The Gauge Has Been Broken Since 2018 and the Model Still Uses It

Station 04185000 sits on a concrete pier under a county road bridge in northwest Ohio, and it has been reporting stage height every fifteen minutes since 1974. In April 2018 a piece of agricultural equipment struck the bridge. The gauge housing was replaced that summer. The datum — the reference elevation the readings are measured from — was not re-surveyed.

The consequence is that every reading since then is offset by about eleven centimetres. Not randomly: consistently, in one direction, in a way that makes the creek look shallower than it is.

I found this in 2023 while trying to reconcile two datasets that should have matched and did not. I want to describe what happened next, because the interesting part is not the error. Errors like this are ordinary and unavoidable in a network of eight thousand stations maintained on a budget that has been flat in nominal terms for a decade. The interesting part is how long it took the correction to reach the places the data had already gone.

I reported it in June 2023. The station was re-surveyed in October. The historical record was flagged and corrected in the public database in January 2024, which is a nineteen-month lag from my report and a sixty-nine-month lag from the strike, and honestly that is a reasonable turnaround given the queue.

The problem is what happened in those sixty-nine months.

The readings from that station fed a regional flood frequency analysis that was published in 2021. That analysis fed the update to two county floodplain maps. Those maps determine which parcels require flood insurance, what a developer must do to build, and what a bank will lend against. By the time the datum was corrected, the numbers derived from it were embedded in a regulatory instrument that has its own revision cycle, measured in years, and no mechanism at all for propagating an upstream data correction.

A correction to a measurement does not travel to the decisions the measurement produced. There is no channel. We built one direction of a pipe.

This is the thing I have become preoccupied with, and it is not a story about one gauge. It is structural. Environmental data flows outward from monitoring networks into models, from models into assessments, from assessments into rules — and every one of those steps is a copy, taken at a moment, cited by identifier and date. When the source is later amended, nothing walks back down the chain. The 2021 flood frequency analysis will cite station 04185000 forever. It does not know the station was wrong. Nothing will tell it.

I asked, at a conference, how many published analyses in the region relied on stations with known post-hoc datum corrections. Nobody knew, because answering it would require joining the correction log to the citation record, and those live in different institutions with no shared key.

Let me be precise about the size of what I am claiming, because eleven centimetres is not a catastrophe and I do not want to inflate it. In the flood frequency context it shifts the estimated hundred-year stage by a small amount — enough to move a handful of parcels across a line, not enough to make the map fictional. If this were the only such error, the right response would be a shrug.

The reason it is not a shrug is the base rate. In the same network, in the same decade, I can point to gauges affected by channel aggradation that outpaced the rating curve update, gauges where a nearby bridge replacement changed the hydraulics, and at least two other datum problems reported by other people. None of those were negligence. Field crews are stretched, the network is old, and the surprising thing is how well it works, not how often it drifts. But drift is normal, corrections are normal, and therefore the absence of a correction channel is a systematic problem rather than a story about one bad April.

There is a fix and it is not technically hard. Every published analysis that uses gauge data should record the station identifiers and the retrieval date in machine-readable form, in a registry. Every correction to a station's record should be published with an effective date range. A weekly job could then produce a list: these analyses used data that has since been amended in a way that overlaps their period. Not an alarm, not a retraction — a list, sent to the authors and the agency.

I have described this to three different agency staff and every one of them agreed it was sensible and none of them owned the problem, which is the actual obstacle. It sits between the monitoring network, which is responsible for the data, and the assessment programmes, which are responsible for the analyses, and neither is responsible for the join.

I drove out to the bridge last spring. The new housing is a slightly different grey from the old pier. There is a survey mark on the concrete from October 2023, and about four inches below it, faintly, an older one from 1974 that somebody was careful enough to leave alone.

Update. One of the two county floodplain maps has since been revised on its normal cycle and the revision incorporates the corrected datum. Nobody propagated anything — the cycle simply came round and the new analysis pulled current data. The other map's cycle is in 2029. I am recording this because a partial fix arriving by coincidence is the outcome my argument predicted, and it would be dishonest to leave the piece implying nothing moved.