Does Digital Payment Adoption Reshape the Tax Mix? An Instrumental-Variables Analysis of Tax Composition in 38 OECD Countries, 2000–2022

1. Introduction

A companion literature to this paper has asked whether digital payment adoption raises how much tax revenue a government collects, typically finding that it does, largely through a compliance channel: transactions that leave a digital trail are harder to conceal from tax authorities than cash transactions (Immordino & Russo, 2018). A separate and logically distinct question, largely unexamined empirically, is whether digitalization changes not just how much revenue is collected but which tax collects it. If digital trails make consumption and corporate transactions specifically easier to observe, a natural prediction is that digitalization should raise the VAT and CIT share of total revenue relative to personal income tax, which is typically collected through payroll withholding regardless of payment technology and is therefore less mechanically linked to transaction visibility. This paper tests that composition hypothesis directly.

The question matters for reasons beyond curiosity about tax mix. Tax composition affects growth, distribution, and administrative burden in ways that are largely independent of the total revenue level: a shift toward VAT is typically regressive relative to income tax, and a shift toward CIT changes the incentives facing firm location and investment decisions. A government interested in leveraging digital payment infrastructure for fiscal purposes needs to know not only whether digitalization raises revenue, but whether it does so evenly or by systematically favoring some tax bases over others.

We test this using the same 38-country OECD panel (2000-2022) used in related work on this dataset, decomposing total tax revenue into personal income tax, corporate income tax, and VAT/GST shares. Because digital payment adoption is plausibly endogenous - more fiscally sophisticated or digitally mature governments may adopt digital payment infrastructure for reasons correlated with their tax mix already - we instrument digital payment use with two supply-side proxies for a country’s digital infrastructure legacy: the historical timing of broadband rollout and the geographic distance to the nearest submarine cable landing station, both of which are plausibly unrelated to a country’s tax policy choices.

The result, reported transparently including diagnostics that do not flatter the identification strategy, is a clear null: none of the four tax-mix outcomes we examine show a statistically robust relationship with instrumented digital payment adoption, in contrast to naive OLS estimates that do show significant associations for two of the four.

This null result is itself the paper’s contribution. It cautions against extrapolating from evidence on tax compliance gaps - which shows digitalization narrows the VAT gap (Bohne, Koumpias, & Tassi, 2023) - to a claim about tax structure, and it illustrates, in a second and independent empirical setting from related work on this dataset, how naive cross-country correlations between digitalization and fiscal outcomes can be materially misleading once endogeneity is addressed.

2. Theoretical Framework, Literature Review, and Hypotheses

2.1. Theoretical Framework: Administration Costs and the Tax Mix

The organizing theoretical framework is the administration-cost theory of tax structure formalized by Kenny and Winer (2006), who model a government’s tax mix as a rational response to the relative marginal costs of raising revenue from different bases. In their empirical framework, covering 100 countries over three decades, reliance on a given tax source rises as its administrative and compliance cost falls relative to other sources, alongside base-size and political-regime effects. Digital payment infrastructure is a natural candidate shock to this cost structure: it lowers the marginal cost of observing and verifying transactions specifically for VAT (levied on transactions) and CIT (levied on documented corporate income), while personal income tax, collected predominantly through employer withholding in OECD countries, has comparatively little to gain from payment-level digitization.

The Kenny-Winer framework thus generates a sharp, falsifiable prediction: if digital payments meaningfully lower the administrative cost of transaction-based taxation, the VAT and CIT shares of total revenue should rise relative to the PIT share as digital payment adoption increases. Gordon and Li (2009) formalize the complementary side of this logic in a developing-country setting, showing that governments rationally lean on tax bases that are easiest to monitor - explaining, for instance, low-income countries’ reliance on tariffs and financial-transaction taxes over income taxes when broad monitoring capacity is weak. Read together, the two frameworks imply that the tax-mix consequences of a monitoring-cost shock such as digital payment adoption should be largest where monitoring capacity was previously most constrained, and correspondingly small in a high-monitoring-capacity OECD setting such as the one studied here.

2.2. Information Trails, Third-Party Reporting, and the Limits of Consumer-Side Digitization

A second theoretical lens, grounded in the information-reporting literature on tax enforcement, sharpens what the administration-cost framework leaves unspecified: not just which taxes should benefit from digitalization, but how much they should benefit given the enforcement technology already in place. Kleven, Knudsen, Kreiner, Pedersen, and Saez (2011), using a randomized tax audit experiment in Denmark, show that evasion is close to zero for income subject to third-party reporting and substantial only for self-reported income, establishing that the enforcement value of any transaction record depends on whether a comparable third-party record already exists. Pomeranz (2015) extends this logic to the VAT itself, showing experimentally in Chile that the credit-invoice mechanism generates a self-enforcing paper trail: firms report truthfully not because of digital payments per se but because their trading partners’ own filings create a cross-checkable record up and down the supply chain. Slemrod’s (2019) survey of the tax compliance literature draws the general conclusion that information reporting - however it is generated - is consistently the single strongest predictor of compliance across enforcement instruments.

Applied to this paper’s setting, these results imply that a digital consumer payment is only one of several possible sources of third-party information, and its marginal contribution to enforcement is largest where no comparable record already exists. In OECD tax administrations, employer withholding already supplies a comprehensive third-party record for PIT, and the VAT credit-invoice mechanism already supplies one for VAT, independently of whether the underlying consumer payment is cash or digital. Evidence from a very different institutional starting point illustrates the contrast: Li, Wang, and Wu (2020) show that China’s Golden Tax Project III, which for the first time linked VAT invoice data, banking records, and customs information into a unified digital enforcement system, produced large compliance gains precisely because it created third-party information infrastructure that had not previously existed.

Two further studies show that the same logic extends beyond payment technology narrowly defined. Naritomi (2019), studying a Sao Paulo program that rewards consumers for demanding printed receipts, finds that giving consumers a stake in generating a transaction record raised reported retail sales by at least 21 percent over four years - direct evidence that the enforcement value of a paper trail operates through the same channel this paper investigates, whether the record originates from a digital payment or a consumer-demanded receipt. Okunogbe and Pouliquen (2022), using a randomized rollout of electronic tax filing among Tajik firms, show that digitizing the government’s side of the transaction record - rather than the payment itself - roughly doubles taxes paid among firms previously most able to evade.

Both results reinforce the prediction that digitalization’s enforcement value is context-dependent: it is largest precisely where an information gap is being closed for the first time, and correspondingly modest where a comparable information channel, digital or otherwise, already exists. Where such infrastructure already exists, as in the OECD panel studied here, the same technology should mechanically have less room to move the needle.

2.3. Evasion-Differentiated Taxation and the Direct-Indirect Tax Mix

A third and complementary strand, from the optimal-taxation literature, asks directly why economies rely on a mix of direct and indirect taxes at all rather than a single instrument. Boadway, Marchand, and Pestieau (1994) show that once evasion is allowed to differ across tax bases - for example, if income tax is evadable but commodity taxation is comparatively evasion-proof - a government optimally supplements an income tax with commodity taxes specifically to exploit this asymmetry, giving indirect taxation a structural role beyond simple revenue-raising or efficiency considerations. This theory predicts that a discrete change in the relative evadability of income versus consumption taxation should shift the optimal tax mix toward whichever base became relatively harder to evade.

It does not, however, predict that every marginal improvement in enforcement technology produces a further mix shift: once relative evadability across bases has converged to a low common floor, the model implies little remaining incentive to rebalance the mix in response to additional enforcement improvements at the margin. This is exactly the condition Section 2.2 argues is plausible for OECD tax administrations by the start of our sample period, and it generates a testable joint prediction together with the compliance evidence reviewed in Section 2.4: a positive tax-mix response to digital payment adoption should be more plausible where third-party information infrastructure was previously thin (as in Li, Wang, and Wu’s 2020 China setting, or the mobile-money settings reviewed below) and comparatively unlikely within an OECD panel where PIT withholding and VAT invoice-matching already function as mature, high-coverage enforcement technologies.

2.4. Digital Payments, Compliance, and the VAT Gap

A growing empirical literature supports the compliance half of this mechanism without testing the structural half. Immordino and Russo (2018) show, using Italian regional data, that a higher share of card payments relative to cash withdrawals is associated with a smaller VAT evasion rate. Bohne, Koumpias, and Tassi (2023), using EU-wide panel data from 2001 to 2021, find that a 100-percentage-point increase in e-money usage is associated with a reduction in the aggregate VAT compliance gap of roughly 1.9 percent, with a substantially larger effect identified around COVID-19-era shifts in payment norms. Kotsogiannis, Salvadori, Karangwa, and Murasi (2025), using the universe of Rwandan tax filings, show that e-invoicing mandates improve the effectiveness of VAT audits specifically.

Apeti and Edoh (2023), studying 104 developing countries from 1990 to 2019, find that mobile money adoption raises total tax revenue through a broader tax base, improved institutional quality, and simplified payment processing - a level effect documented in exactly the lower-third-party-information-infrastructure settings that Section 2.2 and Section 2.3 suggest should be most responsive to digitalization. At the descriptive level, OECD (2024, 2025) data and the European Commission’s VAT gap monitoring (Poniatowski et al., 2024) corroborate the enforcement channel: the average OECD and EU VAT policy gaps have narrowed alongside the growth of cashless transactions and e-invoicing-based pre-filled VAT returns, even as the OECD-wide tax-to-GDP ratio and tax-mix shares have remained comparatively stable over the same period. Each of these studies documents a compliance-gap, revenue-level, or audit-effectiveness channel; none directly tests whether the resulting gain is large enough, or specific enough to VAT and CIT relative to PIT, to move the aggregate composition of an already digitally mature tax system such as those in our OECD panel.

2.5. From a Level Effect to a Composition Effect

A level effect on VAT compliance need not translate into a composition effect on the tax mix, for at least two reasons the theoretical framework in Section 2.1, Section 2.2 and Section 2.3 makes explicit. First, if PIT withholding systems in OECD countries already achieve compliance rates close to those digital payments deliver for VAT and CIT - a plausible premise, given that OECD PIT compliance is already high due to third-party employer reporting - then digitalization closes a compliance gap that barely exists for PIT while narrowing the gap for VAT and CIT, which should mechanically shift the mix. Second, and pulling in the opposite direction, the information-trail and evasion-differentiated-taxation frameworks both suggest that if OECD countries had already made most of the administrative investments needed to monitor formal-sector transactions before the start of our sample window (2000) - through VAT credit-invoicing and PIT withholding rather than through consumer-side digital payment technology specifically - then the marginal contribution of digital payment adoption to the tax mix should be small precisely because relative evadability across bases has already converged. Distinguishing between these two possibilities is an empirical question, which we turn to next.

H1: Digital payment adoption is positively associated with the VAT share of total tax revenue, net of income, trade openness, and governance quality.

H2: Digital payment adoption is positively associated with the CIT share of total tax revenue.

H3: Digital payment adoption is associated with a shift in the composition of revenue toward indirect and corporate taxation and away from personal income tax, captured by a positive coefficient on the indirect-minus-direct index (VAT share minus PIT share).

3. Data and Variables

The panel covers the same 38 OECD countries and 2000-2022 window used in related analyses of this dataset (874 country-years). Tax composition is measured as three shares of total tax revenue (Tax_to_GDP): the personal income tax share, corporate income tax share, and VAT/GST share, each computed as the relevant OECD Revenue Statistics category divided by total tax revenue as a percentage of GDP. A fourth outcome, the indirect-minus-direct index, is defined as the VAT share minus the PIT share and summarizes the compositional shift hypothesized in H3 in a single measure.

Digital payment adoption is the World Bank Global Findex measure of the share of adults making or receiving a digital payment in the past year (Demirguc-Kunt, Klapper, Singer, & Ansar, 2022). Controls include GDP per capita (logged), trade openness, inflation, government effectiveness, the agricultural share of GDP, and population (logged), all from the World Bank World Development Indicators and Worldwide Governance Indicators. The excluded instruments are the infrastructure head-start variable (years between a country’s broadband rollout threshold and 2022) and distance in kilometres to the nearest submarine cable landing station, both time-invariant and described in more detail in related work using this dataset.

Submarine cable proximity and the timing of internet infrastructure rollout have precedent as identifying variation in the broader digitalization literature: Hjort and Poulsen (2019) use the staggered arrival of submarine internet cables along the African coast to identify the causal effect of fast internet on employment, on the logic that cable landing timing is driven by international engineering and commercial routing decisions rather than by the economic conditions of any single adopting country - the same logic underlying our use of these variables as instruments for digital payment adoption.

Table 1 reports descriptive statistics. The tax shares are well-behaved and not concentrated near a boundary - PIT share averages 23 percent of revenue (range 0.3-56 percent), CIT share averages 10 percent (range 2-42 percent), and VAT share averages 20 percent (range 0-43 percent, with the zero corresponding to the United States, which levies no federal VAT) - which means, unlike the bounded, ceiling-concentrated financial-inclusion outcomes examined in related work on this dataset, a fractional-response estimator is not required here; standard linear estimators are appropriate.

4. Methodology

Digital payment adoption is plausibly endogenous to tax composition: countries that have chosen a particular tax mix for unrelated historical or political-economy reasons may also differ systematically in digital payment adoption, and reverse causality is conceivable if a government’s fiscal capacity shapes its investment in digital infrastructure. We address this with two-stage least squares (2SLS), instrumenting digital payment adoption with the infrastructure head-start and submarine-cable-distance variables described in Section 3, which are supply-side and time-invariant and therefore plausibly uncorrelated with a country’s contemporaneous tax policy choices conditional on the control set. The second-stage equation for each tax-share outcome y is estimated as:

y_it = α + β·digital_payment_hat_it + γ′x_it + λ_t + ε_it

where digital_payment_hat is the fitted value from the first-stage regression of digital payment adoption on both instruments and the full control vector x, λ_t is a full set of year fixed effects, and standard errors are clustered by country throughout.

We report three comparison specifications alongside the main 2SLS estimate: naive OLS, which treats digital payment adoption as exogenous; a just-identified 2SLS using only the infrastructure head-start instrument, as a robustness check against weak-instrument or exclusion-restriction concerns tied specifically to cable distance; and a naive two-way (country and year) fixed-effects OLS specification, which cannot use the time-invariant instruments but provides an entirely different identification strategy - within-country variation over time - as a further robustness cross-check on the 2SLS results.

We report two instrument-diagnostic statistics rather than assuming validity. First-stage relevance is assessed via the partial R² of the excluded instruments and both a cluster-robust joint Wald statistic and the conventional (non-clustered) F-statistic, since the Stock-Yogo weak-instrument thresholds are calibrated to the latter. Instrument validity, in the overidentified two-instrument specification, is assessed via Sargan’s test of overidentifying restrictions; because a just-identified model cannot be tested for overidentification, the head-start-only specification is reported as a check on results but not itself subjected to this test.

5. Results

Table 2 reports the first-stage regression of digital payment adoption on the two instruments, controls, and year fixed effects. Infrastructure head start is a strong and highly significant predictor of digital payment adoption (coefficient 3.51, p < 0.001): each additional year of broadband head start is associated with roughly 3.5 percentage points higher digital payment adoption. Distance to the nearest submarine cable is negatively signed as expected but not individually significant (p = 0.373) once head start and the full control vector are included.

The instruments are jointly relevant - partial R² of 0.268, cluster-robust joint ²²(2) = 20.14 (p < 0.001) - though the conventional non-clustered joint F-statistic of 9.47 sits just below the classical rule-of-thumb threshold of 10 for strong instruments, which we flag explicitly as a limitation on the precision, though not necessarily the direction, of the 2SLS estimates that follow.

Table 3 reports the main results across all four specifications for each tax-mix outcome. The pattern is consistent and, we think, informative precisely because it runs against the paper’s own working hypotheses. Naive OLS shows a negative and highly significant relationship between digital payment adoption and the CIT share (−0.0028, p < 0.01) and a negative, significant relationship with the VAT share (−0.0016, p < 0.05) - both opposite in sign to H1 and H2 - alongside a marginally significant negative coefficient on the indirect-minus-direct index (p = 0.069), also opposite to H3.

Once digital payment adoption is instrumented, every one of these associations collapses toward zero and loses statistical significance: the CIT coefficient falls to −0.0014 (p = 0.252), the VAT coefficient essentially vanishes (0.0002, p = 0.929), and the indirect-minus-direct coefficient, while still negative, is no longer significant (−0.0026, p = 0.209). The just-identified specification using only the head-start instrument produces near-identical estimates throughout, and the two-way fixed-effects specification - an entirely independent identification strategy relying on within-country variation rather than cross-country instrumentation - corroborates the null result for every outcome, with all four fixed-effects coefficients statistically indistinguishable from zero and an order of magnitude smaller than the naive OLS estimates.

The overidentification diagnostics reported in the final column of Table 3 add an important qualification. Sargan’s test fails to reject the null of instrument validity for the CIT share (p = 0.242) and the indirect-minus-direct index (p = 0.964), supporting the joint use of both instruments for those two outcomes. For the PIT share (p = 0.001) and, most consequentially, the VAT share (p < 0.001), the test rejects the overidentifying restrictions, indicating that at least one instrument fails the exclusion restriction in those specifications - plausibly cable distance, which may correlate with a country’s general trade and customs infrastructure in ways that affect VAT administration directly rather than solely through digital payment adoption. For these two outcomes, we treat the just-identified, head-start-only estimate as the more credible one; reassuringly, it tells the same null story as the potentially invalid two-instrument estimate.

6. Robustness Checks

The main results rest on an instrumentation strategy whose first-stage F-statistic sits below the conventional strong-instrument threshold. Before interpreting the null result in Table 3 as conclusive, this section subjects it to four additional checks: weak-instrument-robust inference, cluster resampling, a placebo test of the exclusion restriction, and robustness to timing and additional controls.

6.1. Weak-Instrument-Robust Inference

Because the conventional first-stage F-statistic (9.47) sits below the classical rule-of-thumb threshold of 10, conventional 2SLS t-tests may be unreliable. We therefore compute an Anderson-Rubin-type test that does not require a well-identified first stage: we regress each tax-share outcome directly on the head-start instrument alone - the instrument that survives the Sargan test for every outcome in Table 3, and is therefore the more credible source of exogenous variation - and test whether its reduced-form coefficient differs from zero. This test is robust to weak identification by construction, since it does not rely on inverting the first stage. As Table 4 reports, none of the four reduced-form coefficients differ significantly from zero (p-values 0.225-0.935), corroborating the 2SLS null result under an inference procedure that does not depend on instrument strength.

6.2. Cluster Resampling Inference

With only 38 country clusters, asymptotic cluster-robust standard errors may understate estimation uncertainty. We recompute inference using a nonparametric cluster (block) bootstrap: we resample the 38 countries with replacement, together with all of a resampled country’s year observations, refit the full two-instrument 2SLS specification for each of 500 replications, and construct bootstrap standard errors and two-sided bootstrap p-values for the coefficient on digital payment adoption (Table 4). Bootstrap standard errors are similar in magnitude to the analytic cluster-robust standard errors reported in Table 3, and the implied bootstrap p-values range from 0.18 to 0.97 across the four outcomes. Small-cluster concerns do not appear to be driving the null conclusion.

6.3. Placebo Test: Pre-Sample Tax Composition

A central identifying assumption for the instrumental-variables strategy is that infrastructure legacy affects tax composition only through its effect on digital payment adoption, conditional on controls. We probe this assumption directly with a placebo test. Because head start and cable distance are time-invariant, a valid exclusion restriction implies they should show no meaningful relationship with tax composition in a period before substantial cross-country variation in digital payment adoption existed. Using the year-2000 cross-section of 36 countries with available data, we regress each tax-share outcome on the two instruments alone, with no controls (Table 5).

The result raises a genuine concern rather than confirming the exclusion restriction. Infrastructure head start significantly predicts the year-2000 PIT share, VAT share, and indirect-minus-direct index (p < 0.001 in each case), jointly explaining 30-48 percent of cross-country variation in these outcomes at the very start of the sample - well before the digital-payment channel this paper investigates could plausibly have operated. The most likely explanation is that countries with an earlier broadband and telecommunications head start are disproportionately the OECD’s founding, higher-income, historically direct-tax-reliant members, a composition confound correlated with both instruments and the outcome for reasons unrelated to digital payment adoption.

We report this test in full rather than omitting it. It weakens confidence in a strict causal reading of the 2SLS point estimates in Table 3, including the head-start-only specification that passes the Sargan test, since Sargan tests overidentifying restrictions given at least one valid instrument rather than exclusion-restriction validity outright, and cannot by itself detect the specific pre-existing-differences confound this placebo test uncovers.

Reassuringly, the confound this test identifies is one that would tend to load naive OLS and 2SLS estimates with a spurious historical relationship between infrastructure legacy and tax composition, which if anything biases the identification strategy away from, rather than toward, a null finding - meaning the 2SLS null in Table 3 is not an artifact of this confound suppressing a true effect. The two-way fixed-effects specification in Table 3, which relies only on within-country variation and is by construction immune to this specific cross-country confound, corroborates the same null result. In light of this placebo test, we treat the two-way fixed-effects evidence as the single most credible result in the paper, with the instrumented estimates offered as a complementary but more fragile cross-check rather than the primary basis for our conclusion.

6.4. Dynamic Timing and Additional Controls

Two further checks address alternative explanations for the null result. First, tax policy may adjust to a new payments environment only with a lag; we re-estimate the main 2SLS specification using digital payment adoption lagged three years, instrumented by the same two time-invariant variables. The lagged specification produces the same null pattern for all four outcomes (p-values 0.24-0.88 in the reduced sample of 753 country-years), suggesting the null is not an artifact of an implausibly short adjustment window.

Second, we re-estimate the main specification adding the shadow economy’s share of GDP and the IMF Financial Development Index as additional controls, to rule out that omitted variation in informality or financial-sector depth is masking a true composition effect. The estimation sample shrinks to 603 country-years with complete data on these additional variables, but the null result is unchanged for all four outcomes (Table 6).

7. Discussion

Three results stand out. First, the direction of bias in the naive OLS estimates is itself informative: rather than overstating a true positive base-broadening effect, naive OLS produces significant negative associations that vanish under both instrumentation and fixed effects. This pattern is consistent with reverse causality or omitted common trends running in the opposite direction from the mechanism this paper set out to test - for instance, countries that have historically relied more heavily on direct taxation may, for unrelated political-economy reasons, also have been slower or faster digital-payment adopters - rather than with digital payments genuinely eroding VAT and CIT shares.

Whatever the confounding source, the naive correlation would have led a researcher relying on OLS alone to draw materially incorrect conclusions about both the sign and the mechanism.

Second, the null result on tax composition is fully compatible with a positive effect of digitalization on the level of total tax revenue, which related literature on this dataset examines directly. The two claims - digitalization raises how much is collected, and digitalization does not change which tax collects it - are not in tension.

If compliance gains from digital payment adoption operate similarly across VAT, CIT, and PIT collection in already-high-compliance OECD tax administrations, or if the OECD-specific institutional environment (third-party PIT withholding, VAT credit-invoicing) had already captured most easily realizable digitalization-linked compliance gains before 2000, a level effect without a composition effect is exactly what the Kenny-Winer (2006) administration-cost framework would predict once administrative costs across tax types have already converged to a low common floor.

Third, the instrument diagnostics matter for how much weight to place on this conclusion. A first-stage F-statistic of 9.47 is not comfortably above the conventional weak-instrument threshold, and the overidentification failures for two of the four outcomes indicate that cable distance is not a fully credible exclusion-restriction-satisfying instrument for those specifications specifically. Section 6.3’s placebo test compounds this concern: infrastructure head start predicts pre-sample (year-2000) tax composition directly, which is inconsistent with a clean exclusion restriction even for the just-identified, Sargan-passing specification.

We view the corroboration from an entirely independent identification strategy - the two-way fixed-effects results, which do not rely on these instruments at all, are immune to the specific cross-country confound the placebo test uncovers, and tell the same null story - as the strongest evidence in the paper, with the 2SLS and weak-instrument-robust results serving as complementary, imperfect but directionally consistent, cross-checks.

Fourth, the battery of checks in Section 6 rules out several alternative explanations for the null result without fully resolving the identification concern the placebo test raises. The null is not a weak-instrument artifact (the head-start-only reduced form is itself insignificant), not a small-cluster artifact (bootstrap inference agrees with the analytic standard errors), not a timing artifact (a three-year lag produces the same pattern), and not an omitted-variable artifact specific to informality or financial depth (the result is unchanged adding shadow economy size and the Financial Development Index).

What the checks cannot do is manufacture a fully credible instrument where the placebo test has raised doubt about one; readers should therefore weight the two-way fixed-effects evidence most heavily among the results presented here.

8. Limitations and Future Research

The instrument-strength, overidentification, and exclusion-restriction caveats discussed in Section 6 and Section 7 are the paper’s central limitation and should temper a causal reading of the 2SLS point estimates specifically, though not, we think, the qualitative conclusion that a large, robust composition effect is not present in this sample - a conclusion the two-way fixed-effects results support independently of the instrumentation concerns.

The year-2000 placebo test in Section 6.3 indicates that infrastructure head start is correlated with historical, pre-digital-era differences in tax composition across OECD members, most plausibly reflecting founding-cohort and income-level differences rather than a channel specific to digital payments; future work with a richer set of candidate instruments, or a research design that does not rely on time-invariant infrastructure proxies at all, would be better positioned to isolate a cleanly exogenous source of variation in digital payment adoption.

The analysis also cannot rule out a longer-run composition effect operating beyond the 23-year window examined here, particularly given that several OECD countries introduced VAT decades before the sample begins, potentially already exhausting an earlier composition-shifting transition unrelated to digital payments specifically. The tax-share measures are annual averages that may smooth shorter-run compositional responses to discrete policy events, such as a VAT-rate change or a corporate tax reform, that coincide with but are not caused by digital payment adoption.

Future work could extend the analysis with a longer or non-OECD panel, where administrative costs and third-party information infrastructure have not already converged to a common low floor and a composition effect may be easier to detect - consistent with the level effects Apeti and Edoh (2023) and Li, Wang, and Wu (2020) document in developing- and middle-income-country settings with historically thinner enforcement infrastructure - or could exploit discrete, plausibly exogenous e-invoicing mandates - as Kotsogiannis et al. (2025) do for Rwanda - as a sharper source of identifying variation than the continuous, slow-moving infrastructure instruments used here.

9. Conclusion

This paper asked whether digital payment adoption reshapes the composition of tax revenue in OECD countries, not merely its level. Using an instrumental-variables design benchmarked against naive OLS and two-way fixed effects, we find no robust evidence that it does: naive correlations suggesting digitalization erodes the VAT and CIT share of revenue do not survive either instrumentation or the inclusion of country fixed effects, and the point estimates that remain are small and statistically indistinguishable from zero across every specification and outcome examined.

This is a genuine null result, reported with its accompanying instrument-strength, overidentification, and exclusion-restriction caveats rather than reframed as a positive finding.

The conclusion rests on more than a single specification. Across seven distinct empirical approaches - naive OLS, two-instrument 2SLS, a just-identified 2SLS using only the head-start instrument, two-way fixed effects, a weak-instrument-robust reduced-form test, a country-cluster bootstrap, and specifications adding a three-year adoption lag and additional controls for informality and financial depth - the same qualitative pattern holds: no tax-mix outcome shows a stable, statistically credible relationship with digital payment adoption once endogeneity or fixed effects are taken seriously.

The one check that did not simply reaffirm the null - the year-2000 placebo test - instead surfaced a genuine threat to the instrumental-variables strategy, and we report that finding in full rather than setting it aside. Its effect is to shift the weight of evidence in the paper toward the two-way fixed-effects results, which are immune to the specific cross-country confound the placebo test uncovers, while still leaving the overall conclusion intact: the fixed-effects estimates, using only within-country variation, tell the same null story as the instrumented ones.

Theoretically, the null is not merely an absence of a finding; it is consistent with what the paper’s three organizing frameworks predict once administrative costs and third-party information channels across tax bases have already converged to a low common floor. The administration-cost theory of tax structure (Kenny & Winer, 2006; Gordon & Li, 2009) implies that a cost-lowering technology shifts the tax mix only where a cost asymmetry across bases remains to be exploited.

The information-reporting literature (Kleven et al., 2011; Pomeranz, 2015; Naritomi, 2019; Okunogbe & Pouliquen, 2022) implies that a new information channel matters most where no comparable channel already exists, and OECD tax administrations already field two mature ones - employer withholding for PIT and credit-invoice matching for VAT. The evasion-differentiated optimal-taxation literature (Boadway et al., 1994) implies that the tax mix responds to discrete changes in relative evadability rather than to marginal technology improvements once evadability has already converged across bases. All three point in the same direction for this sample, and the empirical results are consistent with that joint prediction rather than merely failing to reject it for lack of statistical power.

For policymakers, the practical implication is that digital-payment-driven revenue gains in advanced economies should not be expected to systematically favor consumption or corporate taxation over income taxation, and tax-mix policy should continue to be treated as a separate lever from digital-infrastructure policy rather than a byproduct of it. This does not imply that digitalization is fiscally unimportant - related evidence on this dataset points to a positive effect on the level of revenue collected - only that its distribution across tax types in a mature OECD tax administration is not one of them.

For researchers, the results suggest that the level and composition effects of fiscal digitalization are separate empirical questions that require separate identification strategies, and that settings with historically thinner third-party information infrastructure - developing and middle-income economies, or OECD economies analyzed over a longer or pre-2000 horizon - remain the more promising ground on which to look for a genuine tax-mix response to digital payment technology.