Quantifying the collective influence of social determinants of health using conditional and cluster modeling

Objectives Our objective was to analyze the collective effect of social determinants of health (SDoH) on lumbar spine surgery outcomes utilizing two different statistical methods of combining variables. Methods This observational study analyzed data from the Quality Outcomes Database, a nationwide United States spine registry. Race/ethnicity, educational attainment, employment status, insurance payer, and gender were predictors of interest. We built two models to assess the collective influence of SDoH on outcomes following lumbar spine surgery—a stepwise model using each number of SDoH conditions present (0 of 5, 1 of 5, 2 of 5, etc) and a clustered subgroup model. Logistic regression analyses adjusted for age, multimorbidity, surgical indication, type of lumbar spine surgery, and surgical approach were performed to identify the odds of failing to demonstrate clinically meaningful improvements in disability, back pain, leg pain, quality of life, and patient satisfaction at 3- and 12-months following lumbar spine surgery. Results Stepwise modeling outperformed individual SDoH when 4 of 5 SDoH were present. Cluster modeling revealed 4 distinct subgroups. Disparities between the younger, minority, lower socioeconomic status and the younger, white, higher socioeconomic status subgroups were substantially wider compared to individual SDoH. Discussion Collective and cluster modeling of SDoH better predicted failure to demonstrate clinically meaningful improvements than individual SDoH in this cohort. Viewing social factors in aggregate rather than individually may offer more precise estimates of the impact of SDoH on outcomes.


Introduction
Internationally, lumbar spine surgery is typically reserved for individuals who have responded poorly to conservative care or have marked physiological degeneration that has resulted in very high levels of pain, disability and lower levels of function [1,2]. Because spine surgery also has higher incidences of harms and costs, a significant amount of effort has gone into modeling individuals who are good candidates for surgical intervention and conversely, those who are at risk for poor outcomes [3][4][5].
Scientists, clinicians and policy makers have recognized the influence of biopsychosocial factors on self-reported health outcomes [6]. Care pathways and risk stratification schemes commonly take into account biological and psychological factors [7,8], yet noticeably less attention has been paid to social factors such as social determinants of health (SDoH). SDoH are broadly defined as the conditions in which people are born, work, live and play and include areas such as economic stability, education, social and community context, and environment [9]. Because recovery from spine surgery can be upwards of 6 months and correspond to increased psychological distress and decreased activity [10][11][12][13], we believe that the importance of addressing SDoH in this population is heightened.
To date, only small scale studies have evaluated individuals' SDoH for spine surgery [4,[14][15][16][17][18], suggesting that these factors do individually influence outcomes. However, SDoH variables do not routinely exist in singularity. What remains unknown is the collective impact of SDoH on 3-and 12-month outcomes following lumbar spinal surgery. The goals of this study were to analyze the collective effect of SDoH on lumbar spine surgery outcomes utilizing two different statistical methods of combining variables. The findings will provide a better understanding of the role of SDoH, and will outline which method of statistical analysis defines a clearer picture of the role of the impact of SDoH on outcomes post lumbar surgery.

Study design
This study was an observational design utilizing a retrospective review of a lumbar spine database from the Quality Outcomes Database (QOD). The QOD is a prospective registry established to define risk-adjusted morbidity and 12-month clinical outcomes following common surgical spine procedures [19,20]. The registry has been enrolling patients since 2012 from 74 sites across 26 US states [20]. This study protocol was approved by the Duke University Institutional Review Board (Pro00029554) and adheres to the Reporting of studies Conducted using Observational Routinely collected Data (RECORD) guidelines [21].

Participants
Patients aged 18 or older with degenerative disorders (stenosis, spondylolisthesis, disc herniation, scoliosis, kyphosis, or pseudarthrosis) who received a primary lumbar spine surgery (laminectomy, arthrodesis, osteotomy, corpectomy, interbody graft) were eligible for inclusion. Patients who received revision surgery and those who have reported baseline outcome Success thresholds were calculated from the change between baseline and each time point. Thresholds were defined for NRS-BP (1.2 points), NRS-LP (1.6 points), and ODI (12.8 points), using minimal clinically important difference (MCID) values previously reported [34]. To date, no MCID has been reported for the EQ-5D VAS. In this absence, we chose to use median change values (11 points for 3 months and 10 points for 12 months). Success in patient satisfaction was defined as either "Surgery met my expectations" or "I did not improve as much as I had hoped, but I would undergo the same operation for the same results."

Cohort derivation and missing data
Little's missing completely at random (MCAR) test was employed for each variable and suggested that the data were not missing completely at random [35]. Methods for dealing with missing data can include listwise deletion and multiple imputation [36,37]. Because the missing data were present in high-stakes variables, we chose to use listwise deletion to remove missing values in which an entire record is excluded from analysis if any single value is missing.

Statistical analysis
Descriptive statistics were performed to assess differences in baseline variables utilizing linear mixed-effects modeling for continuous variables and Chi-square test for categorical variables [38].
Bivariate analyses. We ran independent analyses for each SDoH variable and each outcome variable. Age, multimorbidity (defined as 2 or more comorbid conditions) [39], surgical indication (spondylolisthesis, disc herniation, stenosis, scoliosis, kyphosis), type of surgery (laminectomy, arthrodesis, osteotomy, corpectomy, interbody graft), surgical approach (posterior only, anterior only, lateral only, two stage) and baseline outcome score were used as covariates as in similar studies [40][41][42]. The strength of association between the independent and dependent variables was expressed with adjusted odds ratios (ORs) with 95% confidence intervals (CIs) and Nagelkerke pseudo R-squared values that reflect the predictive power of the model [42,43]. In our study, ORs above 1.0 indicated the likelihood of not meeting the MCID whereas ORs below 1.0 reflected the likelihood of meeting the MCID. The percentage of participants meeting each condition variable was calculated.
Statistical modeling method 1-Stepwise regression. To determine the associations between collective SDoH we utilized binary logistic regression between conditions of 0 of 5, 1 of 5, 2 of 5, 3 of 5, 4 of 5, and 5 of 5 SDoH and lumbar spine surgical outcomes adjusted for age, multimorbidity, surgical indication, type of surgery, surgical approach, and baseline outcome score as in the bivariate analyses. We converted the inverse of the odds ratios to probabilities of 100 patients achieving success and calculated the difference between individuals with 0 of 5 SDoH conditions and the 4 of 5 SDoH conditions. Statistical modeling method 2-Cluster analysis. To better understand patterns of SDoH, we utilized a two-step cluster analysis to subgroup patients based upon the SDoH variables. Cluster analysis identifies homogenous subgroups who have similar characteristics where the grouping is not previously known [44]. The two-step cluster analysis first identifies groupings by pre-clustering based on dense regions in the attribute-space, then merges them using hierarchical methods [44]. We utilized the Bayesian Information Criterion (BIC) to determine the appropriate number of clusters that was based on the lowest BIC and the largest BIC change between the number of clusters [44]. Silhouette coefficients were used to appraise cluster solution quality with less than 0.2 classified as poor; between 0.2 and 0.5 as fair; greater than 0.5 as good solution quality. We considered good solution quality as acceptable clustering [44]. Two-step clustering has been regarded as a reliable and reproducible way to classify subgroups of individuals [45,46].
We dummy coded each cluster and utilized binary logistic regression modeling to measure the associations between each clustered subgroup and lumbar spine surgical outcomes as in method 1. We converted the inverse of the odds ratios to probabilities of 100 patients achieving success and calculated the difference between subgroups. Significance was set at p < 0.05 and analyses were performed using R (R Foundation for Statistical Computing, Vienna, Austria; version 4.0.2) including the 'rms' package [47] and SPSS version 25.0 (IBM Corporation, Armonk, NY).
Sensitivity analyses. We performed sensitivity analyses (S1 Appendix) with missing values multiply-imputed using a flexible additive imputation model with predictive mean matching for missing values (n = 32,573). This method of imputation takes all aspects of uncertainty in the imputations into account by using the bootstrap to approximate the process of drawing predicted values from a full Bayesian predictive distribution [48]. Predictive mean matching works for binary, categorical, and continuous variables without the need for iterative maximum likelihood fitting for binary and categorical variables, and without the need for computing residuals or for curtailing imputed values to be in the range of actual data [48].

Results
Of the 8,977 individuals included in this study, 7,448 (83.0%) had SDoH whereas 1529 (17.0%) had none (Fig 1). Three thousand nine hundred and fifty-nine (44.1%) had two SDoH, 937 (10.4%) had three, 172 (1.9%) had four, and only 16 (0.2%) had five SDoH factors (S1 Table). Clustering identified four distinct subgroups: 1) older, white, female (OWF; n = 2249, 25.1%), 2) older, white, male (OWM; n = 2066, 23.0%) 3) younger, minority, low socioeconomic status (YML; n = 1952, 21.7%), and 4) younger, white, high socioeconomic status (YWH; n = 2710, 30.2%) with good cluster quality (average silhouette = 0.6). The overall trend was that the YML group had more pre-operative pain, disability, and comorbid conditions and lower QoL compared to the other groups. The YML cluster generally differed from the other groups in terms of level of education and insurance payer (Medicaid), but was similar in terms of baseline symptoms. However, those in the YML cluster were more likely to be smokers and have a higher BMI and COPD. Pre-operative characteristics of the four subgroups are described in Table 1.

Individual SDoH
S2 Table presents the results of binary logistic regressions and outlines associations between each SDoH and failure to achieve success at 3 months. S3 Table presents similar results for 12-month outcomes. Statistically significant associations were noted between each SDoH and each outcome with the exception of gender. Overall, educational attainment, insurance type, and employment status were the strongest predictors of outcomes at 3 and 12 months. Sensitivity analyses revealed no substantial changes in 3-month outcomes (Table A in S1 Appendix) or 12-month outcomes ( Table B in S1 Appendix).
Stepwise modeling S4 Table presents the conditional binary logistic regression results adjusted for age, multimorbidity, surgical indication, type of surgery, and baseline outcome score at 3 months. S5 Table  displays similar results for 12-month outcomes. Statistically significant associations were noted for all outcomes at each time point. Overall, an additive effect for SDoH was observed across all outcome variables at 3-and 12-months post-surgery where the odds of failing to demonstrate success increased as more SDoH were present (S6 Table). The widest differences in outcomes were noted in patient satisfaction followed by leg pain and disability. S1 and S2 Figs show the difference in probability of 100 persons with 0 of 5 SDoH variables compared to those with 4 of 5 SDoH variables having success in each outcome at 3 and 12 months, respectively. Compared to those with 4/5 SDoH variables, between 19 and 31 more individuals with 0/5 SDoH variables out of 100 will experience success after lumbar spine surgery. Table 2 displays binary logistic regression results adjusted for age, multimorbidity, surgical indication, type of surgery, and baseline outcome score for each clustered subgroup. Table 3 presents similar findings for 12-month outcomes. S7 Table presents the mean baseline, 3-month, and 12-month outcomes by cluster. The OWF and OWM subgroups did not have statistically significant differences at 3 and 12 months. The YML subgroup demonstrated increased odds of failing to achieve success in each outcome at 3 and 12 months. In contrast, the YWH subgroup demonstrated decreased odds of failing to achieve success in each outcome at 3 and 12 months. The widest differences in outcomes were noted in back pain, leg pain, and disability. Figs 2 and 3 represent the differences between the probability of 100 persons in the YML compared to the YWH subgroup having success in each outcome at 3 and 12 months, respectively. Compared to individuals in the YML subgroup, between 21 and 27 more individuals from the YMH subgroup out of 100 will experience success after lumbar spine surgery.

Discussion
This study analyzed the collective effect of SDoH on lumbar spine surgery outcomes by analyzing two different statistical methods of combining variables. We targeted individuals undergoing primary lumbar surgery in the hopes of homogenizing the patient population. When controlled for numerous covariates, across both types of modeling the presence of SDoH at baseline was associated with reduced success in improving in back pain, leg pain, disability, quality of life, and satisfaction at 3 and 12-month follow-up. These findings support the integration of SDoH for predictive modeling when determining prognosis following spine surgery. Interestingly, the findings associated with a collective effect when more than one SDoH variable was present was less definitive. To our knowledge, this is the first study to examine 2 methods of modeling the collective impact of SDoH for any musculoskeletal disorder. Because social factors influence health in complex and interrelated ways [49], we elected to investigate two distinct methods of modeling the collective impact of SDoH for spine surgery. Whereas the stepwise regression modeling revealed an additive effect of SDoH (where each additional factor generally increased the odds of failing to demonstrate clinical improvement), it did not substantially outperform individual SDoH factors in predictive ability until 4 of 5 conditions were present. The clinical utility of this finding is limited since the number of patients with 4 of 5 of the measures SDoH represented only 1.9% of the overall sample. The cluster modeling yielded intriguing results. The two-step cluster modeling identified four distinct patterns of SDoH: 1) OWF, 2) OWM, 3) YML, and 4) YWH. Sociodemographic, clinical, and comorbidity variables each differed by group allocation suggesting unique socialbiological phenotypes. The differences observed between the YML and YWH subgroups were the most profound among the subgroups, especially with patient satisfaction, which exhibited the widest variation in success probability. The influence of various SDoH such as insurance and race on patient satisfaction has been previously documented in the surgical literature [17,50]. The disparities seen in pain and disability have not previously been observed and begin to justify the need for more robust methods of quantifying the relationships between various SDoH [4,51]. Overall, the cluster analysis produced subgroups with clearly defined characteristics that may be useful in clinical practice (Fig 4).
Lastly, the findings from this study shed light on potential care pathway structures for those who present with SDoH. Routine pre-operative screening for SDoH should be required to appropriately support patients. [52,53] If a patient is a plausible candidate for surgery but has at least 3 of 5 SDoH variables implying risk or if the SDoH variables match the YML cluster identified, increased pre-and post-surgical community support may assist in mitigating the disparities observed in this study and optimize the risk benefit ratio in the patient's favor. Prior studies have identified increased referral to and use of wraparound services including clinical team members or behavioral health when such pathways are implemented. [54,55] Addressing SDoH in risk stratification models and care pathways is an important step toward improving the equity of outcomes from spine surgery.

Limitations
This study is limited by its use of observational data in which cause and effect cannot be implied. Another limitation was the missing data present in the QOD. However, these missing data were handled through listwise deletion which is an acceptable procedure [56]. The definition of success was chosen based on standard MCID measures, but to date there are not universally agreed-on MCID values for all outcome measures [57]. The performance of the models based upon Nagelkerke pseudo R-squared value was modest to good with an explained variance of 1 to 38 percent. However, the utility of this measure in large behavior-based datasets has been called into question [58,59]. Predictive models may be useful to guide clinician behavior even if the variance explained by the model is low. Finally, the 3-and 12-month time points are relatively short-term follow ups and the influence of SDoH at long-term time points remains unknown.
In this study, the predictor variable was developed by collapsing five variables-race/ethnicity, educational attainment, employment, insurance payer, and gender. Other social factors known to be associated with musculoskeletal disorders including income and place of residence were not available in the QOD [60,61]. These data are especially important in light of recent work indicating that outcomes following microdiscectomy could not be accurately predicted by commonly captured sociodemographic variables [62]. It is unknown how including additional SDoH would affect the present results. Still, the authors hypothesize that the inclusion of additional SDoH variables would increase the precision and magnitude of the association between SDoH and clinical outcomes following spine surgery.

Conclusion
The present study suggests that, in aggregate, SDoH predict failure to achieve success in pain, disability, quality of life, and satisfaction at 3-and 12-month follow-up time points following lumbar spinal surgery. Validation of these models in other populations with musculoskeletal disorders including robust markers of SDoH is warranted.  Table. Association between presence of SDoH at baseline and failing to achieve clinically meaningful improvement on outcome at 3 months. (DOCX) S3 Table. Association between presence of SDoH at baseline and failing to achieve clinically meaningful improvement on outcome at 12 months. (DOCX) S4 Table. Association between presence of SDoH at baseline and failing to achieve clinically meaningful improvement on outcome at 3 months. (DOCX) S5 Table. Association between presence of SDoH at baseline and failing to achieve clinically meaningful improvement on outcome at 12 months. (DOCX) S6 Table. Baseline, 3-month, and 12-month outcomes for each SDoH condition. (DOCX) S7 Table. Baseline, 3-month, and 12-month outcomes for total sample and each cluster.