Noise Health Home 

[Download PDF]
Year : 2007  |  Volume : 9  |  Issue : 37  |  Page : 96--100

Factorial validity of the noise sensitivity questionnaire

Martin Schutte, Stephan Sandrock, Barbara Griefahn 
 Institute for Occupational Physiology at the University of Dortmund, IfADo, Ardeystraße 67, D-44139, Dortmund, Germany

Correspondence Address:
Martin Schutte
Institut für Arbeitsphysiologie an der Universität Dortmund, Ardeystraße 67, D-44139 Dortmund


The Noise Sensitivity Questionnaire (NoiSeQ) aims at the measurement of global noise sensitivity as well as the sensitivity for five domains of everyday life namely 'Leisure', 'Work', 'Habitation', 'Communication' and 'Sleep'. The present investigation examined the factorial validity of the NoiSeQ to determine whether the items of the NoiSeQ cover the different factors as assumed. The analysis was done using the method of Confirmatory Factor Analysis (CFA). The linear structural model took into consideration only the scales of the NoiSeQ for which reliability could be demonstrated, namely, 'Sleep', 'Communication', 'Habitation' and 'Work'. The linear structural model presumed that each of the 28 items has a relation only to one corresponding factor. Furthermore, the model allowed for correlations between the four factors. The data base encompassed 293 persons. Parameter estimation was based on the General Least Square method. The data was checked with respect to the occurrence of multivariate outliers, deviation from multivariate normality and existing collinearities. The data met the overall requirements of a CFA. The evaluation of model fit was based on the relative χ2 -value, the Root Mean Square Error of Approximation, the Goodness of Fit Index, the Adjusted Goodness of Fit Index and the Root Mean Square Residual. All fit indices indicated an acceptable match of the model. As the postulated structure of the NoiSeQ was consistent with the empirical data, the classification of the items as well as the claimed interdependencies between the scales can be maintained. The regression weights of all items as well as the correlations between the latent variables were statistically significant. The estimated reliability of the latent variables took values of ≥0.84. The findings generally justified the conclusion that there is no urgent need to modify the four scales of the NoiSeQ thus, indicating the factorial validity of the NoiSeQ.

How to cite this article:
Schutte M, Sandrock S, Griefahn B. Factorial validity of the noise sensitivity questionnaire.Noise Health 2007;9:96-100

How to cite this URL:
Schutte M, Sandrock S, Griefahn B. Factorial validity of the noise sensitivity questionnaire. Noise Health [serial online] 2007 [cited 2023 Feb 3 ];9:96-100
Available from:

Full Text

When persons are exposed to traffic noise, the predominantly occurring reaction is the experience of annoyance indicating the degree of undesirability of the particular noise. However, the same sound evokes annoyance ratings characterized by substantial interindividual differences. The individual level of noise sensitivity is considered to be one of the main causes for the large variation of annoyance data. Noise sensitivity is understood as a stable personality trait affecting the reactivity towards noise sources. According to the results of psycho-acoustic studies, noise sensitivity has no relation to auditory acuity, auditory reaction time and intensity discrimination but reflects a judgmental, evaluative predisposition towards the perception of sounds. [1] Noise sensitivity thus represents a moderator variable [2] influencing the relation between noise level and annoyance. [3]

The presently available (German) questionnaires for the determination of individual noise sensitivity such as the Weinstein-Scale (WNS) [4] or the "Fragebogen zur Erfassung der individuellen Lδrmempfindlichkeit (LEF)'' [5],[6] focus on global noise sensitivity. Nevertheless, the analysis of the factorial structure of the LEF led to a multidimensional solution indicating that separate measurements for different areas of everyday life might be more appropriate when determining the level of noise sensitivity. Therefore, the "Noise-Sensitivity-Questionnaire (NoiSeQ)" was developed for the measurement of global noise sensitivity as well as for the sensitivity concerning different domains of everyday life, namely, 'Leisure', 'Work', 'Habitation', 'Communication' and 'Sleep' where each subscale comprises seven items.

The reliability analysis was based on the G-Theory. The results of the G-study showed that the noise-sensitivity measurement is not affected by the age or gender of the respondents. Furthermore, the findings proved that a single application of the questionnaire is sufficient for the determination of individual noise sensitivity. [7],[8] If the NoiSeQ is used for determination of global noise sensitivity, the reliability (G-coefficient) for relative and absolute measurements reaches values >0.90. According to ISO 10075-3, the instrument satisfies precision level 1 - "accurate measurement" in this case. [9] Taking into consideration the lower limit of at least 0.70 required for orienting purposes as recommended by ISO 10075-3, the G-coefficients for the subscales exceed this value with the exception of the subscale 'Leisure' which has not proved satisfactory.

The present study analyzed the factorial validity of the NoiSeQ and therefore, it sought to determine whether the items of the NoiSeQ cover the different factors ('Sleep', 'Communication' etc.) as assumed. For these purposes, a Confirmatory Factor Analysis (CFA) was undertaken. The used approach was a strictly confirmatory one because it only examined whether the pattern of variances and co-variances in the data are in accordance with the specified structural model.


Structure of the linear structural model

The linear structural model took into consideration only those scales of the NoiSeQ for which reliability could be demonstrated. These were the subscales 'Sleep', 'Communication', 'Habitation' and 'Work'. Starting from the structure of the questionnaire, the linear structural model presumed that each of the 28 items has a relation only to one factor.

As empirical results concerning the situational specificity of noise sensitivity suggest that the different aspects of noise sensitivity are not completely independent from each other, [5] the model allowed for correlations between the four factors.

For this reason, the following two hypotheses formed the basis of the CFA analysis, namely, whether the four factor model with each variable loading only on one factor fits the data and whether there exist significant co-variances between the four factors of noise sensitivity.

Formal aspects of the model

To get information about the determinability of the model, the number of parameters to be estimated must be calculated. As the metric of the latent variables are unknown, a constraint must be introduced. One possible way is to set the variance of each factor to 1. Such a fixation leads to a standardized solution. [10] This strategy was used because no multiple group analysis was desired. Furthermore, the errors are also latent variables and must be scaled too. In this case, it is usual to set the regression weights to 1, which permits the estimation of error variances.

Accordingly, the specified model requested estimation of 28 variances, 28 regression weights and six co-variances, thus resulting in a total of 62 parameters. The empirical covariance matrix consisted of 28 variances and 378 covariances, that is to say, 406 values (28 × (28 + 1)/2). Accordingly, the model was overidentified and could be tested with 344 degrees of freedom.

CFA is based on analysis of covariances whose estimation presupposes large sample sizes whereas a relation of 10:1 between sample size and the number of variables is thought to be acceptable. [11] The sample of this study consisted of 293 persons resulting in a relation of 10 persons per variable, thus meeting the requirements.

For parameter estimation, the widely-used Maximum-Likelihood (ML) method was not applied as this procedure necessitates sample sizes of >500 persons in order to get reliable results. [12] Hence, the General Least Square (GLS) approach was utilized which is more suitable if the sample size is Analysis of data distribution and outlier detection

At first for each variable, the kurtosis and skewness as well as the corresponding standard errors were calculated in order to check whether the data corresponded to a normal distribution. The ratios of kurtosis and skewness to their corresponding standard errors can be regarded as standardized scores from a normal distribution whereas absolute values exceeding 2 indicate a deviation from normality. For some of the variables, this ratio took values of >2 necessitating a normalization for all variables. [13]

The Mahalanobis distances of all test persons from the group mean were calculated in order to determine whether there were multivariate outliers. The distances were evaluated using the χ2 -distribution with 28 degrees of freedom and p [14] According to this criterion, any case with a Mahalanobis distance exceeding the χ2 -value of 56.88 would represent a multivariate outlier. The maximum of the Mahalanobis distances took a value of 46.59 due to which no outliers could be detected.

Following this step, the assumption of multivariate normality was tested. The method used is based upon the fact that the Mahalanobis distances are in accordance with the χ2 -distribution if the empirical data correspond to multivariate normality. [15],[16] Therefore, the distances had to be sorted in ascending order and the χ2 -values corresponding to the distances determined. The Mahalanobis distances were then plotted against the χ2 -values. Generally, the empirical data fulfill the distribution requirements when the value pairs fall on a nearly straight line. [15] Furthermore, the Kolmogorov-Smirnov Test allows to test whether the peak of the empirical distance distribution conforms to the theoretically expected χ2 -distribution. [17]

Thus, there existed close relationships between the Mahalanobis distances and the χ2 -values [Figure 1] with the exception that in the upper part (distances > 42), the corresponding χ2 -values showed some larger deviations from the main diagonal.

Nevertheless, 48.12% of the distances were less than the theoretical χ2 -value for the 50 th percentile (χ2 = 27.34) indicating multivariate normality because the empirical gained percentage was close to the ideal value of 50%. The test-statistic D of the Kolmogorov-Smirnov Test took a value of 1.05 falling below the critical D-statistic amounting to 1.07 ( p = 0.20). Therefore, the results justified the conclusion that the data correspond to a multivariate normal distribution.

Screening for multicollinearity

In principle, variables included in a CFA-analysis should neither be highly correlated nor be redundant. [14] Accordingly, the correlation matrix was tested for very high interdependencies and thus for bivariate correlations above 0.90. [14] All correlations were below this limit. The highest correlation took a value of 0.68 and appeared between the item "When people around me are noisy, I don't get on with my work" (scale 'Work'-item 5) and the item "I find it very hard to follow a conversation when the radio is playing" (scale 'Communication'-item 5). Additionally the squared multiple correlations (SMC) of each variable with all other variables were calculated as part of a further screening for multicollinearity [Table 1]. The highest proportion of common variance amounted to 66% (scale 'Sleep': items 3 and 4), likewise confirming that multicollinearity was not a problem.

The so-called condition number of the co-variance matrix was then calculated indicating the numerical stability of the matrix. The parameter reflects the relation between the maximal and minimal Eigenvalue of the matrix taking a value of 45.62 (8.014/0.177) showing that the matrix possessed acceptable stability. [18] The data thus fulfilled the conditions precedent to the accomplishment of a CFA.

Proving the model fit

The evaluation of model fit can be based on various indices. Generally, it is recommended that different parameters be reported acting on different conceptualizations of goodness of fit. [19],[20] All coefficients should suggest the same conclusions concerning the adjustment of the model. The relative χ2 -value (χ2Min /df) reflects the discrepancy between the observed and the estimated population covariance matrix. This index should not pass over the upper limit value of 2. A value of 1.66 resulted for the existing model allowing the conclusion that the model has an acceptable fit [Table 2]. The Root Mean Square Error of Approximation (RMSEA) reflects the lack of fit in a model in relation to a perfect model. Values [21]

All regression weights of the items belonging to the subscale 'Work' significantly deviated from zero. With the exception of item 2 ("I have no problems to do routine work in a noisy environment") reaching a regression weight of 0.392, the standardized coefficients of all other items exceeded the lower limit value of 0.60. Accordingly, the squared multiple correlations indicating the proportion of variance in the items accounted for by the factor 'Work' varied between 15 and 61 per cent [Table 3]. The construct reliability (work-related noise sensitivity) took a value of 0.84 even though item 2 possessed a rather small weight.

Comparable findings concerned the subscale 'Habitation'. All regression weights significantly deviated from zero [Table 3]. Other than item 5 ("When other peoples' children are noisy, I would prefer that they should not play in front of my house") reaching a standardized regression weight of 0.520, the remaining items had loadings of 0.60 and more. The proportion of variance in the items explained by the factor 'Habitation' varied between 27 (Item 5) and approximately 60% (item 2). Even if item 5 showed a lower relation to the factor, the estimated reliability of the latent variable 'Habitation' amounted to 0.87.

Similar outcomes arose for the scale 'Communication' for which all regression weights were significantly different from zero. In the same way, one item, namely, number 2 ("When I am absorbed in a conversation, I do not notice whether it is noisy around me") was striking as the regression weight fell below the lower limit value of 0.6. The remaining items were characterized by loadings in the range between 0.611 and 0.771. The proportion of variance which can be traced back to the factor 'Communication' varied between 18 (item 2) and 59% (item 5) [Table 3]. The reliability of the factor added up to 0.84 despite the somewhat deficient characteristics of item 2.

The regression weights obtained for the subscale 'Sleep' also significantly deviated from zero. Likewise, one item, in this case, item 7 ("The sound of loud thunder does not usually wake me up") did not meet the lower limit value of 0.60. The loadings of the other items varied from 0.643 (item 6) to 0.849 (item 3). The corresponding proportions of variance accounted for by the factor 'Sleep' took values ranging from 10 (item 7) to 72% (item 3). The estimated reliability of the factor took a value of 0.89 although item 7 was characterized by an inferior measurement quality.

The postulated interrelations of the subscales could also be confirmed [Table 4] as all correlations were statistically significant ( p [1] then it should be verifiable that noise sensitivity is more relevant for evaluative variables like annoyance than for performance-related variables such as speed of task solutions. These are questions for future studies concerning the analysis of the validity of the NoiSeQ.

Moreover, it must be taken into consideration that the CFA is based only on four scales of the NoiSeQ. Correspondingly, after revision of the subscale 'Leisure', the analysis should be repeated taking into consideration all five subscales. As the NoiSeQ is available in various European languages, [22] it has to be analyzed whether the model is the same for all versions of the NoiSeQ.


1Ellermeier W, Eigenstetter M, Zimmer K. Psychoacustic correlates of individual noise sensitivity. J Acoust Soc Am 2001;109:1464-73.
2Baron RM, Kenny DA. The mediator-moderator variable distinction in social psychological rsearch: Conceptual, strategic and statistical consideartions. J Personality Soc Psychol 1986;51:1173-82.
3Job RF. Noise sensitivity as a factor influencing human reaction to noise. Noise Health 1999;3:57-68.
4Zimmer K, Ellermeier W. A German version of the noise sensitivity scale of Weinstein. Zeitschrift für Lδrmbekδmpfung 1997;44:107-10.
5Zimmer K, Ellermeier W. Design and evaluation of a questionnaire for recording individual noise sensitivity. Diagnostica 1998;44:11-20.
6Zimmer K, Ellermeier W. A short questionnaire for recording noise sensitivity. Umweltpsychologie 1998;2:54-63.
7Schütte M, Marks A. Development of the Dortmunder noise sensitivity questionnaire (DoLe). In : Society of Work Science (GfA), editior. Health + work in efficient work systems. GFA Press: Dortmund; 2004. p. 387-90.
8Schütte M, Marks A, Wenning E, Griefahn B. The development of the noise sensitivity questionnaire (NoiSeQ). Noise Health 2007;9:15-24.
9ISO 10075-3. Ergonomic principles related to mental workload - Part 3: Measurement and assessment of mental workload. International Organization for Standardization: Genf; 2004.
10MacCallum RC. Model specification - Procedures, strategies and related issues. In : Hoyle RH, editor. Structural equation modeling - Concepts, issues and applications. Sage: Thousand Oaks; 1995. p. 16-36.
11Ullman JB. Structural equation modeling. In : Tabachnick BG, Fidell LS, editors. Using multivariate statistics, 4th ed. Allyn and Bacon: Boston; 2001. p.653-771.
12Hu LT, Bentler PM, Kano Y. Can test statistics in covariance structure analysis be trusted? Psychol Bull 1992;112:351-62.
13Harter HL. Expected values of normal order statistics. Biometrika 1961;48:151-65.
14Tabachnick BG, Fidell LS. Using multivariate statistics, 4 th ed. Allyn and Bacon: Boston; 2001.
15Johnson RA, Wichern DW. Applied multivariate statistical analysis, 3 rd ed. Prentice-Hall: Englewood Cliffs NJ; 1992.
16Thompson B. Multinor: A fortran program that assists in evaluating multivariate normality, Educ Psychol Measurement 1990;50:845-8.
17Bortz J, Lienert GA, Boehnke K. Non parametric methods in biostatistics. Springer: Berlin; 1990.
18Sponsel R, Hain B. Numerically unstable matrices and collinearity in psychology. IEC Verlag: Erlangen; 1994.
19Bollen KA. Structural Equations with latent variables. Wiley and Sons: New York; 1989.
20Hair JF, Black WC, Babin BJ, Anderson RE, Tatham RL. Multivariate data analysis, 6th ed. Upper Saddle River. Prentice Hall: New Jersey; 2006.
21Lutz MT. Validity and factorial invariance of a neuropsychological test battery for recording intelligence of patients with epilepsy, unpublished doctoral thesis. Faculty of psychology and sports science at the University of Bielefeld: Bielefeld; 2006.
22Sandrock S, Schütte M, Griefahn B. The reliability of the noise sensitivity questionnaire (NoiSeQ) in a cross-national analysis. Noise Health 2007;9:8-14.