Promoting Healthy Lifestyles With Aging: Development and Validation of the Health Enhancement Lifestyle Profile (HELP) Using the Rasch Measurement Model

Health Enhancement Lifestyle Profile

The HELP consists of two major sections. The first section consists of personal background and health information such as age, gender, ethnicity, marital status, chronic disease or disabling conditions, and self-rated heath. The second section of HELP includes seven scales measuring different aspects of health-related lifestyle: (1) exercise, (2) diet, (3) social and productive activities, (4) leisure, (5) ADLs, (6) stress management and spiritual participation, and (7) other health promotion and risk behaviors (sample items are available online at www.ajot.ajotpress.net [navigate to this article, and click on “supplemental materials”]). Each scale consists of eight items (questions) that examine the frequency of the respondent's engagement in various health-related activities or events. People are asked to respond to each question according to their typical or routine performance during the past 3-mo period. Response categories generally include (1) never, (2) 1–2 days, (3) 3–4 days, (4) 5–6 days, and (5) 7 days (per wk). However, as suggested by participants in the field pretesting study, an additional category “(6) 1–2 days a month” was added. For scoring, each response was given a numerical value reflective of the relative frequency, which results in a 0- to 5-point rating scale. Scores from negatively worded items are to be reversed. A subtotal score can be computed for each HELP scale, ranging from 0 to 40, where a higher score indicates a more favorable level of lifestyle. The total score of HELP ranges from 0 to 280.

Piloting of HELP was completed through the methods of focus group and field pretesting. A focus group led by the author included a convenience sample of eight community-dwelling older adults solicited from three local senior citizen centers. Discussion involved confirming the definition and scope of various aspects of health-related lifestyle as well as reviewing the content of HELP on an item-by-item basis. Accordingly, revisions were made to enhance the relevance and clarity of this measurement tool so as to better reflect the perspectives of healthy lifestyle among older adults. Subsequently, field pretesting of the revised HELP was conducted with another convenience sample of 16 participants representative of the target population. On the completion of the measure, each participant was given a one-on-one debriefing interview in which the participant related how he or she experienced the HELP questions and identified any confusion or suggestions for improving the understandability of the questions. Each briefing interview was led by one of the six occupational therapist senior graduate students specially trained for this research project. As a result, further revisions and modifications were made to different technical aspects of HELP such as question formation (e.g., separation of double-barrel questions), item wording, and response categories. Positive comments and quotes on HELP were also documented throughout the interviews. Some sample quotes recorded are as follows:

Participants and Data Collection Procedures

Participants in this study consisted of community-dwelling (noninstitutionalized) older adults age ≥55 residing in southern California who were cognitively intact and able to communicate in English. No other specific exclusion criteria were considered. The study was approved by the institutional review board of the California State University, Dominguez Hills. Convenience sampling and snowball sampling methods were used to recruit participants from a diversity of community sites, including senior citizen centers, senior residential communities, independent living facilities, adult day health care centers, local senior social and activities groups, and religious groups or organizations. HELP was completed by each participant through one of the following methods: (1) a direct interview in which a graduate student or a staff member read through questions and recorded the responses for a participant, (2) on-site paper-and-pen administration to an individual or a group of participants, and (3) mail survey.

Data Analysis

Rasch analysis of the Andrich’s Rating Scale Model (Linacre, 2009) was used to determine the internal validity and clinical applicability of the HELP scales. Analysis was conducted through the WINSTEPS Rasch model computer program, Version 3.68.0 (SWREG Inc., 2009). Psychometric characteristics, including unidimensionality (by means of analyses of principal components of residuals, fit statistics, local dependency, and differential item functioning), item hierarchy, person separation index and reliability, and rating scale validity, were examined for each scale of HELP. Details of these Rasch procedures are described along with their findings in the Results section.

Unidimensionality and Data–Model Fit

Unidimensionality of the HELP scales was first confirmed by principal components analysis (PCA) of standardized residuals after an initial Rasch factor has been extracted. The following criteria were adopted to indicate the unidimensional feature of a scale: (1) a cutoff of 60% of the variance explained by the initial Rasch factor, (2) the variance explained by the first contrast in the residuals less than 10%, and (3) the eigenvalue of the first contrast smaller than 3.0 (Linacre, 2009; Smith, 2002). Table 1 depicts the results of PCA with each HELP scale. All but the Social and Productive Activities scale demonstrated >60% of variance explained by the primary Rasch factor, and all seven scales met the other criteria for the variance of residuals (<10%) and eigenvalues (<3.0). The proposed PCA criteria for unidimensionality of HELP were generally met.

Rasch goodness-of-fit statistics further determined how well the items of each HELP scale fit the one-dimensional model of linear measure. In Rasch analysis, Infit MnSq (i.e., information-weighted mean square residual goodness-of-fit statistic) has been considered a more critical and sensitive indicator of unidimensionality than Outfit MnSq (i.e., outlier-sensitive mean square residual goodness-of-fit statistic), because Outfit MnSq is less threatening to measurement and easier to manage (Bond & Fox, 2001; Linacre, 2002). Given the sizes of the study sample and the scales, a HELP item with an Infit MnSq >1.3 or <0.7 in combination with a ZStd (i.e., standardized mean square) >2.0 was used to indicate the lack of fit in unidimensionality (Bond & Fox, 2001; Linacre, 2002). In particular, an Infit MnSq >1.3 indicates 30% more variance than expected and thus suggests that the item measures a construct different from the overall scale. Table 2 demonstrates fit statistics for HELP, including measures (item calibration: logits), standard errors, Infit MnSqs, and ZStds. Of the 56 items from the seven HELP scales analyzed, 3 were identified as a misfit to the Rasch measurement model, based on the criterion used: “How often during a week do you go to work (paid work)?” in the Social and Productive Activities scale, “How often during a week do you talk with a special someone to share with your life, happiness, and sadness?” in the Stress Management and Spiritual Participation scale, and “How often during a month do you take pain medicine to control any form of body pain (such as migraine headache, arthritic pain, or back pain)?” in the Other Health Promotion and Risk Behaviors scale.

The correlation of standardized residuals was used to detect “local dependence” between pairs of items, namely, a subsidiary dimension in the measurement that is not accounted for by the initial Rasch factor (Linacre, 2009). In particular, negative local dependence is often an indicator of multidimensionality. The Social and Productive Activities scale demonstrated a low negative residual correlation (r = –.38) between the items “How often during a week do you go to paid work?” and “How often during a week do you go to volunteer work?” That is, the endorsement of either item may contradict the response to the other. Nevertheless, all other residual correlations within the HELP scales remained at the negligible level (r _s = –0.20 to 0.20).

The testing of differential item functioning (DIF) further identified potential misfit to the Rasch model through the inspection of the differences in item calibration within the data. DIF occurs when item difficulty estimates vary across demographic groups, thus indicating that the scale is not sufficiently unidimensional and that other exogenous factors may sway the responses (Linacre, 1998). For the DIF analysis, the participants were divided by (1) gender; (2) age, based on the median of 72 (“young-old,” 55–72 yr, vs. “old-old,” 73–92 yr); and (3) race (White vs. non-White), respectively. The ethnicity categories were dichotomized because of the relatively small and disproportionate samples among the non-White subgroups. The criteria used for the DIF analysis were DIF contrast >0.50 (a difference of 0.5 logits in item difficulty calibrations between two groups) and p < .006 (adjusted for multiple comparisons by a Bonferroni correction; Linacre, 2009). The HELP scales were mostly free of DIF. However, one item in the Leisure scale regarding “carpentering, auto/boat/house fixing, or any other mechanical work for your hobby” was easier to endorse for male participants (DIF = 0.58 logits, t[251] = 2.87, p = .003). The previously identified problematic item “How often during a week do you go to paid work?” in the Social and Productive Activities scale was also found to be easier to endorse for “young-old” participants (DIF = 0.79 logits, t[251] = 3.40, p = .001).

Item Hierarchy

The Rasch model establishes the hierarchy of items and people along an equal-interval continuum. The item measures expressed in logits (see Table 2) were used to arrange the items within each HELP scale in their hierarchical order of difficulty from easiest at the top to most difficult at the bottom. For example, for the Diet scale, the first item, “healthy foods rich in protein” that was the easiest to endorse had a difficulty estimate of –1.19 logits. Consecutively, the second item, “healthy foods rich in calcium,” demonstrated a difficulty estimate of –0.72 logits, and, proceeding with the continuum, the last (most difficult to endorse) item, “two servings of sweets or desserts,” revealed the highest difficulty estimate (1.33 logits) of the scale.

Person Separation Index and Reliability

The person separation index (PSI), strata, and reliability coefficient of each HELP scale are summarized in Table 3. The number of strata was calculated using the formula Strata = (4PSI + 1)/3 (Wright & Masters, 1982). The greater the strata, the more distinct levels of healthy lifestyle were revealed from the sample. Acceptable to good separation and reliability indexes were demonstrated, and the resultant distinct number of healthy lifestyle strata for each HELP scale approximated to three or more.

Rating Scale Validity

Table 4 reports the analysis of the functioning of rating scale structure (0- to 5-point scale) for each HELP scale. The results include the observed average measure, model expected measure, MnSq fit statistics, and step calibration of each rating category.

First, the observed average measure is an empirical indicator of the context in which the rating category is used (Linacre, 1999a, 2009). In general, observations in higher categories must be consistently produced by higher measures. That is, no category disordering should be found; otherwise, the meaning of the rating scale is indecisive for that data set, and consequently any derived measures can be of minimal utility. As seen in Table 4, the average measures by category advance monotonically up the rating scale. This pattern of increase is an empirical confirmation of scale functioning that higher rating categories indicate more of the latent variable measured (i.e., healthy lifestyle). Correspondingly, the expected measures are the values that the Rasch model predicts for the measures of various rating categories (Linacre, 1999a, 2009). The residual difference between the expected values and observed measures assists in determining the degree to which data obtained using the existing rating scale would fit the Rasch model. As shown in Table 4, most rating categories in the HELP scales appear to have their observed average measures reasonably close to the expected values.

MnSq fit statistics for each rating category are also reported in Table 4. For rating scales, a high MnSq fit suggests that there is more unexplained noise than explained noise, thus indicating the existence of misinformation in the observations on the basis of the use of the rating scale. In general, a rating category is thought to be used in an unexpected manner if its unweighted MnSq fit statistics (Outfit MnSq) is >2.0 (Linacre, 1999a, Wright & Masters, 1982). As can be seen in the table, all the rating categories in the HELP scales were found to have an Outfit MnSq smaller than the critical value of 2.0.

Step calibration (or Rasch–Andrich threshold) is a Rasch-model parameter that indicates a relative pairwise measure of the transition between the adjacent rating categories (Linacre, 1999b). A positive value of step calibration indicates that the lower of the pair of categories is more likely to be observed; conversely, a negative value indicates that the higher category of the pair is more likely to be observed. The step calibration is expected to increase with category value. As shown in Table 4, there was no step disordering; the criterion of monotonic advance was met for each set of step calibration in the HELP scales.