Introduction

Advances in medicine and injury management have helped to decrease the risk of mortality associated with spinal cord injury (SCI) in the first year post injury. However, mortality rates thereafter appear to have remained stable with long-term mortality estimated at around 1.3–2.5% annually.1, 2, 3 If long-term mortality is to decrease, prevention strategies are needed, targeting individuals at risk for excess mortality. This study addresses this issue by identifying the risk and protective factors that may become targets for intervention.

In recent years, investigators have begun to focus on a wider array of risk and protective factors for mortality in addition to the biographic and injury characteristics that have been most widely investigated. Krause4 developed a theoretical risk and prevention model (TRPM) to guide studies of mortality after traumatic neurologic injury, proposing a series of mediational relationships between the different sets of risk and protective factors, with health factors being most proximal and highly related to mortality. A series of analyses on data from the SCI Model Systems (SCIMSs) in the United States has been used to evaluate different sets of predictors.5, 6, 7 Unlike studies using SCIMS data to monitor patterns of mortality over time primarily as a function of biographic and injury characteristics at injury onset,1, 8 this research has utilized follow-up data from a subset of participants. In the first study,5 several sets of risk factors were structured post-hoc according to the theoretical risk model and evaluated among 5497 participants. Hospitalization in the previous year, grade 3 or 4 pressure ulcer, community integration, type of insurance and household income at or below the poverty level were all predictive of mortality and related to substantial variations in life expectancy.5

Several follow-up studies have been conducted using this data and addressing one or more aspects of the preliminary findings. Strauss et al.6 replicated the original study5 using updated data and a larger participant cohort (n=7331), only focusing on economics. They found a significant effect for economics, but smaller than obtained previously.5 However, two recent subsequent analyses7, 9 focused on socioeconomic factors (income, education and employment), using a yet larger participant cohort (n=8027 and n=7955) and correcting for a limitation in the earlier studies, found prominent economic effects consistent or greater than that of by Krause et al.5

A series of studies using data from a clinical USA cohort10, 11, 12, 13, 14, 15 found general support for the TRPM and the importance of each type of predictor. In a simultaneous analysis of all predictors, four health factors were predictive of mortality (that is, hospitalizations, fractures/amputations, surgeries for pressure ulcers, probable major depression), whereas two types of health behaviors (prescription medication use, binge drinking) and one socio-environmental variable (income) were significant. Other studies have found respiratory disease and cardiovascular disease highly related to mortality after SCI.16, 17

Taken together, these studies have identified several risk factors for excess mortality. Nevertheless, there are limitations in the scope of risk and protective factors, and study methodologies. Specifically, the last comprehensive report of the SCIMS data was 2004, and the sample size has been substantially augmented since that time, creating an opportunity for inclusion of additional predictors.

Purpose and hypotheses

Our purpose was to utilize the TRPM to structure analysis of multiple risk and protective factors for excess mortality, including three health factors not previously investigated. There were several study enhancements including an increase in the sample size, use of repeated measured predictors, a refined measure of participants’ economic status and the addition of the three secondary health condition predictors. Our key hypotheses were:

  1. 1

    Three newly added secondary health conditions and the new economic status measure will relate to excess mortality.

  2. 2

    Socio-environmental factors will mediate the association between mortality and injury severity.

  3. 3

    Health and secondary conditions will mediate the association between mortality and socio-environmental factors.

Materials and methods

Participants and procedures

The National Spinal Cord Injury Statistical Center Database contains data reported from SCIMS rehabilitation hospitals since 1973. Institutional review board approval was obtained locally at each center prior to data collection. Eligibility criteria included: admission within 1 year of injury, traumatic SCI, residence within the catchment area and discharge with some neurologic deficit. We excluded those <18 years old, as socio-environmental variables (income) are not measured by SCIMS for this group. The subset of data was restricted to those years during which the socio-environmental and health predictors were assessed (1995–2006). Twenty SCIMS hospitals contributed data to this study. There were a total of 8183 participants, in contrast with 5947 in the original study5 and 7331 utilized by Strauss et al.6 Among 8183 participants, 4773 had only one follow-up and the remaining 3410 had two or more follow-ups.

Measures

Mortality status was determined by routine follow-up at each SCIMS and by Social Security Death Index, with the most recent searches conducted in June 2011. We also used NDI search to confirm 340 participants’ mortality status. Participants not found deceased were presumed to be alive.

Predictors from the original 2004 study5 include: sex, age, race, neurologic level of injury, ASIA Impairment Scale (AIS) grade, ventilator dependency, marital status, four participation subscales from the Craig Handicap Assessment and Reporting Technique (CHART; physical independence, mobility, occupation and social integration), worker compensation insurance, interaction term between age and worker compensation insurance, self-perceived health (original 5-point scale was combined into three categories: poor, fair and good/excellent), health status compared to 1 year ago (worse/much worse versus others), hospitalization in the past year and pressure ulcer grade. A more complete description of these variables may be found in the original manuscript.5 The economic sufficiency scale was replaced by a single item on annual familial income using the following categories consistent with recent research:7 <$25 000, $25 000–$74 999 and ⩾$75 000. Three additional health variables were included based on a history of the following conditions within the 12 months prior to assessment: pneumonia, deep vein thrombosis (DVT) and kidney calculus.

Analysis

The data were analyzed by the logistic regression model on person–year observations.18, 19 We broke each individual’s event history into a set of separate observations, one for each year until death or censoring (the last available date known to be alive). For example, a person who participated in our study in 2000 and died in 2008 would contribute eight observations (or eight person–years). For each of these observations, we coded the outcome variable as 1 if deceased during that time unit, otherwise 0. The predictors were either time-invariant or time-variant (repeated measurement available). In our model, gender, race, injury level, AIS grade and ventilator usage were time-invariant predictors, whose values remained the same until censoring or deceased. All other predictors were time-variant in the model. They took on whatever value occurred during the time of measurement, and remained the same until the next measurement. We pooled these observations and estimated a logistic regression model by maximum likelihood.

Based on the TRPM, three logistic regression models were analyzed hierarchically. The first model included basic demographic and injury severity predictors. Socio-environmental factors were added to the second model, and health and secondary conditions were added to the final model. The generalized rescaled R2, a coefficient of determination, was used to measure how well we can predict mortality based on independent variables’ values. We cannot interpret this R2 as a proportion of variance explained, therefore it was used for comparison purpose only. As its upper bound is <1, it is possible its value is lower than the R2 of the linear regression model.20, 21 We calculated the proportion of concordant, discordant and tied pairs for each model.

Statement of ethics

We certify that all applicable institutional and governmental regulations concerning the ethical use of human volunteers were followed during the course of this research.

Results

Of the 8183 participants, 1381 deaths occurred during the 76 262 person–years of follow-up. The mean length of time between injury date and the date of entering the study was 7.7 years, and the average follow-up time was 9.3 years. Of the participants, 79.5% were male, 72.0% were white, 50.6% had a cervical injury and 49.4% were neurologically complete injuries.

Univariate analysis

The mortality rates of those with each of the health conditions were higher than those who did not have the condition (Table 1). For instance, the mortality rate of those with pneumonia in the previous year was 46.1 compared with 17.3 for those who did not report pneumonia (per 1000 person–years).

Table 1 Characteristics of participants by mortality status
Full size table

Multivariate logistic regression

There was a modest increase in the percentage of concordant pairs and the generalized rescaled R2 with each stage in the analysis (Table 2). The percentage of concordant pairs increased from 71.4% for model 1 to 76% for model 3. A generalized rescaled R2 increased from.085 to 118, indicating a greater model fit at each stage.

Table 2 Predictive power of each logistic regression model
Full size table

All predictors were significant in the first stage (Table 3). Participants who were male, older and Black had a higher level of injury, a more severe injury or were ventilator-dependent at discharge, were more likely to be deceased at follow-up.

Table 3 Three-stage multivariate logistic regression models
Full size table

In the second stage, several new variables were significant including each of the four scales from the CHART, family income and marital status (Table 3). Worker’s compensation insurance and the interaction term between age and worker’s compensation insurance were not statistically significant, and were eliminated from the analysis, as was race. The odds ratios for several indicators of injury severity decreased after the inclusion of the new set of predictors. For instance, the odds ratio of ventilator dependency decreased from 2.73 to 2.39, a 13% change (that is, ((2.39−2.73)/2.73) × 100). The odds ratios of C1–4 level injury, C5–8 level injury, AIS A injury, AIS B injury and AIS C injury declined by 27, 15, 21, 18 and 14%, respectively.

Several health variables were significant in the third stage, including each of those used in previous analyses. Those who had pneumonia during the past year had 1.33 times greater odds of dying. Kidney calculus was marginally significant (P=0.05), as the odds of dying were 1.30 times greater among those reporting the condition. However, DVT was not statistically significant and was eliminated from the final model. The odds ratios of several variables dropped further after the addition of the health measures. For instance, the odds ratios of ventilator dependency and AIS A injury declined by 5 and 6%, respectively. Odds ratios decreased for physical independence by 7%, for mobility by 10%, and for low income by 7%.

Discussion

This study systematically builds upon earlier research using data from the SCIMS to predict excessive mortality by evaluating several predictors in relation to mortality. After using repeated measured variables, our results reaffirm and expand upon findings from previous studies.5, 6 There was evidence to at least partially support each of the three study hypotheses.

The first study hypothesis was partially confirmed, as two of the three new secondary condition predictors (pneumonia and kidney calculus) were significant risk. The presence of either condition should therefore be taken as a serious risk factor for future mortality. However, a history of DVT was not significantly related to future mortality. This apparent inconsistency with the well-established literature22 no doubt relates to the type of analysis. Without question, DVT causes excess mortality in SCI. However, the current findings suggest that this relationship may be acute and that, once successfully treated, having a history of DVT does not necessarily indicate risk of excess future mortality. Alternatively, there were only 222 DVT recorded, therefore the number of deaths was limited.

The decrease in odds ratios with the addition of socio-environmental and health predictors provides confirmation for a mediating effect of each set of variables, consistent with the TRPM and hypotheses 2 and 3. The decrease in odds ratios for injury severity variables after introduction of socio-environmental factors was particularly striking, although this no doubt reflects, to some degree, the nature of two of the participation scales (independence and mobility) and their relationship with injury severity. Smaller, yet notable, additional decreases in odds ratios were observed after introducing health and secondary conditions, which also is consistent with the TRPM. In essence, this indicates that more sophisticated and more immediate predictors of mortality take on greater importance and help to explain the observed relationships between injury severity and mortality.

There are several important clinical implications of the findings related to both socio-environmental and health factors. Many of the health conditions that related to mortality are preventable, to varying degrees, with good health care and appropriate health maintenance behaviors. Clinicians should attempt to identify and prevent these conditions using multiple prevention approaches including, but not limited to, medical, educational and behavioral strategies. Furthermore, socio-environmental factors, such as low family income, may also become the focus of interventions. Owing to the well-documented association between unemployment and elevated risk of mortality, we need to be cautious about the possible unintended consequence led by policies providing health and financial disincentives to return to work and obtaining earned income. From a clinical perspective, enhancing participation may have the added benefit of enhancing longevity. Our results indicate the protective effect of marriage. However, marriage rates are low and divorce rates high among the spinal cord injured.23 This suggests the importance of interventions that enhance opportunities to develop interpersonal relationships and enhance overall participation.

Methodological considerations

First, we included more participants than the original studies using the subsample of SCIMS data and included more years of follow-up.5, 6 Second, we used repeated measures predictors (that is, time-dependent covariates). Third, we evaluated several new variables. Lastly, we used the more accurate NDI search to confirm mortality status for a group of participants.24, 25

There are also several study limitations. First, as all participants were enrolled at follow-up, there may be systematic differences between those who were and were not enrolled in the study. Second, the findings apply only to those who survived the first year. Third, time-dependent covariates were only available for 42% of the participants. Fourth, as SCIMS hospitals only record secondary health conditions occurring 12 months prior to assessment, we cannot assess complications happening outside of this window. Lastly, our data did not have psychological and behaviors factors available to test the mediational effects of behavioral factors suggested by the TRPM.

Future research

Additional research should address a wider array of predictive factors including more refined socio-environmental predictors, such as access to health care, and yet more diverse health factors. It should also include health behaviors, while maintaining the focus on modifiable risk and protective factors that may become the focus of intervention strategies. Lastly, there is a need to link risk and protective factors to specific causes of death, particularly those with elevated risk after SCI including septicemia, influenza and pneumonia.

Conclusion

Besides demographic and injury severity predictors, risk of mortality is related to social participation, income, general health and several secondary health conditions. A history of pressure ulcers, pneumonia and kidney calculus represents significant risk for excess mortality. Socio-environmental factors and health factors mediate the effect of demographic and injury severity on mortality.

Data archiving

There were no data to deposit.