English: "To compare these probabilities with the original distribution of socially useless jobs, another model is estimated using only occupations as independent variables. Figure 4 displays the predicted probabilities from both these models. Results from the fully specified model are depicted in black and those from the ‘occupations only’ model are shown in grey.

Figure 4 displays several findings. First, one notices that most confidence intervals overlap and that there are no statistically significant differences between most occupations in these models. This is due to the small group sizes if all occupations are examined individually. These results should therefore be interpreted with caution. Nevertheless, they offer some valuable insights. Thus, one also notices that most predicted probabilities move closer to zero once control variables are included. This is not surprising because the control variables can explain some of the variance in the dependent variable. Thus, the remaining effect of working in an occupation represents the variance between occupations that is not explained by control variables, which also affects the predicted probabilities. This leads to different results for different occupations. ‘Transportation and material moving’ occupations, for example, display the highest value in the model without controls since they have the highest share of socially useless jobs. When control variables are included, however, the predicted probability shrinks by more than 50%. This means that the high share of socially useless jobs in this group can be explained to a large part by the control variables and the occupation itself is not that strongly associated with socially useless jobs. Interestingly, however, Graeber’s five occupations all end up on top of the list now, thus displaying the strongest association with being perceived as socially useless when control variables are included. Only ‘computer and mathematical’ occupations have a similarly high predicted probability.5 Assuming that all relevant factors are controlled for, this further supports Graeber’s theory."

"Notes: Both models are based on the full sample of 1811 workers. The model with control variables includes all variables from model 1 in Table 4. Pseudo R2 (McFadden) is 0.05 for the model without control variables and 0.18 for the model including control variables. Predicted probabilities are estimated while fixing all other independent variables at their mean values."