Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
37,704 characters · 24 sections · 15 citation commands
Digital Maturity and Technical Efficiency in NHS Acute Trusts: Cross-Sectional Evidence from England
\thispagestyle{empty}
\setcounter{page}{1}
NHS acute hospitals face sustained pressure to improve productivity within constrained budgets. Capital investment in digital health technology is frequently proposed as a route to efficiency gains, and NHS England's Digital Transformation programme commits substantial resources on this premise nhse2021digital. Whether such investment translates into measurable productivity improvements at the trust level remains, however, poorly quantified.
The theoretical mechanisms through which digital maturity might improve efficiency are well established. Digitally integrated clinical workflows reduce duplicated effort; electronic prescribing reduces medication errors and associated costs; real-time patient flow data supports better bed management; and population health analytics enable preventive pathways that reduce emergency demand. Technology adoption is nonetheless costly, implementation disrupts existing workflows, and benefits may accrue unevenly across organisational contexts.
Prior evidence on hospital digital maturity and efficiency is modest and methodologically heterogeneous. Studies from the United States report positive associations between health information technology adoption and financial performance and care quality outcomes bardhan2013electronic, but the UK context differs substantially in funding mechanisms, incentive structures, and the nature of digital investment. NHS-specific evidence is largely limited to descriptive analyses and aggregate productivity indices that do not control for input usage nhsi2019tech.
In this paper, the association between digital maturity and technical efficiency is estimated for all NHS acute non-specialist trusts in England in 2024/25 using Bayesian stochastic frontier analysis (SFA). The Battese-Coelli one-stage approach bc1995 is adopted to estimate the efficiency equation simultaneously with the production function, avoiding the two-stage bias that arises when efficiency scores are regressed on covariates post-estimation. The model is estimated in a Bayesian framework, allowing prior information and full posterior uncertainty quantification for all parameters.
Three contributions are made. First, fully audited financial data from NHS Trust Annual Accounts are used to construct production function inputs, avoiding the cost-approximation assumptions of prior NHS efficiency studies. Second, physical capital is explicitly included as a production function input and its orthogonality to digital maturity is demonstrated empirically, directly addressing a potential confound. Third, a hierarchical Bayesian model is employed for pillar-level digital maturity effects, enabling differential returns across the seven Digital Maturity Assessment domains to be estimated with appropriate uncertainty quantification.
The analytical sample comprises 111 acute non-specialist NHS trusts in England in 2024/25. Two categories of trust are excluded. Acute specialist trusts (16 organisations) are excluded because their output mix is not comparable to general acute hospitals; such trusts focus on single clinical domains and their cost-weighted activity cannot be placed on a common frontier with general acute providers. Acute multi-service trusts (7 organisations) are excluded because they provide integrated acute and community care services under a single organisational structure. Their total operating expenditure, as recorded in NHS Trust Annual Accounts, therefore encompasses community nursing, social care liaison, and other non-acute activities, while the CWA captures only acute hospital activity. This creates a systematic input-output mismatch: inputs are inflated by community service costs that generate no measured output, producing spuriously low efficiency estimates that reflect the accounting structure rather than productive performance. A ratio of total operating expenditure to cost-weighted activity of up to 2.1 is observed in this group, compared with approximately 1.0--1.1 for standard acute trusts. Both exclusions are defined a priori on grounds of comparability rather than on observed results. The remaining sample covers all acute large, medium, small, and teaching trust subtypes as classified by the NHS England Oversight Framework nhse2025oversight.
The output measure is MFF-adjusted cost-weighted activity (CWA) from the NHS National Cost Collection 2024/25 nhse2025ncc. CWA aggregates healthcare resource group activity weighted by national reference costs and adjusted for the Market Forces Factor to account for input price variation across geographies. It is a standard output measure in NHS productivity analysis castelli2011.
Four inputs are included in the production function. Clinical full-time equivalents (FTE) comprises medical staff plus professionally qualified clinical non-medical staff, sourced from NHS Electronic Staff Records (ESR) September 2024 nhse2025esr. Administrative FTE comprises support to clinical staff and NHS infrastructure support from the same source. Non-labour expenditure (NLE) is computed as total operating expenditure minus actual staff costs, with both components derived from NHS Trust Annual Accounts (TAC) 2024/25 nhse2025tac, ensuring that a common accounting framework is used for both numerator and denominator. Physical capital is the sum of property, plant and equipment and right-of-use assets from TAC balance sheet data. All inputs are log-transformed and centred at their sample means; centring improves sampler geometry without affecting the elasticity interpretation of coefficients.
Digital maturity is measured by the NHS Digital Maturity Assessment (DMA) 2025 nhse2025dma, a structured self-assessment covering seven domains (`pillars'): Well-Led, Smart Foundations, Safe Practice, Support Workforce, Empower People, Improve Care, and Healthy Populations. Domain scores are averaged to form a composite score, which is standardised to zero mean and unit variance before inclusion in the model. It is acknowledged that the DMA is a survey instrument and may not fully capture operational digital capability. In particular, organisations with long-standing clinical system investment may be underscored relative to externally validated measures such as the HIMSS Electronic Medical Record Adoption Model, potentially attenuating estimated efficiency effects toward zero. A previous Digital Maturity Assessment was conducted in 2023/24. The 2024/25 assessment is used here as it incorporated revised question sets addressing limitations identified in the earlier iteration, providing a more refined measure of organisational digital capability. Results from the 2023/24 assessment are not used.
A related limitation concerns temporal ordering. The cross-sectional design uses contemporaneous digital maturity and efficiency measures, which is appropriate if digital capability and operational efficiency are jointly determined in a stable organisational equilibrium, but precludes assessment of whether efficiency gains lag digital investment. An earlier Digital Maturity Assessment was conducted in 2017; however, the instrument underwent substantial revision between 2017 and 2024/25, with changes to question sets, pillar definitions, and scoring methodology that render the two assessments non-comparable. Using 2017 scores as a lagged exposure for 2024/25 efficiency would therefore introduce measurement error from instrument change, likely exceeding any methodological benefit from temporal separation. Identification of lagged effects awaits a sufficiently long time series of comparable assessments.
The inefficiency equation includes the Index of Multiple Deprivation score for each trust's catchment population nhse2026imd, sourced from NHS acute hospital trust catchment population tables (2024, all admissions), a binary indicator for teaching trust status, and a binary indicator for financial deficit status, both from the NHS Oversight Framework nhse2025oversight.
In 2024/25, 89 of 111 trusts (80.2%) were recorded as being in financial deficit, reflecting the broader NHS financial position in that year. The limited variation in this indicator reduces its power to identify an independent deficit effect in the inefficiency equation.
A Cobb-Douglas production function is specified:
where $\widetilde{\ln x}$ denotes mean-centred log inputs for clinical FTE ($c$), administrative FTE ($a$), non-labour expenditure ($n$), and capital ($k$); $v_i \sim \mathcal{N}(0, \sigma_v^2)$ is a symmetric noise term; and $u_i \geq 0$ is technical inefficiency, so that the trust-specific efficiency score is $\text{TE}_i = \exp(-u_i) \in (0,\, 1]$.
Informative priors are placed on the production function coefficients, with prior means set to TAC-derived cost shares: $\beta_{\text{clinical}} \sim \mathcal{N}(0.378,\, 0.15^2)$, $\beta_{\text{admin}} \sim \mathcal{N}(0.197,\, 0.15^2)$, $\beta_{\text{NLE}} \sim \mathcal{N}(0.376,\, 0.15^2)$, and $\beta_{\text{capital}} \sim \mathcal{N}(0.049,\, 0.10^2)$. Cost shares are computed as the ratio of each input's cost to total expenditure using TAC data averaged across the analytical sample; these shares sum to unity by construction. Prior standard deviations of 0.15 (0.10 for capital, which has a smaller share) are selected to be permissive of substantial updating from the data while providing regularisation in the presence of high input collinearity (condition number 273). Cost shares are computed by allocating actual staff costs proportionally across FTE categories, which assumes a uniform average cost per FTE across staff groups; prior standard deviations of 0.15 allow substantial updating from the data if this assumption is violated. The intercept prior is set to $\alpha \sim \mathcal{N}(\overline{\ln\text{CWA}},\, 2^2)$, centred at the sample mean of log output. Noise and baseline inefficiency variance are given weakly informative priors: $\log\sigma_v \sim \mathcal{N}(-2.5,\, 1^2)$ and $\log\sigma_{u,0} \sim \mathcal{N}(-3,\, 1^2)$. Full prior specifications for the CES and translog robustness models are provided in the Supplementary Material.
Adapting bc1995, technical inefficiency is modelled as:
where $d_i$ is the standardised digital maturity composite score, $T_i$ is a teaching trust indicator, $D_i$ is a financial deficit indicator, and $m_i$ is the standardised IMD score. A negative $\gamma$ implies that higher digital maturity is associated with lower inefficiency variance, that is, with greater technical efficiency. The prior is specified as $\gamma \sim \mathcal{N}(0,\, 0.5^2)$, reflecting prior ignorance about the direction and magnitude of the effect. This prior places 95% of prior mass between $-1.0$ and $+1.0$: a value of $-1.0$ would imply that a trust one standard deviation above average in digital maturity has an inefficiency variance approximately 37% of the baseline, which is regarded as an upper bound on plausible effect sizes. Controls are given the same weakly informative prior: $\gamma_T, \gamma_D, \gamma_M \sim \mathcal{N}(0,\, 0.5^2)$.
Under the half-normal parameterisation, the expected technical efficiency of trust $i$ is $E[\exp(-u_i)] = f(\sigma_{u,i})$, a decreasing function of $\sigma_{u,i}$. A negative $\gamma$ therefore implies that higher digital maturity is associated with lower $\sigma_{u,i}$, which in turn implies higher expected technical efficiency. Mechanistically, this is consistent with digitally mature trusts exhibiting more standardised clinical processes, reduced process variation, and fewer extreme inefficiency realisations rather than a uniform shift of all trusts toward the frontier. The variance parameterisation follows bc1995.
Three model variants are estimated. Model A includes the composite digital maturity score in the inefficiency equation. Model B augments Model A with a dispersion term (the within-trust standard deviation across the seven pillar scores, standardised), testing whether balanced digital development matters independently of the composite level. Model C replaces the composite score with a hierarchical model over the seven DMA pillars. Pillar coefficients are drawn from a common distribution with mean $\gamma_\mu$ and standard deviation $\gamma_\sigma$, estimated from the data, so that:
This non-centred parameterisation enables partial pooling of pillar-specific estimates toward a common mean, providing regularisation when individual pillar effects are imprecisely estimated. The prior on $\gamma_\mu$ is $\mathcal{N}(0,\, 0.3^2)$ and $\gamma_\sigma$ is given a $\text{HalfNormal}(0.3)$ prior with a floor of 0.05 to prevent posterior degeneracy near zero.
Models are estimated using Hamiltonian Monte Carlo via PyMC 6.0.1 abril2023pymc. Four chains of 5,000 draws each are retained after 8,000 tuning iterations with target acceptance rate 0.97 (0.99 for Model C). Convergence is assessed via $\hat{R}$ statistics and effective sample size. All $\hat{R} \leq 1.010$ for parameters of interest; effective sample size for $\gamma$ exceeds 650 in all primary specifications. Log-likelihood is computed for leave-one-out cross-validation model comparison.
Functional form robustness is assessed via constant elasticity of substitution (CES) and translog specifications; details are provided in the Supplementary Material. Prior sensitivity is assessed via four specifications for $\gamma$: baseline $\mathcal{N}(0, 0.5^2)$, diffuse $\mathcal{N}(0, 1.0^2)$, sceptical $\mathcal{N}(0, 0.2^2)$, and enthusiastic $\mathcal{N}(-0.1, 0.5^2)$. Capital confounding is assessed by computing Pearson correlations between physical capital (and capital intensity) and digital maturity scores across all seven DMA pillars.
Descriptive statistics for the 111-trust analytical sample are presented in Table (ref). Mean CWA is \pounds 685 million; mean clinical FTE is 4,637; mean administrative FTE is 2,350; mean NLE is \pounds 353 million; and mean physical capital is \pounds 442 million. Digital maturity composite scores range from 1.67 to 3.44 with a sample mean of 2.5. Physical capital is found to be weakly and non-significantly correlated with the digital maturity composite ($r = 0.161$, $p = 0.090$) and orthogonal to six of the seven DMA pillars; one pillar (Empower People) reaches nominal significance ($r = 0.289$, $p = 0.002$) but does not survive Bonferroni correction for multiple comparisons ($p < 0.006$ threshold for 16 simultaneous tests), and one significant result in 16 tests is consistent with the expected false discovery rate under the null hypothesis. All correlations between capital intensity and DMA scores are small and non-significant. Full results including correlation heatmaps are presented in the Supplementary Material (Table S(ref) and Figure S(ref)). Capital endowment is therefore not regarded as a material confound for the efficiency premium, though the possibility of a weak association between physical capital stock and the Empower People pillar is noted.
Production function estimates are stable across all three model variants (Table (ref)). In the primary specification (Model A), estimated output elasticities are: clinical FTE 0.423 (95% CrI 0.25--0.59), administrative FTE 0.204 (0.09--0.32), NLE 0.264 (0.14--0.39), and capital 0.055 ($-$0.03--0.14). Returns to scale are estimated at 0.946 (95% CrI 0.896--0.997), indicating mildly decreasing returns consistent with coordination costs at larger trust scale. The clinical FTE elasticity is approximately twice that of administrative FTE, consistent with the relative cost shares and with clinical staff being the primary productive input in acute hospital care. The capital coefficient spans zero, reflecting limited independent variation after conditioning on labour and NLE inputs (pairwise log-input correlations 0.82--0.97); its inclusion does not materially affect the digital maturity coefficient.
The composite digital maturity score is negatively associated with technical inefficiency in all three model variants. In Model A, $\hat{\gamma} = -0.612$ (95% CrI $[-1.289, +0.005]$, $P(\gamma < 0) = 0.974$). The credible interval nearly excludes zero, with strong posterior probability concentrated below zero. Model B yields $\hat{\gamma} = -0.660$ ($P(\gamma < 0) = 0.971$); the dispersion term spans zero ($\hat{\gamma}_{\text{disp}} = -0.025$, 95% CrI $[-0.570, +0.490]$), indicating that the balance of digital development across pillars is not found to matter independently of the composite level. LOO-CV weights are indistinguishable across Models A, B, and C (maximum ELPD difference 1.0 standard error), and Model A is preferred on parsimony.
Trust-level efficiency scores range from 82.0% to 98.8% with a sample mean of 96.1% (standard deviation 2.6 percentage points). The Pearson correlation between the digital maturity composite and the posterior mean efficiency score is 0.711 (Figure S(ref) in the Supplementary Material). Mean efficiency by DMA quartile is: Q1 93.2%, Q2 96.2%, Q3 97.2%, Q4 98.0% (Figure (ref)). The Q1--Q4 gap of 4.8 percentage points implies approximately \pounds 20 million of additional cost-weighted activity at mean output levels per trust, or \pounds 1.1 billion in aggregate across the lowest-quartile trusts. Catchment deprivation is not found to have an independent efficiency effect in any specification ($\hat{\gamma}_{\text{IMD}}$ spans zero in all models), indicating that the digital maturity efficiency premium is not a proxy for serving a less deprived population.
In Model C, the mean pillar effect is estimated at $\hat{\gamma}_\mu = -0.211$ (95% CrI $[-0.450, +0.019]$), confirming a consistent negative association across pillars. Pillar-specific estimates, ordered by magnitude, are presented in Table S1 of the Supplementary Material. Healthy Populations is found to have the largest association ($\hat{\gamma} = -0.296$, $P(\gamma < 0) = 0.961$), followed by Safe Practice ($-0.266$) and Improve Care ($-0.258$, $P(\gamma < 0) = 0.931$). No individual pillar excludes zero after accounting for hierarchical shrinkage, consistent with limited power to distinguish pillar effects at $n = 111$. The consistent ordering is suggestive that population health management and care pathway optimisation capabilities have the strongest associations with efficiency, though this interpretation is exploratory given the available sample size.
Robustness check results are summarised in Table (ref). Under a CES functional form, $\hat{\gamma} = -0.628$ (95% CrI $[-1.289, -0.041]$, $P(\gamma < 0) = 0.981$), consistent with the primary specification. The substitution parameter $\rho$ spans zero (posterior mean 0.544, 95% CrI $[-1.081, +2.104]$), confirming that the Cobb-Douglas unit elasticity restriction is not rejected. LOO-CV weights are similar across Cobb-Douglas and CES (0.53 vs 0.47), providing no evidence against the primary specification. Under a translog specification, $\hat{\gamma} = -0.630$ (95% CrI $[-1.300, -0.059]$); LOO-CV weights again favour Cobb-Douglas marginally (0.56 vs 0.44). Full details of the CES and translog specifications and results are provided in the Supplementary Material.
Prior sensitivity results are presented in Table (ref). The baseline, diffuse, and enthusiastic priors all yield credible intervals that exclude or nearly exclude zero. Under the sceptical prior $\mathcal{N}(0, 0.2^2)$, which constrains 95% of prior mass to within $\pm 0.39$, the posterior mean is $-0.190$ (95% CrI $[-0.530, +0.170]$, $P(\gamma < 0) = 0.856$), spanning zero. This prior imposes strong a priori scepticism that may be regarded as conservative for a novel empirical question where no prior evidence exists to justify such tight shrinkage.
Evidence from this analysis indicates that digital maturity is associated with technical efficiency in NHS acute hospitals. The estimated association is negative in sign across all model specifications and functional forms, and is robust to alternative prior assumptions except under the most restrictive specification. The magnitude of the association implies that trusts in the highest digital maturity quartile operate closer to their estimated production frontier than those in the lowest quartile. These results are consistent with the hypothesis that organisational digital capabilities are correlated with observed variation in technical efficiency.
The pillar-level analysis, while underpowered to distinguish individual effects, suggests heterogeneity in the association across domains of digital maturity. Capabilities related to population health management and care pathway optimisation are estimated to have larger negative coefficients than those related to administrative or governance functions, though credible intervals for all individual pillars include zero after hierarchical shrinkage. This pattern is consistent with mechanisms operating through clinical pathway coordination and demand management rather than through administrative substitution, but the pillar-level results should be interpreted as exploratory given the available sample size.
A central concern in cross-sectional efficiency studies is that the variable of interest may capture unobserved organisational characteristics rather than the effect of a specific mechanism. Several aspects of the analysis reduce the plausibility of this interpretation. Physical capital, a potential proxy for organisational resources or investment environment, is not found to be systematically related to digital maturity across the DMA domains. Teaching status, financial deficit, and catchment deprivation are included directly in the inefficiency equation and are not estimated to have independent effects, suggesting that the digital maturity coefficient does not reflect these observable dimensions of organisational context. In addition, the association is stable across alternative prior specifications, including those that impose substantial shrinkage toward zero.
These considerations are not sufficient to establish causal interpretation in the absence of exogenous variation in digital maturity. Unobserved factors such as managerial capability, governance structures, or organisational culture may remain correlated with both digital maturity and efficiency. The results should therefore be interpreted as a conditional association that is consistent with, but does not identify, a causal effect of digital capability on technical efficiency.
Several further limitations are acknowledged. The NHS DMA is a fairly new and therefore immature self-reported survey instrument. Comparison with externally validated assessments indicates that some trusts with long-standing clinical system investment may be underscored by the DMA, which would attenuate estimated effects toward zero, implying that reported estimates are likely conservative. The sceptical prior sensitivity result ($P(\gamma < 0) = 0.856$) reflects the modest sample size of 111 trusts and should be interpreted accordingly; the association is present in the data but is not sufficiently strong to dominate aggressive prior shrinkage. inally, CWA captures acute inpatient and outpatient activity but excludes community services, research, and education, outputs that may be relevant for some trusts and that could introduce residual measurement error in the production function.
Digital maturity is found to be associated with technical efficiency in NHS acute hospitals. Trusts in the highest DMA quartile are estimated to operate approximately 5 percentage points closer to their production frontier than those in the lowest quartile. The association survives adjustment for input costs, physical capital, catchment deprivation, teaching status, and financial position, and is robust to functional form. The most conservative prior specification yields weaker evidence, reflecting a modest sample and the inherent difficulty of identifying organisational effects in cross-sectional data. Population health management and care pathway optimisation capabilities are found to have the strongest pillar-level associations with efficiency, though these results are exploratory. The results are consistent with digital capability being a contributor to technical efficiency in NHS acute hospitals. Causal identification would require exogenous variation in digital investment, for example from a natural experiment or staggered rollout, which is not available in this cross-sectional setting. Pending such evidence, the results nonetheless provide a quantitative benchmark for the scale of efficiency differences associated with digital maturity, and suggest that population health management and care pathway optimisation capabilities warrant particular attention in future investment prioritisation.
All data used in this study are publicly available. Cost-weighted activity and MFF-adjusted expenditure: NHS National Cost Collection 2024/25 (\url{https://www.england.nhs.uk/national-cost-collection}). Workforce data: NHS Workforce Statistics, September 2024 (\url{https://digital.nhs.uk/data-and-information/publications/statistical/nhs-workforce-statistics}). Financial data: NHS Trust Annual Accounts 2024/25 (\url{https://www.england.nhs.uk/financial-accounting-and-reporting/nhs-provider-accounts}). Digital Maturity Assessment: NHS England (\url{https://www.england.nhs.uk/digitaltechnology/connecteddigitalsystems/digital-maturity}). Catchment IMD scores: NHS Acute Hospital Trust Catchment Populations, April 2026 (\url{https://www.england.nhs.uk/publication/nhs-acute-hospital-trust-catchment-populations}). NHS Oversight Framework: NHS England (\url{https://www.england.nhs.uk/publication/nhs-system-oversight-framework}).
No conflicts of interest are declared.
This research received no specific funding.
Large language model assistance (Claude AI Sonnet 4.6) was used to assist the author with manuscript typesetting (LaTeX generation) and to perform code double-checking. AI was not used in the design of this work and did not influence the results or interpretation.