Handling PO Boxes and Non-Addressable Locations

This page solves one exact task: identifying addresses that are mailing destinations rather than places — post office boxes, rural routes, general delivery, military mail — and routing each to the treatment it deserves before it reaches a geocoder that will happily return a coordinate for it.

The failure this prevents is quiet and specific. A post office box geocodes to the post office, which is a real building at a real coordinate, and nothing in the result says the customer does not live there. Aggregate a few thousand of those and one census block group acquires a phantom population cluster that shows up in every demographic join and every trade-area estimate that follows.

Prerequisites #

• Python packages: pandas and regex (or the standard re) are sufficient; usaddress if you are detecting during parsing, which is the recommended point.
• Parsed address components from parsing and standardizing US store addresses — detection on raw strings works but misses the cases where the box designator sits in a unit field.
• A decision per use case. This is the part that cannot be automated: what a box means for a store record, a customer record and a delivery record are three different answers.

Configuration and execution parameters #

The three policy rows are the whole point of the page. A single global rule — drop them, or geocode them anyway — is wrong for at least one of the three cases, and the cost of being wrong differs: dropping customer boxes understates rural markets, while geocoding store boxes puts a storefront in the wrong town.

Annotated implementation #

Detection is a small amount of pattern matching applied to normalized components, plus a policy table that turns the detection into a decision. Keeping the two separate means the patterns can be improved without renegotiating the policy.

from __future__ import annotations

import re
from enum import Enum

import pandas as pd


class AddressKind(str, Enum):
    STREET = "street"                # a physical, addressable location
    PO_BOX = "po_box"                # mail held at a post office
    RURAL_ROUTE = "rural_route"      # a delivery route, not a point
    MILITARY = "military"            # APO/FPO/DPO — no domestic ground location
    GENERAL_DELIVERY = "general_delivery"


PATTERNS = [
    (AddressKind.PO_BOX, re.compile(r"\b(P\s*O\s*BOX|POB|POST\s+OFFICE\s+BOX|BOX)\s+\d+", re.I)),
    (AddressKind.RURAL_ROUTE, re.compile(r"\b(RR|R\s*R|RURAL\s+ROUTE|HC|HIGHWAY\s+CONTRACT)\s*\d+", re.I)),
    (AddressKind.MILITARY, re.compile(r"\b(APO|FPO|DPO)\b", re.I)),
    (AddressKind.GENERAL_DELIVERY, re.compile(r"\bGENERAL\s+DELIVERY\b", re.I)),
]

# What each use case does with each kind. The only entries that matter are the
# ones that are NOT "street" — everything else follows the normal path.
POLICY = {
    ("store", AddressKind.PO_BOX): "reject",
    ("store", AddressKind.RURAL_ROUTE): "reject",
    ("store", AddressKind.MILITARY): "reject",
    ("store", AddressKind.GENERAL_DELIVERY): "reject",
    ("customer", AddressKind.PO_BOX): "postal_resolution",
    ("customer", AddressKind.RURAL_ROUTE): "postal_resolution",
    ("customer", AddressKind.MILITARY): "exclude_from_geography",
    ("customer", AddressKind.GENERAL_DELIVERY): "postal_resolution",
    ("delivery", AddressKind.PO_BOX): "keep_flagged",
    ("delivery", AddressKind.RURAL_ROUTE): "keep_flagged",
}


def classify(components: dict) -> AddressKind:
    """Check the street line AND the unit field — boxes hide in both."""
    haystack = " ".join(str(components.get(k, "")) for k in
                        ("street", "unit", "number")).strip()
    for kind, pattern in PATTERNS:
        if pattern.search(haystack):
            return kind
    return AddressKind.STREET


def apply_policy(df: pd.DataFrame, use_case: str) -> pd.DataFrame:
    out = df.copy()
    out["address_kind"] = [classify(c or {}) for c in out["parsed"]]
    out["policy"] = [
        POLICY.get((use_case, k), "normal") for k in out["address_kind"]
    ]
    # Only STREET addresses proceed to a site-level geocode.
    out["geocode_site_level"] = out["policy"].isin({"normal"})
    return out

The detail that catches people is checking the unit field as well as the street line. A record whose street is a real street and whose unit is BOX 214 is a mailing address at that street’s post office, and a detector that only reads the street line will pass it through as a rooftop-quality record.

Failure modes and debugging #

A street genuinely named Box. Streets called Box Canyon Road and businesses at Boxwood Lane exist, and a loose pattern turns them into post office boxes. Anchoring the pattern on a following number and requiring the token to be at the start of the component removes most false positives; the remainder show up as a small, stable list that can be allow-listed by hand.

Dual addresses. Many rural businesses have both a physical address and a box, and some source systems concatenate them into one field. The record is not a box — it is a street address with a box attached — so the detector should extract the street portion rather than rejecting the record. This is the case where detecting during parsing rather than on the raw string pays off, because the parser has already separated the components.

Military addresses treated as domestic. An overseas military address carries a domestic-looking postal code, so it geocodes to a location in the United States that has nothing to do with where the person is. Excluding them from geography entirely, rather than degrading them to postal resolution, is the only correct treatment.

Silent policy drift. A new use case appears — a marketing extract, say — and inherits whichever policy the code happened to default to. Making the use case an explicit argument with no default, as above, forces the decision to be made rather than inherited.

Verification #

• Count by kind, per source, per import. A source whose box rate jumps overnight has changed a field mapping, not acquired rural customers.
• Confirm no rejected record reached the geocoder. The assertion is trivial and it is the one that matters: a rejected store address that was geocoded anyway will be indistinguishable from a real store downstream.
• Spot-check the false positives. Pull every record classified as a box whose street component also looks like a real street, and read them. The list should be short and stable; a growing list means the patterns need anchoring.
• Compare market totals with and without degraded records. In rural markets the difference is large, and knowing its size is what allows a planner to interpret a coverage figure honestly.

Frequently Asked Questions #

Related #

• Parsing and Standardizing US Store Addresses — where detection should happen.
• Geocoding & Address Normalization — precision tiers and the degraded-resolution path.
• Best Practices for Securing PII in Customer Location Datasets — handling customer addresses once they are classified.

← Back to Geocoding & Address Normalization