Descriptive data on the characteristics of homeless individuals, studies of risk factors for homelessness in veterans, and the results of housing interventions strongly suggest that the veteran homeless population is heterogeneous. This heterogeneity likely underlies limitations of the existing homeless programs to achieving higher levels of program completion and avoiding subsequent episodes of homelessness. Failure to appreciate the complexity of factors that contribute to homelessness, and especially chronic homelessness, stands in the way of attempts to develop more specifically tailored housing interventions that might improve housing intervention outcomes and of strategies to optimize the delivery of existing healthcare services. Higher successful completion rates in homeless programs could be achieved, in part, by utilization of an assessment instrument that would provide a comprehensive evaluation of the homeless veteran that would drive decisions regarding the type, duration, and intensity of homeless interventions. In this research we developed an alpha version of the Homelessness Severity Index (HSI), an instrument designed to meet this assessment need.
The purpose of this one-year year pilot project is to develop an alpha version of a multiscale instrument designed to provide a global estimate of the severity of homelessness and estimates of standing on subscales measuring important domains of homelessness (e.g., alcohol abuse, health, social adjustment).
We used a multi-step method to develop an alpha version of the HSI, which included the following sequential steps: 1) development of item pool for eight domains (e.g., alcohol use, housing history); 2) review by expert panel to evaluate domain coverage, suggest additional items, recommend item changes and deletions, edit items; 3) collect data on 121-item version with sample of homeless veterans; 4) conduct classical test theory (CTT) and Mokken analyses of sample data; 5) review item/scale characteristics produced by CTT and Mokken scaling methods to determine final item composition for each scale; and 6) examine characteristics of resulting scales.
In the CTT analyses items in each domain are treated as a potential scale and item (descriptive statistics, intercorrelations, corrected item-total scale correlations (CITCs), and squared multiple correlations (SMCs) and scale (descriptive statistics, internal consistency reliability coefficients) characteristics are examined. An iterative process is used for each potential scale in which non-contributory or weak items are individually deleted from the scale item pool and item/scale characteristics are re-examined. This process is continued until the scale internal consistency reliability coefficient (alpha) is maximized. In the CTT approach, extreme scale homogeneity (internal consistency reliability .90) is not beneficial, indicating a scale with limited focus that reduces validity. The CTT approach aims to examine items that represent different facets of the domain. Scales developed with this approach are therefore often not factorially pure (e.g., the Beck Depression Inventory).
Mokken scaling is based on the idea that items meaningfully form a scale if they satisfy the conditions of the monotone homogeneity model (MHM; unidimensionality, local independence, and monotonicity). Under these conditions it makes sense to summarize the set of items with a single score because having a higher score on the instrument will (in most cases) indicate a higher probability of having a larger value of the latent trait. One implication of satisfying the MHM is that each pair of items in the scale will have a sizable positive covariance with each other item; if this holds in the observed test data then it is a sign that the underlying model is at least approximately MHM. In Mokken scaling the sample covariances are rescaled into scalability coefficients that have a maximum possible value of one. Scalability is measured at the item pair level (every item pair should have a positive item pair scalability Hij), at the item to scale level (every item should have a reasonably large scalability with the scale as a whole, say Hi 0.3), and as an entire scale (an H of 0.3 = weak scale, 0.4 = moderate scale, 0.5=strong scale).
Following IRB approval, an item pool of 130 items reflecting eight domains was assembled and reviewed by an expert panel of 10 experienced housing intervention program case managers. After item deletions/additions and editing, a 121-item version was formatted and used in individual interviews to collect data from 107 homeless veterans receiving homeless services. These data were entered into a database and CTT analyses were subsequently conducted using SPSS. Mokken analyses were conducted by a consultant specialist using a modification of Sijtsma and Molenaar's AISP algorithm.
Initial CTT item analyses revealed skewed distributions for several items. These items were recoded into dichotomous or limited range likert format (3-4 distribution points) based on percentile scores. It was expected that this procedure would reduce item variance and attenuate scale item intercorrelations and internal consistency reliability coefficients. Given this restriction, a target goal of alpha .60 was set.
In initial iterations, the identification/history (e.g., years of education, combat), alcohol, work/finance, housing, and substance abuse scales did not achieve the target alpha. The work/finance and housing scales were therefore combined to form a work/housing scale and the alcohol and substance abuse scales were combined to form a single alcohol/substance abuse scale. Subsequent iterative analyses provided support for the five resulting scales:
Health (8 items, alpha = .65, score range = 0-13, scale M = 5.46 (SD = 2.78),
Mental Health (11 items, alpha = .80, score range = 0-13, scale M = 3.61 (SD = 3.35),
Social Adjustment (9 items, alpha = .79, score range = 0-10, scale M = 4.13 (SD = 2.81),
Work/Housing (11 items, alpha = .64, score range = 2-16, scale M = 9.03 (SD = 3.41),
Alcohol/Substance Abuse (12 items, alpha = .72, score range = 0-13, scale M = 4.25 (SD = 2.99).
In the Mokken analyses, no reliable scales were formed for the Work, Housing, Health or Alcohol drug sets of items. Mokken analyses did produce a reliable 10-item scale from the pool of Mental Health items (H=0.52, minHi=0.36, and a=0.82). Ten of these items overlapped with the items of the CTT-development Mental Health Scale. Analyses of the Social Adjustment items produced a 7-item scale with good reliability (H=0.65, minHi=0.44, and a=0.80). All of these items overlapped with items in the CTT-developed Social Adjustment Scale.
Principal component analyses of the five CTT-developed scales revealed all to be multidimensional. In each case factors represented facets of the overall domain (e.g., for the Mental Health scale: symptoms, emotional control). An HSI total score was calculated by summing the individual raw scores for the five scales. Correlations with individual scales ranged from .46 to .72. This score was normally distributed, suggesting that it would provide an adequate measure of individual differences in severity of homelessness condition.
The project successfully developed an alpha version of a severity scale to measure individual differences in the status of homeless veterans relevant to housing intervention. Further development focusing on concurrent and predictive validation studies in various homeless populations would allow determination if the instrument has potential value in triage and treatment planning for homeless veterans.
External Links for this Project
Grant Number: I01HX000571-01A1
None at this time.