A government shutdown doesn't just furlough workers. It deletes data. And deleted data, it turns out, doesn't just disappear — it gets replaced with estimates, and those estimates carry assumptions that almost nobody reads.
That's the quiet story inside a recent Bureau of Labor Statistics Monthly Labor Review article on how the BLS handled missing Consumer Expenditure Survey (CE) data caused by the 2025 government funding lapse. The piece is methodologically careful and admirably transparent. It is also, for anyone who uses CE data downstream — which is nearly everyone who cites household spending statistics — a reminder that the numbers you're quoting have a production process, and that process has failure modes.
The article is not a scandal. It's something more instructive: a case study in what happens when a primary dataset has a gap, how statisticians patch it, and why the patch introduces its own uncertainty that rarely makes it into the headline figure.
The Gap Nobody Announced
The Consumer Expenditure Survey is the foundational dataset for how Americans spend money. It feeds the CPI market basket weights, informs poverty research, and shows up in virtually every serious analysis of household economics. When the BLS can't collect CE data — because, say, the government runs out of appropriated funds — that's not a minor administrative inconvenience. It's a hole in the denominator of a lot of downstream claims.
The BLS article describes the challenge directly: the 2025 appropriations lapse interrupted CE data collection, creating a period where survey responses were either not collected or collected at reduced rates. The BLS then had to decide what to do with that gap — impute values, reweight the remaining sample, or flag the affected period as unreliable.
This is a genuinely hard problem. Any solution involves assumptions. Imputation assumes the missing households look like the observed ones. Reweighting assumes the remaining sample is representative after the lapse. Neither assumption is obviously true, and neither is obviously false. The honest answer is: we don't know exactly how much the gap distorted the estimates, and the BLS, to its credit, says so.
What the BLS does not control is what happens after publication. The downstream users — journalists, think tanks, policy analysts — will cite the resulting figures without the asterisk. The CE data will flow into analyses of household spending on food, housing, healthcare, and energy. Some of those analyses will be used to make arguments about inflation, inequality, or the adequacy of social programs. None of those arguments will note that the underlying survey had a collection gap in 2025 and that the gap was addressed through statistical imputation.
That's not a criticism of the BLS. It's a description of how statistical information degrades as it travels from primary source to public claim.
What Imputation Actually Does (and Doesn't Do)
The word "imputation" sounds technical enough that most readers skip past it. It shouldn't. Imputation is the practice of filling in missing data with estimated values derived from the data you do have. It's standard, defensible, and widely used across government statistics. It's also a source of uncertainty that is almost never communicated when the resulting numbers are cited.
Here's the core issue: imputation works well when data is missing at random — when the households that didn't respond to the CE survey during the lapse period are statistically similar to the ones that did. If the lapse disproportionately affected certain geographic areas, income groups, or household types (which government shutdowns plausibly could, if field interviewers in certain regions were furloughed at different rates), then the imputed values will be systematically off in ways that are difficult to detect after the fact.
The BLS Monthly Labor Review piece is transparent about the methodological choices made. That transparency is genuinely valuable — it's the kind of documentation that allows researchers to assess the data's limitations. But transparency in a technical journal article is not the same as transparency in the headline figure that gets cited in a congressional hearing or a news story about consumer spending trends.
The gap between "what the methodology section says" and "what the number implies to a casual reader" is where most statistical misinformation lives. It's not fabrication. It's context that got left behind somewhere between the technical documentation and the tweet. ProPublica's investigative team has documented a version of this problem in other domains — the difficulty of getting transparent, comprehensive information out of institutions that technically publish it but bury it in formats nobody navigates. The CE methodology notes are a textbook example: they exist, they're accurate, and they're effectively invisible to most data consumers.
The June Jobs Report and the Measurement Stack
This matters right now because the June 2026 Employment Situation report — released July 2 — is being read against a backdrop of economic data that includes CE-derived figures. The jobs report itself is a separate survey with its own methodology, but it sits in an ecosystem of labor market statistics where the CE data plays a supporting role.
The June report showed the unemployment rate and payroll figures that generated the usual round of competing interpretations. What it didn't show — what no single release can show — is the full uncertainty stack underneath the headline numbers. Every major economic statistic is built on surveys with response rates, weighting schemes, seasonal adjustments, and revision cycles. The CE lapse is a visible example of a problem that exists, less visibly, in every dataset.
The BLS is, by global standards, an exceptionally rigorous statistical agency. Its documentation is thorough, its methodologies are peer-reviewed, and its revisions are transparent. The point here is not that BLS data is unreliable. The point is that "BLS data" is not a monolith — it's a collection of surveys with different sample sizes, different response rates, and different vulnerability to disruption. When someone cites "BLS figures" to support a claim about household spending trends, the appropriate question is: which survey, what sample, what time period, and were there any collection disruptions that required imputation?
Most of the time, nobody asks.
The Census Pipeline and the Coming Data Releases
The Census Bureau's July 10 tip sheet previews several upcoming data releases that will generate their own round of headline numbers. Among them: the 2024 Small Area Health Insurance Estimates (SAHIE), described as "the only source for single-year estimates of the number of people under age 65 with and without health insurance coverage in each of the nation's 3,143 counties." Scheduled for release in the coming weeks, this data will almost certainly produce stories about the uninsured rate — stories that will cite percentages without always noting that SAHIE is a model-based estimate, not a direct count.
The June 26 Census tip sheet also flagged the launch of USA Trade Online: Reimagined, a retooled platform carrying the same underlying trade data as its predecessor. That release is a useful reminder that even when the underlying data doesn't change, a platform transition creates a window where saved reports break, comparisons get disrupted, and users who don't notice the migration may be working from stale exports. The number looks the same. The pipeline shifted underneath it.
The July 10 tip sheet additionally flags the release of 2024 county-level internet adoption estimates from the Local Estimates of Internet Adoption (LEIA) program, scheduled for July 30. These are described as an update to 2022 experimental estimates. "Experimental" is doing a lot of work in that sentence. Experimental estimates are produced when the Census Bureau is still developing and validating the methodology — they're useful, but they carry more uncertainty than the bureau's standard products. When those numbers get cited as "Census data on internet access," the experimental qualifier tends to disappear.
This is the pattern. Primary statistical agencies produce careful, documented, appropriately hedged data. The hedges get stripped in transit. The number arrives at its destination wearing the authority of the source but none of its caveats.
The Denominator Problem Is Also a Documentation Problem
The BLS article on the CE lapse is, in a narrow sense, a story about one survey and one disruption. In a broader sense, it's a demonstration of something that applies to every major government dataset: the methodology is the story, and the methodology is almost never read.
Consider what it would take to responsibly cite CE-derived household spending figures for the period affected by the 2025 lapse. You'd need to know that the lapse occurred, that it affected data collection, that the BLS addressed the gap through imputation, what assumptions the imputation made, and whether those assumptions are likely to hold for the specific population or spending category you're analyzing. That's five layers of context that sit between "the BLS reported X" and a defensible claim about household spending.
Most citations skip all five layers. That's not laziness — it's a structural feature of how statistical information moves through public discourse. The documentation exists. It's in the Monthly Labor Review. It's in the technical notes appended to every BLS release. It's in the Census Bureau's methodology pages. Nobody reads it, not because they're incurious, but because the system isn't designed to make the caveats travel with the numbers.
The fix isn't complicated to describe, even if it's hard to implement: treat the methodology section as load-bearing, not decorative. When a number comes from a survey that had a collection gap, say so. When an estimate is model-based rather than directly measured, say so. When a figure is experimental, say so. These aren't obscure technical distinctions — they're the difference between a number that supports a claim and a number that merely resembles one.
What to Watch in the Coming Weeks
The SAHIE health insurance data, when it drops, will be the first real test of whether the "experimental vs. standard estimate" distinction survives contact with the news cycle. Watch for stories that cite county-level uninsured rates without noting the model-based methodology. Watch for comparisons to prior years that don't account for methodological changes between vintages.
The internet adoption estimates due July 30 are similarly worth tracking — particularly any claims about the "digital divide" that treat LEIA county-level figures as precise measurements rather than model-derived approximations with confidence intervals.
The WHO's weekly respiratory surveillance updates offer a useful contrast in how uncertainty can be communicated at scale: the agency publishes positivity rates with explicit geographic scope, distinguishes between elevated and low activity thresholds, and names the specific influenza subtypes driving regional trends. It's a model for how to attach context to a number rather than letting the number travel alone. Government economic statistics could learn something from that discipline.
And the CE data, now that the BLS has documented its imputation approach, will flow into analyses of household spending for months. Every time someone cites a figure about what American households spent on food or healthcare in 2025, there's a reasonable chance that figure was touched by the imputation process described in that Monthly Labor Review article. The article exists. The caveat is there. The question is whether it travels.
It usually doesn't. That's the denominator problem, and it's not going anywhere.
