I pre-registered a test, ran it, and it killed my own favorite hypothesis. Then I ran it a second time to be sure it wasn't a fluke. Honestly the best thing that happened to this paper.
Here's the setup. Accounting research has spent decades ranking firms on "conservatism" .. how fast a company books bad news, and how much of that news sticks around. Two different properties: timeliness and durability. Everybody measures them. Nobody separates them, and it turns out you can't. That's the theorem. Two firms can behave completely differently .. one fast and forgetful, one slow and permanent .. and file literally identical numbers. Same series, same everything.
This is the part I want to be clear about, because it's the part people get wrong when I explain it at a bar: it is not a power problem. More data doesn't fix it. The identified set is a continuum, and every instrument the field currently uses is reading the product of two things and reporting it as one thing.
What survives is less than I wanted, but it's real, and §9 is where the registered test took my preferred reading of the null out behind the shed. Paper's here. Fair warning, it's 60 pages and it earns them.
Here's the setup. Accounting research has spent decades ranking firms on "conservatism" .. how fast a company books bad news, and how much of that news sticks around. Two different properties: timeliness and durability. Everybody measures them. Nobody separates them, and it turns out you can't. That's the theorem. Two firms can behave completely differently .. one fast and forgetful, one slow and permanent .. and file literally identical numbers. Same series, same everything.
This is the part I want to be clear about, because it's the part people get wrong when I explain it at a bar: it is not a power problem. More data doesn't fix it. The identified set is a continuum, and every instrument the field currently uses is reading the product of two things and reporting it as one thing.
What survives is less than I wanted, but it's real, and §9 is where the registered test took my preferred reading of the null out behind the shed. Paper's here. Fair warning, it's 60 pages and it earns them.