Explaining the intention to participate in a web survey: a test of the theory of planned behaviour

Abstract

Even though web surveys have become increasingly popular, considerable efforts are necessary to obtain acceptable response rates. This explains the proliferation of experiments aimed at improving levels of participation in web surveys. The current study is not aimed at increasing web survey response by means of experimental research, but instead uses a general psychological theory to explain web survey response. More specifically, the theory of planned behaviour (TPB) is used to explain the intention to participate in a web survey. Previous studies have used this theory to explain survey participation in specific populations (students), but the current study extends that scope by targeting a general population. The results show that the TPB is capable of explaining people’s intentions to participate in a web survey in the context of a general population.

Keywords:

• survey response
• online survey
• structural model
• LISREL

Acknowledgements

The authors would like to thank Caroline Roberts from City University London for her much appreciated assistance in preparing this manuscript for publication. The authors would also like to thank the anonymous reviewers for their constructive comments. This publication was made during the last part of a Research Fund Grant (BOF) of the Katholieke Universiteit Leuven and during the first part of an FWO Flanders post‐doctoral research grant (both grants extended to Dirk Heerwegh).

Notes

1. In the remainder of this article, the terms ‘web survey response’ and ‘response rate’ will refer to unit response.

2. Please note all questions have been translated from Dutch. Some of the original meaning of the questions may have been lost in the process of translation.

3. After this question, it was explained to respondents that no emails would actually be sent to them and they were not asked to provide their email address.

4. This assumes that the sample was properly drawn and that it is also representative of the target population. A separate analysis that is not reproduced here showed that this was indeed the case.

5. The asymptotic covariance matrix was calculated using PRELIS, as were the polychoric correlations. These matrices were used to estimate the measurement and structural models (these estimation procedures used the weighted least squares method), as this is the appropriate method for analysing ordinal data (Jöreskog, 2005). The polychoric correlation matrix is reproduced in Appendix 1. The appropriate tests regarding the polychoric correlations revealed no violations of the assumed underlying bivariate normality using the RMSEA > 0.1 criterion as suggested by Jöreskog (2005, p. 18).

6. Because subjective norm was measured by two variables, this part of the measurement model is statistically under‐identified. Therefore, the model imposed a restriction on these loadings in that they were made to equal each other. This decreases model fit but avoids statistical under identification (cf. Byrne, 1998).

7. NNFI = non‐normed fit index; IFI = incremental fit index; CFI = comparative fit index; GFI = goodness of fit index; RMSEA = root mean square error of approximation; SRMR = standardised root mean square residual; χ ² = Chi‐squared; df = degrees of freedom. Conventional threshold values indicating model fit are >0.9 and >0.95 for NNFI, IFI, CFI, and GFI; <0.08 and <0.06 for RMSEA; <0.08 for SRMR; and χ ²/df values smaller than 2 (Beauducel & Wittmann, 2005, p. 51). Only the SRMR value suggests inadequate model fit.

8. Moreover, it should be noted that the previous studies were conducted in different countries. Therefore, the similarities between study results could be misleading. Cross‐cultural differences combined with different sample types may be responsible for seemingly similar results (opposite effects may have occurred which cancel out each other). Therefore, one should not hastily conclude without further evidence that convenience samples provide an equally firm basis as random samples to test the TPB.