In each of the following sections, the analytical procedure is divided into two parts.
First, the descriptive analysis provides a comparison across countries and genders. Then, the linear regression is run to compare the association between variables, across countries and genders.
Country and gender comparisons. In the descriptive analysis section, I compare variables across countries and genders. When deemed informative I also provide country comparison by gender, or gender comparison by country. For country comparison, I use a simple one-way ANOVA with country as a factor, followed by a post-hoc Scheffe-test at the 0.05 level, to determine country subsets. Country’s mean-values and standard deviations are presented according to the Scheffe-test subsets. The ANOVA results are qualified with the F-value and the effect size (η2). When the ANOVA is not significant, only p-value is provided. For gender comparison, I use a t-test for equality of means, with unequal variance assumed. Each gender’s mean-values and standard deviations are provided. The gender difference effect size is qualified by Cohen’s d. When the t-test is not significant, only p-value is provided. Categorical variables are compared using chi-square, and Cramer’s V provides an estimate of the strength.
Mediation. The mediation provides a causal pathway to explain how social norms impacts dating “via pragmatism.” The moderated mediation (next section) explains for whom the mediation is significant. Moderated mediation is used to test the main hypotheses H1 to H5, with the PROCESS software (Hayes, 2012; Hayes & Preacher, 2014; Preacher, Rucker, &
Hayes, 2007). The first set of hypotheses (H1 and H2) predict that social norms, via pragmatism hinders the dating outcome and increase difficulties in dating for Northeast Asians (Figure 4.1).
Figure 4.1: Indirect (via pragmatism) and direct effects of social norms on dating outcome and difficulties (X is the predictor, M is the mediator and Yis are the tested outcomes)
49
The model (Figure 4.1) provides a mechanism to explain how social norms have an indirect effect on dating, through pragmatism. The first stage (XM) tests if pragmatism stems from social norms. The second stage (M Yi) tests if pragmatism is used in dating (i.e., affect the dating outcome, and causes difficulties in dating). Finally, XYi tests a possible direct effect of social norms on dating. Although causality is implied in the model, the analysis only tests the association between variables30. In the study, X, M and Yi are continuous variables and the various effects between variables (labeled a, b and c) are estimated using the ordinary least squares (OLS) method with two sets of equations:
M = i1 + a x X + eM , (1)
Y = i2 + b x M + c x X + eY (2)
(i1 and i2 are the equations’ intercepts. eM and eY are the error terms)
What is of interest for the current study is the indirect effect of X to Yi (XMYi), quantified as a x b, and the direct effect of X to Yi, quantified as c. The indirect effect (a x b) serves as an estimate of the mechanism of which X affects Yi, through M. And the direct effect (c) serves as an estimate of the effect of X on Yi, independently from M.
A traditional approach to estimate the indirect effect is the causal step approach (Baron &
Kenny, 1986), which focuses on running significance tests for each pathway, and infer logically from the significant individual paths if the mediation occurs. New methods favor an inferential test on the product a x b, using a bootstrap estimate of the confidence interval for a x b (which should not straddle zero to be significant; Hayes, 2009; Hayes & Preacher, 2014)31.
There are several advantages to the bootstrap estimate. First, the indirect effect is directly measured and is not inferred from multiple tests (Hayes, 2009). Then, the bootstrap estimate disregards situations where the individual effects (a and b) differ from zero, but only consider the significance of the indirect effect, independently from individual steps. Although the individual steps are informative, the indirect effect itself (through pragmatism), is the object of interest. Finally, the indirect effect may also exist if there is no direct association between social
30 It would be necessary to test for temporal precedent of social norms compared to pragmatism, in order to validate causality, which is not done is this cross-sectional study.
31 Another estimation of a x b, the Sobel test, requires the indirect effect to have a normal sampling distribution, which is rarely true. The bootstrap method is however robust to deviation from normality (Hayes, 2009).
50
norms and the dating outcome, as the latter may involve other pathways, sometimes opposed in sign (for a discussion see Hayes, 2009).
Moderated mediation32 (Figure 4.2). The test of moderated mediation examines for which group in particular the hypothesized indirect effect described above is applicable33.
In the first stage (XM), I hypothesizes that pragmatism in dating stems from social norms for Northeast Asians, but not for French participants. Thus, I predict that the effect of social norms on pragmatism differs by country. This would be supported by a significant interaction (a x n1) between the social norms and the moderator “country”. To run this test, I compare French and Taiwanese (France = 0, Taiwan = 1), and then I compare French and Japanese (France = 0, Japan = 1).
Figure 4.2: Conditional indirect (via pragmatism) and direct effects of social norms on dating outcome and difficulties (a, b, c, ni, and gi are unstandardized regression coefficients)
Second, I hypothesize a possible gender difference in the association between social norms and pragmatism. A significant interaction (a x g1) would support this gender difference. Gender is coded as women = 0 and man = 1. In addition, there could be differences among men and women within a specific country, therefore the interaction a x n1 x g1 is also considered. The
32 A model where only the first stage (X M) is moderated, is typically called “mediated moderation.” Since both stages are moderated “moderated mediation” is preferred here (Edwards & Lambert, 2007; Preacher et al., 2007).
33 Structural Equation Modeling (SEM) could have been used to test the indirect effect for each group separately, but tests conducted group-by-group have lower statistical power, and testing indirect effects in each sub-group separately does not indicate if there is a difference between groups, thus the current method was preferred.
51
effect of social norms, country, and gender on pragmatism are estimated using the OLS method with equation (3).
M = i1 + a x X + n1 x X + g1 x X (3) + a x n1 x X + a x g1 x X
+ a x n1 x g1 x X + n1 x g1 x X + eM
In the second stage (MYi), I predicted that the effect of pragmatism on the dating outcome occurs only among Northeast Asians because they (but not the French) use pragmatism in dating. This is tested by a significant interaction (b x n2) between pragmatism and country. I also hypothesized a possible gender difference in the usage of pragmatism. A significant interaction (b x g2) would support this gender difference. And in each country separately, there may a gender difference in the usage of pragmatism, therefore the interaction b x n2 x g2 is also considered. Finally, I do not predict culture or gender differences on the direct effect of social norms on dating outcome (XYi). Therefore moderation is not included here. The indirect effect of pragmatism, country, and gender on dating outcome (and difficulties) are estimated using the ordinary least squares (OLS) method with the equation (4).
Y = i2 + c x X + b x M + n2 x M + g2 x M (4) + b x n2 x M + b x g2 x M
+ b x n2 x g2 x M + n2 x g2 x M + eY
To test the conditional indirect effect for each of the moderator value country x gender, the PROCESS software provides a bootstrap estimate of a x b. This final test is independent on whether a or b are individually significant, and on whether there is a country or a gender difference in the two stages. A similar model is used to examine hypothesis H3 (Figure 6.2), H4 (Figure 7.1) and H5 (Figure 7.2).
For each set of analysis, I report the unstandardized regression coefficients for each predictor and each interaction. I also report the model’s R2 which indicates the variance explained by the model, and the model’s F-value, which indicates the strength of the predictors used.
The OLS method assumes variables are normally distributed and homoscedasticitic, and residual errors are uncorrelated to each other and to other variables (Edwards & Lambert, 2007;
Hayes, 2009). First, the bootstrap estimate is robust to deviation from normality (Edwards &
Lambert, 2007) and it does not require error terms to be uncorrelated, as estimates are not
52
calculated from variance, but are estimated and validated with bootstrapping (Hayes, 2009).
The assumption is important for the estimation of the product term for indirect effects, as the distribution of product is non-normal, even if each product term is normally distributed (Edwards & Lambert, 2007). Finally the PROCESS software used in the tests provides a heteroscedasticity-consistent standard errors’ calculation module (labelled “HC3”), that reduces bias in situations of heteroscedasticity (Hayes, 2012), and was used to confirm results (results differed in one case: see Table A.17). In addition, the use of moderators in mediation analysis is conditional to the absence of measurement error on mediators, and to the fact the dependent variable does not cause the mediator (Baron & Kenny, 1986), which are both verified here.
53
CHAPTER FIVE: DATING EXPERIENCE, DATING PYRAMIDS AND DATING DIFFICULTIES
In this section, in line with research questions RQ1, I examine if Northeast Asians would date less than the French, make less progress on the dating steps, and perceive more difficulties in the process of finding a partner.