RESEARCH ARTICLE
A Preliminary Risk Stratification Model for Individuals with Neck Pain Chad Cook1* PT, PhD, Jason Rodeghero2 PT, PhD, Joshua Cleland3 PT, PhD & Paul Mintken4 PT, DPT 1
Duke University, Durham, NC, USA OSF Saint James – John W. Albrecht Medical Center, Pontiac, IL, USA
2 3
Franklin Pierce University, Manchester, NH, USA
4
University of ColoradoSchool of Medicine, Aurora, CO, USA
Abstract Introduction. The aim of the present study was to identify predictive characteristics related to patients with neck impairments who have a high risk of a poor prognosis (lowest functional recovery compared to visit utilization) as well as those who are at low risk of a poor prognosis (highest functional recovery compared to visit utilization). Methods. A retrospective cohort of 3,137 patients with neck pain who were seen for physiotherapy care was included in the study. All patients were seen at physiotherapy clinics in the United States and were provided with care in a manner in which the physiotherapists felt was appropriate and necessary. Univariate and multivariate multinomial regression analyses were used to identify significant patient characteristics predictive of treatment response. Results. Statistically significant predictors of high-risk categorization included longer duration of symptoms, surgical history and lower comparative levels of disability at baseline. Statistically significant predictors of low-risk categorization were younger age, shorter duration of symptoms, no surgical history, fewer comorbidities and higher comparative disability levels of function at baseline. Discussion. Few studies have analysed risk stratification models for neck pain, and the findings of the present study suggest that predictors of poor success are similar to those in most musculoskeletal prognostic models. Limitations of the study included those inherent in secondary analysis and the inability to identify the diagnoses of the patients. Conclusions. Future research should continue to examine the variables predictive of treatment response in patients with neck pain. Copyright © 2015 John Wiley & Sons, Ltd. Keywords Risk stratification; neck pain; physical therapy; prognosis *Correspondence Chad Cook, 2200 W Main Street, Duke University, Durham, NC, 27708, USA. Tel: +1 919 684 8905; Fax: +1 919 684 1846. Email: [email protected]
Published online 2 March 2015 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/msc.1098
Introduction In the United States, low back- and neck-related pain was responsible for nearly $86 billion in incremental healthcare costs in 2005 (Martin et al. 2008). Approximately 36% of these costs were associated with outpatient-oriented care (Martin et al. 2008). In 2008,
Musculoskelet. Care 13 (2015) 169–178 © 2015 John Wiley & Sons, Ltd.
nearly 13.6 million adults had an ambulatory visit for spinal care, an increase of 15% over a nine-year period (Davis et al. 2012). In both 2005 and 2008, there were notable increases in costs for nearly every provider type, including physiotherapy services (cost increases of 11% from 1997 to 2005) (Martin et al. 2008). The
169
Cook et al.
Risk Stratification and Neck Pain
mean adjusted expenditure for physiotherapy services in 2008 was $1,543, although costs varied notably across different neck-related cases (Davis et al. 2012). Unlike for low back pain, there are no multidiscipline-endorsed guidelines for the treatment of neck pain. This may be one of the reasons for the extensive variability in the provision of care for spine pain among physiotherapists (Gellhorn et al. 2012), with care for neck pain often consisting of interventions that lack evidence for clinical use (Goode et al. 2010). Identifying optimal neck pain treatment strategies that provide the most cost-effective intervention has been recognized as a priority for physiotherapists (Korthals-de Bos et al. 2003). One option that is designed to direct treatment interventions based on outcomes is prognostic risk stratification (PRS). PRS modelling is a strategy involving the identification of patients during the initial diagnosis who are either at risk for a poor outcome or who are likely to have a very good outcome regardless of the intensity of care received (Shanafelt et al. 2004). PRS methods have the capacity to direct care to those who truly need an intervention and are likely to respond to it, and withhold unnecessary care to those who are unlikely to respond, or simply do not need formal interventions. Past studies have shown that limiting care to the most appropriate individuals is not only a useful mechanism for controlling costs (Hill et al. 2008; Foster et al. 2014), but also has the capacity to generate similar outcomes to those achieved using usual care (Rethnam et al. 2008; Srinivas et al. 2012). PRS modelling requires very large sample sizes in order to include the population characteristics represented and to enable appropriate precision in the statistical analysis (Shanafelt et al. 2004). PRS modelling methods must focus on variables that have been shown to influence outcomes in clinical practice. For example, longer pain durations (Skargren and Oberg 1998; Hill et al. 2008), higher levels of fear of movements (Hill et al. 2008), poorer report of quality of life (Hill et al. 2008) and catastrophizing behaviours (Hill et al. 2008) are just some of the variables that have been related to poorer overall outcomes in individuals with mechanical neck pain in studies with small to moderately sized samples. Individuals with whiplashassociated disorder with higher general health questionnaire scores, widespread pain throughout the body and higher initial disability scores demonstrated a risk of persistent neck pain (Atherton et al. 2006). 170
To our knowledge, no studies have used PRS modelling for individuals with neck pain managed by physical therapists and no studies have evaluated prognostic factors in large sample sizes, such as those found in repositories or outcomes databases. Large outcomes databases offer an opportunity to explore trends well beyond those possible within a clinical trial or a traditional observational study (Lohr 2012). Large datasets often contain variables that are consistent across all databases, which improves the transferability of the predictive findings. The objective of the present study was to identify a preliminary set of variables, through PRS, that identified individuals who were either at risk for a poor outcome or were likely to have a very good outcome regardless of the intensity of care received. The preliminary findings could serve as an initial investigation toward validating a risk stratification tool for targeting specific treatment intensities for individuals with neck pain seen by physiotherapists.
Methods Study design The study involved retrospective analyses of outcomes data from a clinical cohort of patients with neck pain who were formally seen for physiotherapistadministered care. It was approved by the Institutional Review Board at Franklin Pierce University, Manchester, NH, USA. Patient population A de-identified patient outcomes dataset, representing one full year of data collection (June 2012 to June 2013) from multiple physiotherapy clinics, was provided from Focus On Therapeutic Outcomes, Inc. (FOTO), an international medical rehabilitation outcomes database management company (Knoxville, TN, USA). All patients aged at least 18 years, with musculoskeletal impairments (all body regions) and complete data (intake scores, discharge scores and number of physiotherapy visits) were included in the initial dataset. Those with a primary diagnosis associated with cervical pain only were included (N = 3,137). Personal patient information was removed and each patient was assigned a patient identification number. All patients received care in outpatient physiotherapy clinics that used FOTO for collecting/measuring patient outcomes. Musculoskelet. Care 13 (2015) 169–178 © 2015 John Wiley & Sons, Ltd.
Cook et al.
Clinical data collection process Clinical data collection was longitudinal and consisted of data capture at baseline, during progress visits and at the time of discharge. During admission to physiotherapy, patients entered demographic data and completed self-report surveys using Patient Inquiry®, a computer program developed by FOTO. Patients completed these surveys throughout the episode of care and at discharge, depending on his/her clinicians’ case management strategies. Demographic data included age, state of residency and gender. Case-related data included baseline levels of pain, disability and fear avoidance beliefs; duration of symptoms; exercise status; surgical history; payer type, comorbidities and medication use.
Variables used in the modelling Our goal was to include variables in the PRS modelling that were available in most commercial or clinically available datasets. This required the inclusion of standard demographic and patient-reported outcomes data. We also included data on comorbidities, although there is a possibility that these variables may not have been captured in all datasets. Within the dataset, patient age and intake functional status (FS) score (intake FS score was initially rated 0– 100%, with 100% being full function) were coded as continuous variables, whereas the rest of the variables were treated as categorical variables. The FOTO FS survey to determine patient-reported functional level has been previously researched, with reliability and validity established (Hart et al. 2006, 2011). The FS scale is a self-reported measure administered using computeradaptive testing, which is designed to reflect the most appropriate item selections for each individual patient. Each item is queried on a five-point scale (0–4), with selections including ‘no difficulty’ = 4; ‘little difficulty’ = 3; ‘some difficulty’ = 2; ‘much difficulty’ = 1; and ‘I can’t do this’ = 0, and ‘didn’t do before’ = N/A. Symptom acuity, operationally defined as the number of calendar days from the date of onset of the condition being treated in therapy to the date of initial therapy evaluation, was categorized as acute (90 days). Surgical history was categorized into five categories (none, one, two, three, four or more surgical interventions related to the impairment being treated). Payer type was Musculoskelet. Care 13 (2015) 169–178 © 2015 John Wiley & Sons, Ltd.
Risk Stratification and Neck Pain
categorized into 16 payer sources (e.g. health maintenance organization, preferred provider organization). Number of functional comorbid conditions was assessed using a list of 30 conditions common to patients entering an outpatient rehabilitation clinic (e.g. heart attack, sleep disturbance, overweight and cancer). Exercise history prior to receiving therapy was categorized as exercising three times a week or more, 1–2 times a week, or seldom or never. Use of medication at intake was a dichotomous response (yes, no).
Recoding strategy As stated, the majority of codes within the dataset were categorical but involved multiple categorical options. As a research team, we opted to recategorize variables in an attempt to dichotomize the variables based on clinical sensibility. Clinical sensibility is the concept that the recategorized variables reflect findings typically encountered or derived from clinical practice. For example, although multiple categories were present for ‘duration of symptoms’, we dichotomized the variable into 22 days (sub-acute or chronic). Percentage change for function was calculated by taking the difference in the functional score (from baseline to discharge), dividing it by the baseline score and then multiplying by 100. The end product was a positive or negative, percentage change. Others have advocated using percentage change from baseline to determine improvement in outcomes (Dworkin et al. 2008). An example includes the Initiative on Methods, Measurement, and Pain Assessment in Clinical Trials (IMMPACT), which recommended a 30% reduction in pain from baseline as a lower threshold of success, with a 50% reduction a substantially clinically important change (Dworkin et al. 2008). Surgical history was dichotomized into ‘none’ (no) and ‘1 or greater’ (yes). Exercise status was dichotomized into 2 times a week (moderate or more). Multiple forms of payer type were represented and we endeavoured to categorize by similarity. Payer type was clustered into a) automotive, litigation and worker’s compensation; b) Medicare (all types) and Medicaid (based on care entitlement of use in the United States) and c) all others. As age has no obvious clinically sensible categories, it was retained in a continuous format. Lastly, patients were classified into high (elevated) or low 171
Cook et al.
Risk Stratification and Neck Pain
(non-elevated) fear avoidance beliefs, based on the screening items developed by Hart et al. (2011).
Risk stratification categorization PRS categorization can occur through many methods (Miller et al. 2001). Different processes may include providing a weighted score assigned for each criteria, a priori creation of risk categories or creation of risk categories based on management options (Miller et al. 2001). We elected to create a risk category based on management options – in other words, based on the outcomes and intensity of services provided (with intensity operationally reflecting the number of physiotherapy visits provided for the same condition). Our goal was to define who had a very good functional outcome with a low intensity of services and who had a poor outcome despite highly intensive services. Because we planned on creating a variable that incorporated the two unique elements of a functional outcome and intensity of services, it was not appropriate to analyse quartiles, standard deviations and other methods. As stated, we created three new categories by analysing the interplay between functional change (outcome) and total number of physiotherapy visits (intensity of services). In particular, we were interested in identifying individuals in each tail of the distribution: a) a group reflecting excellent functional outcome (by percentage change) with a minimal total number of physiotherapy visits and b) a group representative of a poor overall functional outcome despite multiple physiotherapy visits. After several iterations (e.g. 20%, 25%.....40%) of combining percentage distributions for FS outcomes and visit utilization, we identified that the highest 35% for visit utilization and the bottom 35% for functional outcome created the most notably remarkable tails. If individuals demonstrated both conditions they were uniquely coded as a high-risk (poor outcome) group. By contrast, to represent the low-risk group, we captured the highest 35% for functional percentage recovery and the lowest 35% for visit utilization. If individuals met both conditions, they were uniquely coded as a low-risk (good outcome) group. For completion of coding, we created a single code trichotomized into three categories: a) high-risk (low outcome, high number of physiotherapy visits); b) medium-risk (all individuals not in high- or low-risk categories) and c) low-risk (high outcome, low number 172
of physiotherapy visits). There were no instances in which one person was coded in both categories. After categorizing the groups, we found that the high-risk group comprised 17.9% of the total sample, whereas the low-risk group comprised 9.0% of the sample. Our medium-risk group comprised the majority of the total sample, at 73.1%.
Determining the appropriate number of observations per variable Homer and Lemeshow (2000) recommended that the minimum observation-to-variable ratio for a univariate multinomial or logistic regression is ten but cautioned that this recommendation is likely to lead to the overfitting of a model. With our sample size of greater than 3,100, and our current selection of ten variables, we were in no danger of overfitting the model.
Data analysis All analyses were performed using Statistical Package for the Social Sciences, version 22.0 (SPSS Inc., Chicago, IL, USA). Descriptive statistics for the full sample were calculated, including the values associated with the total number of physiotherapy visits and total episode duration. Continuous variables were represented by original means and standard deviations, whereas categorical variables were presented by frequencies using the recoded values. Differences among the high-, medium- and low-risk groups were calculated using chi-square and analysis of variance tests where appropriate. Univariate multinomial regression analyses were performed for each of the predictor variables for the trichotomized risk stratification variable. The referent variable used was medium-risk stratification. Multinomial logistic regression is used to predict the probability of category membership for a dependent variable with two or greater classifications, based on multiple independent variables (Domínguez-Almendros et al. 2011). Multinomial logistic regression analyses use maximum likelihood estimation to estimate model parameters. Selection of independent variables was performed manually to identify the strongest model. Multicollinearity for each the ten independent variables was evaluated by analysing correlation matrixes. According to Cohen (1988), correlational coefficients of the order of 0.10 are small, those of 0.30 are medium Musculoskelet. Care 13 (2015) 169–178 © 2015 John Wiley & Sons, Ltd.
Musculoskelet. Care 13 (2015) 169–178 © 2015 John Wiley & Sons, Ltd.
1,941=12–24 score 10.11 (7.20) 47.15 (50.71)
Total number of visits Total episode duration (days)
15.12 (8.42) 67.07 (50.29)
54.10 (15.27) 195 = Male 365 = Female 4.81 (3.07) 76 = 2 days a week 314 = 2 or less per week 70 = Group A 159 = Group B 331 = Group C 284 = Yes 276 = No 55.69 (14.03) 214 = 0–11 score 346 = 12–24 score
High risk (N = 560; 17.9%) frequency
9.65 (6.53) 45.56 (51.05)
53.50 (15.52) 841 = Male 1,453 = Female 4.84 (3.27) 490 = 2 days a week 1,380 = 2 or less per week 273 = Group A 617 = Group B 1,404 = Group C 1,231 = Yes 1,063 = No 51.31 (12.95) 871 = 0–11 score 1,423 = 12–24 score
Medium-risk (N = 2294; 73.1%) frequency
3.94 (1.00) 20.60 (29.52)
50.40 (16.80) 88 = Male 195 = Female 4.41 (2.62) 120 = 2 days a week 165 = 2 or less per week 34 = Group A 72 = Group B 177 = Group C 167 = Yes 116 = No 49.07 (9.70) 111 = 0–11 score 172 = 12–24 score
Low-risk (N = 283; 9.0%) frequency
A Preliminary Risk Stratification Model for Individuals with Neck Pain Chad Cook1* PT, PhD, Jason Rodeghero2 PT, PhD, Joshua Cleland3 PT, PhD & Paul Mintken4 PT, DPT 1
Duke University, Durham, NC, USA OSF Saint James – John W. Albrecht Medical Center, Pontiac, IL, USA
2 3
Franklin Pierce University, Manchester, NH, USA
4
University of ColoradoSchool of Medicine, Aurora, CO, USA
Abstract Introduction. The aim of the present study was to identify predictive characteristics related to patients with neck impairments who have a high risk of a poor prognosis (lowest functional recovery compared to visit utilization) as well as those who are at low risk of a poor prognosis (highest functional recovery compared to visit utilization). Methods. A retrospective cohort of 3,137 patients with neck pain who were seen for physiotherapy care was included in the study. All patients were seen at physiotherapy clinics in the United States and were provided with care in a manner in which the physiotherapists felt was appropriate and necessary. Univariate and multivariate multinomial regression analyses were used to identify significant patient characteristics predictive of treatment response. Results. Statistically significant predictors of high-risk categorization included longer duration of symptoms, surgical history and lower comparative levels of disability at baseline. Statistically significant predictors of low-risk categorization were younger age, shorter duration of symptoms, no surgical history, fewer comorbidities and higher comparative disability levels of function at baseline. Discussion. Few studies have analysed risk stratification models for neck pain, and the findings of the present study suggest that predictors of poor success are similar to those in most musculoskeletal prognostic models. Limitations of the study included those inherent in secondary analysis and the inability to identify the diagnoses of the patients. Conclusions. Future research should continue to examine the variables predictive of treatment response in patients with neck pain. Copyright © 2015 John Wiley & Sons, Ltd. Keywords Risk stratification; neck pain; physical therapy; prognosis *Correspondence Chad Cook, 2200 W Main Street, Duke University, Durham, NC, 27708, USA. Tel: +1 919 684 8905; Fax: +1 919 684 1846. Email: [email protected]
Published online 2 March 2015 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/msc.1098
Introduction In the United States, low back- and neck-related pain was responsible for nearly $86 billion in incremental healthcare costs in 2005 (Martin et al. 2008). Approximately 36% of these costs were associated with outpatient-oriented care (Martin et al. 2008). In 2008,
Musculoskelet. Care 13 (2015) 169–178 © 2015 John Wiley & Sons, Ltd.
nearly 13.6 million adults had an ambulatory visit for spinal care, an increase of 15% over a nine-year period (Davis et al. 2012). In both 2005 and 2008, there were notable increases in costs for nearly every provider type, including physiotherapy services (cost increases of 11% from 1997 to 2005) (Martin et al. 2008). The
169
Cook et al.
Risk Stratification and Neck Pain
mean adjusted expenditure for physiotherapy services in 2008 was $1,543, although costs varied notably across different neck-related cases (Davis et al. 2012). Unlike for low back pain, there are no multidiscipline-endorsed guidelines for the treatment of neck pain. This may be one of the reasons for the extensive variability in the provision of care for spine pain among physiotherapists (Gellhorn et al. 2012), with care for neck pain often consisting of interventions that lack evidence for clinical use (Goode et al. 2010). Identifying optimal neck pain treatment strategies that provide the most cost-effective intervention has been recognized as a priority for physiotherapists (Korthals-de Bos et al. 2003). One option that is designed to direct treatment interventions based on outcomes is prognostic risk stratification (PRS). PRS modelling is a strategy involving the identification of patients during the initial diagnosis who are either at risk for a poor outcome or who are likely to have a very good outcome regardless of the intensity of care received (Shanafelt et al. 2004). PRS methods have the capacity to direct care to those who truly need an intervention and are likely to respond to it, and withhold unnecessary care to those who are unlikely to respond, or simply do not need formal interventions. Past studies have shown that limiting care to the most appropriate individuals is not only a useful mechanism for controlling costs (Hill et al. 2008; Foster et al. 2014), but also has the capacity to generate similar outcomes to those achieved using usual care (Rethnam et al. 2008; Srinivas et al. 2012). PRS modelling requires very large sample sizes in order to include the population characteristics represented and to enable appropriate precision in the statistical analysis (Shanafelt et al. 2004). PRS modelling methods must focus on variables that have been shown to influence outcomes in clinical practice. For example, longer pain durations (Skargren and Oberg 1998; Hill et al. 2008), higher levels of fear of movements (Hill et al. 2008), poorer report of quality of life (Hill et al. 2008) and catastrophizing behaviours (Hill et al. 2008) are just some of the variables that have been related to poorer overall outcomes in individuals with mechanical neck pain in studies with small to moderately sized samples. Individuals with whiplashassociated disorder with higher general health questionnaire scores, widespread pain throughout the body and higher initial disability scores demonstrated a risk of persistent neck pain (Atherton et al. 2006). 170
To our knowledge, no studies have used PRS modelling for individuals with neck pain managed by physical therapists and no studies have evaluated prognostic factors in large sample sizes, such as those found in repositories or outcomes databases. Large outcomes databases offer an opportunity to explore trends well beyond those possible within a clinical trial or a traditional observational study (Lohr 2012). Large datasets often contain variables that are consistent across all databases, which improves the transferability of the predictive findings. The objective of the present study was to identify a preliminary set of variables, through PRS, that identified individuals who were either at risk for a poor outcome or were likely to have a very good outcome regardless of the intensity of care received. The preliminary findings could serve as an initial investigation toward validating a risk stratification tool for targeting specific treatment intensities for individuals with neck pain seen by physiotherapists.
Methods Study design The study involved retrospective analyses of outcomes data from a clinical cohort of patients with neck pain who were formally seen for physiotherapistadministered care. It was approved by the Institutional Review Board at Franklin Pierce University, Manchester, NH, USA. Patient population A de-identified patient outcomes dataset, representing one full year of data collection (June 2012 to June 2013) from multiple physiotherapy clinics, was provided from Focus On Therapeutic Outcomes, Inc. (FOTO), an international medical rehabilitation outcomes database management company (Knoxville, TN, USA). All patients aged at least 18 years, with musculoskeletal impairments (all body regions) and complete data (intake scores, discharge scores and number of physiotherapy visits) were included in the initial dataset. Those with a primary diagnosis associated with cervical pain only were included (N = 3,137). Personal patient information was removed and each patient was assigned a patient identification number. All patients received care in outpatient physiotherapy clinics that used FOTO for collecting/measuring patient outcomes. Musculoskelet. Care 13 (2015) 169–178 © 2015 John Wiley & Sons, Ltd.
Cook et al.
Clinical data collection process Clinical data collection was longitudinal and consisted of data capture at baseline, during progress visits and at the time of discharge. During admission to physiotherapy, patients entered demographic data and completed self-report surveys using Patient Inquiry®, a computer program developed by FOTO. Patients completed these surveys throughout the episode of care and at discharge, depending on his/her clinicians’ case management strategies. Demographic data included age, state of residency and gender. Case-related data included baseline levels of pain, disability and fear avoidance beliefs; duration of symptoms; exercise status; surgical history; payer type, comorbidities and medication use.
Variables used in the modelling Our goal was to include variables in the PRS modelling that were available in most commercial or clinically available datasets. This required the inclusion of standard demographic and patient-reported outcomes data. We also included data on comorbidities, although there is a possibility that these variables may not have been captured in all datasets. Within the dataset, patient age and intake functional status (FS) score (intake FS score was initially rated 0– 100%, with 100% being full function) were coded as continuous variables, whereas the rest of the variables were treated as categorical variables. The FOTO FS survey to determine patient-reported functional level has been previously researched, with reliability and validity established (Hart et al. 2006, 2011). The FS scale is a self-reported measure administered using computeradaptive testing, which is designed to reflect the most appropriate item selections for each individual patient. Each item is queried on a five-point scale (0–4), with selections including ‘no difficulty’ = 4; ‘little difficulty’ = 3; ‘some difficulty’ = 2; ‘much difficulty’ = 1; and ‘I can’t do this’ = 0, and ‘didn’t do before’ = N/A. Symptom acuity, operationally defined as the number of calendar days from the date of onset of the condition being treated in therapy to the date of initial therapy evaluation, was categorized as acute (90 days). Surgical history was categorized into five categories (none, one, two, three, four or more surgical interventions related to the impairment being treated). Payer type was Musculoskelet. Care 13 (2015) 169–178 © 2015 John Wiley & Sons, Ltd.
Risk Stratification and Neck Pain
categorized into 16 payer sources (e.g. health maintenance organization, preferred provider organization). Number of functional comorbid conditions was assessed using a list of 30 conditions common to patients entering an outpatient rehabilitation clinic (e.g. heart attack, sleep disturbance, overweight and cancer). Exercise history prior to receiving therapy was categorized as exercising three times a week or more, 1–2 times a week, or seldom or never. Use of medication at intake was a dichotomous response (yes, no).
Recoding strategy As stated, the majority of codes within the dataset were categorical but involved multiple categorical options. As a research team, we opted to recategorize variables in an attempt to dichotomize the variables based on clinical sensibility. Clinical sensibility is the concept that the recategorized variables reflect findings typically encountered or derived from clinical practice. For example, although multiple categories were present for ‘duration of symptoms’, we dichotomized the variable into 22 days (sub-acute or chronic). Percentage change for function was calculated by taking the difference in the functional score (from baseline to discharge), dividing it by the baseline score and then multiplying by 100. The end product was a positive or negative, percentage change. Others have advocated using percentage change from baseline to determine improvement in outcomes (Dworkin et al. 2008). An example includes the Initiative on Methods, Measurement, and Pain Assessment in Clinical Trials (IMMPACT), which recommended a 30% reduction in pain from baseline as a lower threshold of success, with a 50% reduction a substantially clinically important change (Dworkin et al. 2008). Surgical history was dichotomized into ‘none’ (no) and ‘1 or greater’ (yes). Exercise status was dichotomized into 2 times a week (moderate or more). Multiple forms of payer type were represented and we endeavoured to categorize by similarity. Payer type was clustered into a) automotive, litigation and worker’s compensation; b) Medicare (all types) and Medicaid (based on care entitlement of use in the United States) and c) all others. As age has no obvious clinically sensible categories, it was retained in a continuous format. Lastly, patients were classified into high (elevated) or low 171
Cook et al.
Risk Stratification and Neck Pain
(non-elevated) fear avoidance beliefs, based on the screening items developed by Hart et al. (2011).
Risk stratification categorization PRS categorization can occur through many methods (Miller et al. 2001). Different processes may include providing a weighted score assigned for each criteria, a priori creation of risk categories or creation of risk categories based on management options (Miller et al. 2001). We elected to create a risk category based on management options – in other words, based on the outcomes and intensity of services provided (with intensity operationally reflecting the number of physiotherapy visits provided for the same condition). Our goal was to define who had a very good functional outcome with a low intensity of services and who had a poor outcome despite highly intensive services. Because we planned on creating a variable that incorporated the two unique elements of a functional outcome and intensity of services, it was not appropriate to analyse quartiles, standard deviations and other methods. As stated, we created three new categories by analysing the interplay between functional change (outcome) and total number of physiotherapy visits (intensity of services). In particular, we were interested in identifying individuals in each tail of the distribution: a) a group reflecting excellent functional outcome (by percentage change) with a minimal total number of physiotherapy visits and b) a group representative of a poor overall functional outcome despite multiple physiotherapy visits. After several iterations (e.g. 20%, 25%.....40%) of combining percentage distributions for FS outcomes and visit utilization, we identified that the highest 35% for visit utilization and the bottom 35% for functional outcome created the most notably remarkable tails. If individuals demonstrated both conditions they were uniquely coded as a high-risk (poor outcome) group. By contrast, to represent the low-risk group, we captured the highest 35% for functional percentage recovery and the lowest 35% for visit utilization. If individuals met both conditions, they were uniquely coded as a low-risk (good outcome) group. For completion of coding, we created a single code trichotomized into three categories: a) high-risk (low outcome, high number of physiotherapy visits); b) medium-risk (all individuals not in high- or low-risk categories) and c) low-risk (high outcome, low number 172
of physiotherapy visits). There were no instances in which one person was coded in both categories. After categorizing the groups, we found that the high-risk group comprised 17.9% of the total sample, whereas the low-risk group comprised 9.0% of the sample. Our medium-risk group comprised the majority of the total sample, at 73.1%.
Determining the appropriate number of observations per variable Homer and Lemeshow (2000) recommended that the minimum observation-to-variable ratio for a univariate multinomial or logistic regression is ten but cautioned that this recommendation is likely to lead to the overfitting of a model. With our sample size of greater than 3,100, and our current selection of ten variables, we were in no danger of overfitting the model.
Data analysis All analyses were performed using Statistical Package for the Social Sciences, version 22.0 (SPSS Inc., Chicago, IL, USA). Descriptive statistics for the full sample were calculated, including the values associated with the total number of physiotherapy visits and total episode duration. Continuous variables were represented by original means and standard deviations, whereas categorical variables were presented by frequencies using the recoded values. Differences among the high-, medium- and low-risk groups were calculated using chi-square and analysis of variance tests where appropriate. Univariate multinomial regression analyses were performed for each of the predictor variables for the trichotomized risk stratification variable. The referent variable used was medium-risk stratification. Multinomial logistic regression is used to predict the probability of category membership for a dependent variable with two or greater classifications, based on multiple independent variables (Domínguez-Almendros et al. 2011). Multinomial logistic regression analyses use maximum likelihood estimation to estimate model parameters. Selection of independent variables was performed manually to identify the strongest model. Multicollinearity for each the ten independent variables was evaluated by analysing correlation matrixes. According to Cohen (1988), correlational coefficients of the order of 0.10 are small, those of 0.30 are medium Musculoskelet. Care 13 (2015) 169–178 © 2015 John Wiley & Sons, Ltd.
Musculoskelet. Care 13 (2015) 169–178 © 2015 John Wiley & Sons, Ltd.
1,941=12–24 score 10.11 (7.20) 47.15 (50.71)
Total number of visits Total episode duration (days)
15.12 (8.42) 67.07 (50.29)
54.10 (15.27) 195 = Male 365 = Female 4.81 (3.07) 76 = 2 days a week 314 = 2 or less per week 70 = Group A 159 = Group B 331 = Group C 284 = Yes 276 = No 55.69 (14.03) 214 = 0–11 score 346 = 12–24 score
High risk (N = 560; 17.9%) frequency
9.65 (6.53) 45.56 (51.05)
53.50 (15.52) 841 = Male 1,453 = Female 4.84 (3.27) 490 = 2 days a week 1,380 = 2 or less per week 273 = Group A 617 = Group B 1,404 = Group C 1,231 = Yes 1,063 = No 51.31 (12.95) 871 = 0–11 score 1,423 = 12–24 score
Medium-risk (N = 2294; 73.1%) frequency
3.94 (1.00) 20.60 (29.52)
50.40 (16.80) 88 = Male 195 = Female 4.41 (2.62) 120 = 2 days a week 165 = 2 or less per week 34 = Group A 72 = Group B 177 = Group C 167 = Yes 116 = No 49.07 (9.70) 111 = 0–11 score 172 = 12–24 score
Low-risk (N = 283; 9.0%) frequency
