Exercise, Functional Limitations, and Quality of Life: A Longitudinal Study of Persons With Multiple Sclerosis

Data Collection Procedures 

We recruited the sample for this longitudinal study from a group of persons with MS (initially recruited through contacts with 2 chapters of the National Multiple Sclerosis Society in the southwestern United States and advertisements in rural newspapers) who had participated in a cross-sectional survey in 1996−1997. Following institutional review board approval in 1999, we mailed study information and consents for an additional 5-year follow-up to the 749 participants who remained eligible and had not been lost to follow-up due to death, institutionalization, or loss of contact. We received 621 (82.9%) useable questionnaires. Those who did not return questionnaires and indicate their willingness to continue in the study for 5 years did not receive subsequent mailings. Those who returned questionnaires in 1999 were enrolled in the longitudinal study and received study surveys over each of the following years unless they requested to be dropped from the study (n=12), were ineligible due to institutionalization (n=7), reported they were too ill to continue participation (n=10), changed diagnosis (n=10), were deceased (n=19), or moved without follow-up address (n=3). At time 2 (2000), 89.6% of the mailed questionnaires were returned (n=557), at time 3 (2001) 88.6% of the questionnaires were returned (n=530), at time 4 (2002) 87.4% returned questionnaires, and at time 5 (2003) 85.3% returned questionnaires. At the end of the fifth year of the longitudinal data collection 560 of the original 621 participants in the longitudinal study remained enrolled. As might be expected because study dropouts include those who have died and/or been institutionalized, the group of 61 dropouts was significantly older (52y vs 46.9y) and had greater functional limitations scores (21.1 vs 16.8) than those still enrolled in the longitudinal study.

Sample Characteristics 

The sample used for the longitudinal analysis described below included 611 persons with MS (10 of the original participants were not included because they reported after their recruitment and participation that their physician had subsequently told them that they did not have MS). At the start of the longitudinal study (time 1), the participants ranged in age from 21 to 80 years (mean ± standard deviation, 49.4±10.2y). The majority were women (83%), white/non-Hispanic (93%), and married (73%). Most (85%) had completed high school and 35% had completed college. Only 25% were currently employed full-time, 32% reported that they were unemployed due to disability, and 13% were retired. With respect to illness characteristics, 41% reported having relapsing-remitting MS, 18% had primary progressive MS, 17% had secondary progressive MS, and 11% had progressive relapsing MS. The remainder reported either benign MS or did not know what type they had. The average length of diagnosis was 13.5 years. This sample corresponds closely to estimates for the population of persons with MS in the United States.26 Using National Health Interview Survey data, it is estimated that among those with MS, 73% are women, 95% are white or non-Hispanic, 85% are married, and 82% had completed high school.

Those who chose to enroll in the longitudinal study were, statistically speaking, younger and reported less limitation, better QOL, fewer secondary conditions, more frequent health behaviors, greater social support, and more economic resources than the nonresponders from the initial cross-sectional survey. Although significant in this large sample, the effect size of all differences between those who participated and those who chose not to was small.

Procedures 

We sent questionnaires to the participants at each time point unless they withdrew from the study after the longitudinal study started. If participants did not return the questionnaire within 30 days, they were sent a new questionnaire. The questionnaire format remained consistent over the course of the study and addressed a range of constructs from the explanatory model. Variables of interest for these analyses included demographic variables, functional limitations, perceived QOL, and exercise behaviors.

This study used self-report measures to provide a more complete understanding of the progression of functional limitations and QOL over time. Although performance-based measures of functioning are often assumed to be superior to self-reported measures, no extensive empirical evidence exists to support this assumption. For example, the Expanded Disability Status Scale,27 the most frequently used outcome measure in MS studies, has been widely criticized for its lack of sensitivity to detect change, low interrater reliability (especially at the lower ends of the scale), and the skewed distribution and ordinal nature of the resulting scale scores.28 In a direct comparison of self-report and performance measures of basic and instrumental activities of daily living, Myers et al29 found that relative to self-report questionnaires, performance measures were not psychometrically superior, more acceptable to participants, easier to interpret, or easier to administer. In fact, self-report measures may be preferable to performance measures, which may be vulnerable to differences in effort and influenced by testing situations. Lower self-reported function predicts a number of adverse outcomes including increased hospitalization, greater use of services, and mortality.30

We used a background information sheet to collect information on a variety of demographic and disease characteristics that were used to describe the sample. Data were collected about participants’ age, race and ethnicity, marital status, children, and education and employment status. Participants were also asked to indicate their type of MS and the year in which MS was first diagnosed by a physician.

Individual ratings on the Incapacity Status Scale (ISS)31 provided a measure of functional limitations due to MS. Kurtzke’s original structured interview form was adapted to a self-administered questionnaire format to provide an assessment of the degree of impairment as reflected in 16 aspects of personal functioning (eg, ambulation, personal care, vision, bladder and bowel functioning) represented on the ISS. Each of the items is rated on a 5-point scale, with 0 indicating normal functioning and 4 indicating complete inability to perform the activity. The Cronbach α for the scale at time 1 was .88. In a subsample (n=73) who completed the MS Functional Composite,32 total scores on the ISS were significantly related to the performance measures of ambulation (r=−.49) and upper-extremity coordination and control (r=−.57).

Participants completed the 52-item Health Promoting Lifestyle Profile II (HPLP-II) annually. This scale33 assesses the frequency with which individuals report engaging in activities to increase their level of health and well-being. Responses are scaled from 1 (never) to 4 (routinely), with higher scores indicating more frequent practice of a health behavior. For this study, the sum of scores on the 8-item exercise/physical activity subscale of the HPLP-II were used to measure exercise behaviors over time. Higher scores indicate more frequent engagement in exercise and physical activity behaviors. Examples of subscale items include how frequently the respondent follows a planned exercise program, engages in light to moderate physical activity, engages in leisure activity, does stretching exercises, and gets exercise during daily activities. The Cronbach α for this subscale for the current sample at time 1 was .84. In earlier research,34 scores on the exercise subscale were significantly associated with measures of strength and flexibility in persons with MS.

We measured QOL by the Quality of Life Index (QLI)–MS version.35 This instrument was designed to assess general satisfaction with and degree of importance placed on components that add to QOL. This measurement approach is consistent with the World Health Organization36 definition of QOL as a person’s perception of their position in life in the context of that person’s culture and value system and in relation to his/her goals, standards, and concerns. Composed of 72 items, the QLI is comprised of 2 parts. Part 1 assesses satisfaction across a variety of life domains; part 2 measures the degree of importance attributed to these domains. Respondents rate each item on a 6-point ordinal scale, ranging from “very satisfied” to “very dissatisfied” for part 1 and “very important” to “very unimportant” for part 2. Total scores are computed by weighting each satisfaction response with its corresponding importance response. Thus, combinations of high satisfaction and high importance ratings produce higher total scores, indicating more positive perceptions relating to life quality. The Cronbach α at time 1 was .96. Convergent validity has been supported by a correlation of .77 between the QLI and an index measuring overall satisfaction with life.37

Longitudinal Data Analysis 

We used latent curve modeling to analyze self-reported, longitudinal measures of functional limitations, QOL, and exercise behaviors. Latent curve models are studied under different names including hierarchical linear models, multilevel models and mixed-effects models for longitudinal data. Latent curve models are widely used to study change in a single repeated measure, where change in an outcome is considered at the population level as well as the individual.38, 39 For example, a variable that follows a linear trend over time may be well described by a model that includes an intercept (the value of the outcome measure when time is equal to zero) and a linear time effect. In a latent curve model, a fixed intercept and linear time effect describe change at the population level. In many cases, people vary with respect to their levels or rates of change in the response measure. Individual differences in response trajectories are handled by a model that allows the intercept and linear time effect to vary across individuals. The assumption is that the coefficients (eg, intercept, linear time effect) of the people sampled are representative of a larger population of coefficients, and so they are treated as random in the model. This assumption allows the trajectories of the people to vary from one another in the population. Latent curve models may also be used to study the moderating effects of person or contextual variables on individual trajectories.40 For example, individual differences in the level or the rate of change in a variable may be accounted for in part by sex. Moderating effects are studied by regressing coefficients at the first level of the model (eg, individual-level intercept or slope) on variables measured at the second level (eg, sex).

Latent curve models have also been applied to multiple variables observed longitudinally.41, 42 In a multivariate latent curve model, characteristics of change in 1 variable are related to those of another variable. For the current study, this approach was used to study how both the level of response at time 1 and the rate of change in functional limitations were related to response levels at time 1 and the rate of change in reported levels of exercise behaviors and QOL. A multivariate latent curve model was estimated in which these 3 outcome variables were considered together to study the correlations between change characteristics (eg, response levels at time 1 and linear time effects) between these variables. Covariates were also included in the model to adjust for differences in characteristics of change due to age, sex, years since diagnosis, residence type (rural vs urban), and subject attrition (described below).

We considered a growth model for annual observations spanning the duration of the longitudinal study.⁎ In a latent curve model, the interpretation of the model’s intercept is dependent on the meaning of the value of zero for each predictor variable. That is, the intercept represents the level of the response variable when the predictors are equal to zero. At the occasion level, time was coded so that the intercept of a model for a particular response variable reflected the response level at the start of the longitudinal study. Time was coded 0, 1, 2, 3, and 4 to represent the annual assessments with zero corresponding to the start of the study. The time effect thus represented an annual rate of change. At the individual level, sex (labeled hereafter as “male”) was coded 0 for female and 1 for male. Residence status (labeled hereafter as “rural”) was coded 0 for urban and 1 for rural. For age, a quantitative variable, zero had no practical meaning. Therefore, age was transformed by subtracting from each person’s observed age the sample mean age of 49 years at time 1. The effect of this transformation of age meant that the zero point represented 49 years after the transformation. Similarly, years-since-diagnosis was transformed by subtracting 5 years, the sample mode at time 1,† from each person’s observed score. We chose to center both age and years-since-diagnosis to values that were best representative of the current sample. Together, these predictors were included in the models to account for individual differences in the levels of the outcome variables at the start of the study as well as in the rates of change in measures over the 5-year period.

Moderating Effects of Person-Level Variables on Change Characteristics 

Person-level variables, age, sex, years since diagnosis, and residence type were considered as potential moderators of time 1 response levels and annual change rates at the individual level. To study these relations, we treated time 1 response levels and annual change rates as criterion variables in an intercepts-and-slopes-as-outcomes model40 with these background variables treated as predictors.

Reasons for missing data included death, institutionalization, illness, relocation, and refusal to participate further. When data are missing in a random manner, a latent curve model may yield unbiased results.40 When data are not missing at random, the method may produce biased estimates.43 Given the nature of the current study, it was possible that unmeasured processes giving rise to the missing data were related to the missing values so that missing data could not be ignored. For example, individuals who dropped from the study may have had response patterns in their measures of functional limitations, QOL, and exercise behaviors over time that differed from those who remained in the study. To address this possibility, we adopted a pattern-mixture random-effects model44 by including in the latent curve model variables that represented different missing data patterns. Two specific patterns of missing data were coded by creating the following indicator variables: miss1 equals 1 if an individual dropped by the second or third measurement occasion (n=64), and miss2 equals 1 if an individual dropped by the fourth or fifth occasion (n=42). Indicator variables were then coded 0 if the particular missing data pattern did not apply. Other possible patterns yielded very small sample sizes and so were not considered. These 2 variables were included in the model to adjust for effects due to drop out. Mplus version 345,a was used to estimate the models. All statistical tests were based on 2-tailed tests using a significance level of .05.

Trajectories of Individual Outcome Variables 

Each outcome measure was best described by a linear trajectory. Therefore, each outcome was summarized by a model that included an intercept representing the level of response at time 1 and a linear time effect denoting the annual rate of change. Random coefficients (ie, random intercept, linear time effect) allowed for individual differences in these change features. Different assumptions about the errors at the occasion level were considered. A common assumption is that these errors are independent with constant variance across time. For each measure, this assumption was tested by comparing the fits of alternative models that either allowed the error variances to vary across time (ie, heterogeneity of variance) or for covariances between errors at temporally adjacent time points (ie, autocorrelation). The assumption of homogeneity of variance for the errors was tenable for both QOL and functional limitation scores, whereas, for exercise scores, allowing the errors to covary between adjacent time points was preferred.

Estimates that describe mean change in the three measures are given in table 1. The estimated mean levels at time 1 for functional limitation scores, exercise behaviors scores, and QOL scores were 18.4, 15.4, and 20.3, respectively, and the estimated average annual rates of change were .236, −.040, and .032, respectively. Only for functional limitations was the average annual rate of change statistically different from zero. That is, on average, exercise behaviors, and QOL scores showed no change, but functional limitations increased on average over the 5-year period.

Table 1. Estimated Mean Levels at Time 1 and Time Effects and Variances for Corresponding Random Effects for 3 Outcome Variables

	Mean Effects
	Outcome	Effect	Mean Estimate ± SE	t Ratio	P⁎
	Functional limitations	Level at time 1	18.4±.400	45.99	<.001
		Time	.236±.056	4.18	<.001
	Exercise behaviors	Level at time 1	15.4±.219	70.58	<.001
		Time	−.040±.050	−0.81	.421
	QOL	Level at time 1	20.3±.200	101.12	<.001
		Time	.032±.038	0.86	.390

	Random Effects
	Outcome	Effect	Variance Estimate ± SE	t Ratio	P⁎
	Functional limitations	Level at time 1	91.7±5.60	16.39	<.001
		Time	.712±.109	6.54	<.001
	Exercise behaviors	Level at time 1	23.4±1.72	13.60	<.001
		Time	.388±.103	3.76	<.001
	QOL	Level at time 1	21.1±1.41	14.92	<.001
		Time	.185±.052	3.58	<.001

NOTE. Values are maximum likelihood ± standard error (SE).

⁎ Based on 2-tailed significance tests.

Individual differences in the response levels at time 1 and the annual rates of change were evident due to the large variances corresponding to the random intercepts and time effects. The estimated variances of scores at time 1 (ie, variances of the intercepts) for functional limitations, exercise behaviors, and QOL scores were 91.7, 23.4, and 21.1, respectively. Each of these variances was statistically different from zero, indicating interindividual variability in response levels at time 1. The estimated variances of the annual rates of change (ie, variances of the slopes) for functional limitations, exercise behaviors, and QOL scores were .712, .388, and .185, respectively. Each of these variances was statistically different from zero, suggesting interindividual variability in the annual rate of change in scores across the 5-year period. As discussed earlier, the average annual rates of change in exercise behaviors and QOL scores were statistically not different from zero, suggesting that scores did not change over the study period at the population level. However, individual differences in the individual-level change rates suggest scores did change at the individual level for both variables. For these 2 measures, given that the population change rate was essentially close to zero, evidence of individual differences in the change rates suggested that about half of the individuals had scores that increased over time and about half had scores that decreased.

Tests of Moderating Effects of Person-Level Variables on Change Characteristics 

Potential moderating effects of age, sex, years since diagnosis, residence type, and drop-out types on individual response levels at time 1, and annual change rates in the outcome variables were studied by allowing the change characteristics to vary as functions of the background variables. Results from this analysis are presented in table 2. Corresponding to each variable is an estimated mean time 1 response level and annual change rate, both at the population level, when all the individual-level predictors are equal to zero. Due to the coding of the background variables as discussed earlier (ie, how 0 was defined for each variable), the estimated mean time 1 response levels and annual change rate for each outcome variable represent those values for women who were 49 years old, were living in urban areas, remained in the study at the last assessment, and were diagnosed 5 years prior to the start of the study. This particular group was not of specific interest, so the estimated mean time 1 response level and annual change rates for this group are not considered further. The effects relating to male, rural, miss1, and miss2 are the estimated mean differences in time 1 responses and annual change rates between men and women, rural and urban residence types, early drop-out and no drop-out, and late drop-out and no drop out, respectively, holding constant the other predictors in the model. The effects relating to years-since-diagnosis and age are the estimated moderating effects of these measures on time 1 responses and annual change rates, holding constant the other model predictors. All of these effects are discussed with regard to specific outcome variables below.

Table 2. Moderating Effects of Background Variables on Time 1 Response Levels and Annual Change Rates

	Outcome	Background Variable	Effect Estimate	t Ratio	P⁎
	Functional limitations
	Time 1 response level		15.5±.749	20.60	<.001
		Years since diagnosis	.120±.058	2.06	.040
		Male	1.02±1.01	1.01	.313
		Age	.153±.042	3.63	<.001
		Rural	2.08±.881	2.36	.019
		Miss1	6.18±1.26	4.89	<.001
		Miss2	3.25±1.39	2.34	.020
	Annual change rate		.323±.111	2.90	.004
		Years since diagnosis	−.014±.009	−1.54	.124
		Male	.252±.152	1.65	.099
		Age	.004±.006	0.64	.511
		Rural	−.020±.131	−0.16	.877
		Miss1	1.60±.836	1.91	.057
		Miss2	.355±.256	1.39	.165
	Exercise behaviors
	Time 1 response level		16.1±.429	37.49	<.001
		Years since diagnosis	−.013±.033	−0.38	.702
		Male	−.059±.579	−0.10	.920
		Age	−.024±.024	−0.98	.326
		Rural	−.580±.502	−1.16	.247
		Miss1	−1.99±.739	−2.69	.007
		Miss2	−2.35±.799	−2.94	.003
	Annual change rate		.000±.097	0.00	1.000
		Years since diagnosis	−.000±.008	−0.05	.958
		Male	−.057±.133	−0.43	.665
		Age	−.001±.006	−0.23	.822
		Rural	−.263±.115	−2.30	.022
		Miss1	−.300±.729	−0.41	.680
		Miss2	.730±.232	3.14	.002
	QOL
	Time 1 response level		20.6±.392	52.54	<.001
		Years since diagnosis	.040±.030	1.32	.187
		Male	−1.28±.528	−2.42	.016
		Age	.027±.022	1.20	.231
		Rural	−.733±.459	−1.60	.110
		Miss1	−1.67±.673	−2.48	.013
		Miss2	−1.99±.727	−2.74	.006
	Annual change rate		.010±.075	0.14	.893
		Years since diagnosis	.003±.006	0.42	.672
		Male	−.047±.102	−0.46	.646
		Age	.002±.004	0.53	.597
		Rural	−.042±.088	−0.47	.636
		Miss1	−.727±.622	−1.17	.242
		Miss2	.258±.183	1.42	.156

NOTE. Values are parameter estimates ± SE.

⁎ Based on 2-tailed tests of significance.

For functional limitation scores at time 1, the effects relating to years-since-diagnosis and age were significant, suggesting a tendency for those with a greater number of years since diagnosis and who were older to have higher levels of functional limitations at time 1. Although there was no sex difference in functional limitation scores at time 1, those living in rural areas had on average a functional limitation score that was about 2 points higher that those living in urban areas, those who dropped from the study early had on average a functional limitation score that was about 6 points higher than those who remained by the last occasion, and those who dropped from the study late had on average a functional limitation score that was about 3 points higher than those who remained by the last occasion. With regard to the annual change rate in functional limitation scores, statistical tests did not suggest any associations between the background variables and the annual change rate.

For exercise behavior scores at time 1, those who dropped from the study early had on average an exercise behavior score that was about 2 points lower than those who remained by the last occasion, and those who dropped from the study late had on average an exercise behavior score that was about 2.4 points lower than those who remained by the last occasion. Statistically, there was no difference in the effects of drop-out on time 1 exercise scores for those who dropped early versus late. No other background variables were related to exercise behavior scores at time 1. With regard to the annual change rate in exercise behavior scores, those living in rural areas showed on average a change in scores of about .3 points lower per year relative to those living in urban settings, and those who dropped from the study late had on average a change in scores of about .7 points higher per year relative to those who remained in the study by the last occasion. Statistical tests did not suggest associations between the annual change rate in exercise scores and the remaining background variables.

For QOL scores at time 1, men reported on average lower QOL scores (1.3 points lower) relative to women, those who dropped from the study early had on average a QOL score that was about 1.7 points lower than those who remained by the last occasion, and those who dropped from the study late had on average a QOL score that was about 2 points lower than those who remained by the last occasion. Statistically, there was no difference in the effects of drop-out on time 1 QOL scores for those who dropped early versus late. No other background variables were related to QOL scores at time 1. With regard to the annual change rate in QOL scores, statistical tests did not suggest any associations between the background variables and the annual change rate.

Correlations Among Change Characteristics of Different Outcome Variables 

Correlations between change characteristics of each outcome variable, adjusting for effects due to background variables, are given in table 3. In the table are the estimated correlations between the random effects corresponding to time 1 response levels and annual change rates. At time 1 and adjusting for background variables, functional limitation scores were negatively correlated with both exercise behaviors (r=−.34) and QOL (r=−.50) scores, suggesting that at the start of the study higher functional limitations tended to correspond with lower exercise and QOL levels. Additionally, exercise behaviors and QOL scores at time 1 correlated positively (r=.42), suggesting a tendency for higher exercise levels to be associated with high levels of QOL at this time. With regard to the rates of change in the outcome measures, increasing rates of change in functional limitations correlated with decreasing rates of change in both exercise behaviors (r=−.25) and QOL (r=−.35) scores, suggesting that increases in functional limitations tended to correspond to decreases in exercise and QOL levels over the study period. Also, time 1 exercise behavior scores were negatively correlated with the annual change rate in functional limitations (r=−.17), suggesting a tendency for higher exercise levels at the start of the study to be related to less change in functional limitation scores over the study period. The remaining correlations did not differ statistically from zero.

Table 3. Correlations Among Individual-Level Time 1 Responses and Annual Change Rates for 3 Outcome Variables, Adjusting for Background Variables⁎

	Outcome	Individual-Level Effect	Functional Limitations		Exercise Behaviors		QOL
			Time 1 Response Level	Annual Change Rate	Time 1 Response Level	Annual Change Rate	Time 1 Response Level	Annual Change Rate
	Functional limitations	Time 1 response level	1.000
		Annual change rate	−.002	1.000
	Exercise behaviors	Time 1 response level	−.338	−.174	1.000
		Annual change rate	−.109	−.250	−.158	1.000
	QOL	Time 1 response level	−.503	.023	.417	−.017	1.000
		Annual change rate	.014	−.352	−.085	.420	−.144	1.000

⁎ In the multivariate latent curve model, the errors at the occasion level were allowed to covary between outcome variables within occasions.