Field notes: the cohort that outlived three satellite missions

I pulled up the exposure assignment code for a long-running cohort last month and realized the comment block at the top still referenced MOD04_L2 collection 5. The cohort had been enrolled in 2005. The code hadn't been touched since 2011. Nobody had told it that Terra's orbit has drifted enough to shift its overpass time by nearly forty minutes since launch, or that the algorithm version it's calling now bears only a loose resemblance to what ran when the first wave of baseline visits happened.
This is the part of cohort work nobody warns you about in grad school. You estimate exposure over and over across a cohort's lifetime, re-running it each time a satellite mission ages out and gets replaced by something with a different equator-crossing time, a different swath, a different retrieval algorithm entirely. A cohort that runs fifteen or twenty years doesn't sit still under one consistent instrument. It sits under a relay team.
Our cohort started under Terra and Aqua MODIS, moved through the Collection 6.1 reprocessing that changed aerosol optical depth values retroactively for every day already in the database, picked up VIIRS on Suomi NPP partway through follow-up, and is now, in theory, supposed to account for JPSS-2. Each handoff comes with its own quiet set of decisions that somebody has to make and then defend at a grant renewal five years later. Do you harmonize retroactively, re-running the whole exposure history under the newest algorithm so early and late person-time are comparable? Or do you stitch eras together and hope the discontinuity falls somewhere that doesn't correlate with your outcome?
I've done it both ways, on different studies, and I don't think either is wrong. I do think whichever you pick needs to be written down somewhere more permanent than a code comment, because the person who inherits your cohort in 2032 will not know which choice you made unless you tell them. Orbital drift alone will get you. A sensor launched to cross the equator at a fixed local time in the morning doesn't hold that schedule forever, and if your health outcome has any diurnal structure to the pollutant you're estimating, that drift is not a footnote.
The thing that actually keeps me up before a site visit isn't the algorithm version, though. It's the residential histories. A participant who moved three times over fifteen years of follow-up has three different pixels assigned to them, from three different mission eras, each with its own retrieval quality and its own gaps on cloudy days. Multiply that by a cohort of a few thousand and you have an exposure matrix that is less a clean grid and more an archaeological section, with sediment layers from each instrument generation stacked on top of each other.
What I've started doing, on the studies where I have any say in the pipeline, is treating the satellite record as raw material rather than the deliverable itself, and building one consistent product on top of it for the exact geography the cohort lives in, rather than inheriting whatever boundary the original grant proposal happened to specify. Air Quality Exposure is built around that idea, a daily retrieval downscaled to neighborhood resolution over the footprint you draw, so the mission history underneath it is somebody else's maintenance problem and not yours.
If your cohort has outlived a satellite mission or two of its own, it might be worth a conversation about how the exposure side of it gets handled from here.