Response Conversion for Improving Comparability of International Physical Activity Data

Abstract

Many questionnaires for measuring physical activity (PA) exist. This complicates the comparison of outcomes.
In 8 European countries, PA was measured in random samples of 600 persons, using the IPAQ as a 'bridge' to historical sets of country-specific questions. We assume that a unidimensional scale of PA ability exists on which items and respondents can be placed, irrespective of country, culture, background factors, or measurement instrument. Response Conversion (RC) based on Item Response Theory (IRT) was used to estimate such a common PA scale, to compare PA levels between countries, and to create a conversion key. Comparisons were made with Eurobarometer (IPAQ) data.
Appropriateness of IRT was supported by the existence of a strong first dimension established by principal component analysis. The IRT analysis resulted in 1 common PA scale with a reasonable fit and face validity. However, evidence for cultural bias (Differential Item Functioning, DIF) was found in all IPAQ items. This result made actual comparison between countries difficult.
Response Conversion can improve comparability in the field of PA. RC needs common items that are culturally unbiased. Wide-scale use of RC awaits measures that are more culturally invariant (such as international accelerometer data).

Full text

29
Journal of Physical Activity and Health, 2012, 9, 29 -38
© 2012 Human Kinetics, Inc.
Hopman-Rock, Dusseldorp, Chorus, and Jacobusse are with
the Dept of Prevention and Healthcare, TNO Quality of Life,
Leiden, Netherlands. Ruetten is with the Dept of Sport Sci-
ence, Insitute of Sport Science and Sport, Friedrich-Alexander
University, Erlangen, Germany. van Buuren is with the Dept
of Methodology and Statistics, University of Utrecht, Utrecht,
Netherlands.
Response Conversion for Improving Comparability
of International Physical Activity Data
Marijke Hopman-Rock, Elise Dusseldorp, Astrid Chorus, Gert Jacobusse,
Alfred Ruetten, and Stef van Buuren
Background: Many questionnaires for measuring physical activity (PA) exist. This complicates the comparison
of outcomes. Methods: In 8 European countries, PA was measured in random samples of 600 persons, using
the IPAQ as a ‘bridge’ to historical sets of country-specic questions. We assume that a unidimensional scale
of PA ability exists on which items and respondents can be placed, irrespective of country, culture, background
factors, or measurement instrument. Response Conversion (RC) based on Item Response Theory (IRT) was
used to estimate such a common PA scale, to compare PA levels between countries, and to create a conver-
sion key. Comparisons were made with Eurobarometer (IPAQ) data. Results: Appropriateness of IRT was
supported by the existence of a strong rst dimension established by principal component analysis. The IRT
analysis resulted in 1 common PA scale with a reasonable t and face validity. However, evidence for cultural
bias (Differential Item Functioning, DIF) was found in all IPAQ items. This result made actual comparison
between countries difcult. Conclusions: Response Conversion can improve comparability in the eld of PA.
RC needs common items that are culturally unbiased. Wide-scale use of RC awaits measures that are more
culturally invariant (such as international accelerometer data).
Keywords: item response theory, cultural bias, differential item functioning, IPAQ
Researchers in the eld of physical activity (PA) and
(public) health gather data on PA by a wide variety of
measurement instruments. A few broad classes of such
instruments are: questionnaires, accelerometers, and eld
tests. Even within these broad classes, the variability of
instruments is great and likely to increase over time as
new instruments are being developed. Each new genera-
tion of instruments attempts to remedy the deciencies
of older ones, ideally converging into tools that are free
of the most obvious aws. On the other hand, the actual
situation is nowhere near this ideal. Different instru-
ments express PA in different units (frequency, duration,
intensity, energy expenditure), and there is no easy way
to convert one measure into the other. For example, both
the International Physical Activity Questionnaire (IPAQ)1
and the Short Questionnaire to Assess Health-Enhancing
Physical Activity (HEPA)2 aspire to measure PA in
community samples. Yet, one cannot simply convert or
compare their scores. There is also no accepted validation
paradigm (except research with the expensive doubly
labeled water method) by which we judge the quality
and comparability of the outcomes of an instrument. This
hampers progress in the area of international comparabil-
ity of PA levels of communities and individuals.
In this paper, we present a technique known as
Response Conversion (RC).3–5 This method is based
on rm theoretical principles and detects and repairs
comparability problems that arise out of differences in
the formulation of survey questions and response cat-
egories. RC attempts to translate responses obtained on
the same topic but with different questions into scores
on a common underlying unidimensional scale. Scores
on this scale are meant to be comparable, although they
were derived from different questionnaires measured in
different populations at different times.
RC is based on the assumption that instruments
measure the same continuum (eg, PA), but do so in differ-
ent ways. There is a clear analogy to physics, where the
distance between 2 points may be measured by a ruler,
by a difference between viewing angles, or by the time
taken to reect sound. As long as one knows how the
resulting values (cm, degrees, seconds) can be expressed
in terms of a common distance unit, it is possible to scale
the outcome on the same continuum. The RC technique is
a method to unveil such conversion rules using a linkage
diagram with ‘bridge’ items (ie, common items) as a start
to result in a routinely applicable conversion key. RC is
a test linking technique based on Item Response Theory.
(For a detailed description of test equating and test linking

30 Hopman-Rock et al
in the eld of physical activity, see Zhu6,7). RC has been
applied successfully for dressing and walking disability.8,9
RC is appropriate for linking questionnaire items (that are
assumed to measure the same continuum) from several
separate databases, where subjects completed at least 2
items, and where databases are linked by bridge items.
The current approach within the eld of PA research
is to express activity in MET-minutes10,11 and then linking
test data if possible. Use of METs-values requires extra
analyses and computations and some questions are not
suitable to transit in METs values. In addition, it is not
possible to correct for cultural bias in the outcomes after
the linking.
This paper explores the use of RC in the eld of
PA measurement. We will apply RC to an international
European data set, the EUPASS data,12 and compare the
results to the Eurobarometer study.13 RC has facilities to
correct for cultural bias (technically known as Differential
Item Functioning, DIF) This paper evaluates the potential
of RC for current items on PA, examining cultural bias,
and validates results in an example comparing PA in
different age groups.
Methods
Study Design and Measures
The EUPASS study gathered data in the year 2000 using
the IPAQ1 in 8 different countries (Belgium, Finland,
France, Germany, The Netherlands, United Kingdom,
Italy, and Spain). In addition, the EUPASS study included
existing country-specic PA questions. Within each of the
8 countries, about 600 adult respondents were randomly
selected (total database N = 4976), and interviewed using
a computer-aided telephone interview. Precise details can
be found elsewhere.12
The IPAQ questions were asked in all countries.
The IPAQ was explicitly designed to be cross-culturally
equivalent.1 In total, 9 IPAQ items and 40 national items
were available for analysis. Conforming to IPAQ instruc-
tions,10 we computed continuous compound variables for
vigorous activity, moderate activity, and walking, using
the items “hours a day,” “minutes a day,” and “days per
week.” We removed subjects with unlikely answers (ie,
subjects who report more than 3 hours of vigorous activ-
ity, moderate activity, or walking a day). This resulted
in a substantial reduction of the sample size to N = 3597
[lost for analyses is 27.7%; this value varied between
countries from 11.7% (Italy) to 38.5% (Belgium)].
Subjects that remained in the sample had slightly worse
health (16%, against dropout group 22%; Chisquare 39.5,
df = 4, P = .00), were more female (58% against 52%
in dropout group; Chisquare 15.2, df = 1, P = .00) and
were slightly older (mean age 46 years against 44 years in
dropout group; F = 14.5, P = .00). Continuous compound
variables were expressed in total minutes per week. Two
variables measuring sitting behavior either on a weekday
or on a weekend day were merged together according to
the IPAQ manual. Table 1 contains an overview of the 49
items in the form of a diagram that shows which items
were administered in which countries and how the link-
age was established.
Table 1 Linkage Diagram of the EUPASS Data; Overview of All Items and for Which Countries Each Item
Is Applicable
Country
Questionnaire items
RoaRua
BE FI GE IT NL UK SP FR
1. IPAQ At what pace usually walk* 3 3 X X X X X X X X
2. IPAQ How much PA in place of work last 7 days 3 3 X X X X X X X X
3. IPAQ How much PA for purpose of transportation last 7 days 3 3 X X X X X X X X
4. IPAQ How much PA in and around home last 7 days* 3 3 X X X X X X X X
5. IPAQ How much PA recreation, sport, leisure time 3 3 X X X X X X X X
6. IPAQ how much time in usual week doing vigorous PA C 7 X X X X X X X X
7. IPAQ how much time in usual week doing moderate PA C 7 X X X X X X X X
8. IPAQ how much time in total you spend on walking C 7 X X X X X X X X
9. IPAQ sitting: sum in minutes for 1 day REVERSEDb* C 5 XXXXXXXX
10. On how many days sweating at least 1 time per week 8 7 X
11. Leisure time PA for at least half an hour (at least l sw) 7 7 X
12. Minutes a day walking, running or riding a bicycle to/of work 6 5 X
13. Demanding job physically 4 4 X
14. How much exercise or PA in free time 4 4 X
(continued)

31
Country
Questionnaire items
RoaRua
BE FI GE IT NL UK SP FR
15. How often engaged in sports/ strenuous activities 5 5 X
16. Get out of breath after climbing 3 oors REVERSEDb2 2 X
17. How often do you participate in sports 5 5 X
18. Time spend per day sleeping (Monday to Friday) REVERSEDbC 5 X
19. Time spend per day sitting (M-F) REVERSEDbC 7 X
20. Time spend per day light activities* (M-F) C 8 X
21. Time spend per day moderate activities (M-F) C 8 X
22. Time spend per day strenuous activities (M-F) C 8 X
23. Time spend per day sleeping (Weekend) REVERSEDbC 5 X
24. Time spend per day sitting (Weekend) REVERSEDbC 7 X
25. Time spend per day light activities* (Weekend) C 8 X
26. Time spend per day moderate activities (Weekend) C 8 X
27. Time spend per day strenuous activities (Weekend) C 8 X
28. How long are you engaged in sports / strenuous act. 4 4 X
29. Regular sporting activities in free time 2 2 X
30. Occasional sporting activities in free time 2 2 X
31. How many month in total C 4 X
32. Consider all the sporting activities over past 12 months 6 6 X
33. Any type of physical activity at least twice a year 4 4 X
34. How many activities 5 5 X
35. Sporting activities requiring payment 2 2 X
36. Practice requiring payment (lessons) 2 2 X
37. Annual (periodic) fee for sport club 2 2 X
38. Number of times PA participation past 14 days c 5 X
39. How many times a day do you walk c 5 X
40. How many times sports or exercise c 5 X
41. Sum of minutes heavy PA yesterday c 7 X
42. Sum of minutes moderate PA yesterday c 7 X
43. Sum of minutes light PA yesterday* c 7 X
44. How many sports 5 3 X
45. Did you sport yesterday 2 2 X
46. Gardening, dig or building work done in the past 4 weeks* 2 2 X
47. Any exercise or sport during the last 4 weeks 2 2 X
48. Was the effort or activity usually makes you out of breath 2 2 X
49. Walking of a quarter of a mile done locally or away from 2 2 X
a Ro: Response scale of original items, the number of response categories is given, and ‘c’ is used for continuous items; Ru: number of response categories used
in the Rasch model; BE, Belgium; FI, Finland; GE, Germany, IT, Italy; NL, The Netherlands; UK, United Kingdom; SP, Spain; FR, France.
b Items are coded such that a lower score reects less physical activity (for example, more sitting).
* Removed from Rasch analysis.
Note. For overview of all questionnaires, see: http://www.public-health.tu-dresden.de/dotnetnuke3/Portals/5/Projects/EUPASS/appendix%20b.pdf.
Table 1 (continued)

32 Hopman-Rock et al
All PA variables were coded such that a higher value
reected a higher activity level. For example, the coding
of sitting items was reversed so that much sitting indicated
low PA. Continuous PA variables were categorized for
application of the Rasch model. For example, the IPAQ
items expressed in total minutes of activity per week
were categorized into average number of half hours per
day (ie, divided by 210 and then rounded). The number
of categories of some categorical variables was reduced,
because of low frequencies in the extreme values. For
example, categories 6 and 7 of the item “on how many
days sweating at least 1 time per week” (item 10, Table
1) were merged to 1 category: 5 days or more. Thus, the
number of categories for categorical measured variables
varied between 5 and 8. The correlation between all cat-
egorized and corresponding original variables was always
higher or equal to 0.90. Table 1 gives for each item the
number of original response categories and the number
of categories used in the IRT analysis.
On the basis of the IPAQ activity measures expressed
in MET-minutes, the following 3 categories were com-
puted by using the IPAQ manual: HEPA 1, 2, and 3. These
categories reect the percent of people at low (HEPA 1:
sedentary/inactive), moderate (HEPA 2: not sufcient),
and high levels (HEPA 3: sufcient) of PA (to accumu-
late to 100% of the population). The high level of PA is
similar to the “sufcient total activity” level used in the
Eurobarometer study.13 This categorical representation
of PA enables us to compare the results of the EUPASS
data with the Eurobarometer data (please note that the
EUPASS and the Eurobarometer study are separate
studies, only similar in using the IPAQ in an European
sample). In addition, we will compare it with the results
from the RC analyses.
Statistical Analysis
We used the polytomous Rasch (IRT) model14,15 to
estimate the relative position (often interpreted as ‘dif-
culty’) of the items on the PA ability scale. The Rasch
model describes the probability that a person responds
into a category conditional on the location of the person
on the continuum of PA (which can be interpreted as a
scale that indicates how physically active a person is).
The model has 1 or more difculty parameters for each
item (there are m difculty parameters for an item with
m+1 categories). For a dichotomous item (with categories
‘no’ and ‘yes’), the difculty parameter indicates how
much ability a respondent needs to achieve a 50/50%
chance of scoring ‘yes.’ An afrmative answer to a
difcult question like ‘do you go running for at least 3
hours a week’ is generally associated with higher PA
level than an afrmative response to an easier item like
‘do you walk at least 1000 m a week.’ For a polytomous
item, the difculty parameters can be considered as step
difculties associated with the transition from one cat-
egory to the next.14 A positive response, especially to a
‘difcult’ question, results in a higher respondent score
on the PA scale. A negative response, especially to an
easy question, results in a lower PA score. A better or
more precise estimate of the ability of a person can be
calculated from his or her responses to a series of items.
We opted for the Rasch model since that model is the
only one in which estimation of the difculty parameters
is independent of the distribution of PA in the reference
population. Thus, the choice of the reference sample is
not critical to parameter estimates.
An important assumption underlying the Rasch
model is that items measure the same continuum (ie, that
they are unidimensional). We checked this assumption
by categorical principal components analysis using SPSS
CATPCA16 on the 9 IPAQ items, where we assumed
ordered categories.
We used RUMM 202017 to estimate item difculty
parameters. The estimation method is based on the
pairwise conditional approach, and has been described
in detail by Andrich and Luo.15 This approach generally
works well with incomplete and sparse data.15 Using the
Bayes rule, the parameters can be used to calculate the RC
key. The RC key is a simple table in which it is possible
to transform the original category of a questionnaire into
a place on the new PA scale. For additional information
and examples of this procedure see Jacobusse et al.18
The reliability of the Rasch model was measured
by the person separation index (also called the person
separation reliability.19 It range between 0 and 1, and the
interpretation is similar to that of Cronbach’s α. Further-
more, the item t was measured by the t residual statistic
per item (given by RUMM). When the Rasch model is
true, this measure follows a standard normal distribution
(mean near 0, and standard deviation near 1); values
higher than 2 indicate that unexpected deviations from
the model occur.20 Because of the large sample size of
our study, we chose a more liberal criterion for mist of
items, that is, a standard residual greater than 3.5, and we
excluded mistting items from the analysis. Similarly as
for items, respondents with a person-t residual statistic
greater than 3.5 were excluded from the nal analysis.
The IPAQ items acted as bridge items between the country
samples (see Table 1). The assumption is that 9 bridge
items measure PA in the same way in different countries
(without cultural bias). If this assumption is false, the
item has Differential Item Functioning (DIF),21 and we
cannot use it in its original form to equate items. We
tested for DIF by ANOVA using a Type I error rate of P
< .001 because of large sample size. If DIF was present,
we retted the data under the more relaxed model where
deviating countries obtained separate difculty parameter
estimates, an action known as item splitting. DIF was
tested again in the remaining items, until an acceptable
solution was found without DIF. This procedure is known
as the Stocking and Lord iterative procedure.22
Results
Table 2 summarizes the main characteristics of the
selected sample (N = 3597) from the EUPASS data. The
Netherlands and Belgium had older people than the other
countries in the sample. Note that the responses to the

Common PA Scale Development 33
general health question also varied considerably across
countries.
The rst dimension based on categorical princi-
pal components analysis on the recoded IPAQ items
explained 21.5% of the total variance. For comparison,
a linear PCA with the same variables explained 20.7% of
the variance. The percentage was 22.0% for the unrecoded
continuous data, indicating that the loss by categorization
is negligible. The rst eigenvalue of the PCA solution
was large compared with the second (1.9 vs. 1.1), and
the eigenvalues other than the rst were about the same
size (between 1.1 and 0.7). These results pointed to a
dominant rst factor underlying the PA items.6
The int statistics of the estimated Rasch model
with all 49 items ranged from –6.6 to 5.9. After 4 item
removal steps in which questions with a t statistic over
3.5 were dropped, all items had a t statistic lower than
3.5. The highest remaining int statistic was 2.5. The
following items were removed: 3 IPAQ items (items 1, 4,
and 9; Table 1), 3 items referring to light activities (item
20, 25, and item 43), and 1 item measuring gardening,
do-it-yourself, or building work (item 46). Most of these
mistting items had no or limited conceptual overlap
with the other items. After removing these items, none
of the persons had a person-t statistic greater than 3.5.
The overall goodness-of-t index (the person-separation-
index) of the Rasch model was reasonable: 0.68.17 Figure
1 shows the distribution of the estimated common PA
scale based on this model. The scale was transformed
in such a way, that the mean score was set at 50 and a
standard deviation of 10. Parameter estimates indicated
that the easiest item (the item that most people were most
likely to respond “yes” to) was “walking a quarter of a
mile or more in the past 4 weeks, either locally or away
from home” (no/yes; item 49, Table 1). The most dif-
cult transition (this means only positively answered by
respondents with relatively high ability levels) occurred
for the item “sum of minutes heavy physical activity
yesterday” (item 41): the transition from 0 to 10 minutes
(category 1) to 10 to 40 minutes (category 2).
We investigated DIF of the remaining 6 IPAQ
items. All 6 items had statistically signicant DIF (P <
.001), indicating that their interpretation differed across
countries. Figure 2 shows how DIF between countries
manifests itself in IPAQ question ‘How much time in
total you spend on walking” (item 8, Table 1). This item
has 7 response categories (0 to 6), representing the aver-
age number of half hours walking per day. If there was
no DIF, the response curves would be located closely
to each other. For most countries, this is the case. How-
ever, it appears that persons from the Netherlands score
consistently higher at the same level of PA (especially
at the lower end of the scale). In other words, item 8 is
more “easy” for the Dutch than for the other countries.
The consequence is that we cannot use the responses
on item 8 in the Netherlands in the same way as for the
other countries. To correct for this, we estimated separate
item parameters for the Netherlands. This item splitting
procedure was performed for all items, until there was
no signicant DIF left.
To get more insight into the consequences of the item
splitting procedure, we compared the results of the Rasch
model without correcting for DIF (Figure 3, panel a) with
those of the model while correcting for DIF (Figure 3,
panel b). The black lines in the boxes in Figure 3 repre-
sent the country medians. Italy and UK have a median
score below or equal to the overall median of 52.5 in both
models. And additionally, The Netherlands and Belgium
have a mean score below the overall mean of 50.0 in
both models. So, the correction of DIF seems to have
limited inuence. The standardized difference in means
is large (effect size d = 0.7–0.8) between Germany and
United Kingdom (Figure 3, panel a), and between Spain
and the United Kingdom (Figure 3, panels a and b). The
Table 2 Main Characteristics of the Respondents by Country (N = 3597)
Country
Belgium Finland France Germany Italy Netherl. Spain UK Total
n respondents 376 379 390 384 530 467 500 571 3597
Mean age (SD) 51.5
(18.1)
46.4
(15.3)
40.3
(16.7)
42.8
(15.2)
44.3
(15.9)
51.3
(18.8)
46.1
(18.4)
43.8
(17.0)
45.8
(17.4)
% Women 55 61 57 56 56 63 58 57 58
Mean BMI 24.5 25.0 23.0 24.3 24.2 24.6 24.4 24.3 24.3
Health (%)
Very good/good – 63 66 75 55 80 70 62 67
Fair/poor/bad – 36 34 26 45 20 30 39 33
Occupation (%)
Working 45 58 52 63 53 42 50 60 52
Retired 36 27 16 16 19 29 18 20 22
Other 19 15 16 22 29 29 33 20 27
Note. In Belgium, the general health question was not administered.
Abbreviations: Netherl., The Netherlands; UK, United Kingdom.

34
Figure 1 — Distribution of the individual ability scores on the common PA scale obtained by the Rasch analysis (N = 3579). Mean
PA score is normalized as 50, and standard deviation is 10.
Figure 2 — Differential Item Functioning (DIF) indicating cultural differences in responding by country in IPAQ item: “How much
time in total you spend on walking” (item 8, Table 1).

Common PA Scale Development 35
other differences are moderate (d = 0.5–0.6) to negligible
(d = 0.1). Most differences between country means are
statistically signicant, but not all. For example, the
differences between the means of Finland and Germany
and between those of the Netherlands and Italy are not
signicant for both models. The standard deviation of
the scores (on both PA scales) was higher than 10 in the
United Kingdom, Belgium, and the Netherlands.
Some countries were more responsible for DIF than
others. The Netherlands showed DIF on 5 of the 6 IPAQ
items, Italy on 4 items, and Spain, Finland, and Germany
on 3 items each. The differences between the results of
the model without DIF and with DIF were highest for
Spain and the Netherlands. The mean PA score of Spain
was lower in the model without DIF (Figure 3, panel a)
than with DIF (Figure 3, panel b), and for the Netherlands
the reverse was true.
We also compared the rank order of the countries to
the prevalence of the HEPA categories estimated from the
Eurobarometer study (collected in October to December
2002; see Table 3). Both Germany and Finland were in
the top 3 of physically active countries, according to the
continuous PA scale as well as according to the “sufcient
activity” category (HEPA 3). This applies for both the
Eurobarometer study and this study (Table 3). However,
according to the common PA scale, Spain belongs also
to this top 3 (Figure 3), whereas according to the HEPA
3, the Netherlands is one of the most physically active
countries (Table 3). Spearman rank-order correlation
between HEPA 3 of the EUPASS and HEPA 3 of the
Eurobarometer was 0.55 (P < .001), indicating a moder-
ate level of agreement.
With regard to the 3 least active countries, both repre-
sentations of PA and both studies agree that Belgium and
the United Kingdom fall in this category (Figure 3, panel
a and Table 3, HEPA 1 columns). However, according to
the common PA scale and HEPA 1 of this study, Italy is
also one of the least active countries (Figure 3 and Table
3), whereas according to HEPA 1 of the Eurobarometer
study, France is the least active country. Spearman rank-
order correlation between HEPA 1 of the EUPASS and
HEPA 1 of the Eurobarometer was 0.52, indicating a
moderate level of agreement.
In general, we expect a decline in PA with age.
With higher age, the mean PA score remains the same
for France (Figure 4, panel a), shows a small decline for
Spain, Finland, and Germany, and shows a large decline
in the elderly (from 64 yrs old) for the Netherlands,
Figure 3 — Distribution of the common PA scale estimated by the Rasch model without correcting for DIF (panel a) and by the model with
correcting for DIF (panel b). Both PA scales are normalized with mean 50, and standard deviation 10. A higher score means a higher level
of physical activity.

36
Table 3 Prevalence of 2 HEPA Categories of Physical Activity in the EUPASS Study and the
Eurobarometer Study (Derived From Sjöström, 2006); Those Countries From the Eurobarometer
Study Were Selected That Were Also Measured in This Study; Prevalences With Highest Rank Order
(1) Are in Bold Face
HEPA 1
(sedentary/inactive %) (rank) HEPA 3
(sufficient PA %) (rank)
EUPASS Eurobarometer EUPASS Eurobarometer
Belgium 30.1 (3) 39.8 (2) 31.9 (6) 25.0 (7)
Finland 18.2 (6) 23.8 (7) 45.1 (2) 32.5 (3)
France 25.6 (4) 43.1 (1) 37.9 (5) 24.1 (8)
Germany 15.1 (7) 24.1 (6) 52.6 (1) 40.2 (2)
Italy 38.9 (1) 35.3 (4) 18.1 (8) 25.8 (5)
Netherlands 22.9 (5) 19.3 (8) 44.3 (3) 44.2 (1)
Spain 10.8 (8) 31.2 (5) 39.0 (4) 25.2 (6)
United Kingdom 35.4 (2) 37.4 (3) 25.0 (7) 28.7 (4)
Note. EUPASS data uncorrected for DIF. Spearman correlation HEPA 3 for EUPASS and Eurobarometer = 0.55 (P = .00).
Figure 4 — Relationship between age and the PA scale (panel a) and between age and sufcient total activity (HEPA 3; panel b). Age is
divided into the following categories: 1 = 15–29 yrs; 2 = 30–38 yrs; 3 = 39–49 yrs; 4 = 50–63 yrs; 5 = 64–93 yrs. Each category contains
approximately 20% of the respondents.

Common PA Scale Development 37
Belgium, Italy, and the United Kingdom. Note that these
trends are not present in the graph showing the relation-
ship between age and HEPA 3 (Figure 4, panel b). These
results suggest that the common PA scale could be a more
sensitive measure of PA.
Discussion
This article introduced Response Conversion to improve
international comparability of PA data. The PA data
used in this study (the EUPASS data; 12) included both
items from a relatively new instrument (IPAQ) as well
as ‘old’ existing items from locally used questionnaires
in 8 European countries.
Our results demonstrated that 1) the PA items
included in the EUPASS study satised the assumption
of unidimensionality, 2) a ‘Physical Activity Scale’ could
be estimated by a Rasch model with reasonable t, and
3) the Physical Activity Scale was more sensitive to the
age-effect than the HEPA by showing a clear pattern of
decreasing PA with age (thus showed face validity).
All IPAQ bridge items appeared to suffer from DIF.
This means that items are not fully comparable between
countries. For a similar conclusion see Bauman et al,23
who included 20 worldwide countries using the IPAQ.
Thus, the IPAQ might be less suitable for comparing
populations from different countries because of cultural
bias. This is disappointing because the IPAQ items were
designed to be free of cultural bias. This suggests that it
is difcult to develop items that are free of DIF. Another
problem with the IPAQ was the relatively high loss of
subjects that apparently had difculty in understanding
the questions. For example, 3.3% of the subjects indi-
cated more than 7 hours of vigorous activity a day (up to
20 hours). We expect that respondents have interpreted
this as activity per week. The difculty of the questions
was also observed by Heesch et al,24 who performed a
qualitative study in older people that completed the IPAQ.
These authors emphasized that most items are very dif-
cult to answer.
The ability to recognize and quantify DIF is a major
methodological advance of RC. As the validity of the RC
key relies on linked databases which are free from DIF,
we have decided to withhold its publication until better
quality data become available. One possibility is the
GPAQ developed by the World Health Organization,25
which will eventually be combined with international
accelerometer data.
We refer to the common underlying scale resulting
from RC as the “Physical Activity Scale.” This scale is
conceptually different from the energy expenditure scale,
expressed in METs, that is often used to summarize
IPAQ or other PA items. Physical activity is a somewhat
more general concept than energy expenditure, relating
to actual behaviors rather than the amount of energy
required to perform these behaviors. An advantage of
such a broader concept is that more PA indicators can be
related to it (such as sitting behavior). This enables us to
encompass a wider range of items, eventually resulting
in increased measurement precision.
To check the validity of the data we made compari-
sons with data from the Eurobarometer study.13 Country-
specic percentages of activity categories as used by
the Eurobarometer were in range with our ndings. An
unexpected nding was that compared with the other
countries, the data from the Netherlands showed a low
level of PA, especially in older people. In the Netherlands
the year 2000 was—according to the national trend
report—the year with the lowest level of PA in the older
population (as measurement started in 2000). Since that
time the national PA levels have been improved signi-
cantly for all age groups.26
We conclude that Response Conversion is a promis-
ing technique to improve comparability in the eld of PA
applying to both existing as well as new databases. The
technique is able to integrate various components of PA
into a common ‘Physical Activity Scale.’ This PA scale
is a valuable addition to the concept of METs. However,
wider application of the new technique requires better
quality data with less cultural bias. We expect that new
and improved measures will be developed in the future,
thus making the benets of RC available to the eld of PA.
Acknowledgments
This study was supported by a grant (SI2.131854 /99CVF3-
510) from the European Commission DG SANCO Health
Monitoring Programme.
References
1. Craig CL, Marshall AL, Sjöström M, et al. International
Physical Activity Questionnaire (IPAQ): 12-country reli-
ability and validity. Med Sci Sports Exerc. 2003;35:1381–
1395.
2. Wendel-Vos GC, Schuit AJ, Saris WH, Kromhout D.
Reproducibility and relative validity of the short question-
naire to assess health-enhancing physical activity. J Clin
Epidemiol. 2003;56(12):1163–1169.
3. Hopman-Rock M, Van Buuren S, de Kleijn-de Vrankrijker
MW. Polytomous Rasch analysis as a tool in the revision
of the severity of disability scale of the ICIDH. Disabil
Rehabil. 2000;22:363–371.
4. van Buuren S, Eyres S, Tennant A, Hopman-Rock M.
Response conversion: a new technology for compar-
ing existing health information. TNO Report PG/
VGZ/2001.097. Leiden: TNO Prevention and Health.
ISBN 90-6743-813-8, 2001.
5. van Buuren S, Hopman-Rock M. Revision of the ICIDH
Severity of Disabilities Scale by data linking and item
response theory. Stat Med. 2001;20:1061–1076.
6. Zhu W. An empirical investigation of Rasch equating of
motor function tasks. Adap Phys Act Qu. 2001;18:72–89.
7. Zhu W. Scaling, equating, and linking to make measures
interpretable. In: Wood TM, Zhu W, eds. Measurement
theory and practice in kinesiology. Champaign, IL: Human
Kinetics; 2006:93–112.

38 Hopman-Rock et al
8. van Buuren S, Eyres S, Tennant A, Hopman-Rock M.
Assessing comparability of dressing disability in different
countries by response conversion. Eur J Public Health.
2003;13(3, Suppl 1):15–19.
9. van Buuren S, Eyres S, Tennant A, Hopman-Rock M.
Improving comparability of existing data by response
conversion. J Off Stat. 2005;21(1):53–72.
10. IPAQ manual, 2005. http://www.ipaq.ki.se/scoring.htm
11. Ainsworth BE, Haskell WL, Whitt MC, et al. Compen-
dium of physical activities: an update of activity codes
and MET intensities. Med Sci Sports Exerc. 2000;32(9,
Suppl):S498–S504.
12. Rütten A, Ziemainz H, Schena F, et al. Using different
physical activity measurements in eight european coun-
tries: results of the European Physical Activity Surveil-
lance System (EUPASS) Time Series Survey. Public
Health Nutr. 2003;6: 371–376. See also: http://www.
public-health.tu-dresden.de/dotnetnuke3/eu/Projects/
PastProjects/EUPASS/tabid/337/Default.aspx
13. Sjöström M, Oja P, Hagströmer M, Smith BJ, Bauman AE.
Health-enhancing physical activity across European Union
countries: the Eurobarometer study. J Public Health. 2006,
4:291–300. See also: http://www.euro.who.int/document/
e89490.pdf
14. Embretson SE, Reise SP. Item response theory for psy-
chologists. Mahwah, NJ: Lawrence Erlbaum; 2000.
15. Andrich D, Luo G. Conditional pairwise estimation in
the Rasch model for ordered response categories using
principal components. J Appl Meas. 2003;4:205–221.
16. Meulman JJ, Heiser WJ. SPSS. SPSS Categories 10.0.
Chicago: SPSS Inc.; 1999.
17. RUMM LABORATORIES. Rumm 2020. Rasch Unidi-
mensional Measurement Models, 2003. www.rummlab.
com.au
18.
Jacobusse G, van Buuren S, Chorus AMJ, Hopman-Rock
M. Physical activity. In: Buuren S van, Tennant A (ed).
Response conversion for the Health Monitoring Program.
Leiden: TNO Prevention and Health, Publ. nr. 04.145. ISBN
90-5986-082-9, 2004 (available at www.stefvanbuuren.nl).
19. Wright BD, Masters GN. Rating scale analysis. Chicago:
MESA Press; 1982.
20. Smith RM. Application of Rasch measurement. Chicago:
MESA Press; 1992.
21. Holland PW, Wainer H, eds. Differential item functioning.
New York: Lawrence Erlbaum; 1993.
22.
Stocking ML, Lord FM. Developing a common metric in item
response theory. Appl Psychol Meas. 1983;7(2):201–210.
23. Bauman A, Bull F, Chey T, et al. The International Preva-
lence Study on Physical Activity: results from 20 countries.
Int J Behav Nutr Phys Act. 2009;6(1):21.
24. Heesch KC, van Uffelen JG, Hill RL, Brown WJ . What
do IPAQ questions mean to older adults? Lessons from
cognitive interviews. Int J Behav Nutr Phys Act. 2010,
11;7:35.
25. Bull FC, Maslin T. Final report on reliability and validity
of the Global Physical Activity Questionnaire (GPAQ v1).
Geneva: World Health Organization; 2006.
26. Hildebrandt VH, Ooijendijk WTM, Hopman-Rock M.
Trendrapport bewegen en gezondheid 2006/2007. Leiden,
The Netherlands: TNO; 2008. [Trend report Physical
Activity and Health].