Ambulatory Care Fragmentation and Total Healthcare Costs

Abstract

Background:

The magnitude of the relationship between ambulatory care fragmentation and subsequent total healthcare costs is unclear.

Objective:

To determine the association between ambulatory care fragmentation and total healthcare costs.

Research Design:

Longitudinal analysis of 15 years of data (2004 – 2018) from the national Reasons for Geographic and Racial Differences in Stroke (REGARDS) study, linked to Medicare fee-for-service claims.

Subjects:

13,680 Medicare beneficiaries ≥65 years old.

Measures:

We measured ambulatory care fragmentation in each calendar year, defining high fragmentation as a reversed Bice-Boxerman Index ≥0.85 and low as <0.85. We used generalized linear models to determine the association between ambulatory care fragmentation in one year and total Medicare expenditures (costs) in the following year, adjusting for baseline demographic and clinical characteristics, a time-varying co-morbidity index, and accounting for geographic variation in reimbursement and inflation.

Results:

The average participant was 70.9 years old; approximately half (53%) were women. One-fourth (26%) of participants had high fragmentation in the first year of observation. Those participants had a median of 9 visits to 6 providers, with the most frequently seen provider accounting for 29% of visits. By contrast, participants with low fragmentation had a median of 8 visits to 3 providers, with the most frequently seen provider accounting for 50% of visits. High fragmentation was associated with $1,085 more in total adjusted costs per person per year (95% CI $713 to $1,457) than low fragmentation.

Conclusions:

Highly fragmented ambulatory care in one year is independently associated with higher total costs the following year.

INTRODUCTION

Ambulatory care in the U.S. is highly fragmented.¹ Patients routinely see multiple ambulatory care providers without a dominant provider.^2,3 Such patterns of care may be clinically appropriate, but they create challenges; for example, providers do not consistently communicate with each other regarding the patients they have in common.⁴ As a result, highly fragmented care can lead to gaps in clinical information.⁵ This, in turn, can independently increase the risk of drug-drug interactions,⁶ repeat tests,⁷ excess procedures,⁸ emergency department visits,^9–11 and hospitalizations.^2,12 Based on this, one would expect more ambulatory care fragmentation to be associated with higher total healthcare costs. However, empirical evidence of the precise relationship between ambulatory care fragmentation and total healthcare costs in the U.S. has been limited.

Some experts have estimated that “when patients fall through the slats in fragmented care,” this costs American society on the order of $25 billion to $45 billion each year.¹³ Other experts place the estimate even higher, on the order of $158 billion to $226 billion annually.¹⁴ While these estimates underscore the expected magnitude of this problem, they are based on expert opinion, rather than quantitative evidence. One study that was quantitative found that more ambulatory care fragmentation was associated with higher total costs, but that study may have been limited because it relied exclusively on claims (which lack clinical detail) and because it was restricted to patients with congestive heart failure, chronic obstructive pulmonary disease, or diabetes.¹⁵ Another study attempted to measure the association between ambulatory fragmentation and costs but substituted the average fragmentation score of other patients (who had the same primary care provider) for an index patient’s fragmentation score,¹⁶ leaving uncertainty about the actual relationship between ambulatory care fragmentation and total healthcare costs.

We sought to determine the association between ambulatory care fragmentation and total healthcare costs, using data from a national cohort study with detailed clinical information from primary data collection, linked with Medicare claims. Elucidating the relationship between fragmentation and costs is important because fragmentation is potentially modifiable.⁵ Moreover, measuring the magnitude of the association between fragmentation and costs can reveal the potential amount of cost savings that could be achieved if fragmentation were decreased.

METHODS

Overview.

We conducted a longitudinal cohort study, using data from the ongoing REasons for Geographic and Racial Differences in Stroke (REGARDS) study. Institutional review boards of the participating institutions approved the protocol. All participants provided written informed consent.

Study population.

Between 2003 and 2007, the REGARDS study enrolled 30,329 community-dwelling, English-speaking adults in the U.S. who were ≥45 years old.¹⁷ The study involved oversampling of adults living in the Southeastern U.S., balanced sampling of White and Black individuals, and balanced sampling of men and women.¹⁷ Participants were not selected based on any clinical risk factors for stroke (besides race and sex).

Data sources.

We used data collected by the REGARDS study at baseline, which involved a computer-assisted telephone interview and an in-home study visit, including a physical examination, blood and urine tests, an electrocardiogram, and a medication inventory.¹⁷ We also used de-duplicated final Medicare fee-for-service claims, starting with the first full calendar year for which claims were available (2004). The REGARDS participants with linked Medicare fee-for-service claims have been shown to be representative of the national population of Medicare fee-for-service beneficiaries.¹⁸ The Medicare claims (Inpatient, Outpatient, Carrier, Skilled Nursing, Home Health, and Durable Medical Equipment) included the total amount Medicare paid in dollars for each service. These files also included the total amount paid by the beneficiary (in terms of deductibles, co-insurance, and co-pays) and the total amount paid by the primary payer if not Medicare (e.g., Medicaid). We did not include Part D claims.

Variables. Healthcare fragmentation.

To measure fragmentation of ambulatory care, we first identified ambulatory visits in Medicare claims, using a modified version of the definition by the National Commission for Quality Assurance (NQCA), which we have used previously.^7,19 Using Clinical Procedure Terminology (CPT) codes, this modified definition restricts ambulatory visits to in-person, evaluation-and-management visits for adults in an office or clinic setting.^7,19 The NCQA definition does not include emergency department visits.

We calculated fragmentation of ambulatory care using the Bice-Boxerman Index (BBI).²⁰ This Index incorporates the total number of visits, the total number of providers, and the distribution of visits across those providers to yield a single score, ranging from 0 (each visit with a different provider) to 1 (all visits with the same provider). We reversed the Index (rBBI), calculating 1 minus BBI, so that higher scores reflect more fragmentation.¹⁰ We calculated the rBBI for each participant for each calendar year in which the participant had ≥4 ambulatory visits (as calculating fragmentation with <4 visits can lead to unreliable estimates).¹² That is, we recalculated a participant’s fragmentation score for each eligible calendar year. Within each year, we defined high fragmentation as an rBBI score ≥0.85 and low fragmentation as rBBI < 0.85, as this cutoff has been found to be an independent predictor of hospitalization.²

Total healthcare costs.

We tallied the total expenditures for each participant for each calendar year (for Inpatient, Outpatient, Carrier, Skilled Nursing, Home Health, and Durable Medical Equipment expenditures, including Medicare, beneficiary, and other primary payer payments), and we refer to these total expenditures as total costs.²¹ We standardized total costs for geographic variation in Medicare reimbursement, using methods recommended by the Centers for Medicare & Medicaid Services.^22–24 We also standardized total costs to 2018 dollars to account for inflation.

Potential confounders.

We used the following variables as potential confounders, all of which were collected by REGARDS at baseline: demographic characteristics (age, sex, race, marital status, education, annual household income, geographic region, and residence in an urban area); medical conditions (hypertension, dyslipidemia, diabetes, myocardial infarction, atrial fibrillation, and stroke); medications (number of medications, anti-hypertensive medication, insulin use, and statin use); health behaviors (smoking, alcohol use, exercise); psychosocial variables (being a family caregiver, social support, depressive symptoms, self-rated general health, self-rated physical health, and self-rated mental health); and physiologic variables (body mass index, systolic blood pressure, total cholesterol, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, glucose, estimated glomerular filtration rate, urinary albumin-to-creatinine ratio, and c-reactive protein). The definitions of these variables are shown in Appendix 1, Supplemental Digital Content 1, http://links.lww.com/MLR/C798.

In order to carefully account for the burden of illness, we calculated a time-varying Charlson co-morbidity index, adapted for claims, for each participant.^25–27 The Charlson co-morbidity index includes 19 chronic conditions (Appendix 2, Supplemental Digital Content 1, http://links.lww.com/MLR/C798).²⁵ Points are assigned for each condition a participant has, with the points weighted for the severity of the condition and then tallied.²⁵ This index has been extensively validated and widely used.²⁸ The rationale for incorporating the Charlson co-morbidity index into this analysis was two-fold. First, the index includes diseases that were not a focus of the REGARDS study, such as cancer and liver disease.²⁵ Second, using the Charlson co-morbidity index allowed us to update the burden of co-morbidities each year, which is not possible using REGARDS data. We calculated two versions of the Charlson co-morbidity index: one version using each year of claims data separately (that is, counting only those diseases that appeared in claims for a given year) and a second version using each year of claims data to create a cumulative score (that is, assuming that diseases that appeared in earlier claims persisted, even if they were not explicitly mentioned in subsequent claims). We designed this cumulative score because nearly all diseases included in the Charlson co-morbidity index are chronic conditions that do not necessarily resolve. We used the first version of the index (with each year of claims data considered separately) for descriptive statistics only; we used the second, cumulative version for descriptive statistics and for modeling.

Data analysis. Derivation of the study cohort.

We included REGARDS participants who were ≥65 years old at any time during the study period and whose cohort data were linked to Medicare claims. We excluded participants who qualified for Medicare on the basis of end-stage renal disease, as those beneficiaries have distinct ambulatory utilization patterns.²⁹ We also excluded participants with <24 months of continuous Medicare fee-for-service coverage, including those with non-fee-for-service products (i.e., Medicare Advantage), as we did not have access to Medicare Advantage claims. We excluded participants who had fewer than 4 ambulatory visits in every year of observation, as fragmentation could not be reliably calculated.¹² We also excluded participants for whom we could not calculate geographically standardized total costs in any year due to missing geographic data.

Descriptive statistics.

We used descriptive statistics to characterize patterns of ambulatory care in the first year of observation, calculating the median and interquartile range for the number of ambulatory visits, number of ambulatory providers, proportion of visits with the most frequently seen provider, and fragmentation score. We compared the characteristics of participants with high vs. low fragmentation in the first year of observation, using Wilcoxon rank sum tests, t tests, and chi-squared tests.

Association between ambulatory fragmentation and total costs.

For all measures of association between ambulatory fragmentation and total costs, we considered fragmentation (as the exposure) in one calendar year and total costs (as the outcome) in the following calendar year. This method of using the exposure in one year and the outcome in the following year is based on the methods of the case mix literature, which has validated comorbidity indices by using the presence of chronic conditions in one year to predict costs in the following year.³⁰ Measuring fragmentation in one year and costs in the following year also ensured that the ambulatory visits that counted toward the fragmentation score preceded in time the utilization that counted toward total costs. We used Wilcoxon rank sum tests for unadjusted tests of association between fragmentation in one year and total costs in the subsequent year.

To enable adjustment for potential confounders, we used generalized linear models (GLM) with a Poisson distribution and log-link function, incorporating robust standard errors to account for multiple observations per participant.³¹ We conducted separate models for each pair of years (i.e., exposure year + outcome year) and also conducted one model that combined all years of data. We included in multivariable models those co-variates that had a bivariate association with fragmentation (bivariate p < 0.10). If two variables had bivariate p-values <0.10 but one was embedded in the definition of another (e.g., systolic blood pressure was embedded in the definition of hypertension), we only adjusted for the broader variable. We adjusted for calendar year and for the cumulative Charlson co-morbidity index, as measured through and including the exposure year; this was a more conservative approach than adjusting for a non-cumulative co-morbidity index, which may have underestimated the full extent of the disease burden. We did not adjust for self-reported health, because this variable was strongly correlated with the Charlson co-morbidity index (Spearman’s correlation coefficient > 0.001) it is only available at baseline, whereas Charlson comorbidities are available in each year. Missing values for baseline co-variates were imputed using stochastic regression, a single imputation method that adds a residual error term to each imputed predicted value, thereby preserving uncertainty while reducing bias.³² We chose this approach instead of multiple imputation due to the complexity of repeated measures in our dataset. The most frequently missing variable was income (missing for 13%), and the next most frequently missing variable was C-reactive protein (missing for 6%). We used the GLM models to estimate absolute adjusted costs (that is, the adjusted mean total costs in each fragmentation group as well as the difference in adjusted mean total costs between groups). The same models also generated relative results expressed as adjusted cost ratio, which indicate the relative increase in cost among those with high fragmentation.^33,34

Sensitivity analysis.

We conducted one sensitivity analysis using beneficiary-level fixed effects, to determine the robustness of our results. Because fixed effects inherently adjust for observed and unobserved baseline characteristics, the sensitivity analysis needed to adjust only for the time-varying co-morbidity index and age. We conducted a separate sensitivity analysis, using propensity scores to express the likelihood of having highly fragmented care, using the same co-variates as in the fully adjusted model described above. We then assessed the association between fragmentation and total costs, while using inverse probability weighting to account for the propensity to have highly fragmented care.

Statistical software.

Analyses were conducted with SAS (version 9.4; SAS Institute, Cary, NC) and Stata (version 14; StataCorp, College Station, TX). P-values <0.05 were considered statistically significant.

RESULTS

Sample characteristics and ambulatory utilization.

We identified 13,680 participants who met our inclusion criteria (Appendix 3, Supplemental Digital Content 1, http://links.lww.com/MLR/C798). The average age of the sample was 70.9 years (SD 5.9). Approximately half (53%) were women. Nearly one-third (32%) were Black. Approximately half (48%) rated their general health as excellent or very good. At baseline, the average participant had a Charlson co-morbidity score of 1.1 (SD 1.4). By the end of the study period, the average participant had a cumulative Charlson co-morbidity score of 3.4 (SD 2.7) (Appendix 4, Supplemental Digital Content 1, http://links.lww.com/MLR/C798). Additional participant characteristics are shown in Table 1.

Table 1.

Characteristics of study sample, overall and stratified by fragmentation score in the first year of observation

Characteristic*	Overall (N = 13,680)	Low fragmentation during first year of observation (N = 10,072)	High fragmentation during first year of observation (N = 3,608)	p-value
Demographic characteristics
Age, years, mean (SD)†	70.9 (5.9)	71.0 (5.9)	70.5 (5.7)	<0.001
Sex, female (%)	53.1	52.9	53.5	0.53
Race
White (%)	67.7	65.3	74.4	<0.001
Black (%)	32.3	34.7	25.6
Marital status, married (%)	59.9	58.9	62.7	<0.001
Education, high school diploma or more (%)	87.6	86.6	90.3	<0.001
Annual household income, ≥$35,000 (%)	50.7	48.6	56.7	<0.001
Geographic region (%)
Stroke Belt‡	36.3	35.6	38.2	<0.01
Stroke Buckle§	23.3	23.8	21.7
Neither Stroke Belt nor Stroke Buckle	40.5	40.6	40.1
Residence in urban area (%)	75.4	75.5	75.3	0.84
Medical conditionsǁ
Hypertension (%)	62.1	63.7	57.7	<0.001
Dyslipidemia (%)	63.1	63.6	61.8	0.06
Diabetes (%)	22.4	23.2	20.0	<0.001
Myocardial infarction (%)	14.4	14.4	14.3	0.85
Atrial fibrillation (%)	9.8	9.7	10.2	0.35
Stroke (%)	6.8	7.1	6.0	0.03
Medications
Number of medications, median (25th, 75th percentiles)	6 (3, 9)	6 (3, 8)	6 (3, 9)	<0.001
Anti-hypertensive medication use (%)	56.2	57.8	51.8	<0.001
Insulin use (%)	5.8	5.7	5.9	0.74
Statin use (%)	36.4	36.2	37.0	0.41
Health behaviors
Current smoker (%)	11.1	11.7	9.7	<0.01
Alcohol use, none (%)	63.5	64.7	60.2	<0.001
Exercise frequency, 0 times per week (%)	34.6	34.7	34.4	0.93
Psychosocial variables
Cares for a family member with a chronic illness or disability (%)	11.8	11.8	11.8	0.99
Lack of social support (%)	4.0	3.9	4.3	0.24
Depressive symptoms (%)	6.3	6.4	6.2	0.74
Self-rated general health (SF-1) (%)
Excellent	16.6	16.1	18.3	0.03
Very good	31,1	31.1	31.1
Good	35.0	35.4	33.8
Fair	14.1	14.3	13.5
Poor	3.2	3.2	3.3
Physical Component Score from the SF-12, mean (SD)	46.2 (10.5)	46.2 (10.4)	46.1 (10.8)	0.59
Mental Component Score from SF-12, mean (SD)	54.9 (7.9)	54.9 (7.9)	54.7 (7.9)	0.27
Physiological variables
Body mass index, kg/m2, mean (SD)	29.0 (5.8)	29.1 (5.9)	28.7 (5.7)	<0.01
Systolic blood pressure, mm Hg, mean (SD)	128.6 (16.3)	129.1 (16.3)	127.3 (16.2)	<0.001
Total cholesterol, mg/dL, mean (SD)	190.1 (39.7)	190.5 (40.1)	189.0 (38.5)	0.05
Low-density lipoprotein cholesterol, mg/dL, mean (SD)	111.6 (34.1)	112.1 (34.5)	110.1 (32.9)	<0.01
High-density lipoprotein cholesterol, mg/dL, mean (SD)	51.5 (16.2)	51.3 (16.1)	52.0 (16.3)	0.03
Glucose, mg/dL, mean (SD)	104.2 (34.4)	104.8 (34.9)	102.5 (32.7)	<0.001
Estimated glomerular filtration rate, mL/min/1.73 m2, median (25th, 75th percentiles)	83.7 (69.8, 94.1)	83.8 (69.8, 94.4)	83.7 (70.3, 93.8)	0.92
Urinary albumin-to-creatinine ratio, mg/g, median (25th, 75th percentiles)	7.8 (4.9, 16.9)	8.0 (4.9, 17.6)	7.4 (4.7, 15.0)	<0.001
C-reactive protein, mg/L, median (25th, 75th percentiles)	2.1 (1.0, 4.8)	2.2 (1.0, 4.8)	2.0 (0.9, 4.5)	0.001
Composite co-morbidity index
Charlson comorbidity index, adapted for claims, mean (SD)	1.1 (1.4)	1.1 (1.3)	1.2 (1.6)	<0.001

^*Missing data: education (N = 4), income (N = 1726), hypertension (N = 38), dyslipidemia (N = 457), diabetes (N = 480), atrial fibrillation (N = 278), stroke (N = 51), number of medications (N = 11), current smoker (N = 46), alcohol use (N = 245), body mass index (N = 72), urinary albumin-to-creatinine ratio (N = 600), and C-reactive protein (N = 817).

^†Age at entry into this analysis (that is, age at baseline or when aged into Medicare and became eligible for inclusion).

^‡Stroke Belt = North Carolina, South Carolina, Georgia, Tennessee, Mississippi, Alabama, Louisiana, and Arkansas (except for 153 coastal counties in North Carolina, South Carolina, and Georgia that constitute the Stroke Buckle).

^§Stroke Buckle = 153 coastal counties in North Carolina, South Carolina, and Georgia.

^ǁSee methods section for detailed definitions of variables.

Key: High fragmentation indicates reversed Bice-Boxerman Index (rBBI) ≥0.85; low fragmentation indicates rBBI<0.85. SD = standard deviation. SF = short-form survey. Percentages may not sum to 100 due to rounding.

One-fourth (26.3%) of participants had high fragmentation in the first year of observation. Those participants had a median of 9 visits to 6 providers, with the most frequently seen provider accounting for 29% of the visits, yielding a median fragmentation score of 0.90. By contrast, participants with low fragmentation in the first year of observation had a median of 8 visits to 3 providers, with the most frequently seen provider accounting for 50% of visits, yielding a median fragmentation (rBBI) score of 0.70 (Table 2).

Table 2.

Ambulatory utilization in the first year of observation, overall and stratified by fragmentation score^*

	Overall (N = 13,680)	Low fragmentation (N = 10,072)	High fragmentation (N = 3,608)	p-value
	Median (IQR)
Ambulatory visits	8 (6, 13)	8 (5, 12)	9 (6, 15)	<0.001
Ambulatory providers	4 (3, 6)	3 (3, 5)	6 (5, 8)	<0.001
Proportion of visits with the most frequently seen provider	0.47 (0.33, 0.60)	0.50 (0.43, 0.67)	0.29 (0.25, 0.33)	<0.001
Fragmentation score	0.78 (0.61, 0.86)	0.70 (0.54, 0.80)	0.90 (0.87, 0.93)	<0.001

^*Low fragmentation is defined as a reversed Bice-Boxerman (rBBI) score of <0.85. High fragmentation is defined as an rBBI score ≥0.85.

Key: IQR = interquartile range.

Participants with high fragmentation were more likely to be younger (p < 0.001), more likely to be White (p < 0.001), and more likely to perceive that they are in excellent health (p = 0.03), but also more likely to have a higher Charlson comorbidity score (p < 0.001). Additional differences in characteristics between participants with high vs. low fragmentation are shown in Table 1.

Ambulatory care fragmentation and total healthcare costs.

When we conducted separate models for each pair of years, we found that unadjusted total healthcare costs (standardized for geographic variation and inflation) were consistently higher for those with high vs. low fragmentation (Table 3). Participants with high fragmentation had total unadjusted costs that were between $629 and $2,336 more per person per year, compared to participants with low fragmentation; these comparisons were statistically significant for 11 of the 14 outcome years. When we adjusted for co-variates, costs for participants with high fragmentation were between $242 and $1,765 more per person per year than costs for participants with low fragmentation, with comparisons statistically significant for 5 of 14 outcome years (Figure).

Table 3.

Unadjusted mean total Medicare costs per participant by calendar year, overall and stratified by fragmentation score^*

Outcome year	N	Medicare costs in dollars				p-value
		Overall, mean (sd)	Low fragmentation, mean (sd)	High fragmentation, mean (sd)	Difference in mean cost (high fragmentation minus low)
2005	6509	10,727 (21,147)	10,460 (21,117)	11,516 (21,220)	+1,056	0.08
2006	6668	12,195 (21,900)	11,735 (21,052)	13,440 (24,007)	+1,705	<0.01
2007	6630	12,636 (23,588)	12,165 (23,247)	13,788 (24,372)	+1,623	0.01
2008	6666	13,080 (25,091)	12,614 (25,262)	14,215 (24,638)	+1,601	0.02
2009	6865	13,660 (25,459)	13,121 (25,050)	14,877 (26,324)	+1,756	<0.01
2010	6932	13,837 (27,956)	13,362 (29,226)	14,881 (24,908)	+1,519	0.04
2011	7184	13,610 (27,854)	13,154 (27,684)	14,564 (28,188)	+1,410	0.05
2012	7288	13,248 (25,150)	12,460 (24,871)	14,796 (25,623)	+2,336	<0.001
2013	7254	12,447 (22,731)	12,217 (22,838)	12,846 (22,542)	+629	0.26
2014	7015	12,824 (24,181)	12,269 (23,448)	13,710 (25,285)	+1,441	0.02
2015	6611	12,730 (23,135)	12,091 (22,341)	13,655 (24,212)	+1,564	<0.01
2016	6358	13,352 (24,641)	12,700 (24,596)	14,209 (24,680)	+1,509	0.02
2017	5941	13,323 (23,420)	13,002 (23,852)	13,708 (22,888)	+706	0.25
2018	5506	14,268 (26,934)	13,291 (22,662)	15,359 (30,984)	+2,068	<0.01

^*Fragmentation scores, as a potential predictor of total Medicare costs, were calculated in the year prior to the outcome year listed. Fragmentation scores are based on the reversed Bice-Boxerman Index, with scores <0.85 indicating low fragmentation and scores ≥0.85 indicating high fragmentation. Total Medicare costs were standardized to 2018 dollars and standardized for geographic variability in Medicare reimbursement but are otherwise unadjusted in this table.

Figure. Average adjusted total cost per Medicare beneficiary for 2005–2018, stratified by high vs. low fragmentation of ambulatory care*.

*p < 0.05. In this analysis, fragmentation of ambulatory care is the exposure, and total cost of care is the outcome. Fragmentation of ambulatory care is measured in the year preceding the year in which total costs (the outcome) is measured. Fragmentation scores are based on the reversed Bice-Boxerman Index, with scores <0.85 indicating low fragmentation and scores ≥0.85 indicating high fragmentation. Results were generated using generalized linear models with a Poisson distribution using a log-link function and robust standard errors. The fully adjusted model accounts for geographic variation in Medicare reimbursement and also adjusts for age, sex, race, marital status, education, income, region, hypertension, dyslipidemia, medication count, smoking, alcohol use, body mass index, c-reactive protein, and cumulative Charlson co-morbidity index. The fully adjusted model includes stochastic imputation for missing co-variates at baseline.

When we used all years of data, high fragmentation was associated with 12% higher unadjusted total healthcare costs (cost ratio 1.12; 95% confidence interval [CI] 1.09, 1.15) (Table 4). After adjustment for baseline characteristics, cumulative co-morbidity, and calendar year, the association persisted, with a cost ratio of 1.09 (95% CI 1.06, 1.12). Having high fragmentation was associated with $1,085 more in adjusted total costs per person per year (95% CI $713 to $1,457), compared to having low fragmentation.

Table 4.

Association between high ambulatory care fragmentation and total healthcare costs^*

	Low fragmentation	High fragmentation
	Cost Ratio (95% confidence interval)		p-value
Crude	[ref]	1.12 (1.09, 1.15)	<0.001
Fully adjusted	[ref]	1.08 (1.05, 1.11)	<0.001
Fully adjusted with imputation	[ref]	1.09 (1.06, 1.12)	<0.001
	Predicted mean cost in dollars (95% confidence interval)		Predicted difference in mean cost
Crude	12,465 (12,186, 12,744)	13,953 (13,582, 14,324)	+1,488 (+1,089, +1,887)
Fully adjusted	12,426 (12,161, 12,690)	13,435 (13,098, 13,771)	+1,009 (+627, +1,392)
Fully adjusted with imputation	12,599 (12,341, 12,856)	13,684 (13,357, 14,011)	+1,085 (+713, +1,457)

^*These results were generated by considering the association between ambulatory care fragmentation in one year and total healthcare costs in the following year. The models include data from 2004 – 2018 and allow participants to contribute data for multiple years. The results were generated using generalized linear models with a Poisson distribution using a log-link function. Robust standard errors are used to account for multiple observations per participant. The crude model is standardized for geographic variation in Medicare reimbursement and adjusts for outcome year but is otherwise unadjusted. The fully adjusted model accounts for geographic variation in Medicare reimbursement and also adjusts for age, sex, race, marital status, education, income, region, hypertension, dyslipidemia, medication count, smoking, alcohol use, body mass index, c-reactive protein, cumulative Charlson co-morbidity index, and outcome year. The fully adjusted model with imputation adds stochastic imputation for missing co-variates at baseline.

The sensitivity analysis using beneficiary-level fixed effects showed an adjusted cost ratio of 1.02 (95% CI 1.00, 1.05; p = 0.09). This suggests that accounting for unobserved sources of individual-level heterogeneity reduces the estimated association between fragmentation and healthcare costs somewhat, though the positive coefficient is consistent with the direction of the main analyses.

The sensitivity analysis using propensity scores yielded results that were the same as the base case analysis: an adjusted cost ratio of 1.09 (95% CI 1.06, 1.12).

DISCUSSION

In this national study of 13,680 Medicare beneficiaries 65 years and older, having highly fragmented ambulatory care in one year was associated with $1,085 more in total adjusted costs per person the following year (95% CI $713 to $1,457), compared to having low fragmentation. This was equivalent to a 9% increase in adjusted total costs for high vs. low fragmentation (95% CI 6% to 12%). This finding was consistent across the 15 years of our study (2004–2018), accounting for inflation and geographic variation in Medicare reimbursement.

The magnitude of our finding cannot easily be compared to other quantitative studies due to differences in methodology. For example, another study measured the difference in the cost of an episode of care for every 0.1-point difference in the fragmentation score (rather than the difference in the total cost of care for high vs. low fragmentation).¹⁵ That study found for every 0.1-point increase in fragmentation score, the cost of an episode of care for congestive heart failure, chronic obstructive pulmonary disease, or diabetes increased by 4.7% to 6.3%, depending on the condition.¹⁵ The direction of our finding is consistent with this.

In post-hoc analyses, we found that the largest category of cost in each year was inpatient costs, which accounted for an average of 36.5% of total costs. By contrast, ambulatory visits accounted for an average of 6.3% of total costs. Although this study is not granular enough to determine the exact mechanism by which fragmented ambulatory care may be associated with total costs, this study complements other work that has found that fragmented ambulatory care is independently associated with an increased risk of hospitalization.^2,12,15

To illuminate the implications of this study, it is important to discuss the possibility of self-selection bias and consider evidence that may make this possibility less likely. Self-selection bias would occur if Individuals who are unmeasurably sicker and more prone to higher healthcare costs self-select into more fragmented care. However, we previously conducted a study that asked patients and providers why some patients receive care from many providers and others do not.⁵ We found that there were more than 40 different reasons why some patients receive care from many providers, only a few of which were related to medical need.⁵ Examples of reasons not related to medical need included: patient preference for convenience due to geography (e.g. sometimes seeing one physician at an academic medical center and other times seeing the same type of physician closer to home), providers’ time constraints (including the pressure to see more patients faster, resulting in lack of time for explaining why a referral might not be necessary), organizational processes (such as automatically scheduled follow-up appointments with specialists after hospital discharge, bypassing providers seen pre-hospitalization), and the increasing availability of urgent care centers.⁵ This prior work – combined with our use in this study of a time-varying co-morbidity index – arguably lessen concerns about possible self-selection into more highly fragmented care.

Several limitations warrant discussion. First, this study is observational; we cannot infer causality or rule out unmeasured confounders. Second, we cannot measure communication between providers, so the presence of fragmentation should not be interpreted as definitive evidence of the absence of communication. Third, we cannot measure the clinical appropriateness of the care delivered. Fourth, we did not include pharmacy claims in our analysis, based on the premise that better preventive care may be associated with higher pharmacy costs (which is a direction of association that is distinct from the other hypothesized patterns); separate studies are needed to understand the relationship between fragmentation and pharmacy costs. Fifth, this study did not measure health outcomes, as health outcomes require separate study, often with disease-specific inclusion and exclusion criteria.^35,36 Sixth, this study did not measure indirect costs (such as the time and effort of patients and caregivers spent navigating fragmented care). Finally, the results may not be generalizable to Medicare Advantage.

Despite these limitations, the implications of this study are important, because randomized trials of fragmented care are both potentially unethical and unlikely to occur. This rigorously conducted observational study quantifies what experts have estimated and suggests the magnitude of savings that could potentially be achieved if the frequency of highly fragmented care were reduced. Fragmentation of care is about the organization and delivery care and, as such, is inherently modifiable. Given that previous work has shown that most causes of fragmented care are not related to medical need,⁵ the opportunities to reduce unnecessary fragmented care are likely numerous.

In conclusion, in this national, 15-year study of 13,680 fee-for-service Medicare beneficiaries 65 years and older, high ambulatory care fragmentation was associated with $1,085 more in total costs per person per year than low ambulatory care fragmentation, after adjusting for potential confounders. This magnitude of association is considerable and warrants serious attention, given the urgent need to address the high costs of healthcare in the U.S.