The pivot you can count is not the pivot founders report
Nov 01, 2026
10 min read
15 views
Glasp Research. Keisuke Watanabe, Glasp Inc.
Take 3,262 American businesses founded in 2004 and followed by survey for eight years. In each follow-up they were asked whether their main line of business was still the same as before. By their own answers, 678 of them changed. The industry code that researchers use to detect exactly that kind of change flags 270 of those 678. It does not flag the other 408, which the cross-tabulation classifies as self-report-only: 60 percent of the self-reported changes of business.
The full cross-tabulation, firm counts:
Said main business changed (678): code changed 270, code unchanged 408.
Said it did not (2,584): code changed 6, code unchanged 2,578.
All firms (3,262): code changed 276, code unchanged 2,986.
Every number in this article that is not quoted from a published source, those counts included, comes from my own scripts run on the public Kauffman Firm Survey file. Automated reruns reproduced them, but no human has checked a source row or validated how the variables were built.
Start from the firms whose code changed and the measure looks excellent: 270 of 276, 97.8 percent, agree with the firm's own answer, and false alarms are almost absent. Start from the firms that said their business changed and it looks poor: 270 of 678, 39.8 percent. Both numbers are true of the same counts, and which one you quote decides whether the proxy looks good or bad. Thomas and Bandera (2023), the paper this proxy comes from, report neither.
What the data is, and what the measure is supposed to do
A pivot in startup language is a change of direction: the company keeps something (the team, the technology, the market insight) and changes something else (who the customer is, what the product does, how it makes money). Eric Ries, who popularised the word, defined it as a structured course correction designed to test a new fundamental hypothesis about the business. I have not read Ries's book; that definition reaches me second-hand through Pillai, Goldfarb and Kirsch (2020), and I name the intermediary rather than pretend otherwise.
A pivot defined that way is a story, and stories do not come in survey columns, so researchers look for something countable that moves when one happens. The usual candidate is the industry code. In the United States every business is classified under the North American Industry Classification System (NAICS), a six-digit code whose first two digits name the broad sector: construction, retail, information, professional services. A firm coded as a retailer in year one and a software company in year three changed something fairly fundamental. That is the sector-change proxy: a pivot is recorded whenever the two-digit sector code differs from what it was.
The dataset is the Kauffman Firm Survey (KFS), a panel study of 4,928 American businesses all founded in 2004, surveyed at baseline and then yearly through 2011. It is free to download without registration, it follows the firms that failed rather than only the survivors, and it records each firm's sector code and its status in every wave. That is why it keeps being used here.
The counts above use the KFS public file, a window of the first two years after founding, and the survey's own main-business question. The analysis sample is the 3,390 firms that answered the second follow-up, matching exactly the 3,390 "Complete" responses the Kauffman documentation reports for that wave; 3,262 of them answered that question in both of the first two follow-ups, and they are the firms counted.
Two more ways the measure moves without the business moving
One-year code changes mostly undo themselves. In the single-year window, 540 firms in the main sample show a sector change between founding and the first follow-up. By the second follow-up 387 of them are back on their original code, and 343 of those had told the survey their main business was the same all along. Such reversals may reflect coding practice, temporary business changes, or both; the public file cannot distinguish them. That is why the two-year window is the one used above.
Firms that stop answering look like firms that never changed. In the second follow-up, 775 firms in the wider sample did not respond: refusals, cut-off interviews, firms that could not be located. They still carry a sector code in the file, and for 756 of them it is identical to their founding code. Carried forward, not collected. Farhat and Robb's KFS methodology guide names last-observation-carried-forward as a method the survey uses, for legal status and business location explicitly, and says that in the longitudinal file it fills the whole fixed core of questions this way for firms that were located but gave no data. It does not say so for the sector code, so for the public file this is an inference, a strong one.
So whoever uses the file has to decide what to do with those firms, and that decision is not visible in the output. Non-respondents land in the "no change" group almost without exception, which means that group contains firms whose change, if any, was never collected. One small demonstration of what the choice carries: taking the outcome "business no longer operating", which in this survey includes being sold or merged, the exit risk ratio for firms with a changed code against firms without one is 0.96 when the non-respondents are kept with their carried-forward codes read as "no change", and 1.04 when the sample is restricted to the firms that answered the wave. Both are imprecise, both intervals contain 1, and the direction flips between them. Neither is a finding about what pivoting does to a company, and this article does not try to answer that question. The point is only that a handling choice, one Thomas and Bandera do not state, decides which way such a number points.
Where the proxy comes from, and what that paper was measuring
The proxy has a published home. Thomas and Bandera (2023), in the HICSS-56 proceedings, study pivot timing on the same data: "the public version of the Kauffman Firm Survey (KFS), a longitudinal survey of 4928 ventures formed in 2004 across eight years to 2011", and "as a proxy for a pivot, we use any change in a company's North American Industry Classification System (NAICS) sector code (the most two significant digits of the six-digit NAICS code)" (p. 4767). Same data, same proxy, same period. I read it in full from the Hawaii repository; it is six pages, open access under CC BY-NC-ND.
The relationship is a difference in measurement target, not a flaw to score. Their unit of analysis is the firm-year and their model runs on all 39,424 of them, which is 4,928 firms times 8 years: the complete rectangular panel, nothing dropped. In spreadsheet terms, every firm contributes eight rows, one for each year, including the years after it stopped answering the survey. The analysis here instead gives each firm one row and one outcome, which is why the two sets of numbers are not comparable even though they come from the same file. Their outcome is a binary "venture survival" per firm-year, which the paper does not operationally define; the words "exit", "merger", "weight" and "cluster" do not appear in it at all. Their Table 2, captioned "Survival and pivoting activity per year", has a first column headed "Missing" that reaches 1,962 by 2011, introduced by the sentence "each year several hundred additional companies ceased to operate" (p. 4767). The non-missing firm-years sum to 32,226 while the model keeps 39,424, so those missing firm-years are in the regression, which means the model implicitly treats a missing firm-year as nonsurvival. The rule is not stated.
A rectangular panel is appropriate to the question they ask, which is when in a venture's life pivots happen. It is not built to answer the question here, whether the proxy measures what its name says, and for that question the difference between a firm that closed and a firm that stopped picking up the phone is the thing being measured. The KFS documentation keeps them apart, with separate status codes for "Dropout (unit non-response) during follow-up t", "Permanently stopped operations during follow-up t" and "Merged, or sold during follow-up t", and it states that unit non-response is handled through weighting adjustments.
On validity, Thomas and Bandera give one sentence: a firm could pivot without a code change, "although an NAICS change would remain being a strong indicator of radical pivot" (p. 4769). Starting from the firms whose code changed, that is right. Starting from the firms that said their business changed, the sentence addresses positive predictive agreement but not sensitivity. The paper offers no concordance evidence either way, which is normal for a limitations paragraph and is exactly the gap a concordance table fills.
What this is not
It is a measurement observation, not a causal finding, and not a validity statistic in the strict sense. Calling the 408 firms "missed pivots" assumes the self-report is the right standard, and that has not been shown. The Kauffman documentation describes the first follow-up's version of the question as a comparison against a commercial database record rather than against the previous year, so some of its "no" answers may not be post-founding changes at all; the questionnaire wording was never checked, and the two measures may not cover identical windows. The two risk ratios quoted earlier are not a result either: the outcome was looked at after the direction was already known, and they appear only to show that the handling choice moves the sign.
The literature behind the framing is thinner than a reference list suggests. Only 4 of 19 papers or proceedings were read beyond metadata or summaries, and some of those four were read only in part; 11 have a free version, and 7 free ones were never opened. For this article I opened two, quoted above; anything from elsewhere is labelled with its intermediary. On the data side, four automated reviewers re-downloaded the file and reproduced the output byte-for-byte, which establishes that the code does what it says and nothing about whether the variable definitions are right. No human has checked a row.
What would change the picture
The questionnaire. The decisive unknown is whether answering "my main business changed" triggers the interviewer or the coder to assign a new sector code. If it does, the 97.8 percent agreement is mechanical rather than independent confirmation, and this becomes an audit of one file's processing rather than a statement about the measure. It was never established. The Kauffman documentation does show the mechanism exists in the instrument for one analogous item: for legal status, "if the respondent confirms the legal status to be the same as in the previous year, the legal status of the business is copied from the previous year", and a new value is collected only when the respondent says it changed. Whether the sector code works that way is not documented in anything I could open, and a questionnaire may not settle back-end coding practice at all.
Digit depth. The public file carries only the first two digits, so a change within a sector is invisible by construction. Thomas and Bandera flag this and point at the confidential file, which has all six. The same concordance at four and six digits would say whether the missing 60 percent are finer-grained pivots or something else.
Full concordance reporting. Sensitivity, specificity and predictive values with intervals, wave by wave, both measures' time windows stated, and a declared reference standard or an explicit admission that neither measure is one.
Until then the practical reading is the modest one. A study that counts pivots by industry code is counting a relatively rare recorded sector-code change: on this data, 8.4 percent of firms in their first two years, against 12.9 percent who said their main business had changed in the latest year alone. It is a minority of what founders themselves call a change of business, and how many such changes a file appears to hold depends on how non-response is handled, which Thomas and Bandera do not state.
Data: Kauffman Firm Survey, public-use microdata, baseline through follow-up 7, free from the Kauffman Foundation. This article is a plain-language write-up of a longer Japanese working draft and its assessment. It reports a measurement observation, has not been peer reviewed, and makes no causal claim. The literature summarised here was gathered with AI assistance, and the reading depth of each source is stated where it matters.
Thomas, E. and Bandera, C. (2023). On the timing of pivots: jumping the gun or late out of the gate? Proceedings of the 56th Hawaii International Conference on System Sciences, 4765-4770. Read in full, CC BY-NC-ND 4.0, from
https://hdl.handle.net/10125/103213
https://doi.org/10.24251/hicss.2023.580
Farhat, J. and Robb, A. Analyzing the 2004-2011 KFS multiply imputed data. Ewing Marion Kauffman Foundation.
https://ssrn.com/abstract=2367300
Pillai, S. D., Goldfarb, B. and Kirsch, D. A. (2020). The origins of firm strategy: learning by economic experimentation and strategic pivots in the early automobile industry. Strategic Management Journal, 41(3), 369-399.