Introduction

Each year in the United States, an estimated 11 000 individuals experience traumatic spinal cord injury (SCI).1 These individuals will incur short- and long-term expenses for medical treatment and rehabilitation. Emergency treatment reduces the damage caused by SCI (immobilization of the spine, maintenance of body functions and prevention of infection), and surgery can improve function in some situations (for example, to address compression of the spinal cord by a herniated disk). Pharmacological therapies aimed at improving long-term function, however, are limited. For example, results from two randomized trials lead the American Academy of Neurological Surgeons to conclude that nerve growth agent GM-1 monosialoganglioside (for example, Sygen, produced by Fidia Pharmaceuticals, Abano Terme, Padova, Italy) fails to confer a clinical benefit.2 Conclusions regarding methylprednisolone (Solu-Medrol, produced by Pfizer Pharmaceuticals, New York, NY, USA) have been less definitive, but here too the American Academy of Neurological Surgeons concluded that the evidence does not support a clinical benefit.2 Individuals who sustain a SCI are at risk for secondary complications resulting in increased use of health care resources. Additionally, persons with substantial disability may require assistance with activities of daily living and have a reduced capacity for paid work. Therefore, a major consequence of disability because of SCI is increased cost to the patient, the health care system and society.

With limited options available to improve the condition of SCI patients, there has been little impetus to systematically and quantitatively estimate the benefits of potential SCI therapies. As a practical matter, it is likely that the long-term benefits of novel therapies would have to be estimated by projecting gains as a function of short-term outcomes measured in clinical trials. In this paper we estimate the association between the motor domain of the Functional Independence Measure (FIM) at the time of discharge from rehabilitation and longer-term quality of life and care costs. Although the target of novel therapies is likely to be neurological function, we chose the FIM as a more functionally relevant measure. Understanding the association between short and long-term outcomes will begin to aid estimation of the economic impact of improved functional ability in persons with a traumatic SCI.

Materials and methods

Data source

The National Spinal Cord Injury Statistical Center (NSCISC) database at the University of Alabama, one of the most extensive registries of SCI patients, contains data on patients treated since 1973 at federally designated model SCI care systems throughout the United States.3 A total of 26 care centers have contributed to the NSCISC since its inception. Funding for continued participation is subject to renewal every 5 years, so the composition of the contributing centers shifts. At this time, there are 14 centers contributing to the NSCISC, centers that together treat ∼15% of all new cases in the United States of SCI.4

Using data through March 2010 for patients admitted no earlier than 1 October 1988 (when FIM scores were first recorded), we excluded subjects for whom (1) age was 5 years or less at time of injury (FIM score is not recorded below the age of 6 years), (2) injury impairment was classified as E using the American Spinal Injury Association (ASIA) scale5 or the corresponding level as measured using the Frankel scale (that is, we excluded subjects with normal motor function), (3) death occurred before discharge from rehabilitation or (4) etiology was gunshot wound or penetrating wound. We excluded gunshot and penetrating wound etiologies, because subjects with these injuries are often excluded from SCI pharmaceutical clinical trials (for example, see Bracken et al.6 and Geisler et al.7). Retaining this population would therefore make projections developed here less applicable to results from these and other clinical trials.

The FIM score

The FIM is a popular measure in the United States for assessment of SCI patients, because it is required in order to obtain reimbursement from Medicare.8 The FIM’s properties have been extensively documented,9, 10, 11, 12, 13 and its ability to detect meaningful changes in function during rehabilitation has been verified.14, 15 The motor domain of the FIM score is the sum of the ratings for each of 13 tasks.16 Each task rating ranges from one (full assistance required) to seven (complete independence). Hence, the total motor domain FIM score (hereafter referred to as the FIM) ranges from 13 (least favorable) to 91 (most favorable). As there is a high correlation among the 13 FIM motor tasks,16 this analysis uses the sum of these scores. To adjust for a systematic decrease in FIM scores following adoption of a prospective payment system by inpatient rehabilitation facilities in January, 2002, we decreased pre-2002 total FIM scores by 2.77.17

Analysis

Outcomes

The analysis investigated the association between FIM measured at the time of release from rehabilitation and both one-time outcomes and outcomes measured at repeated follow-up times. One-time outcomes included: (1) residential status at discharge from rehabilitation, categorized as ‘community’ (reported in the NSCISC database as private residence, group living situation or hotel/motel) or ‘professional care’ (reported in the NSCISC database as hospital or nursing home) and (2) survival. Follow-up outcomes reported potentially at 1, 5, 10, 15 and 20 years post injury included: (3) FIM score, (4) residential status (again, dichotomized as ‘community’ or ‘professional care’), (5) number of hospitalizations during the previous year, (6) number of days hospitalized during the previous year, (7) hours of paid care assistance received per day, (8) total hours of care assistance received per day (paid and informal unpaid) and (9) paid work hours per week. As explained in the Technical Appendix (in Supplementary Information), we dichotomized the last three of these outcomes for statistical reasons (0 or >0).

Outcome-specific inclusion criteria

For paid hours of care assistance per day (outcome 7), total hours of assistance per day (outcome 8), and paid hours worked per week (outcome 9), we restricted the analysis to observations for which residential status was classified as ‘community’, because these outcomes have either no meaning or a noncomparable meaning among subjects who are institutionalized.

Statistics

We regressed outcomes against the FIM using statistically appropriate techniques (see Technical Appendix (in Supplementary Information) for details). Binary outcomes included residential status at discharge (outcome 1) and at follow-up (outcome 4). For statistical reasons, we also converted paid hours of assistance (outcome 7), total hours of assistance (outcome 8) and paid work hours (outcome 9) to binary variables, dichotomizing these quantities as 0 or >0. Negative binomial outcomes included number of hospitalizations (outcome 5) and number of days hospitalized (outcome 6). We modeled FIM at follow-up using linear regression (outcome 3). Finally, we used a Cox proportional hazards model to model survival. For follow-up outcomes potentially measured on multiple occasions for each subject (outcomes 3–9), we used a generalized linear model with a robust variance estimator.

Controlled variables

In addition to the FIM, the analysis controlled for the following characteristics. First, it controlled for both ASIA grade (injury completeness) and neurological level (that is, spinal column location where the injury occurred). In cases where the Frankel injury completeness level was reported, we replaced this information with the corresponding ASIA grade. As noted earlier, injury completeness and neurological level are correlated with outcomes of interest.18, 19, 20 As the goal of this study is to estimate the incremental implications of changes in the FIM score (independent of changes in injury completeness and neurological level) in terms of longer-term outcomes, our analysis controls for these characteristics.

In order to capture possible interactions between injury completeness and neurological level while creating a reasonable number of joint categories, we used categories based largely on Sipski et al.21 For example, the first injury completeness/neurological level category includes all subjects with injuries at cervical levels 1–4 (that is, C1–C4) and completeness grades of A (complete) or B (incomplete with sensory function only preserved below the point of injury).

We also controlled for demographic characteristics because these factors can influence outcomes of interest, such as survival and the number of rehospitalizations.18, 19, 20, 21 Specifically, we controlled for gender, race, Hispanic ethnicity and marital status for all outcomes. In addition, for outcomes measured repeatedly at follow-up, we controlled for age at follow-up, while for outcomes not measured repeatedly at follow-up (survival time and residential status at discharge), we controlled for age at the time of injury. Finally, the analysis controls for year of admission to rehabilitation to address the possibility that practices change. For example, studies have reported a decline in the length of stay for rehabilitation over time.22

As there are a large number of observations in the data set and hence ample degrees of freedom, we retained all predictors in the models regardless of the statistical significance of their regression coefficients.

Imputation of missing values

As explained in greater detail in the Technical Appendix (in Supplementary Information), we imputed missing values for FIM measured at discharge and for all outcomes except for date of death. Next, we created five imputed data sets, eliminating from each the imputed follow-up values for time points beyond the date when the subject died or beyond the date when the data set was frozen (March 2010). We conducted each regression on each of the five imputed data sets and used appropriate techniques to combine the resulting estimates quantifying the association between FIM and each outcome.

Absolute impact of FIM score change on outcomes

This study estimated the absolute impact of a one-point change in FIM score at discharge from rehabilitation on outcomes, because health economic outcomes of interest are typically proportional to these absolute changes. For example, savings stemming from reduced institutionalization is proportional to the absolute change in the probability of institutionalization. The impact on the odds ratio cannot be used to directly inform economic analyses. Nonlinear regression outcomes for which we computed the absolute impact of a one-point FIM score change included residential status at discharge from rehabilitation, residential status at follow-up, hospitalizations per year, days hospitalized per year, any paid care assistance, any assistance and any paid work.

As explained in the Technical Appendix (in Supplementary Information), the absolute impact of a one-point FIM score change varies across subjects. To estimate a single value for the absolute impact of a one-point FIM score change, we therefore averaged the impacts estimated individually for every subject in the NSCISC database included in our analysis.

Statement of ethics

Use of the de-identified NSCISC data was approved by the Institutional Review Board at Tufts Medical Center.

Results

We retrieved data for 14 620 subjects from the NSCISC database. Of these subjects, we excluded 2935, because of at least one of the following exclusion criteria: their injury impairment was classified as ‘normal’ (ASIA grade of ‘E’) (n=214), death occurred before discharge (n=20) or the injury was caused by a gunshot or deep penetrating wound (n=2712) (note that the exclusion categories are not mutually exclusive). Tables 1 and 2 summarize the characteristics of the retained sample. As detailed in Table 2, the available number of subjects drops from 9644 at 1-year follow-up to 191 at 20 years follow-up. The number of subjects with entries in any particular field at each follow-up time point was below these limits.

Table 1 Subject characteristics—characteristics measured once
Full size table
Table 2 Subject Characteristics: characteristics measured at follow-up
Full size table

Table 3 reports the impact on each outcome of a one-point FIM score change at discharge from rehabilitation and the central estimates for the regression coefficients. As detailed in the Technical Appendix (in Supplementary Information), Table A-1, the confidence intervals are narrow, extending no more than 23% above and below the central estimate value in all cases. The Technical Appendix (in Supplementary Information) also shows that for all but two outcomes, estimates of the impact of a change in the FIM score developed using the imputed data did not differ substantially from estimates developed using the original data. For hospitalizations per year and days hospitalized per year, however, the difference between the impacts computed using the original data were substantially greater than the impacts computed using the imputed data, approaching nearly a factor of two for days hospitalized per year.

Table 3 Regression results: change in outcomes associated with one-point change in FIM score at discharge from rehabilitation
Full size table

Finally, Table 4 reports the impact of one-point FIM score change on follow-up outcomes by year of follow-up. Technical Appendix Table A-2 (in Supplementary Information) lists the confidence intervals for the central estimates in Table 4.

Table 4 Change at follow-up associated with a one-point change in FIM score at discharge from rehabilitation
Full size table

Discussion

The results from this analysis show that higher values of FIM measured at the time of discharge from rehabilitation are associated with both intermediate and long-term outcomes. In the short term, a higher FIM score is associated with a greater probability of living in the community, rather than in an institutionalized setting. Over the longer term, a higher FIM at rehabilitation discharge is associated with longer survival, improved function at follow-up (that is, higher FIM scores at follow-up), reduced care burden (lower probability of institutionalization, less hospitalization and fewer hours of both paid and informal assistance) and a greater probability of having paid work.

Heterogeneity across treatment centers and among individuals makes it inappropriate to use our findings to predict outcomes for specific individuals. Nonetheless, the findings described here can be used to project both intermediate and long-term population average benefits of therapy based on a clinical trial reporting only the much shorter term FIM score at discharge from rehabilitation. For example, a hypothetical therapy that on average achieves a five-point improvement in the motor domain total FIM score would reduce the proportion of individuals institutionalized at discharge by 1.7%, and the proportion of individuals institutionalized at follow-up from between 0.6 to 0.8%. Hospitalizations per year per individual would drop by 0.02 (that is, 2 per 100 individuals), and days hospitalized per year would on average drop by more than 0.3. The proportion of individuals receiving paid assistance each day would decrease by 3.5%, while the proportion of individuals receiving any assistance each day (paid or unpaid) would decrease by more than 4%. Finally, the proportion of individuals with any paid work would increase by 2%.

Health care savings resulting from these hypothetical improvements can also be estimated. Taking the average cost of a semi-private room in a nursing home facility to be $198 per day23 implies that the 1.7% reduction in institutionalization following rehabilitation would save an average of more than $1200 per treated individual in the first year following injury. The 0.6–0.8% institutionalization reduction during subsequent follow-up years would save from $430 to $580 annually per treated individual. Hospitalization stays for an individual with SCI cost on average $20 000 (the average charge is $40 000).24 Hence, reducing the number of hospital stays by 0.02 per year would save $400 annually per treated individual. A home health aid costs $21 per hour.23 Results in Table 2 indicate that all individuals with SCI receive an average of 2.6 h per day of paid assistance in the first year following their injury but that this help is used by 32.4% of individuals with SCI. Taken together, these results imply that conditional on receiving any paid assistance, individuals with SCI receive an average of ∼8 h per day of paid help (that is, 2.6 h divided by 32.4%). Estimates for other follow-up years are generally similar. Those individuals receiving any paid home help therefore cost an average of ∼$60 000 per year. A 3.5% reduction in paid home health aid use would therefore save ∼$2200 annually per treated individual.

These findings are subject to certain limitations. First, the multiple imputation algorithm that was most appropriate for this analysis treated all variables as normally distributed. In the case of dichotomous outcomes, imputed values had to be rounded before the regression analysis was conducted (values <0.5 were converted to zero, whereas other values were converted to one). Although this issue does introduce some bias, Bernaards et al.25 showed that omitting incomplete observations can introduce even greater bias. Importantly, the results we computed using the imputed data were in most cases very similar to the results we computed using the original data. The most extreme difference arose in the analysis of days hospitalized per year at follow-up. Even in this case, the result computed using the imputed data and the result computed using the original data were within a factor of <2. The results based on the imputed data should be viewed as more valid as patients lost to follow-up may differ systematically from those who are followed.

Second, adjusting for injury completeness may have resulted in over control of confounders. Recall that a key motivation for this analysis is projecting long-term benefits associated with short-term improvements in the FIM so that the FIM can be used to more comprehensively evaluate the benefits of SCI therapies. If novel therapies can influence injury completeness, then controlling for injury completeness, as done here, would obscure benefits estimated using the FIM and the relationships developed in this paper. The approach used here can hence be regarded as being conservative, because it potentially understates the magnitude of the long-term benefits associated with FIM score improvements.

Third, the estimated absolute impact of FIM score on outcomes was calculated based on the assumption that the NSCISC database is representative of the United States population of individuals with SCI. Go et al.26 found the NSCISC database to be generally similar to the population with SCI as documented in state population registries, although more severe injuries, non-Caucasians, and injuries resulting from acts of violence were very modestly over-represented in the NSCISC database. Changes in the participating care centers over time has also limited follow-up in some cases, a factor we attempted to address by use of multiple imputation techniques, as discussed above. In any case, the NSCISC database is the most extensive data set in the United States for the SCI population.

Finally, this analysis excluded individuals with SCI resulting from gunshot wounds or deep penetrating injuries, because this group is often excluded from therapeutic clinical trials for SCI treatments. As a result, our findings cannot be generalized to this group.

Despite these limitations, this work provides a useful starting point for using FIM data collected in a relatively short-term clinical trial to project longer-term outcomes. In the future, it may be possible to improve upon the work described here by using an instrument better designed to evaluate function in people with SCI. For example, the Spinal Cord Independence Measure is more sensitive to functional changes in this population than the FIM,27, 28 but has yet to be recorded over an extended time period in a large population. We therefore encourage incorporation of the Spinal Cord Independence Measure into data collection efforts designed to track large numbers of individuals with SCI over an extended time period. In the meantime, the relationships reported here provide a useful projection of long-term outcomes so that decision makers can better understand the overall value of therapies for the SCI population.