Female political leaders, the view of the world. <English translation of the title, maximum two lines incl. subtitle>

Transcription

1 Kandidatuppsats Statistiska institutionen Bachelor thesis, Department of Statistics Nr 204:2 Female political leaders, the view of the world <English translation of the title, maximum two lines incl. subtitle> August Andersen Självständigt arbete 5 högskolepoäng inom Statistik III, HT204 Handledare: Dan Hedlin

2 Contents Introduction Background Method Delimitations Model Missing values World Values Survey History Sample design Cultural map The WVS cultural map of the world Variables Description of variables Analysis Missing values Missing values per survey question by country Characteristics Characteristics of view on female opinion Characteristics of view on homosexuality Characteristics of view on abortion Characteristics of level of education Characteristics of perceived social class Characteristics of immigrant Model fit Pattern of missing values Residual Likelihood ratio Estimates Fixed effect estimates Fixed effect estimates using multiple imputation Random effect estimates Country specific random effects estimates using imputation Country specific random effects estimates using delete records Diagram of country specific random effect and common intercept Discussion Bibliography Appendix Multiple Imputation Output of parameter estimates using multiple imputation Output of parameter estimates using delete records... 33

3 Introduction. Background Gender equality is one of the central questions of the 2 st century. With females gaining the right to vote in Sweden first in 2. Yet after almost a century, has a female head of state not been elected in Sweden - all while neighboring countries such as Finland, Iceland and Norway have. With the first female head of state in Europe being Margaret Thatcher elected in 7 has the following female head of state in Europe been few and far between relative to the number of male leaders. One may wonder why, why if the population proportion is equal between the sexes would this not be reflected in the political positions of power? Gender inequality, according to Dorius & Firebaugh (200) can be described as when either men or women enjoy a disproportionally large share of some valued good such as political power [ ]. If gender inequality can be expressed as political power does that mean that it can also be measured in the form of capability as political leadership? We know for a fact that the amount of political leaders are disproportional among the genders, does that imply a distrust for female political leaders? An explanation to this irregularity and in turn the disproportionality of female leaders is the congruity theory of prejudice toward female leaders described by Eagly, H. & Karau, J.(2002) The theory describe two types of prejudice toward female leaders women are perceived as less favorable leaders than men and that women underperform in roles that require leadership. These prejudices has the consequence of decreasing the attitude toward female leaders which in turn increases the difficulty for women to become and succeed as leaders. This prejudice exist in sociological theory but measuring what increase or decreases the view whereby influence the general view on female leaders remain ambiguous. If one were to measure this effect would one have to take in to account numerous different environmental and cultural effects unique to the individual. Identifying and estimating each individual effect and factor would in itself be another problem with another frame. With lacking theory can the prejudice of female leaders be hard to estimate unless some of the variability has a known cause. World Value Survey have presented findings using factor analysis that support a frame were the worlds countries can be categorized in to cultural cluster in which countries share similar and some, less similar views on different human traits. This gives us an opportunity of another approach were we can estimate the effect when many potential factors are at play and hence avoid the problematic nature of finding each individual effect when there is no measurable factors backed by rigid theory. We can now instead describe the differences in terms of country and by doing so, use country specific characteristics to describe the magnitude of the differences on the view on female political leadership. This estimation is done using logistic regression, due to the dichotomous nature of the response variable were either the respondent agree or disagree on the statement that females make worse political leaders. This gives us a response in an interpretable output were the probabilities interpretation is interchangeable with country proportions. The final question then become how one should handle the complex nature of a random variable which only

4 varies between countries and which tries to encompass all of the hidden effects of prejudice against female leaders. The random coefficient is then the answer, the random coefficient which take in to account the multilevel nature of country wide variation. With missing values being an issue common to survey items of sensitive nature resulting in biased estimates - Shall an appropriate imputation method be used in an attempt to acquire non-biased estimates. The question of the thesis have settled on the estimation of one effect unique to each of the countries included in the sample. What is the opinion of female political leadership in countries of different cultural regions? Are these countries significantly different according to their cultural belongings?.2 Method To analyze and answer this research question will survey data collected by World Value Survey be used this data will then be cleaned to only contain relevant research variables, based on the cultural map presented by World Value survey findings. The data will then be subject to multiple imputation as to decrease the level of missing values in an attempt to increase the accuracy of the estimates. Descriptive statistics will be computed to describe each population of the countries under analysis. The hierarchical logistic regression model will compute estimates, this model will be evaluated using test-statistic relevant to the model used. Lastly will inference be drawn based on the estimate of the random coefficient, country-effect..3 Delimitations The thesis will be limited to studying six countries while retaining the data of the other countries in each respective data set. Furthermore will only one survey wave be used for estimation, ranging from the years 200 to 204. The choice of variables and countries is based on the cultural map, supported by World value survey findings - this is described further in detail in the chapter concerning the data from World value Survey. The six countries under study will be Sweden, United States, South Korea, Belarus, Uruguay, Spain, Hong Kong and Iraq. With the selection of countries being done as each and one of them is to be representative of their respective cultural cluster of the cultural map. It should be noted that this selection may not be considered fully representative of the clusters countries may vary greatly within each cultural cluster. 3

5 2 Model When analyzing the response variable, opinion on female political leaders as the response variable is dichotomous since y ij can take on the value of either one or zero where one represents a positive view on female political leadership and zero a negative view. This inclines that a model which handles the binomial and in turn, nonlinear link between the dependent variable and explanatory variable be used. The logistic regression model solves our first problem but an extension is needed, this is extension is the multilevel modelling specifically using a random coefficient with the intent of capturing the unexplained variation. Equation ( ) displays the model in its natural state, the left side is devised of the natural logarithm of the odds and the right hand side is devised of the explanatory variables. (Sharma 6) This is the basic logistic regression model which also serves as our level one model. Were p ij represents the probability that a respondent disagrees on the statement that female make wore political leaders and β 0j represents the random effect which varies over countries, in this thesis will country be the random intercept effect - as individuals are nested within countries. While the remaining explanatory coefficients β to β h represent the fixed effects which do not vary cross country. This means that an increase on the view is abortion justifiable will have the same impact no matter the country the individual belongs to - the coefficients does not vary cross country. The country specific random effect has its own random clustering effect, this error term δ 0j is assumed to be N~(0, σ 2 ) the level one model can then be expressed as equation ( ) see Leeuw, J. (2008) and Goldstein () for further details. p ij log ( ) = β p 0j + β c x ij ij h c= ( ) β 0j = β 0 + β 2 x j + δ 0j ( 2 ) Where i =, 2,, k. Where j =, 2,, n. Where c =, 2,, h. Equation ( 2 ) is the level two model of a hierarchical model were only one explanatory variable is allowed to vary cross level, the random intercept. As level one variance is fixed, written explicitly in equation one, this means that further addition of level one covariates will lead to larger random effect variances in absolute values shown by Snijders and Bosker (). If the purpose is then to estimate a general difference among countries in the form a random effect would over-inclusion of variables be contra productive. Finding the individual effects would require and exploratory analysis with clear theory on the topic of increases and decreases the trust in female political leadership. Where subscript i, j and c denotes the individual respondent identifier, the country in which the respondent belong to (referred to as the level two) and c the fixed effect explanatory variable e.g. view on abortion respectively. The extension of a hierarchical structure does not directly affect interpretation of the coefficients. The kernel function which give the relationship between the probability and the

6 covariates will be that of the natural logarithm. This means that the curve of the probability is constrained to values between zero and one. This kernel also creates the relationship of equation ( 3 ) presented in chapter 4.4, which eases the interpretation of the coefficients. It should be noted that the interpretation of the independent coefficients should be done with the respect to the log of odds and not to the probability p ij. (Sharma 6). An increase in an ordinal categorical variable should only take in to account the increase of its own estimate not in to addition of its lower ranked estimates. In this thesis will these estimates be calculated using the SAS.4 software and specifically the Glimmix procedure. For more details refer to the appendix chapter or the SAS guide. 2. Missing values One of the more common and problematic issues with survey data is the case of missing values. In the data used in this thesis can missing values take on the expression of different codes (see appendix for WVS codebook). The responses coded in the survey as; not asked, inappropriate response, no answer and don t know were all considered as missing values. It should be noted that when the data extraction process occurred from World Value survey had the data already been cleaned for missing values and the response don t know had already been coded as missing value. This might affect the accuracy of the imputation as the true frequency of don t know remains unknown. If the frequency would have been low would it been viewed as missing values, this can inflate the frequency of missing values. Non-response can have many mechanisms as described by Rubin (87) and may cause biases in the estimates if not treated properly. As WVS employ a face to face interviewing approach in their data collection process and due to the sensitive nature of the questions asked in this thesis is there to be expected that a social desirability bias occurs expressed in missing values see Biemer & Lyberg (pp 2, 2003 ) for an explanation of these mechanisms. The underlying causes of the missing data is described by Rubin (87) as three types of patterns, Missing at random (MAR), missing completely at random (MCAR) and not missing completely at random (NMAR). MAR loosely defined as the conditional probability of response not being related to the variables under study, given the observed data does the missingness mechanism not depend on the unobserved data. MCAR described as the probability of response being independent of the response explanatory variables. A common approach is deleting the records with missing observations, this is something that can cause biases in the estimates and great loss in information if the missingness is related to the characteristics of the population. Another popular way of handling the missingness, given MAR is multiple imputation. Used by Nasser (204) and described in depth by Rubin (87). Multiple imputation can be decomposed in to three steps. Firstly is m complete data sets created whereby the missing values are filled in m times. Secondly are the m complete data sets analyzed using the procedure of choice, in our case logistic regression and lastly are one of the imputed data sets parameter estimates analyzed. Following Nasser (204), is the method of fully conditional specifications (FCS) used. FCS is a semi-parametric model that imputes the data on a variable-by-variable basis. These imputed 5

7 values are based on a series of conditional models, starting from a single imputation are further imputation done each iteration based on the conditional imputation model. All of the explanatory variables and the dependent variable will be subject to this imputation. 3 World Values Survey 3. History World values survey is a network of social scientist that study human values and their sequential impact on life. With the self-proclaimed intent of helping policy makers understand changes in attitudes and beliefs making it possible for better balanced decisions to be made. Starting with their first survey in 8, six waves of surveys have followed spanning over 00 countries with 0% of the human population being covered - rendering World value surveys the largest noncommercial and cross-national investigation of human beliefs ever executed. The main questionnaire is developed in English and since the survey is conducted in numerous different countries in turn mean that the survey is translated to various languages. WVS employ a pre-testing to check for translation problems, even going so far to remove questions which have problematic translations. The questionnaire is according to WVS under continuous development, were the less useful questions are removed to create space for more informative research items while trying to maintain the questions relevant for the population. 3.2 Sample design WVS employ a complex survey design that can vary extensively depending on the country being surveyed. The samples are drawn from the entire population of 8 years and older. In some countries e.g. Sweden are national registers used to draw a representative sample, using a simple random sampling. Other countries e.g. China use more complex sample designs such as stratified random sampling. When stratified sampling is used are firstly primary sampling units (PSU) selected, these sampling units are probabilistic and based society statistical regions such as districts and election units. Furthermore is the rate of urbanization taken into account when needed in the form of clusters found within each of the primary sampling units. It should be noted that when stratified sampling is used is the minimum amount of primary sampling units 30. It should be noted that if unbiased estimates are to be computed must one take in to account the complex sampling design used for each country. World value survey do not provide any cluster or primary sampling units variables in their data, this will bias the estimates of the country specific random intercept for the countries which used complex sampling design. This bias of variances entails that a full scale comparison of all the countries contained in the survey wave of interest would require that each country sample design is taken in to account. One wave of data was used for this thesis. Wave 6, which were surveyed through the years involving 67 countries with a sample of Six countries were selected for an in-depth analysis Sweden, United States, South Korea, Belarus, Uruguay, Spain, Hong Kong and Iraq. With the countries specific survey year being 20, 20, 200, 20, 20, and 202 respectively.

8 3.3 Cultural map When deciding on which variables to be included in the original model is the underlying theme the cultural map, based on World values survey findings explicitly the findings report by Inglehart & Welzel (20). The cultural map differentiate the world in two cultural dimensions. Traditional values contra secular-rational values which encompass the importance of religion in society. Survival values contra self-expression values which encompass the transition from a pre-industrialized country to a post-industrialized country, the state of becoming a knowledge society. Typical traditional societies have a greater emphasize on of traditional family values, deference to authority, absolute standards, high levels of national pride and nationalistic outlook while secular-rational societies have opposite preferences. Societies with high values of self-expression are countries were survival is taken for granted for a large proportion of the population. This have led to an emphasis on subjective well-being and freedom of self-expression The WVS cultural map of the world Figure 3.: Inglehart, R., & Welzel, C. (20). The WVS cultural map of the world Each country is placed accordingly to their proximity in values and not according to their geographical location, this creates cultural neighbors which in turn create cultural clusters. 7

9 These two dimensions according to Ronald Inglehart & Chris Welzel explain more than 70 percent of the cross-national variance. This map is constructed by analyzing the aggregated national-level data from WVS using an exploratory factor analysis with the intent of finding large and coherent cross-cultural differences. Using the survey items from their data were two dimensions (factors) constructed which includes all the countries on the spectrum of Traditional and survival values. See Inglehart & Welzel (2005) for an in-depth description of the creation and methodology of the cultural map. 3.4 Variables Eight variables were included from the data collection process. These variables include the dependent variable view on political female leaders, the level 2 identifier variable of country and the six explanatory variables. The variables included below try to encompass these dimensions in one way or another while the choice of countries was done by taking one country from each of the cultural clusters Description of variables Variable Dependent variable On the whole, men make better political leaders than women do Independent Variables Sex Descriptions Positive/Negative Male/Female Is homosexuality justifiable Never justifiable, 2, 3, 4, 5, 6, 7, 8, Always justifiable Is abortion justifiable Never justifiable, 2, 3, 4, 5, 6, 7, 8, Always justifiable Highest attained level of education No formal education, Incomplete primary school, Complete primary school, Incomplete secondary school, Complete secondary school, Some university-level education, University-level education with degree. People sometimes describe themselves as belonging to the working class, the middle class, or the upper or lower class. Would you describe yourself as belonging to the: Upper class, Upper middle class, Lower middle class, Working class, Lower class

10 On this list are various groups of people. Could you please mention any that you would not like to have as neighbors?(immigrants) Mentioned, not mentioned Table 3.2: Description of variables. What follows is an explanation of the survey questions appertained to the variables used in this thesis and any transformation done to the variable in question. The dependent variable of this research, opinion of female political leaders. The original survey question written in English as: On the whole, men make better political leaders than women do. The variable of interest was reduced from its four possible answers of; Agree strongly, Agree, Disagree strongly, Disagree. To a more convenient Agree or disagree. The new answer of Agree encompasses Agree and Agree strongly, while disagree encompasses Disagree and Disagree strongly. This renders the variable dichotomous. The respondent s gender were kept as explanatory variables. Gender was reported as binary, were represents male and 0 female. The variable of homosexuality, original survey question written in English as: Tell me if abortion/homosexuality can always be justified, never or something in between. Constrained to values of one to 0, 0 meaning always justifiable and one never justifiable. The variable of abortion, original survey question written in English as: Tell me if abortion can always be justified, never or something in between. Constrained to values of one to 0, 0 meaning always justifiable and one never justifiable. The variable of the respondent s education level. Were a higher value equals a higher level of education. Constrained to values of one to. The variable of perceived social class, original survey question written in English as: People sometimes describe themselves as belonging to the working class, the middle class, or the upper or lower class. Would you describe yourself as belonging to the. Constrained to values of one to five, one being upper class and five lower class. The final explanatory variable is immigrants and foreign workers mentioned as unwanted neighbors. Original survey question written in English as: on this list are various groups of people. Could you please mention any that you would not like to have as neighbors? Were one represents mentioned and zero not mentioned. For a detailed description concerning the phrasing and coding of response see appendix for the codebook used by World Value Survey.

11 4 Analysis The analysis will be comprised of two parts, the first part is a descriptive part and the second will contain model diagnostics and the model parameter estimates. The descriptive part will be comprised of descriptive statistics for the missing values by each countries under study, followed by proportional characteristics of the explanatory and response variables. In the second part will the model fit be evaluated followed by interpretation and evaluation of the parameter estimates computed by the logistic regression model. 4. Missing values It can be seen that the percent of missing values of each survey item per country seem to keep a steady amount of variation. With Sweden having a maximum of 6.4 percent on the item perceived social class and with a minimum of.8 percent at the item of education. United states have a lower amount of missing values with the maximum being 3.3 percent on the item view on homosexuality and with a minimum of. percent on the research item female leader. South Korea have lower amount missing values than both United States and Sweden with the maximum being 2 percent belonging to the research item of female leader with a minimum belonging to view on abortion of 0.4 percent. Belarus have a similar level of missing values to that of South Korea with a maximum of percent belonging to the item of view on homosexuality and a minimum belonging to view on abortion of half a percent. The following country is Uruguay with the highest volatility of missing values of all the countries under study. With a maximum of percent belonging to the research item of female leader and with a minimum of 0. percent belonging to level of education. Spain has the highest proportion of missing values due to the total exclusion of response on the item of chief earner, besides earner having a maximum of 5.5 percent belonging to view on homosexuality and with a minimum of 0.5 percent belonging to level of education. 4.. Missing values per survey question by country Variable Sweden Percentage of frequency % United South Belarus Uruguay Spain Hong Iraq states Korea Kong Female leader Gender Homosexuality Abortion Earner Education Social class Immigrant, Neighbors Table 4.. Missing values. Missing values of all the countries under study per variable of interest. Hong Kong have the lowest proportion of missing values cross survey items with a maximum of 0.6 percent belonging to the research item of female leader and with a minimum of 0. percent at three separate survey items. Finally there is Iraq with a maximum of 4 percent

12 belonging to view on homosexuality and with a minimum of 0.2 percent belonging to level of education. It should be noted that the table above only present s item non-response by country, there is no unit non-response in the data used. Continuing on by missingness per survey item, beginning with the dependent variable Female leader can it be seen that the amount of missing values have a quite low frequencies with only one country, Uruguay going above five percent with percent. Gender has no missing values, since gender is only a confounding variable and is not used in the imputation model will this not affect the estimates. The survey item of view on homosexuality has a larger amount of missing values compared to the response variable. With a maximum of 8.4 percent for Uruguay and minimum of Hong Kong and South Korea with.4 and.7 percent respectively. The survey question concerning abortion pertains the same pattern as view on homosexuality with the highest proportion of missing values belonging to Uruguay at 6.3 percent and the lowest missing value country wise belonging to South Korea and Hong Kong at 0.4 and 0. percent respectively. Continuing on to the variable of chief wage earner, can there be seen that the same pattern of previous variables continuous except the addition of the two extreme values belonging to Spain and Iraq. Spain having a complete no-response on the survey item while Iraq has a complete response. As the data had been previously been cleaned is it impossible to say what is responsible to the missing values appertained to Spain. Level of education is also one of the standard questions in survey design and retains a low value of missing values in the survey. The item of perceived social class have one of the lowest proportions of missing values cross country, of the variables included in the study, with the highest value appertained to Sweden with 6.4 percent and the lowest belonging to Hong Kong at 0. percent. The item of immigrants as unwanted neighbors have a missing value of zero cross countries. This is due to the survey design were the response option of immigrants is one of the alternatives, if the respondent did not choose to mention Immigrants/foreign workers was this noted as a response.

13 4.2 Characteristics 4.2. Characteristics of view on female opinion Characteristics Agree/Agree strongly that men make better political leaders than women do Disagree/Disagree strongly that men make better political leaders than women do Percentage of frequency % South Sweden United Iraq Belarus Uruguay Spain Hong Korea states Kong Table 4.2. Characteristics by country. Percentage of frequency by survey item view on female political leaders. South Korea and Hong Kong have an almost equal proportion of agree and disagree of the dependent variable with 43./32.5 percent of whom agree and 53./66. percent that disagree. With a difference of almost 0 and 30 percent, not taking into account the proportion of missing values, both of which belonging to the Confucian cultural cluster. Sweden, United States, Uruguay and Spain have a large response proportion belonging to disagree with 85.4/7.4/82.6/8 percent while only 0.7/.4/8.3/5.3 percent agree on the survey statement. There seem to exist a similarity in view between catholic Europe, protestant Europe, Latin America and North America countries all of which can be classified as cultural neighbors. Iraq have the largest proportion of respondents agreeing on the statement with 7.2 percent, while.2 percent disagree. This is the largest proportion of disagreement, followed by Belarus with a 0 percent difference in proportion, with Iraq belonging to the Islamic cluster and Belarus the orthodox. Both of which have a high score on the traditional values in the cultural map Characteristics of view on homosexuality Characteristics Belarus Homosexuality is never justifiable Percentage of frequency % Hong Kong Iraq South Korea Spain Sweden United states Uruguay sometimes

14 Homosexuality is Always justifiable 2.4 Table 4.3. Characteristics by country. Percentage of frequency by survey item view on homosexuality. On the survey item if homosexuality is justifiable can a pattern resembling that of female leaders be found, with Sweden having the highest proportion, 5.2 percent of respondents saying that homosexuality is always justifiable. Followed once again by its cultural neighbors Spain, United States and Uruguay with 32.4, 2.5 and 27.8 respectively. There is however a larger spread within the scale between the countries, with United States having the highest proportion of the proportion never justifiable at 24 percent followed by Sweden, Uruguay and Spain at 7.4, 7.6 and 7.7 percent respectively. The cultural clusters seem to persist for the other countries as well. With the proportion of respondents saying that homosexuality is never justifiable in South Korea and Hong Kong being 42.2 and 28.4 percent respectively. Although there is a small proportion that view homosexuality as always justifiable at 5 and 2.4 percent, there is however a larger proportion that answer 5 5 which in this ordinal measurement could be viewed as sometimes. Iraq and Belarus remain close to each other in proportions with Belarus having a proportion of 57.2 percent saying that homosexuality is never justifiable and only 0.5 saying sometimes. Iraq having the same response pattern with 70 percent saying that homosexuality is never justifiable and only 3.5 percent saying sometimes Characteristics of view on abortion Characteristics Belarus Abortion is never justifiable Percentage of frequency % Hong Iraq South Spain Sweden United Uruguay Kong Korea States sometimes Abortion is always justifiable Table 4.4. Characteristics by country. Percentage of frequency by survey item view on abortion. The survey item abortion share a similar pattern as of that of view on Homosexuality but with a distribution more spread between the responses alternatives. Starting with Sweden, do 43.5 percent of the respondents view abortion as always justifiable with 4.7 percent saying the opposite and two clusters at sometimes and 8 of 0. percent. Spain share a similar 3

15 distribution to Sweden with sometimes having 20 percent while the two extremes only having 5 percent each. United States once again follow and share a similar pattern but with a larger proportion allocated at never justifiable with 22.4 percent and sometimes having a proportion of 23.2 percent. The earlier similarity between Iraq and Belarus seem to have diminished somewhat with Uruguay, Belarus, Hong Kong and South Korea now instead sharing a similar distribution. With never justifiable being where the majority of the respondents are allocated with 30.4, 38, 36.4 and 42 percent. This meanwhile having many respondents responding sometimes at 23.5, 8.8, 7.4 and 4.3 percent respectively. While Iraq share a similar patter but with the mass of the respondents answering never at 47. percent or at any of the lower alternatives. The cultural neighbors do not follow as closely as before, something that indicates that the view on abortion is more strongly affected by another of the cultural dimension. Most likely due to the religious link to abortion Characteristics of level of education Characteristics Belarus No formal education Incomplete primary school Complete primary school Incomplete secondary school Higher level education. Percentage of frequency % Hong Iraq South Spain Sweden United Uruguay Kong Korea states ,3 83,6 4,3 78, ,3 7,4 43,6 Table 4.5. Characteristics by country. Percentage of frequency by survey item level of education. Please note that higher level education is a summarization of many different higher level educations, refer to table 3.3. for further details. Belarus have the largest proportion of higher education at 8.3 percent with only 0.5 percent having no formal education. Hong Kong and South Korea once again share a similar pattern of distribution with 83.6 and 78.7 having a higher education. Iraq, Spain and Uruguay share the similarity of having a similar proportion of higher education at 4.3, 48 and 43.6 percent respectively with differentiated proportions at primary school incomplete and complete. United States and Sweden both have a large density at higher education with a small proportion at lower levels Characteristics of perceived social class Characteristics Belarus Hong Kong Percentage of frequency % Iraq South Korea Spain Sweden United States Uruguay Lower class Working class

16 Lower middle class Upper middle class Upper class Table 4.6. Characteristics by country. Percentage of frequency by survey item perceived social class. The survey item social class show no apparent pattern between the cultural neighbors. The proportion of upper class is very similar cross country with the maximum being 3.3 percent belonging to Iraq and the minimum 0.8 percent of Sweden. Alike the proportion of upper class is lower class similar cross country with three countries going above 0 percent. The majority of respondents being affiliated with the middle answers of working class, lower middle class and upper middle class. The proportions have the look of a Gaussian distribution albeit skewed Characteristics of immigrant Percentage of frequency % Immigrants/foreign Belarus Hong Iraq South Spain Sweden United Uruguay workers Kong Korea states Not mentioned Mentioned Table 4.7. Characteristics by country. Percentage of frequency by survey item immigrants/foreign workers as unwanted neighbors. Finally is there the survey item of unwanted neighbors. Once again can the similarities between the cultural neighbors be observed with South Korea and Hong Kong having a proportion of 2. and 44. percent respectively. While Sweden, Uruguay and Spain have proportions of a lesser magnitude at 3.4,.7 and 7.5 percent respectively. United States differentiate itself from the previous similarities with protestant and catholic Europe while Uruguay is closer in proportion. Belarus and Iraq again show their cultural likeness with the difference in proportion being 0 percent at 33.4 and 3 percent respectively. 5

17 4.3 Model fit A total of 2336 observations were imputed in the dataset, leaving a total of observations requiring no imputation. The table below present the pattern of missing values, sorted by the dependent variable view on female political leaders. It can be seen that the pattern is monotone when sorted by the response variable, on the other hand is the same pattern displayed if sorted by any of the other explanatory variables, this indicates MAR to some degree. As the missingess appears to be equally present to all of the involved covariates. Please refer to the appendix for the complete table of the pattern Pattern of missing values Missing Data Patterns Group leader Homo abort edu Socio Immigrant Freq Percent X X X X X X X X X X X X X X X. X X X X. X X X X X.. X X X. X X X X X. X. X X X.. X X X X... X X. X X X X Table 4.8. Missing Data Patterns. Incomplete table of missing values. As the logistic model ( ) was tested and run, two of the original variables were excluded from the final model; perceived social class and view on abortion. They were dropped as none of these gave significant p-values well above the arbitrary 5 percent level. The two variables were dropped separately to ensure none of the variables created a conflict between each other. The table below presents the residual log likelihood of the three models. As the likelihood ratio is a relative number should a decrease be interpreted as such, the change is relative to the original model likelihood ratio. There is a slight decrease which indicate a better fit of the model and as unnecessary explanatory variable inclusion may increase the confidence interval and hence decrease the accuracy of the random effect Residual Likelihood ratio Model test Excluding abortion and social class Excluding abortion Including abortion and social class -2 Res Log Pseudo- Likelihood Table 4.. Residual likelihood ratio. Ratio for the three tested models with the model to the left being the final. This leaves us with a final model including the country specific random effect with the fixed effects of homosexuality, level of education, immigrants as unwanted neighbors and gender. The table above presents the three models tested. It should also be noted that the convergence criterion was satisfied after three iterations, refer to appendix for full output.

18 4.4 Estimates The two following sections consist of two parts, first will the estimates for the fixed effect be presented followed by the country specific random effect using complete record method and multiple imputation lastly will the probability each country be calculated based on the previously presented parameter estimates. The two following tables below present the computed estimates of the country specific random effect and the solutions for the fixed effect of the final model. Please refer to the appendix for computed estimates pre-imputation and for the complete output of country specific random effects. We shall begin with looking at the estimates of the fixed effect. The final model as stated earlier excluded both the parameter of perceived social class and view on abortion. Leaving us with four explanatory variables excluding the random effect of country. It should also be noted that the estimates for the random and fixed effects were computed under the assumption that the outcome variable will be zero. That the probability that the respondent disagree on the statement that females on a whole make worse political leaders. Furthermore is the interpretation of the fixed effects not as straightforward when a random effect is introduced, the fixed effects presented below should be interpreted as marginal increases in response variable. Note that both of the tables include -probability this is the marginal probability that the respondent disagree on the dependent variable. The estimate has undergone the conversion of equation ( 3 ) presented below to ease the interpretation of the coefficients. p = +e (β 0j + h c= β c) ( 3 ) 4.5 Fixed effect estimates The common intercept related to all of the countries is significant at a 5 percent level and increases the positive view on female leadership by 60 percent, this can be interpreted as a base level for the model and its estimates. This is can be regarded as a quite high probability but it should be remembered that this probability is only relative to the other fixed effect estimates and the random coefficients. View on homosexuality has a marginal increase of 2 percent when never justifiable, this parameter increase with 2 percent as the view changes to always justifiable, which can be regarded as a relative small increase. Education share a similar increase with a starting values of 34 percent with no formal education and increases by 5 percent with a final value when a university degree is present at 4 percent. This can be interpreted as a higher level of education leads to a lower mistrust in female political leaders. The marginal increase is 56 percent when immigrants and foreign workers are not mentioned as unwanted neighbors. It should however be remembered from the previous descriptive chapter that the proportion of mentions were quite small cross country and this might increase the effect. 7

19 Finally is there the parameter of gender, if the respondent is female increases the positive view on female political leadership marginally by 56 percent. It should be noted that the table for the fixed effect estimates using complete record cleaning (CC) are not presented. The estimates for the individual fixed effects for when CC was used were each significant at a 5 percent level except for the intercept and education class 8. The same scenario as for when imputation was used, the two estimates seem to have retained their insignificance suggesting that missing values were not the cause for the insignificance. With the insignificant intercept having a decrease in its standard error from 0.36 to 0,27 and education class 8 from to respectively Fixed effect estimates using multiple imputation Solution for Fixed Effects Effect Class Estimate -Probability Standard Error Pr > t Intercept 0,438 0,606 0,27 0,003 Homosexuality -0,86 0,27 0,0374 <, ,73 0,3 0,04472 <, ,727 0,325 0,04704 <, ,7057 0,330 0,05024 <, ,4565 0,387 0,047 <, ,5573 0,364 0,0452 <, ,552 0,365 0,05353 <, ,45 0,388 0,05356 <,000-0,3572 0,4 0,0634 <,000 Homosexuality(ref) 0 0 0,, Education -0,678 0,336 0,0434 <, ,636 0,346 0,0408 <, ,4562 0,387 0,037 <, ,3363 0,46 0,0356 <, ,673 0,458 0,02665 <, ,27 0,432 0,03403 <, ,222 0,444 0,02643 <, ,0382 0,40 0, ,2645 Education(ref) 0 0,, Immigrant 0 0,228 0,557 0,0883 <,000 gender 0,565 0,637 0,0577 <,000 Table 4.0. Fixed effect estimates. The fixed effect estimates of the final model, only the significant parameters included.

20 4.6 Random effect estimates 4.6. Country specific random effects estimates using imputation Solution for country specific Random Effect Subject Estimate -probability Std Err Pred Pr > t Belarus -0,6876 0, ,328 <,000 Hong Kong 0,6432 0, ,388 <,000 Iraq -,75 0, ,42 <,000 South Korea 0,255 0, ,348 0, Spain,404 0, ,44 <,000 Sweden,5066 0, ,5 <,000 United States,2 0,8372 0,33 <,000 Uruguay,8506 0,0740 0,543 <,000 Table 4.. Country specific random effect estimates. The estimated random effects using multiple imputation. The probabilities shown in the table above are the transformed log odds with the addition of the common intercept, no fixed effect is included in the calculations. Here we can see the various degrees of view on female political leadership cross country and in turn cross cultures. Orthodox Belarus with a high score in survival values and secular-rational values have a probability of 43 percent. Confucian Hong Kong and South Korea with an even higher placement on the secular-rational values and survival values have a probability of 74 and 65 percent respectively. It should however be noted that South Koreas random intercept is not significant at the arbitrary five percent level. Islamic Iraq with strong traditional and survival values have a probability of 32 percent the lowest probability of all the countries under study. Catholic Spain, protestant Sweden, English speaking United States and Latin American Uruguay share a similar probability with the highest difference between United States and Uruguay being seven percent. All of which have high values of self-expression and secular rational values. The marginal probabilities for each country is different among the cultural clusters, with the cultural neighbors sharing similar probabilities and views on female leadership. The key factors which decide the probabilities have not been identified and further research is needed.

21 4.6.2 Country specific random effects estimates using complete records Solution for country specific Random Effect Subject Estimate Std Err Pred Pr > t Belarus <.000 Hong Kong <.000 South Korea Spain <.000 Sweden <.000 United states <.000 Uruguay <.000 Table 4.2. Country specific random effect estimates. The estimated random effects using complete record. Above are the country specific random effect presented using the complete record method. As neither social class, earner nor abortion was included in the final model Spain was not excluded as an estimate - the variables were instead dropped from the analysis. It can be seen that the complete record method yield larger standard errors for all the random effects with South Korea still being insignificant and yielding an even greater P-value. The computed estimates both for the random and the fixed effect kept a similar magnitude of the standard error implying that countries which were significant before the imputation stayed such with no increase in accuracy. On the other hand countries which had non-significant random effect before imputation appertained their status but with a greater probability for type- II error. This is most likely due the fact that countries which had insignificant estimates used complex survey designs which the statistical model did not take in to account, for total cross country significant estimates must the clusters and primary sampling units appropriate for each country be used if significant estimates are to be attained.

22 Uruguay Sweden Netherlands Spain Germany Peru United states Mexico Ecuador Trinida New Zealand Slovenia Austral Colombia Brazil Taiwan Argenti Chile Hong Kong Cyprus Poland Rwanda S,Korea Romania Singapore Japan Thailand China Ukraine Estonia South Africa Zimbabwe India Lebanon Philippines Morocco Russia Kyrgyzstan Armenia Kazakhstan Turkey Azerbai Malaysia Belarus Bahrain Pakistan Tunisia Algeria Libya Nigeria Ghana Iraq Uzbekistan Kuwait Jordan Palestine Yemen Qatar Egypt The table below show the estimated country effect in addition to the common intercept expressed as probabilities to ease interpretation. Although not all of the estimates are significant due to varying complexity of sample design the diagram does display the variety of opinion with values ranging from 0 percent (Uruguay) to 20 percent (Egypt). The table also display the clustering of cultural neighbors according to the view on female political leaders. As not all of the countries had significant random effect estimates should the table be interpreted with this in mind Diagram of country specific random effect and common intercept Country effect specific random effect plus common intercept expressed as probabilities 0, 0,8 0,7 0,6 0,5 0,4 0,3 0,2 0, 0 Diagram 4.. Diagram of the country specific random effect with the common intercept. Each country is colored with its respective cluster. Note: the random effect estimate appertained to countries China, Estonia, Japan, Romania, South Korea, Singapore, South Africa, Thailand, Ukraine and Zimbabwe were each insignificant at a five percent level. The diagram above display the cultural clusters described by the cultural map, were the Islamic, orthodox European and African countries are grouped at lower values at the right tail. While protestant European, Latin American, English speaking and Catholic European countries are grouped at the left hand tail with higher probabilities. Finally are the south Asian and Confucian countries grouped in the middle of these two extremes. The cultural differences seem apparent, very strong and can be differentiated somewhat in the form of view on political leadership. 2

23 5 Discussion We will begin the discussion with the problems with the study as to keep fresh in mind what can and what cannot be concluded from the study. The first problem is that of the sample design. The sample design presented in chapter 3.. has two different approaches depending on if national registers were available or not. This could be the cause for the insignificant estimates of the random effect for various countries, as ignoring the complex survey design of primary sampling units will lead to large confidence intervals and hence non-accurate estimates. An example could be Zimbabwe with no national registry were the survey had to be conducted using different regional sampling units. The choice of initial variables also proved to be less significant than priory assumed, identifying the effects which influence the view on female political leadership would require another research approach, an exploratory one. The structure of this thesis was not that of an exploratory one instead confirmatory, the cultural differences and in turn different views on female leadership were a priori assumed. It would be interesting to extend the research following the literature on gender equality in an attempt to identify and measure these effects, factor analysis could be appropriate for this, as factor analysis has previously been used on the data to identify what affect other views of different countries. The estimated fixed effect which were kept in the final model all proved to be significant, even before imputation had been done, this was contrary to my belief as one would expect the levels of missingness would lead to a greater effect on the significance of the estimates. The thesis used data of secondary type, the collection was executed by World Values Survey. This should not have caused any problems if it were not for the structure of the data itself. When the data was extracted from world value survey was missing values not differentiated from response in the form of don t know, this mean that even though some respondents had answered were this inputs still viewed as missing values. The research variable of View on female political leaders, originally phrased in English as On the whole, men make better political leaders than women do have a quite ambiguous interpretation. Political leadership may differ greatly across countries, based on governmental type. Political leadership may take the form of prime minister, dictator, regional leader and so on. The right of a female becoming a political leader may be apparent in Sweden but it should be noted (which is also apparent in the study) that this is not true for the rest of the world. A world in which some countries does not even allow females to vote. Female political power may be so very low that the proposition of a female political leader may present itself as outrageous. The survey years presented in chapter 3.4 also present a problem of timeliness as the survey years for some of the countries date back to the beginning of the collection period. The opinion of countries may differ greatly over time. With World Value survey acknowledging a paradigm shift in view of the world as of the year 2000, with countries becoming knowledge states were self-expression and secular-rational values having a general increase cross all developed countries. Concluding, yes there are differences among the countries with various strengths of a country wide effect and there seem to exist groupings based on the cultural map presented by World Value Survey.