Data Stories · Corporate Disclosure

The Survivors' Curve

We reported that the corporate risk section tripled in fifteen years. Before believing our own chart, we re-ran it on the 1,619 companies that filed in every year — to tell firms getting wordier apart from a changing cast of filers.

Who this data represents The same corpus as the original piece — US SEC registrants filing a substantive Item 1A, FY2006–2020. The balanced panel is the 1,619 firms present in all fifteen years: by construction, larger and longer-lived than the average filer. It answers a within-firm question, not a market-wide one.

An earlier article in this series, The Risk Section That Ate the 10-K, made a clean-looking claim: the median corporate risk section grew from 3,788 words in 2006 to 11,008 by fiscal 2020 — a 2.9× expansion that never once reversed. That number was computed the way most such numbers are: across every filing in each year, whoever happened to file.

But filers enter and leave. A firm goes public, another gets acquired, a third delists after a bad decade. If the companies entering the pool write longer risk sections than the ones leaving it, the aggregate can "triple" without any individual company changing a word. So which is it — the same firms getting wordier, or a changing cast? The honest way to find out is to fix the cast: keep only the 1,619 companies that filed in all fifteen years, and watch them alone.

1,619

continuous filers — present every year 2006–2020

72–84%

of the tripling is within-firm — on the ratio scale (~58% in raw words)

2.4× vs 2.9×

survivors' median growth vs the full pool's

~¾

of the readability decline is within-firm too (68–79%)

Most of the tripling is real

Hold the cast fixed and the risk section still balloons. The 1,619 survivors went from a median of 3,637 words in 2006 to 8,759 in 2020 — a 2.4× expansion of the same companies' own disclosures. The full-pool figure is larger, 2.9×, and the gap between the two lines is the part of the story that belongs to the changing cast, not to any individual firm.

Figure 1 · The survivors' curve

Median Item 1A length: all filers vs. the 1,619 continuous filers, by fiscal year

The audit largely holds. The lower (survivor) line is what the same firms did to their own risk sections; the shaded wedge above it is composition — the pool tilting toward longer-winded newcomers. Most of the rise sits below the wedge.

How much of the increase is within-firm depends on how you measure "length." Using the published metric — the median section — 71% of the 2006→2020 growth is the survivors' own. On the multiplicative scale a word like "tripled" implies, it runs higher: between 72% comparing average lengths and 84% averaging each firm's own log-growth, with a firm-fixed-effects regression near 76% — and that ratio never falls below 70% across two dozen robustness variations. On ratios, the finding is solid. Two caveats keep it from being airtight, and they belong in the open: in raw average word count only 58% is within-firm, and the median figure is tender at the edges — start the clock in 2008 rather than 2006, before Item 1A had finished phasing in, and it slips toward 64%. The tripling is mostly the same companies — but "mostly" is three-quarters on the ratio scale, not the near-certainty a single tidy number would imply.

Why do ratios and raw counts disagree? The firms that join the pool over time file unusually long risk sections. That lifts the average word count a lot — so composition looks large in levels — while barely moving the typical firm's multiple, so it stays small in ratios.

Figure 2 · What the same firms did, vs. the changing cast

Share of the 2006→2020 change that is within-firm vs. compositional

Within-firm growth is the larger share on every defensible measure — the published median, the ratio of average lengths, and a firm-fixed-effects estimate of readability all land around 70%. Composition is real but secondary. (In raw mean words, not shown, the within share falls to 58% — see text.)

But the reason isn't the one we'd have guessed

The tidy version of the composition story is a crisis story: the 2008–2010 delisting wave flushed out failing firms, leaving a survivor pool that looks artificially stable. It is a good guess. It is also wrong here. The pooled and survivor lines track each other almost exactly through 2011 — the survivors' median even runs ahead of the pool in 2008–2010 — and only begin to separate from about 2012 on, widening most in 2014, 2018, and 2020. The compositional gap is a steady, modern divergence, not a financial-crisis artifact.

Figure 3 · The cast turns over every year

Filers entering and leaving the pool, by fiscal year

2008 brought the largest single intake of new filers (nearly 1,900) — yet that is exactly when the two length curves are closest. If composition drove the trend, the gap would have opened here. It didn't, which is why the delisting-wave explanation fails.

Even the prose got denser on its own

The original piece also reported that risk sections grew harder to read. That, too, could have been composition — verbose newcomers dragging the average fog index up. It mostly wasn't: depending on the estimator, between about two-thirds and four-fifths of the 2006→2020 rise in the fog proxy is within-firm (a firm-fixed-effects estimate puts it near 68%; the balanced panel nearer 75–79%). The same companies did not just write more; they wrote it more densely. The claim that readability declined is not entirely a within-firm effect — something like a quarter of it is the changing cast — but it is predominantly the same lawyers, writing worse, year after year.

Why publish an audit of your own chart

Because the interesting outcome was not guaranteed. We pre-registered three questions before running any of this: whether the length finding would survive a balanced panel, whether composition would prove to be a small crisis-era effect, and whether the readability decline was within-firm. The first passed, and had it failed — had the tripling turned out to be mostly a changing cast — that would have been the more surprising and more important article. It didn't. But the second failed on its mechanism and the third came back "mostly, not entirely," and reporting those honestly is what earns the first one its credibility. A number you have tried to break and couldn't is worth more than a number you only admired.