Continuity of Care with a Primary Care Physician and Mortality in Older Adults

Abstract

Background

We examined whether older adults who had continuity of care with a primary care physician (PCP) had lower mortality.

Methods

Secondary analyses were conducted using baseline interview data (1993–1994) from the nationally representative Survey on Assets and Health Dynamics among the Oldest Old (AHEAD). The analytic sample included 5,457 self-respondents 70 years old or more who were not enrolled in managed care plans. AHEAD data were linked to Medicare claims for 1991–2005, providing up to 12 years of follow-up. Two time-dependent measures of continuity addressed whether there was more than an 8-month interval between any two visits to the same PCP during the prior 2-year period. The “present exposure” measure calculated this criterion on a daily basis and could switch “on” or “off” daily, whereas the “cumulative exposure” measure reflected the percentage of follow-up days, also on a daily basis allowing it to switch on or off daily, for which the criterion was met.

Results

Two thousand nine hundred and fifty-four (54%) participants died during the follow-up period. Using the cumulative exposure measure, 27% never had continuity of care, whereas 31%, 20%, 14%, and 8%, respectively, had continuity for 1%–33%, 34%–67%, 68%–99%, and 100% of their follow-up days. Adjusted for demographics, socioeconomic status, social support, health lifestyle, and morbidity, both measures of continuity were associated (p < .001) with lower mortality (adjusted hazard ratios of 0.84 for the present exposure measure and 0.31, 0.39, 0.46, and 0.62, respectively, for the 1%–33%, 34%–67%, 68%–99%, and 100% categories of the cumulative exposure measure).

Conclusion

Continuity of care with a PCP, as assessed by two distinct measures, was associated with substantial reductions in long-term mortality.

MANY proposals exist about how best to restructure the American health care delivery system, and the medical home is at the heart of most. Promoted by politicians (1,2), physicians (3,4), and insurers (5), the medical home builds on the primary and chronic care models (6–9), which share continuity of care as a dominant feature. But as Bindman (10, p. 352) noted earlier this year, the discussion “about and the implementation of the medical-home delivery model is progressing much faster than the accumulation of evidence about its effectiveness.”

Because medical homes will receive additional reimbursement, considerable effort is being spent on developing assessment tools for whether physicians’ practices qualify (11–14). Such tools should focus on key primary care concepts, have face validity, have demonstrable predictive validity of desirable outcomes, and be measurable using administrative records if possible (5). The administrative burden of current tools, however, seems excessive (15).

In this article, we step back from the rush to develop assessment tools for certifying medical homes and focus on the more fundamental issue—does continuity of care, a core component of the medical home, really make a difference (10)? Specifically, we assess the impact of continuity on mortality among older adults using two time-dependent adaptations of our recently developed measure (16) that can be implemented solely using administrative claims. To the best of our knowledge, no other study has examined the association of continuity of care with mortality, which may be considered the ultimate predictive validity criterion.

METHODS

The AHEAD Study

Participants in the Survey on Assets and Health Dynamics among the Oldest Old (AHEAD) were identified from two sources: (a) the 1992 household screening conducted for a companion study of preretirement (i.e., 50–64 year old) adults and (b) a supplemental sample from the Medicare Master Enrollment File of persons 80 years old or more (17). Because African Americans, Hispanics, and Floridians were oversampled, the data were weighted to adjust for the unequal probabilities of selection. Baseline interviews were conducted in 1993–1994 with 7,447 respondents 70 years old or more, and an 80.4% response rate was obtained.

Exclusion Criteria

Of the 7,447 AHEAD participants, 802 (10.8%) could not be linked to their Medicare claims, 665 (8.9%) were in managed Medicare at or during the 2 years prior to baseline, and 523 (7.0%) were proxy respondents, leaving 5,457 self-respondents in our analytic sample (73.3% of the total AHEAD sample). Participants in managed Medicare care were excluded because the data reporting requirements for them differ from those in fee-for-service Medicare plans (18). Proxy respondents were excluded because cognitive and psychosocial data were not collected from them. Participants were censored at the first of three competing risks—death, enrollment into managed Medicare, or January 1, 2006.

Mortality

Vital status was taken from the Medicare denominator files, which indicate on a monthly basis whether the beneficiary is alive or dead. Because the exact day of the month of death is not given, we assigned the death day to be the middle of the month in which the death was registered. We used multivariable proportional hazards regression with competing risks (19) to model time to death. Model development and evaluation followed standard procedures (20,21).

Defining Continuity of Care

Consensus in defining continuity of care involves (a) the many positive outcomes that should result from it and (b) the multidimensionality of the concept (22–24). Continuity is expected to result in “improved doctor–patient relationships, enhanced physician knowledge of the patient, greater rapport and disclosure, increased compliance, reduced hospitalization rates, increased patient and physician satisfaction, reductions in disability levels, costs, and missed appointments, and improved problem recognition and management” (16, p. S160). Continuity is also presumed to have several dimensions, including informational, longitudinal, interpersonal, geographic, team based, and familial (23).

Based on this literature, we defined continuity of care as “an ongoing relationship with a particular [primary care] physician in the outpatient setting with sufficient frequency for that physician to assume primary responsibility for both the patient’s basic health care needs and her overall disease and care management” (16, p. S161). Because we relied solely on Medicare claims, we cannot determine whether the physician actually assumes primary responsibility for her patient or that the patient recognizes that physician as her primary physician. Instead, our conceptualization assumes that there is a sufficient frequency of visits between a particular physician and a particular patient for continuity to have occurred.

Measuring Continuity of Care

We used two time-dependent measures based on our prior definition of continuity as having no more than an 8-month interval between any two outpatient visits during a 2-year period to the same primary care physician (PCP) (16). Face validity was established in consultation with several general internists, geriatricians, and gerontologic health services researchers (25). The consensus was that “semiannual” visits were the desirable minimum for older adults. We relaxed that threshold to at least one visit every 8 months to account for vacations and scheduling difficulties. The 2-year window ensured longitudinal continuity. Prior analyses found that varying the visit interval between 4–12 months had no meaningful impact on measuring continuity (16).

To identify PCP visits, we used a multistage approach. First, we deleted all inpatient-related line items and claims (i.e., Part A or inpatient standard analytic files [SAF] claims). Second, we deleted all Part B (i.e., carrier SAF) claims in which the “from and through” dates overlapped with hospital stays, except claims that occurred on the day of a hospital admission because these likely reflected diagnostic visits triggering that hospitalization. Third, we identified PCP line items or claims in the carrier SAF as those that were from general practitioners, family practitioners, geriatricians, internists, preventive medicine specialists, unclassified multispecialty clinic or group practices (very few of these led to patients having continuity), nurse practitioners, certified clinical nurse specialists, and physician assistants using provider specialty codes and unique provider identification numbers. Fourth, we deleted all line items or charges without evaluation and management (E&M) codes because without E&M codes it was unlikely that patients had been seen by the PCP. Finally, we deleted all duplicate line items and claims (same day, same provider). All remaining line items qualified as PCP visits.

We refer to our first time-dependent measure as “present exposure.” It dynamically calculated whether the definition of continuity of care was satisfied on each postbaseline interview day. That is, for each day after the participant’s baseline interview, this measure looked back in time for 2 years to determine whether the definition was met, and the measure was categorized as yes or no for that day. Then, this process was repeated each subsequent day, until censoring or death. The major benefit of using this present exposure measure (i.e., indexing continuity of care as a time-dependent covariate recalculated on a daily basis in the context of a proportional hazards model) is that it avoids the problem of immortal time bias that frequently occurs when administrative claims data are used to calibrate risk factor exposure (26). That is, it considers the participant as presently exposed to continuity of care only for those days on which the definition is satisfied and accurately accommodates participants having oscillating (on or off) exposures to continuity of care.

A drawback to the present exposure index of continuity of care, however, is that it does not reflect on the individual participant’s history of (i.e., their “cumulative exposure” to) continuity of care up to that day. Therefore, our second time-dependent indicator is a cumulative exposure measure, which places participants into one of five categories based on the daily updated percentage of prior postbaseline days for which they met the continuity of care definition. That is, this measure calculates the cumulative percentage of follow-up days for which the continuity of care definition was met prior to that day. Thus, this measure reflects cumulative exposure to continuity of care prior to any given day. The five categories were 0%, 1%–33%, 34%–66%, 67%–99%, and 100% of the follow-up period. We chose the two extreme categories because of their straightforward interpretation (i.e., never vs always having continuity), whereas the middle categories were chosen to yield roughly equivalent group size. A set of time-dependent dummy variables (having continuity for 1%–33%, 34%–67%, 68%–99%, or 100% of the follow-up days vs the reference category of never having had continuity of care) was then constructed, and this set of time-dependent dummy variables are recalculated daily (i.e., they are time dependent).

Statistical Methods

The outcome of interest is mortality, and it is measured as the number of days from baseline to death. Two proportional hazard regression models with competing risks were used to model the hazard rate of death in terms of the time-dependent continuity of care measures, along with a participant-specific set of covariates that were used to adjust for potential confounders. The unit of analysis was the participant for both of the models that we estimated. The first model used our time-dependent present exposure (PE_t) continuity of care measure, which is recalculated on a daily basis (i.e., the t subscript reflects day t) to index whether the continuity of care definition was met during the 2-year period immediately preceding that day for each participant. The second model uses our time-dependent cumulative exposure (CE_t) measure of continuity of care, which is a set of time-dependent dummy variables (i.e., CE1_t, CE2_t, CE3_t, and CE4_t reflecting the four contrast categories vs never having had continuity of care) that are also recalculated on a daily basis (i.e., the t subscript reflects day t) to index the percentage of days from each participant’s baseline interview to the present day for which the continuity of care definition was met. In both models, participants were censored at the end of the observation period (January 1, 2006) or at the day on which they entered a managed care plan, whichever came first.

Covariates

Analyses adjusted for a number of covariates. Sociodemographic factors included age, sex, race, veteran status, marital status, subjective life expectancy, and population density. Social support measures included living arrangements, living children, and access to future helpers. Socioeconomic factors were employment status, education, income, and Medicaid status. Alcohol consumption, smoking, and weight tapped health lifestyle behaviors. Morbidity was indicated by self-rated health; difficulty walking; activities of daily living (ADLs); instrumental activities of daily living (IADLs); depressive symptoms; cognitive status; and reporting arthritis, cancer, diabetes, a heart condition, hypertension, stroke, or psychological conditions. Hospitalization in the year before baseline was taken from the Medicare claims.

Selection Bias

Our exclusions (nonlinkage to Medicare claims, proxy respondents, and participation in managed Medicare) created the potential for selection bias, which we addressed using propensity score methods to reweigh the data (27–30). Among all 7,447 AHEAD participants, we estimated (data not shown) a multivariable logistic regression model of inclusion in the analytic sample using an extensive set of baseline interview data. The fit of this model was good (C statistic = .72; Hosmer–Lemeshow statistic = .15) (31,32), indicating that an effective reweighing of the data to adjust for any potential selection bias would be effective. Within propensity score (predicted probability) deciles, we determined the average participation rate (i.e., inclusion in the analytic sample or P) and used the inverse (1/P) to reweigh the data. This adjusted the existing weights traditionally used in analyzing the AHEAD data for the potential selection bias. The propensity score weights were then rescaled so that the final weighted N was equal to the actual number of participants in the analytic sample. As an added safeguard, we also estimated our survival models without using the propensity score adjustment and found the results to be equivalent. Thus, our results are not artifacts of adjusting for potential selection bias.

RESULTS

Descriptive Data

Table 1 contains the percentages (or means) for each covariate. The mean age at baseline was 77 years; 38% were men, 10% were African American, 4% were Hispanic, and 41% were widowed. One fourth had only been to grade school, and mean income was $25,417. One fourth had arthritis, 9% had angina, 13% had cancer, 12% had diabetes, 46% had hypertension, and 7% had psychological problems. The mean number of ADL and IADL difficulties was 0.34 and 0.41, respectively. About 18% had been hospitalized in the year prior to the baseline interview. The total number of person-years of surveillance was 40,713 with a mean of 7.5 years per person. During the 2-year prebaseline period, there were 70,244 physician visits (M = 6.4 per person-year). Over the 12-year follow-up period, there were 351,614 physician visits (M = 8.6 per person-year). By the end of the follow-up period, 2,964 (54.3%) participants in the analytic sample had died.

Table 1.

Weighted Means and Percentages on the Covariates, Overall, and by Percent of Continuity of Care Across the Entire Period

	Overall	Percent of Continuity of Care Across the Entire Period
		0	1–33	34–67	68–99	100
Sample size	5,457	1,472	1,677	1,096	774	438
Sociodemographics
Age	77.4	77.9	77	77.3	77.3	78.3
Men, %	38.1	48.3	35.3	32.9	32.7	38.0
Race, %
White (reference group)	84.8	80.9	87.5	87.8	83.0	82.9
Hispanic	3.9	4.2	2.8	3.9	5.7	4.4
African American	10.2	13.4	8.7	7.4	11.2	10.9
Other race	1.1	1.6	0.9	0.9	0.1	1.8
Veteran, %	22.4	28.2	21.8	17.8	20.6	19.8
Marital status, %
Married (reference group)	50.4	53.3	49.6	49.6	48.6	49.0
Never married	3.3	3.3	2.9	3.5	4.9	2.0
Separated or divorced	4.9	4.8	5.8	4.3	4.2	4.8
Widowed	41.4	38.7	41.7	42.6	42.4	44.2
Subjective life expectancy
0%–50% (reference group)	59.8	56.2	58.9	61.3	63.9	63.5
>50%	23.5	25.9	24.6	22.5	21.8	16.8
No answer, %	16.8	17.9	16.6	16.1	14.3	19.8
County population, %
>1 million	20.7	26.1	19.8	16.8	18.6	20.0
600,000–1 million	15.8	17.0	14.5	13.5	16.9	20.4
250,000–600,000 (reference group)	18.4	17.7	18.8	19.4	18.6	16.5
50–250,000	20.5	18.8	20.4	22.4	22.1	19.2
<50,000	24.6	20.4	26.5	27.8	23.9	23.8
Social support
Living alone, %	36.5	34.6	37.8	37.6	37.0	34.4
Has living children, %	86.2	84.9	86.7	87.0	85.2	88.3
No future helper, %	39.1	36.6	40.1	40.2	41.3	36.7
Socioeconomic
Working, %	8.8	9.5	10.4	6.8	9.1	5.0
Education, %
Grade school	25.5	27.9	24.4	24.9	24.3	25.5
High school (reference group)	47.6	43.9	45.6	51.0	50.0	54.1
Some college	26.9	28.2	29.9	24.1	25.6	20.5
Low income, %	45.6	47.8	42.6	46.0	43.7	51.9
On Medicaid, %	9.0	7.5	8.6	10.7	9.9	9.1
Lifestyle
Should reduce drinking (CAGE cut down), %	14.3	16.0	13.6	13.1	11.7	19.1
Ever smoked, %	51.9	56.6	50.1	49.9	47.6	56.3
Weight, %
Underweight	3.8	4.9	3.2	3.6	2.6	4.8
Normal or overweight (reference group)	82.7	83.8	82.7	81.8	83.1	80.8
Obese	13.5	11.3	14.1	14.6	14.3	14.4
Morbidity
Fair or poor self-rated health, %	35.4	33.6	30.2	37.1	37.6	52.8
Difficulty walking, %	37.4	36.0	32.2	37.4	41.8	54.1
Activities of daily living, %
0 (reference group)	80.7	80.6	83.3	81.8	80.6	68.9
1	10.5	9.2	8.9	11.3	11.0	17.5
2	4.4	4.9	4.3	3.6	3.5	7.3
≥3	4.4	5.2	3.6	3.3	4.9	6.3
Instrumental activities of daily living, %
0 (reference group)	78.8	76.4	81.6	81.5	79.9	67.4
1	10.5	10.6	10.2	9.9	10.9	12.9
2	5.0	5.9	4.1	4.6	4.3	8.0
≥3	5.6	7.1	4.1	4.0	5.0	11.7
Depressive symptoms (CESD-8 count), %
0 (reference group)	37.7	37.3	42.2	37.8	34.9	26.6
1–3	44.6	45.8	41.1	47.1	45.6	46.7
>4	17.6	16.9	16.6	15.0	19.4	26.7
Cognitive status (TICS-7 score), %
≤10	28.3	31.7	26.0	26.3	26.1	35.0
11–13 (reference group)	31.8	30.1	32.5	34.0	30.8	31.1
≥14	39.9	38.2	41.5	39.7	43.1	33.9
Arthritis, %	24.7	18.8	21.1	29.5	32.5	32.5
Cancer, %	12.9	11.3	12.4	13.2	13.4	18.1
Diabetes, %	12.5	10.3	8.1	11.6	18.9	27.3
Lung disease, %	9.4	10.0	7.3	8.6	9.0	17.6
Heart condition, %	28.9	27.6	23.9	28.0	34.6	44.4
Hypertension, %	45.5	38.2	39.9	50.6	57.1	57.4
Stroke, %	9.6	9.0	7.7	10.1	11.1	14.7
Psychological conditions, %	7.0	5.8	6.9	7.1	8.4	8.5
≥2 of the 8 diseases above, %	44.9	39.2	34.7	49.0	57.1	70.4
Claims-based health services use
Hospitalized prior to baseline, %	18.0	14.5	16.3	18.8	22.0	26.6

CAGE = Cut, Annoyed, Guilty, Eye; CESD-8 = Center for Epidemiologic Studies Depression Scale, 8-items; TICS-7 = Telephone Interview for Cognitive Status, 7-items.

The Prevalence of Continuity of Care

Table 1 also shows the distributions on the cumulative exposure measure of continuity of care when aggregated across the follow-up period. That is, in Table 1, the continuity of care category into which a participant is placed was based on their individual percentage distribution of continuity over all of their particular follow-up period up to their time of their death or study censoring (which occurred at their entrance into managed Medicare care or January 1, 2006, whichever came first). In the analytic models (shown in Table 2), however, the category (percentage of follow-up days with continuity of care) is calculated based on each participant’s continuity of care experiences prior to every given day, with this calculation iteratively repeated on a daily basis going forward. Therefore, please note that these distributions cannot be presented at all because they vary daily. That said 27.0% of the analytic sample never satisfied the continuity of care definition during their follow-up period, whereas 30.7% had continuity for 1%–33% of their follow-up days, 20.1% had continuity for 34%–67% of their follow-up days, 14.2% had continuity for 68%–99% of their follow-up days, and 8.0% had continuity on all of their follow-up days. The mean number of follow-up days within categories of the continuity of care percentages shown in Table 1 were equivalent for those with 1%–34%, 35%–67%, or 68%–99% (i.e., 3,123 vs 3,132 vs 3,011 days, respectively), whereas they were noticeably lower for those who either never or always had continuity of care (i.e., 2,050 vs 1,581 days, respectively).

Table 2.

Crude and AHRs From the Propensity Score Reweighed Mortality Models

Continuity of Care Measures	Crude HRs			AHRs
				Time-Dependent Model			Time-Dependent Cumulative Exposure Model
	HR	p Value	95% Confidence Interval	HR	p Value	95% Confidence Interval	HR	p Value	95% Confidence Interval
Time-dependent measure*	0.87	.0007	0.81–0.95	0.84	<.0001	0.77–0.91
Time-dependent cumulative exposure measure†
0% of follow-up days (reference group)	1.00						1.00
1%–33% of follow-up days	0.27	<.0001	0.23–0.31				0.31	<.0001	0.27–0.36
34%–67% of follow-up days	0.40	<.0001	0.35–0.45				0.39	<.0001	0.34–0.45
68%–99% of follow-up days	0.47	<.0001	0.41–0.54				0.46	<.0001	0.40–0.53
100% of follow-up days	0.75	<.0001	0.66–0.84				0.62	<.0001	0.55–0.70

Notes: AHRs are adjusted for all variables shown in Table 1. AHRs = adjusted hazards ratios; HRs = hazards ratios.

^*This measure is coded 1 = yes, has continuity of care for the prior 2-year period, versus 0 = no.

^†This measure is coded as the percentage of follow-up days on which continuity of care existed for the prior 2-year period.

Correlates of Continuity

As shown in Table 1, the distributions of the covariates varied across the categories of the cumulative exposure measure of continuity when it was calculated for the entire follow-up period. The largest variations indicate that having continuity was more often likely among those with lower subjective life expectancy; high school graduates; those with fair or poor self-rated health; difficulty walking; higher levels of depressive symptoms; and having arthritis, cancer, diabetes, lung disease, heart conditions, hypertension, stroke, psychological conditions as well as among those hospitalized in the year prior to baseline. In nearly every within-row comparison, the highest percentage of these characteristics was found for those having continuity all the time, suggesting that one reason older adults have continuity is to deal with their increased health burdens.

Modeling Mortality

Table 2 contains the results of the multivariable, competing risks proportional hazards mortality models. Column one contains the crude “hazards ratios” (HR), column two contains the “adjusted hazards ratios” (AHRs) when using the time-dependent present exposure continuity measure, and column three contains the AHRs when using the dummy variable set of time-dependent cumulative exposure continuity measures. For simplicity, the AHRs for the covariates are not shown. Please note that because both the time-dependent present exposure and cumulative exposure continuity measures used in the models shown in Table 2 were iteratively calculated on a daily basis going forward, it is not possible to present traditional Kaplan–Meier curves because these time-dependent curves by definition change on a daily basis.

Both the present exposure and the cumulative exposure measures were statistically and substantively significantly associated with mortality. On days when the present exposure measure was “on” (i.e., when the continuity definition was met for the 2 years immediately prior to that day), the risk of mortality was reduced by 16% (AHR = 0.84, p < .0001). Similarly, the risk of mortality was statistically and substantively significantly reduced at all levels of the cumulative exposure continuity measure, with AHRs of 0.31, 0.39, 0.46, and 0.62 (all p values < .0001), respectively, for the 1%–33%, 34%–67%, 68%–99%, and 100% of the time categories.

DISCUSSION

At the heart of the medical home are the core principles that make up the primary and chronic care models. Continuity is central to both. We have demonstrated the reduction of mortality associated with two time-dependent continuity measures that were calculated using administrative claims. The magnitude of the relationships of the continuity measures with mortality risk was not appreciably affected by adjustment for self-reported baseline sociodemographic, social support, socioeconomic, health lifestyle, and morbidity measures. To the best of our knowledge, this is the first study to demonstrate the significantly protective association of continuity of care with mortality. Therefore, we submit that our results provide considerable evidence that continuity of care matters for the ultimate criterion—life or death.

One finding warrants special mention. For the cumulative exposure measure, one might have expected to observe a dose–response relationship between the percentage of prior follow-up days with continuity and subsequent mortality risk such that the greater the cumulative exposure to continuity of care, the lower the mortality risk. Although all the cumulative exposure dummy variables (vs never having continuity of care) protected against mortality risk, what we found was that the greater the cumulative exposure, the lesser the protection. Three explanations of this relationship warrant consideration. First, it might be that the functional form of the relationship between the cumulative exposure to continuity of care measure and mortality involves a low threshold met by all the cumulative exposure categories. If that was the case, however, we should have found equivalent AHRs across cumulative exposure categories rather than the dose–response pattern shown in Table 2.

A second explanation is that the smaller protective effect of having continuity all the time may reflect comorbidity confounding. For example, as patients become more seriously ill and especially as they enter the end stages of their conditions, their PCPs may be more likely to see them more regularly. Although we do not have data on disease severity per se, we were able to conduct additional analyses to explore this possibility. Simply put, we replicated the results of the cumulative exposure continuity measure separately within strata based on the number of chronic conditions reported by the participants at baseline. Our four strata reflected those with no chronic conditions (N = 1,254), those with only one (N = 1,814), those with two (N = 1,355), and those with three or more chronic conditions (N = 1,034). These results (data not shown) indicated that among those with no chronic conditions, the effect of having continuity all the time (AHR = 0.36, p < .0001) was equivalent to that observed for each of the other levels of having continuity (mean AHR = 0.32, p < .0001). In contrast, among those with one, two, or three or more chronic conditions, the protective effect of having continuity all the time (mean AHR = 0.59, p < .0001) was notably smaller than that for all the other levels (mean AHR = 0.36, p < .0001), although still statistically and substantively significant. This may indicate that when older adults with no chronic conditions have continuity all the time, they enjoy its full benefits (as do those with any level of continuity). Similarly, when older adults with chronic conditions have continuity all the time, their benefit is diminished because part of their continuity may stem from the need for treatment of their chronic conditions rather than from the intrinsic value of continuity.

A third possible explanation is that our results may reflect confounding with the volume of visits to PCPs. To address this possibility, we replicated our analyses separately within five strata based on quintiles of the number of PCP visits per year of follow-up. The quintiles of the number of PCP visits per year of follow-up were 0.00–1.22, 1.23–2.52, 2.53–3.80, 3.81–5.70, and ≥5.71. Continuity could not have an effect in the lowest (first) quintile because the threshold for continuity could not be met at this rate of PCP visits per year of follow-up. In the second, fourth, and fifth quintiles (data not shown), the protective effects of the extreme categories of the cumulative exposure measure (i.e., 1%–33% and 100% of the time) were much closer in magnitude than those shown in Table 2, whereas the protective effects for the middle quintile were similar to those shown in the same. Moreover, as the rate of PCP visits increased, so did the protective effects of continuity. Thus, there is evidence of some moderation effect associated with the volume of PCP visits such that the greater the volume of PCP visits, the greater the mortality benefit that accrues from having continuity of care.

Our work has limitations and three warrant mention. The first is that we don’t know whether the physicians and patients involved would agree that the spirit of the medical home was met in their relationships. The second limitation is that because we relied on Medicare claims, our results only relate to older adults. Finally, we did not evaluate or adjust for the other suggested characteristics of the medical home.

Those limitations notwithstanding our findings have several important implications for health care reform. First, the measures of continuity used in this analysis were calculated from administrative data and could easily be used to assess the degree to which individual medical home practices were, indeed, providing continuity of care. Such assessments could then be used in determining reimbursements. Second, and perhaps more importantly, our findings highlight the beneficial impact of continuity of care on outcomes for vulnerable elders. Moreover, these findings provide much needed empirical data to support health care reform based on rebuilding the nation’s primary care infrastructure in a manner that promotes continuity.

That said, however, it would be prudent to obtain stronger (i.e., experimental or quasi-experimental) evidence that continuity of care and all the other components of the medical home actually have their expected effects before any major restructuring of the health care delivery system (33–35). To be sure, although the promise of the medical home is considerable and a number of demonstration projects are underway (36), the evidentiary base remains quite limited (10). Moreover, the initial evidence is mixed (37) despite the success of selected early adopters (38). And that is very reminiscent of the situation in the 1970s and 1980s during the big push for health maintenance organizations (HMOs) based on the long-standing experience of the pioneering HMOs that was shown not to readily transfer to many subsequent entrants (39,40). Thus, due diligence suggests that at least some caution is prudent in believing that the medical home is the proper panacea for health care reform.

FUNDING

This study was approved by all applicable oversight groups, including the AHEAD restricted data access board (#2003-006; 2-20-2003), the University of Iowa Institutional Review Board (#200303008; 03-24-2003), and the Center for Medicare and Medicaid Services (DUA #14807; 3-3-2005). Support was provided by National Institutes of Health grants AG-022913, AG030333, and AG031307 to F.D.W. G.E.R. is the PI of the Center for Research in the Implementation of Innovative Strategies in Practice (CRIISP) at the Iowa City Veterans Administration Medical Center, and M.P.J. is the Senior Statistician at CRIISP. CRIISP is funded through the Department of Veterans Affairs, Veterans Health Administration, Health Services Research, and Development Service (HFP 04-149). The sponsors of this research had no role in the design of the study, the analyses conducted, or the interpretation and presentation of the results. The opinions expressed here are those of the authors and do not necessarily reflect those of any of the funding, academic, or governmental institutions involved.