How to scale a sample-region machine count to a national total

A wide open-pit quarry with heavy machinery working among terraced rock layers.

You've got a clean count. Maybe your team ran a pilot over one province, one state, or a 50-mile radius around a dealer cluster, and you came out with a real site-level machine count for that patch of ground. Now someone in the strategy meeting asks for the national number, and the instinct is to just scale it up. Multiply by the ratio of national acreage to sample acreage, or by population, or by dealer count, and call it a day.

A flat multiplier like that is where most installed-base estimates fall apart, usually at the exact step nobody double-checks.

Why a flat multiplier doesn't hold up

A straight area-ratio or population-ratio multiplier assumes your sample region looks like the rest of the country. It almost never does. Machine density tracks terrain, crop mix, construction activity, rental penetration, and dealer territory boundaries, none of which scale linearly with land area or headcount. A sample pulled from a dense agricultural basin will overstate a national total if you apply its density to range land and suburban fringe alike. A sample from a construction-heavy metro will do the opposite once you stretch it across rural counties with a fraction of the build activity.

There's a second problem underneath the first: your sample count itself is rarely a true count. High-res imagery misses machines under tree cover, inside structures, or parked close enough together to read as one blob instead of two. If you don't correct for that occlusion before you scale, you're propagating an undercount across every region you extrapolate into, and the error compounds with distance from your sample.

This is the gap between a number you can put in an internal slide and a number you can defend when a board member, an analyst, or a competitor's filing pushes back on it.

A scale-up method that survives scrutiny

A defensible regional scale-up method starts the correction work before the extrapolation step, so the errors don't ride along into every region you project into. In practice that looks like four steps:

  1. Get the site-level count right first. Run the imagery pass over your sample region and apply an occlusion correction for the things imagery systematically misses: tree cover, shadow, indoor storage, tight clustering. This is the number everything else inherits, so errors here don't average out, they multiply out.

  2. Stratify before you scale. Instead of one national multiplier, break the country (or the region you're sizing) into strata that actually predict machine density, such as crop type, construction permit volume, climate zone, or dealer territory. Pull a sample count in each stratum, or at minimum in the strata that carry the most fleet.

  3. Apply the ratio within strata, not across them. Scale each stratum's sample count to its own full extent, then sum. A flat multiplier skips this step entirely, and it's usually where the biggest error hides.

  4. Validate against a second, independent sample. Pull a count from a region you didn't use to build the scaling factors and check whether your stratified model predicts it within a range you're willing to put your name on. If it doesn't, the strata are wrong, not just the multiplier.

This is roughly the same logic behind a site-level count scaled into a regional installed base with coverage and occlusion correction built in, rather than one blanket ratio applied to a single sample.

Where this fits next to a dealer survey estimate

A dealer survey estimate has its own value: dealers know their own books, their trade-in pipeline, their local competitors. What it doesn't give you is an independent cross-check, since it's built from the same channel whose share you're trying to measure. An imagery-based, stratified count is slower to build but doesn't depend on who answered the phone that week. Most strategy teams end up wanting both, with the imagery-based number as the denominator they can defend if the dealer number gets challenged.

If you're putting together a market-share denominator for a region you haven't sampled yet, that's the use case this is built around.

Request early access

← Back to the blog