The hospitalization of nursing home residents has emerged as an important area of concern for policy makers. These hospitalizations are already frequent, and they are becoming more so.1,2 They result in complications, morbidity, and Medicare expenditures that amount to more than a billion dollars annually.1⇓–3
Empirical research suggests that both the quantity and type of nursing home staff members, especially physicians, might have an impact on the number of potentially avoidable hospitalizations.4 In particular, the lack of physicians at many nursing homes during off hours might be one cause of inappropriate hospitalizations.5 If a medical issue arises during the evening or weekend that cannot be addressed over the phone, the on-call physician can either travel to the facility or recommend that the nursing home resident be transferred to a hospital. All too often, the on-call physician recommends sending the resident to the emergency department.6
Telemedicine makes real-time medical consultation available to nursing home patients and their families via two-way videoconferencing.7 By providing patients with this direct contact, telemedicine could prevent costly hospitalizations of nursing home residents.
This study was designed to answer two questions. First, did the residents of nursing homes that were randomly chosen to receive off-hours physician coverage by a telemedicine service experience a lower rate of hospitalization, compared to residents of homes that received standard physician coverage? And second, if the nursing homes with telemedicine coverage did have lower rates of hospitalization, did they realize substantial savings?
Study Data And Methods
Study Design And Setting
We studied the introduction of telemedicine in a Massachusetts for-profit nursing home chain in the period October 2009–September 2011. The nursing home chain signed a contract with a telemedicine provider to introduce service in eleven nursing homes to cover urgent or emergent calls on weeknights (5:00–11:00 p.m.) and weekend days (10:00 a.m.–7:00 p.m.). As part of this study, the nursing home chain agreed to randomly stagger the introduction of telemedicine coverage. We asked the leaders of both the nursing home chain and the telemedicine provider to blind their staffs to the fact that we would be studying the hospitalizations of residents in these facilities.
All eleven facilities were Medicare and Medicaid certified and cared for a mix of postacute and long-stay residents. Importantly, during the study period the facilities were not engaged in any other intervention that was specifically designed to reduce residents’ hospitalizations, such as the INTERACT program.8
We assigned the nursing homes to categories based primarily on the facilities’ scores on the Five-Star Quality Rating System of the Centers for Medicare and Medicaid Services (CMS)9 and secondarily on their bed turnover rate (that is, total number of admissions per bed). By matching nursing homes on the bed turnover rate, we captured the differences across facilities in the shares of postacute and long-stay residents.10
We randomly assigned six facilities to receive the telemedicine intervention in November 2010, while the other five facilities were scheduled to receive it in October 2011. Thus, the thirteen-month period of October 2009 through October 2010 was designated as the pre-intervention period, and the eleven-month period of November 2010 through September 2011 was designated as the post-intervention period. By examining the first phase of the staggered introduction of telemedicine coverage at randomly selected nursing homes using a pre-post design, we were able to control for secular trends in nursing homes during the study period.
All of the residents in the participating nursing homes received their primary care through physician group practices. Thus, prior to the intervention, evening or weekend calls were directed to the covering physician in the group practice, with off-hours care typically provided by telephone from a remote location.
Before the telemedicine service was introduced into the six nursing homes, separate training sessions were held for direct care staff members and physicians at each facility. The goals of these sessions were twofold. The first was to teach the staff members how to use the service. The second was to educate the physicians about the service and convince them to sign over their off-hours coverage to it.
Across the six treatment facilities, 90 percent of the physicians signed over their off-hours coverage. Because an off-hours phone consultation would not typically generate any reimbursement for the physician, this shifting of calls to the telemedicine service did not generally lead to lower revenue for the physician.
The intervention consisted of introducing into the nursing home a cart with equipment for two-way videoconferencing and a high-resolution camera for use in wound care. When a nursing home resident had an off-hours medical problem, a staff member brought the cart into the resident’s room and contacted the telemedicine service.
The service’s medical call center was staffed by a medical secretary and three providers: a registered nurse, a nurse practitioner, and a physician. Calls were triaged by the medical secretary to the appropriate provider at the call center.
Data for this study were obtained from multiple sources. From the nursing home chain’s electronic health record system, we obtained the following data, aggregated at the facility level: bed size; residents’ demographic and health data; and admissions, hospital transfers, and resident days in the facility per month. From the telemedicine provider, we obtained aggregate monthly data by facility on the number of and reasons for calls to the service.
In addition, for the purposes of categorizing the facilities, we obtained each facility’s CMS five-star rating,9 number of beds, and staffing levels from the CMS Nursing Home Compare website. We also obtained information from the website on all nursing homes in Massachusetts that did not participate in our study, to compare them to the participating facilities at baseline.
The key outcome of interest was the number of residents hospitalized, by nursing home and month. Because of the nature of the nursing home chain’s billing system, this measure captured only the hospitalizations with a stay that included midnight.
Importantly, our outcome measure did not include only the hospitalizations that occurred during the evening and weekend hours. This approach allowed us to incorporate into our results any possible spillover effects on daytime hospitalizations—for example, the telemedicine service could simply delay nighttime hospitalizations until the next day. Based on a recent study,2 we assumed that Medicare paid $10,000 per hospitalization.
We generated descriptive statistics on the frequency and types of telemedicine calls by month and facility. Based on our analysis of these statistics, we categorized certain facilities as “more engaged” and other facilities as “less engaged” in the telemedicine intervention. Next, we compared the treatment and control nursing homes with each other and with all nursing homes in Massachusetts at baseline along several characteristics, to evaluate the study design’s internal and external validity.
In examining nursing home residents’ hospitalizations, we first evaluated the unadjusted pre-post difference for both the treatment and the control groups. To assess the impact of the intervention, we next conducted a difference-in-differences calculation in which we compared the difference in pre-post hospitalizations between the treatment and the control facilities.
To analyze the data in a statistically efficient yet flexible manner, we treated the observed number of hospitalizations in a month as a Poisson distributed random variable. The key variable of interest was the interaction of a facility’s treatment or control status and the time period (before or after the intervention).
We controlled for facility and month fixed effects and an offset for the log of the facility-monthly average census count. The inclusion of the offset variable allowed us to model the per capita rate of hospitalizations in terms of a Poisson regression model, thus respecting the natural discrete distribution of the data while still modeling the ratio of total hospitalizations to population—the true quantity of interest.
To account for the clustering of observations within nursing homes, we used generalized estimation equations with a working correlation matrix that allowed different nonzero correlations between observations from a nursing home that were one, two, three, and four months apart and zero correlation between observations further apart (a four-period dependent structure) to appropriately calibrate standard errors, confidence intervals, and statistical tests.
In an additional analysis, we categorized nursing homes that received the telemedicine intervention according to the extent to which they were engaged in it. The resulting model then had two treatment effects, one for a less engaged treatment facility and the other for a more engaged treatment facility.
Furthermore, because there was some ambiguity about whether calls for emergent cases involved any discretion (that is, hospitalization is relatively certain in such cases), it was not clear whether our measure of the level of engagement should incorporate calls for emergent cases. The results differed minimally according to whether or not we included the emergent cases. As a result, we retained them in the definition of engagement in the analyses reported here.
This analysis is limited in various ways. First, because our data came from eleven nursing homes in a single for-profit chain in Massachusetts during a two-year study period, the results might not be generalizable to other nursing homes or time periods.
Second, the nursing home billing data that we used to record hospitalizations did not provide the time of the resident’s transfer to the hospital. Nor did the billing data include emergency department visits, which might also be influenced by the use of telemedicine. Unfortunately, the billing data also did not allow us to distinguish between hospitalizations for short-stay residents and those for long-stay residents.
Third, although randomization provided a strong study platform to evaluate telemedicine versus on-call coverage at the treatment and control facilities, various unmeasured selection or confounding effects could be associated with which nursing homes became engaged in the telemedicine intervention. Thus, any differences we observed between more- and less-engaged facilities could be spurious artifacts of differences in the value of some unmeasured predictor of engagement and frequency of hospitalization.
However, we included nursing home fixed effects in our regression analyses. This allowed us to control for any omitted time-invariant factors such as proximity to the hospital, facility average case-mix, percentage of physicians who signed over their off-hours coverage to the telemedicine service, and the presence of on-site point-of-care testing (for example, oximetry) that might be correlated with both facility engagement and hospitalizations.
Finally, we were not able to look at other outcomes that might have been related to telemedicine, including the quality of care, the resident’s overall health, staff retention, staff satisfaction, and the satisfaction of the resident or his or her family. Therefore, we assumed that an avoided hospitalization was a positive event, but we were unable to evaluate the health implications associated with that avoidance.