that each town is compared with itself, and you get neat longitudinal comparisons.

Of course, such a method inherits the flaws of its inputs, as births and deaths that are never registered end up being counted as migration.

Scatter comparing our estimates with the 2018 national survey — There is no independent benchmark to grade these estimates against. The only national migration survey (2018) asked people where they lived five years earlier, but a person who left a province cannot be interviewed in it, so the survey misses out-migration; its own regional net flows add up to +388,000 people instead of zero. The survey and this dataset agree on the strongest patterns: the capital region loses people and Calabarzon gains them.

Limitations

Where the estimates are weak

Be careful about reading the BARMM rows too deeply. I suspect the estimates for Basilan, Sulu, Tawi-Tawi, Lanao del Sur, and Maguindanao mostly measure the reach of the civil-registration system there, as opposed to migration. PSA's 2020 census found that 77% of BARMM's population had a registered birth, against 96.6% nationally, and in the residual arithmetic a birth missing from the registry looks identical to a person moving in. The region's 2024 census growth (3.4% a year, per PSA) is four times the national rate. With records this incomplete, it's hard for us to disentangle counting effects from migration. Every affected row carries a flag.

Line chart of registered births per 1,000 people: BARMM provinces run far below every other province and the national rate — Registered births per 1,000 people. Every gray line is a province; the orange lines are BARMM provinces, running at a third of the national rate. The gap between orange and blue is unregistered births, and it is exactly what the migration estimates absorb. Maguindanao's spike after 2016 is backlog registration from the registration drives running in the region, not a birth wave.

Two more things to keep in mind. First, the estimates do not sum to zero across the country (+2.8 million for 2015–2020, +0.5 million for 2020–2024). The remainder absorbs international migration, unregistered births and deaths, and differences in how completely the two censuses counted, so comparisons between places are more reliable than any single town's level. Second, small towns have noisy estimates; rows with rates beyond ±50 per 1,000 per year are flagged as extreme in the data.

Uses & gaps

What you can do with this, and what is still missing

Things the dataset can answer today:

Which towns are growing because people are arriving, not because of births. Those are the towns that will need schools, clinics, housing, and water connections soonest.
What a disaster did to a place's population. The census panel goes back to 1960, so you can compare any dated shock against similar towns that were spared. The Haiyan and Pinatubo results above are worked examples.
Whether public works change where people move. The estimates join to any spending or project dataset through PSGC codes. We are using this to study the economic impacts of flood control projects.
Whether global population models get the Philippines right. WorldPop and similar products publish modeled migration estimates, and this is census-based data to check them against.
Population change at the barangay level. 35,634 barangays have verified 2010 and 2020 populations, the smallest units in the country where change can be measured.

What is missing, and where help is welcome:

Where people move between. These are net numbers, each town's arrivals minus its departures. No public origin-to-destination table exists at any useful scale. The census tables that could partly fill this are listed in the data inventory.
Who moves. There is no age or sex breakdown yet. Census age tables would support estimates by age group.
Anything between censuses. Estimates exist only for census windows. Annual records like school enrollment or voter registration could fill the years in between.
Older municipal records. Births and deaths before 2017 exist only in PDF publications, so a freedom of information request to PSA is drafted. The primary 2024 municipal census file would replace our verified transcription.

Download

The files

municipal_master.csv

Start here: one row per town and city with its full census history 1960–2024, registered births and deaths, and the migration estimate. Everything below is joined in this one file.

net_migration_municipal_2020_2024.csv

Net migration for 1,626 towns and cities, 2020–2024, with a flag column for the caveats above.

net_migration_province.csv

The same for provinces and highly urbanized cities, for 2015–2020 and 2020–2024.

population_census_municipal.csv

Every town's population in the 8 censuses from 1960 to 2020, checked across two PSA publications.

population_barangay_2020.csv

All 42,041 barangays with their 2020 population. 35,634 of them also have their 2010 population, verified against official municipal totals.

vital_statistics_municipal.csv + vital_statistics_provincial.csv

Registered births and deaths: towns 2017–2024, provinces 2006–2024. Missing years are left blank and flagged, never filled in.

municipalities.csv + psgc_changes.csv

The reference list of 1,642 towns and cities, plus 4,688 boundary and name changes since 1977.

atlas/

Ranked tables of the largest gains and losses, and event tables for Haiyan and Marawi.

Everything can be rebuilt from the raw sources with one command: python3 pipeline/run_all.py. The origin and checksum of every downloaded source file is logged in data/provenance.jsonl. License: ODbL for the data, MIT for the code.

Citation

Citing this dataset

@misc{africa2026philippinemigration,
  author = {Africa, David Demitri},
  title  = {Philippine Internal Migration Dataset},
  year   = {2026},
  url    = {https://github.com/DavidDemitriAfrica/philippines-internal-migration},
  note   = {Net internal migration estimates for every Philippine city and
            municipality, built from public PSA data}
}

If you're interested in the other work I do, you can check out my website here or my Google Scholar here. If you're interested specifically in work I do on Filipino or Tagalog, we released Batayan (Batayan: A Filipino NLP Benchmark for Evaluating Large Language Models) last year and PACUTE (Phonology-, Affix-, and Character-level Understanding of Tokens for Filipino) last month. I also did some work on meta-learning to improve adaptation to low-resource languages in small LMs (Learning Dynamics of Meta-Learning in Small Model Pretraining and Meta-Pretraining for Zero-Shot Cross-Lingual Named Entity Recognition in Low-Resource Philippine Languages).