6. Results
6.2.4. Results from other data sets and additional operationalisations
Appendixes C and D presents the models of hypotheses 3, 4, and 5 run with ABII and WVS6 data. The central information provided by these two other data sets, ABII and WVS6, is consistent with the results from ABIII. Generally being unemployed does not have a
statistically significant impact on the likelihood of an individual to take part in protests. When there are statistically significant results they primarily indicate that the unemployed are less active protesters compared to the employed. In addition, based on ABII and WVS6 there are no statistically significant results indicating that the unemployed would protest more probably in the four countries most affected by the Arab Spring, namely Egypt, Tunisia, Libya, and Yemen. When relative and absolute income are used to measure employment and all countries are studied together, both ABII and WVS6 show a statistically significant relationship, again consistently with ABIII; the more people earn the more likely they are to protest.
Interestingly it seems that according to each data set, out of four separately studied countries, Tunisia is the country where the unemployed are most active protesters compared to the employed. This notion can be made based on the use of the employment status question in hypothesis 3. As the results are statistically insignificant this is not a robust finding. Still, in ABIII the Tunisian unemployed are practically as likely to protest as the employed, the
coefficient is positive but very low and its p-value is 0.92, in ABIII also the coefficients for the other three separately studied countries are positive, but the coefficients are higher and their p-values are respectively lower, reaching statistical significance in Libya. In ABII only Tunisia, Egypt, and Yemen are included, but the situation is same, for Tunisia the coefficient of this variable is positive and low with a p-value of 0.98, whereas for other two countries the coefficients are greater and their p-values lower. In WVS6 again three countries are included, this time Tunisia, Libya, and Yemen and there Tunisia’s coefficient is negative but
statistically insignificant, in Libya and Yemen the coefficients are according to WVS6 positive.
It is also interesting, that as with ABIII, also WVS6 indicates that in the sample of “other countries” the unemployed are statistically significantly more active protesters than the employed when measured by the employment status question. In contrary, when income questions are used, WVS6 indicates that the unemployed are statistically significantly less active protesters compared to the employed. And the third data set, ABII, does not indicate that the unemployed of the “other countries” would be more active compared to the
provide any strong evidence to support our hypothesis, but I think it indicates that there are some differences between countries in how active the unemployed are. Partly the countries included in the sample of the “other countries” in ABIII and WVS6 are the same. Thus, possibly a more detailed analysis would show that statistically significant supportive results of “other countries” appear because there are same countries in the samples of “other countries”.
I have also tested how some changes or alternative operationalisations in the models of research question II affect the results. These variations are tested with the seven samples of the ABIII data set. I have not included their results as tables in the thesis nor in the
appendixes. First I tried to add marital status as control variable, because it could be assumed, for example, that bachelors protest more often. I tried two ways to have this as a binary variable, first coding bachelors as their own group and having others as the reference group, and secondly coding the married as their own category and having others as the reference group. In chi-square both these operationalisations indicated a statistically significant relationship. However, when this was added as a control variable in the base model of
hypothesis 3, it lost its statistical significance whichever operationalisation was used. It might be that the higher protest participation among bachelors which was evident in the chi-square tests, is partly explained by their age; young respondents protest more often and are more often bachelors. Adding marital status as a control variable neither appreciably increased the coefficient of determination nor affected the coefficients or the statistical significance of other variables. Thus I have left marital status out from my analyses.
In hypothesis 5 I divided the data to study the educated unemployed. Another way to study the interaction effect of two variables would be to include an interaction term in the regression model. In practice the interaction term for two variables means that these variables are multiplied with each other. In this case, it is not reasonable to add an interaction term to the current base model for the two variables measuring employment status, namely Employed and Other. Doing so would actually test the interaction between education and being
employed or belonging to the group of “others”, and based on this it would be challenging to interpret whether there is an interaction between being unemployed and educated. An
alternative is to exclude everybody but the employed and the unemployed from the analysis, and code the binary variable as 0 for the employed and 1 for the unemployed, and this is what I tested with ABIII. As a result I expected a positive coefficient for the interaction term, but in four samples out of seven the coefficient was negative, and more importantly, it was not statistically significant in any sample. However, this analysis observes whether there is an interaction between unemployment and education, it does not compare the educated
unemployed to the educated employed, which is what we actually wanted to do. The results still support each other; the educated unemployed do not appear to be active protesters no matter whether they are compared to the educated employed or whether this is studied by means of an interaction term.
I also tried using an interaction term for age and unemployment status in hypothesis 4. There we expect the coefficient of the interaction term to be negative. Adding this interaction term brought hardly anything new, but confirmed what was seen when comparing results from hypotheses 3 and 4. There were two samples where the interaction term was statistically significant, these were the samples of all countries and the sample of Tunisia. In both these cases the coefficient was negative. And, not surprisingly, these were the two samples where the unemployed were not statistically significantly more active in protesting than the
employed in hypothesis 3 but the unemployed youth was more active in protesting than the employed youth in hypothesis 4.
I made an additional model where I had 25 years as a limit for youth in hypothesis 4 instead of 35 years. This did not change much. In general all the results were statistically less significant compared to having 35 years as a limit. This can be explained with the sample sizes, which were naturally reduced considerably when studying only respondents 25 years old and younger. Other than that the results were concordant with those presented in table 7.