From reporting gaps to hospital cost drivers to enhance digital health decision making: A machine learning-assisted analysis of national hospital data

Abstract

Objective

Hospitals across the United States face growing operational and financial strain, resulting in closures that threaten healthcare access and system resilience. This study aimed to identify significant predictors of hospital total facility expenditures and to evaluate the performance of multiple imputation methods for incomplete data in the American Hospital Association (AHA) Annual Survey Database.

Methods

The de-identified 2022–2023 AHA survey data (n=12,359) comprising 34 financial, structural, and operational features was analyzed. Missing data were addressed using the Multivariate Imputation by Chained Equations (MICE) framework, comparing regression-based and machine learning–based algorithms. Random Forest (RF) imputation was selected for its superior accuracy based on fivefold cross-validation. Linear regression models were fitted on five RF-imputed datasets to identify key determinants of total facility expenditure (EXPTOT).

Results

RF-based imputation achieved the lowest error and highest consistency across variable types. Regression results identified full-time registered nurses (FTRNTF), facility size (GFEET), and property and equipment costs (PLNTA) as the strongest predictors of hospital expenditure (p<0.001). Hospitals with community designations, oncology or research services, and Joint Commission accreditation had significantly higher expenditures, whereas rural and community trauma centers reported lower costs. Geographic visualization revealed substantial disparities in hospital resources and expenditures, especially in rural areas.

Conclusion

Machine learning–based multiple imputation improves data completeness and modeling accuracy for hospital operations research. Findings highlight critical cost drivers and geographic inequities, informing data-driven policymaking and resource allocation in health system management.

Keywords

multiple imputation machine learning hospital expenditures healthcare operations American hospital association (AHA) annual survey random forest imputation geospatial analysis

Introduction

Over the past two decades, hospitals in the United States have faced mounting financial and operational pressures that have contributed to widespread closures, with profound consequences for healthcare access in affected communities.¹ For example, over 146 hospitals in rural areas closed due to this period, often due to vulnerabilities in occupancy rates, fluctuations in operating costs, and the structural limitations of small facilities.¹ In urban areas, hospital closures were linked with hospital financial structure, such as the higher likelihood of for-profit hospitals to close, as well as policy considerations such as if the state did or did not participate in Medicaid expansion.² In 2022, the American Hospital Association (AHA) reported a loss of $8.4 billion due to rising inflation, reduced reimbursement, and increasing labor costs.³ As recent constraints on federal sources of hospital funding continue, heightened attention has been placed on identifying indicators of hospital operational health and prediction thresholds to sound early warning of instability in providing care access.⁴ Given the central role hospitals play in emergency response, population health, and disease control, systematic monitoring of operational functionality has become an urgent priority for policymakers, health systems, and researchers.⁵

The AHA Annual Survey Database serves as one of the most comprehensive and authoritative sources of hospital data in the United States, and is extensively used by federal agencies, researchers, and policymakers.⁶ As a trusted longitudinal resource, it provides detailed information across a wide range of domains, including hospital demographics, service lines, utilization, workforce composition, financial performance, information technology, and telehealth capabilities.⁶ The database offers more than 1,300 standardized data elements for over 6,200 hospitals and 400 health systems nationwide, enabling robust analyses of hospital operations, trends, and system-level characteristics.⁷ While other databases are available to purchase hospital-level data, such as SK&A Healthcare Databases and from the Internal Revenue Service, the AHA Annual Survey databases focuses on those hospitals which are interconnected or at the juncture of systems ranging from decentralized to centralized, affording an overview which acknowledges resource and information exchange across geographic areas by organizational structure rather than by individual provider or physician groups.⁸ The database has been used to position the agenda of Centers for Medicare and Medicaid Services (CMS) surrounding care pricing transparency for patients prior to receiving care while reducing administrative burden in processing claims.⁹

The increasing digitization of healthcare systems has generated large-scale administrative datasets, such as from the AHA, which support research, policy evaluation, and operational decision-making. However, the issue of data completeness aligns with broader data quality concerns in the hospital sector; for example, a recent report noted that about three-quarters of hospitals faced at least one challenge to public health reporting in 2022.¹⁰ Despite its breadth and utility, the AHA survey often suffers from inaccuracies, reporting biases, and limited methodological transparency, which can undermine hospital strategic planning and governance policymaking.^11,12

It is essential to consider the mechanisms of missingness as described in Rubin’s framework when addressing incomplete data: data may be missing completely at random (MCAR), where the probability of missingness is unrelated to observed or unobserved features; missing at random (MAR), where the probability of missingness depends on observed features but not on unobserved values; or missing not at random (MNAR), where missingness is systematically related to unobserved values themselves.¹³ Multiple imputation (MI) offers a principled solution under the assumption of MCAR and MAR by generating several complete datasets in which missing values are replaced with plausible estimates derived from the observed data structure.^11,13,14 MI has been widely used and appraised to impute missing values in healthcare studies with higher precision and less bias in statistical analysis compared with traditional imputation methods and complete case analysis.^11,14–21 Recent studies have shown that machine learning–based approaches, such as random forest multiple imputation, outperformed traditional methods for electronic health records data by providing lower bias and more robust estimates across different missingness mechanisms.²²

In this analysis, our response variable, total expenditure, is fully observed, while many covariates have missing values. Accordingly, we assume that the missing covariates are MAR. This assumption is supported by the observation that the survey collects a large number of hospital characteristics related to staffing, size, operations, and organizational structure that are likely associated with reporting behaviors which may contribute to the pattern of missingness. Consequently, it is likely that the variables influencing missingness are captured by the observed data.

Missing data in healthcare datasets have traditionally been treated as a statistical problem addressed through imputation techniques. Methods such as Multivariate Imputation by Chained Equations (MICE) and machine learning-based approaches, including random forest imputation, have been widely adopted to recover incomplete observations and improve predictive modeling performance.^22–25 However, most studies have applied these methods primarily as a preprocessing step to enable downstream analyses, with relatively limited attention to the structural causes or implications of missing information in hospital reporting systems.^22,24,25 As healthcare systems increasingly rely on digital data, such as AHA, to support analytics and policy development, understanding reporting disparities becomes critical to decrease bias in the analysis.

The purpose of this study is to examine hospital reporting gaps within a national administrative dataset and to evaluate how incomplete reporting may influence health system analytics and policy-relevant interpretations. To achieve this objective, we develop an analytical framework that integrates three components: (1) characterization of missing data patterns across key hospital operational indicators, (2) recovery of incomplete observations using the best performed MICE imputation method after comparing a series of imputation methods, and (3) evaluation of how recovered data influence the analysis of hospital operational and financial indicators.

Using the 2022 and 2023 AHA Annual Survey Database, we analyze selected indicators of hospital operational functionality and examine variation across geographic location, hospital type, degree of system centralization, hospital size, and financial expenditures. In addition to assessing differences in these indicators, the study investigates patterns of data missingness and explores how incomplete reporting may affect statistical modeling and interpretation of hospital system characteristics.

Method

This retrospective observational study utilized hospital-level data obtained from the AHA database. The study was conducted using data from healthcare institutions across the United States, covering the period from 2022 to 2023. A total of 12,359 surveys were included in this study (n₂₀₂₂= 6,193, n₂₀₂₃=6,166). To observe the trend from 2022 to 2023, a feature to indicate the year of data selected was included in the study. Selected features (p=34) from AHA surveys were included in this study to identify the hospital's financial determinants (Table 1). To distinguish health systems based on the differentiation and centralization of their hospital services, physician arrangements, and provider-based insurance products, an identification system was jointly developed by AHA’s Health Research and Educational Trust, Health Forum, and the University of California-Berkeley. This system identified five groups with shared strategic and structural features: Centralized Health System, Centralized Physician/Insurance Health System, Moderately Centralized Health System, Decentralized Health System, and Independent Hospital System (Appendix I). The Core Based Statistical Area (CBSA) codes, a standardized set of geographic delineations in the United States developed by the U.S. Office of Management and Budget, were used to create visualization plots to observe the hospital financial distribution across the U.S.²⁶ Features and identified survey questions for individual purchase were selected based on literature review from previous research surrounding indicators of hospital operational function (e.g., expenditures, size, structure, geographic location, and labor).^12,27,28 A summary of 34 features’ key statistical inferences and their data completion rate were shown in Table 1. Data completeness varied, with only 11 of the 34 total features (32.4%) containing complete records (Table 1).

Table 1.

Features selected from AHA’s annual survey.

Features	Feature summaries¹	Data completion²
Continuous features (p=11)
EXPTOT (Total facility expenses, excluding bad debt)	238.4 Million (489.5 Million)	100%
HOSPBD (Total hospital beds)	150 (197)	100%
VEM (Emergency department visits)	23,403 (34,957)	100%
FTMDTF (Full-time physicians and dentists)	28.1 (117)	100%
FTRNTF (Full-time registered nurses)	239 (470)	100%
ADC (Average daily census (Inpatient Days))	99.8 (153)	100%
PLNTA (Property, plant and equipment at cost)	308.3 Million (623.9 Million)	53.6%
GFEET (Total gross square feet of your physical plant)	544,401 sq ft (958,631 sq ft)	51.0%
CEAMT (Total capital expenses)	19.9 Million (52.7 Million)	49.5%
VIDVZ (Number of video visits)	23,262 (88,8835)	44.5%
PRPM (Number of patients monitored through remote monitoring)	2,289 (20,541)	41.1%
Binary features (p=20)
CHC (Community hospital code)		100%
Yes	10,361 (83.8%)
MAPP1 (Accreditation by The Joint Commission)		100%
Yes	8,064 (65.2%)
MAPP18 (Critical Access Hospital)		100%
Yes	2,701 (21.9%)
MAPP20 (Sole Community Provider)		100%
Yes	575 (4.7%)
IINSPT (Partnership with an insurance provider/health plan)		54.8%
Yes	1,186 (9.6%)
CMRPAY (Performance-based contracts with commercial payers)		54.7%
Yes	3,880 (31.4%)
FAMADV (Patient and family advisory council)		52.7%
Yes	3,389 (27.4%)
COUTRHOS (Own community outreach)		63.5%
Yes	5,480 (44.3%)
EMDEPHOS (Own on-campus emergency department)		63.5%
Yes	6,407 (51.8%)
FITCHOS (Own fitness center)		63.5%
Yes	2,166 (17.5%)
HLTHSHOS (Own health screenings)		63.5%
Yes	5,698 (46.1%)
HLTRHOS (Own health research)		63.5%
Yes	2,257 (18.3%)
NUTRPHOS (Own nutrition program) Yes	5,826 (47.1%)	63.5%
ONCOLHOS (Own oncology services)		63.5%
Yes	4,057 (32.8%)
PALHOS (Own palliative care program)		63.5%
Yes	3,242 (26.2%)
SOCEHR (Record social needs screening results in EHR)		48.9%
Yes	5,636 (45.6%)
OUTMTX (Assess social needs interventions by outcomes)		55.0%
Yes	4,373 (35.4%)
WFAIPPD (Use AI or ML to predict patient demand)		46.5%
Yes	1,117 (9.5%)
COLLCLI (Collaborate with academia on community social determinants of health initiatives)		50.7%
Yes	2,487 (20.1%)
Year (Year of data collected)		100%
2023	6,166 (49.9%)
Multi-level categorical features (p=3)
SCNED (Hospital or health systems social needs screening)		54.8%
Yes, for all patients	3,684 (29.8%)
Yes, for some patients	2,467 (20.0%)
No	619 (5.0%)
Group (Group of health systems that share common strategic/structural features, see Appendix I)		67.5%
Centralized Health System	1,152 (9.3%)
Centralized Physician/Insurance Health System	499 (4.0%)
Moderately Centralized Health System	2,371 (19.2%)
Decentralized Health System	1,796 (14.5%)
Independent Hospital System	2,524 (20.4%)
TRAUML90 (Trauma center level owned/provided by my facility)		25.6%
Regional resource trauma center	526 (4.3%)
Community trauma center	622 (5.0%)
Rural trauma center	1,068 (8.6%)
Other, specific to some states	950 (7.7%)

¹Mean (Standard Deviation) for continuous features; Frequency (Relative Frequency) for binary and categorical features.

²The percentage of non-missing data points: completed data points/total sample size.

A descriptive analysis (statistical inferences: mean, standard deviation, frequency and relative frequency) was performed for the retrospective data collected from AHA between 2022 and 2023 using R programming language (version 4.2.0).²⁹ These analyses were conducted to characterize the distribution and patterns of missing values across hospital characteristics. Understanding these patterns provides insights into whether missing data may be associated with structural reporting differences across institutions. A geographic map of the United States was constructed using the ggmap package in R to examine spatial patterns in hospital expenditures and resource allocation across regions.³⁰

The MICE package in R was used to address missing data, which performs multiple imputations by iteratively modeling each feature with missing values as a function of other features in the dataset.²³ This method provides a robust framework for generating plausible values that preserve multivariate relationships in the data.²³ The performance of all appropriate methods outlined in the MICE package was evaluated for each feature type, such as continuous, binary, and categorical features (Table 2). MICE includes both parametric regression-based approaches, commonly used as standard implementations within the multiple imputation framework, and machine learning-based algorithms, which often outperform traditional methods when relationships among variables are more complex.²³Five-fold cross-validation was used to evaluate the performance of different imputation methods for each variable in the model. For each given variable, one-fifth of its observed values were repeatedly held out. The held-out observations were predicted five times using each imputation method. The goal of this comparison was not only to recover missing values but also to evaluate the feasibility of the best MICE imputation method for improving completeness of national hospital datasets used in digital health research.

Table 2.

Compared imputation methods in MICE.

Methods	Codes	Compared data	Description
Predictive Mean-Based Methods	pmm	Continuous Binary Categorical	Predictive mean matching; finds donors with similar predicted values.
	midastouch	Continuous Binary Categorical	Weighted predictive mean matching using nearest neighbors.
	sample	Continuous Binary Categorical	Random sampling from observed values.
Tree and Forest Methods	cart	Continuous Binary Categorical	Classification and regression trees for flexible imputation.
Tree and Forest Methods	rf	Continuous Binary Categorical	Random forest–based imputations for complex, nonlinear relationships.
Linear and Bayesian Regression Methods	mean	Continuous	Simple unconditional mean imputation.
	norm	Continuous	Bayesian linear regression imputation.
	norm.nob	Continuous	Linear regression ignoring model error.
	norm.boot	Continuous	Linear regression using bootstrap sampling.
	norm.predict	Continuous	Imputation with predicted values from regression.
LASSO-Regularized Regression Methods	lasso.norm	Continuous	LASSO-regularized linear regression.
	lasso.select.norm	Continuous	LASSO variable selection + linear regression.
	lasso.logreg	Binary	LASSO-regularized logistic regression.
	lasso.select.logreg	Binary	LASSO variable selection + logistic regression.
Logistic Regression Methods	logreg	Binary	Standard logistic regression for binary variables.
Logistic Regression Methods	logreg.boot	Binary	Logistic regression using bootstrap sampling.
Categorical and Ordered Methods	polyreg	Categorical	Polytomous logistic regression for unordered categories.
Categorical and Ordered Methods	lda	Categorical	Linear discriminant analysis for categorical imputation.

Multiple statistical techniques were applied to analyze the relationships among key features. Pearson correlation was used to assess the strength and direction of linear associations between continuous features, providing a measure of their co-movement.³¹ For categorical data, Cramér’s V was calculated to evaluate the strength of association between categorical features.³² Multiple linear regression was then conducted to examine the unique contributions of hospital characteristics to total facility expenditures while controlling other factors, thereby identifying the most significant cost drivers.³³ The regression model was used to examine how imputed data influences estimates of hospital expenditure drivers. Comparing model results before and after imputation provides insight into how reporting gaps may affect conclusions drawn from incomplete national hospital datasets. The significance level was 0.05 for all statistical analyses.

This study was reviewed and determined to be exempt by the Institutional Review Board (IRB) at Montana State University (Protocol No. 2024-1850-EXEMPT; approval date: July 3, 2025).

Results

As shown in Table 1, there are 23 features with at least 30% missing rate. Although the missing patterns varied across features, the presence of consistent red lines across the missing feature plot indicated that many features shared a similar pattern of missingness (Figure 1). This suggests that multiple hospital records were missing information across several features at the same time. Missingness was more often from smaller hospitals with less resources and expenses. The study was limited by a significant number of incomplete cases; only 1,182 out of 12,359 observations (9.56%) contained all the necessary information for a complete case analysis.

Figure 1.

Missing features distribution.

A correlation analysis was conducted using Pearson’s r for continuous features and Cramér’s V for categorical features (Figure 2). The results revealed strong to very strong linear correlations among several continuous predictors. For example, there was a correlation of r > 0.90 between HOSPBD and ADC, and r > 0.80 between FTRNTF and both PLNTA and GFEET, which suggested potential multicollinearity issues if all are included in a regression model (Figure 2). Similarly, strong associations among categorical features were observed, such as EMDEPHOS with COUTRHOS (V = 0.58) and HLTHSHOS (V = 0.56), HLTHSHOS with COUTRHOS (V = 0.57), and TRAUML90 with HLTRHOS (V = 0.61) (Figure 2).

Figure 2.

Pearson’s r and Cramér’s V Correlation Matrix.

The missing data was imputed using multiple imputations via the MICE algorithm. Table 3 presents the results for each imputation method averaged across associated variables. For this dataset, RF-based imputation was consistently among the best-performing methods, with CART-based imputation performing similarly well (Table 3). Normal linear regression also performed strongly for continuous variables (Table 3). Performance of the imputation methods was generally consistent within each variable type (continuous, binary, categorical), i.e., for a given variable type, the relative performance of each method did not substantially vary across individual variables.

Table 3.

Average performance of imputation methods across variables, with the top-performing method for each variable type shown in bold.

Method	Continuous RMSE	Binary accuracy	Categorical accuracy
norm.predict	1.59
cart	1.61	0.82	0.52
rf	1.61	0.82	0.53
pmm	1.71	0.78	0.43
lasso.norm	1.73
norm.boot	1.73
norm.nob	1.74
norm	1.74
lasso.select.norm	1.73
midastouch	1.72	0.34	0.39
mean	2.47
sample	2.72	0.65	0.32
lasso.logreg		0.78
logreg.boot		0.78
logreg		0.78
lasso.select.logreg		0.78
lda			0.46
polyreg			0.44

RF-based imputation was used to impute missing values for the hospital total facility expenses prediction model. Missing values were imputed using multiple imputation by chained equations with the mice package in R and random forest imputation (method = “rf”). Five imputed datasets were generated with 25 iterations per dataset. Random forest models were fitted using the package’s default hyperparameters (ntree = 10, mtry = √p). Convergence diagnostics demonstrated stable and flat trends for all features as iterations increased, indicating successful convergence with no evidence of drift or divergence (Figure 3).

Figure 3.

Convergence diagnostics for the imputed values with RF

The distribution of U.S. hospital total facility expenses (EXPTOT) from 2022 and 2023 was skewed due to the presence of extremely large expenditure values. Therefore, a logarithmic transformation was applied in all prediction models to reduce skewness, stabilize variance, and improve model performance. Other continuous features, such as HOSPBD, VEM, FTMDTF, FTRNTF, ADC, PLNTA, GFEET, CEAMT, VIDVZ, and PRPM, were also transformed with a logarithmic function to achieve the same benefits.

A generalized Variance Inflation Factor (GVIF) analysis was conducted on all five imputed datasets with RF-based method to assess and reduce multicollinearity in the linear regression models, specifically accounting for both continuous and categorical features. The adjusted GVIF (GVIF^{1/(2*Degree of Freedom)}) was used to compare the GVIF between categorical and continuous features. The results consistently identified HOSPBD, VEM, ADC, and EMDEPHOS as collinear features (adjusted GVIF>3) across all models. Compared with static infrastructure measures such as bed capacity or facility size, ADC provides a more direct measure of healthcare service utilization and was therefore retained in the final models to reduce redundancy while preserving the most relevant variation. After removing HOSPBD, VEM, and EMDEPHOS, all other features’ adjusted GVIFs were less than 2.

On average, the linear regression after imputation performed well in terms of MSE (0.15), MAE (0.29) and R2 (92.35%), and residual diagnostics (Figure 4). Since the residual plots from each imputed dataset were similar to each other, only one representative 4-in-1 residual plot was illustrated in Figure 4. Although the normal Q-Q plot of the residuals indicated slightly heavy tails, with a large sample size (n = 12,359), the coefficient estimates are likely reliable, though confidence intervals may be slightly anti-conservative.

Figure 4.

Model performance and residual diagnosis.

The linear regression model suggests that FTRNTF, GFEET, and PLNTA were the most significant continuous features when predicting the EXPTOT (Table 4). A 1% increase in the number of FTRNTF was the strongest predictor of hospital facility expenditures, associated with an approximate 0.52% increase in EXPTOT (β=0.5167, p<0.001), followed by GFEET with a 0.09% increase in EXPTOT (β=0.0934, p<0.001), and PLNTA with a 0.088% increase in EXPTOT (β=0.0882, p<0.001) (Table 4). As the most significant categorical feature, CHC was positively associated with costs, leading to an approximate 16.3% increase in expenditures compared to hospitals without this designation (β=0.1509, p<0.001). Hospitals designated as rural trauma centers or other state-specific trauma center classifications had expenditures approximately at least 12.64% lower (p<0.001) than the regional resource and community trauma center (Table 4). The presence of oncology services (ONCOLHOS) and health research services (HLTRHOS) were associated with an approximate 11.64% and 8.16% increase, respectively, in hospital facility expenditures (p<0.001) (Table 4). Owning palliative care program (PALHOS) and nutrition program (NUTRPHOS), accredited by the Joint Commission (MAPP1), partnering with an insurance vendor (IINSPT), applying AI/ML prediction model (WFAIPPD) and providing social needs screening services (SCNED) moderately increased the hospital expenses (β>0.049, p<0.001) (Table 4).

Table 4.

Linear regression summary aggregated with Rubin’s rule.

Features	Coefficient	Standard error	CI of coefficient	Test statistics	P-value
FTRNTF	0.5167	0.0082	(0.5006, 0.5328)	62.8227	<0.0001*
CHC (Yes)	0.1509	0.0139	(0.1237, 0.1781)	10.8713	<0.0001*
TRAUML90 (Other, specific to some states)	-0.1368	0.0450	(-0.2250, -0.0486)	-3.0405	0.0249*
TRAUML90 (Rural trauma center)	-0.1264	0.0301	(-0.1854, -0.0674)	-4.2036	0.0033*
TRAUML90 (Community trauma center)	-0.0768	0.0272	(-0.1301, -0.0235)	-2.8279	0.0186*
ONCOLHOS (Yes)	0.1164	0.0119	(0.0931, 0.1397)	9.8170	<0.0001*
GFEET	0.0934	0.0090	(0.0758, 0.1110)	10.4224	<0.0001*
PLNTA	0.0882	0.0043	(0.0798, 0.0966)	20.7420	<0.0001*
ADC	0.0843	0.0060	(0.0725, 0.0961)	14.0034	<0.0001*
HLTRHOS (Yes)	0.0816	0.0123	(0.0575, 0.1057)	6.6537	<0.0001*
MAPP1 (Yes)	0.0763	0.0088	(0.0591, 0.0935)	8.6865	<0.0001*
PALHOS (Yes)	0.0756	0.0140	(0.0482, 0.1030)	5.4081	<0.0001*
SCNED (Yes, for some patients)	0.0647	0.0154	(0.0345, 0.0949)	4.2036	0.0001*
SCNED (Yes, for all patients)	0.0390	0.0141	(0.0114, 0.0666)	2.7745	0.0061*
NUTRPHOS (Yes)	0.0529	0.0127	(0.0280, 0.0778)	4.1490	0.0005*
IINSPT (Yes)	0.0509	0.0124	(0.0266, 0.0752)	4.0995	<0.0001*
FTMDTF	0.0501	0.0031	(0.0440, 0.0562)	16.1306	<0.0001*
WFAIPPD (Yes)	0.0495	0.0172	(0.0158, 0.0832)	2.8719	0.0120*
Year (2023)	0.0377	0.0076	(0.0228, 0.0526)	4.9792	<0.0001*
CLUSTER (Decentralized Health System)	-0.0371	0.0174	(-0.0712, -0.0030)	-2.1319	0.0410*
CLUSTER (Independent Hospital System)	-0.0349	0.0178	(-0.0698, -0.0000)	-1.9598	0.0587
CLUSTER (Moderately Centralized Health System)	-0.0165	0.0180	(-0.0518, 0.0188)	-0.9158	0.3705
CLUSTER (Centralized Physician/Insurance Health System)	-0.0081	0.0222	(-0.0516, 0.0354)	-0.3630	0.7178
HLTHSHOS (Yes)	-0.0346	0.0135	(-0.0611, -0.0081)	-2.5675	0.0131*
CEAMT	0.0344	0.0031	(0.0283, 0.0405)	10.9569	<0.0001*
SOCEHR (Yes)	0.0329	0.0183	(-0.0030, 0.0688)	1.7960	0.0789
OUTMTX (Yes)	-0.0315	0.0115	(-0.0540, -0.0090)	-2.7319	0.0152*
FAMADV (Yes)	0.0212	0.0127	(-0.0037, 0.0461)	1.6641	0.1257
COUTRHOS (Yes)	0.0195	0.0135	(-0.0070, 0.0460)	1.4415	0.1619
COLLCLI (Yes)	0.0108	0.0101	(-0.0090, 0.0306)	1.0674	0.2900
VIDVZ	0.0078	0.0015	(0.0049, 0.0107)	5.0421	0.0001*
MAPP18 (Yes)	-0.0073	0.0132	(-0.0332, 0.0186)	-0.5483	0.5844
MAPP20 (Yes)	-0.0070	0.0190	(-0.0442, 0.0302)	-0.3697	0.7119
FITCHOS (Yes)	0.0036	0.0112	(-0.0184, 0.0256)	0.3226	0.7491
PRPM	0.0019	0.0021	(-0.0022, 0.0060)	0.8958	0.3784
CMRPAY (Yes)	-0.0016	0.0097	(-0.0206, 0.0174)	-0.1606	0.8727

*Significant features with p-value less than 0.05.

In order to evaluate the importance of imputing missing data, we compared results from the RF-imputed datasets to a complete-case analysis which included only observations with fully observed variables. The estimated coefficients from both models are shown in Figure 5, illustrating how reliance on complete cases affects the fitted model. For visualization purposes, the data were standardized using complete case means and standard deviations to place coefficients on comparable scales and thus differ to those presented in Table 4. Figure 5 includes the relative differences in the estimated coefficients, calculated by $100 \cdot \frac{(imputed - complete case)}{| complete case |}$ , and the absolute differences, calculated by $(imputed - complete case)$ . In particular, the confidence interval for CHC1 is extremely wide because the complete-case dataset contains only one observation with CHC = 0 and 1,176 with CHC = 1. Likewise, there were no observations with SCNED = 0, so coefficients for the SCNED variable were estimated only in the imputed model. While further variable selection could improve the complete-case results, substantial information would still be lost.

Figure 5.

Estimated coefficients and 95% confidence intervals from complete case (pink) and RF-imputed (cyan) datasets.

A sensitivity analysis was conducted using the delta method to assess potential bias from MNAR missingness mechanisms. Violations of the MAR assumption could occur if hospitals with unobserved characteristics were differentially likely to have missing covariate values.

For each continuous covariate with missingness (PLNTA, GFEET, CEAMT, VIDVZ, PRPM), systematic bias adjustments (delta parameters) were applied to each imputed dataset, ranging from $- 2$ to $+ 1$ standard deviations of the complete case data.^34,35 Specifically, imputed values were adjusted for each covariate $j$ by adding $δ \cdot S D (j)$ to originally missing values, where $δ$ is the chosen adjustment parameter and $S D (j)$ is the standard deviation for covariate $j$ from complete cases. This adjustment was applied separately to each covariate and each imputation m = 1, $\dots$ , 5. Negative delta values simulated scenarios where missing values were systematically lower than imputed values, while positive values correspond to systematically higher missing values. Regression models were re-estimated under each delta scenario using the imputed datasets, pooled results using Rubin’s rules, and examined whether the coefficient estimates and statistical significance changed across scenarios. A tipping point analysis identified the delta range within which each variable maintained statistical significance at p < 0.05.

The tipping point analysis is shown in Figure 6. While coefficient estimates changed with respect to delta, the results were largely stable and only the coefficient corresponding to PRPM changed sign and significance as delta increased. When delta adjustments were applied to one covariate (e.g., GFEET), coefficients for unadjusted variables (e.g., PRPM) remained unchanged, confirming that MNAR adjustments operated independently across variables.

Figure 6.

Tipping point analysis for continuous covariates across a range of delta values. Red dashed line indicates MAR assumption (delta = 0). Blue and orange lines indicate significance and non-significance at p < 0.05, respectively.

Moran’s I was calculated using distance-based spatial weights across multiple variables and present the results in Table 5.³⁶ Complete case model residuals exhibited significant positive spatial autocorrelation (Moran’s I = 0.0736, p < 0.001), indicating that nearby CBSAs have correlated prediction errors. Importantly, across all five imputed datasets, Moran’s I statistics for residuals ranged from 0.1180 to 0.1277 (all p < 0.001), showing stronger spatial clustering in imputed data compared to complete cases. It is difficult to determine if the higher spatial correlation is due to the imputation process or a higher spatial correlation in observations with missing data. The significant positive Moran’s I statistics indicate residual spatial dependence across CBSAs, suggesting that conventional ordinary least squares standard errors may underestimate uncertainty. Consequently, confidence intervals and p-values reported in Table 4 should be interpreted cautiously, as spatial clustering may result in anti-conservative statistical inference.

Table 5.

Moran’s I statistics for spatial autocorrelation of complete case residuals, imputed residuals (five imputations), and EXPTOT across CBSAs using distance-based spatial weights.

Model/Variable	Moran’s I	P-value
Complete Case Residuals	0.0736	<0.001
Imputation 1 Residuals	0.118	<0.001
Imputation 2 Residuals	0.1186	<0.001
Imputation 3 Residuals	0.1233	<0.001
Imputation 4 Residuals	0.1277	<0.001
Imputation 5 Residuals	0.1189	<0.001
EXPTOT	0.0515	<0.001

The outcome variable also demonstrated significant positive spatial autocorrelation (Moran’s I = 0.0515, p < 0.001), indicating that hospital expenditures cluster geographically, likely reflecting regional economic variation and healthcare market structure.

Rural hospitals showed higher missingness, likely due to less standardized data collection and limited administrative capacity. The imputation approach in this study borrowed information from geographically similar facilities, which may inadvertently amplify rural-specific missingness patterns. While sensitivity analyses suggested main findings were robust, rural facility estimates should be interpreted cautiously, as predicted metrics reflect both facility characteristics and shared structural factors affecting both missingness and expenditures. Future work with explicit rural-urban stratification could improve inference for this important subgroup.

Discussion

The research offers a foundation of discussion for hospital administrators and policymakers to develop strategic plans for resource allocation and cost containment during periods of economic headwinds and volatility in healthcare demands. This study examined hospital reporting gaps and operational indicators within the AHA Annual Survey dataset using an analytical framework that integrates three components: characterization of missing data patterns, machine learning–assisted multiple imputation, and evaluation of the influence of recovered data on hospital expenditure analysis. By combining these steps, the study moves beyond treating missing data solely as a technical challenge and instead situates data completeness as a key factor influencing health system analytics and policy-relevant interpretations.

This study characterized missingness across the 2022 and 2023 AHA Annual Survey datasets, addressed incomplete reporting using machine learning–based multiple imputation, and evaluated how recovered data influenced analyses of hospital operational and financial indicators.

Trauma center designation was negatively associated with expenditures relative to regional trauma centers, which may reflect differences in service complexity, patient volume, or resource allocation patterns across trauma center classifications. Cost is a common proxy given the volume of resource utilization in trauma centers particularly the complexity of injuries, surgical cases, and prolonged hospitalization. Rural regions experience long interfacility transit times and limited specialty providers or services which can delay treatment, therefore increasing costs as well as likelihood of poor outcomes especially if communication is disjointed between rural and trauma center care teams.³⁷ Similar findings have been described in the literature with research in orthopedic outcomes via machine-learning–based mortality prediction models surrounding resource utilization variables.³⁸ These findings support the interpretation that cost-related indicators in predictive models may capture underlying clinical complexity rather than purely financial characteristics of hospital care, reinforcing the clinical relevance of cost metrics in outcome prediction for trauma center cases, as an example. However, electronic health records may be incomplete or inconsistently structured across departments and facilities, such as specialties (orthopedics) and general care found in rural areas. Imputation techniques, such as random forest, have been deployed for heterogeneous, incomplete clinical datasets containing both continuous and categorical variables.³⁹ However, uncertainty remains when imputation is necessary with incomplete datasets, and thus limits generalizability outside typologies of healthcare systems or means of data management at the organizational level without external validation from more consistent sources.³⁹

Using the completed dataset, several hospital-level characteristics were identified as significant predictors of total facility expenditures. These findings suggested that hospital expenditures tend to scale with staffing capacity (FTRNTF), infrastructure footprint (GFEET & PLNTA), and the provision of specialized clinical and research services (CHC, ONCOLHOS & HLTRHOS). Conversely, hospitals designated as lower-level trauma centers were associated with comparatively lower facility expenditures relative to regional-level trauma centers.

Several of these indicators were highly correlated, particularly nursing staffing capacity (FTRNTF), averaged daily census inpatient days (ADC), hospital beds (HOSPBD), facility size (GFEET), and capital expenditures (PLNTA). The strong correlation among these variables suggests that hospital expenditures are closely tied to overall organizational scale and infrastructure intensity.

As the most significant indicator with largest influence on hospital facility expenses, FTRNTF was used as a representing indicator of hospital size and resources usage to be compared with total hospital facility expenses in Figure 7. The bivariate choropleth maps of selected top features, including FTRNTF, CHC, ONCOLHOS and HLTRHOS, associated with hospital total facility expenses were presented in Figure 7 to show how hospitals in each area performed across CBSAs in the U.S. High expenditure and high labor resource CBSAs clustered in major metropolitan regions, such as Northeast, Southeast and Pacific coastal states and regions while rural areas, such as Intermountain West and Midwest, often showed lower expenditures and fewer hospital-related resources (Figure 7). A nationwide disparity was observed in the availability of oncology and research services, with a notable absence of these services in rural areas (Figure 7). These disparities reflect what Probst and colleagues' term structural urbanism, which refers to systemic biases rooted in a market-based health care model that prioritizes population-dense areas and creates inefficiencies in delivering services to sparsely populated rural communities.⁴⁰ For example, a profoundly rural state, Montana, where over 34% of residents live in rural areas, reported only 7 CBSAs with community hospitals and oncology services (Figure 7). These services are highly concentrated on the west side of the state, with only 4 CBSAs offering research services (Figure 7). For rural residents, this geographic barrier can create significant costs, such as hundreds or even thousands of dollars due to long-distance traveling to seek care or meet regular clinical trial appointments, which in turn can compromise patient safety and limit access to care.⁴¹

Figure 7.

Top Significant Hospital Characteristics and Hospital Total Facility Expenses (EXPTOT) Distribution across U.S. (2022-2023).

The establishment of oncology services within hospitals has been shown to significantly improve patient care quality by increasing timely diagnosis, treatment planning, and access to radical or adjuvant therapies, while reducing emergency admissions and cases of unknown primary cancers.⁴² These findings underscore the critical role oncology departments play in enhancing patient outcomes and ensuring comprehensive cancer care delivery. However, the parallel growth of advanced technologies and novel therapies has led to rising costs of cancer care, disproportionately burdening patients, and healthcare systems.⁴³ To address these inequities with a more cost-friendly strategy, health systems must either bring digital health strategies directly to rural patients through strategies such as tele-oncology, mobile clinics, and hub-and-spoke models, or provide the resources necessary to help patients reach specialized centers, including travel support and physician availability.⁴¹ By enabling real-time communication between rural providers and specialized centers, these digital health interventions can improve access to care while mitigating infrastructure constraints faced by smaller hospitals.

These findings underscore geographic disparities in hospital resources and expenditures, suggesting that while expenditures tend to align with greater staffing, facility capacity, and capitalized costs, notable exceptions may point to differences in efficiency, resource allocation, or care delivery models across CBSAs (Figure 7). Understanding the drivers of hospital costs is especially important given the growing threat of closures in financially vulnerable hospitals. For instance, research shows that rural hospital closures not only reduce access to care but also lead to higher prices at nearby surviving hospitals, thereby affecting affordability.⁴⁴ Moreover, broader funding instability, such as disruptions to community health centers and hospital reimbursement streams, can exacerbate resource shortfalls and hasten institutional exit from the market.⁴⁴ Thus, linking cost‐drivers to closure risk provides actionable insight for policymakers and health system leaders seeking to safeguard access in underserved geographies.

Importantly, the analytical framework used in this study highlights the role of data completeness in supporting such system-level planning. Missing or incomplete reporting within national hospital datasets can obscure disparities in resource distribution and hinder the development of targeted policy interventions. Integrating machine learning assisted imputation into health system analytics can therefore serve as a complementary strategy for improving the usability of incomplete datasets while more comprehensive reporting infrastructures are developed.

It is important to distinguish between descriptive associations identified in this study and causal relationships. The regression analyses presented here identify statistical associations between hospital characteristics and facility expenditures but do not establish causal mechanisms underlying these relationships. Factors such as hospital staffing levels, facility size, and service offerings may co-evolve as part of broader organizational strategies rather than directly causing changes in expenditure levels. Future research using longitudinal designs, quasi-experimental methods, or causal inference frameworks may help clarify the mechanisms linking hospital resource allocation, service availability, and financial outcomes.

From a policy perspective, the findings highlight several opportunities for improving hospital data infrastructure and supporting evidence-based decision-making. First, national hospital reporting systems could benefit from automated data validation and complete monitoring tools integrated into digital reporting platforms.^45,46 Such systems could flag incomplete submissions and provide real-time feedback to reporting institutions, thereby improving data reliability.^45,46 Second, machine learning-assisted imputation frameworks similar to those used in this study may support interim analyses when incomplete reporting cannot be immediately resolved, enabling policymakers and health system leaders to generate more accurate situational awareness from available data. Third, integrating geospatial analytics with more complete datasets can support targeted investments in healthcare infrastructure, particularly in underserved rural regions where service gaps remain substantial.

Limitations and future work

This study has several limitations that should be acknowledged. First, the regression models were developed using a subset of features from the AHA survey, meaning that the results reflected only the features included and may not capture the full range of hospital characteristics available in the dataset. Second, the quality of the AHA survey data itself presents constraints, as issues of missingness and limited data accuracy could have influenced the reliability of the findings. Thirdly, many machine learning methods have been used with multiple imputations to address the missing, but only RF was used in this study. Finally, while multiple imputations were employed to address missing data and enhance the robustness of the analyses, this approach relies on assumptions that may reduce model accuracy, particularly when data are not missing at random and the AHA survey provides limited contextual information to fully characterize the mechanisms of missingness. Future studies could expand the analysis by incorporating a broader range of features from the AHA survey and other relevant data sources to provide a more comprehensive understanding of hospital facility expenditures. In addition, applying advanced statistical and machine learning methods may help capture nonlinear relationships that are not fully addressed by regression models. Causal analysis is needed to explore the nonlinear effects and discover causal relationships among hospital financial indicators.

While random forest imputation demonstrated strong cross-validation performance within this dataset, downstream regression estimates may vary across alternative imputation strategies. External validation using additional survey years and independent hospital datasets is necessary to assess the generalizability and stability of the identified expenditure predictors. Prior machine learning studies in heterogeneous healthcare datasets have similarly highlighted the importance of external validation when applying predictive frameworks across institutional settings and populations.

Although GVIF-based feature selection was used to reduce multicollinearity, correlated hospital characteristics may still influence coefficient stability. Future work should incorporate model-agnostic feature importance approaches, such as permutation-based importance measures or information-theoretic methods, to further validate predictor rankings and assess robustness across analytic frameworks.⁴⁷

Residual diagnostics revealed substantial departure from normality, with excess kurtosis values ranging from 4.78 to 5.24 across imputations (indicating heavy tails relative to a normal distribution), consistent with the presence of outliers. However, with n > 12,000 observations, the ordinary least squares estimator remains consistent and asymptotically normally distributed regardless of the normality assumption such that coefficient estimates and confidence intervals are reliable despite this departure. ⁴⁸

The significant spatial autocorrelation in residuals indicates a violation of the standard regression independence assumption. Although coefficient estimates remain unbiased under spatial dependence, residual autocorrelation may lead to underestimated standard errors and overly narrow confidence intervals. Addressing this issue would require spatial regression models or spatially robust variance estimators that explicitly account for geographic correlation structures. However, implementing and validating such models was beyond the scope of the current analysis, which focused on covariate selection and imputation methodology. Future work should incorporate spatial regression approaches to provide more precise inference and better characterize how regional factors drive hospital expenditure variation.

Lastly, the MNAR sensitivity analysis was restricted to continuous variables with missingness (PLNTA, GFEET, CEAMT, VIDVZ, PRPM). While binary and categorical variables exhibited substantial missingness, formal MNAR sensitivity analysis for categorical data lacks a standardized approach analogous to the delta method for continuous variables. Extension of the MNAR framework to categorical covariates would require specification of missingness mechanisms that fundamentally alter classification probabilities, requiring a choice that is not grounded in a principled sensitivity parameter like delta for continuous data. Future work should employ pattern-mixture or related approaches specifically designed for categorical missingness mechanisms.

Conclusion

In conclusion, this study contributes to the literature by addressing the challenge of missingness in the AHA survey through the application of multiple imputation, thereby enabling more reliable analyses of hospital data. By identifying the key factors associated with total facility hospital facility expenditures, the findings highlight critical drivers of cost variation across healthcare systems. Moreover, the integration of imputed data with geospatial visualization provides a valuable framework for region-specific case studies and policy simulations, offering actionable insights to support decision making and resource allocation among healthcare administrators and policymakers.

Footnotes

Acknowledgements

AI tools were used to assist with proofreading and minor language editing of the manuscript.

ORCID iDs

Jiahui Ma

Elizabeth Johnson

Ethical considerations

2024-1850-EXEMPT.

Author contributions

Jiahui Ma: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft, Writing – review & editing, Visualization, Supervision, Project administration, Validation.

Ian Laga: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft, Writing – review & editing, Visualization, Supervision, Validation.

Elizabeth Johnson: Conceptualization, Methodology, Validation, Investigation, Data curation, Writing – original draft, Writing – review & editing, Supervision, Validation, Funding acquisition, Resources.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Funding for this study made possible with Montana State University College of Nursing THRIVE intramural funding.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The datasets analyzed during the current study are not publicly available because they were obtained under license from the American Hospital Association but may be available from the AHA upon reasonable request and purchase. h. Data Guarantor: Elizabeth Johnson. *

Appendix

Table 6.

Hospital groups identification system specification.

Code	Label	Description
1	Centralized Health System	A delivery system in which the system centrally organizes individual hospital service delivery, physician arrangements, and insurance product development. The number of different products/services that are offered across the system is moderate.
2	Centralized Physician/Insurance Health System	A delivery system with highly centralized physician arrangements and insurance product development. Within this group, hospital services are relatively decentralized with individual hospitals having discretion over the array of services they offer. The number of different products/services that are offered across the system is moderate.
3	Moderately Centralized Health System	A delivery system that is distinguished by the presence of both centralized and decentralized activity for hospital services, physician arrangements, and insurance product development. For example, a system within this group may have centralized care of expensive, high technology services, such as open-heart surgery, but allows individual hospitals to provide an array of other health services based on local needs. The number of different products/services that are offered across the system is moderate.
4	Decentralized Health System	A delivery system with a high degree of decentralization of hospital services, physician arrangements, and insurance product development. Within this group, systems may lack an overarching structure for coordination. Service and product differentiation is high, which may explain why centralization is hard to achieve. In this group, the system may simply serve a role in sharing information and providing administrative support to highly developed local delivery systems centered around hospitals.
5	Independent Hospital System	A delivery system with limited differentiation; hospital services, physician arrangements, and insurance product development. These systems are largely horizontal affiliations of autonomous hospitals.

References

Rupasingha

Cho

. 146 rural hospitals closed or stopped providing inpatient services from 2005 to 2023 in the United States. Economic Research Service, 2025.

Turbow

Lom

Ali

. Where and what separates rural from urban hospital closures? J Hosp Med 2024; 19(9): 812–815. https://doi.org/10.1002/jhm.13438

American Hospital Association . The cost of caring: challenges facing America's hospitals in 2025. [cited 2025 Oct 10]. Available from. https://www.aha.org/costsofcaring (2025).

Andalo

. Health centers face risks as government funding lapses. KFF Health News, 2025. [cited 2025 Oct 10]. Available from. https://kffhealthnews.org/news/article/community-health-centers-government-shutdown-state-cuts-funding-risks/

Kruse

Jeurissen

PPT

. For-profit hospitals out of business? Financial sustainability during the COVID-19 epidemic emergency response. Int J Health Policy Manag 2020; 9(10): 423–428. https://doi.org/10.34172/ijhpm.2020.67

American Hospital Association . The American Hospital Association Annual Survey. [cited 2025 Feb 10]. Available from. https://www.aha.org/american-hospital-association-annual-survey

AHA Data & Insights . The go-to destination for reliable and consistent data about the nation’s hospitals. [cited 2025 Feb 10]. Available from. https://www.ahadata.com/why-aha-data

Cohen

Jones

Heeringa

, et al. Leveraging diverse data sources to identify and describe U.S. health care delivery systems. EGEMS (Wash DC) 2017; 5(3): 9. https://doi.org/10.5334/egems.200

American Hospital Association . AHA comments on CMS RFI on hospital price transparency accuracy and completeness. [cited 2025 Oct 10]. Available from. https://www.aha.org/lettercomment/2025-07-22-aha-comments-cms-rfi-hospital-price-transparency-accuracy-and-completeness (2025).

10.

Richwine

. Progress and ongoing challenges to electronic public health reporting among non-federal acute care hospitals. ASTP Health IT Data Brief. Washington (DC): Office of the Assistant Secretary for Technology Policy 2012; 66: 1–17.

11.

Cummings

. Missing data and multiple imputation. JAMA Pediatr 2013; 167(7): 656–661. https://doi.org/10.1001/jamapediatrics.2013.1329

12.

Mullner

Chung

. The American Hospital Association’s annual survey of hospitals: a critical appraisal. J Consum Mark 2002; 19(7): 614–618. https://doi.org/10.1108/07363760210451438

13.

Rubin

. Inference and missing data. Biometrika 1976; 63(3): 581–592. https://doi.org/10.2307/2335739

14.

Austin

White

Lee

, et al. Missing data in clinical research: a tutorial on multiple imputation. Can J Cardiol 2021; 37(9): 1322–1331. https://doi.org/10.1016/j.cjca.2020.11.010

15.

. The use of multiple imputation to handle missing data in secondary datasets: suggested approaches when missing data results from the survey structure. Inquiry 2022; 59: 00469580221088627. https://doi.org/10.1177/00469580221088627

16.

Junaid

Kiran

Gupta

, et al.

How much missing data is too much to impute for longitudinal health indicators?

Popul Health Metr 2025; 23(1): 2.

17.

Zhang

Lyman

, et al. The HCUP SID imputation project: improving statistical inferences for health disparities research by imputing missing race data. Health Serv Res 2018; 53(3): 1870–1889. https://doi.org/10.1111/1475-6773.12704

18.

Newgard

. The validity of using multiple imputation for missing out-of-hospital data in a state trauma registry. Acad Emerg Med 2006; 13(3): 314–324. https://doi.org/10.1197/j.aem.2005.09.011

19.

Penny

Atkinson

. Approaches for dealing with missing data in health care studies. J Clin Nurs 2012; 21(19–20): 2722–2729. https://doi.org/10.1111/j.1365-2702.2011.03854.x

20.

Woods

Gerasimova

Van Dusen

, et al. Best practices for addressing missing data through multiple imputation. Infant Child Dev 2024; 33(1): e2407. https://doi.org/10.1002/icd.2407

21.

Zhou

Shi

Stein

, et al. Missing data matter: an empirical evaluation of the impacts of missing EHR data in comparative effectiveness research. J Am Med Inform Assoc 2023; 30(7): 1246–1252. https://doi.org/10.1093/jamia/ocad066

22.

Getz

Hubbard

Linn

. Performance of multiple imputation using modern machine learning methods in electronic health records data. Epidemiology 2023; 34(2): 206–215. https://doi.org/10.1097/EDE.0000000000001578

23.

van Buuren

Groothuis-Oudshoorn

. mice: multivariate imputation by chained equations in R. J Stat Softw 2011; 45(3): 1–67. https://doi.org/10.18637/jss.v045.i03

24.

Shah

Bartlett

Carpenter

, et al. Comparison of random forest and parametric imputation models for imputing missing data using MICE: a CALIBER study. Am J Epidemiol 2014; 179(6): 764–774. https://doi.org/10.1093/aje/kwt312

25.

Qiu

. Multiple imputation by chained equations for missing data in UK Biobank. 2022 6th Annual International Conference on Data Science and Business Analytics (ICDSBA). IEEE, 2022, pp. 72–82. https://doi.org/10.1109/ICDSBA57203.2022.00026

26.

Knoedl

. Core based statistical areas. Congressional Research Service, 2024. Report No.: IF12704.

27.

Bao

Bardhan

. Hospital productivity and value in pay-for-performance healthcare programs. Health Syst 2024; 14(2): 131–144. https://doi.org/10.1080/20476965.2024.2421533

28.

Rutter

Park

. Relationship between hospital characteristics and value-based program measure performance: a literature review. West J Nurs Res 2020; 42(12): 1010–1021. https://doi.org/10.1177/0193945920920180

29.

R Core Team . R: a language and environment for statistical computing. : R Foundation for Statistical Computing, 2021. Available from. https://www.R-project.org/

30.

Kahle

Wickham

. ggmap: spatial visualization with ggplot2. R J 2013; 5(1): 144–161. https://doi.org/10.32614/rj-2013-014

31.

Schober

Boer

Schwarte

. Correlation coefficients: appropriate use and interpretation. Anesth Analg 2018; 126(5): 1763–1768. https://doi.org/10.1213/ane.0000000000002864

32.

McHugh

. The SAGE encyclopedia of educational research, measurement, and evaluation. : SAGE Publications, 2018.

33.

Aiken

West

Pitts

. Multiple linear regression. In: Schinka

Velicer

(eds). Handbook of psychology. : Wiley, 2003, pp. 481–507. https://doi.org/10.1002/0471264385.wei0219

34.

Lipkovich

Ratitch

O'Kelly

. Sensitivity to censored‐at‐random assumption in the analysis of time‐to‐event endpoints. Pharm Stat 2016; 15(3): 216–229. https://doi.org/10.1002/pst.1738

35.

Carpenter

Kenward

. Missing data in randomized controlled trials: a practical guide. Birmingham: Technology Assessment Methodology Programme, 2007.

36.

Moran

. Notes on continuous stochastic phenomena. Biometrika 1950; 37(1/2): 17–23.

37.

Johnson

Galatzan

. The communication conundrum: a pilot cross-sectional descriptive examination of family nurse perspectives surrounding patient information exchange during interfacility patient transfers in Montana. Res Theory Nurs Pract 2024; 38(3): 382–405. https://doi.org/10.1891/RTNP-2023-0115

38.

Carvalho

Gavaia

Brito Camacho

. OrthoMortPred: predicting one-year mortality following orthopedic hospitalization. Int J Med Inform 2024; 192: 105657. https://doi.org/10.1016/j.ijmedinf.2024.105657

39.

Carvalho

Gavaia

. Enhancing osteoporosis risk prediction using machine learning: A holistic approach integrating biomarkers and clinical data. Comput Biol Med 2025; 192(Pt B): 110289. https://doi.org/10.1016/j.compbiomed.2025.110289

40.

Probst

Eberth

Crouch

. Structural urbanism contributes to poorer health outcomes for rural America. Health Aff (Millwood) 2019; 38(12): 1976–1984. https://doi.org/10.1377/hlaff.2019.00914

41.

Unger

McAneny

Osarogiagbon

. Cancer in rural America: improving access to clinical trials and quality of oncologic care. CA Cancer J Clin 2025; 75(4): 341–361. https://doi.org/10.3322/caac.70006

42.

Quero

Martín-García

Rivas-Ruiz

, et al. Comparison of oncological care in a hospital center before and after integration of specialized oncology services. Clin Transl Oncol 2025; 27(9): 3811–3818. https://doi.org/10.1007/s12094-025-03902-4

43.

Meropol

Schrag

Smith

, et al. American Society of Clinical Oncology guidance statement: the cost of cancer care. J Clin Oncol 2009; 27(23): 3868–3874. https://doi.org/10.1200/jco.2009.23.1183

44.

Carroll

Chang

. Rural hospital closures led to increased prices at nearby ‘surviving’ hospitals, 2012–22. Health Aff (Millwood) 2025; 44(5): 563–571, [online first]. https://doi.org/10.1377/hlthaff.2024.00700

45.

Bos

. Monitoring the integration of hospital information systems: how it may ensure and improve the quality of data. Med Care Compunetics 3 2006; 121: 176.

46.

Hoeijmakers

Beck

Wouters

MWJM

, et al.

National quality registries: how to improve the quality of data?

J Thorac Dis 2018; 10(Suppl 29): S3490–S3499. https://doi.org/10.21037/jtd.2018.04.146

47.

Carvalho

Gavaia

. Robustness of Osteoporosis Risk Prediction Models with Enhanced Statistical Analyses. Comput Biol Med 2025; 196: 110711. https://doi.org/10.1016/j.compbiomed.2025.110711

48.

Huber

. The behavior of maximum likelihood estimates under nonstandard conditions. Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, 1967.