TITULO III. DE LA PARTICIPACIÓN SOCIAL.
4 CAPÍTULO 4 EVALUACIÓN DE LA SITUACIÓN ACTUAL DE LOS COMPONENTES
4.1.1 Caracterización general del cantón Manta.
To test for the statistical significance of the wave-on-wave changes, a repeated measures ANOVA was carried out on those respondents who answered the relevant questions in Wave 1 and all subsequent waves. The results which emerge from this analysis, shown in Table 7.3 below, reveal interesting differences from those for the full-wave sample in each year.:
Wave 1 2 3 4 5 6 Average all
waves COMPLEX 4.13 4.05 4.04 4.10 4.27 4.23 4.14 NUSKILLS 4.83 4.64 4.50 4.55 4.67 4.56 4.62 USESKILL 5.56 5.42 5.47 5.48 5.52 5.45 5.48 OWNTASK 4.99 4.89 4.91 4.92 4.97 4.88 4.93 HAVESAY 4.56 4.45 4.52 4.53 4.66 4.58 4.55 WORKFLOW 3.60 3.61 3.67 3.67 3.80 3.78 3.69 Skill-intensity 14.54 14.11 14.01 14.14 14.48 14.26 14.26 Task discretion 13.15 12.96 13.09 13.14 13.44 13.25 13.17 Table 7.3
Estimated marginal means, Waves 1-6
(repeated measures ANOVA, respondents who answered in all waves)
Change over the whole five years was found to be statistically significant at the .01 level for all six variables and both scales, but failed in many cases to reach the .05 level of
differences occurred in the estimated marginal means for each variable between each wave and all the others, rather than just the changes from one wave to the next. This information is important to understand the degree of variation that occurred in each variable over the full period. Wave 1 2 3 4 5 6 Partial η2 N Skill intensity 2, 3, 4, 6 1, 5 1, 5, 6 1, 5 2, 3, 4* 1, 3, 5 .037 2961 Task discretion 5 5, 6 5 5 1, 2, 3,4 2 .014 2980 COMPLEX 3*, 5 5, 6 1*, 5, 6 5, 6 1,2,3,4 2, 3, 4 .027 3001 NUSKILLS 2, 3, 4, 5, 6 1, 3 1, 2, 5 1,5 1,3,4,6 1, 5 .043 3005 USESKILL 2, 3*, 4* 1, 5 1* 1* 2, 6* 5* .012 3009 OWNTASK 2, 6* 1 6* 5* .008 3010 HAVESAY 2 1, 6 5 5 2,3,4 2 .015 3008 WORKFLOW 5, 6 5, 6 5, 6* 5,6 1, 2, 3, 4 1, 2, 3*, 4 .015 3010 Table 7. 4
Significant differences between estimated marginal means, Waves 1-6 (repeated measures ANOVA – respondents who answered in all waves)*
*Findings marked with a single asterisk are significant at .05. All others, including η2, are
significant at .01.
The partial eta-squared figure in the second-last column is an indicator of effect size over the full period, based on Wilks’ Lambda. Effect sizes of this magnitude would be
considered trivial if the purpose were to determine the short-term impact of a purposive intervention, but can be regarded as meaningful if small when applied, as here, to effects arising from an unknown variety of causes over an unknown but possibly quite extended period. The main reason for including them here is identify which of the variables or scales show the strongest and weakest change over time. The provide a more statistically rigorous confirmation of the impression gained from the findings of the first two analyses that skill- intensity showed more movement over time than task discretion, and that USESKILL is the individual variable which shows most year-to-year change. On the other hand they show that the variation over the six waves for COMPLEX is much more pronounced, at least for this more limited sample, than would appear from either the raw or the binned scores. The findings reached by this method permit a number of substantive conclusions, some of which reinforce trends appearing in the aggregate means and binned scores, while others contrast with them:
• While both scales show a declining trend in aggregate mean scores over the full period, the task discretion scale for this more limited population trends slightly but significantly upwards from Wave 3 on;
• Movement on the task discretion scale is nevertheless much smaller than for the skill-intensity scale. Without the spike in scores for Wave 5 this scale would have
• The rise in Wave 5 continues to stand out on both scales and most variables. However, where the aggregate figures suggest a return to a lower trendline after a single-year spike, the relative paucity of significant changes between Waves 5 and 6 leaves open the possibility that on some variables at least, this rise may signal the start of a sustained rising trend;
• The drop following Wave 1 also remains evident but less pronounced, with
significant change between Wave 1 and 2 affecting only four out of the six variables, and only the skill-intensity scale;
• COMPLEX and NUSKILLS show the greatest relative volatility, with the latter having the greater range of variation, as shown by its higher (though still low) eta- squared coefficient. NUSKILLS is also the only variable for which significant differences occur between a single wave and all the rest. The variables exhibiting greatest stability are OWNTASK and USESKILL, with the former showing no significant change over the middle three years;
• Significant change between adjacent waves (shown in bold in the table) is relatively uncommon, especially in the middle years, suggesting that trends over this timescale emerge either very slowly or unevenly.
Many of the discrepancies between these marginal means and the aggregate means (i.e. the means for all respondents in each wave) can be explained by the drastic sample loss which is the price of moving to a sample that remains genuinely constant from wave to wave. Although the size of the remaining sample is fully adequate to support the kinds of inferential analysis for which it is used here, the choice brings an inevitable tradeoff between validity of inference (i.e. confidence that the changes observed reflect genuine changes in the construct of interest rather than random variations in sample composition) and representativeness (confidence that the remaining sample reflects the composition of the full population as accurately as the full achieved sample in an individual wave). The latter must be in question given that there is no reason for confidence that the non- response is randomly distributed. It was noted in Chapter 4 how the managers of HILDA have calculated that panel attrition is concentrated among those members most likely to be poorly educated, low-paid and precariously employed. If this assumption is extended to item non-response in individual waves, one might expect the all-waves sample to represent a slightly different population from the designed or indeed the full achieved sample, one which was likely to be in higher-skilled or higher-quality work, and hence to score higher on most of the key indicators. This is precisely the impression that emerges from Figure 7.5, where the movements in the two sets of means are plotted together.
While the movements of the two curves on each scale, particularly skill-intensity, are generally in the same direction and of roughly comparable magnitude, those respondents who answered in all waves score consistently 0.5 to 0.75 higher than the full set of respondents in wave. Such a finding would be consistent with a population more likely than the average randomly selected member of the public to be employed, permanently employed, educated, and engaged in challenging or responsible work.
12.50 13.00 13.50 14.00 14.50 15.00 1 2 3 4 5 6 Skill-intensity (observed) Skill-intensity (est. marginal) Task discretion (observed) Task discretion (est. marginal) Figure 7.5
Composite scales: observed mean scores (all respondents in wave) vs. estimated marginal means
(repeated measures ANOVA, respondents who answered in all waves)
In addition to these possible differences in the key parameters, the all-waves sample is more likely than later-wave recruits, later-wave dropouts or sporadic responders to feel the
impact of any panel conditioning that has taken place: for example, to have become “better” at answering surveys, more familiar with the individual questions and their place in the respective sequences, and more attuned to the issues about which they are being asked to respond. Conversely, they may be closer to experiencing survey fatigue than those who have been in the panel for fewer years or less continuously. While the data offer no clear evidence of panel conditioning in the HILDA panel except for the first two waves, these two risks together make it necessary to look closely at the patterns of non-response before any assessment can be made of the relative merits of the three approaches to estimating change over time.