Systematic review on measurement properties of questionnaires assessing the neighbourhood environment in the context of youth physical activity behaviour

In this review, we considered all neighbourhood environmental attributes potentially relating to physical activity behaviours in children and adolescents including aspects such as accessibility, safety, convenience, attractiveness and distances to physical activity and recreational facilities or areas, aspects of urban design and traffic features [12, 15].

We conducted a systematic computerized literature search to identify all relevant articles on measurement properties of self- and proxy-report questionnaires concerning the perceived physical environment as a predictor of physical activity behaviour in children and adolescents. On July 4th 2012, we searched the following relevant electronic databases for English-language peer-reviewed journal articles: in topics in Web of Science (limits: English language, articles) and Medline (limits: Journal articles), in keywords in TRID (Transportation Research International Documentation; limits: English language) and in abstracts in SportDISCUS, PsycARTICLES and PsycINFO. The search term consisted of five types of related terms:

At least one term of every term type had to be met. In accordance with recommendations for systematic reviews on measurement properties [22], we screened reference lists and citations of included articles to identify additional relevant studies.

We used the following eligibility criteria: (i) The study evaluated an instrument on any aspects of the neighbourhood environment related to physical activity behaviour also including active transport, walking or cycling; (ii) the instrument was a self- or proxy-report questionnaire for assessment of the neighbourhood environment on the level of individuals; (iii) the main aim of the study was to evaluate at least one measurement property of the questionnaire, and information on measurement properties were collected intentionally; and (iv) the age range of subjects was 3 to 18 years or their mean age was within this range.

Two independent reviewers (AKR, FM) conducted the stepwise literature search. First, all articles were screened based on titles. In a second step, abstracts of potentially relevant articles were reviewed. If the abstract indicated that the study fulfilled the eligibility criteria or the abstract did not provide sufficient information for selection decision, both reviewers assessed the full texts of articles for eligibility. If necessary, supplementary files were also reviewed for additional information. Discrepancies between article selections were resolved after discussion at the end of the selection process. Additionally, both reviewers screened all reference lists and citations of included articles listed in Web of Science using the same procedure. In one case, the corresponding author of the reviewed article was contacted to request additional information on general characteristics of the instrument that were necessary for selection decision.

We extracted the target data from the full text articles and from webpages specified therein, from additional files and, if necessary, from related publications. The extracted data included: general characteristics of the instruments, characteristics of the studies and study populations, methodological quality and the results of the studies on measurement properties.

Two reviewers (AKR, FM) conducted independent methodological quality assessments to determine the evidential basis of the included studies. We adapted the COSMIN (COnsensus-based Standards for the selection of health Measurement INstruments) [23] checklist–a standardized tool for evaluating the methodological quality of studies on measurement properties of health-related patient-reported outcome measures. We used the taxonomy, terminology, and definitions of measurement properties for health-related patient-reported outcomes defined in the Delphi study within the COSMIN study [24]. The COSMIN checklists considered internal consistency, reliability, measurement error, content validity, structural validity, hypotheses testing, cross-cultural validity, criterion validity and responsiveness. We rated the quality of each measurement property of a study according to the items of the corresponding adapted COSMIN checklist. Only measurement properties of the COSMIN checklists were rated. The four response options in the COSMIN scoring system with the ‘worst score counts’ algorithm [25] did not reveal any differences in the methodological quality between studies. Therefore, we used dichotomous response options [26] and calculated the percentage of items with positive ratings across each checklist to represent the number of quality criteria fulfilled by each study. Because of different numbers of items for each checkbox, missing an item had a different impact on methodological quality scores (MQS) across measurement properties.

Hypotheses testing was defined as the degree to which the scores of an instrument are consistent with the hypotheses based on the assumption that the instrument validly measures the construct to be measured and is subdivided into convergent, discriminant and discriminative validity [24]. Because the included studies only addressed convergent validity, the more common term ‘convergent validity’ will be used in this paper [27].

In accordance with other authors [28, 29], the COSMIN manual suggests that internal consistency and structural validity statistic can only be interpreted if the construct is based on a reflective model [26]. In reflective models, the items are manifested by the construct, and the direction of causality is from the construct to the items [30]. Thus, variation in the construct causes variation in the indicators, the items are interchangeable and the items presumably have a high inter-correlation [28, 31]. In comparison, in formative constructs the latent variable is formed and defined by the items and the causality is from the items to the construct [30]. Furthermore, the variation in the construct does not necessarily cause variation in the indicator, the items are not interchangeable and the items are not necessarily correlated with each other [28, 31]. Statistical procedures assuming inter-item correlation, such as factor analysis or internal consistency, are not appropriate for expressing measurement properties of questionnaires assessing formative constructs [29, 32]. We assume that the neighbourhood environment is a formative construct because attributes of the neighbourhood environment (e.g. accessibility of parks or pavements of streets) form the construct of the neighbourhood environment. Changes in the neighbourhood environment do not necessarily cause changes in all indicators and the items may not be inter-correlated. For instance, changes in accessibility to parks causes changes in the neighbourhood environment but is not necessarily associated with changes in pavements of streets. Furthermore, pavements of streets and accessibility of parks may not be inter-correlated. According to these arguments, estimates of internal consistency and structural validity have not been considered for evaluating questionnaires, and the methodological quality of the studies on internal consistency and structural validity were not rated in this review.

Characteristics of the 13 studies included in this review are presented in Table 1. Seven studies were conducted in the United States [36‐42], two in Europe [35, 43], two in Australia [44, 45], one in Hong Kong [46] and one in Iran [47]. Three studies targeted only female participants [37, 40, 47]. Eight studies evaluated a self-report measure [37, 39, 40, 42‐44, 46, 47], two a parent-report measure [35, 45] and three included both administration forms [36, 38, 41]. The study samples were recruited mainly from schools (n = 8) [39, 40, 42‐47] or communities (n = 4) [35, 36, 38, 41]. None of the studies explicitly stated the response rates for each measurement property.

General characteristics of the instruments included in our systematic review are presented in Table 2. The target populations ranged from preschool children to adolescents. The number of items ranged from 4 [42, 47] to 66 [36]. Only some questionnaires focused on specific physical activity behaviours, one instrument examined environmental and safety barriers in relation to physical activity in parks and streets [38], one focused on active transport to school [40], one on walking and cycling to three different neighbourhood destinations [41], and one on walking in the neighbourhood [36]. Two studies evaluated the same environment scale originally used in the Amherst Health and Activity Study in English [42] and translated into Persian [47]. Most questionnaires employed Likert scales (Table 2).

Nine articles examined the questionnaires’ internal consistency [35‐37, 41‐44, 46, 47], ten their test-retest reliability [36, 38‐42, 44‐47] and three their inter-rater reliability [36, 37, 41]. Estimates and MQS of studies on reliability are presented in Table 3. Overall, three studies provided information about the percentage of missing values [39, 44, 46] and two studies described how missing values were handled [37, 47]. The MQS of the test-retest reliability studies ranged from 33 to 70 where 100 was the maximum possible score. The test-retest intervals ranged from 6 to 27 days. The studies on inter-rater reliability [36, 37, 41] examined the agreement of ratings from adolescents and their parents. All inter-rater reliability studies reached MQS of 29. The main flaw of test-retest reliability studies was not ensuring similar test conditions for both measurements (second survey was sent back with a return envelope) [36, 38, 41, 45, 47] and that of inter-rater reliability studies was not ensuring independent measurements (e.g. when participants answered the questionnaires at home) [36, 37, 41].

Six studies examined the convergent validity [36‐38, 40, 41, 43], six studies the structural validity [35, 38, 41, 43, 46, 47] and one study the criterion validity [35] of the questionnaires. Estimates and MQS of validity studies are presented in Table 4. None of the studies provided the percentage of missing values. Two studies described how missing values were handled [37, 47]. One study examined criterion validity by using a telephone interview as a ‘gold standard’ [35] (not shown in table) with a MQS of 50, and an agreement between telephone interview and questionnaire responses between 62.5 and 93.8% (Κ = 0.00–0.54) on the item level. Six studies assessed the structural validity using an exploratory factor analysis [13, 20, 24, 33, 41, 43]. Ommundsen and colleagues [43] also conducted a confirmatory factor analysis to perform a cross-cultural, age and gender validation of the instrument. The MQS of studies on convergent validity ranged from 29 to 71. The relationships with non-specific [36, 37, 40, 43] and behaviour or context-specific physical activity measures [36, 38, 40, 41] were tested. Only one study examined the relationship of neighbourhood environmental attributes with objectively measured physical activity [43]. All other studies used subjective measures of physical activity.

The evidential basis and thus the generalizability of the results of the studies on measurement properties was low partly because of selective samples, non-reporting of response rates, missing values and handling missing values. The main strengths of most studies were that they reported detailed characteristics of participants and had adequate sample sizes.

Many studies recruited selective samples only from urban areas, and sampling procedures were mainly based on school clusters or community based convenience samples, which might lead to clustered samples within similar neighbourhood environments. As discussed by some authors [45], convenience samples might not cover the entire target populations. However, Ommundsen and colleagues [43] used a two-stage cluster sampling procedure for recruiting a representative sample, and Huang and colleagues [46] recruited a sample with a reasonable representation of families from different socioeconomic status areas. Ommundsen and colleagues [43] examined the measurement properties in international samples (i.e. Norway, Denmark, Portugal and Estonia) and demonstrated cross-cultural invariance of their measure. Most other studies were conducted in the United States, and it is thus unknown if the employed questionnaires are appropriate for study populations outside of the United States. Our results further emphasize the previously stated [16, 48] need for international studies on measurement properties and additional evaluation studies conducted outside the United States for examining cross-cultural appropriateness of these measures, improving generalizability and facilitating international comparison of research findings of neighbourhood environmental impacts on physical activity behaviour.

Few of the reviewed studies reported response rates and percentages of missing values and how missing values were handled. Item and unit nonresponse may indicate selection bias and hence limit the generalizability of the results [26, 49]. High percentages and inappropriate handling missing values can lead to bias in parameter estimates, lower sample sizes and lower statistical power [50]. Non-random missing items may bias the results and lead to misinterpretation and misjudgement of measurement properties of instruments [26]. In addition, a high percentage of missing values on one item may indicate that this item is not relevant for the target population or that it is ambiguously formulated [23].

Considering the flaws of previous studies, we recommend that future research should accurately report response rates and provide information on handling missing values. In addition, randomness of nonresponse should be examined and reported in future studies.

Four test-retest reliability studies had high methodological quality [39, 44, 46, 51] and three of these studies also yielded substantial reliability estimates on all indices [39, 44, 46]. Thus, there is sufficient evidence supporting substantial test-retest reliability of the 9-item neighbourhood environment and the 11-item convenient facilities scales of Erwin [39], the 5-item neighbourhood safety and 5-item sports facilities scales of Huang and colleagues [46] and the 15-item physical environment, 9-item aesthetics and 5-item safety scales of Hume and colleagues [44]. These questionnaires are recommended based on their test-retest reliability. Nevertheless, to date the validity of these questionnaires has not been evaluated (except the structural validity in the safety scale of Huang and colleagues [46]) and should be assessed in further studies.

The test-retest intervals of the included studies on test-retest reliability ranged from 6 to 27 days and varied within some of the studies. To the best of our knowledge, to date an appropriate test-retest interval in reliability testing of questionnaires on physical environmental attributes has not been specified. Considering the relative stability of perceived physical environmental constructs such as facilities or street connectivity, potential memory effects should be avoided by choosing sufficiently long test-retest intervals. Moreover, one should consider that the first assessment may have drawn the participant’s attention to specific aspects of their neighbourhood environment and thereby potentially influence the responses of the second assessment [52].

Only three studies examined inter-rater reliability, and the methodological quality scores of these studies were low. Thus, to date no conclusion about acceptability of inter-rater reliability estimates can be drawn. It is possible that the agreement between self- and proxy-responses is an inadequate criterion of measurement quality because the perception of the environment may differ between parent and child [53].

A common flaw of reliability studies is that the questionnaire administration was not supervised by a researcher (e.g. when providing a home administration questionnaire). Unsupervised questionnaire administration leads to uncertainty about the time of administration (test-retest intervals), administrators and test conditions. This issue could be addressed by administering the instrument in the classroom under the supervision of a researcher, choosing a specified test-retest interval (e.g. one week) and conducting the second measurement exactly at the same time of the day as the first measurement.

Vague definitions of the area of interest (e.g. the neighbourhood) may have caused low test-retest and inter-rater reliability estimates in some studies [54]. Administrator responses could be influenced by different interpretations of the neighbourhood and therefore lead to divergent answers because, for instance, the term ‘neighbourhood’ might be associated with various geographical extensions. Therefore, questionnaires should define the exact area of interest. For example, Rosenberg and colleagues [36] and Huang and colleagues [46] defined the neighbourhood as “the local area around [the adolescent’s] home, within a 10–15 minute walk in any direction” and “the area within a 20-minute walk or drive from [the children’s] home”, respectively.

Therefore, we recommend that the neighbourhood should be concisely defined in questionnaires of neighbourhood environmental attributes and the administration of questionnaires should be standardized and supervised.

McMinn and colleagues [35] evaluated the criterion validity describing the degree to which the scores of an instrument are an adequate reflection of a ‘gold standard’. In this study, questionnaire responses were compared with interview responses of the same persons, that is interviews including the same items as the questionnaire were conducted with the participants who answered the questionnaire. Kappa estimates of this study were not substantial and did not support criterion validity of the perceived local environment questionnaire. The small number of studies on criterion validity may be related to the lack of a well-established ‘gold standard’ that could be used as a reference for new measures [35]. Objective environmental measures such as GIS data do not adequately reflect the same constructs as subjective environmental measures [18] and thus cannot be accepted as a ‘gold standard’. Several studies have shown low agreement between subjectively and objectively measured features of the physical environment [18, 55]. Physically active people may know their neighbourhood better then physically inactive people and hence achieve higher agreement between subjectively and objectively measured aspects of their neighbourhood [56]. The criterion validity study included in this review used an interview method as a ‘gold standard’ because this method enables a greater insight into participants’ answers and provides important feedback. In addition, both types of instrument administrations (questionnaire and interview) capture perceptions of the same target person and this procedure could produce correlated bias of both measurements and is prone to social desirability. The small sample size of the criterion validity study (N = 24), which is presumably related to the high cost and time needed for interview methods, queries the level of evidence of the results [35].

With regards to convergent validity, the studies of Ommundsen and colleagues [43] and Rosenberg and colleagues [36] had high methodological quality scores in combination with acceptable estimates of convergent validity. Therefore, sufficient evidence supports convergent validity of the 3-item opportunity and 2-item facility scales in the study by Ommundsen and colleagues [43] and the 66-items NEWS-Y in the study by Rosenberg and colleagues [36]. Because the participants of the study by Ommunden and colleagues [43] were randomly selected from four European countries, this measure seems to be appropriate for usage in different countries across Europe. Both questionnaires [36, 43] are recommended based on their convergent validity. Nevertheless, the reliability of the questionnaires developed by Ommundsen and colleagues [43] should be evaluated in further studies. Although the reliability of the NEWS-Y [36] has been assessed, the methodological quality of these reliability studies was limited. Thus, the reliability of the NEWS-Y should be evaluated in studies with high methodological quality.

Ding and Gebel [16] suggested to focus on conceptually matched associates between environmental attributes and domains of physical activity behaviour, and thus neighbourhood environmental measures should be behaviour- and context-specific [7]. For example, neighbourhood environmental barrier scales were developed to measure behaviour and context-specific environmental attributes such as barriers for walking and cycling to parks or barriers for physical activity in local parks [38, 41]. Common objective methods for assessing physical activity such as accelerometry on their own are not capable of capturing context-specific physical activity behaviours. The use of a more time consuming but feasible log booklet [57] or the ecological momentary assessment [58] are alternatives to these objective measures.

We recommend that convergent validity of instruments should be evaluated by examining theoretically linked neighbourhood environmental constructs and physical activity measures. Specific hypotheses regarding expected correlations with other constructs must be formulated a priori when developing a new measure.

As described above, we assume that neighbourhood environment is based on formative constructs. Some authors propose formative measurement models in physical activity environment measures [27, 59, 60], and low internal consistency across studies [35, 42‐44, 47] indicate that the underlying constructs are formative. Consequently, statistical procedures such as calculating Cronbach’s Alpha or factor analyses may not be appropriate for estimating internal consistency and structural validity. However, this issue is only rarely addressed in the scientific literature on physical environmental constructs in relation to physical activity behaviour. The studies on internal consistency and structural validity included in this review did not appropriately consider this issue when evaluating the measurement properties of the questionnaires. Future studies evaluating neighbourhood environmental measures should consider the measurement model of the underlying construct because of its importance for investigating measurement properties of the questionnaires [29]. Formative models should not be evaluated using statistical concepts such as internal consistency and structural validity [29]. Alternative methods for psychometric testing of neighbourhood environments include ecometric approaches [61‐63].

The major strength of this systematic review is the independent literature search and rating of the methodological quality of studies on measurement properties by two independent researchers, hence considering the evidential basis of the included studies. However, we only evaluated peer-reviewed journal articles that were published in English and did not include grey literature and articles not listed in the screened databases. In addition, we did not consult experts for ensuring that all relevant articles were included. Finally, we could not distinguish between low reporting and low methodological quality of studies, and hence low scorings of methodological quality may reflect weak reporting or weak study designs.