194
R. Maritz et al.
to the model is achieved, an interval-scaled metric can
be derived from ordinal scales (14, 15).
Earlier analysis of the FIM™ using Rasch analysis
in the 1990s indicated that the FIM™ 18-item version
incorporates 2 different constructs, represented by a
motor scale and a cognitive scale, each of which should
be scored separately (16). However, in clinical practice
both the reporting of 2 separate motor and cognitive
total scores and the reporting of a single total score of
the FIM™, is evident (7, 9, 11). Since this first Rasch
analysis of the FIM™, many others have been publish
ed, mostly on its motor subscale (17), but also on
adaptations of the FIM™ (18, 19). More recently, the
issue of so-called local item dependency has received
attention (20). Local item dependency occurs when
instrument items remain correlated when conditioned
on the trait, what is functional independence in the
case of the FIM™. Local dependency is indicated by
significant correlation of the standardized analysis
residuals. Fit of the FIM™ motor scale to the Rasch
model has been shown to be seriously affected by local
item dependency, which, once accommodated, resulted
in adequate model fit (17).
Thus, given the recent methodological developments
with regards to addressing the issue of local depen-
dency in health scales, and inconsistency in reporting
the FIM™ in practice, a review of the FIM™ 18-item
version seemed appropriate, in order to address the
following question: Is it possible to add all FIM™
items together to obtain a valid unidimensional total
score, taking into account the local dependency in its
item set? The objective of this study was therefore to
revisit the question of whether the FIM™ can be re-
ported as a unidimensional interval-scaled metric when
local dependency is taken into account. Two specific
aims in relation to the study’s objective were: (i) to
explore the metric properties of the FIM™; and (ii) to
determine whether an interval-scale scoring system of
the FIM™ 18-item version can be made available and,
if so, to create an interval-scale transformation of the
FIM™ raw scores when administered in the context
of national quality monitoring in neurological and
musculoskeletal rehabilitation.
METHODS
Subjects and setting
Data collected routinely for the Swiss national quality reporting,
coordinated by the ANQ, was used for secondary analysis. All
64 Swiss rehabilitation clinics that provided data to the ANQ
in 2016 for musculoskeletal or neurological rehabilitation were
contacted, of which 30 voluntarily agreed to provide their ANQ
datasets. Since the clinics can choose between different assess-
ment tools in ANQ data collection, not all datasets contained
FIM™ data. Thus, this study used datasets from 23 rehabilitation
www.medicaljournals.se/jrm
clinics, with 11,103 complete cases in total, representative of
3 different Swiss language regions (German, French, Italian).
The FIM™ was administered at admission and discharge. Ethics
approval for the study was requested from the Swiss Ethics
Commissions, which stated in a declaration of no objection that
the project fulfils the general ethical and scientific standards for
research with humans and poses no health hazards.
Functional Independence Measure
The FIM™ is an assessment tool comprising 18 items. Thirteen
items belong to the motor subscale and 5 items belong to the
cognitive subscale. All items are scored from 1 (total assistance)
to 7 (complete independence). The FIM™ item scores are sum-
med up to a total score, ranging between 18 and 126, or total
motor score ranging between 13 and 91 and between 5 and 35
for the cognitive total score (4). The ANQ used German, French
and Italian translations of the FIM™ based on its official English
version, on which a translation agreement was made with the
Uniform Data System for Medical Rehabilitation (UDSMR).
As this is common practice, the translations have not been
authenticated by the UDSMR. In order to qualify to administer
the FIM™ , the health professionals received training provided
by the ANQ according to the respective UDSMR policy.
Sampling
A random stratified calibration sample was created using R (21),
since type I errors, i.e. rejecting a hypothesis even if it was true,
are likely to appear with a large sample size in Rasch analysis
(22). The aim was to create a sample of approximately 1,000
cases, representing 4 equally sized subsamples, each with suf-
ficient sample size for a stable item calibration and statistical
interpretation (23, 24). Each subsample focused on one of the 2
different time-points of measurement, and one of the 2 different
health condition groups of musculoskeletal and neurological
rehabilitation: musculoskeletal cases at admission (MSKt1),
musculoskeletal cases at discharge (MSKt2), neurological cases
at admission (NEURt1) and neurological cases at discharge
(NEURt2). To obtain precision across the whole range of scores
(total score range 108; 18–126) and representation of language
regions, a random sample was taken from each available total
score per subsample and language region group. Cases that
were selected from the admission subsamples were excluded
and not selected for the discharge subsamples (25). Prior to the
random selection all cases with missing values in a person’s
contextual factors of interest (described in more detail below)
and all cases that scored an extreme score (18 or 126), were
deleted, since they are excluded from the calculation of item
difficulties by the Rasch measurement model. The sampling
strategy is shown in Fig. 1.
Data analysis
To summarize basic sample characteristics and response dist-
ributions of the FIM™, descriptive statistics were conducted
with Stata Version 14.2 (26). In order to achieve the study’s first
specific aim Rasch analysis was conducted using RUMM2030
(27). The analytical focus gave reference to local response
dependency represented by residual correlations. High residual
correlations indicate that items are measuring the same thing too
closely (13). Furthermore, threshold disordering was examined,
which indicates that the different response categories of an item
are not in a successive order, i.e. do not represent an increasing
level of functional independence. In addition, differential item