Greece: 37% of the 2011 population has no LAU (EL_NA sentinel = 3.9M)

[MaartenHilferink]

Found while validating the LAU age × sex layer for EuroCrisp prep. More than a third of the Greek population in the 2011 layer never reaches a LAU polygon, and almost all of it sits under a literal not-available sentinel.

This is not #15 — nothing here is Iceland, Switzerland or the Canaries/Azores, and none of it is filtered out by IsContinentalEU. These GISCO_IDs are absent from LAU_RG_01M_2021_3035.gpkg altogether, so they fall out at AGEGROUP5_SEX_perLAU/Result/LAU_rel before any study-area selection happens.

Measurement

POP_AGE_SEX_2011_AGE_5y_v210525.csv against the GISCO_ID column of LAU_RG_01M_2021_3035.gpkg (98 188 rows, all with geometry):

	rows	persons
census total 2011	4 072 298	447 681 223
unmatched (LAU_rel is null)	39 754	4 045 464

Every one of the 149 unmatched GISCO_IDs is Greek. Restricting to Greece:

	distinct GISCO_ID	persons	share of EL
EL in census	5 759	10 816 286	100 %
matched to a polygon	5 610	6 770 822	62.6 %
unmatched	149	4 045 464	37.4 %

The census total for Greece, 10.82 M, is right for 2011 — so the data is there, it just cannot be placed.

The 3.9 M of it is one sentinel value

EL_NA          3 899 452 persons   15 437 rows

EL_NA is 5 characters where every real code is 11 (EL_53010201). It is a "not available" placeholder covering roughly 367 LAU-equivalents (15 437 rows ÷ 42 rows per unit), and it alone is 36 % of Greece. It should almost certainly be dropped or redistributed rather than silently discarded — right now it is neither, it just vanishes into a null LAU_rel.

The other 148 are structurally valid 11-character codes totalling ~146 012 persons, headed by:

GISCO_ID	persons
EL_53010201	61 938
EL_56010201	6 869
EL_35010601	4 301
EL_32010901	3 620
EL_14040201	3 399
EL_53011301	2 406
EL_53010701	2 171
EL_56010301	1 900
EL_32010801	1 719
EL_53011001	1 570

The EL_32* and EL_53* clustering suggests a Kallikratis-era renumbering rather than 148 unrelated gaps — the gpkg holds 6 137 EL LAUs against 5 759 in the census, so 527 polygons have no census row either. A crosswalk is probably the fix for these, and a decision (drop, or redistribute pro rata within the country) for EL_NA.

Full list of the 149 with populations available on request — it is a 150-line csv, happy to attach it here or drop it on the share.

Why it matters for EuroCrisp

EuroCrisp projects age × sex pyramids per NUTS3 in 5-year steps off exactly this layer. Left as is, Greece enters the IPF with 63 % of its people, so every Greek NUTS3 marginal is wrong and the downscaling inherits it. Worth settling before the 2021 → 2011 backcast exercise leans on it.

Measured with GeoDms 20.19.1.m on main, with the reader fix from the companion issue applied (without it the 2011 table does not read at all).

[MaartenHilferink]

Confirmed from the model side, on a full EU run (GeoDms 20.19.1.m, main @ 735cad7, PopYear = '2011', 100 m), with two additions.

The 527 polygons without a census row are real, and 529 of them show up as such in the model. Checks/Qr_0_and_Pr_gt0 is true for 529 of the 6 133 Greek LAUs in the study area, so they run on the NUTS3 fallback of #12 rather than on their own pyramid.

One of them loses its people entirely: EL_99010000. The gpkg has it as Ψευδοκοινοτητα Αγιο Ορος (Αυτοδιοίκητο), the Mount Athos pseudo-municipality, with POP_2021 = 1 811. It has no census rows at all, and the NUTS3 it sits in has no census population either, so Q_asr/Total is null and all 1 798 people ARDECO puts there are dropped from the output. Same failure path as Paris; see #20 and #19.

What the 4.05 M actually does to the result. It does not change the Greek total, because ARDECO drives the cell totals: Greece goes in with 10 798 465 and comes out with 10 796 485, the 1 980 difference being EL_99010000 and rounding. What it changes is the structure: the age x sex distribution of the 62.6 % that does match a polygon is applied to 100 % of the Greek population. If the unmatched share is not demographically representative - and 3.9 M under one sentinel plus a EL_32*/EL_53* cluster suggests it is not - every Greek pyramid in the output is biased, without anything in the run reporting it.

So the crosswalk matters even though the totals look right.