The present paper examines the PPS and IPS, as well as the complete DPS, AIP, and GPS instruments in a global data set with 4,169 participants. Using Confirmatory Factor Analyses (CFI) and Item Response Theory (IRT), we compare the different scales, examine their factor structures, and suggest simplified scales to measure procrastination. We also assess measurement invariance over gender and age, internal reliability, as well as correlations between the instruments.

The first purpose of the present study was to examine the possibility that the PPS addresses three rather different facets of procrastination (Svartdal et al., 2016), and that the middle five items of the PPS correspond to IPS in measuring “irrational” delay. As the GPS seems to measure this construct also, the full GPS as well as reduced versions were examined. Lay (1986) hypothesized the GPS to measure a unidimensional construct, but as mentioned, subsequent studies have not supported this assumption and have instead suggested different factor structures (Argiropoulou and Ferrari, 2015) or a reduced version (Klingsieck and Fries, 2012). We examine these possibilities, the latter being particularly interesting because the five implemental items of the PPS (4–8) are in fact GPS items. Thus, the possibility that the GPS could be reduced to five items is tested. Finally, the DPS and AIP were also examined. Recall from Table 2 that the DPS focuses on decisional procrastination, whereas the AIP contains several items that focus on timeliness and lateness. As the PPS includes items from both these scales, we ask how these PPS items perform compared to the full scales. In effect, we pursue the possibility that the DPS and AIP could be reduced to fewer items, likely corresponding the PPS items, and in consequence that the PPS could replace the DPS, AIP, and the GPS scales.

The sample consisted of 4169 respondents (57.4% women) completing an online survey. Mean age was 37.4 years, the most frequent age group being 20–30 years (1200 respondents). Most participants were located in the United States (68.1%), 5.9% in Canada, 4.4% in the United Kingdom, 2.4% in Australia, 1.6% in Italy, with the rest distributed among a large number of countries worldwide with 1–40 respondents/country. Respondents were recruited to participate in a study on regret when visiting a procrastination-themed website.

All respondents answered a questionnaire consisting of standard demographic questions followed by items from the complete DPS, GPS, AIP, and IPS scales. The DPS (Mann, 1982, unpublished; Mann et al., 1997) contains five items that primarily focuses on delay in planning and decision making, e.g., “I waste a lot of time on trivial matters before getting to the final decisions,” though has one item related to implementation, “Even after I make a decision I delay acting upon it” (DPS 2). Internal reliability for the DPS is relatively high, α = 0.70–0.83 (Mariani and Ferrari, 2012). The GPS (Lay, 1986) encompasses 20 items focusing primarily on implemental delay, e.g., “Even jobs that require little else except sitting down and doing them, I find that they seldom get done for days” (GPS 7). Two versions of the GPS exist, a general version and a version adapted for students specifically. The general version was used here. It has a good internal consistency, α = 0.86 (Lay, 1986). The AIP (McCown et al., 1989) contains a mix of items addressing decisional and implemental delay, as well as lateness (see Table 2). Test-retest reliability of this scale is relatively high, r = 0.71, as is internal consistency, α = 0.86 (Ferrari et al., 2005). The IPS (Steel, 2010) is a nine-item scale focusing on implemental delay, e.g., “I delay tasks beyond what is reasonable” (IPS 7). The IPS demonstrates good internal reliability, α = 0.91 (Steel, 2010). Of note, the PPS was not included as a separate scale, as this scale is composed of 12 items from the DPS, GPS, and AIP. Steel (2010) reported internal consistency of the PPS at α = 0.92. For discriminant validity purposes, respondents answered the five-item Satisfaction with Life Scale (SWLS; Diener et al., 1985). All items were answered on a common 1–5 scale, 1 = “Very seldom or not true of me,” 5 = “Very often true, or true with me.” All answered a total of 159 items. First, respondents answered the demographic questions, then the procrastination scales of the present study and finally the SWLS. Items were presented in fixed order, one scale at a time.

Participation was voluntary, anonymous, and confidential. Participants read a consent form describing the nature and purpose of the study and then provided written informed consent before responding. No payment was provided. The project of which this study was a part received ethics approval from the Conjoint Faculties Research Ethics Board (CFREB) at the University of Calgary.

The item scores were first examined for skewness and kurtosis. Then multivariate normality was assessed for all scales, in particular multivariate kurtosis, which is important to parameter estimation in CFA (Byrne, 2008). Non-normality was apparent in each scale according to the Mardia skewness and kurtosis tests. Hence, we report the Satorra–Bentler scaled chi-square statistic which is robust to multivariate non-normality (Satorra and Bentler, 2001). Configural fits to the suggested models were evaluated in CFA according to the root mean square error of approximation (RMSEA), the Bentler comparative fit index (CFI), and the standardized root mean square residual (SRMR) (Byrne, 2001). Acceptable goodness of fit adopted the standard criteria of RMSEA < 0.08, CFI values in the 0.90–1.00 range, and SRMR < 0.08 (Brown, 2015; Kline, 2016). After having established acceptable configural baseline models for the PPS and IPS, those models were tested for measurement invariance over gender and age groups, using standard procedures to test for configural, metric, and scalar invariance (Byrne, 2001; Gregorich, 2006; Brown, 2015). Scales reflecting a single latent construct were also analyzed by IRT using the graded response model (GRM), focusing on parameter a (discrimination) and the difficulty parameter (e.g., Fraley et al., 2000). CFAs and IRTs were performed using the SEM and IRT modules in STATA 14.2¹.

As discussed, four different factor models have been suggested for the PPS. These are shown in Table 3 along with the CFA fit indices for the present data. As is seen in the table, the three-factor model for the PPS – items 1–3 measuring decisional procrastination, items 4–8 measuring implemental delay, items 9–12 measuring timeliness, and promptness – was superior to the other suggested models. As the one-factor model and the suggested three-factor models are nested, a Δ Chi squared comparison between these models indicates whether one model demonstrates a better fit (Brown, 2015). This difference was significant, Δ Chi squared = 2062.85, Δ df = 3, p < 0.001. Also, the Δ CFI between these models was 0.07, well above the 0.01 criterion suggested by Cheung and Rensvold (2002).

An examination of modification indices of the three-factor solution indicated a path to be added between the PPS factor timeliness/delay and PPS item 1. This connection is reasonable because this item explicitly addresses timeliness/lateness (“I delay making decisions until it’s too late”). Adding this path improved fit, RMSEA = 0.065, CFI = 0.968, SRMS = 0.029.² In the next iteration of modification indices analysis, a path from PPS factor implemental delay and item 9 was suggested. This item (“I find myself running out of time”) has previously been argued to be conflated with the busyness construct and not procrastination per se (Steel, 2010; Svartdal et al., 2016) and recommended for deletion from procrastination scales. Deleting it improved overall fit, RMSEA = 0.062, CFI = 0.974, SRMS = 0.028. In the final iteration, modification indices indicated a path from PPS factor implemental delay and PPS item 3 (“I waste a lot of time on trivial matters before getting to the final decisions”), improving fit even more, RMSEA = 0.043, CFI = 0.989, SRMS = 0.015. In this model, shown in Figure 1, correlations between PPS factors were all < 0.75, indicating discriminant validity. As a more formal test of discriminant validity, we compared the squared correlations (SC) between factors with the average variance extracted (AVE) by the latent variables (Brown, 2015). All AVE values were higher than the SC values, indicating discriminant validity, and all AVE values were higher than 0.05, indicating convergent validity.

Addressing the PPS parts specifically, PPS items 4–8 should measure implemental delay satisfactorily. A CFA of these items indicated good fit, RMSEA = 0.064 (0.053), CFI = 0.991 (0.992), SRMR = 0.016. Regarding the IRT analysis of this construct, parameter a coefficients were > 2 for all items, item 7 demonstrating the lowest coefficient (2.15) and item 4 the highest (2.71). All items covered the range of the latent construct quite well, from -3 to +2, indicating that this short scale measures implemental procrastination well in individuals in the normal range of the latent trait. Note that the scale discriminates rather poorly in the higher end of the latent construct, speaking for cautious use in clinical settings. Test Information Function (TIF) and Item Information Functions (IFF) graphs are shown in Appendix. A corresponding examination of PPS items 9–12 (lateness/timeliness items from the AIP) indicated an excellent fit, RMSEA = 0.056 (0.049), CFI = 0.996 (0.997), SRMR = 0.011. The IRT demonstrated discrimination coefficients between 1.72 (item 12) and 3.50 (item 10). Examination of the TIF graph (Appendix) again indicated rather poor discrimination in the higher end of the latent construct. Finally, examination of the decisional part of PPS, items 1–3, demonstrated an excellent fit, RMSEA = 0.000, CFI = 1.00, SRMR = 0.000. The IRT discrimination coefficients ranged from 2.18 (PPS item 1) to 3.85 (item 2). Again, the TIF graph (Appendix) indicated less reliability in the higher end of the latent construct. In summary, the three-factor model of the PPS, as well as reduced models focusing on three unidimensional constructs, decisional, implemental delay, and lateness/timeliness, all appear to work well psychometrically.

Given the basic configural model of the PPS, we tested invariance across gender and age groups. Both gender and age differences (i.e., 30 years and above versus below) have been discussed repeatedly in the literature (e.g., Steel and Ferrari, 2013; Beutel et al., 2016), but as scalar measurement invariance is a prerequisite for meaningful comparisons of means over populations (Gregorich, 2006; Brown, 2015), conclusions about such differences cannot be settled until invariance has been established. As shown in Table 4, a multigroup men vs. women CFA indicated configural as well as metric invariance, but not scalar invariance. Hence, comparisons of PPS means with gender is problematic. Also note that configural fit was improved for participants > 30 years of age. The results further indicated that gender differences appeared in the decisional and lateness parts of the PPS (z = -3.65 and -6.58, p < 0.000), but not in the implemental part (z = -0.28, p = 0.777). Hence, invariance tests of PPS items 4–8 demonstrated a similar pattern to that of the complete PPS, with the important exception that scalar equivalence was now observed for the age group > 30 (see Table 5). In summary, the PPS results indicate that the complete scale does not attain full invariance across gender, and furthermore the model fit was better for participants greater than 30 years of age. For the reduced PPS (items 4–8, i.e., the implemental part), gender differences were minimal and these items also demonstrated full measurement invariance for participants over 30 years, indicating that this part of the scale permits comparisons of means scores for adults.

The IPS is hypothesized to measure a single construct, “irrational delay,” and the present data indicates that it does, RMSEA = 0.075 (0.066), CFI = 0.971 (0.973), SRMR = 0.032. In accordance with prior findings (Svartdal et al., 2016), modification indices indicated that the reversed items should be correlated. This resulted in an improved fit, RMSEA = 0.058, CFI = 0.984, SRMR = 0.019. Omitting the reversed items improved fit slightly. These analyses thus support the hypothesis that the IPS confirms to a single latent construct, implemental or irrational delay. Omitting the reversed items improves fit indices and provides an instrument that is more easily administered and scored. Supporting this, the IRT indicated good parameter a (discrimination) coefficients for all procrastination-consistent items (range 2.24–3.14), item 4 being lowest but covering the higher range of the latent construct better. The reversed items – and particularly items 2 and 9 – demonstrated the lowest coefficients (item 9 = 1.41; item 2 = 1.49). As for the PPS subscales, the TIF graph (Appendix) indicated rather poor discrimination in the higher end of the latent construct.

Testing IPS measurement invariance over gender and age groups indicated somewhat better fit for age over 30 years, but as is apparent from Table 6, even in the older group scalar invariance did not appear, indicating that care should be taken in comparing mean IPS scores between genders and age groups.

Given that items 4–8 of the PPS measure implemental or “irrational delay,” this part of PPS should correlate highly with IPS, whereas the two other factors of the PPS should demonstrate more moderate correlations. As is seen from Table 7, this was the case, r = 0.83 vs. 071 and 0.76. Further, IPS and PPS item means 4–8 should be comparable, and for the present sample they were, at 3.62 in both cases. These results indicate that PPS items 4–8 and IPS address the same unidimensional construct, implemental delay.

As mentioned, Lay (1986) proposed the GPS as a scale measuring a unidimensional construct procrastination, but Argiropoulou and Ferrari (2015) suggested a two-factor solution (delay and procrastination domains), and a German study, testing the student version of the GPS, proposed a reduced version – GPS-K – consisting of items 1, 2, 7, 12, 14, 15, 18, 19, and 20 (Klingsieck and Fries, 2012). These items (except items 2 and 14) are identical in the general version of GPS. In the present study, the one-factor model for the complete GPS did not demonstrate a good fit, RMSEA = 0.088, CFI = 0.824, SRMS = 0.058; the two-factor model indicated somewhat better fit, RMSEA = 0.070 (0.064), CFI = 0.89 (0.89), SRMS = 0.058. In both cases, however, the CFI criterion was not acceptable. The German reduced model, excluding items 2 and 14, indicated a somewhat better fit, although not acceptable, RMSEA = 0.088 (0.076), CFI = 0.96 (0.97), SRMS = 0.031.

As the other general procrastination scales discussed, the AIP is hypothesized to measure a single latent construct, procrastination, and Mariani and Ferrari (2012) reported support for a single-factor latent model in an Italian sample. The present data did not indicate a good fit for this model, RMSEA = 0.135, CFI = 0.761, SRMS = 0.076. Mariani and Ferrari (2012) reported an even better fit when errors of items 13, 4, 7, and 8 were allowed to correlate. This is theoretically reasonable, as these items concern things to do before a deadline. Again, this model did not improve fit indices in the present data. Thus, the present data did not support either of the suggested factor solutions for the AIP. Analysis of individual items indicates that the AIP focuses on rather different aspects of procrastination (see Table 2), which in part may explain why this scale did not do well in the CFA analyses.

The DPS demonstrated a poor fit for a one-factor solution, RMSEA = 0.259, CFI = 0.916, SRMS = 0.076. Modification indices suggested correlations between errors for items 4 and 5 (two items with quite similar wording), and then between items 1 and 2. This model indicated an excellent fit, RMSEA = 0.017, CFI = 1.000, SRMS = 0.002.

Table 7 shows the Cronbach’s alphas, correlations and mean scores for the DPS, AIP, and GPS scales as well as for the IPS and PPS (complete and subscales). SWLS is also included to evaluate divergent validity. Overall, good convergent validity was observed between the procrastination scales, and divergent validity to the SWLS was apparent for all instruments. Note that the complete PPS correlates highly with the GPS, AIP, and DPS (all correlations > 0.81), making the PPS a briefer alternative to these scales. Also note that the DPS total scale correlates very highly with the first factor of the PPS, r = 0.97, effectively making the DPS part of the PPS equivalent to the complete DP scale. Similarly, the complete AIP correlates highly with the PPS factor containing AIP items, r = 0.86, making these four items comparable to the complete AIP.

As scalar measurement invariance was demonstrated in the PPS 4–8 subscale for age > 30, we plotted mean PPS 4–8 subscale scores over age (decades). This is shown in Figure 2. The figure indicates a slight reduction of procrastination over age decades 40–70, supporting the view that procrastination decreases with age (Steel and Ferrari, 2013; Beutel et al., 2016). For illustrative purposes we also plotted the other scales and subscales in the figure. Note that all scales agree to an overall decrease over decades, one deviation being the PPS items 9–12 subscale, indicating that timeliness/lateness forms of procrastination increase until 40 years of age, then decreases. However, this result must be interpreted with great caution, as scalar invariance was not observed for other scales or subscales.