Why restaurant failure rates disagree

Every figure below was read from the page that published it, not from a search result. Each answers some version of “how often do restaurants fail”.

The first two rows answer the same question about the same industry and differ by a factor of nearly twenty. That is not a rounding disagreement. It is three separate things being called one word.

What counts as failure. A restaurant that closes because its owner retired, sold, rebranded or moved two doors down is a closure in a licence record and not a failure in any useful sense. A study following a cohort forward counts the exit. An analyst verifying closures from review sites often counts the sign coming down. Neither is wrong; they are not the same quantity.

Which restaurants are in the denominator. A rate over “restaurants operating this year” is dominated by established businesses and will always look low. A rate over “restaurants that opened in year X” follows the riskiest period of a business’s life and will always look high. Both get published as “the failure rate”.

When the closure becomes visible. This is the one almost nobody states, and it is the reason we publish no failure rate of our own. A closure does not appear in a licence record when it happens. It appears when the licence lapses past its renewal window, which takes roughly three years. Any count of recent closures is therefore an undercount, and it is an undercount that shrinks as the data ages. Measure the same year twice, three years apart, and you get two different answers with no error in either.

We hold the Chicago licence record: 21,585 restaurant attempts, each one an individual business with a start date and, where the record shows it, an end. What that supports is survival, measured only over attempts old enough to have had the full horizon in which to fail.

• 85.7% are still licensed at two years, over 20,220 attempts.
• 60.9% at four years, over 18,755.
• 47.0% at six years, over 17,402.

What we will not do is subtract the first of those from 100 and call the remainder a two-year failure rate. That figure would be 14.3%, it would look like an answer, and it would be the exact move this page exists to take apart. The section below says why the complement of a survival rate is not a closure rate. The paragraph after this one says why the two-year row in particular cannot carry it.

The clearest illustration of visibility lag we have is our own number, so here it is rather than someone else’s.

The Bureau of Labor Statistics tracks establishment survival by opening year from payroll tax records. For Accommodation and Food Services, averaged over the same opening cohorts ours covers, it puts two-year survival at 73.6%. Our licence record says 85.7%. That is a 12.1-point disagreement about the same industry over the same years, and it is not a disagreement about Chicago. It is a licence lapsing slowly. At four years, once the lag has worked through, the two series land 1.0 points apart: 59.9% against our 60.9%.

So a first-year failure rate drawn from a licence record is not a small correction away from right. It is measuring how fast a city processes paperwork. 2 of the 3 published figures at the top of this page are first-year numbers, and no source among them says how long its underlying record takes to notice a closure. That is the whole problem, and we are not exempt from it.

The full curve, cohort and method are published separately.

The complement of a survival rate is not a closure rate. A 60.9% four-year survival does not mean 39.1% closed. Part of the remainder is attempts whose outcome is not yet observable. Closures take roughly three years to surface in the licence record, so any recent-period closure figure would be an undercount, which is why none is published.

We could publish one. It would rank, it would be quoted, and it would be wrong in the same direction as everyone else’s: low, because the closures have not surfaced yet. The honest version of this page is the one that explains why the number you came for does not exist in the form you wanted it.