Development of a Predictive Algorithm to Identify Adults with Mobility Limitations Using VA Healthcare Administrative Data

Abstract

An estimated 31.5 million Americans have a mobility limitation. Healthcare administrative data could be a valuable resource for research on this population but methods for cohort identification are lacking. We developed and tested an algorithm to reliably identify adults with mobility limitation in US Department of Veterans Affairs healthcare data. We linked diagnosis, encounter, durable medical equipment, and demographic data for 964 veterans to their self-reported mobility limitation from the Medicare Current Beneficiary Survey. We evaluated performance of logistic regression models in classifying mobility limitation. The binary approach (yes/no limitation) had good sensitivity (70%) and specificity (79%), whereas the multi-level approach did not perform well. The algorithms for predicting a binary mobility limitation outcome performed well at discriminating between veterans who did and did not have mobility limitation. Future work should focus on multi-level approaches to predicting mobility limitation and samples with greater proportions of women and younger adults.

Introduction

Approximately 13% of US adults have a mobility limitation (Courtney-Long et al., 2015) and the rate is even higher among US military veterans, 18% (Centers for Disease Control and Prevention, 2020). People with mobility limitations are the largest subgroup within the population who report a disability (Courtney-Long et al., 2015). Mobility limitations include difficulty with several physical tasks, including walking, climbing stairs, and transferring in various environments (Patla & Shumway-Cook, 1999). Compared to the general adult population, adults with mobility limitation have higher rates of poor self-rated health (Froehlich-Grobe, Jones, Businelle, Kendzor, & Balasubramanian, 2016; Reichard, Stolzle, & Fox, 2011) and chronic conditions including heart disease, diabetes, and hypertension (Reichard et al., 2011), and are at increased risk of injuries and mortality (Guralnik et al., 1993; Hardy, Kang, Studenski, & Degenholtz, 2011). Health system and community barriers can limit healthcare access and quality of care for people with mobility limitation (Krahn, Walker, & Correa-De-Araujo, 2015). Thus, people with mobility limitation have been described as a health disparity population. Research is needed to identify promising interventions and evaluate the efficacy of policies aimed at reducing disparities for this large and growing population.

Population-based research on mobility-limited individuals faces significant barriers due to limitations in available data. National surveys generally include small numbers of adults with mobility limitation, requiring pooling of several years of data to achieve an adequate sample size. (C. Carroll, Cochran, Guse, & Wang, 2012). National surveys that include disability identifiers are often only available as cross-sectional samples (Livermore, 2007), which limits longitudinal research that can help answer important questions about how new policies/practices affect people with mobility limitations and how health status changes over time. Use of healthcare administrative data can address limitations related to sample size and lack of repeated measures found in survey data. But a barrier to their use is that mobility limitation is not a routinely coded condition. Therefore, identifying individuals with mobility limitation is a challenge and a significant barrier to the generation of evidence on how to improve the health of this large and growing population.

Use of administrative data to identify populations with mobility limitations

Previous research has used administrative data to identify similar constructs, but they have been confounded with more general functional limitations, limited to specific health conditions, or only, in part, related to mobility limitation. These included identifying people with a disability related to any limitations in activities of daily living (ADLs) and instrumental activities of daily living (IADLs) (Ben-Shalom & Stapleton, 2016; Davidoff et al., 2013) or focusing on aspects of frailty (Faurot et al., 2015). However, grouping mobility limitation with all other functional limitations or frailty may lead to aggregation bias and misunderstanding of the nature of relationships that exist for people with mobility limitation specifically. In other words, people with different types of functioning (cognitive vs mobility or mobility vs. getting dressed) may have different outcomes that would be missed when aggregated together. For example, the relationship between mobility limitation and obesity is quite different from the relationship between limitations with managing money (an IADL) and obesity. Furthermore, there are varying degrees of mobility limitation from mild to severe. Those with mild and moderate mobility limitation are outside of the construct of frailty and so would not be identified by the frailty algorithm.

Identifying mobility limitations in healthcare administrative data is also challenging because many people with mobility limitation have no directly related diagnosis (e.g., osteoarthritis, paralysis, multiple sclerosis) nor received an assistive mobility device (e.g., cane, walker), either of which is easily identifiable in claims or other administrative data and might be used as a proxy indicator of mobility limitation. Further, adults who do have a diagnosis often or usually associated with mobility limitation vary tremendously in their degree of mobility limitation (Iezzoni, 2002). The inability to reliably identify mobility limited populations is a significant barrier to realizing the potential of healthcare administrative data to address important research questions aimed at improving this population’s health outcomes (Iezzoni, 2002).

The purpose of this study was to develop and evaluate an approach for identifying veterans with mobility limitation using VA healthcare administrative data. The central research question was whether healthcare administrative data alone can reliably discriminate between those who do and do not have a mobility limitation. We evaluated multiple approaches to operationalizing mobility limitation including binary and multi-level mobility limitation severity. Our analysis utilizes a robust data set from the U.S. Department of Veterans Affairs (VA) that includes data not tested in previous research, such as neighborhood socio-economic status (SES) and census division. It is important to develop and test algorithms that are customized for different health systems that have different levels of availability and types of data, such as in the VA. The VA provides the opportunity to incorporate both age-related and military-services related causes of mobility limitation. Policies for durable medical equipment, such as wheelchairs, walkers, canes, differ from Medicare as they are based on medical need only and there is no copay or deductibles (Hubbard Winkler et al., 2006). Our motivation for the algorithm development was our desire to study personal and environmental moderators of the relationship between mobility limitation and obesity over time among military veterans.

New Contribution

To the best of our knowledge, this is the first study to 1) introduce a tool for identifying adults with mobility limitation in healthcare administrative data and 2) work towards development of a multi-level mobility limitation algorithm. Prior research has only used codes specific to health conditions or binary algorithms for having any ADL or IADL. The mobility limitation algorithm developed in this paper facilitates studies that seek to understand outcomes and utilization of adults with mobility limitation, a particular type of functional limitation associated with health risks different from those associated with other functional limitations. The algorithm provides new opportunities for researchers across many fields (e.g. public health, rehabilitation, occupational therapy, disability) studying health services and outcomes for veterans with mobility limitation.

Design and Methods

Overview

To develop the algorithm, we used a broad and comprehensive set of administrative data linked to survey data for the same individuals. The International Classification of Functioning, Disability and Health or ICF (World Health Organization, 2001), a biopsychosocial model of health and function, served as our conceptual framework. Based on the ICF, mobility limitation is an activity limitation that is influenced by environmental factors, personal factors, health conditions, and body functions and structure. Potential predictors of mobility limitation were selected from ICF domains and factors. Self-reported mobility limitation was used as the gold standard against which to evaluate the accuracy of the algorithm.

Sample

The cohort used in this study came from the Weight and Veteran’s Environment Study (WAVES) (Zenk et al., 2018), a retrospective observational study of US military veterans enrolled for VA healthcare. The sample included veterans 20-80 years old who were receiving primary healthcare in the VA in 2009-2014. Veterans were excluded who had no primary care visit within the two years preceding the first visit in 2009-2014 (the “look-back” period for obtaining baseline health information), had no geocodable home address for any of the years, or had a long nursing home stay at baseline (>90 days) (Zenk et al., 2018). For this study, we focused on a sub-sample of 964 veterans from WAVES that had also participated in the Medicare Current Beneficiary Survey (MCBS) for the years 2010 to 2013.

Data sources

We utilized data that are housed in the VA Corporate Data Warehouse, sourced from electronic health records and Medicare claims, and from the National Patient Prosthetics Database (NPPD), a resource developed to keep an accurate record of all prosthetics, medical equipment, and assistive devices, including those related to mobility (Department of Veterans Affairs, 2014). In the NPPD, Healthcare Common Procedure Coding System codes used for classifying assistive devices are organized into groups of related codes called ‘NPPD lines’ (i.e., manual wheelchairs), and further into ‘NPPD groups’ (i.e., wheelchairs and accessories) (Department of Veterans Affairs, 2014). Supplementary Table 1 provides a summary of the types of data and their sources.

The self-reported mobility limitation data came from the MCBS, which had been linked to administrative records using Social Security Numbers by the VA Information Resource Center. The MCBS is an ongoing survey of Medicare enrollees who participate in a 4-year panel (Adler, 1994). We used the survey items from the Health Status and Functioning – Community module (RIC_2) from the Cost and Use release of the data.

Dependent variable: mobility limitation

The MCBS asked four questions on difficulty walking. The first question was “because of a health or physical problem, do you have any difficulty walking?”. Respondents who said yes or that they are ‘unable to walk’ were subsequently asked if they “use special equipment or aids to help you with walking?” and “receive help from another person with walking?”. Respondents were also asked if they have “no difficulty at all, a little difficulty, some difficulty, a lot of difficulty, or are not able to walk a quarter mile (2 to 3 blocks)?”. We coded each of the difficulty walking questions as (0) for no or (1) for yes and dichotomized the ‘difficulty walking a quarter mile’ question as (1) for those reporting some difficulty, a lot of difficulty, or unable to walk a quarter mile, and (0) for those reporting a little difficulty or no difficulty.

In Figure 1, we provide an overview of the study design and analysis. There were 345 (36%) of the 964 MCBS respondents who reported any walking difficulty or reported that they were unable to walk. Of these, 221 (64%) used specialized equipment for walking and 44 (13%) needed help from another person with walking. There were 399 (41%) respondents who reported having at least some difficulty walking a quarter mile.

Figure 1:

Model for Classifying Mobility Limitation Severity Using Questions on the Medicare Current Beneficiary Survey (MCBS) (adapted from Shumway-Cook et al. (2005)

MCBS questions (blue rectangles)
Any difficulty walking		Because of a health or physical problem, do you have any difficulty walking?
Help from person		Do you receive help from another person with walking?
Use special equipment		Do you use special equipment or aids to help you with walking?
Any difficulty walking ¼ mile		Would you say you have no difficulty at all, a little difficulty, some difficulty, a lot of difficulty, or are not able to do it (walk ¼ mile, that is 2-3 blocks).
Approaches to operationalizing mobility limitation (grey rectangles)
Approach #	Categorization	Definition
1	Categorical (4-levels)	1. Develop 4 separate logistic regression models (Orange trapezoids 1-4) for each of the 4 MCBS questions 2. Generate predicted probabilities for each question 3. Choose a cut-off to use to create a binary predicted response 4. Combine predicted responses into a 4-level measure – none, mild, moderate, severe- based on diagram above. 5. Compare predicted level to the actual self-reported answer.
2	Categorical (3-levels)	1. Classify each person as none/mild, moderate, severe 2. Develop 3 separate logistic regression models (Orange trapezoids 5-7) 3. Generate predicted probabilities for each severity level 4. Calculate z-standardized predicted probabilities 5. Classify as predicted none/mild, moderate, severe based on highest z-standardized value. 6. Compare predicted level to originally classified level of mobility limitation
3	Binary moderate to severe mobility limitation	1. Develop a logistic regression model to predict those with moderate to severe mobility limitations (Orange trapezoid 8) 2. Generate predicted probabilities 3. Choose a cut-off to use to create a binary predicted response 4. Compare predicted moderate-severe limitation to actual moderate-severe limitation
4	Binary mild to severe mobility limitation	1. Develop a logistic regression model to predict those with mild to severe mobility limitations (Orange trapezoid 9) 2. Generate predicted probabilities 3. Choose a cut-off based to use to create a binary predicted response 4. Compare predicted mild-severe limitation to actual mild-severe limitation

The responses about walking difficulty are arranged using the same operational definition as Shumway-Cook, Ciol, Yorkston, Hoffman, and Chan (2005). In their approach, walking difficulty responses are organized based on a physical test, called the Functional Independence Measure (FIM) (Fiedler & Granger, 1996). The “underlying assumption [of the FIM] is that individuals requiring equipment are more restricted than those who do not use equipment but less restricted than those requiring the personal assistance of another” (Shumway-Cook et al., 2005, p. 1218).

Based on the combination of responses to these four questions, we developed a 4-level mobility limitation outcome (none, mild, moderate, severe), a 3-level mobility limitation outcome (none/mild, moderate, severe), and two binary outcomes. The legend of Figure 1 summarizes the steps used for developing the outcome variables, regression models, and comparison between predicted and actual outcomes, which are also described in more detail later in the methods section.

Predictor variables

For each of the four mobility limitation outcomes, the same set of potential predictor variables were used in developing logistic regression models.

ICF - Environmental factors

Assistive mobility devices:

We used a combination of NPPD lines and NPPD groups (Hubbard Winkler et al., 2012) to code six dummy variables: 1) ‘artificial leg’ for any prosthetic from foot to whole leg; 2) ‘surgical’ for surgical implants (e.g. in foot, knee, hip etc.); 3) ‘standing mobility’ included cane, walker, and walking aid accessories; 4) ‘orthotics’ included different types of braces and apparatuses that attach to the lower body; 5) ‘seated mobility’ included manual wheelchairs, power wheelchairs and scooter; and 6) ‘immobility’ for items related to home mobility aids and devices, such as a hospital beds, patient lifts and ramps into the home. Records were dropped if they were listed as ‘incomplete’, which meant the device was never picked up by the patient. We applied the same NPPD line coding system and indicators to Medicare data for patients who also were seen under Medicare and appended it to the NPPD data. (see Supplemental Table 2 for the NPPD codes and groups used)

Healthcare utilization:

Counts of inpatient stays, outpatient primary care visits and specialist visits were obtained from VA and Medicare data.

Geography and neighborhood SES:

Aspects of the environment such as weather and topography as well as policies or healthcare related practices that are specific to geographic regions can potentially affect mobility limitation and care related to mobility limitation as is discussed in the Dartmouth Atlas of Health Care (Wennberg, Fisher, Goodman, & Skinner, 2008). To approximate these important factors we included a variable for the census division the veteran resided in. We included measures on the percentage of the population with income below the federal poverty line and median household income at the census tract level.

ICF - Personal factors

We included demographic information including gender (male, female), race/ethnicity (white, black, Hispanic, other), age (continuous and categorical by ten-year age groups), marital status (married, single, widowed), and VA priority group. VA priority group is assigned based on service-connected disability, income, and identified special statuses (e.g., recently discharged, former prisoner of war). Priority group determines copayment obligations with the highest priority group, veterans who have a 50-100% service-connected disability, paying no copayments. Other veterans (including those with a <50% service-connected disability) are required to pay copayments for some services and others pay copayments for all services. We collapsed eight priority groups to 3 to conserve power. The 3 groups coincide with the levels of copayment obligation, but they also reflect degree of service-connected disability (which may include a mobility limitation), as determined by VA (US Department of Veterans Affairs, 2017).

ICF - Health conditions

We used a comprehensive set of ICD-9 codes and groups of codes for physical disability (mostly mobility) developed by an expert panel in Khoury et al. (2013). Because of our small sample and low prevalence of many conditions, we used ICD-9 code groups instead of individual codes. We supplemented the list by Khoury et al. (2013) with some additional diagnoses not on their list but often related to difficulty walking, including stroke, osteoarthritis, and ALS (amyotrophic lateral sclerosis, also called Lou Gehrig’s disease) (Hoenig, Pieper, Zolkewitz, Schenkman, & Branch, 2002; Hubbard Winkler et al., 2010). We included two general diagnoses related to difficulty walking that may be used to diagnose a more general state but not specific to a condition. Finally, we included twenty additional chronic conditions affecting other body systems that could also affect mobility, such as heart disease, COPD, and liver disease.

To strengthen the validity of ICD-9 code sets we used as predictors of mobility limitation, we applied a set of rules. First, we obtained codes only from provider encounter records, excluding those on records for ancillary services such as laboratory or imaging. Second, to reduce the risk of identifying a “rule-out” diagnosis as a definitive diagnosis, our operational definition for each condition required at least two occurrences of a diagnosis code on separate dates, at least 30 days apart. In inpatient records, only one occurrence was required.

We included a variable for the individual’s body mass index (BMI) in that year and also coded an indicator variable for morbidly obese (BMI≥ 40). BMI was calculated from patient visits as the modal height, and the mean weight in the 2nd half of the year. (Zenk et al., 2018)

ICF - Body functions and structure

Body functions and structure related variables included having an amputation, knee replacement, and hip replacement and a code group for paralytic conditions, such as paraplegia, quadriplegia, hemiplegia, and hemiparesis.

Development of the analytic dataset

We selected the first year of MCBS responses (out of four) for which healthcare data were available. For 16 of the veterans, all four potential years had no health data on BMI, demographics, healthcare utilization, census division, and census tract poverty. We used data from the two years before the first year in the MCBS and in the case of healthcare utilization variables took the average because they vary over time. Five subjects had no available weight data and so had missing values for weight related measures.

Our approach required us to set a window of time (see Figure 2) whereby diagnoses made in that window were considered potential predictors of the self-reported mobility limitation. Based on each interview date, dummy variables for each health condition or assistive mobility device were coded (1) if the date of diagnosis or receipt of device was within the 24 months prior to the interview date or 6 months afterwards. For predictors in the subdomains of healthcare utilization, geography and neighborhood SES, personal factors, and BMI, an annual measure was used coinciding with the year of the MCBS interview.

Figure 2:

Timeline of data used for Development of Regression Models for predicting Mobility Limitation using the Medicare current Beneficiary survey (MCBS)

Some health conditions were combined because of low frequencies (under ten). Variables were only combined if it made conceptual sense, such as three different types of cancers and diabetes with and without complications. We examined collinearity between predictor variables and removed one variable from a pair that was highly collinear and conceptually very similar.

Statistical analysis

We calculated descriptive statistics and bivariate correlations between each predictor and the outcome variable using Wilcoxon rank-sum (continuous with 2 groups) or Kruskal-Wallis (>2 groups) test, and Pearson’s chi-squared test for binary and categorical variables.

We used an iterative process to develop multivariate logistic regression models to predict two binary, one 3-level, and one 4-level mobility limitation outcome variables. Similar to Davidoff et al. (2013) and Faurot et al. (2015), in stage one, we used stepwise backwards logistic regression to identify the variables that strongly predicted each outcome by retaining only variables with a p-value of <0.05. In stage two, we started with the variables kept in the backwards elimination process and tested whether adding different forms of variables (categorical vs. binary) that were removed in the backwards stepwise regression models resulted in any new significant predictors that could be kept. If the new variable we added was significant at the p<0.05 level, it was retained. In stage three, we tested the inclusion of interaction terms with age, standing mobility devices, seated mobility devices, and race because the interaction of these variables with others may result in a greater likelihood of reporting walking difficulty. An interaction term was retained if both the interaction term and main effect terms were significant. We did not find any significant interaction terms to retain. In stage four, we used a bootstrap procedure to test the consistency of the statistical significance of variables using 200 repetitions (Nemes, Jonasson, Genell, & Steineck, 2009).

For each model, we calculated the Area Under the Curve (AUC) and its 95% confidence interval. The higher the AUC, the better the model is at minimizing false positives and maximizing true positives. An AUC of 0.70-0.79 is considered acceptable, 0.80 – 0.89 is good and 0.90 or above is excellent (Streiner & Cairney, 2007).

Model evaluation

After running the models, we generated predicted probabilities of having each mobility limitation outcome. To evaluate each model’s predictive performance, it was necessary to establish a cut-off value for the predictive probability. People whose predicted probability was above the cut-off were considered to have the condition while those with values below did not. Youden’s J was calculated to maximize the sum of the sensitivity and specificity (Youden, 1950), and so did not favor one over the other. Comparing the self-reported (our gold standard) outcomes to the predicted outcomes, we calculated the sensitivity, specificity, positive predictive value, negative predictive value, likelihood ratio positive, likelihood ratio negative, and percentage correct. For the three- and four-level measures, we evaluated the percentage correctly classified. The legend in Figure 1 describes the multi-level outcome evaluation.

Results

In Table 1, we assembled the frequencies and number of cases for each predictor variable. Given that this is a Medicare population, the average age was 70 and most subjects were in the range of 65-79. The sample was mostly white non-Hispanic (73%), almost completely male (96%) and most were married (65%). The largest category of assistive devices was canes and walkers (13%) and the lowest was home modifications (5%) The most common type of mobility related health condition was disorders of the Musculoskeletal system (36%) followed by diabetes (35%).

Table 1.

Bivariate Correlations of Potential Predictors with Any Walking Difficulty Among Veterans Who Completed the Medicare Current Beneficiary Survey, 2010-2013

Factor	No walking difficultya	Yes walking difficulty	p-valueb
N	619	345
Age 26-64	95 (15.3%)	125 (36.2%)	<0.001
Age 65-79	418 (67.5%)	167 (48.4%)
Age 80-85	106 (17.1%)	53 (15.4%)
Age, median (interquartile range)	73.0 (67.0, 78.0)	68.0 (63.0, 78.0)	<0.001
Married (compared to other non-married categories)	405 (65.4%)	219 (63.5%)	0.54
Non-Hispanic white (compared to other race/ethnicities)	453 (73.2%)	248 (71.9%)	0.66
VHA copayment group 1	139 (22.5%)	145 (42.0%)	<0.001
VHA copayment group 2	253 (40.9%)	132 (38.3%)
VHA copayment group 3	227 (36.7%)	68 (19.7%)
Home modifications (lift, standing frame, hospital bed)	--	39 (11.3%)	<0.001
Prosthetics or orthotics	22 (3.6%)	57 (16.5%)	<0.001
Wheelchairs (manual and electric)	15 (2.4%)	49 (14.2%)	<0.001
Canes, forearm crutches	22 (3.6%)	39 (11.3%)	<0.001
Walker	33 (5.3%)	47 (13.6%)	<0.001
Arthritis	85 (13.7%)	75 (21.7%)	0.001
Asthma	14 (2.3%)	12 (3.5%)	0.26
Heart failure	19 (3.1%)	31 (9.0%)	<0.001
Diseases of the central nervous system	20 (3.2%)	46 (13.3%)	<0.001
Chronic obstructive pulmonary disease	--	14 (4.1%)	0.001
Cerebrovascular disease	26 (4.2%)	26 (7.5%)	0.028
Depression	74 (12.0%)	94 (27.2%)	<0.001
Dementia	13 (2.1%)	--	0.61
Difficulty walking or abnormal gait	17 (2.7%)	43 (12.5%)	<0.001
Injuries and joint replacements	18 (2.9%)	21 (6.1%)	0.016
Diseases of the musculoskeletal system and connective tissue	171 (27.6%)	176 (51.0%)	<0.001
Osteoporosis	--	--	0.64
Peripheral vascular disease	36 (5.8%)	35 (10.1%)	0.014
Disorders of the peripheral nervous system	41 (6.6%)	62 (18.0%)	<0.001
Renal disease	56 (9.0%)	52 (15.1%)	0.004
Substance abuse	36 (5.8%)	23 (6.7%)	0.60
Cancer grouped	93 (15.0%)	37 (10.7%)	0.061
Diabetes without complications	187 (30.2%)	132 (38.3%)	0.011
Diabetes with complications	39 (6.3%)	47 (13.6%)	<0.001
Diabetes with OR without complications	192 (31.0%)	142 (41.2%)	0.002
Liver disease	13 (2.1%)	--	0.82
Morbid obesity	23 (3.7%)	39 (11.3%)	<0.001
Inpatient hospital stay (>0)	20 (3.2%)	29 (8.4%)	<0.001
0 visits to a specialist	294 (47.5%)	110 (31.9%)	<0.001
1-3 visits to a specialist	162 (26.2%)	92 (26.7%)
4 or more visits to a specialist	163 (26.3%)	143 (41.4%)
Census Divisions:
New England	20 (3.2%)	--	0.020
Middle Atlantic	78 (12.6%)	36 (10.4%)
East North Central	107 (17.3%)	48 (13.9%)
West North Central	62 (10.0%)	43 (12.5%)
South Atlantic	160 (25.8%)	77 (22.3%)
East South Central	54 (8.7%)	31 (9.0%)
West South Central	43 (6.9%)	49 (14.2%)
Mountain	49 (7.9%)	31 (9.0%)
Pacific	45 (7.3%)	24 (7.0%)
Census Division Missing	--	--
Metropolitan County	465 (75.1%)	241 (69.9%)	0.077
percent of census tract below poverty	12.2 (6.7, 19.2)	12.8 (7.6, 20.4)	0.17
median household income (census tract)	$48,250 (37,404, 63,424)	46,135 (36,275, 60,169)	0.094

^aThe interview question on walking difficulty in the Medicare Current Beneficiary Survey was “Because of a health or physical problem, do you have any difficulty walking?”

^bSignificance of correlations was assessed for continuous variables using Wilcoxon rank-sum (2 groups) or Kruskal-Wallis (>2 groups) test, and using Pearson’s chi-squared test for binary and categorical variables.

-- Indicates that there is too few subjects and results are not shown for privacy regulations from the VHA

In Table 1, we also show the bivariate correlations between each predictor and the ‘any difficulty walking’ MCBS question. Age was inversely correlated with ‘any difficulty walking’ (p<0.001), reflecting Medicare eligibility requirements for individuals under 65 (principally, disability). All of the assistive mobility device types were highly correlated with any difficulty walking (p<0.001). Many of the health conditions were also correlated with having any walking difficulty, such as COPD, heart failure, and depression. Although many people who have had a stroke have mobility limitation, cerebrovascular disease was not associated with having any difficulty walking.

In Supplemental Table 2, we show the coefficients and significance levels for variables kept in models in each of the nine final logistic regression models developed across the four approaches used to operationalize mobility limitation. Variables in all of the ICF domains and factors used (environmental factors, personal factors, health conditions, and body functions and structure ) were found to be significant predictors. The most common was from environmental factors, such as home modifications.

Model performance results

The model performance properties are summarized in Table 2. The AUC value, its standard errors, and 95% confidence intervals are listed for each outcome along with the performance properties. Overall, the AUC values were considered acceptable ( >0.7) and some were close to or above 0.8. The highest AUC was for the severe category of the 3-level outcome (0.841). The binary moderate to severe outcome had the highest sensitivity (71%) and needing help from another person had the highest specificity (85%).

Table 2.

Comparison of Model Performance of Final Logistic Regression Models Across the Four Mobility Limitation Outcomes Using Healthcare Administrative Data Among Veterans Who Completed the Medicare Current Beneficiary Survey Between 2010-2013

Mobility limitation classificationa	Question/level	AUC (SE) and CI	Sensitivity	Specificity	PPV	NPV	LR+	LR−	%correct overall
Multi-level #1 (4-levels)	Any walking difficulty	0.772 (0.016) CI: (0.740- 0.803)	71%	72%	59%	82%	2.58	0.40	72%
	Difficulty walking ¼ mile	0.779 (0.015) CI: (0.750- 0.809)	66%	79%	69%	77%	3.14	0.43	73%
	Uses specialized equipment	0.734 (0.026) CI:(0.683- 0.784)	61%	78%	83%	53%	2.78	0.50	67%
	Needs help from a person to walk	0.817 (0.036) CI: (0.746-0.888)	68%	85%	18%	98%	4.51	0.37	84%
Multi-level #2 (3-levels	None/mild	0.794 (0.016) CI: 0.764- 0.824)	n/a	n/a	n/a	n/a	n/a	n/a	n/a
	Moderate	0.758(0.0168) CI: 0.725 - 0.791)	n/a	n/a	n/a	n/a	n/a	n/a	n/a
	Severe	0.841(0.034) CI: 0.773 - 0.908)	n/a	n/a	n/a	n/a	n/a	n/a	n/a
Binary #1	Moderate to severe	0.812 (0.0170) CI: (0.778 - 0.845)	79%	72%	46%	92%	2.82	0.29	73%
Binary #2	Mild to severe	0.800 (0.0144) CI: (0.772 - 0.828	70%	79%	74%	75%	3.33	0.38	75%

^aSee figure 1 for full description of each mobility limitation outcome.

In Tables 3a and 3b, we compare the % correct of 4-level categorical outcomes and 3-level categorical outcomes respectively. The predicted 4-level outcome did not match up well to the actual 4-level reported by respondents. The percentage of those who were predicted to have no limitation was correct for 75% of those reporting no limitation, but the percentage correct for the other levels were all below 50%. In the 3-level outcome, it was a similar pattern, as 80% of veterans were correctly classified as having none or mild mobility limitation, 51% were correctly classified as having moderate limitation, and 33% correctly classified as having a severe limitation.

Table 3A:

Comparison of Predicted to Actual Severity of Mobility Limitation for the Four Level Mobility Limitation Outcome Using Healthcare Administrative Data Among Veterans Who Completed the Medicare Current Beneficiary Survey Between 2010-2013^a,b

Limitation	predicted none	predicted mild	predicted moderate	predicted severe	total
None	368	93	26	29	516
% correct	75%	44%	23%	19%
mild	86	72	29	36	223
% correct	17%	34%	26%	24%
moderate	32	40	54	54	180
% correct	7%	19%	49%	36%
severe	6	6	2	30	44
% correct	1%	3%	2%	20%
Total	492	211	111	149	963

^aMobility limitation severity levels were derived from questions on walking difficulty.

^bSee figure 1 for full description of the 4-level mobility limitation outcome.

Table 3B:

Comparison of Predicted to Actual Severity of Mobility Limitation for the Three Level Mobility Limitation Outcome Using Healthcare Administrative Data Among Veterans Who Completed the Medicare Current Beneficiary Survey Between 2010-2013 ^a,b

Limitation	predicted none/mild	predicted moderate	predicted severe	total
none/mild	485	128	5	618
% correct	80%	44%	7%
moderate	112	149	41	302
% correct	19%	51%	59%
severe	8	13	23	44
% correct	1%	4%	33%
total	605	290	69	964

^aMobility limitation severity levels were derived from questions on walking difficulty.

^bSee figure 1 for full description of the 3-level mobility limitation outcome.

Discussion

The objective of this study was to determine whether healthcare administrative data alone could be used to reliably identify people with mobility limitation. We drew on data from the VA and developed several alternate algorithms based on different forms of a mobility limitation outcome and tested their performance. We showed that the algorithms for predicting a binary mobility limitation outcome performed well at discriminating between people who did and did not have mobility limitation. However, for the multi-level mobility limitation outcome, the predicted levels of mobility limitation severity did not match well with the actual severity levels.

Comparison of multi-level and binary approaches

Several of the binary mobility limitation outcomes we studied had an AUC near or above 0.8, which means the models performed well at discriminating between those with and without mobility limitations. Due to the absence of similar studies on mobility limitation in the extant literature, we compare our results to those that focused on identification of any limitations in ADLs. Previous studies predicting any limitation in ADLs had similar performance as ours (Ben-Shalom and Stapleton (2016) (AUC = 0.75) and Faurot et al. (2015) (AUC= 0.85)). Our focus on mobility limitation specifically can support the identification and study of this population and answering policy relevant questions targeted to people with mobility limitation.

Similar to the limitations in any ADLs studies, we found predicting mobility limitation severity challenging. Davidoff et al. (2013) reported their models could not distinguish a middle ‘some limitation’ category from their ‘none’ or ‘a lot’ of limitation categories. The difficulty discriminating between severity levels may stem from an inherent limitation in claims data; diagnosis codes do not capture severity of a condition. For instance, from a single diagnosis code, we are not able to distinguish between mild and severe osteoarthritis or where someone might be in the progression of multiple sclerosis that leads to greater mobility limitation. The challenge in predicting severity may also be, in part, due to the temporal nature of mobility limitation. In a study about the consistency of the disability questions used on the Current Population Survey, Ward, Myers, Wong, and Ravesloot (2017) showed that only 39% of respondents with mobility limitation consistently reported having one within the same year. The middle or “some” category may be people with temporary or milder mobility limitations that can vary over time.

Role of environmental factors

Assistive mobility devices proved to be critical for correctly identifying mobility limitation as they were the strongest predictors in every model in our study. Davidoff et al. (2013) and Faurot et al. (2015) also found that indicators of having a wheelchair or hospital bed were strong predictors in their limitations in any ADLs models. The robust predictive value of assistive mobility devices may reflect the fact that they are used by people across health conditions who have a mobility limitation. However, among those with each type of assistive mobility device, there were between 1.1% and 5% who responded on the MCBS that they do not have difficulty walking. This unexpected finding shows how relying simply on assistive devices is not a reliable approach to identifying a sample with mobility limitations. Some respondents who used assistive devices may feel they do not have a limitation when they walk with their devices. Alternatively, these could be people who received a device as part of rehabilitation and then no longer use it.

Neighborhood SES may play an important role in contributing to mobility limitation. The percentage of people below the federal poverty line and median household income were significant predictors in some of the models we tested. The contributing role of neighborhood SES aligns with how the ICF identifies environmental factors as interacting with personal factors and health conditions to affect mobility limitation (Schneidert, Hurst, Miller, & Ustun, 2003). In our study, people living in a census tract with a higher percentage of individuals below poverty were more likely to report mobility limitation. Similarly, being in a census tract with a higher median income was associated with a lower likelihood of reporting a mobility limitation. Previously research has not included neighborhood SES data, which may be picking up the effect of individual SES on mobility limitation. These findings affirm the correlation between poverty and disability that is well documented in the literature (Lauer & Houtenville, 2018).

Census division, and in particular the West South Central division, was also a consistent predictor of mobility limitation. Veterans in Texas, Louisiana, Oklahoma, and Arkansas may be more likely to report difficulty walking even when many other factors are controlled for. Additionally, there may be certain aspects of the policies and treatments (or lack thereof) for mobility limitation that differ by census division. Different state policies, or different healthcare practices that vary by Veterans Integrated Services Networks (VISN) (Hubbard Winkler et al., 2010), may account for some of this regional variation.

Sensitivity analyses showed similar results to our main analyses (see Supplemental Table 3). We tested a shorter time window around the MCBS interview date (12 months prior and three months post interview) instead of the two-year window in the period before the interview and the six months after the interview for the binary mild to severe model. The AUC for the shorter time window was slightly lower (0.753) and also had lower sensitivity (70%) and specificity (71%). We also examined an alternate model removing variables that may not be readily available outside the VA: marital status, race/ethnicity, copayment group status, and BMI. This alternative model also had similar performance properties. As a next step, it is necessary to validate the algorithm in a different sample that has a higher percentage of women and younger age groups. A similar approach as was used in this paper could be tested to expand the mobility limitation algorithm to people with acquired and congenital mobility limitations. Future research can work to identify when mobility limitation decreases in severity or is no longer a problem. This would help in distinguishing between those with temporary and permanent mobility limitations.

Implication for policy and practice

There are many public health and healthy aging strategies that could benefit from using healthcare administrative data to evaluate outcomes for people with mobility limitation. The algorithm developed in this paper has implications not only for our current study – WAVES- but also for the extensive portfolio of research needs within the VA studying healthcare utilization and outcomes. The algorithm can be applied to healthcare administrative data from the VA to estimate the number of veterans with mobility limitation who have chronic disease, such as diabetes, COPD, and heart disease and these same individuals can be tracked over time. Given the high rates of obesity and physical inactivity among people with disability (An, Andrade, & Chiu, 2015), previous research has argued for policies to make accessible weight scales more available (Iezzoni, 2011) and training physicians to recommend physical activity for patients with mobility limitations (D. Carroll et al., 2014). However, evaluating whether such interventions are effective using healthcare administrative data requires identification of a cohort of people with mobility limitation. Specific to data from WAVES, we plan to examine how changes in neighborhood walkability impact obesity over time for veterans with mobility limitation. The algorithm developed in this paper meets these requirements for studying veterans and overcomes the gaps of survey data, which are usually cross-sectional and insufficient for following a cohort of people with mobility limitations over time and in relation to a specific policy change or intervention (Livermore, 2007). The approach to algorithm development can be replicated in other health systems to understand if the algorithm has similar predictive variables and performance.

Limitations

There could be mistakes in coding of diagnosis by healthcare providers (Peabody, Luck, Jain, Bertenthal, & Glassman, 2004), but we attempted to deal with these by requiring two codes greater than 30 days apart for outpatient visits.

Some veterans may have received an assistive device for a temporary limitation and if their mobility improved, they may then report no difficulty walking. However, we did not include devices that are specifically designed for temporary conditions, such as crutches. Because of the smaller sample, we had to combine some diagnosis codes. With a larger sample, it may be possible to use more individual ICD-9 diagnosis codes to tease out any heterogeneity within ICD-9 code groups. However, we were able to show that using ICD-9 code groups had good predictive properties. The sample was predominantly older male and so further research is needed with females and younger adults with mobility limitation. The analysis was run on the development sample, as we did not have a large enough N for a development and validation sample. Additional testing is needed to validate the algorithm and refine it so that it is generalizable beyond the VA. The methods and scope of predictors used are likely transferable to other populations given that our list of candidate predictors was organized around the ICF, a tool for classifying disability used internationally.

Conclusion

In this study, we showed that in the absence of self-reported mobility limitation data, it is possible to identify adults with mobility limitations using a predictive algorithm composed only of healthcare administrative data. The algorithm developed in this paper can facilitate health services and outcomes research by defining cohorts of people who have a mobility limitation and studying their health over time. In order to address health disparities that have been documented for people with mobility limitations, the mobility limitation algorithm can be used to evaluate environmental, policy, or systems changes that are implemented in Veterans Health Administration, and with some additional refinement, in other health care settings as well.