social_ES

A Python library to ingest, clean, and transform up-to-date Spanish demographic, socioeconomic, and other social-related datasets from multiple data source entities. At this moment, only INE (Instituto Nacional de Estadística — Spanish National Statistics Institute) is supported.

🎯 Overview

social_ES automates the discovery, download, and cleaning of official Spanish statistics from INE's data portal. Instead of manually navigating INE's website and wrangling raw CSV/TSV exports, each function in the library scrapes the relevant dataset, normalizes column names and geographic codes (autonomous community, province, municipality, district, census tract), and returns a dictionary of pandas.DataFrames keyed by geographic level (e.g. "Census tracts", "Districts", "Municipality", "Province", "National"), with the level(s) available depending on the dataset.

Downloaded and processed data is cached locally in your working directory, so subsequent calls reuse the cached files instead of re-downloading from INE.

Key Features

• 📥 Automated INE Scraping: Discovers and downloads the latest published data directly from INE's dissemination portal, no manual exports needed
• 🧹 Cleaning & Normalization: Consistent column naming, locale-aware numeric parsing (INE publishes the same table in Spanish 24.900 and English 24,900 formats depending on the province), and geographic code harmonization
• 🗂️ Local Caching: Results are persisted as .tsv/.parquet files under your working directory to avoid redundant downloads
• 🌍 Multi-level Geography: Data available at autonomous community, province, municipality, district, or census-tract level depending on the dataset
• 🔎 Filtering: Most functions support filtering by municipality_code and years
• 🔗 hypercadaster_ES Integration: EssentialCharacteristicsOfPopulationAndHouseholds links Census 2021 indicators to building-level data exported from hypercadaster_ES
• 🏚️ Historical Census Series: EmptyAndSecondaryDwellingsCensus puts the 2001, 2011 and 2021 dwelling-use counts in one table, keeping the field-census and the 2021 electricity-based classifications apart rather than pretending they are the same variable
• 🧩 Cross-dataset Join Keys: HouseholdIncomeDistributionAtlas carries an inflation-adjusted Household income group that matches the bands of TimeUseSurvey, so income and time-use data join directly
• 🗺️ Boundaries & Maps: AdministrativeBoundaries downloads INE's census-tract cartography for any published year and dissolves it into districts, municipalities, provinces and autonomous communities; MapVariable joins any variable of any dataset to it and writes a standalone interactive HTML choropleth

🚀 Installation

Install from PyPI:

pip install social_ES

# with the boundaries and mapping functions, which need geopandas
pip install "social_ES[geo]"

Or, for development from source:

git clone https://github.com/BeeGroup-cimne/social_ES.git
cd social_ES
pip install .

📖 Quick Start

from social_ES import INE

# Define a working directory where downloaded/processed data will be cached
wd = "/path/to/your/data"

# Household income distribution at census-tract level (returns dict with "Census tracts", "Districts", "Municipality")
atlas = INE.HouseholdIncomeDistributionAtlas(wd=wd)
atlas_sections = atlas["Census tracts"]  # census-tract level data

# Population census, filtered to a specific municipality and years (returns dict with "Census tracts", "Districts", "Municipality")
population = INE.PopulationCensus(wd=wd, municipality_code="08019", years=[2021, 2022])
population_sections = population["Census tracts"]  # census-tract level data

# Education and employment census (relative shares; returns dict with "Census tracts", "Districts", "Municipality")
education = INE.EducationAndEmploymentCensus(wd=wd, mode="relative")  # mode defaults to "relative"
education_sections = education["Census tracts"]

# Population and dwellings census 2021 (returns dict with "Census tracts", "Districts", "Municipality")
census_2021 = INE.DwellingsAndPopulationCensus(wd=wd)
census_2021_sections = census_2021["Census tracts"]

# Empty and secondary dwellings across the 2001, 2011 and 2021 censuses
# (returns dict with "Municipality", "Province", "Autonomous community")
dwelling_use = INE.EmptyAndSecondaryDwellingsCensus(wd=wd, municipality_code="08019")
dwelling_use["Municipality"][["Year", "Percentage of dwellings ~ Comparable use:Main",
                              "Percentage of dwellings ~ Comparable use:Secondary",
                              "Percentage of dwellings ~ Comparable use:Empty"]]  # one series across the 3 censuses
dwelling_use["Autonomous community"]  # regional figures — complete, unlike summing the municipal table

# Consumer Price Index by category (returns dict with "National")
cpi = INE.ConsumerPriceIndex(wd=wd)
cpi_national = cpi["National"]

# Join income data to time-use profiles: both datasets share the same income banding
time_use = INE.TimeUseSurvey(wd=wd)
weekly_by_tract = atlas["Census tracts"].query("Year == 2021").merge(
    time_use["WeeklySchedule"], on=["Autonomous community code", "Household income group"])

Full documentation — one page per function, with the levels it returns, its parameters and its columns.

See examples/get_ine.ipynb for a full worked example, including the output format of each function.

📊 Available Datasets

Function	Description	Geographic level	Key arguments
RelationAutonomousCommunityAndProvince()	Static lookup table mapping autonomous community codes/names to province codes/names	Province	—
MunicipalityNamesToMunicipalityCodes()	Official INE dictionary of municipality names and codes	Municipality	—
HouseholdIncomeDistributionAtlas(wd, municipality_code, years)	Household income distribution (Atlas de Distribución de Renta de los Hogares)	Census tract / district / municipality	municipality_code, years
PopulationCensus(wd, municipality_code, years)	Population census counts	Census tract / district / municipality	municipality_code, years
EducationAndEmploymentCensus(wd, municipality_code, years, mode)	Education level and employment status	Census tract / district / municipality	mode="relative"|"absolute"
DwellingsAndPopulationCensus(wd, municipality_code, years)	Population and dwellings census 2021: population profile (sex, age, nationality, education, labour force status, marital status) and dwelling, tenure and household size counts	Census tract / district / municipality	municipality_code, years (fixed to 2021)
EmptyAndSecondaryDwellingsCensus(wd, municipality_code, years, predict)	Empty, secondary and non-main dwellings in the 2001, 2011 and 2021 censuses, as counts and as shares of the area's dwellings	Municipality / province / autonomous community	municipality_code, years (2001, 2011, 2021)
HouseholdsPriceIndex(wd, municipality_code, years)	Housing price index (whole, new, and second-hand market)	Province (derived from autonomous community), quarterly	municipality_code, years
HouseholdsRentalPriceIndex(wd, municipality_code, years)	Housing rental price index	Municipality / district	municipality_code, years
AggregatedElectricityConsumption(wd, municipality_code, years)	Aggregated electricity consumption percentiles (2021)	District (municipality code retained)	municipality_code, years
ConsumerPriceIndex(wd, years)	Consumer Price Index (CPI) broken down by COICOP category	National	years
EssentialCharacteristicsOfPopulationAndHouseholds(wd, hypercadaster_ES_input_pkl_file, hypercadaster_ES_input_gdf)	Census 2021 population and household characteristics, linked to building-level data	Building (via hypercadaster_ES)	hypercadaster_ES_input_pkl_file, hypercadaster_ES_input_gdf
TimeUseSurvey(wd, municipality_code, years, reference_year)	Time Use Survey (EET 2009-2010): weekly hourly activity shares + yearly daily holiday schedule, linked to census tracts via autonomous community + household income band	Census tract (matched to autonomous community + income band)	municipality_code, reference_year (note: years accepted but ignored — survey fixed to 2009-2010)

Most functions accept municipality_code as either a single code (str) or a list of codes, and years as a list of years to filter to. When omitted, the full dataset is returned.

2021 census indicators: DwellingsAndPopulationCensus reads INE's single C2021_Indicadores.csv file, published at census-tract level only; the "Districts" and "Municipality" tables are aggregated by social_ES, summing the counts and averaging each share weighting by the population it is a share of (total population, population aged 16 and over, or active population, depending on the indicator), so that all three levels are read the same way. Shares are returned as percentages (0-100) rather than the proportions of the source file, and Average age is left in years. INE suppresses the indicators of the smallest census tracts for statistical confidentiality: those rows keep their Population and are NaN elsewhere, and they are skipped when the coarser levels are built, so the totals fall slightly short of the published national ones.

Empty and secondary dwellings: the 2021 census dropped the classification the 2001 and 2011 ones used, so EmptyAndSecondaryDwellingsCensus returns two side by side and never mixes them.

Dwellings ~ Dwelling type:* is the field-census classification, where an agent visiting the building sorted each dwelling into main, secondary (occasional use, e.g. holidays) or empty. Main and Non-main are published in all three censuses; Secondary and Empty exist in 2001 and 2011 only, plus an Other non-main residual in 2001. The 2021 census is built from administrative registers alone, with nobody to ask whether a dwelling is a second home, so it publishes only main/non-main and both are NaN.

Dwellings ~ Electricity use:* is what INE publishes in its place for 2021, derived from the yearly electricity consumption of each dwelling: Empty (no supply contract, or less than the equivalent of 15 days a year for that municipality), Very low consumption (up to 250 kWh), Sporadic use (251-750 kWh, roughly one to three months a year) and Regular use. The four partition the total.

Dwellings ~ Comparable use:* aligns the two into a single series that can be read across the three censuses, filled from whichever classification each census published:

Comparable use	2001 / 2011 (field census)	2021 (electricity)
Main	Dwelling type:Main	Electricity use:Regular use
Secondary	Dwelling type:Secondary (+ Other non-main, in 2001)	Electricity use:Very low consumption + Sporadic use
Empty	Dwelling type:Empty	Electricity use:Empty

The three classes partition the total in every census, so their shares always add to 100, and they are returned as integer counts of dwellings. One choice is worth spelling out: 2021's Very low consumption (up to 250 kWh, about a month of use) counts as secondary, not empty — a dwelling used a month a year is what the earlier censuses recorded as a second home, and reading Sporadic use alone would understate it. A row takes all three classes from one classification or from none, so the series is never half-filled from one census and half from another.

2001 is the only census that splits an "otro tipo" residual out of the non-main dwellings, and it is counted in Dwelling type:Secondary rather than returned on its own — which is what INE itself does in its published 2001–2011 comparison (3,360,631 + 292,332 = 3,652,963). Keeping it apart would make the 2001 secondary count read lower than the later ones for no reason other than the questionnaire.

Filling the gaps INE leaves (predict=True): the 2011 split is published only for municipalities over 2,000 inhabitants and the 2021 electricity classification only for those over 1,000, so most municipalities carry NaN. What is missing is never the whole picture, though — the total each missing piece belongs to is published for the municipality (2011 gives every non-main count, 2021 every dwelling count), the province total of every missing column is published too, and the municipal counts of those totals add up to their province exactly in all 52 provinces. So a model only has to say how a known quantity divides, and gradient-boosted trees do that from what INE publishes everywhere: the dwelling counts of all three censuses, the 2001 split (the only one that exists for every municipality), and the 2021 census indicators describing the place. Each predicted share is then scaled until its province adds up to the published figure exactly, so summing the municipal table reproduces the regional one — which, with the gaps left as they come, it does not.

Held out five ways, the shares are predicted at R² 0.72 (2011 secondary) and 0.76 (2021 empty), against 0.00 for giving every municipality the national rate. It is still an estimate: a municipality of forty dwellings is estimated no better than that. The Predicted column says of every row whether any of its figures were modelled, MapVariable turns it into a show/hide control on the map, and without predict=True nothing is filled in at all.

⚠️ This is a bridge, not an identity. The 2021 numbers come from electricity meters, the earlier ones from a census agent's judgement at the door, so any movement between 2011 and 2021 along these lines mixes a real change in use with the change of instrument. It bites hardest in tourist municipalities, where a second home occupied for a full summer consumes well over 750 kWh and is classified as Main. When the published figures are what matter, use Dwelling type:* and Electricity use:*, which are always one column away.

Nationally the harmonised series reads:

Comparable use	2001	2011	2021
Main	14,187,169	18,083,693	19,336,136
Secondary	3,652,963	3,681,566	3,459,265
Empty	3,106,422	3,443,365	3,828,307
Total	20,946,554	25,208,624	26,623,708

Municipal coverage follows each census's methodology and is deliberately left as NaN where INE publishes nothing:

Census	Dwellings, Main, Non-main	Secondary / Empty	Electricity use:*	Comparable use:*
2001	8,108 municipalities (all)	8,108 (all)	—	8,108 (all)
2011	8,116 municipalities (all)	2,308 (over 2,000 inhabitants)	—	2,308
2021	8,131 municipalities (all)	—	3,139 (over 1,000 inhabitants, ~97% of the population)	3,139

The municipalities left out are only released aggregated per province, under pseudo-codes like 01999 Resto de Araba/Álava, which are dropped since they are not municipalities.

⚠️ Do not sum the municipal table to get regional figures. For the partially-published columns it falls short by construction — nationally it recovers only 85.9% of the 2021 empty dwellings, and as little as 48% in Castilla y León and 50% in Aragón, whose parks are concentrated in small municipalities. Use the "Province" and "Autonomous community" tables, which read those columns from INE's own regional tables and are complete.

The "Province" and "Autonomous community" tables reproduce INE's published regional figures exactly (verified against every autonomous community for the three censuses and against the four Catalan provinces): 2001 to the unit, 2021 to the unit, and 2011 to within the ±1 of INE's own rounding of its sample-based estimates.

Every count is echoed as Percentage of dwellings ~ ..., its share of the area's Dwellings, in percent (0-100). Municipality codes are the ones each census was published with, so municipalities created, merged or renamed between 2001 and 2021 do not line up across years — see INE's list of alterations.

One more comparability trap: INE's own 2001–2011 series counts the 2001 Other non-main residual as secondary, so reproducing it means adding Secondary + Other non-main (3,360,631 + 292,332 = 3,652,963 nationally). They are kept apart here because 2001 published them apart.

Joining income data: HouseholdIncomeDistributionAtlas adds a Household income group column (plus its Household income group label and the Average household net income (2010 EUR) it is derived from). Each row's nominal average net household income is deflated to 2010 prices with the general CPI (ConsumerPriceIndex) and then bucketed into the four bands the Time Use Survey respondents answered in — 1,200 € or less, 1,201 to 2,000 €, 2,001 to 3,000 €, More than 3,000 € — so Atlas rows of any year can be joined straight onto the TimeUseSurvey schedules on ["Autonomous community code", "Household income group"].

Deflating matters: a tract whose income grows only with inflation keeps the same group across years, instead of drifting upwards. In 2023 nominal terms the band edges sit at roughly 1,529 €, 2,549 € and 3,823 € per month.

ℹ️ The EET microdata are anonymised with Ceuta and Melilla merged into a single community, so their joint time-use profile is published under both standard codes (18 and 19). Melilla tracts join like any other; just note that their schedules are shared with Ceuta rather than measured separately.

🗺️ Boundaries and Maps

Two functions turn the tables above into geography. They need geopandas, which the base install deliberately leaves out: pip install "social_ES[geo]".

Function	Description
AvailableBoundaryYears()	The years INE publishes cartography for (2001, and 2003 onwards), and the file of each
AdministrativeBoundaries(wd, year, level, municipality_code, province_code, autonomous_community_code)	A GeoDataFrame of boundaries in WGS84, at any of the five levels
MapVariable(data, variable, wd, ...)	Writes an interactive HTML choropleth of a dataset — one variable, or every one of them behind a picker
BoundaryTiles(wd, year, level)	Builds and caches the PMTiles vector-tile archive of a level, for MapVariable(tiles=True)
ServeMaps(wd, port)	Serves the written maps and their tile archives over HTTP, byte ranges included

from social_ES import INE

wd = "/path/to/your/data"

# The 2021 census tracts of Barcelona, ready to join any census-tract table onto
tracts = INE.AdministrativeBoundaries(wd=wd, year=2021, level="Census tracts", municipality_code="08019")

# A map of one variable, straight from what a dataset function returned
census = INE.DwellingsAndPopulationCensus(wd=wd, municipality_code="08019")
INE.MapVariable(census, "Percentage of population aged 16 and over ~ Educational level:Tertiary education",
                wd=wd, title="Tertiary education, Barcelona")

# Leave the variable out and every one of them goes into the page, behind a picker
INE.MapVariable(census, wd=wd)

# Several years become a slider, with the classes computed over all of them at once
dwellings = INE.EmptyAndSecondaryDwellingsCensus(wd=wd)
INE.MapVariable(dwellings, "Percentage of dwellings ~ Comparable use:Empty",
                wd=wd, level="Autonomous community")

# A value read against a reference rather than as a magnitude
INE.MapVariable(dwellings, "Percentage of dwellings ~ Comparable use:Empty", wd=wd,
                level="Municipality", year=2021, palette="diverging", center=15)

Where the boundaries come from: INE publishes the georeferenced contours of every census tract of the country, one national file per year, at its open data portal — for 2001 and for every year from 2003 to the current one. Only census tracts are published: districts, municipalities, provinces and autonomous communities are exact unions of tracts, and social_ES builds them by dissolving, so every level nests inside the coarser ones without slivers or mismatched coastlines. The 2021 file yields 36,333 tracts, 10,479 districts, 8,131 municipalities, 52 provinces and 19 autonomous communities.

The published files come in two layouts, both normalised to the column names the dataset functions use and reprojected to WGS84 (EPSG:4326): from 2011 on, a single national shapefile in ETRS89 / UTM 30N carrying the codes and the community, province and municipality names; before that, two shapefiles in different UTM zones (peninsula plus Balearics, and the Canaries) carrying the codes alone, whose names are filled in from the library's own lookups.

⚠️ Boundaries are redrawn every year. Tracts are split as population grows and municipalities merge or are renamed, so a code only means the same area within a year or two of its cartography. AdministrativeBoundaries takes the year rather than choosing one, and MapVariable defaults it to the most recent year being mapped. Areas of the data with no boundary in that year are reported and left off the map.

What the map is: a single self-contained HTML file — the geometry is embedded, so nothing but the basemap tiles and MapLibre itself is fetched when it is opened. The areas are drawn as vectors on the GPU, so they stay sharp at any zoom and the zoom itself is continuous rather than a ladder of whole levels. Hovering an area gives its name, code and value; several years become a slider, with the classes computed over every year at once so the colours mean the same thing at each stop; and the values are also written out as a table, since a colour is not a readable number. The polygons use a single-hue blue ramp for a magnitude (palette="sequential", the default), blue-to-red around a midpoint for a value read against a reference ("diverging", with center), and a fixed categorical order for a non-numeric variable, which is detected and switched to automatically. "greens", "oranges" and "purples" are the same ramp in another hue — the lightness steps are shared, so only the colour changes and never how far apart two classes look — and "viridis" runs the other way, dark to light, stopping short of its yellows, which are the part of it a light basemap swallows.

A diverging map mirrors its classes around center, which defaults to the median of the values. Zero is the midpoint of a change or a difference, but of nothing that is only ever positive: centred on zero, a percentage spends half its classes below the smallest value there is, and the legend reads in negative numbers. Pass center=0 for the variables whose midpoint really is zero. The mirror reaches only as far as the shorter side of the data allows, and the two outer classes carry whatever lies past it.

Opacity, borders and the tooltip are page controls too. opacity sets how solid the areas are drawn (areas with no value stay fainter, at the same ratio); borders draws the white hairline between areas, and defaults to off on a tiled map, where every area is cut at the tile boundaries and outlining the pieces would draw the tile grid across the map; the tooltip that follows the pointer can be switched off from the legend.

Changing the map without rewriting the file: the legend carries the classification itself — the palette, the number of classes, and whether they are cut at quantiles or at equal intervals — plus the opacity, the borders and the tooltip, and the page cuts the values again as those change. palette, bins and classification set where the controls start rather than settling the matter, so trying six classes instead of four, or linear instead of quantiles, costs a click rather than another call. Class boundaries passed in by hand (classification=[0, 100, 1000]) are offered as Custom and are what the map opens on; asking for any other cut replaces them. A non-numeric variable has no boundaries to move, so the controls are hidden for it.

A table carrying a Predicted column — EmptyAndSecondaryDwellingsCensus(predict=True) — gets one more: a checkbox that drops the modelled areas from the map and from the table, so what INE actually published can be seen on its own. The tooltip says so on the area itself. The button in the map's top corner folds the panels away entirely, and panels=False opens the page with them already folded.

Size and detail: past 200 KB the payload is gzipped and base64'd into the page, and unpacked in the browser with DecompressionStream (Chrome/Edge 80+, Firefox 113+, Safari 16.4+); below that it stays plain JSON, readable in an editor. Because compression pays for detail, boundaries are simplified by how much geometry there is rather than by how much of the country is on screen: anything inside the vertex budget keeps a 1 m tolerance, which is a pixel at zoom 17 and invisible, and only what overruns is coarsened. Census tracts of a city and provinces are drawn at 1 m and 5.5 m; the 8,131 municipalities of the country, four million vertices of coastline, at 77 m. Pass simplify_tolerance=0 for exactly what INE published. The result is both finer and smaller than reading the geometry uncompressed:

Page	Before	Now	Tolerance
Barcelona census tracts, one variable	0.41 MB	0.14 MB	5 m → 1 m
Barcelona census tracts, all 36 variables	0.76 MB	0.26 MB	5 m → 1 m
Autonomous communities, all variables, 3 years	0.86 MB	2.60 MB	100 m → 4.4 m
Spain municipalities, one variable	9.29 MB	4.95 MB	100 m → 77 m

The areas are one GPU layer rather than one element each, and the data reaches them as feature state — the class each area falls in — with the colours following from a paint expression. Changing the palette, the number of classes or the year rewrites that expression instead of restyling thousands of shapes, and never touches the geometry.

Vector tiles, for the maps that don't fit

A page can only carry so much geometry. The 36,333 census tracts of the country are close to seven million vertices, and coarsening them enough to embed would need a tolerance of 222 m against areas whose median extent is 822 m — which flattens a census tract into a blob. MapVariable refuses that rather than drawing it, and points here.

tiles=True cuts the geometry into vector tiles the browser fetches as it draws them, so what a page holds stops depending on how much of the country it covers:

census = INE.DwellingsAndPopulationCensus(wd=wd)

# Builds {wd}/INE/AdministrativeBoundaries/census_tracts_2021.pmtiles the first time,
# then reuses it for every later map of that level and year
INE.MapVariable(census, wd=wd, level="Census tracts", tiles=True)

server = INE.ServeMaps(wd=wd)          # http://127.0.0.1:8000/Maps/
# ... open the printed URL ...
server.shutdown()

The archive is built over the whole country, whatever the map that asked for it draws, so one archive serves every map of its level and year — a map of one city and a map of all of Spain read the same file. It is cached beside the cartography it comes from and rebuilt only if deleted. Values stay in the page and are joined to the tiles by the area code each tile feature carries, which is what lets one archive serve maps of different variables, years and datasets.

⚠️ A tiled map has to be served. A page opened from a file:// URL is not allowed to fetch anything, tiles included — so tiles=True maps are opened through ServeMaps (or any server that answers byte-range requests), not by double-clicking. Maps without tiles=True stay self-contained files. The page reads the PMTiles archive in place, asking for the ranges holding the header, the directory and the tiles it draws, so the level stays the one cached file it was built as and nothing has to unpack it on the way out.

Level (2021, whole country)	Areas	Embedded page	Tiled page	Archive, built once
Districts, one variable	10,479	~10 MB	0.16 MB	21.7 MB, 13,201 tiles, 121 s
Census tracts, all 36 variables	36,333	refused	3.85 MB	29.9 MB, 13,201 tiles, 205 s

(Both archives hold the same 13,201 tiles: the tiles are the ones covering Spain, and only what is inside them differs.) The archive is built to zoom 12, where a tile unit is about 2 m; past that the browser scales the tiles it has, so zooming to street level costs nothing further.

Leaving variable out puts every mappable column of the table into the page, behind a picker, so the map can be read without deciding beforehand which variable was the interesting one — the heading, the legend, the colours, the tooltips and the table all follow it. Each variable is classified on its own, so a count and a percentage each get their own classes, and a text column is drawn as categories while its neighbours are drawn as magnitudes. Columns that identify the area (codes, names, Year) are left out, as are columns that hold nothing and text columns with more than 50 distinct values. variable also takes a list, to offer a chosen few. Geometry dominates the file size, so the picker is close to free: the 1,068 census tracts of Barcelona are 140 KB with one variable and 260 KB with all 36.

MapVariable accepts either the whole dictionary a dataset function returned — picking the table by level, or the finest one it holds — or a single DataFrame, whose level it reads off the code columns. It refuses to write more than max_areas (25,000) areas, or more than max_cells (500,000) areas × variables × years, into one page — a browser would struggle to open either; filter the data, or name the variables you want.

Return format: Most functions return a dictionary with keys like "Census tracts", "Districts", "Municipality", "Province", or "National" holding the corresponding pandas.DataFrame. The exception is EssentialCharacteristicsOfPopulationAndHouseholds, which returns a dictionary, sorted by key, where each key is an English snake_case topic name (e.g. "main_dwellings_heating_type") and each value is a building-level DataFrame whose columns append the class after a ~ (e.g. "main_dwellings_heating_type~individual"). ECEPOV publishes a few dwelling attributes twice, once counting households and once counting main dwellings; only the main_dwellings_* variant is returned, since it carries the same information (this drops households_n_rooms and households_floor_area).

💾 Caching Behavior

On first call, each function downloads the relevant data from INE and stores a processed copy under:

{wd}/INE/{FunctionName}/

Depending on the function, cached files may be .tsv, .parquet, or .pkl (for metadata dictionaries). Subsequent calls with the same wd read from this cache instead of hitting INE again. To force a refresh, delete the corresponding cache file/folder.

Cartography is the big one: the first call to AdministrativeBoundaries for a year downloads a ~60 MB shapefile and writes one GeoParquet per level under {wd}/INE/AdministrativeBoundaries/, about 345 MB per year in total, in exchange for every later call being a local read. Maps are written to {wd}/INE/Maps/ unless output_file says otherwise.

Cache versions: some cache files carry a _v2/_v3 suffix. When a parsing fix makes previously cached data wrong, the suffix is bumped so the next call rebuilds it instead of silently reusing bad values; the superseded file is left in place and can be deleted by hand.

• HouseholdIncomeDistributionAtlas/df_v3.tsv supersedes df_v2.tsv. INE serves this dataset in two number formats depending on the province — Spanish (24.900) and English (24,900) — and the earlier parser read the English ones as a thousandth of their value (Córdoba, Guadalajara, Lleida and Asturias were almost entirely affected), besides truncating values whose trailing digits were zeros. Expect a one-time full re-download on the first call after upgrading.
• EssentialCharacteristicsOfPopulationAndHouseholds/*_v2.pkl supersedes the unsuffixed metadata caches, which were built with a variable-splitting bug.
• TimeUseSurvey/*.tsv hold the profiles exactly as the microdata support them, so they contain no community 19: Melilla is filled in from Ceuta when the files are read, and no rebuild is needed.

🎯 Key Applications

• Urban & Regional Analysis: Combine income, population, and housing indicators at municipality or census-tract level
• Building-level Socioeconomic Profiling: Join Census 2021 household characteristics to cadastral building data via hypercadaster_ES
• Energy & Social Studies: Cross-reference aggregated electricity consumption with income and demographic indicators

📄 License

This project is licensed under the EUPL v1.2. See the license for details.

Authors

• Jose Manuel Broto - jmbroto@cimne.upc.edu
• Gerard Mor - gmor@cimne.upc.edu

Copyright (c) 2025 Jose Manuel Broto, Gerard Mor