How many sites do you need to sample for a regional installed-base estimate?

A vast wheat field stretching to the horizon, representative of the farmland an installed-base survey would cover.

Every OEM strategy analyst who's tried to build a market-share denominator from scratch hits the same wall: a VP asks "how many sites did we check," and there's no clean answer, because sample size for an installed-base estimate isn't a single number you look up. It's a function of how much variance sits inside the fleet you're trying to count, and how much error you can live with in the final denominator.

That's a different question than the one most people start with. Most people start with "is 50 sites enough" or "do we need 200." Wrong place to start.

Start with the strata, not the count

A regional fleet is really several populations stacked on top of each other. Dealer density varies by sub-region. Equipment mix varies by end use, construction versus ag versus aggregates. Site size varies wildly, a single large yard can hold more units than twenty small ones combined.

If you sample 150 sites at random across a region and most of them happen to be small contractor lots, your estimate will undercount the big distribution yards that hold a disproportionate share of the fleet. The fix is stratifying first: split the region into bins by density and site-size class, then set a sample target per bin. A region with a long tail of small sites and a handful of large ones needs heavier sampling in the tail, because that's where the variance lives. Uniform coverage everywhere wastes budget on sites that look alike and skips the ones that don't.

This is the part dealer-survey-based estimates usually skip, because getting dealer self-reports stratified by true site size is close to impossible. You're relying on who answers the phone, not who holds the most units.

What moves the error bar

Once you've got strata, the sample size question becomes tractable: for each bin, how much does machine count vary site to site, and how tight does your denominator need to be. If you're using the number internally to rank yourself against two competitors, a looser estimate with wider error bars is fine. If it's going into a board deck as "we hold 22% share in this region," you need the error bar tight enough that a competitor's internal number, if they built one the same way, would land somewhere close to yours.

A few things inflate the required sample size beyond what people expect:

  • High site-to-site variance. Regions with a mix of mega-dealers and tiny lots need more sites per stratum than a region where every yard looks roughly the same.
  • Occlusion. Any site-level count, whether it's from a drive-by, a dealer call, or an aerial pass, misses some fraction of units parked under cover, stacked behind other equipment, or just out of frame. If you don't correct for that at the site level before you scale, your sample size calculation is built on an undercount, and no amount of additional sampling fixes that.
  • Coverage gaps. If entire sub-regions have zero sampled sites, you're guessing instead of interpolating. That's a sampling design flaw, worth fixing before you add a single site elsewhere.

None of this is unique to overhead-imagery methods. It's the same logic a market research team would apply to any multi-stage sampling design, applied to equipment yards instead of consumer panels.

Why this is easier to get right with a scaling approach

A dealer survey makes stratified sampling hard because you don't control who responds. A site-level count scaled up with coverage and occlusion correction, the kind of approach behind Installed Fleet Estimation, lets you set the stratification up front, because you're choosing which sites get imaged and counted rather than waiting on a response rate. You can oversample the sparse, high-variance tail of a region without needing anyone's cooperation to do it.

The honest version of the sample size question is this: how much variance sits in your region, how tight does the denominator need to be for the decision riding on it, and how are you correcting for what a single pass over any site inevitably misses. Answer those three and the number of sites falls out of the design. It isn't something you pick first.

If you're scoping a denominator for next quarter's share numbers, request access to see how the scaling method handles stratification for your region.

Request early access

← Back to the blog