Platformization — broadly understood as one strand in the evolution of economic coordination mechanisms — manifests primarily through two- and multi-sided markets (Rochet and Tirole, 2003) organized around digital firms (Kapelyushnikov and Zinchenko, 2025). By most accounts, the phenomenon has been developing since the late 1990s, entering an accelerated corporate phase in the mid-2010s. Regulatory responses, in Russia and internationally alike, have gathered momentum since the early 2020s: the Russian Federal Law “On Certain Aspects of Platform Economy Regulation,” the European Union’s Digital Markets Act and Digital Services Act, and analogous regulatory initiatives in BRICS countries and China in particular (Avdasheva and Korneeva, 2019; Remington et al., 2024). Over this period, digital multi-sided platforms (hereafter DMPs) have evolved from an unconventional, niche form of production organization based on internet technologies into an established, large-scale industry capable of influencing macroeconomic indicators and financial markets. Nevertheless, despite their apparent ubiquity, the question of the extent to which economic activity has migrated from “open” markets to platform-mediated exchange — and which indicators are associated with this migration — remains open, which motivates the present study.

Measuring platformization is far from trivial. There exists no generally accepted national or international methodology, nor indeed a consensus definition of “platformization” itself. The intermediary role of DMPs poses additional obstacles to market-share estimation. Moreover, the industries subject to platformization extend well beyond electronic commerce: de facto, most consumer-facing sectors are amenable to platform intermediation (Table 1), since the matching mechanism linking supply and demand sides is applicable to virtually any form of exchange. Accordingly, this study treats platformization as encompassing all B2C industries. Furthermore, the DMPs under consideration operate on a global scale across diverse national economic contexts, which permits treating platformization as a universal phenomenon with both common manifestations and common drivers. It is on this premise that the present paper attempts to construct a MIMIC model.

Despite sectoral diversity, DMPs share a common set of economic traits: the reduction of information asymmetries through algorithmic matching; the redistribution of capital expenditure from the supply side to the intermediary (and the associated transformation of fixed costs into variable costs); and the generation of positive network externalities that increase nonlinearly with the number of participants (Avdasheva and Geliskhanov, 2025). These patterns are reflected in macroeconomic indicators: internet penetration, the volume of trade in techno­logy-intensive goods and services, and the maturity of digital infrastructure, which motivates the use of a unified set of variables for cross-country estimation of platformization (see Table 4).

These measurement challenges stem from the absence of a generally accepted statistical source comparable to the World Bank’s World Development Indicators: publicly available datasets are typically incomplete, and platform companies’ financial disclosures may omit the requisite operational metrics. The question is therefore: how can an economy’s degree of platformization be assessed using indirect indicators? In what follows, we employ two definitions of this concept. The first is a full, monetary formulation, expressed as the ratio of transactions constituting final household consumption C_B2C in GDP that are intermediated by platforms. The second is a reduced, population-based formulation, expressed as the share of the population engaged in platform activity (whether on the supply or demand side). The latter is used for modeling, as it satisfies the constraints imposed by the chosen specification. The objective of this paper is to present a robust model for estimating the level of platformization using publicly available macroeconomic indicators and to evaluate its convergence properties.

Existing studies of platformization, reviewed below, share two common limitations. First, a tendency to focus on individual platform markets rather than treating platformization at a global scale. Second, reliance on survey-based approaches, which impede cross-country comparability. The approach proposed here overcomes both limitations through the application of structural equation modeling, specifically, the Multiple Indicators Multiple Causes (MIMIC) framework, augmented with a Mundlak decomposition. This permits (1) the derivation of comparable platformization estimates for 86 countries, (2) the separation of short-run from long-run determinants of platformization, and (3) the avoidance of survey-induced bias. The digital economy and platformization have similarly been treated as unobserved variables in various studies (Brynjolfsson et al., 2023; Brynjolfsson and Collis, 2019; Kuzminov et al., 2025b; Milyakin et al., 2025; Watanabe et al., 2018); however, to the best of our knowledge, structural equation modeling methods have not previously been applied to the measurement of platformization. Earlier attempts to estimate the size of the platform economy arithmetically (Voytov and Polyakov, 2022) placed it at approximately 13% (the share of “platformized” industries) in Russia in monetary terms as of late 2021. However, the approach has become infeasible since 2022 following the cessation of public financial reporting by several major Russian companies.

Given the growing regulatory salience noted above, model-based estimates may serve government and independent bodies primarily in assessing the scale of the platform economy at the level of individual industries or the economy as a whole to inform regulatory decisions, whether independently of or as a complement to survey results. Moreover, the proxy indicators considered here may be employed for nowcasting purposes.

The remainder of the paper is organized as follows. Section 2 reviews definitions and characteristics of platformization that serve as the basis for variable selection in the MIMIC model, and existing survey estimates. Section 3 describes the MIMIC model specification. Section 4 presents the data and variable selection grounded in the Aghion–Howitt framework. Section 5 reports the estimation results, country-level estimates, and robustness checks. Section 6 concludes with a discussion of limitations and directions for future research.

Despite the extensive literature on two- and multi-sided markets (Armstrong, 2006; Rochet and Tirole, 2003), scholars, national statistical offices, and international organizations have yet to converge on a consensus definition, although certain universal features recur across studies: network effects, the platform operator as intermediary, reliance on digital network technologies, and the presence of aggregated supply and demand sides. The terminology employed in the literature includes, among others: (digital) platform economy, digital platforms, and two- and multi-sided markets (Table 2).

An evolutionary trajectory is apparent: from two-sided to multi-sided markets (1990–2000s), then — as technology advanced — to platforms (2010–2015), and subsequently to platforms aggregating activity across multiple multi-sided markets simultaneously (2015–present). In the present paper, we employ the following terms:

1. • Digital multi-sided platforms (DMPs): organizations (firms) that facilitate ­algorithmically mediated interaction among two or more groups of users (supply and demand sides) through digital technologies.
2. • Platform economy: system of economic relations and activities within DMPs, including the production, distribution, and consumption of goods and services through platforms acting as intermediaries.
3. • Platformization: the process whereby the platform economy’s role in total economic activity increases. Operationally defined as the ratio of household expenditure on goods and services acquired through DMPs to total consumer spending. It denotes the transition from direct bilateral interaction between supply and demand to trilateral interaction with a DMP serving as the organizer and guarantor of the transaction (purchase of goods, taxi rides, music streaming, medical services, and so forth).

These definitions allow us to distinguish among the object of study (DMPs), the domain under investigation (the platform economy), and the process being measured (platformization). By definition, DMPs are global actors; moreover, the multi-sided market model is replicable, a feature exploited by certain countries­ to promote the development of national platforms. However, infrastructure readiness, namely digital, social, and logistical, varies considerably, which accounts for cross-country differences in the pace and level of platformization. DMPs are typically digital companies developing the platform model (Watanabe et al., 2018), striving for or occupying near-monopolistic, duopolistic, or oligo­polistic positions and operating at transnational scale. The status quo may, of course, vary: a given DMP may begin as a marginal player, but in the long run such cases either exit the market or evolve toward a recognizable form of imperfect competition. The Schumpeterian mechanism on which the present analysis rests does not require literal monopoly — quality-ladder rents accruing under oligopolistic competition (Yandex–Google in Russian-language search; Ozon, Wildberries, Yandex Market, and MegaMarket in Russian e-commerce; Grab and GoTo in Southeast Asia) are sufficient to fund subsequent R&D investment.

It is worth emphasizing that the global presence of DMPs is not synonymous with uniform penetration. Despite comparable access to global platforms (e.g., Amazon, Uber, or Spotify, services of which are available in dozens of countries), actual usage levels are determined by local factors: logistical infrastructure maturity, digital payment penetration, household incomes, and the regulatory regime. E-commerce penetration, for instance, ranges from below 5% of retail turnover in several African countries to over 30% in South Korea and China — a pattern that motivates the use of country-level determinants in the modeling below.

As noted, DMPs typically operate across several complementary consumer verticals (see Table 1). A canonical example is Indonesia’s GoTo, which combines transport services, e-commerce, and financial services, with each vertical constituting a distinct platform; by some estimates, the conglomerate accounts for up to 2% of the country’s GDP. The underlying technology and business model rest on the internet (we henceforth assume that any offline component, where present, is also indirectly intermediated through the internet).

Computing platformization penetration, despite the apparent clarity of the formulation, is impeded by the difficulties outlined above: first, monetary metrics may not be disclosed in corporate reporting; second, country-level disaggregation is frequently unavailable (given that DMP coverage is assumed to be global); and third, B2C turnover is not always separable from B2B turnover. The attendant methodological difficulties and constraints are discussed in Voytov and Polyakov (2022). In addition, the methodological difficulty of separating transactional from non-transactional platform activity, and of defining product-market boundaries when network effects are present, is documented in detail by Shastitko and Markova (2020) in the context of antitrust market-definition tests.

Thus, platformization (P) is defined below in two alternative formulations. The target (monetary) formulation defines platformization as the ratio of consumer expenditure intermediated by platforms to total consumer expenditure:

$P_{\text{monetary}}=\frac{GM{V}_P}{C_{B2C}}$, (1.1)

where GMV_P denotes the gross merchandise volume of digital multi-sided platforms and C_B2C denotes household final consumption expenditure. The monetary formulation captures direct transactional DMPs whose revenue is generated as a commission on gross merchandise turnover. Pure attention-market platforms (search engines, social networks, short-form video) have advertisers, not households, as their primary demand-side counterparty; the corresponding revenue does not enter household final consumption expenditure C and is therefore excluded from GMV_P​​ as defined in (1.1). Their activity is nonetheless assumed to facilitate household consumption indirectly, by reducing search costs and accelerating the matching of consumers to transactional platforms and to non-platform retailers alike. Accordingly, (1.1) constitutes a conservative lower bound on the total economic activity associated with DMPs: the consumer-surplus and matching-efficiency contributions of attention-market platforms, examined separately in the GDP-B literature (Brynjolfsson et al., 2023; Brynjolfsson and Collis, 2019), lie outside the present operationalization by design. Computing GMV_P at the country level is impeded by the factors discussed above. Accordingly, the reduced (population-based) formulation is used for modeling:

$P_{pop}=\frac{\text{platform users}}{\text{population}}$, (1.2)

The transition from (1.1) to (1.2) is justified on the grounds that all MIMIC model variables are expressed as population-based ratios (internet-user share, employment share, etc.) without standardization, so that the latent variable η inherits their unit of measurement. This circumvents the calibration problem characteristic of MIMIC estimates of the shadow economy (Breusch, 2016) and avoids potential difficulties arising from dimensional mismatches between the reflective and structural equations’ feature spaces. The monetary interpretation (1.1) is used for validation purposes, benchmarking against the point estimates in Table 3 through an adjustment for the share of household final consumption in GDP. Under the target formulation, platformization is the ratio of consumer expenditure intermediated by platforms to total household consumer expenditure (C_B2C); under the reduced formulation employed for modeling, it is the share of the population using (involved in) platforms. The latter is expected to substantially exceed the former; (1.1) is then calculated as a product of share of household consumption and model η from (1.2).

In Russian literature, estimates of the platform economy’s size are reported in Kuzminov et al. (2025a), Koshel et al. (2025), Kuzminov et al. (2025b), measured as the share of platforms in total consumer turnover in Russia, estimated at 5.5%. Kapelyushnikov and Zinchenko (2024, 2025) estimate a related indicator, the share of employment in the corresponding sector, at 4%. The primary impetus for interest in this topic is the observation that regulation substantially lags the actual level of DMP development. Other researchers (Milyakin et al., 2025) find that a 3–5 percentage point increase in platformization (as the share of platform-intermediated turnover relative to total household consumer spending) is associated with a 1 percentage point increase in GDP, estimated using an input–output method. Citing internal data, Sberbank reported a figure of 3% in 2021.¹ Furthermore, according to statements by Russian government officials, the share of the platform economy is no less than 5% of GDP.²

In the international literature, the platform economy is usually treated as a component of the digital economy and assessed within the paradigm of “unobserved” GDP growth (Brynjolfsson et al., 2023; Watanabe et al., 2018). To this end, new metrics have been proposed — for instance, GDP-B (Brynjolfsson and Collis, 2019) to construct a monetary subjective valuation of platform utility based on surveys, which is then added to official figures. The average magnitude of this addition is estimated at 5.95% of GDP. For OECD countries, the direct and indirect contribution of “intermediaries” was estimated at 8.35% of cumulative GDP over 2009–2013 (Nielsen et al., 2013), or 1.67% per annum on average. Point estimates of the platform economy as a share of U.S. disposable income, based on data from JPMorgan Chase (Farrell and Greig, 2017), place the supply-side share at approximately 4% and the combined share at approximately 18% as of 2016. In the World Bank research (2023), the share of firm turnover attributable to platforms, based on survey results for 2019–2022, averages 6%, while the share of consumer expenditure is 30.9%, with developing countries predominating in the sample. Additionally, the OECD regularly conducts surveys across 32 member countries (covering the share of the population and enterprises buying and selling goods online, among other indicators), which suggest that approximately 16% of enterprise retail turnover is platform-related.³

Existing estimates of the platform economy’s size may be summarized as follows (Table 3). The median estimate is approximately 5.5% of GDP. It is noteworthy that statistical and survey-based estimates exhibit a degree of consensus. In the aforementioned study (Voytov and Polyakov, 2022), the share of “­platformized” industries­ in Russia (i.e., those in which transactions can be conducted via platforms) was estimated at 13% as of late 2021; however, when all consumer industries are included, counting those where platforms were not yet present, the estimate falls to 7%. Moreover, survey methods predominate in the cited studies, underscoring the need for a unified phenomenological approach independent of questionnaire design particularities, the results of which should nevertheless be taken into account for validating model-based estimates. Although research on the contribution of digitalization to GDP, including its unobserved component, has been conducted since the 2010s (Watanabe et al., 2018), empirical estimates of digital platforms’ share in aggregate consumption appear in the international literature only from the second half of the 2010s, and in the Russian literature from the early 2020s, whereas the theoretical foundations of platform markets have been under development since the early 2000s. We note that the platformization estimated below is not synonymous with the “gig economy” (i.e., full platform employment without alternative; Kapelyushnikov and Zinchenko, 2025).

As noted above, the principal obstacle to measuring platformization is the limited­ data availability, a constraint similar to that encountered in the shadow economy literature. However, whereas in the latter case data scarcity arises from agents’ motivation to conceal their activities, platformization may be characterized­ as a latent construct with certain qualifications. First, the volume of corporate information disclosed by the largest digital platforms, which may seek to avoid antitrust enforcement (Avdasheva and Geliskhanov, 2025), is typically limited. Second, platformization manifests partly as the restructuring of existing industries through transaction-cost reduction and partly as the generation of novel economic categories — the data brokerage layer, programmatic digital advertising, and the demand pull on specialized hardware (GPUs and TPUs) — that did not exist as discrete industries prior to DMP development. Both effects exhibit features of latent influence in the sense relevant for MIMIC estimation: they manifest through the same set of macroeconomic indicators (digital infrastructure, service-sector employment, ICT goods imports, new business registrations) regardless of whether the underlying mechanism is incumbent displacement or category creation, which is the substantive justification for treating platformization as a single latent construct rather than two separate phenomena.

The most widely used econometric tool for estimating such unobserved variables is structural equation modeling (SEM), originating in the LISREL framework of the 1970s and extensively applied in psychometric research (Muthén, 2002), and its reduced form, the MIMIC model (Multiple Indicators Multiple Causes; Zellner, 1970), which is suited to cases lacking natural units of measurement for the construct and exhibiting heterogeneity in residuals.

It must be acknowledged that this approach is not free of substantial limitations. The most systematic critique of the MIMIC approach in a macroeconomic context (shadow economy estimation), building on (Dell’Anno and Schneider, 2006), as an outcome of that discussion, and presented in (Breusch, 2016), may be distilled into four principal objections: (a) obstacles to direct interpretation in the statistical identification of the latent variable as a specific phenomenon; (b) non-robustness of estimates to changes in specification (hypersensitivity); (c) the arbitrariness of the benchmarking procedure (also referred to as “calibration”; Dybka et al., 2019) and of variable transformations (the requirement to apply identical transformations to provide a substantive interpretation of the resulting dimensionless indicator); and (d) the requirement that indicators be mutually uncorrelated conditional on the latent variable (conditional independence). Furthermore, MIMIC is a confirmatory rather than an explanatory model: coefficients are fitted within a researcher-specified structure of causes and indicators, meaning that the resulting estimates express associative rather than causal relationships. Interpreting parameters as explanatory is therefore inappropriate, and the theoretical basis justifying the chosen specification is of paramount importance.

In light of this critique, the present paper undertakes a series of measures to mitigate each limitation:

(a) Identification of the latent variable. Unlike shadow economy research, where the latent construct is defined by negation (production that cannot be observed), platformization has a more circumscribed theoretical content: it is a measurable process of DMP penetration into economic activity through identifiable channels (digital transactions, infrastructure readiness, and usage patterns) that could in principle be estimated mechanically (and likely will be in the future as available data improves), by analogy with market share estimation, but is currently unavailable for direct measurement. This narrows the space of alternative interpretations of the latent factor. Substantive identification is provided by the Aghion–Howitt theoretical framework of “creative destruction” described below, while statistical identification is achieved by fixing the loading of the indicator “share of the population with internet access” at 1, which establishes a natural upper bound on DMP penetration and preserves a direct interpretation of the latent variable η (hereafter Λ₁ = 1 in equation (3)).

(b) Measurement homogeneity and the absence of arbitrary transformations. As Breusch (2016) observes, a significant problem in MIMIC studies is the use of opaque variable transformations (standardizing indicators while leaving causes unstandardized, etc.), whereby the latent factor becomes an artifact of the chosen scale. In the present paper, causes, indicators, and the estimated latent variable are all expressed in uniform population-based ratio metrics. For variables exhibiting pronounced right-skewness, logarithmic transformations were applied uniformly to preserve monotonicity (see Table 4). Notably, these adjustments did not involve standardizing the indicators relative to the causal predictors.

(c) Sensitivity to specification. A systematic leave-one-country-out (LOCO) procedure is employed, along with bootstrapping for the computation of confidence intervals. The stability of the latent variable’s trajectory under changes in specification is interpreted as evidence in favor of a substantive, rather than statistical, origin of the extracted factor. The resulting estimates are also benchmarked against those reported in the literature reviewed above.

(d) Conditional independence of indicators and cross-country heterogeneity. The assumption of mutual uncorrelatedness of indicators conditional on the latent variable proves to be the most difficult to verify: each variable is expected to correlate individually with η while not correlating with other variables. In the context of panel data, a substantial portion of the observed covariation among indicators is attributable not to conditional independence violation per se, but rather to cross-country heterogeneity: countries at different levels of development exhibit systematically different levels of all indicators and causes simultaneously, creating an appearance of co-movement not mediated by the latent factor. In the literature, combinations of fixed and random effects with SEM have been applied in longitudinal studies in the health sciences (Bollen and Brand, 2010); in the present paper, a Mundlak decomposition is employed to separate within-country and between-country variance while preserving the random-effects structure necessary for MIMIC model identification (Mundlak, 1978). To our knowledge, the combination of Mundlak decomposition with a MIMIC specification has not previously been applied in a macroeconomic context, although analogous within–between decompositions have been employed with structural equation models in longitudinal health and sociological research; it permits the retention of unit-specific characteristics while achieving the requisite conditional independence of variables. Following this correction, the conditional independence assumption is required only for within-country variation of indicators over time — a considerably more tenable condition than global conditional independence in the pooled sample.

The reduced form of SEM, the MIMIC model, may be represented as a structural equation for computing the value of the latent (structural) variable (2) and a measurement (reflective) model for assessing its manifestation through observed indicators (3):

η_it = γ' x_it + ζ_it, (2)

y_it = Λ η_it + ϵ_it, (3)

where x_it is the vector of observed causes (causal or structural indicators) serving as regressors of the latent variable; η_it is the scalar latent variable; y_it is the vector of observed indicators (reflective indicators through which the latent variable is manifested); ζ_it is the disturbance in the equation for η; ϵ_it is the vector of indicator measurement errors; γ' is the coefficient vector linking causal variables to the latent variable in the structural equation (2); and Λ = (λ₁, ..., λ_q)' is the vector of factor loadings of the latent variable on the reflective indicators.

The estimation mechanism can be represented schematically as:

x ^γ→ η ^λ→ y,

and is used to substantiate the selection of Aghion–Howitt-inspired model para­meters. Let us denote Σ_x = Var(x_it), then:

Var(η_it) = ψ = γ' Σ_x γ + σ_ζ². (4)

The intuition behind the model is the assumption of an empirical relationship between causes and indicators, reproduced in the assumed specification of the MIMIC model with parameters Σ ≈ Σ(θ), in which case the model and empirical covariances are close in value. Under the standard assumption of uncorrelated measurement errors with causes Cov(ϵ_it, x_it) = 0, the matrix expressions of covariances are:

Cov(y_it, x_it) = Λ(γ' Σ_x), Cov(y_it) = ΛψΛ' + Ψ_ϵ,

where Ψ_ϵ = diag(σ_ϵ1², ..., σ_ϵq²) for uncorrelated measurement errors. In component form, for individual indicators r, s, and causes j:

$\mathrm{Cov}(y_{it}^{\left(r\right)},x_{it}^{\left(j\right)})={\Lambda }_r\sum _m{\gamma }_m{\Sigma }_{x,mj},\mathrm{Cov}(y_{it}^{\left(r\right)},y_{it}^{\left(s\right)})={\Lambda }_r{\Lambda }_s\psi +\mathrm{Cov}({ϵ}_{it}^{\left(r\right)},{ϵ}_{it}^{\left(s\right)})$.

The model covariance matrix of observable variables is written as:

$\Sigma \left(\theta \right)=\left(\begin{array}{ll}{\Sigma }_x& {\Sigma }_{y,x}\\ {\Sigma }_{y,x}& \Lambda \psi {\Lambda }^{\prime }+{\Psi }_{ϵ}\end{array}\right),{\Sigma }_{y,x}=\Lambda \left({\gamma }^{\prime }{\Sigma }_x\right),{\Sigma }_{y,x}={\Sigma }_{y,x}^{\prime }={\Sigma }_x\gamma {\Lambda }^{\prime },$

and the model hypothesis is that the empirical covariance matrix Σ is repro­ducible by the model matrix Σ(θ) for some set of parameters θ.

The current specification uses nine causal indicators (x_it) and seven effect indicators (y_it) for the combination of countries (i) and years (t) described below. The empirical covariance matrix Σ has a dimension of 16×16 and contains all pairwise covariances between variables:

$\Sigma ={\left[\begin{array}{ll}{\Sigma }_{xx}& {\Sigma }_{xy}\\ {\Sigma }_{yx}& {\Sigma }_{yy}\end{array}\right]}_{16\times 16}$.

The model covariance matrix is expressed through the MIMIC model parameters and has the same block structure. For example, the model and empirical covariance between “R&D expenditure” (x₂) and “new businesses” (y₃) is ­described as:

$\mathrm{Cov}(x_2,y_3)=\sum _{k=1}^9\mathrm{Cov}(x_2,x_k)\cdot {\gamma }_k\cdot {\lambda }_3$, (5)

where the amount of R&D expenditure correlates with latent “platformization” (γ_k), which manifests itself in new businesses emergence (λ₃), but this relationship also takes into account correlations with other causes.

The Mundlak decomposition, which reduces the conditional independence requirement to within-country variation, is formalized as follows:

$x_{it}=\underset{\text{Between}}{\underbrace{{\overline{x}}_i}}+\left(\underset{\text{Within}}{\underbrace{x_{it}-{\overline{x}}_i}}\right)$, (6)

where the first component (between) captures the country-specific time-average of the variable (structural differences between countries) and the second component (within) captures the deviation from the country mean (temporal changes within a country). The structural equation then takes the form:

${\eta }_{it}=\sum _{k=1}^9{\gamma }_k^W(x_{kit}-{\overline{x}}_{ki})+\sum _{k=1}^9{\gamma }_k^B{\overline{x}}_{ki}+{\zeta }_{it}$, (7)

where γ_k^W are the within-country effect coefficients (short-run dynamics) and γ_k^B are the between-country effect coefficients (long-run structural differences). The reflective model remains unchanged:

y_jit = λ_j η_it + ε_jit, j = 1, ..., 7.

The complete specification of the model consists of a modified structural equation:

η_it = γ₁^W(x_1it – x‾_1i) + ... + γ₉^W(x_9it – x‾_9i) + γ₁^B x‾_1i + ... + γ₉^B x‾_9i + ζ_it,

and a measurement equation (reflective model):

y_1it = 1.0 ∙ η_it + ε_1it, (fixed)

y_2it = λ₂ η_it + ε_2it

⋮

y_7it = λ₇ η_it + ε_7it. (8)

Parameter estimation is carried out by minimizing the discrepancy function using diagonally weighted least squares (DWLS), which is standard for structural equations.

The in dicators for estimating platformization draw on the Aghion–Howitt model of economic growth (Aghion and Howitt, 1992), which extends the theory of endogenous growth (Romer, 1990), both being heirs to Schumpeter’s concept of “creative destruction.” The choice of this theoretical framework is guided by three requirements. First, the model must interpret economic growth as a pheno­menon linked to imperfect competition, of which DMPs are prominent exemplars (the supernormal profits of participants in this market constitute a source of technology financing). Second, its parameters must be operationalizable through universal economic indicators available for any country. Third, the model must be endogenous, so that the selected parameters can be empirically estimated.

The Aghion–Howitt model, extending Romer’s “quality ladders,” describes a three-sector economy: a final goods market (consumption), an intermediate goods market (capital, production, and medium-skilled labor), and an R&D sector (highly skilled labor). Growth is endogenous and driven by monopolistic competition: the incumbent monopolist’s incentive to invest in R&D diminishes as current supernormal profits exceed the expected return on innovation. Over time, an innovator emerges who displaces the incumbent through technological superiority, whereupon the cycle repeats. The transmission mechanism propagates sequentially through the three sectors: R&D investment improves the quality of intermediate goods, which increases output and consumption in the final sector and redistributes market power in favor of the technological leader. The probability of an innovation arising as a product of R&D activity directly affects the rate of sustainable growth:

$g^{\ast }=\frac{\lambda L_R(\gamma -1)-r}{\theta }$,

where λ is R&D productivity, L_R is labor in the R&D sector, γ > 1 is the size of the quality improvement, r is the rate of intertemporal preference, and θ is the elasticity of intertemporal substitution. In other words, the greater the activity of the R&D sector, the higher the quality and productivity of the intermediate sector and, consequently, the greater consumption in the final sector.

In the context of the platform economy, DMPs assume the role of Schumpeterian monopolists. The mechanism of “creative destruction” is reproduced as follows: a platform achieves a temporary monopolistic position through network effects and the reduction of transaction costs, attracting the demand side, which in turn attracts the supply side. Accumulated monopoly rents are reinvested in R&D (cloud infrastructure, artificial intelligence, logistics technologies), raising the quality ­parameter λ (not to be confused with MIMIC λ). The cycle is reproduced when a more efficient competitor-platform emerges offering lower transaction costs — a “quality step” (γ > 1) in the Aghion–Howitt terminology. Notably, platform-based “creative destruction” has a distinctive feature: the monopolistic position may be sustained not only through technological superiority but also through self-reinforcing network effects, whereby comprehensive representation of the supply side itself constitutes a form of transaction cost reduction for the demand side. Thus, the causal variables in the MIMIC model represent the economic precursors for platformization, operationalizing the parameters of the Aghion–Howitt model (λ, L_R, institutional environment), while the indicators represent the observable manifestations of “creative destruction” across the three sectors of the economy. Causes capture the degree of readiness for platformization (technological and social capital), whereas indicators reflect the de facto level of technology penetration (Table 4).⁴

Empirical evidence from Russian retail confirms this dual mechanism: Andronova et al. (2021), applying Cox proportional-hazards survival analysis to approximately 130,000 firms over 2004–2018, find that aggregator platforms raise firm survival probabilities by reducing information asymmetries, while transformer platforms lower survival probabilities through direct competitive displacement (a race for efficiency within the supply side) — both effects consistent with the Schumpeterian mechanism described above.

The cause/indicator partition follows three criteria. A variable is treated as a cause if it (i) belongs to the R&D or intermediate sector and entails high capital expenditure with long payback — such that its deployment precedes platform activity rather than scales with it; (ii) constitutes a natural barrier to entry with limited reusability by competing market participants, helping sustain incumbents’ near-monopolistic positions; and (iii) is technologically antecedent to platform-mediated networks. A variable is treated as a reflective indicator if it (i) belongs to the final-consumption sector, (ii) admits high competition and ease of multihoming for end users, and (iii) reflects digital-corporation development and the expansion of retail consumption.

Under these criteria, fixed broadband and mobile cellular subscriptions occupy asymmetric positions despite their apparent similarity. Fixed broadband is a sunk infrastructure asset deployed by telecommunications operators in advance of household platform adoption, satisfying all three cause-criteria. Mobile subscriptions, by contrast, are a household-level expenditure that scales endogenously with platform demand, exhibits low switching costs, and reflects rather than enables platform activity. ATM density is treated analogously as a cause: it proxies for formal-banking penetration and deposit-account ownership — preconditions for digital-payment participation — rather than for the payment networks themselves. Payment networks (Visa, Mastercard, UnionPay, Mir) are themselves canonical two-sided markets (Rochet and Tirole, 2003) and are therefore omitted from both sides of the model: they are part of the latent construct under measurement, and including their volumes would conflate the latent variable with one of its manifestations. The same partition principle is consistent with prior latent-variable approaches to the digital economy (Brynjolfsson et al., 2023; Brynjolfsson and Collis, 2019; Watanabe et al., 2018).

It is important to note that the reflective indicators are deliberately distributed across the three sectors of the Aghion–Howitt model: the R&D sector (represented on the causal side — educational attainment and R&D expenditure — rather than through reflective indicators), the intermediate sector (new business registrations, ICT services exports, ICT goods imports, secure servers), and the final (consumption) sector (internet access, mobile subscriptions, services employment). This diversification minimizes the likelihood of co-movement among indicators arising from sectoral shocks not mediated by the latent factor, thereby strengthening the conditional independence assumption discussed above.

The data source is the World Bank’s World Development Indicators (WDI). The raw data span 2000–2024. However, the effective estimation sample covers 2000–2023, as the most recent two years are subject to substantial publication lags and incomplete coverage. Countries with ≥ 30% missing values across the estimation period were excluded, yielding a panel of 86 countries, since any imputation strategy could lead to biased MIMIC estimates. Grouping by region and income level follows the World Bank methodology. A detailed description of indicators and their descriptive statistics is provided in Appendix 1.

The colored sectors of the economy (Fig. 1) are those subject to the platformization transmission process described above. The direction of the arrows indicates whether a variable belongs to the group of causes or indicators. It is assumed that causes (R&D expenditure, educational attainment, credit expansion) precede platformi­zation, with fixed broadband connectivity serving as an infrastructural precondition and regulatory quality as an institutional prerequisite. The reflective indicators, in turn, capture the degree of utilization of these preconditions (internet usage, development of the tertiary sector and technology exports by the intermediate and final sectors) in the course of platformization.

The model fit statistics (Table 5) indicate a good correspondence between the data and the theoretical structure of the Aghion–Howitt model. The comparative fit index (CFI = 0.984) and the Tucker–Lewis index (TLI = 0.982) substantially exceed the threshold values of 0.95 and 0.9, respectively, indicating excellent incremental fit. The normed fit index (NFI = 0.952) also confirms significant improvement over the null model. The RMSEA (0.031) satisfies the conservative Browne and Cudeck criterion (< 0.05), indicating close fit. The GFI (0.952) and AGFI (0.947) provide additional confirmation of model quality. Given the multilevel nature of the data (86 countries, 24 years) and the within–between decomposition of effects, the overall fit statistics support the use of the model for examining the mechanisms of platformization through the lens of “creative destruction.”

The estimated causal coefficients (Table 6) reveal a pronounced asymmetry between structural (between) and dynamic (within) effects, permitting a separation of short-run from medium- and long-run effects on platformization.

Domestic credit to the private sector retains the largest between-effect (γ̂^B = 6.823, z = 7.87, p < 0.001) with a non-significant within-effect (γ̂^W = −1.466, z = −0.56, p = 0.574), which is consistent with the Aghion–Howitt model’s emphasis­ on the importance of R&D financing by the intermediate goods sector. The structural depth of the financial market, rather than its short-run fluctuations, determines an economy’s capacity to support platform investments, which are characterized by long payback periods and high initial costs (infrastructure development, attainment of critical user mass). This may, in particular, explain the interest of international investors in platforms in emerging markets, whose financing may to some extent substitute for insufficient domestic credit (e.g., India’s Jio Platforms, whose largest shareholders include U.S.-based DMPs, as well as Southeast Asian platforms such as Shopee, Grab, and GoTo, which have actively attracted Western venture capital).

Fixed broadband internet access exhibits significant effects in both dimensions: a dominant within-effect (γ̂^W = 1.268, z = 8.88, p < 0.001) and a significant between-effect (γ̂^B = 0.329, z = 6.49, p < 0.001). This variable serves as a determinant of the digital infrastructure of platformization: the expansion of broadband access within a country directly increases the potential for platform-mediated interactions by providing the bandwidth necessary for cloud computing, data processing, and multimedia services. The significance of both effects indicates that broadband access is a factor of both dynamic transformation and structural cross-country differences, unlike mobile connectivity.

Regulatory quality exhibits the largest between-effect among institutional variables (γ̂^B = 0.414, z = 11.76, p < 0.001) with a non-significant within-effect (γ̂^W = −0.107, z = −0.93, p = 0.351), which is indicative of institutional inertia: the rule of law, which reduces transaction costs in two-sided markets, is a charac­teristic of long-run structural differences among countries. The internal and external lags of short-run regulatory changes do not exert a significant effect on platformization. Regulatory quality here serves as a proxy for social development more broadly, which in the long run plays a significant role in platformization.

Urbanization is significant in both dimensions: the within-effect (γ̂^W = 0.999, z = 2.41, p = 0.016) and the between-effect (γ̂^B = 0.248, z = 9.09, p < 0.001). This reflects the role of physical infrastructure: urban agglomerations concentrate human capital, provide the critical mass of users necessary for two-sided markets, and reduce coordination costs. The significant within-effect indicates that ongoing urbanization within a country catalyzes platformization.

Trade openness has a significant between-effect (γ̂^B = 3.822, z = 4.87, p < 0.001) with a non-significant within-effect (γ̂^W = −4.546, z = −1.02, p = 0.308). Integration into global value chains thus serves as a structural condition for platformization: countries with historically high trade openness possess the infrastructure for cross-border platform interactions (logistics systems, payment gateways, regulatory compatibility). Short-run fluctuations in trade volumes, whether cyclical or conjunctural, are not associated with the dynamics of platformization.

The share of the population with tertiary education exhibits a significant within-effect (γ̂^W = 0.653, z = 3.61, p < 0.001) with a non-significant between-effect (γ̂^B = 0.077, z = 1.51, p = 0.131). This result suggests that the accumulation of human capital within a country is more important for platformization than the average level of education over the entire period, which is consistent with the concept of endogenous growth.

ATM density, serving as a proxy for offline-to-online financial access, is significant via the within-effect (γ̂^W = 5.359, z = 2.16, p = 0.031) with a non-significant between-effect (γ̂^B = 1.013, z = 1.47, p = 0.140). The expansion of physical financial infrastructure within a country is associated with platformization: an increase in financial access points facilitates the population’s connection to digital payment systems, forming a “bridge” between traditional and digital financial intermediation. The non-significance of the between-effect may be attributable to the fact that in developed economies, physical financial infrastructure is being superseded by digital infrastructure, and the role of ATMs exhibits a nonlinear relationship with the level of development.

Access to electricity is significant via the within-effect (γ̂^W = 0.600, z = 2.35, p = 0.019) with a non-significant between-effect (γ̂^B = −0.031, z = −0.63, p = 0.528). This result underscores the threshold nature of basic infrastructure: the expansion of electrification within developing countries creates a necessary precondition for the functioning of digital platforms. The non-significance of the between-effect is explicable by the fact that for countries that have already achieved full electrical coverage, this factor ceases to be differentiating.

R&D expenditure exhibits a negative between-effect (γ̂^B = −1.948, z = −4.18, p < 0.001) with a non-significant within-effect (γ̂^W = 0.835, z = 0.39, p = 0.698). The positive sign of the within-coefficient is consistent with the Aghion–Howitt prediction that R&D investment promotes innovation, but the effect is estimated imprecisely, likely owing to the slow-moving nature of R&D expenditure within the panel’s time dimension. The negative between-effect reflects an asymmetry in the global platform economy: R&D is concentrated in a small number of DMP-exporting countries (principally the United States and China, and to a lesser extent Israel, South Korea, and selected European economies), whereas platformization diffuses globally through the adoption of foreign platform models. Countries with lower average R&D intensity exhibit higher platformization precisely because they are technology recipients rather than originators: the presence of dominant foreign DMPs — whose competitive advantages rest on R&D conducted elsewhere — crowds out the domestic incentive to invest in platform-related innovation. This mechanism is analogous to the center–periphery dynamic described in the dependency literature: the R&D rents accrue to the originating economy, while the recipient economy absorbs the platform model without internalizing the associated innovation capacity. The robustness of the negative sign under bootstrap (the 95% confidence interval lies entirely below zero) supports the interpretation as a structural feature of the global platform economy rather than a statistical artifact.

The indicator-variable estimates (Table 7) reveal varying degrees of association between platformization and sectors of the economy, consistent with the three-sector structure of the Aghion–Howitt model.

Service sector employment has the largest factor loading (λ̂ = 0.638, z = 21.7, p < 0.001), which supports the hypothesis of structural economic transformation through the mechanism of creative destruction (Aghion and Howitt, 1992): by reducing transaction costs in the tertiary sector (logistics, finance, media, education), platforms expand service sector employment, displacing less efficient organizational forms through Schumpeterian competition.

The number of secure internet servers (λ̂ = 0.119, z = 21.7, p < 0.001) reflects the infrastructure component of platformization: the development of DMPs is underpinned by cloud infrastructure and computing scaling. New business registrations (λ̂ = 0.122, z = 11.7, p < 0.001) also exhibit a significant but moderate association, which may point to the dual influence of platforms on entrepreneurship: while lowering barriers to entry, platforms simultaneously intensify competitive pressure through network externalities on the supply side.

ICT goods imports (λ̂ = 0.051, z = 4.44, p < 0.001) are significant, whereas ICT services exports (λ̂ = 0.014, z = 0.52, p = 0.602) are not. This asymmetry is substantively meaningful: platformization is primarily associated with the consumption of digital technologies rather than their production. Countries that consume DMPs (without possessing national DMPs but with significant penetration of foreign counterparts) may exhibit high platformization without substantial ICT exports.

Mobile connectivity (λ̂ = 0.003, z = 7.76, p < 0.001), while significant, has the smallest factor loading, suggesting proximity to saturation: the indicator dif­ferentiates countries less than in prior periods, by analogy with access to electricity.

The population-based model results reported above are translated into monetary terms by multiplying by the share of household final consumption in GDP — under the assumption that all households contribute to this share in equal proportion (an assumption adopted for the purposes of computing the monetary indicator). The analysis of platformization dynamics by geographic group (Table 8, Fig. 2) reveals both a tendency toward relative convergence among regions and persistent inequality between countries by income group.

A methodological caveat is warranted regarding the translation of population-based estimates into monetary terms. The latent variable η, calibrated with Λ1 = 1 against the share of the population using the internet, captures the breadth of potential platform engagement rather than the intensity of platform-mediated transactions. Multiplying η by the share of household final consumption in GDP yields an upper-bound estimate of the monetary platformization ratio: it implicitly­ assumes that all digitally engaged individuals transact through platforms, and that their platform expenditure is proportional to average household consumption. Both assumptions are strong. Accordingly, the population-based η should be regarded as the primary output of the model, while the monetary estimates in Table 8 serve an illustrative and benchmarking function. For calibration against the literature, the monetary estimates for high-income countries (median ≈ 37% in 2022) should be interpreted as an upper envelope; the point estimates in Table 3 (median ≈ 7.2%) likely correspond to a fraction of η that reflects active, regular platform use rather than potential coverage. The ratio of these two figures (≈ 0.19) suggests that approximately one-fifth of the digitally connected population in high-income countries engages in regular platform-mediated consumption — a proportion broadly consistent with OECD survey data on online purchasing frequency.

All country groups for which estimates are inferred exhibit growth in the median value of the platformization (population-based ratio) from 2000 to 2023: high-income countries (HIC) — from 48.41 to 76.14 (+57%), upper-middle-income countries (UMC) — from 9.02 to 54.67 (+506%). Missing values for certain groups are attributable either to the absence of consumption data in the GDP structure in WDI or to an excessively high share of missing values; the convergence conclusions presented below should be regarded as preliminary for these groups. The UMC group exhibits substantially higher growth rates than HIC, which is consistent with the hypothesis of conditional β-convergence: controlling for structural determinants (financial depth, institutional quality), countries with lower initial levels of platformization grow faster. Nevertheless, the absolute gap persists: the HIC median in 2023 (76.14) exceeds the UMC median (54.67) by 39%, which constitutes evidence against absolute convergence. The apparent decline in the EAS indicator visible in Fig. 2 reflects the entry of new countries into the dataset and constitutes a compositional artifact, not a genuine reversal.

Within-group differentiation reveals heterogeneous dispersion dynamics. For HIC, the interpercentile range of the population-based ratio (p10–p90) narrowed from 29.45 (2000) to 27.34 (2023), which may constitute preliminary evidence of σ-convergence, although a more rigorous test would require the coefficient of variation, a measure less sensitive to outliers. For UMC, the range widened from 6.26 to 12.54, indicating increasing within-group heterogeneity as platformization proceeds: some countries in this group (e.g., China, Turkey, Brazil) have undergone a structural leap, while others (e.g., several countries in Central Asia and North Africa) have lagged substantially behind, which may be determined by initial institutional conditions.

Regional dynamics exhibit analogous tendencies. Europe and Central Asia (ECS) show median growth from 22.16 to 68.93 (+211%), and Sub-Saharan Africa (SSF) — from 5.49 to 57.45 (+946%). The narrowing of the interpercentile range in SSF from 8.79 to 4.45, concurrent with a substantial increase in absolute values, formally suggests features of σ-convergence within the group; however, it should be noted that the majority of observations are accounted for by a single country (South Africa). This result must be interpreted with caution, as the composition of countries with adequate data coverage differs between periods, which may generate an artifact of dispersion compression through the exit of the most lagging countries from the panel. For ECS, the range widened from 29.38 to 37.64, reflecting growing polarization between Western and Eastern European economies, apparently amplified by differences in institutional quality and financial market depth.

The persistently low dynamics of low-income countries (LIC) are associated with the dominance of between-effects in the estimated model: the structural depth of the financial market (β̂^B = 6.823), institutional quality (β̂^B = 0.414), and trade openness (β̂^B = 3.822) — variables on which LIC systematically lag behind other groups — determine cross-country differences in platformization. We emphasize that this interpretation is associative in character. Nevertheless, the result is consistent with the theoretical prediction of institutional complementarity (Acemoglu, 2009): improvement of individual infrastructure components (e.g., electrification, β̂^W = 0.600) may prove insufficient without accompanying systemic transformations of the financial and regulatory environment, as indicated by the non-significance of the within-effects of credit (p = 0.574) and regulation (p = 0.351).

Thus, the trajectory of platformization exhibits signs of conditional convergence: controlling for structural determinants, growth rates are inversely related to the initial level, yet absolute gaps between income groups persist, reproducing a pattern characteristic of technology diffusion under institutional heterogeneity. Platformization, unlike a generalized measure of economic growth, exhibits a more pronounced dependence on digital infrastructure (the within-effect of broadband access, β̂^W = 1.268) and financial depth (the between-effect of credit), which preserves the specificity of the construct and distinguishes it from the general­ level of development.

Two procedures are employed to assess the stability of the results: bootstrap analysis (500 iterations) and leave-one-country-out analysis (86 iterations).

The results of bootstrap analysis permit the classification of structural parameters­ into three groups by degree of stability (Table 9, Fig. 3).

The robust core comprises five causal variables that retain significance under resampling: domestic credit, broadband access, regulatory quality, urbanization, and R&D expenditure. Notably, this set represents the subset of point-estimate significant variables (see Table 6) that survives resampling — the within-effects of tertiary education, ATM density, and electricity do not — which provides evidence in favor of a substantive, rather than statistical, origin for the identified associations. The negative sign of γ̂_rnd is robust to resampling — the confidence interval lies entirely in the negative domain — which supports the crowding-out interpretation: R&D investment is concentrated in a small number of DMP-exporting economies, while the platform functionality scales to recipient countries where domestic R&D incentives are correspondingly diminished. At the same time, the width of the confidence interval for credit indicates substantial sensitivity of this estimate to sample composition, which is to be expected given the heterogeneity of financial systems across country groups; the same applies to trade openness.

Borderline variables — trade openness and tertiary education — are characterized by confidence intervals that marginally cross zero. For trade openness, the bootstrap median (2.74) remains substantively meaningful given a point estimate z = 4.87, and the loss of significance may be attributable to mechanism heterogeneity: the influence of trade integration on platformization varies substantially depending on the structure of trade (commodity exports versus integration into technology-intensive value chains). For tertiary education, the non-significance is likely attributable to differences in the character of higher education across countries.

Unstable variables — ATM density and access to electricity — are characterized by weak relationship with DMP development. ATM density exhibits the greatest instability among all parameters: the sign divergence between the mean (+0.01) and the median (−0.04) suggests a weak relationship between platform-based financial systems and national financial infrastructure. Access to electricity also loses significance, consistent with the threshold character of this factor: for the countries in the sample, the variation in electrification may be insufficient for stable identification of the effect, representing a sine qua non condition.

The reflective model demonstrates sound stability: five of seven indicators retain significance. Service sector employment, secure servers, and new business registrations exhibit narrow confidence intervals, confirming their informativeness as manifestations of the latent variable. Mobile connectivity, despite its small absolute loading, is significant — the interval does not include zero. The non-significance of ICT exports is robust: a symmetric interval around zero verifies the conclusion above that platformization is predominantly associated with the consumption, rather than the production, of digital technologies. ICT imports lose significance: the lower bound of the interval is essentially zero, classifying this indicator as borderline. Substantively, this may reflect the substitution of physical ICT equipment imports by cloud services as platform infrastructure develops.

Model fit statistics are robust to resampling: the median values of CFI (0.969) and TLI (0.965) exceed the thresholds for adequate fit even at the lower bounds of their confidence intervals (0.942 and 0.934, respectively). The median RMSEA (0.066) exceeds the point estimate of the main model (0.031) and slightly exceeds the conservative criterion.

Leave-one-country-out (LOCO) analysis (Fig. 4) indicates stability of the model’s­ conclusions under the exclusion of individual countries, with South Africa emerging as noteworthy: as the only large economy in the SSF group with adequate data coverage, its dominance in the regional sample limits the conclusions’ generalization regarding convergence within Sub-Saharan Africa. Nevertheless, no single country destabilizes the model as a whole, which constitutes evidence in favor of the substantive interpretability of the model.

The MIMIC estimates situate the present study within the existing literature along three axes.

First, the synchronization of monetary estimates. For high-income countries the model yields a median ratio (≈ 37%) marginally above the World Bank’s (2023) 31%. For Russia, however, the model-implied monetary share (≈ 34.5%) substantially exceeds the survey-based 5.5% reported by Kuzminov et al. (2025b) and the 4% in Kapelyushnikov and Zinchenko (2025). The discrepancy­ is the most substantively interesting result of the cross-country exercise: the Russian platform economy as a share of household consumption may be materially underestimated by survey-based methods, possibly owing to ­underreporting on the supply side and to the exclusion of attention-market and infrastructure-mediated activity from existing operationalizations or by focusing on rigid platform occupation.

Second, the negative R&D between-effect. The robust negative coefficient on R&D expenditure (γ̂^B = –1.95) does not contradict the existing view that DMPs are associated with technologically advanced firms; rather, it qualifies that view by indicating that innovation rents accrue predominantly to the country of origin. For recipient economies, the presence of foreign DMPs may exert direct competitive pressure on domestic R&D incentives and indirect pressure through the displacement of skilled labor toward platform-operated low- and mid-skill employment. The greatest economic gain from platformization — a point frequently underemphasized in the literature — accrues to the originator, not to the adopter.

Third, methodological contribution. The application of a Mundlak decomposition within a MIMIC specification, to our knowledge novel in the platformization literature, separates structural cross-country differences from within-country dynamics. The result is the empirical isolation of two distinct regimes: long-run platformization is governed primarily by financial depth, institutional quality, and trade openness, while short-run platformization responds to broadband expansion, urbanization, and human capital accumulation.