What the Census Data Actually Show
This chapter examines the Census study's measurement problem: what counts as custom software, how firm scope is observed, and why adoption and investment intensity must be read separately. That distinction supports a specific judgment about the evidence. The data show a strong relationship between multi-establishment firms and custom-software use, but they do not by themselves show that software caused firms to expand or markets to concentrate.
A narrow measure by design
The study combines two U.S. Census data sources. The Annual Capital Expenditures Survey records firm-level investment in structures, equipment, and software. The Revenue-Enhanced Longitudinal Business Database supplies information on the universe of employer establishments, including employment, payroll, industry, location, sales, and the firm identifiers that allow separate establishments to be grouped under one owner. Linking the sources makes it possible to compare what a firm buys with how widely it is organized.
That link is more consequential than a simple software-spending survey. A conventional technology statistic might tell us whether a business has a website, uses cloud services, or purchased an application. The Census study instead asks whether a firm capitalized a particular class of internal-use software and how that investment relates to the number of establishments under the same firm. The measure is designed to fit the paper's economic question: can a capability developed for the firm's own needs be used across several operating units?
The focal category is custom software, defined here as the sum of vendor-customized and own-account software. Vendor-customized software is developed externally but tailored to the firm's needs. Own-account software is created by the firm's employees for internal use. The survey instructs firms to exclude software they plan to sell in the market and to report capitalized investments—assets recorded on the balance sheet and depreciated or amortized. Prepackaged software is excluded from the focal measure.
That exclusion is not a claim that off-the-shelf software has no economic value. It is a measurement choice. Prepackaged software may be expensed rather than capitalized, which makes it harder to compare consistently with the investment categories in the survey. Custom software is therefore a narrower and more observable proxy for a non-rival internal capability. The result is a cleaner connection to the model, but also a narrower field of inference. Findings about this category should not be silently converted into findings about every digital tool a business uses.
The sample follows the same logic. Firms with more than 500 employees are automatically sampled into the Annual Capital Expenditures Survey; smaller firms are stratified by industry and payroll and randomly selected. The researchers apply the survey weights to make the results nationally representative, then remove observations with missing or inconsistent financial and investment data and winsorize extreme values. The final matched sample contains approximately 384,000 observations and accounts, on average, for about 71% of total software investment and 66% of total capital expenditure reported in the public survey totals.
Those figures indicate substantial coverage, not universal visibility. Some firms may record software differently, and the paper's accounting discussion concerns capitalized internal-use investments rather than all technology spending. A firm can be deeply dependent on a subscription service without reporting a custom-software investment in this measure. Another firm can capitalize an internal system whose value is difficult to see from outside the organization. The data are useful precisely because they make one important category measurable; they are not a complete census of digital capability.
The principal measure of firm scope is the number of establishments a firm operates. An establishment is a physical or operating unit associated with the firm, so the measure captures the difference between a one-site firm and a multi-site firm. Scope is not identical to employment, revenue, market share, product breadth, or geographic reach. A firm may have many employees at one site, or several small sites, and those structures can have different management costs even when their total employment is similar.
The authors recognize this limitation and test alternative measures, including employment, sales, and the number of industries in which a firm operates. The reported relationships remain in those checks, but the variables are not interchangeable. The number of establishments is the measure that most directly matches the paper's mechanism: software can be shared across operating units that would otherwise require separate local systems.
There is also an important aggregation choice hidden in the word “firm.” The database observes establishments as operating units but groups them through firm identifiers. That grouping allows the study to ask whether software is shared inside one organization, rather than treating every site as an independent business. It also means that the measured relationship is about the boundary of ownership and control. Two independent businesses in the same building may face similar local problems but cannot necessarily pool a software investment in the way two establishments belonging to one firm can.
The sample and definitions therefore do two jobs at once. They make the reuse question observable, and they exclude many forms of digital activity that do not fit the question cleanly. A positive result is stronger when it survives alternative measures of scope, but a robustness check does not make employment, sales, and establishments conceptually identical. Each measure captures a different way of being large. The establishment count remains the most direct bridge between the statistical evidence and the later claim about a firm's choice of organizational scope.
The measurement pipeline can be represented simply:
The diagram also shows why the study needs more than one outcome. Adoption asks whether a firm reports a positive investment in custom software. Intensity asks how large that investment is relative to the firm's total capital expenditure, conditional on adoption. Scope is the organizational structure that may make reuse valuable. Concentration is an industry-level or economy-level distributional outcome. A single percentage cannot stand in for all four.
The two-margin pattern
The first central result concerns the extensive margin, the decision whether to adopt at all. In the matched sample, firms are called adopters when they report positive investment in custom software and non-adopters when they report zero. Only about 3% of firms are adopters by this definition. The small overall share is important: an average across all firms is dominated by non-adopters and can conceal a large difference within the selected group that does invest.
Adoption becomes more common as scope increases. Relative to single-establishment firms, firms with two to four establishments are 5.5 percentage points more likely to invest in custom software. The relationship grows across the establishment-size bins; firms with more than 100 establishments are over 40 percentage points more likely to invest than single-establishment firms. The exact comparison is estimated with industry-year controls, so it is not merely a reflection of one high-tech industry being larger on average. It remains a conditional relationship between establishment scope and the probability of a positive custom-software investment.
The second result concerns the intensive margin, which is estimated only among adopters. Here the pattern reverses. The average adopter devotes 37.8% of total capital expenditure to custom software, but firms with more than 100 establishments devote a share about 21 percentage points lower than single-establishment firms. The direction is not an error in the data. It says that, once a firm is already using custom software, the relative composition of its capital spending changes with scope.
The two results can be true at the same time. Consider two firms that report positive custom-software investment. The larger one may spend more dollars because it has more activity to support, more data to process, or a more elaborate system to maintain. But its total capital budget may grow even faster because it also buys buildings, equipment, and other physical assets for many establishments. The software investment is then larger in level terms but smaller as a fraction of the whole. A share answers a different question from a level.
This distinction is central to the paper's interpretation. If software is reusable, a large firm does not need to purchase a fully separate copy of the underlying capability for every establishment. The firm can still spend more as it grows—for development, maintenance, integration, and complementary equipment—without software's share rising one-for-one with the number of operating units. The data do not yet prove that this is the cause; they show a pattern that a reuse-based model can be asked to explain.
The factor-share evidence adds another layer. Among adopters, the custom-software cost share declines with the logarithm of establishments. The labor cost share also declines, while the capital cost share rises. For non-adopters, the labor relationship runs in the opposite direction. In other words, scope is associated not just with how much a firm invests, but with how it combines the inputs used to produce output. The paper describes this as evidence consistent with a non-homothetic production structure: factor proportions are not fixed as the firm expands.
These cost shares require care. The data provide investment rather than a complete stock of custom software, so the authors use investment as a proxy under assumptions about depreciation. Rental rates for capital and software also enter the calculation. The paper reports robustness to alternative depreciation assumptions and to different treatments of developer wages and vendor payments, but the cost-share regression remains a conditional relationship. It does not by itself identify how much software displaced labor or caused capital deepening.
The word “share” deserves equal care. A declining software share can result from software becoming cheaper, from other capital spending rising faster, from production changing, or from several of those forces together. It does not mean that software has become irrelevant to the firm's performance. Likewise, a rising capital share among adopters does not tell us whether the additional capital is a substitute for software, a complement to it, or simply the physical capacity required to open more establishments. Those are mechanism questions for the model, not conclusions that can be read directly from the regression coefficient.
The cost-share pattern is still valuable because it rules out a very simple picture in which every firm uses the same production recipe at every size. If scope were merely a label for multiplying an identical establishment, the composition of spending would not need to change in this systematic way. The observed gradients indicate that the organizational unit and its input mix move together. They do not tell us exactly why, but they narrow the kind of explanation that can fit the facts.
A separate cross-industry pattern points in the same direction but carries an even clearer causal warning. Industries with greater custom-software investment per employee—Information, Finance, and Utilities are examples—also tend to have higher sales and establishment concentration. In the displayed regression, a one-log-point increase in industry custom-software investment is associated with an 8.3 percentage-point increase in the share of establishments owned by the top 1% of firms. The estimate includes controls for prepackaged software, non-software equipment, structures, employment, and year effects, and the relationship is reported to hold across 3-digit industries as well.
Controls make the comparison more informative, but they do not turn it into an experiment. The displayed regression has 70 industry observations. Industries that invest heavily in custom software may also have different demand structures, regulations, skill mixes, intangible assets, or competitive conditions. Those characteristics could affect both software adoption and concentration. The association is therefore evidence that the variables move together after the listed controls, not proof of a one-way causal effect.
The paper performs several checks on the firm-level pattern. The positive relationship between scope and adoption, and the negative relationship between scope and software intensity among adopters, remain when scope is measured through employment, sales, or number of industries. The intensive-margin result also survives alternative intensity measures, and the authors report similar patterns in a sample with repeated observations of large firms and in a public-firm sample subject to formal accounting rules. These checks make a simple measurement artifact less likely.
They do not remove selection. Firms that adopt custom software may already be more productive, more capable, or more likely to plan expansion. Larger firms may be better able to finance a long implementation process. A firm may add establishments because its management system is effective, then invest in software to support that structure. The data show the relationship and its robustness; the model supplies a way to examine whether non-rivality can generate it under explicit assumptions.
That is the empirical puzzle to carry forward. The extensive margin says scope and adoption rise together. The intensive margin says software's share of capital spending falls with scope once adoption has occurred. The next question is not whether one of these findings should be discarded. It is how a single reusable investment changes the cost of adding establishments, while labor, physical capital, management, and competition still have to be paid for each expansion.