Health Information Technology Capital Intensity and Hospital Financial Performance: Evidence from U.S. Hospitals, 2018–2023

1. Introduction

1.1. Background and Motivation

Hospitals’ adoption and implementation of electronic health records accelerated after major federal policy initiatives—most notably the Health Information Technology for Economic and Clinical Health (HITECH) Act—intended to speed EHR uptake through incentive payments tied to “meaningful use.”[1] The Meaningful Use regulation operationalized these incentives by specifying program objectives and associated reporting requirements, shaping hospitals’ health IT priorities.[2]

Despite sustained spending, the relationship between HIT investment and hospital financial performance remains unsettled. Evidence syntheses have generally found positive effects of HIT on quality and process outcomes, but financial returns are less consistently demonstrated and may vary by context, implementation approach, and horizon. [3,4,5] Mechanisms plausibly operate in both directions: HIT can improve efficiency, coordination, and revenue cycle performance, but it also imposes up-front capital and ongoing implementation costs (training, workflow redesign, maintenance, vendor lock-in) that may depress margins—particularly during periods of operational stress. [4,5,6]

The question is especially salient in the post-2018 period, which combines a more mature HIT environment with major shocks to hospital operations and finances. COVID-19 created abrupt revenue disruptions (e.g., elective procedure pauses) and simultaneous demand for digital capabilities (telehealth, analytics, coordination), intensifying concerns about hospital liquidity and financial resilience. [7] These conditions make it timely to re-examine how HIT capital intensity relates to profitability and whether associations differ during COVID-era years.

1.2. Gap in the Literature

Prior work leaves four specific gaps that motivate this study.

• Capital-stock framing. Many studies operationalize HIT as adoption or use (e.g., EHR presence), while fewer measure HIT as a balance-sheet intensity concept—HIT capital stock relative to total assets—more directly aligned with financial statements and capital allocation decisions. [4,5,6]
• Distributional/nonlinear effects. Average effects may obscure nonlinearity: moderate HIT capital intensity may be compatible with positive returns, while very high intensity could reflect costly overinvestment, integration burden, or correlated structural features (system affiliation, service mix) that are themselves margin-relevant. Empirical tests of nonlinearity remain limited. [6]
• Recent period and COVID-era heterogeneity. Evidence focused on earlier adoption eras may not generalize to 2018–2023, when HIT is more embedded, and when COVID-era disruptions altered both care delivery and financial conditions. [7]
• Specificity and credibility checks. Associations may reflect broader capital intensity or stable hospital characteristics rather than HIT-specific effects. Placebo-style checks and within-hospital (fixed-effects) contrasts are not routinely deployed in this literature, limiting interpretability. [4,5,6]

1.3. Study Contribution and Overview

We address these gaps using a hospital balance-sheet approach to HIT investment and a design that emphasizes pattern testing (distributional structure, nonlinearity, COVID moderation, and credibility checks). We leverage the RAND-enhanced HCRIS hospital panel—built from Medicare hospital cost reports and value-added fields—because it provides consistent multi-year financial statement elements needed to construct capital intensity measures. [8] Cost reports are a core administrative source for hospital financial and utilization information, maintained by CMS in HCRIS. [9]

1.4. Research Questions (RQs)

RQ1. 

Among hospitals reporting HIT-designated capital stock, how is HIT capital intensity associated with total margin in each year 2018–2023?

RQ2. 

Is the association nonlinear (concentrated at higher intensity levels)?

RQ3. 

Does the association differ during the COVID years (2020–2021) versus other years?

RQ4. 

To what extent are observed associations consistent with within-hospital change (vs time-invariant hospital differences) and with HIT-specific effects (vs general capital intensity)?

1.5. Hypotheses

H1 (cross-sectional association):

Higher HIT capital intensity—especially the top quintile—will be associated with lower total margin (relative to Q1) in at least some years.

H2 (nonlinearity):

Any negative association will be concentrated at higher intensity levels (upper quintiles / upper spline segment).

H3 (COVID moderation):

The negative margin gap at higher intensity will be attenuated during 2020–2021 (positive interaction terms).

H4 (credibility checks):

If associations reflect time-invariant confounding or broader capital dynamics, then (a) hospital fixed-effects estimates will attenuate toward zero, and/or (b) placebo capital-intensity measures will show similar patterns.

2. Methods

2.1. Data Source and Sample

We used the RAND HCRIS calendar-year hospital panel (2024 release), which harmonizes Medicare hospital cost-report information from the Healthcare Cost Report Information System (HCRIS) into a longitudinal analytic file. [8,9] We restricted the study period to calendar years 2018–FY2023 to capture a mature post-adoption HIT environment and the COVID-era financial disruption years.

Within each fiscal year, we restricted the sample to hospitals with total assets ≥ $1,000,000 and selected one observation per Medicare provider-year using a completeness rule prioritizing nonmissing HIT stock, HIT payments, and total margin; remaining ties were resolved by selecting the record with the largest total assets. Primary analyses were restricted to “HIT reporters,” defined as hospital-years with nonmissing and nonnegative HIT-designated capital stock. This reporter restriction improves measurement of the primary exposure but also implies nonrandom selection into reporting, which we treat explicitly as a limitation (see Robustness/Reporting). The attenuation under hospital fixed effects and the similar placebo patterns suggest that observed associations may reflect time-invariant hospital characteristics or broader capital-intensity dynamics rather than HIT-specific impacts.

Data cleaning and construction. We applied prespecified cleaning steps, including provider-year deduplication using the completeness rule described above, exclusion of hospitals with total assets < $1,000,000, restriction to nonmissing and nonnegative HIT-designated capital stock for the reporter sample, and within-year winsorization of total margin; HIT intensity quintiles and spline knots were computed within year among reporters.

2.2. Variables and Measures

Outcome (profitability). The outcome was total margin. For year-specific models, margins were winsorized at the 1st/99th percentiles within year among HIT reporters to reduce leverage from extreme values, consistent with the distributional characteristics of cost-report financial items. A pooled 5th/95th winsorization was used as a sensitivity specification.

Exposure (HIT capital intensity). The primary exposure was HIT capital intensity, defined as HIT-designated capital stock (ending balance) divided by total assets. Because the distribution of HIT capital intensity is right-skewed among reporters (Figure 1), we operationalized exposure in two complementary ways:

• Quintiles (Q1–Q5) defined within year among HIT reporters (Q1 as reference), and
• Spline specification using ln(1 + HIT intensity) with knots at the 25th/50th/75th percentiles within year, to test whether slope changes are concentrated in the upper range.

Covariates. All models adjusted for hospital size (log total assets) and teaching status (major teaching vs non-teaching, as available). Pooled models included year indicators and a COVID-period indicator (2020–2021). If state fixed effects are included in the pooled specification, we report this explicitly and interpret them as controls for time-invariant state-level differences.

Placebo exposure (specificity check). To assess whether observed patterns are HIT-specific versus reflecting broader capital intensity, we constructed a placebo measure: non-HIT fixed asset intensity (non-HIT fixed assets divided by total assets), estimated in otherwise identical models. Because non-HIT fixed assets may be missing or inconsistently reported for some provider-years, placebo analyses are interpreted as credibility checks rather than definitive falsification tests.

Variable construction and field definitions. All analytic variables were constructed from RAND’s standardized HCRIS-derived fields; a complete crosswalk of source fields and transformations (HIT capital stock, total assets, total margin, teaching indicator, and derived intensity measures).

2.3. Statistical Analysis Plan (tied to RQs/Hs)

All analyses used OLS with robust standard errors and were conducted in Stata 18 (StataCorp LLC). Stata Statistical Software. College Station, TX: StataCorp LLC. [10]

A1 (RQ1/H1): Year-specific quintile models. For each year $t\in \{2018,\dots ,2023\}$, we estimated:

${\mathrm{margin}}_{it}={\displaystyle \sum _{q=2}^5}{\beta }_{q}1({\mathrm{HITQ}}_{it}=q)+\gamma \mathrm{l}\mathrm{o}\mathrm{g}({\mathrm{assets}}_{it})+\delta {\mathrm{teaching}}_{it}+{\epsilon }_{it}$

with Q1 as the reference category. This specification tests whether the highest-intensity group (Q5) differs from the lowest (Q1) within each year.

A2 (RQ2/H2): Nonlinearity via spline models. Within each year, we fit two-segment linear spline models in ln(1 + HIT intensity) with knots at the p25/p50/p75 cut points (estimated among reporters). The nonlinearity test assesses whether the slope in the upper segment differs from that in the lower segment, consistent with diminishing returns or high-intensity strain.

A3 (RQ3/H3): Pooled moderation model. We estimated a pooled model including year indicators and HIT quintile × COVID interactions (COVID = 1 for 2020–2021). The interaction terms test whether the margin gap for higher intensity hospitals is attenuated during the COVID years.

A4 (RQ4/H4): Credibility checks.

• Hospital fixed effects (within-hospital change): We estimated a pooled hospital-fixed-effects model among reporters to assess whether within-hospital changes in HIT intensity quintile track within-hospital changes in margin, recognizing that identification relies on within-hospital variation and controls only for time-invariant unobservables. Standard panel-data reasoning motivates interpreting differences between cross-sectional and fixed-effects estimates as evidence about stable confounding. [11]
• Placebo models: We re-estimated A1–A3 using non-HIT fixed asset intensity quintiles to evaluate specificity.
• Oster bounds: We used the coefficient stability approach to characterize sensitivity to omitted variable bias as an exploratory robustness check, interpreting results cautiously given assumptions required for extrapolation and the fact that stability metrics can behave poorly when the baseline $R^2$ is low, or coefficient movements are small. [12]

Missing data. Analyses used complete-case observations within the HIT-reporter hospital-year sample; missingness in the primary exposure is reflected in HIT stock reporting rates by year, and the placebo analysis was conducted on the subset with non-HIT fixed asset values observed.

2.4. Reporting and Robustness

We report this observational study in accordance with the STROBE statement and the RECORD extension for routinely collected health data (see Table A1 and Table A2). We pre-specify primary versus sensitivity analyses (primary: year-specific reporter-only quintile and spline models with 1/99 winsorization; sensitivity: pooled 5/95 winsorization, fixed effects, placebo, and Oster). We also document potential selection into HIT reporting and its implications for generalizability.

3. Results

3.1. Sample Characteristics

Across 2018–2023, HIT stock reporting among eligible hospitals declined from 24.6% in 2018 (545/2,212) to 19.4% in 2023 (982/5,063). Among reporters, HIT capital intensity was right-skewed (illustrated in FY2023; Figure 1), consistent with a small number of hospitals reporting very high HIT capital intensity values.

Accordingly, we used within-year intensity quintiles and flexible spline specifications for ln(1+intensity) to reduce sensitivity to extreme values and to assess nonlinearity in the descriptive statistics.

Sample characteristics are shown in Table 1; the FY2023 distribution of HIT capital intensity among reporters is shown in Figure 1.

3.2. RQ1/H1: Year-Specific Quintile Associations (Table 2; Figure 2; Figure 3)

In year-specific reporter-only OLS models with winsorized total margin (p1/p99 within year), higher HIT intensity was associated with lower margins in selected years, concentrated at the top quintile.

• 2018: Q5 vs Q1 β = −0.055 (p = 0.008; SE = 0.021; ~95% CI −0.096 to −0.015) and Q3 vs Q1 β = −0.040 (p = 0.011; SE = 0.016).
• 2019: Q5 vs Q1 β = −0.042 (p = 0.016; SE = 0.017; ~95% CI −0.077 to −0.008).
• 2020–2022: estimates were smaller and not statistically distinguishable from zero (e.g., 2020 Q5 β = −0.031, p = 0.293; 2022 Q5 β = −0.004, p = 0.829).
• 2023: Q5 vs Q1 β = −0.034 (p = 0.010; SE = 0.013; ~95% CI −0.060 to −0.008).

Adjusted margins by quintile for FY2023 are shown in Figure 2, and the FY2023 coefficient display (showing the Q2–Q5 indicators and covariates) is shown in Figure 3; both align with the negative top-quintile association observed in 2023.

Year-specific regression estimates are reported in Table 2. FY2023 adjusted margins by quintile are shown in Figure 2, and FY2023 model coefficients are summarized in Figure 3.

3.3. RQ2/H2: Nonlinearity

Spline models of ln(1 + HIT intensity) provided mixed evidence of nonlinearity by year. In 2021 and 2023, the post-knot slope term (l2) was negative and statistically significant for at least some knot choices, consistent with more negative associations concentrated in higher-intensity ranges. For FY2023, l2 was negative and statistically significant across knot placements (p25: p<0.001; p50: p=0.001; p75: p=0.005), indicating a more negative slope at higher levels of ln(1+intensity). In contrast, for 2018–2020 and 2022, slope changes were generally not statistically significant.

Spline results are shown in Table 3,

3.4. RQ3/H3: COVID Moderation

In pooled models with year indicators, higher HIT-intensity quintiles were associated with lower margins in non-COVID years (e.g., Q5 main effect β = −0.041; p<0.001). Interaction terms indicate partial attenuation during COVID years (2020–2021): Q4×COVID β = 0.033 (p = 0.029) and Q5×COVID β = 0.032 (p = 0.073). Interpreted as net COVID-year effects, Q5’s estimated gap vs Q1 was smaller during COVID years (≈ −0.041 + 0.032 = −0.009), while Q4’s net effect was near zero or slightly positive (≈ −0.021 + 0.033 = +0.012).

COVID interaction results are reported in Table 4 (pooled model).

3.5. RQ4/H4: Credibility Checks (FE + Robustness + Placebo)

In the pooled hospital fixed-effects model, HIT quintile indicators were not statistically significant (e.g., Q5 β = −0.015; p = 0.438), consistent with attenuation when using within-hospital rather than between-hospital variation. In robustness analyses using pooled 5th/95th winsorization of the raw margin outcome, the Q5 association remained negative (β = −0.029; p<0.001). In placebo models using non-HIT fixed-asset intensity quintiles (subset), the top quintile was also negatively associated with margins (Q5 β = −0.032; p = 0.006), consistent with the possibility that pooled associations reflect broader capital intensity or hospital type rather than HIT-specific effects.

Fixed-effects and robustness/placebo results are reported in Table 5 (FE), Table 6 (5/95 winsorization), and Table 7 (placebo).

4. Discussion

4.1. Principal Findings

Using a balance-sheet-based HIT capital intensity measure among HIT stock reporters (Table 1), higher HIT capital intensity, especially Q5, was associated with lower adjusted total margin in 2018, 2019, and 2023 (Table 2; Figure 3), with weaker and statistically non-distinguishable associations in 2020–2022. Evidence for nonlinearity was mixed, but upper-range associations were more consistently negative in 2021 and 2023 (Table 3). In pooled interaction models, COVID-year interactions suggested attenuation of the negative association in 2020–2021, but fixed-effects and placebo patterns caution against a HIT-specific causal interpretation.

4.2. RQ1/H1: Interpreting the Year-by-Year Results

The Q5 vs Q1 estimates should be read as conditional cross-sectional associations among reporters, not causal effects (Table 2). The across-year variation—including weaker effects in 2020–2022—fits a literature in which HIT–performance links vary by context, implementation, and complementary capabilities, and are not uniformly positive or negative. [15,16,17] In your data, the “re-emergence” of a negative Q5 association in 2023 is compatible with a high-capital-intensity burden interpretation (e.g., depreciation/financing and broader fixed-cost structure) rather than a clean “HIT payoff” story.

4.3. RQ2/H2: Nonlinearity (Upper-Tail Strain vs Diminishing Returns)

Where spline models show a more negative upper-segment slope (notably 2021 and 2023; Table 3), the most defensible interpretation is upper-tail strain/diminishing returns: at very high capital intensity, incremental digital capital may correlate with higher fixed costs, greater implementation burden, or organizational disruption unless matched by complementary process/workforce investments.20,21 This framing remains hypothesis-generating because the splines do not identify mechanisms, only shape.

4.4. RQ3/H3: COVID Moderation

The pooled interaction results are best framed as suggestive attenuation of the negative Q5 association during 2020–2021 (interaction terms near conventional thresholds). A cautious hypothesis is that pre-existing digital capacity may have supported operational adaptation during disruption (e.g., workflow reconfiguration, rapid scaling of virtual modalities), even if it did not translate into uniformly higher margins. This interpretation is consistent with broader evidence that the value of HIT is context-dependent and often hinges on complementary organizational changes. [1,17]

4.5. RQ4/H4: Why this Should not be Framed as Causal

Two outputs directly weaken causal language: (1) fixed-effects attenuation toward null indicates that between-hospital differences (time-invariant factors like baseline strategy, service mix, market position) likely drive much of the cross-sectional association; and (2) the placebo association with non-HIT fixed-asset intensity implies the pattern may reflect broader capital structure rather than HIT-specific dynamics. Together, the defensible conclusion is that high HIT capital intensity is a marker of lower margins in some years among reporters, not evidence that “HIT reduces margins.”

4.6. Strengths

Recent national cost-report-based panel covering 2018–2023; balance-sheet-based exposure definition (capital-stock framing); explicit tests for nonlinearity; COVID moderation test; and transparent robustness/credibility checks (FE and placebo).

4.7. Limitations and Future Research

Selection into HIT capital-stock reporting is limited and plausibly non-random, constraining external validity. The exposure measure may be heterogeneous across hospitals because “HIT-designated” capital can reflect differences in capitalization and coding practices, raising concerns about measurement comparability. The current specification includes a limited set of covariates, leaving residual confounding plausible; the unbalanced panel also reduces within-hospital leverage, consistent with the attenuation of effects under fixed effects. In addition, the placebo outcome is available only for a subset, and the IV approach is not informative given the extremely small complete-case sample, reinforcing the need to frame results as associative and to motivate stronger designs and richer data in future work.

4.8. Implications

Policy and management implications. The COVID-19 period illustrates why short-run margin effects should not be over-interpreted as “returns” to any single investment category. Peer-reviewed evidence shows that operating margins fell early in the pandemic while total margins were partly buffered by nonoperating revenue and relief funds, with vulnerability concentrated among certain hospital types (notably smaller and rural facilities). [18] At the same time, analyses of the Provider Relief Fund indicate that relief payments largely offset losses rather than systematically increasing profits, and that effects were concentrated among hospitals that were financially vulnerable pre-pandemic. [19] Against this backdrop, our finding that the negative high-intensity margin gap appears to narrow during the COVID years is consistent with the idea that system-wide shocks and temporary financing can materially reshape profitability signals, making it inappropriate to infer durable “margin gains” from higher HIT capital intensity during crisis periods. [18,19] A defensible implication is that if policy aims to sustain digital capability during shocks, support should be targeted toward financially constrained hospitals to prevent underinvestment or destabilization, rather than assuming uniform incentives will translate into improved margins. [18,19] Finally, evidence that administrative costs remain substantial—and that computerization alone does not necessarily reduce administrative burden—reinforces the managerial implication that any value from digital capital is likely contingent on complementary operational changes (workflow redesign, staffing models, and governance), not capital accumulation alone. [20]

5. Conclusion

In this national cost-report panel restricted to hospital-years reporting HIT-designated capital stock, higher HIT capital intensity—particularly the top quintile—was associated with lower adjusted total margin in 2018, 2019, and 2023, but not consistently in 2020–2022. Pooled interaction results suggested partial attenuation of this negative association during the COVID years (2020–2021), though the evidence was modest. Fixed-effects attenuation and a similar placebo pattern using non-HIT fixed-asset intensity suggest that stable between-hospital differences and broader capital-structure dynamics may explain much of the observed association, limiting HIT-specific causal interpretation. Future work should use richer covariates and designs better suited to causal inference (e.g., linked claims and market data, quasi-experimental shocks, and stronger instruments).