Abstract of an unpublished paper that examined distribution of absenteeism data and concluded that in many cases inappropriate methods of analysis failed to establish the connections that actually exist between absenteeism and other variables.
Abstract of an unpublished paper that examined distribution of absenteeism data and concluded that in many cases inappropriate methods of analysis failed to establish the connections that actually exist between absenteeism and other variables.
Abstract of an unpublished paper that examined distribution of absenteeism data and concluded that in many cases inappropriate methods of analysis failed to establish the connections that actually exist between absenteeism and other variables.
The Validity and Analysis of Absenteeism Data: Some Methodological Issues
During a research project in a large Western Australian organisation, a survey
examining various attitudes and behaviours was conducted. Absenteeism was one of the behaviours of interest and respondents were asked to indicate the total number of sick days they had taken during the previous twelve months. The researchers were also provided with absence data relating to the whole organisation for the relevant period. Many respondents voluntarily provided their distinct staff numbers, which enabled a direct comparison to be made between their self-reported measures and their absenteeism according to the organisational records. When the self-reported measure was regressed on data extracted from organisational records, approximately 37.4% of variance was explained. Error, however, was not normally distributed (Kolmogorov-Smirnov z = 5.60, p<.001), with both skewness (6.47) and kurtosis (70.24) being well outside generally-accepted limits. These results suggested that the variables (which themselves were highly skewed, leptokurtic and truncated at the lower end) needed to be transformed if linear regression was to be employed as an analytical tool. A transformation that would reduce the skewness and kurtosis of both variables seemed likely to be required, so square root transformations were employed. Regression using the transformed variables explained 48.1% of variance. The error term still was non-normally distributed (z = 3.20, p<.001), but both the skewness (2.92) and kurtosis (19.49) were markedly reduced. Clearly, the transformations had improved the effectiveness of linear regression as an analytical tool for these data. Several inferences flowed from these analyses. First, it seemed that the validity of self-reported data may not have been particularly high, given that only around 50% of the variance in the measures was common. Secondly, it appeared that many correlational analyses which involved raw absenteeism data may, as a consequence of the distribution of the data, have underestimated the strengths of associations between absenteeism and other constructs. If self-reported measures, rather than organisational records, were employed, it was possible that this problem may have been exacerbated. To test these inferences, regressions of both raw and transformed organisational and self-reported absenteeism data on a number of other constructs were performed. The results of these analyses strongly supported the propositions advanced above. This paper concludes that many of the studies that have failed to find theoreticallypredicted links between absenteeism and other constructs may have failed not because the theory was flawed, but because the measures employed were flawed, the methods of analysis used were inappropriate, or both.