Rasch_Medvedev

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

689603

research-article2017
HPQ0010.1177/1359105316689603Journal of Health PsychologyMedvedev et al.

Article

Journal of Health Psychology

Rasch analysis of the Perceived


2019, Vol. 24(8) 1070­–1081
© The Author(s) 2017
Article reuse guidelines:
Stress Scale: Transformation from sagepub.com/journals-permissions
DOI: 10.1177/1359105316689603
https://doi.org/10.1177/1359105316689603
an ordinal to a linear measure journals.sagepub.com/home/hpq

Oleg N Medvedev1, Christian U Krägeloh1,


Erin M Hill2, Rex Billington1, Richard J Siegert1,
Craig S Webster3, Roger J Booth3
and Marcus A Henning3

Abstract
Rasch analysis was conducted to enhance the precision of the widely used 10-item Perceived Stress Scale
using two datasets (n = 450 each) randomly selected from samples of the New Zealand general population
(n = 1102), New Zealand university students (n = 479) and US university students (n = 396). The best Rasch
model fit (χ2(27) = 29.92, p = .36), good person separation reliability (.80) and coverage (98%) of the sample
by the scale items were achieved when locally dependent items were combined into subtests. These findings
support reliability and internal structural validity of the 10-item Perceived Stress Scale. The instrument
precision can be further improved using the ordinal-to-linear conversion tables published here.

Keywords
measurement, Perceived Stress Scale, psychometrics, Rasch analysis, stress

Introduction
Stress is described as heightened emotional states it reflects the subjective evaluation of environ-
associated with physiological changes (Helton mental events (Bloch et al., 2004), which directly
and Näswall, 2015; McEwen and Stellar, 1993). influences physiological responses responsible
Research has shown that extended exposure to for aversive health effects (LeDoux, 2000;
stress can lead to aversive health effects (Cohen Medvedev et al., 2015).
et al., 1998; Hillhouse et al., 1991). Both effective
stress management and stress reduction are criti-
1Auckland University of Technology, New Zealand
cal in reducing negative stress effects on individ-
2West Chester University, USA
ual’s health. Development of effective methods to 3The University of Auckland, New Zealand
manage and reduce stress requires accurate
assessment of perceived stress levels to evaluate Corresponding author:
and compare unique contributions of various pre- Oleg N Medvedev, Department of Psychology, Faculty of
Health and Environmental Sciences, Auckland University
dictors, situations and behaviours that may trig- of Technology, North Shore Campus, Private Bag 92006,
ger and maintain stress. In particular, precise Auckland 1142, New Zealand.
assessment of perceived stress is critical because Email: oleg.medvedev@aut.ac.nz
Medvedev et al. 1071

The Perceived Stress Scale (PSS) (Cohen gender, ethnicity and education level catego-
and Williamson, 1988) was constructed as a ries. Sharp et al. (2007) tested the performance
subjective measure of perceived stress to assess of the PSS-10 items in a clinical sample of
the extent to which a person’s life is perceived asthma patients and reported that few items
as ‘unpredictable, uncontrollable, overloading’ function differently across ethnic and literacy
(p. 387), relative to the individual’s coping abil- factors. Recently, Taylor (2015) used the graded
ities. The scale is very widely used, exceeding response IRT model and reported satisfactory
11,000 citations by the middle of 2016, accord- functioning of individual items. These incon-
ing to Google Scholar. The PSS has been cross- sistent findings suggest that further investiga-
culturally validated and translated into 25 tion is necessary. Moreover, these findings do
different languages (Cohen, 2013). The original not provide feasible solutions to improve the
PSS version contains 14 items and has good overall precision of the PSS-10.
internal consistency (Cronbach’s alpha > .80) Technically, an ordinal scale such as the PSS
and satisfactory construct validity (Cohen and should not be used with parametric statistics
Williamson, 1988). The authors subjected the without violating their fundamental assump-
14-item PSS to principal component analysis tions. However, researchers have used the PSS
(PCA) and found four items that displayed poor with parametric statistics (Chavez-Korell and
loadings on the first principal component in the Torres, 2014; Gitchel et al., 2011), which could
range of 0.11–0.39, which were removed result- potentially result in misleading conclusions and
ing in the popular 10-item version of the PSS implications of the findings linked to individu-
(called PSS-10). A 4-item PSS version was al’s health and well-being. It has been demon-
also introduced as a quick assessment tool, but strated that adding ordinal scores of individual
the PSS-10 demonstrated better internal con- items together would not produce an accurate
sistency (alpha = .78) compared to the 4-item total estimate because each individual item
PSS version (alpha = .60) and was recom- explains different amount of variance due to a
mended by the authors for use in future research latent trait (Allen and Yen, 1979; Stucki et al.,
(Cohen and Williamson, 1988). The PSS-10 has 1996). Moreover, usage of the ordinal PSS-10
been used with different populations for both in research may affect comparisons with neuro-
clinical assessments and empirical investiga- physiological interval or ratio-level data (e.g.
tions including validation studies, all confirm- electroencephalogram, skin conductance and
ing its satisfactory psychometric properties heart rate), an especially important considera-
(Mitchell et al., 2008; Roberti et al., 2006; tion in stress research. Therefore, it is important
Taylor, 2015). Currently, there is no agreement to improve precision of the PSS-10 up to an
on the factor structure of the PSS-10, with some interval-level scale, which can be conducted
studies including the original validation report using Rasch analysis – an approach that is par-
confirming unidimensionality (Cohen and ticularly suited for this purpose (Rasch 1960;
Williamson, 1988; Cole, 1999), while others Tennant and Conaghan, 2007).
argue that a two-factor solution provides a bet- The dichotomous Rasch model was devel-
ter fit (Barbosa-Leiker et al., 2013; Taylor, oped first (Rasch, 1960), which was followed by
2015; Teh et al., 2015). two models designed for polytomous items,
There are only few studies that used item such as used in the PSS-10, including the Partial
response theory (IRT) to investigate function- Credit version (Masters, 1982) and the Rating
ing of the individual PSS-10 items (Cole, 1999; Scale version (Andrich, 1978). Both models
Sharp et al., 2007; Taylor, 2015). Cole (1999) based on assumption that differences between
used IRT to investigate potential item bias of thresholds of individual items vary. However,
the PSS-10 with a larger US sample (n = 2264). the Rating Scale model assumes that these vari-
Significant, but very small, differences in per- ations are similar across all items but the Partial
formance of several items were reported by Credit model allows thresholds distances to vary
1072 Journal of Health Psychology 24(8)

across items, so that every item has individual applied to a wide range of clinical and non-clin-
rating scale parameters. Both polytomous Rasch ical populations. However, its ability to dis-
models are commonly used in health assessment criminate precisely between perceived stress
to test and improve psychometric properties of levels has not been investigated in sufficient
ordinal measures with two or more response detail. Rasch analysis is a suitable method to
options (Hobart and Cano, 2009; Lundgren investigate the ability of the scale and individ-
Nilsson and Tennant, 2011). The decision which ual items to discriminate on their overarching
polytomous model to use depends on the likeli- latent factor. However, to the best of our knowl-
hood-ratio test, which compares threshold dis- edge, this advanced technique has not yet been
tances between individual items prior to main applied to scrutinize and to improve the psycho-
analysis. The unrestricted Partial Credit model metric properties of the PSS. This study aims to
will be used if these distances are significantly apply Rasch analysis to explore strategies to
different (Tennant and Conaghan, 2007). The improve the psychometric properties of the
end product of Rasch analysis is transformation PSS-10 up to an interval level of measurement.
of scores from an ordinal to an interval scale that This psychometric investigation explicitly
increases precision of measurement, which was focuses on the measurement of the overarching
established both theoretically (Rasch, 1960; latent factor of perceived stress, rather than its
Tennant and Conaghan, 2007) and empirically underlying facets already explored in the litera-
(Norquist et al., 2004). ture (Roberti et al., 2006; Taylor, 2015). To
Rasch analysis has demonstrated its distinct ensure that results are generalizable to diverse
advantages over other more traditional psycho- populations, a combined sample of respondents
metric methods, which have been discussed in from New Zealand and the United States is
detail elsewhere (Rasch, 1960; Wilson, 2005). used, consisting of university students as well
The analysis involves testing several important as respondents from the general population.
psychometric parameters including potential Additionally, the dataset is sufficiently large to
item bias, local independence assumptions (e.g. allow splitting of the combined sample into two
unidimensionality), correct stochastic ordering sets, thus allowing replication of the Rasch
of items and response option ordering in polyto- analysis of one half of the sample with the other
mous items (Tennant and Conaghan, 2007). half. An ordinal-to-linear conversion table was
When data meet these requirements and a fit to generated, which can be used to enhance preci-
the Rasch model is achieved, participants can sion of the PSS-10.
be located according to their ability on a scale
measuring the latent factor (in this case per-
ceived stress). Indeed, in Rasch analysis, both Method
the participants and the items are ordered on an
interval scale using the same log-odds metric
Participants
(Pomplun and Custer, 2005). This results in the This study combined three independently col-
graphical representation of an item-person lected samples. Sample 1 (n = 1102) consisted of
threshold distribution, showing how well the the Auckland (New Zealand) general population
range of the items’ difficulty covers the ability who participated in a postal survey on noise sen-
of a sample. Rasch analysis provides a system sitivity and health (Hill et al., 2014). The mean
that can accurately and rigorously investigate age was 51 (standard deviation (SD) = 16.42),
this perceived stress measure in terms of poten- and 65 per cent were female. Sample 2 (n = 479)
tial item bias, differential item functioning contained New Zealand university students
(DIF) and unidimensionality. enrolled in various health science courses at the
In sum, it is assumed that the PSS-10 is a University of Auckland and Auckland University
measure of perceived stress and has generally of Technology. The majority (76%) were female,
accepted psychometric properties that can be and the overall mean age was 19.96 (SD = 4.47).
Medvedev et al. 1073

Sample 3 (n = 396) consisted of US university research credit. The study was conducted in
students enrolled in first-year psychology compliance with the authors’ institutional ethics
classes at West Chester University in the United committees.
States. The mean age was 19.18 (SD = 2.21), and
45 per cent were female. From each of these
Measures
three samples, 300 participants were randomly
selected to create an overall sample of 900 par- While the questionnaires in the three above-
ticipants from the New Zealand general popula- mentioned studies contained various scales, this
tion, New Zealand university students and US study focuses on the PSS-10 only. The PSS-10 is
university students. Each subset was randomly a 10-item self-report questionnaire of perceived
divided by half to create two samples of 450 par- stress operationalized as subjective evaluation
ticipants, where 150 participants were included of lack of control, unpredictability and overload
from each sample population. The main analy- in participants’ daily life (Cohen and Williamson,
ses were conducted using one of these two sam- 1988). The instrument uses a 5-point Likert-
ples and were then replicated with the second. scale response format (1 = ‘Never’ to 5 = ‘Very
As the ethnic profile for the New Zealand often’), and a total score is calculated after
and US samples was very different, common reverse-coding items 4, 5, 7 and 8 and then add-
categories were created so that DIF by ethnicity ing scores of all 10 items together.
could be compared. In the overall sample of
900, 65 per cent were classified as Caucasian,
Data analysis
5 per cent as Polynesian, 9 per cent as Asian and
19 per cent as other. Descriptive statistics and reliability of the PSS-
10 were computed using IBM SPSS v.22, and
Rasch analysis used RUMM2030 (Andrich
Procedure et al., 2009). A likelihood-ratio test was con-
The procedure for the New Zealand general ducted first and indicated that differences
population (Sample 1) is reported in detail else- between response options thresholds of indi-
where (Hill et al., 2014). Auckland residents vidual items vary significantly across items
received a questionnaire on noise sensitivity, meaning that the unrestricted (Partial Credit)
perceived stress and health, which they subse- version of the Rasch model should be used for
quently posted back to the researchers using a the current dataset. The Rasch analysis is con-
self-addressed pre-paid envelope. Participants ducted in an iterative way until all individual
in Sample 2 were university students enrolled in items show sequential ordering of response cat-
health science courses at two major universities egories thresholds, satisfactory overall and
in New Zealand completing a survey on moti- individual item fit to the model are achieved
vation to learn, quality of life and perceived and unidimensionality is clearly evident
stress. Students at the University of Auckland (Siegert et al., 2010). Category threshold is dis-
received an invitation to complete an online ordered if a probability to choose the closest
survey and thus responded at a time of their higher response option is lower than any of sub-
convenience. Students at Auckland University sequent higher response options. The Rasch
of Technology were approached in lecture thea- model fit is evaluated by the mean item and per-
tres and completed the questionnaire during the son location, individual item fit residual and the
lecture break or after the lecture. For Sample 3, overall item-trait interaction chi-square test/p
students taking an introductory psychology value using the following criteria (Gustafsson,
class at West Chester University in Pennsylvania, 1980; Tennant and Conaghan, 2007):
United States, completed a research study on
motivation to learn, quality of life and per- 1. The item location mean is used as a base
ceived stress as an option in fulfilling their class and set to zero.
1074 Journal of Health Psychology 24(8)

2. A person location mean ±0.5 indicates a et al., 1983). Combined items show higher reli-
good coverage of a sample by a scale. ability compared to individual items, more
3. In case of an overall perfect fit, both scale points that contribute to accuracy of
item and person fit residual are 0.00 measurement and lower risk of spurious corre-
(SD = 1.00). lations. Moreover, more accurate estimates of
4. Individual items fit residuals should be latent structures were obtained using item par-
in the range from −2.50 to +2.50. cels compared to individual items because
5. The item-trait interaction chi-square is combining items measuring the same construct
an indicator of the overall model fit and reduces measurement error due to an individ-
should be not significant (p > .05) if data ual item.
fit the Rasch model. The basic assumption of the Rasch model is
unidimensionality of a scale (Rasch, 1960,
The person separation reliability, which is 1961). Smith’s (2002) unidimensionality test is
similar to a Cronbach’s alpha numerically commonly used in Rasch analysis with
(range from 0 to 1), is not an index of Rasch RUMM2030 software (Andrich et al., 2009)
model fit but indicates how well individual trait that involves PCA of the residuals and the
levels are spread out along the scale continuum equating t test. Unidimensionality of the scale is
represented by the items (Fisher, 1992). evident if significant t-test comparisons do not
Both the overall and the individual item fit exceed 5 per cent or if the lower bound of a
to the Rasch model could be affected by local binominal confidence interval computed for the
dependency between items, which refers to a number of significant t-tests overlaps 5 per cent
situation when two or more items are associ- cut-off point. Rasch model requires no signifi-
ated in some way. For example, one item about cant differences (DIF) in item functioning due
negative emotions asked about ‘been upset’ to personal factors including gender, age, sam-
and the other about ‘been angered’. If a person ple (e.g. US university students vs New Zealand
is often getting angry, they should also fre- university students vs New Zealand general
quently get upset. Such relationship violates population), ethnic groups and education levels.
local independency assumption, compromises DIF analysis involves comparing distributions
estimation of model parameters and inflates of individual scores aggregated by class inter-
reliability (Wright, 1996). Local dependency val mean scores between groups for each per-
between two or more items is reflected by the son factor and for every individual item using
residual correlations matrix. A residual correla- analysis of variance (ANOVA) and t-tests for
tion with magnitude more than 0.20 compared pairwise comparisons (Bonferroni adjusted for
to the mean of all residual correlations is a sign the number of tests).
of local dependency (Christensen et al., 2017; The first sample of 450 was used for the
Marais and Andrich, 2008). In this study, we main Rasch analysis, which was then replicated
considered removing items as the last resort to using the second half of the sample. Both sam-
achieve the model fit. Instead of removing ples (datasets a and b) contained a large enough
locally dependent items, these items can be number to satisfy the recommended sample size
simply added together into a subtest or testlet estimates for the Rasch analysis (Linacre,
to solve local dependency issues (Lundgren- 1994). To examine if there is any significant
Nilsson et al., 2013; Wainer and Kiely, 1987). difference to participant estimates of the modi-
Using the example with negative emotions, the fied version, participant estimates for the origi-
subtest would be similar to one item measuring nal scale and the modified version were
both aspects (get upset and angered). Subtests compared using a paired-samples t-test. Finally,
in Rasch analysis are analogous to item par- ordinal-to-interval transformation scores were
cels, and their advantages have been well docu- computed that allow users to transform ordinal
mented elsewhere (Little et al., 2002; Rushton data to an interval-level scale.
Medvedev et al. 1075

Results matrix between all PSS-10 items was generated


and confirmed higher correlations between
Cronbach’s alpha for the PSS-10 with the cur- these items in the range of 0.50–0.60. These
rent dataset (N = 900) was .88 indicating good observations together confirmed three groups
internal consistency with all 10 items having of locally dependent items, which were com-
item-to-total correlations in the range from 0.49 bined into three subtests to address local
to 0.74. Table 1 provides summary of the Rasch dependency (Lundgren Nilsson et al., 2013).
model fit statistics for the initial and the final These minor modifications provided good solu-
analysis including both the overall and the indi- tion to improve the overall scale function that
vidual item fit indices for the sample (a) and the satisfied the expectations of the Rasch model in
overall fit indices for replication of the analysis both samples (χ2(27) = 29.92, p = .36 (a),
with sample (b). Overall, the person location χ2(27) = 36.30, p = .11 (b); Table 1). At this
mean was within the acceptable range ±0.50 in stage, the person location mean was close to
both samples suggesting good targeting of the zero in both samples (Table 1) and indicated
sample by the scale items (Table 1). Also, the better coverage of the sample by the subtest
scale showed good person separation reliability items with acceptable person separation relia-
of .88. However, in both samples, the initial bility (.80–.81). Additionally, none of the indi-
overall fit to the Rasch model was affected by vidual subtest items showed misfit to the Rasch
significant item-trait interaction indexed by chi- model (Table 1, 2a). Therefore, the best model
square test: χ2(90) = 129.26 for sample (a) and fit was achieved without the need to remove
χ2(90) = 150.62 for sample (b), p < .001 (Table any of the PSS-10 items.
1). This means that the scale cannot adequately
discriminate between respondents at different
levels of the latent trait (perceived stress). DIF
Although no items with disordered thresholds ANOVA indicated significant DIF effects by
were identified, item 10 displayed significant sample on the subtest one (F(2, 6) = 6.27,
misfit to the Rasch model with fit residual p = .002, Bonferroni adjusted p = .006), but not
below the acceptable cut-off point of −2.50 for the subtests two and three. However, post
(Table 1) in both samples. Additionally, item 4 hoc comparisons between sample groups
showed deviation from the model expectations showed no significant difference with p ranging
with a positive fit residual above 2.50 in sample from .065 to .217. There were no significant
(b) but not in sample (a). DIFs in functioning of subtest items due to
Given the overall poor fit and misfits at other personal factors including gender, age,
individual item level, the residual correlation ethnic groups and education levels (p > .05,
matrix was examined because local dependency Bonferroni adjusted).
between items affects both discrimination
parameters and test information associated with
the Rasch model fit (Lundgren Nilsson et al., Test for unidimensionality
2013). To test the unidimensionality of the final model
solution, the person estimates from the subtest
with the highest positive loadings on the first
Local dependency
principal component were compared with the
Examination of the residual correlation matrix estimates from the subtest with the highest neg-
indicated correlations with magnitude above ative loadings. Unidimensionality was con-
critical value of 0.20 for items 1, 2 and 9; items firmed for both samples with 5.19 per cent
4, 5, 7, and 8; and items 6, 10 and 3; indicating significant t-tests overlapping 5 per cent cut-off
local dependency between those items. To ver- point on the lower bound of the confidence
ify this observation, a Spearman’s correlation interval (3.16%) for sample (a) and 4.74 per
1076

Table 1. Summary of the Rasch model fit statistics for the initial (1a) and final (2a) analysis of the PSS-10 (n = 450) and final analysis of its subsequent
replication (2b; n = 450).

Tests items/scale Item Item fit Person Person fit Goodness of fit Significant t-tests

Location Residual Location Residual χ2 (df) p % LB


1a 1. Been upset −0.23 0.35 – – 17.20 (9) .046 – –
1a 2. Unable to control −0.22 −1.59 – – 7.73 (9) .561 – –
1a 3. Nervous-stressed −1.10 −0.94 – – 9.60 (9) .384 – –
1a 4. Felt confident 0.84 1.60 – – 10.45 (9) .315 – –
1a 5. Things-your way 0.41 0.02 – – 15.87 (9) .069 – –
1a 6. Could not cope −0.13 1.67 – – 4.42 (9) .881 – –
1a 7. Control irritations 0.27 2.12 – – 20.63 (9) .014 – –
1a 8. On top of things 0.36 0.74 – – 9.58 (9) .386 – –
1a 9. Been angered −0.23 1.83 – – 10.09 (9) .343 – –
1a 10. Could not overcome 0.02 −4.03* – – 23.63 (9) .005 – –
1a Full scale: mean (SD) 0.00 (0.52) 0.18 (1.92) −0.25 (1.22) −0.45 (1.44) 129.26 (90) .004 10.67 8.65
1b Full scale: mean (SD) 0.00 (0.42) 0.11 (1.89) −0.22 (1.21) −0.48 (1.47) 150.62 (90) .000 12.67 10.65
2a 1. Subtest (1, 2, 9) −0.12 −0.03 – – 3.52 (9) .940 – –
2a 2. Subtest (4, 5, 7, 8) 0.32 2.30 – – 12.50 (9) .186 – –
2a 3. Subtest (3, 6, 10) −0.20 −1.62 – – 12.88 (9) .168 – –
2a Full scale: mean (SD) 0.00 (0.28) 0.22 (1.97) −0.09 (0.75) −0.54 (1.05) 29.92 (27) .365 5.19 3.16
2b Full scale: mean (SD) 0.00 (0.21) 0.18 (1.64) −0.11 (0.80) −0.57 (1.11) 36.30 (27) .109 4.74 2.71

SD: standard deviation; PSS: Perceived Stress Scale; LB: lower bound of the 95 per cent confidence interval surrounding t-test.
*Significant misfit to the Rasch model.
Journal of Health Psychology 24(8)
Medvedev et al. 1077

Table 2. Converting from a raw PSS-10 score cent of the participants’ abilities on the latent
(10–50) to an interval scale in logit units and using factor (perceived stress).
the original scale metrics.

Ordinal Interval Ordinal Interval Equating t-test


measure measure measure measure
The means of person estimates from the origi-
Raw Logit Scale Raw Logit Scale nal PSS-10 and the modified version were com-
score score
pared by a paired-samples t-test. The difference
10 −3.34 10.00 31 0.19 32.71 between the person estimates of the two ver-
11 −2.66 14.37 32 0.28 33.28 sions was significant (t(449) = 7.22, p < .01),
12 −2.21 17.23 33 0.37 33.85 indicating successful alteration of the ability of
13 −1.92 19.11 34 0.45 34.40 the final model to discriminate between indi-
14 −1.70 20.53 35 0.54 34.96 vidual levels of perceived stress compared to
15 −1.52 21.68 36 0.63 35.52 the original version. This confirms that the
16 −1.37 22.66 37 0.71 36.08 implemented modifications (subtests) resulted
17 −1.24 23.53 38 0.80 36.66 in an improved solution for the PSS.
18 −1.11 24.32 39 0.90 37.25
19 −1.00 25.06 40 0.99 37.87
20 −0.89 25.77 41 1.09 38.52 Ordinal-to-interval conversion table
21 −0.78 26.45 42 1.20 39.21
Table 2 provides interval scores in both logit
22 −0.68 27.12 43 1.31 39.93
units and the original PSS-10 scale format that
23 −0.58 27.77 44 1.43 40.69
24 −0.48 28.43 45 1.56 41.52
allows researchers to convert ordinal raw scores
25 −0.38 29.06 46 1.70 42.45 to interval-level scores.
26 −0.28 29.70 47 1.87 43.51 Researchers who have already used the PSS-
27 −0.18 30.33 48 2.08 44.88 10 to collect data or are planning to use the
28 −0.09 30.94 49 2.40 46.93 scale can apply the results of this study as fol-
29 0.01 31.54 50 2.88 50.00 lows: calculate the raw score by reverse-scoring
30 0.10 32.13 questions 4, 5, 7 and 8 and then adding scores of
all 10 items together. Next, use Table 2 to con-
PSS: Perceived Stress Scale. vert these scores to the corresponding interval-
The raw score is calculated by reverse-scoring ques-
tions 4, 5, 7 and 8 and then adding scores of all 10 items scale scores ranging from 10 to 50, identical to
together. This table cannot be used for respondents with the range of scores of the original PSS scoring
missing data. system. Using the conversion table provided
here, users are able to increase the precision of
cent of significant t-tests and cut-off overlap of the PSS-10. It should be noted that conversion
2.71 per cent for sample (b) (Table 1). from the ordinal-to-interval scale proposed here
does not require altering the original response
format of the PSS-10 scale. This conversion
Item-person threshold distribution
table was independently generated with the sec-
Figure 1 shows the person-item threshold distri- ond sample (b) (n = 450), showing almost iden-
bution of the modified PSS-10 after combining tical results.
locally dependent items into three subtests
(Table 1, analysis 2a). Here, subtest item thresh-
olds and person ability levels on the latent fac-
Discussion
tor measured by the PSS-10 are plotted using The PSS-10 is a widely used instrument to
the same metric in logit units. Distribution of measure perceived stress on an ordinal scale;
persons is close normal, and the modified PSS- however, its precision has not been yet fully opti-
10 item thresholds satisfactorily cover 98 per mized. This study used the subtests methodology
1078 Journal of Health Psychology 24(8)

Figure 1. Distribution of persons and item thresholds set to interval length of 0.20 in logit units making
30 groups (final analysis, sample (a), n = 450).

analogous to Lundgren Nilsson et al. (2013) to PSS-10 is clearly tapping into one overarching
improve the psychometric properties and preci- latent factor. Local dependency found between
sion of the 10-item PSS up to an interval-level items 1, 2 and 9 is not surprising because these
scale. The best model fit was achieved after com- items all contain themes of control over exter-
bining locally dependent items into three sub- nal events. Another subtest included items 3, 6
tests, and no systematic DIF by personal factor and 10, which are explicitly related to perceived
such as gender, ethnicity, education and sample helplessness (e.g. could not overcome or cope)
population were evident. After these minor mod- and hence explain local dependency. These
ifications, the psychometric properties of the clusters of locally dependent items may have
PSS-10 are robust, and transformation from an influenced variability of the earlier factor anal-
ordinal to an interval-level scale can be con- ysis (Roberti et al., 2006; Taylor, 2015) and
ducted using the conversion algorithm provided may even generated spurious factors (Lundgren
in Table 2. This Rasch analysis contributed to the Nilsson et al., 2013). However, the clear evi-
limited number of IRT-based studies (Cole, dence of unidimensionality of the PSS-10 based
1999; Sharp et al., 2007; Taylor, 2015) that on three subtests replicated by two random
focused on the functioning of individual PSS-10 samples together with the large amount of
items by increasing the precision of the PSS-10 shared variance suggest that a total PSS-10
and addressing local dependency issues. interval score reflects perceived stress levels of
Local dependency found between items 4, 5, the majority of people (Figure 1).
7 and 8 is consistent with earlier research The main contribution of this study is that the
(Roberti et al., 2006; Taylor, 2015), where the PSS-10 raw score can now be converted from an
same items were proposed as a second factor in ordinal scale to an interval scale, which means
a two-factor PSS-10 solution. All these four that parametric statistics can be conducted with-
items are negatively worded and thus measure out violating their fundamental assumptions.
coping abilities as opposed to perceived stress, The interval-level estimates of the latent factor
which could explain previous findings of these offer researchers the opportunity to examine the
items loading together as a factor (Roberti et al., effects of mediators and moderators of per-
2006; Taylor, 2015). However, after combining ceived stress in various contexts. Rasch inter-
locally dependent items into subtests, unidi- val–transformed PSS-10 scores can reliably be
mensionality of the PSS-10 was clearly evident used in such models given that potential item
in both the main and the replication samples, biases (DIF) and local dependency issues are
suggesting that, after Rasch modifications, the ultimately resolved by Rasch analysis if data fit
Medvedev et al. 1079

the model expectations. The improved precision Funding


of the instrument is also highly desirable in clin- The author(s) disclosed receipt of the following
ical assessment, where it informs accuracy of financial support for the research, authorship, and/or
diagnosis and treatment of stress-related condi- publication of this article: This study is a part of a
tions. It is important to note that these improve- doctoral work of the first author funded by the Vice-
ments were possible without the need to alter the Chancellor's Scholarship of the Auckland University
original PSS-10 response format meaning that of Technology.
existing datasets can easily be re-analyzed to
provide interval level of measurement. References
The following limitations are acknowledged. Allen MJ and Yen WM (1979) Introduction to
The samples may not reflect the full diversity of Measurement Theory. Monterey, CA: Brooks/
New Zealand’s or the US’s ethnic groups and no Cole Publishing Company.
efforts were made to purposively sample under- Andrich D (1978) Rating formulation for ordered
represented groups. The response rate of 15 per response categories. Psychometrica 43: 561–573.
cent was relatively low for the New Zealand gen- Andrich D, Sheridan B and Luo G (2009) RUMM
eral population sample (Hill et al., 2014), although 2030. Perth, WA, Australia: RUMM Laboratory.
such response rates are not uncommon for Barbosa-Leiker C, Kostick M, Lei M, et al. (2013)
research of this nature in New Zealand (Krägeloh Measurement invariance of the perceived
stress scale and latent mean differences across
et al., 2013). Even though the Rasch-modified
gender and time. Stress and Health 29(3):
PSS-10 covers 98 per cent of the sample abilities, 253–260.
there are individuals uncovered by item thresh- Bloch GJ, Neeleman L and Aleamoni LM (2004) The
olds on both upper and lower level of the scale. salient stressor impact questionnaire (SSIQ): A
Future studies may consider to test whether using measurement of the intensity and chronicity of
more response options with extreme categories stress. Assessment 11(4): 342–360.
will provide better coverage for individuals with Chavez-Korell S and Torres L (2014) Perceived
lower and higher stress levels. stress and depressive symptoms among Latino
adults: The moderating role of ethnic identity
cluster patterns. The Counseling Psychologist
Conclusion 42(2): 230–254.
Stress may affect both physical and mental Christensen KB, Makransky G and Horton M (2017)
Critical values for Yen’s Q3: Identification of
health and its accurate assessment represents an
local dependence in the Rasch model using
ongoing challenge. This study has demonstrated
residual correlations. Applied Psychological
that after minor modifications, the widely used Measurement 41(3): 178–194.
perceived stress measure PSS-10 satisfies the Cohen S (2013) Frequently asked questions.
expectations of the unidimensional Rasch Available at: http://www.psy.cmu.edu/~scohen/
measurement model. The precision of the PSS- Cohen S and Williamson GM (1988) Perceived
10 can be optimized up to an interval-level stress in a probability sample of the United
scale using the ordinal-to-interval conversion States. In: Spacapan S and Oskamp S (eds) The
tables published here, which can be of benefit Social Psychology of Health: The Claremont
for researchers investigating neurophysiologi- Symposium on Applied Social Psychology.
cal, psychological and environmental correlates Newbury Park, CA: SAGE, pp. 31–67.
of stress. This study is best assessed from the Cohen S, Frank E, Doyle WJ, et al. (1998) Types of
stressors that increase susceptibility to the com-
perspective of Rasch analysis.
mon cold in healthy adults. Health Psychology
17(3): 214–223.
Declaration of Conflicting Interests Cole SR (1999) Assessment of differential item
The author(s) declared no potential conflicts of inter- functioning in the Perceived Stress Scale-10.
est with respect to the research, authorship, and/or Journal of Epidemiology & Community Health
publication of this article. 53(5): 319–320.
1080 Journal of Health Psychology 24(8)

Fisher W (1992) Reliability statistics. Rasch sample of patients undergoing treatment for
Measurement Transactions 6(3): 238. stress-related exhaustion: A Rasch analysis.
Gitchel WD, Roessler RT and Turner RC (2011) Health and Quality of Life Outcomes 11: 2.
Gender effect according to item directional- McEwen BS and Stellar E (1993) Stress and the indi-
ity on the perceived stress scale for adults with vidual: Mechanisms leading to disease. Archives
multiple sclerosis. Rehabilitation Counseling of Internal Medicine 153(18): 2093–2101.
Bulletin 55(1): 20–28. Marais I and Andrich D (2008) Effects of varying
Gustafsson JE (1980) Testing and obtaining fit of magnitude and patterns of response dependence
data to the Rasch model. British Journal of in the unidimensional Rasch model. Journal of
Mathematical and Statistical Psychology 33: Applied Measurement 9(2): 105–124.
205–233. Medvedev ON, Shepherd D and Hautus MJ (2015)
Helton WS and Näswall K (2015) Short stress state The restorative potential of soundscapes: A
questionnaire: Factor structure and state change physiological investigation. Applied Acoustics
assessment. European Journal of Psychological 96: 20–26.
Assessment 31(1): 20–30. Masters GA (1982) Rasch model for partial credit
Hill EM, Billington R and Krägeloh CU (2014) scoring. Psychometrica 47: 149–174.
Noise and diminished health: Testing modera- Mitchell AM, Crane PA and Kim Y (2008) Perceived
tors and mediators of the relationship. Noise & stress in survivors of suicide: Psychometric
Health 16(68): 47–56. properties of the perceived stress scale.
Hillhouse JE, Kiecolt-Glaser JK and Glaser R (1991) Research in Nursing & Health 31(6): 576–585.
Stress associated modulation of the immune Norquist JM, Fitzpatrick R, Dawson J, et al. (2004)
response in humans. In: Plotnikoff N, Murgo A, Comparing alternative Rasch-based methods
Faith R, et al. (eds) Stress and Immunity. Boca vs raw scores in measuring change in health.
Raton, FL: CRC Press, pp. 3–28. Medical Care 42(1): I25–I36.
Hobart J and Cano S (2009) Improving the evalu- Pomplun M and Custer M (2005) The construct
ation of therapeutic interventions in multiple validity of the Stanford-Binet 5 measures of
sclerosis: The role of new psychometric meth- working memory. Assessment 12(3): 338–346.
ods. Health Technology Assessment 13(12): Rasch G (1960) Probabilistic Models for Some
1–200. Intelligence and Attainment Test. Copenhagen:
Krägeloh CU, Kersten P, Billington DR, et al. (2013) Danish Institute for Educational Research.
Validation of the WHOQOLBREF quality Rasch G (1961) On general laws and the meaning
of life questionnaire for general use in New of measurement in psychology. In: Neyman J
Zealand: Confirmatory factor analysis and (ed.), Proceedings of the fourth Berkeley sym-
Rasch analysis. Quality of Life Research 22(6): posium on mathematical statistics and proba-
1451–1457. bility: Volume IV: Contributions to biology and
LeDoux JE (2000) Emotion circuits in the brain. problems of medicine. Berkeley, California:
Annual Review of Neuroscience 23: 155–184. University of California Press, pp. 321–333.
Linacre JM (1994) Sample size and item calibration Roberti JW, Harrington LN and Storch EA (2006)
stability. Rasch Measurement Transactions 7: Further psychometric support for the 10-item
328. version of the perceived stress scale. Journal of
Little TD, Cunningham WA, Shahar G, et al. (2002) College Counseling 9(2): 135–147.
To parcel or not to parcel: Exploring the ques- Rushton JP, Brainerd CJ and Pressley M (1983)
tion, weighing the merits. Structural Equation Behavioral development and construct valid-
Modeling 9: 151–173. ity: The principle of aggregation. Psychological
Lundgren Nilsson Å and Tennant A (2011) Past Bulletin 94: 18–38.
and present issues in Rasch analysis: The func- Sharp LK, Kimmel LG, Kee R, et al. (2007)
tional independence measure (FIM™) revis- Assessing the perceived stress scale for African
ited. Journal of Rehabilitation Medicine 43: American adults with asthma and low literacy.
884–891. Journal of Asthma 44(4): 311–316.
Lundgren-Nilsson Å, Jonsdottir IH, Ahlborg G, Siegert RJ, Tennant A and Turner-Stokes L (2010)
et al. (2013) Construct validity of the psycho- Rasch analysis of the Beck Depression
logical general well being index (PGWBI) in a Inventory-II in a neurological rehabilitation
Medvedev et al. 1081

sample. Disability and Rehabilitation 32(1): perceived stress and perceived health. Stress
8–17. and Health 31(1): 71–77.
Smith EV (2002) Detecting and evaluation the impact Tennant A and Conaghan PG (2007) The Rasch
of multidimensionality using item fit statistics measurement model in rheumatology: What is
and principal component analysis of residuals. it and why use it? When should it be applied,
Journal of Applied Measurement 3: 205–231. and what should one look for in a Rasch paper?
Stucki G, Daltroy L, Katz JN, et al. (1996) Arthritis and Rheumatism 57(8): 1358–1362.
Interpretation of change scores in ordinal clini- Wainer H and Kiely G (1987) Item clusters and
cal scales and health status measures: The whole computer adaptive testing: A case for testlets.
may not equal the sum of the parts. Journal of Journal of Educational Measurement 24: 185–
Clinical Epidemiology 49(7): 711–717. 202.
Taylor JM (2015) Psychometric analysis of the Wilson M (2005) Constructing Measures. Mahwah,
ten-item perceived stress scale. Psychological NJ: LEA Publishers.
Assessment 27(1): 90–101. Wright BD (1996) Local dependency, correlations
Teh HC, Archer JA, Chang W, et al. (2015) Mental and principal components. Rasch Measurement
well-being mediates the relationship between Transactions 10: 509–511.

You might also like