Estimating the causal effects of chronic disease combinations on 30-day hospital readmissions based on observational Medicaid data

Abstract

Objective

Demonstrate how observational causal inference methods can generate insights into the impact of chronic disease combinations on patients’ 30-day hospital readmissions.

Materials and Methods

Causal effect estimation was used to quantify the impact of each risk factor scenario (ie, chronic disease combination) associated with chronic kidney disease and heart failure (HF) for adult Medicaid beneficiaries with initial hospitalizations in 2 New York State counties. The experimental protocol: (1) created matched risk factor and comparator groups, (2) assessed covariate balance in the matched groups, and (3) estimated causal effects and their statistical significance. Causality lattices summarized the impact of chronic disease comorbidities on readmissions.

Results

Chronic disease combinations were ordered with respect to their causal impact on readmissions. Of disease combinations associated with HF, the combination of HF, coronary artery disease, and tobacco abuse (in that order) had the highest causal effect on readmission rate (+22.3%); of disease combinations associated with chronic kidney disease, the combination of chronic kidney disease, coronary artery disease, and diabetes had the highest effect (+9.5%).

Discussion

Multi-hypothesis causal analysis reveals the effects of chronic disease comorbidities on health outcomes. Understanding these effects will guide the development of health care programs that address unique care needs of different patient subpopulations. Additionally, these insights bring new attention to individuals at high risk for readmission based on chronic disease comorbidities, allowing for more personalized attention and prioritization of care.

Conclusion

Multi-hypothesis causal analysis, a new methodological tool, generates meaningful insights from health care claims data, guiding the design of care and intervention programs.

INTRODUCTION

Treatment and prevention of chronic diseases are a priority for the United States health care system, as an estimated 171 million adults will have one or more chronic conditions by 2030.^1,2 This is particularly true for the economically disadvantaged (Medicaid) population, for whom a high prevalence of multiple chronic conditions, limited access to care, and unstable social support place individuals at increased risk for mortality, adverse drug events, and unnecessary hospitalizations.^3–9 A better understanding of the interactive effects of multiple chronic conditions on hospital readmission risk is needed to effectively improve health outcomes for this vulnerable population and reduce unnecessary readmissions. We offer a new method for understanding the impact of patient complexity on health outcomes, considering both the quantity of chronic comorbidities and the nature of these ailments. We are motivated by previous work that found that readmission rates increase with the number of identified chronic diseases^10–12 and that some chronic disease combinations produce synergistic effects.^8,13,14

A renewed focus on reducing unnecessary, costly hospital readmissions led to the development of models designed to identify individuals at risk for hospital readmission. Predictive models, such as those reviewed by Kansagara et al.¹⁵ and Zhou et al.,¹⁶ have included numerous covariates believed to impact readmission risk, including measures of comorbidities and length of stay (LOS), derived from administrative and clinical data. Despite these varied approaches to representing patient complexity, the inconsistent, and often poor, predictive abilities of these models have limited their practical usefulness. Therefore, there is a need for new modeling approaches that provide innovative, actionable insights for risk stratification purposes, assuring that limited resources are effectively used to prevent readmissions in a Medicaid adult population.

We conducted observational causal analyses to understand the impact of specific chronic disease combinations on 30-day hospital readmission rates, hereafter referred to as readmissions. We used a causal inference approach rather than a widely used regression-based predictive approach as a means to understand if, and how, the presence of chronic disease comorbidities changes readmission risk compared to single chronic conditions. This new evidence will support the development of more personalized interventions based on readmission risk associated with specific comorbidities, rather than providing broad-scope interventions for all individuals with one or more chronic diseases.

Observational causal inference inquiries are concerned with understanding the impact of explanatory variables on an outcome of interest; the employed methods include components for (1) creating matched risk factor (treatment) groups and comparator (control) groups balanced in selected covariates and (2) estimating the impact of specific chronic disease combinations on readmissions. Causal analysis approaches have been applied in numerous fields, including statistics,¹⁷ economics,¹⁸ social sciences,¹⁹ and medicine,^20,21 supporting its application to our problem.

While there are alternative methods to achieving balance, including those that rely on instrumental variables or structural equation modeling,²² we incorporated a matching method for its clear advantages: the ability to identify areas of insufficient covariate overlap in the risk factor and comparator groups and explain why causal estimation may have questionable quality, and the ability to produce multiple diagnostic measures for assessing covariate balance post-matching.^22–25 Further, our robust matching method ensures that reasonable causal effect estimates are generated even when the underlying propensity score model is incorrectly specified.

We aim to demonstrate how observational causal inference methods can enhance our understanding of the impact of chronic disease complexity on patient outcomes and identify new opportunities for developing effective population health management strategies. By identifying chronic disease combinations with the greatest impact on readmissions, hospital systems can strategically target high-risk patients for timely intervention and develop new models of care that more accurately address individual patient care needs. These advancements would improve health outcomes for all patients, particularly the understudied Medicaid population.^6,26

OBJECTIVE

We examined the impact of select chronic disease combinations on readmissions using claims data extracted from the New York State Medicaid Data Warehouse (NYS-MDW). Our goal was to demonstrate the viability of observational causal inference methods to identify chronic disease combinations that present the greatest opportunity for reducing readmissions and how such information can be used to stratify this high-risk patient population.

MATERIALS AND METHODS

Data source

Claims data were extracted from the NYS-MDW for the 119 422 low-income adults, ages 18–64, who resided in Erie or Niagara County and were enrolled for ≥10 months in a Medicaid plan in 2013. Our interest was to understand issues of chronic disease comorbidity in a nonelderly adult population, therefore we excluded children ≤18 years of age and elderly individuals ≥65 years of age from our study population. Data were extracted to support ongoing readmissions and chronic disease research efforts at the University at Buffalo. Data extraction and deidentification protocols were approved by the University at Buffalo Institutional Review Board. For this study population, we observed 11 846 (9.92%) hospital admissions and 11 733 (8.4%) readmissions following an index hospitalization in 2013.

It is not possible to identify disease severity in the NYS-MDW data, therefore we were unable to distinguish individuals with highly severe (eg, end-stage renal disease) from those with less severe (eg, stage 1 or 2 renal disease) disease states. Recognizing that potential readmission risk may be different based upon disease severity, we eliminated 1% (n = 113) of cases with the longest LOS values for the index hospitalization, a potential indication of higher severity disease states. This data trimming^27,28 reduced average LOS in the population from 5.77 to 5.19 days and, importantly, reduced LOS data variation from 9.12 to 6.18 days. Our final dataset contained 11 733 individuals with a readmission within 30 days of discharge from their index admission.

In addition to LOS, patient demographic covariates (gender, age, county of residence) and payer-related covariates (Medicaid plan type [fee-for-service or managed care], Medicare dual eligibility) were extracted from the NYS-MDW. Pregnancy indicators were added to individuals with an identified labor-and-delivery–related Current Procedural Terminology code. These covariates helped improve algorithm performance in creating matched and balanced risk factor and comparator groups. Age was included as a discrete quantity; all other covariates were binary indicators.

Chronic diseases and comorbidities were identified from International Classification of Diseases, Ninth Revision codes using the Clinical Classification Software categorization scheme.^29–31 We assigned indicator flags when one or more chronic conditions were identified for each of the 12 disease categories considered: chronic kidney disease (CKD), heart failure (HF), coronary artery disease (CAD), diabetes mellitus (DM), chronic obstructive pulmonary disease (COPD), behavioral health conditions, substance abuse disorders, asthma, obesity, hypertension (HTN), lipid disorders (LD), and tobacco abuse (smoking). These conditions were selected for their high prevalence in the United States adult population and associated high readmission rates.^5,10,11

We narrowed our focus to 2 primary chronic conditions and select comorbidities of interest when seeking to reduce readmissions. We considered first CKD and 5 select chronic disease comorbidities: CAD, CKD, DM, HTN, and obesity. This selection ensured that we included the primary condition (CKD), chronic conditions (DM, HTN, and LD) with a clear physiological connection to the primary condition, and related chronic conditions (CAD and obesity) that place patients at increased risk but indirectly impact the primary condition.^32,33 Individuals with chronic conditions associated with CKD have high hospitalization rates and elevated readmission risk, and are at risk for developing end-stage renal disease.^30,34

We also considered HF and 6 select chronic disease comorbidities: CAD, HTN, DM, LD, obesity, and smoking). This selection includes the primary condition of interest, conditions (CAD and HTN) known to have a physiological connection to the primary, and additional chronic conditions (DM, LD, obesity, and smoking) known to increase risk for mortality or cardiac events.^35,36 Cardiac-related chronic diseases, associated with high readmission risk, are highly prevalent in the Medicaid population.¹⁰

The outcome of interest was the binary variable indicating whether a patient was readmitted within the first 30 days following an index hospital admission. Supplementary Table S1 provides detailed information about the dataset.

Study design

Our observational causal inference inquiry (see Figure 1) reflects the perceived relationship of the treatment condition (ie, chronic disease combination), selected covariates, and outcome of interest. We restricted our analyses to combinations of 1, 2, and 3 chronic disease combinations associated with each primary condition. We selected covariates, including county of residence, pregnancy status, and age, to maximize use of the NYS-MDW data, which contains limited clinical and administrative data elements. Covariate selection was limited to those factors that could be directly or indirectly obtained from the claims data.

Figure 1.

Causal inference inquiry for the research study. Note: β represents the causal effect of the selected chronic disease combination (T) on 30-day hospital readmissions (Y).

Modeling approach

We applied a single experimental protocol to each risk factor scenario (ie, selected chronic disease combinations) to consistently assess observational causal effects of the specified disease combination on readmission status. That is, our protocol employed algorithms to create 2 groups, one with the target risk factor and one without, with the algorithms making the 2 groups otherwise as similar as possible. Readmission rates of the groups were compared; any difference in readmission rate was inferred to be due to the target risk factor, which could be a single condition or a set of conditions.

This protocol was based upon GenMatch, a genetic or adaptive search algorithm that incorporates both propensity score and multivariate matching (generalized Mahalanobis distance measure) to generate matched risk factor and comparator groups.²⁵ Previous researchers³⁷ proposed integrating these methods to overcome individual limitations. Mahalanobis distance measures, which focus on minimizing scalar distances between group members for each covariate, are impacted by covariate distrubtions, particuarly in sampled data, which limits finding suitably matched groups. Propensity score measures require, first, specification of the score function and, second, iterative respecification of this function until suitable covariate balance is achieved. Nevertheless, improvements in the propensity score function do not always lead to improvements in covariate balance. The GenMatch algorithm iteratively performs balancing, even when the underlying propensity score model is unknown,^18,25 making it an ideal approach for this work.

The GenMatch algorithm includes 3 key functions: (1) creation of matched groups for those individuals with the risk factor (ie, the specified chronic disease combination for each testing scenario) and those with comparator features (ie, without the specified chronic disease combination for each testing scenario); (2) estimation of the observational causal effect for the risk factor condition; and (3) assessment of covariate balance in the matched risk factor and comparator groups. All analyses were conducted using the Matching package in R software.³⁸ As our goal was to quantify the causal impact of having a particular combination of chronic disease diagnoses on readmission for individuals with a specified disease combination, estimates of the average treatment effect for the treated (ATT) were generated. Measures of standardized mean difference, variance ratios, and t-test P-values were used to assess the ability of GenMatch to create matched risk factor and comparator groups for each covariate.

The standardized mean difference measure compares the difference in means in relation to the pooled standard deviation^39,40; values close to zero indicate good balance. The variance ratio measure considers the ratio of variances in matched groups^25,40,41; a value of 1 indicates perfect matching. The t-test diagnostic considers whether the groups differ based on their mean values; P-values greater than the selected significance level (α = 0.05) indicate that the means of the risk factor and comparator groups are not statistically different (implying suitable balance).^25,38 We imposed no criteria on these diagnostics to indicate acceptable balance, but rather used the diagnostic output to compare the ability of the experimental protocol to consistently produce results while assessing how individual covariates balanced under different risk factor scenarios.

RESULTS

CKD and related comorbidities

We first review ATT estimates for individual chronic diseases, or main effects, associated with CKD. CAD, a chronic condition that individuals with CKD are at risk for developing, found in 7% of the study population (Table 1), had the highest estimated causal effect (5.2%). This is interpreted as the average difference in the expected readmission rate for individuals with CAD compared to those without CAD. Other main effects had similar results, except for obesity. The negative ATT estimate for obesity suggests that we would expect a lower readmission rate (1.1% lower on average); however, the nonsignificant P-value indicates that there is insufficient evidence that this diagnosis substantially alters readmission rates from what would be expected for nonobese individuals.

Table 1.

Main causal effects and select interaction effects for diseases associated with the primary condition CKD (n = 11 733)

Chronic disease diagnosis	Risk factor group size (% of N)	Comparator group size	Baseline readmission rate (% of risk factor group)	ATT estimate (%)	Standard error	P-value for ATT estimate
CKD	679 (6)	11 054	91 (13.4)	3.5	0.016	0.031
CAD	790 (7)	10 943	105 (13.3)	5.2	0.015	0.001
DM	2024 (17)	9709	211 (10.4)	2.1	0.009	0.017
Obesity	1515 (13)	10 218	129 (8.5)	−1.1	0.010	0.293
HTN	3427 (29)	8306	363 (10.6)	1.7	0.007	0.016
CAD + HTN	591 (5.0)	11 142	80 (13.5)	4.6	0.018	0.009
CKD + HTN	482 (4.1)	11 251	66 (13.7)	4.6	0.020	0.021
CKD + CAD	166 (1.0)	11 567	26 (15.7)	8.4	0.031	0.007
CKD + CAD + HTN	140 (1.2)	11 593	20 (14.3)	9.3	0.036	0.010

CKD: chronic kidney disease; CAD: coronary artery disease; DM: diabetes mellitus; HTN: hypertension.

Potentially more interesting are the interaction effects resulting from combinations of 2 and 3 chronic disease diagnoses associated with CKD (see Table 1 and Supplementary Table S2). Again, we found that CAD had a clear impact on readmission risk, as the ATT for CKD + CAD (8.4%) was higher than for both individual conditions (CKD = 3.5%, CAD = 5.02%). The further addition of an HTN diagnosis (CKD + CAD + HTN) increased the ATT to 9.3%, suggesting that circulatory-related chronic conditions significantly increase readmission risk, on average, for individuals with CKD-related chronic conditions.

However, some chronic diseases associated with CKD, particularly those that include only circulatory-related chronic disease combinations such as CAD + HTN, had lower ATT estimates (4.6%) than the single diagnoses (CAD = 5.2%, HTN = 1.7%). While HTN should not be interpreted as an ameliorating condition for CAD with respect to readmissions, it may indicate that additional factors, such as medication and disease management protocols not considered in this study, have a beneficial effect on readmission risk.

We introduce the causality lattice, a graphical representation of relationships among chronic disease combinations and causal effect measures. Lattice representations (Figure 2) are useful for evaluating interactions among multiple variables.^42,43 Applied to CKD-related chronic disease combinations, the impact of chronic disease comorbidities becomes clearer. Adding CAD to a CKD diagnosis (CKD + CAD) increases expected readmission rates by 3.5% on average (8.4%–5.2%). Adding HTN (CKD + CAD + HTN) further increases this expected rate by 0.9% (9.3%–8.4%).

Figure 2.

Causality lattice showing the relationship between chronic kidney disease (CKD), coronary artery disease (CAD), and hypertension (HTN) diagnoses with the associated causal effect estimates (ATTs).

We stratified the population of individuals with CKD-related chronic diseases⁴⁴ by comparing individual causal effect estimates to a baseline causal effect associated with having any combination of the 12 chronic disease diagnoses (ATT = 4.5%, P < 0.001, standard error = 0.004). We considered all chronic disease combinations with causal effect estimates within ±10% of the baseline to have a similar level of readmission risk. We selected this tolerance to illustrate the potential application of these results rather than to represent known clinical relationships. Figure 3 classifies CKD-related chronic disease combinations into 3 segments: disease combinations with ATT estimates lower than, equal to, and greater than baseline. Such classification can inform the development of intervention programs targeting significant chronic diseases, and therefore combinations that influence readmissions.

Figure 3.

Comparison of causal effect estimates for disease combinations associated with CKD with the general effect of having any of the 12 chronic diseases considered in this study. Causal effects within ±10% of the effect for the general chronic disease treatment condition represent the baseline causal effect for the chronically ill adult Medicaid population. Shaded circular data points indicate significant results at α = 0.05, while unshaded diamond data points indicate results that are not significant at α = 0.05. CAD: coronary artery disease; CKD: chronic kidney disease; DM: diabetes mellitus; HTN: hypertension; OBS: obesity.

HF and related comorbidities

Considering main effects once again, we found that the primary condition (HF) affected 6% of the population and had the greatest impact on readmissions (ATT = 9.1%) compared to individuals without this diagnosis. Conversely, HTN affected 29% of the study population and had the smallest impact (ATT = 1.7%) compared to individuals without this diagnosis. This suggests that HTN may contribute to long-term disease complexity but has a minimal effect on short-term readmissions. Obesity and LD conditions increase patient risk for cardiac events and mortality, yet nonsignificant ATT estimates indicated insufficient evidence to suggest that these diagnoses substantially alter readmissions.

Examining interaction effects revealed the dramatic impact that chronic disease comorbidities may have on readmissions for individuals with HF-related chronic conditions (see Table 2 and Supplementary Table S3). Consider smoking, a preventable condition, which affects multiple body systems. Alone, smoking had minimal impact on readmissions (ATT = 1.6%), yet as a comorbidity – HF + smoking (ATT = 13.5%) or HF + CAD + smoking (ATT = 22.6%) – the impact is far greater.

Table 2.

Main causal effects and select interaction effects for diseases associated with the primary condition HF (n = 11 733)

Chronic disease diagnosis	Risk factor group size (% of N)	Comparator group size	Baseline readmission rate (% of risk factor group)	ATT estimate (%)	Standard error	P-value for ATT estimate
HF	755 (6.4)	10 978	132 (17.5)	9.1	0.016	0.000
CAD	790 (6.7)	10 943	105 (13.3)	5.2	0.015	0.001
DM	2024 (17.3)	9709	211 (10.4)	2.1	0.009	0.017
Obesity	1515 (12.9)	10 218	129 (8.5)	−1.1	0.010	0.400
HTN	3427 (29.2)	8306	363 (10.6)	1.7	0.007	0.016
LD	1950 (16.6)	9783	162 (8.3)	−1.7	0.009	0.052
Smoking	2792 (23.8)	8941	308 (11.0)	1.6	0.008	0.046
HF + CAD + Smoking	53 (0.5)	11 680	16 (30.2)	22.6	0.069	0.001
HF + CAD	234 (2.0)	11 499	46 (19.7)	12.8	0.030	0.000
HF + smoking	141 (1.2)	11 592	30 (21.3)	13.5	0.039	0.001
CAD + smoking	218 (1.9)	11 515	29 (13.3)	1.8	0.032	0.563

HF: heart failure; CAD: coronary artery disease; DM: diabetes mellitus; HTN: hypertension; LD: lipid disorder.

A causal lattice (Figure 4) clarifies the relationships among HF-related chronic diseases. Individuals with HF + smoking have a 4.4% higher readmission risk (13.5%–9.1%) compared to individuals with HF who do not smoke. The addition of CAD (HF + CAD + smoking) increases readmission risk by 9.1% on average compared to nonsmokers (HF + CAD).

Figure 4.

Causality lattice showing the relationship between heart failure (HF), coronary artery disease (CAD), and smoking (SMO) diagnoses and the associated causal effect estimates (ATTs). Shading in the CAD + SMO node indicates that the causal effect estimate is not significant at both α = 0.05 and α = 0.10.

Chronic disease combinations associated with HF were stratified into 3 segments, using the procedure described earlier. The most compelling opportunities to reduce readmissions (Figure 5) are found in the 19 disease combinations with higher than baseline readmission risk.

Figure 5.

Comparison of causal effect estimates for disease combinations associated with HF with the general effect of having any of the 12 chronic diseases considered in this study. Causal effects within ± 10% of the effect for the general chronic disease risk factor condition represent the baseline causal effect for the chronically ill adult Medicaid population. Shaded circular data points indicate significant results at α = 0.05, while unshaded diamond data points indicate results that are not significant at α = 0.05. HF: heart failure; CAD: coronary artery disease; DM: diabetes mellitus; HTN: hypertension; SMO: smoking; OBS: obesity; LD: lipid disorder.

Comparison of results

We assessed the opportunity to reduce readmissions by comparing results from all analyses. Table 3 compares causal effect estimates for (1) disease combinations with highest baseline readmission rates and greatest importance to the study population, (2) disease combinations with the highest ATT estimates and greatest opportunity for reducing readmissions, and (3) chronic conditions that most commonly occur in disease combinations with nonsignificant ATT estimates and have minimal influence on readmissions.

Table 3.

Comparison of causal effect estimates for disease combinations associated with CKD and HF, for combinations of 1, 2, and 3 chronic diseases

Primary condition	Number of chronic diseases	Greatest concern in populationa (baseline readmission rate)	Greatest opportunity for readmission reductionb(ATT estimate)	Least influential disease combinationsc(number of insignificant P-values)
	1	CKD (13.4%)	CAD (5.2%)	Obesity (1)
CKD	2	CKD + CAD (15.7%)	CKD + CAD (8.4%)	Obesity (4)
	3	CAD + DM + Obesity (17.6%)	CKD + CAD + DM (9.5%)	Obesity (6)
	1	HF (17.5%)	HF (9.1%)	Obesity (1), LD (1)
HF	2	HF + smoking (21.3%)	HF + smoking (13.5%)	Obesity (5), smoking (5)
	3	HF + CAD + smoking (30.2%)	HF + CAD + smoking (22.6%)	Obesity (13)

^aThe identified disease combinations had the highest baseline readmission rates in the population. ^bThe identified diseases had the highest causal effect estimates as measured by the average treatment effect for the treated (ATT). ^cThe identified diseases were included in the greatest number of disease combinations that had nonsignificant P-values. For example, obesity is included in 6 combinations of 3 chronic diseases that had nonsignificant P-values among chronic diseases associated with the primary condition of CKD.

CKD: chronic kidney disease; CAD: coronary artery disease; DM: diabetes mellitus; HF: heart failure.

Using these comparisons, we found that for renal-related diseases, CKD alone had the highest baseline readmission rate (13.4%), yet CAD had the highest ATT estimate (5.2%), and hence the greatest opportunity for reducing readmissions. For HF and related conditions, disease combinations with the highest baseline readmission rates and highest ATT estimates were the same, suggesting that different strategies may be needed to develop effective interventions for this set of conditions. For example, interventions demonstrating the impact of smoking on readmission risk during discharge planning and throughout post-discharge recovery may effectively reduce avoidable readmissions.

Algorithm performance

This work explored the potential of our method, developed a preliminary causal inference model, and assessed the performance of individual covariates in anticipation of future research. Therefore, after designing our test protocol to achieve improved covariate balance in matched groups compared to unmatched groups, we made no further adjustments for imbalances associated with individual scenarios.

Improvements in balance, after matching values closer to ideal values than before matching, were assessed using standardized mean difference, variance ratio, and t-test P-values. Graphical representations of pre- and post-matching results (see examples in Supplementary Figure S1) were compared for each diagnostic and covariate. Quantitatively, we compared balancing results to thresholds adapted from limits suggested in existing literature.^22,39,40 The threshold used for acceptable standardized mean difference, <2.0, is larger than the suggested limits for this diagnostic; however, we believe it is appropriate for an exploratory study using a single test protocol for multiple risk factor scenarios.

We aggregated individual test measures for each covariate to assess balancing across the scenarios (Supplementary Table S4) to identify covariates that were consistently well-balanced (excellent), consistently poorly balanced (poor), and resulted in suitable balance in some risk factor conditions and poor balance in others (good). While most covariates (72%) achieved good or excellent balance for each diagnostic, covariates with poor to good balance would benefit from an adjusted protocol. These imbalances impact results, so that in each scenario the risk factor and comparator groups can be considered matched only in a subset of the included covariates, rather than all possible covariates.

Identifying variables with potentially complex relationships with readmissions will guide future methodological improvements. Redefining the causal inference protocol, including higher-order terms, or associating different or additional chronic diseases with each primary condition would begin to address the limitation of applying a single protocol to a diverse set of scenarios.

DISCUSSION

Using observational causal inference methods, specifically a genetic matching protocol and ATT measures, we quantified the impact of specific chronic disease combinations on readmissions in an adult Medicaid population. We demonstrated how these results can be used to stratify a high-risk population and guide development of effective readmission reduction programs.

Our approach provides clinicians with the opportunity to quantitatively discuss the impact of chronic disease on health outcomes, offering a new patient education approach and focusing preventive measures on ensuring that new, potentially impactful, comorbidities do not develop. Further, even modest reductions in readmissions can result in significant cost savings. Assuming a readmission costs $10 000 on average, a 1% reduction in readmissions could save more than $1 million in the study population.

Limitations

Dataset limitations

We selected covariates for this study from Medicaid claims data that reflect our current work in this area^12,45 but recognize that the lack of clinical information in this data source, notably visibility of disease severity, limits the generalizability of our results. We removed individuals with potentially more severe conditions by trimming the dataset based on the index hospitalization LOS. While this was a suitable approach to demonstrate the viability of our approach, improved methods for identifying and incorporating disease severity and validating our data-trimming method are desirable.

Multiple test significance issues

The 80 unique treatment scenarios in this study were treated as independent hypothesis tests, therefore there is a 98% probability that at least one result was found to be significant (α = 0.05) simply because of the large number of tests. Wanting only to identify influential disease combinations from a large pool of feasible combinations, we made no further corrections. However, restricting analyses to previously identified influential disease combinations in future efforts would require additional strategies to address test significance concerns.

Study population and choice of covariates

This work focused on understanding the impact of chronic disease comorbidities on patient health outcomes in the understudied low-income adult (18–64 years of age) Medicaid population. Generalizing our results to different populations, such as the elderly (65 years and older) Medicare population or the commercially insured adult (18–64 years) population, is challenging due to differences in chronic disease comorbidities and prevalence, access to care, and utilization patterns.⁴⁶

Further, we only considered factors that could be extracted from the Medicaid claims data, therefore it is possible that important confounding factors, such as social determinants of health, were omitted because they were not contained in these data. Also, we focused on only 2 primary conditions and did not consider the impact of other conditions, such as behavioral health and substance abuse issues, prevalent in Medicaid populations that may have additional effects on readmissions.

We managed the effect of unknown and known confounding variables by: (1) restricting the dataset to create an initial level of homogeneity and (2) using matching methods to create risk factor and comparator groups that differed only in the risk factor conditions. When adapting this approach to new data elements or problems, careful consideration of the selection covariates is needed, as existing observational causal inference methods may become ineffective due to the complexity of the underlying matching optimization problem if too many covariates are included.

Finally, using only 1 year of claims data, there were many test scenarios, or disease combinations, with small risk factor group sizes. Although this was sufficient for demonstrating new opportunities for causal inference research, it is possible that conditions that were insufficiently identified in this population led to inconclusive results. Additional studies with larger populations would clarify the impact of uncommon disease combinations on readmissions.

Contributions

There is a clear connection between chronic disease and readmissions,^{3,4,7,11,14,47,48} yet the current incomplete understanding of this relationship in the adult Medicaid population hampers the development of more effective interventions for these individuals. We demonstrate that causal inference methods can generate new insights regarding the impact of chronic disease on readmissions, improving upon current approaches that examine only the number^11,13,48 or type of chronic disease comorbidities.^8,14

Such efforts would benefit from expanding causal inference models to include additional primary conditions of interest, such as COPD, a highly prevalent chronic condition with high readmission rates and related chronic comorbidities. Further, including chronic diseases not directly associated with either CKD or HF, such as COPD, is important, as their limited role in the models presented herein yielded poor to good balancing in each risk factor scenario, yet these comorbidities may significantly impact patient outcomes. Examining the interactions of diseases associated with different primary conditions of interest and integrating this evidence into existing or new clinical decision support tools would support the development of readmission reduction interventions designed to specifically address the needs of high-risk patients with chronic disease comorbidities. Finally, the continued development of new, fast matching methods designed for big data will make multi-hypothesis causal inference particularly practical.⁴⁹

CONCLUSION

This exploratory work demonstrates how observational causal inference inquiries can identify specific disease combinations that affect readmissions. Causal effect results were used to stratify a high-risk, low-income population with chronic diseases to prioritize intervention and follow-up care. The presented methodology and results form the basis for designing new initiatives focused on reducing readmission risk for individuals with multiple chronic conditions. We have demonstrated the viability of this method and the importance of considering both the quantity and types of chronic diseases for population management and readmission reduction efforts. Applying our approach to new datasets and similarly complex health care analyses, we can continue to improve the effectiveness of health care interventions and reduce the cost of care.

COMPETING INTERETS

The authors have no competing interests to declare.

FUNDING

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sector.

CONTRIBUTORS

SC, LL, SH, and AN conceived the study and its design. SC and AN conducted the research, the primary analysis, and initial drafting of the paper. LL, SH, and AN contributed to the analysis and drafting of the paper. SC, AN, LL, and SH approved the final manuscript. SC is the corresponding author.

SUPPLEMENTARY MATERIAL

Supplementary material is available at Journal of the American Medical Informatics Association online.