One of this package’s seven data sources — one of the two sources
that can be imported beside NAR. vignette("data-sources")
is the overview and puts it in context.
Quebec’s ministry of natural resources publishes the Répertoire québécois des adresses (RQA), the province’s own address register. It is where NAR’s Quebec rows come from — a 2.5-million-address point comparison shows NAR is carrying RQA’s coordinates, not deriving its own — and it is about 308,000 addresses larger than what NAR passes on.
rqa_import() loads it into the same DuckDB file NAR
lives in, as its own tables, which adds the offline "rqa"
tier for Quebec.
library(cangeocode)
library(dplyr)
con <- nar_connection()What this adds to the package
Quebec addresses NAR does not carry, and the register’s own coordinate in place of a guess.
Both halves matter, and the second is the larger one. On a 4,000-filing Quebec sample:
| method | placed | placed on a register point |
|---|---|---|
c("nar", "nar_interpolate") |
88.5% | 82.7% |
c("nar", "rqa", "nar_interpolate") |
90.1% | 89.1% |
258 filings get an RQA point. Only 62 of them were unplaced before — the other 196 were already being interpolated between two NAR neighbours, and the tier replaces that guess with the register’s actual coordinate, a median 26 m away. The headline is the second column, not the first.
It costs nothing measurable: 10.0 s against 10.1 s on the same batch, because the tier only ever sees rows NAR left unplaced.
There is a second, smaller contribution on the normalization side.
RqaStreets is wired into normalize_address()
as a Quebec-only second gazetteer pass over rows the NAR pass could not
settle. On addresses drawn from the gap population — what RQA
holds and NAR does not — it answers 8.9% of rows, all correctly. See the
warning below before reading more into that than it says.
Licence
RQA is CC-BY 4.0, where NAR is the Statistics Canada
Open Licence. Both are attribution licences and they compose, which is
exactly what makes RQA usable as a default tier where OpenStreetMap’s
ODbL is not. The obligation is real, and rqa_attribution()
is the string that discharges it:
rqa_attribution()
#> [1] "Contains information licensed under CC-BY 4.0 from the Repertoire quebecois des adresses, Ministere des Ressources naturelles et des Forets, Gouvernement du Quebec."Using it
Put "rqa" before
"nar_interpolate". That ordering is the point of the tier —
a register coordinate beats an interpolated one:
gap <- c("5510 Saint-Jacques, Montreal, QC", "1650 Chabanel, Montreal, QC",
"431 Courtemanche, Montreal-Est, QC", "45 Gamelin, Gatineau, QC")
geocode(gap, method = c("nar", "nar_interpolate")) |>
select(input, match_method, uncertainty_m)
#> input match_method uncertainty_m
#> 1 5510 Saint-Jacques, Montreal, QC nar_interpolated 58.76047
#> 2 1650 Chabanel, Montreal, QC none NA
#> 3 431 Courtemanche, Montreal-Est, QC none NA
#> 4 45 Gamelin, Gatineau, QC nar_interpolated 14.25068
geocode(gap, method = c("nar", "rqa", "nar_interpolate")) |>
select(match_method, uncertainty_m, lon, lat)
#> match_method uncertainty_m lon lat
#> 1 rqa_geocoded NA -73.60463 45.46885
#> 2 rqa_building 0 -73.65830 45.52960
#> 3 rqa_building 0 -73.51165 45.63419
#> 4 rqa_geocoded NA -75.73602 45.44344match_method reports the register’s own
positional class rather than one label:
rqa_building, rqa_geocoded,
rqa_uncertain, rqa_lot,
rqa_other. That is a field RQA carries and NAR does not,
and it is more informative than NAR’s building/blockface pair.
uncertainty_m is filled in only for
rqa_building, where 0 means what it
means for NAR: this package added nothing to the register’s own error.
Nothing has measured what Géocodée or
Incertaine are worth on the ground, and an invented number
would be indistinguishable from the two that were measured.
What is in the database
tbl(con, "RqaAddresses") |>
count(POS_QUALITY, IN_NAR) |>
arrange(desc(n)) |>
collect()
#> # A tibble: 13 × 3
#> POS_QUALITY IN_NAR n
#> <chr> <lgl> <dbl>
#> 1 Géocodée TRUE 1890979
#> 2 Incertaine TRUE 1522411
#> 3 Bâtiment TRUE 1332465
#> 4 Géocodée FALSE 193591
#> 5 Incertaine FALSE 147425
#> 6 Bâtiment FALSE 96674
#> 7 Front lot TRUE 48293
#> 8 Centre lot TRUE 45985
#> 9 Centre lot FALSE 27161
#> 10 Front lot FALSE 10427
#> 11 Site FALSE 14
#> 12 Accès propriété TRUE 8
#> 13 Accès propriété FALSE 2POS_QUALITY is the register’s own positional class, and
IN_NAR is whether NAR carries the address too. Reading the
FALSE rows gives the gap by quality: it is not concentrated
in the worst class — a substantial share of what NAR is missing carries
a Bâtiment point.
RqaStreets is the companion table, one row per odonyme
and municipality, with N_NOT_IN_NAR counting how many of
that street’s addresses NAR has no row for. That is the street-level
version of the same question, and it is what the normalization pass
reads.
What to watch out for
It sits beside NAR and is not merged into it
Merging the 308,000 missing addresses into Addresses
would be simpler to query, and it is deliberately not done. The
comparison against NAR is the only instrument Quebec’s coverage is
measurable with. Merge the two and every future question — how
much is NAR missing in Quebec, is that share growing, does a new release
close the gap — becomes unanswerable, because there is no longer a NAR
to compare to. The tables are additive;
nar_schema_version() is deliberately not bumped
for them, since bumping it would force everyone to re-download a 5 GB
database to gain an optional Quebec feature.
IN_NAR over-reports by about 14%
IN_NAR is fold equality on (FSA, civic
number, folded street name), not containment. Containment has no
equijoin key and would turn a scan into a product. The consequence is
that the gap it reports is larger than the real one by roughly 14% —
measured independently, not estimated.
| rows | distinct address keys | |
|---|---|---|
| certified register | 5,315,435 | 3,400,913 |
| not in NAR | 475,294 | 356,089 |
The whole register is imported, not just the gap, for the same
reason: the gap is a property of the pair. A subset built
against one NAR release would be silently wrong against the next, so
IN_NAR is computed inside the release’s own file, where it
stays correct by construction.
Comparing raw street names finds a gap three times too large
NAR keeps the particule inside the street name
(de la Montagne); RQA holds it in a column of its own.
Compare the raw strings and you get 1.27 million Quebec addresses
apparently missing, where the real figure is 358,000. Every join in
R/rqa.R goes through a fold that accounts for this.
The tier joins on the match fold — the one that
spells ST out to SAINT and treats a hyphen as
a word boundary — rather than the plain fold, because the addresses the
tier exists for are exactly the ones the gazetteer could not resolve,
and the plain fold is what failed on them.
Do not read a parser gain off NAR’s residual
This is the standing warning here, and it is recorded because this repository got it wrong. It was projected that importing RQA would take Quebec normalization from 81.8% to 88.3%, on the reasoning that 41.3% of Quebec’s failures are addresses NAR does not carry. Measured after building it:
| before | after | |
|---|---|---|
| Quebec confirmed against NAR | 77.5% | 77.5% |
| Quebec confirmed against NAR or RQA | — | 83.0% |
rows the rqa gazetteer pass answered |
— | 4 |
The improvement was a confirmation-set effect, not a parser
gain. The 41.3% was measured over rows that fail to
join NAR — a question about the judge, not about the parse.
Importing RQA makes the judge better. It does not make
normalize_address() read more Quebec strings: it read four
more.
What the parser actually leaves behind in Quebec is not uncovered
addresses but unreadable ones —
20-110 boul. de Mortagne, Bouceherville,
4150 SteCatherine Ouest, 1052 N.P. LAPIERRE,
1603 - 3410, rue Peel. A second register cannot read a
misspelling. The eval harness now judges Quebec against both registers
on separate lines so the two effects cannot be confused again.
DBI::dbDisconnect(con)Where the measurements live
file.show(system.file("notes", "quebec-addresses.md", package = "cangeocode"))That note carries the point-to-point comparison showing NAR is republishing RQA’s coordinates, why Quebec’s 99.8% “building” coverage is not a quality statement, and the six-way split of what is left of Quebec’s normalization failures.