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.
Statistics Canada’s Road Network File (product
92-500-X) is a national street centreline file: 2.25 million segments,
1.17 million kilometres, one LINESTRING each. What makes it
an address source rather than a basemap is four fields —
AFL_VAL/ATL_VAL and
AFR_VAL/ATR_VAL, the civic number range on the
left and right side of each segment.
rnf_import() loads it into the same DuckDB file NAR
lives in, as its own tables, and that adds the offline
"rnf" tier.
library(cangeocode)
library(dplyr)
con <- nar_connection()
rnf_import() # ~340 MB, onceWhat this adds to the package
It is the only tier that reaches streets NAR does not carry at all.
Every other offline tier reads NAR: "nar" matches a row,
"nar_interpolate" fills a gap between two NAR rows on the
same street, "rqa" reads Quebec’s register. All of them
need the street to exist somewhere in NAR. In the standing evaluation,
3.7% of filings fail because it does not — and a street
that exists on the ground is in the road network file whether or not
anyone has ever geocoded a house on it.
Measured on the 379 addresses geocode() leaves unplaced
out of a 5,000-filing draw, the tier places 93 of them,
24.5% — the largest recovery any tier here has offered, offline
or online. Coverage on that draw goes 92.4% → 94.3%. It costs nothing
measurable in time, because it only ever sees what its predecessors
could not place.
Licence
Statistics Canada Open Licence — the same terms as NAR, and recorded in the database’s own metadata when you import:
Using it
Add "rnf" to method. It belongs after the
NAR tiers and before the online ones — it is offline, so it is cheap,
but it interpolates where the NAR tiers match:
hard <- c("185 Deerfield Rd, Newmarket, ON",
"1006 Theatre Rd, Gravenhurst, ON",
"1435 Celebration Dr, Pickering, ON",
"1545 Maley Dr, Sudbury, ON")
geocode(hard, method = c("nar", "nar_interpolate", "rnf")) |>
select(input, match_method, n_matches, uncertainty_m, lon, lat)
#> input match_method n_matches uncertainty_m lon lat
#> 1 185 Deerfield Rd, Newmarket, ON rnf_interpolated 1 95 -79.47162 44.05564
#> 2 1006 Theatre Rd, Gravenhurst, ON rnf_interpolated 1 95 -79.33221 44.93899
#> 3 1435 Celebration Dr, Pickering, ON rnf_interpolated 1 95 -79.07905 43.83198
#> 4 1545 Maley Dr, Sudbury, ON rnf_ambiguous 3 NA NA NAThree of those four are streets NAR has no address on anywhere. The fourth is the subject of the next section.
What to watch out for
It refuses when the answer is ambiguous, and that is the load-bearing part
1545 Maley Dr came back rnf_ambiguous with
n_matches = 3: three separate segments in Sudbury named
Maley claim a range containing 1545. The tier does not pick the
shortest, or the nearest to the municipality centroid, or the first. It
declines to answer.
That refusal is where most of the tier’s quality comes from. Measured against NAR on addresses both can place, the unambiguous half sits p50 26 m from NAR’s building point; the ambiguous half carries essentially the whole gross-error tail. Dropping seven rows to keep ninety-three honest is the trade the tier makes.
It is necessary and not sufficient. Every recovered
row that landed more than 2 km from where the filer’s own postal code
says it should be had n_matches == 1 — the refusal never
saw them. Two of those three are filers whose city and postal code name
different municipalities, so the check rather than the tier is wrong.
The third is a genuine error, and instructive:
10272 County Road 2, Cobourg, ON parses to street
28, type HWY, in Bailieboro, and RNF placed
10272 on that faithfully. A bad parse, laundered into a confident
coordinate. A tier that exists precisely because NAR does not have the
street cannot sanity-check its input against NAR, so this is the one
tier where a parser error survives all the way to an answer. Roughly 1
row in 93.
The ranges are partly guesses, and the file will not say which
This is the reason the tier went unbuilt for so long. Statistics
Canada’s own documentation says the ranges are observed for some
segments, imputed for others, absent for the rest — and there is no
provenance flag in the record layout. CLASS and
RANK describe the road, not the range. You cannot tell a
surveyed range from an interpolated guess by reading the file.
What replaced the missing flag was a measurement. NAR’s building points are an independent second reading of the same streets, so: how often does a civic number NAR holds fall inside the range RNF claims for the side the house is actually on? 89.7%. That is the honest error rate on the fields, and a floor on the tier’s own. It does not say which 10.3% are wrong — but it lets the tier be priced rather than guessed at.
Coverage is uneven, and thinnest where NAR is thinnest
13% of segments are unnamed and are not candidates for anything. Of the named ones, 71.7% carry a range on at least one side, and the spread is the widest number in the whole measurement:
| province | of named segments | of all segments |
|---|---|---|
| Nova Scotia, the best | 89.2% | 81.8% |
| Ontario | 81.4% | 75.2% |
| Quebec | 76.2% | 73.5% |
| British Columbia | 66.9% | 59.4% |
| Manitoba | 52.5% | 43.2% |
| Saskatchewan, the worst | 36.6% | 18.7% |
The tier will not serve rural Saskatchewan. That ordering is close enough to NAR’s own that the tier is thinnest exactly where the thing it is compensating for is thinnest — the two gaps overlap rather than cancelling.
uncertainty_m is a segment-length model, not a
measurement
max(95, 0.35 × len_m): 95 m on a short urban block,
growing with the segment for long rural ones. Interpolating a house
position from a range assumes the numbers are spaced evenly along the
segment, which they are not, so the error scales with how far apart the
endpoints are. Checked against NAR where both answer, that interval
covers the true distance 91.8% of the time.
For comparison, interpolation lands p50 24.3 m from NAR’s building point and p90 93.3 m — which is about the same accuracy as the difference between a building point and a blockface point, and rather better than the 95 m floor suggests. The floor is deliberately pessimistic.
Accuracy on the overlap does not describe the residual
The p50 26 m above is measured on addresses NAR also places, because those are the only ones there is anything to measure against. The rows the tier exists for are the ones NAR cannot place, and they are harder. Against the filer’s own postal code, on urban FSAs:
| n | p50 | within 500 m | |
|---|---|---|---|
| NAR’s answer, on rows the tier also placed | 3,749 | 43 m | 97% |
| the tier, on those same rows | 3,749 | 60 m | 97% |
| the tier, on rows only it placed | 46 | 149 m | 87% |
Reading down: the tier’s own error costs 43 → 60 m, and being an address NAR could not place costs a further 60 → 149 m. The residual is harder by more than twice what the tier’s own error is worth. Whenever a source is validated where it overlaps NAR, that number is an upper bound on how well it does where NAR is silent — which is the only place it is being used.
Take the shapefile
rnf_import() downloads the shapefile, and there are two
independent reasons, both of which cost something to rediscover. It is
the only format published for every release — the
GeoPackage resolves for 2025 alone and 404s on the archive. And 13
features in the GeoPackage are CircularStrings, which
DuckDB’s spatial extension refuses outright, failing the entire read
rather than the rows. The shapefile format cannot express a
CircularString, so those 13 arrive as ordinary linestrings and no
workaround is needed.
DBI::dbDisconnect(con)Where the measurements live
file.show(system.file("notes", "road-network-file.md", package = "cangeocode"))data-raw/probe_rnf.R reproduces every figure above.