3.1 Measurement of representativeness: Organizational- or individual-level

There are two different measures of representatives that echo the two aforementioned perspectives taken in interpreting the occurrence of symbolic representation, which could be one of the impact factors on observed symbolic effects. On the one hand, research shows that the mere presence of bureaucratic representatives at the organizational level, or say, the fact that a bureaucracy “looks like” the clients it serves in an aggregate sense, can send meaningful signals, thus evoking positive citizen responses (e.g., Baniamin & Jamil, 2023; Breslin, 2019; Riccucci et al., 2014). Relevant studies measure passive representation, which commonly serves as the independent variable, through the demographic composition of the workforce in public agencies either at the frontline (e.g., Baniamin & Jamil, 2023) or in managerial positions (e.g., Dantas et al., 2022). It is important to note that in these contexts, signals of bureaucratic representation are presented as manifestations of the organization's image regarding a representative workforce, rather than individual-level representativeness experienced by citizens in substantive service delivery scenarios.

In contrast, another body of research examines the symbolic effects of one-to-one identity congruence between the client and the bureaucrat during actual interactions (e.g., Guul, 2018; Lee & Nicholson-Crotty, 2022; Theobald & Haider-Markel, 2009). In these studies, the measurement of representativeness is largely dichotomous, indicating whether the individual citizen and bureaucrat share a certain identity (e.g., student and their teacher, driver and the police officer they encountered, etc.). Behind these two different ways of measuring are divergent expectations of how bureaucratic representation works symbolically. That is, whether attitudes for public agencies depends on if the specific government official resembles the individual they interact with, or if the organization passively represents its clients as a whole (Keiser et al., 2022). Therefore, in addition to further theoretical explorations on relevant concepts and mechanisms, it is reasonable and useful to quantitatively examine whether the measurement itself could be a source of inconsistencies of evidence accumulated through the accumulating symbolic representation studies.

3.2 Organizational stratification: Frontline or managerial representation

The symbolic effects of bureaucratic representation might vary from the rank (or, say, the organizational level) at which bureaucratic representatives are present. According to the meta-analysis of Ding et al. (2021) on representative bureaucracy, in which studies on active representation comprised the majority of its sample, representation at the frontline rather than managerial levels results in better organizational outcomes. This may be attributed to the fact that bureaucrat-client encounters happen more often at the street level, where frontline administrators exert immediate and visible influences on policy implementation at their discretion (Meier & Bohte, 2001). Unlike active representation that relies on input from bureaucrats, the theoretical foundation of symbolic representation supposes that a lack of interactivity may not cloud the salience of managerial representation (Lucero et al., 2022). Moreover, with a greater degree of authority and a more powerful voice, managerial representation can be perceived as a stronger signal in demonstrating certain democratic values (Huber & Gunderson, 2023).

Empirically, both frontline and managerial representation have been investigated in studies focused on either organizational or individual-level representativeness. For instance, using a survey experiment, Lucero et al. (2022) showed that identity congruence between an individual citizen and their city manager may increase their compliance with city disaster preparedness orders (i.e., measuring individual-level representativeness, with the bureaucratic representative in a managerial position). Or, utilizing observational data, Wang et al. (2019) examined whether a higher presence of minority police officers would improve citizens' perceived criminal justice (i.e., measuring organizational-level representativeness, with the bureaucratic representatives being front-line officers). However, fewer studies have compared symbolic representation effects across organizational stratifications (e.g., Kang et al., 2022; Xu & Meier, 2021). The question of “Who represents?” has been adequately discussed in the context of research on active representation, but is of equal significance to investigations of symbolic representation—an area of research where it has received comparatively less attention. Hence, this study incorporates organizational stratification as a moderator and assesses the extent to which symbolic effects vary across different ranks of bureaucratic representatives.

3.3 Represented identity: Race, gender, or other

Another potential impact factor on the linkage between passive and symbolic representation is the characteristic on which the measured representativeness is based. While the theory of representative bureaucracy demonstrates concerns for diverse representativeness in its essence (Mosher, 1968), existing empirical studies have predominantly focused on the most prominent demographic characteristics, such as race and gender (Kennedy, 2014). This focus is partially due to data limitations but is also supported by evidence that these identities have the greatest impact on political attitudes and policy perceptions (Keiser et al., 2002; Meier & Jr Stewart, 1992). However, it is acknowledged that restricting the analysis to gender and race might present an incomplete image of representation, as other non-observable characteristics that speak to common experiences and may be central to shared values, could be overlooked (Bishu & Kennedy, 2020; Meier, 2019). To move beyond a narrow operationalization of shared identity, scholars have begun to explore identities beyond race and gender, which can still be salient to the represented population in specific contexts (e.g., LGBTQ status, religion, occupation, etc.). For example, Dantas et al. (2022) offered evidence supporting the symbolic representation effects of the “favela” identity in Brazil, which encompasses the intersectional identity of race, place of birth, and socio-economic status and is more salient than the dynamic of race alone. Therefore, as independent studies increasingly capture broader non-observable and intersectional identities to advance the empirical application of representative bureaucratic theory, it is meaningful to examine whether demographic salience in terms of race and gender still indicates a unique or more substantial effect (Ding et al., 2021).

3.4 Representation outcome: Perceptual or substantive

The potential benefits of symbolic representation have been reported from diverse aspects, which can be categorized as perceptual and substantive outcomes in terms of their forms. Original articulation of symbolic representation stressed its potential to enhance citizens' perceptions and attitudes toward bureaucracies, including but not limited to their satisfaction, perceived legitimacy, trustworthiness, and fairness (e.g., Hibbard et al., 2022; Riccucci et al., 2014). Subsequent studies extended the concept and examined the direct impact of bureaucratic characteristics on citizens' behavioral intentions (e.g., Riccucci et al., 2016), actual behaviors related to compliance or cooperation with organizational decisions (e.g., Lee & Nicholson-Crotty, 2022), and tangible policy outcomes for the represented group (e.g., Wright, 2022). For example, Guul (2018) found that job seekers exhibited higher levels of coproduction effort when matched with a bureaucrat of the same gender. However, it is noted that even perceptual impacts may fall short in certain circumstances (e.g., Benton, 2020; Socia et al., 2021). Therefore, to assess whether focusing on perceptional (attitudes, perceptions, and behavioral intentions) or substantive outcomes (behaviors or tangible policy outcomes) of citizens influenced the strength of the detected linkage between passive and symbolic representation in existing studies, this analysis includes the measurement of representation outcome as a moderator.

3.5 Policy context: Salient or not for the represented identity in question

One of the prerequisites for active representation to occur stipulates that a policy must be salient for the demographic characteristic in question (Meier, 1993). That is, policy implementation shall directly benefit the represented population as a class (Wilkins & Keiser, 2006). For example, a policy issue needs to be “gendered” to anticipate the benefit of promoted gender representation, such as policies geared toward preventing sexual assault or domestic violence (e.g., Atkins & Wilkins, 2013). In terms of race, policy areas such as law enforcement and education are deemed highly significant, especially in countries with a deep history of racial inequality (Darling-Hammond, 1998; Weitzer & Tuch, 2006). Furthermore, certain categories of public services, such as those provided to veterans, may be deemed salient to their specific client groups (Gade & Wilkins, 2013). Nevertheless, research increasingly suggests that the potential of bureaucratic representation may not be exclusive to these “typical” policy areas. For instance, Riccucci et al. (2016) found that female leadership improved citizens' willingness to co-produce, even in a government program that was conventionally not gender-specific: recycling. Comparable evidence on racial representation was also noted in diverse contexts such as emergency management and healthcare (Herman, 2007; Lucero et al., 2022). Given these observations, policy context is included as a moderator to investigate whether stronger symbolic effects still appear in policy domains deemed more relevant to the represented population.

3.6 National context: US or Non-US settings

The literature on representation bureaucracy has primarily focused on a US. context (Bishu & Kennedy, 2020). Responding to the challenge of generalizability, recent scholarship has expanded its geographic coverage in examining multiple aspects of representative bureaucracy theory, including the passive-symbolic representation link (see Doornkamp et al., 2023; Guul, 2018; Sievert, 2021). Particularly, evidence from non-liberal democracies, such as Brazil, China, and Bangladesh, has been on the rise in recent years (see Baniamin & Jamil, 2023; Dantas et al., 2022; Xu & Meier, 2021). Given this advancement, it is worthwhile to ask whether the expected outcomes of symbolic representation vary across US and non-US contexts.

3.7 Research design: Observational or experimental

Besides the abovementioned substantive moderators, this meta-analysis also looks at the extent to which research design matters. Given that the majority of earlier research is based on observational data, the increasing application of experimental frameworks has undoubtedly enriched the understanding of the linkage between bureaucratic representation and its symbolic effects (Riccucci et al., 2014; Riccucci et al., 2018; Schuck et al., 2021). To determine if divergences in research design also account for the variation of findings on symbolic representation, this method indicator is included as an additional moderator.

4.1 Data collection

Data collection commenced by using the Web of Science (WoS) SSCI database to search for peer-reviewed journal articles. Search terms such as “symbolic representation,” “symbolic effects of representation,” and “symbolic AND representative bureaucracy” were used in the “topic” field. By limiting the research discipline to “public administration” and then excluding non-English and non-academic articles, the search yielded 70 results as of April 1, 2022. Following previous meta-analyses related to representative bureaucracy (Ding et al., 2021; Ding & Riccucci, 2023), this study utilized the Representative Bureaucracy Research Archive Data Base.¹ By filtering for studies tagged with the keywords “symbolic representation,” 40 studies from the database were collected. After removing duplicates from 110 records, abstracts of the studies were screened to exclude qualitative research and studies that did not focus on symbolic bureaucratic representation, which resulted in 26 relevant studies.

To ensure the comprehensiveness of the search, two supplementary searches were conducted. I first used Google Scholar, and paid particular attention to the references of collected relevant articles as well as studies that cited the original article. Through this process, three published studies were deemed relevant (Guul, 2018; Meier & Nicholson-Crotty, 2006; Wang et al., 2019). The second search was performed via ProQuest's Dissertation and Theses Global Database and Social Science Research Network (SSRN), in an attempt to reach “gray studies” that have not been published, and check whether publication bias exists in the research field of symbolic representation. Three eligible dissertations were identified, while no relevant working papers from the SSRN were found.

Once the relevant studies were collected, I performed full-text reviews using the following inclusion criteria to identify primary studies for the meta-analysis: (1) The study conducts a direct quantitative test of the symbolic outcomes of passive bureaucratic representation. (2) Passive representation, as the independent variable, is measured as the proportion of bureaucratic representatives (at the organizational level) or identity congruence between the individual bureaucrats and clients (at the individual level). (3) The dependent variable, that is, the symbolic effect of representation, is operationalized as citizens' perceptual or behavioral responses to representative agencies, or specific policy outcomes. (4) Sufficient statistical information, including but not limited to sample size, t, Z, chi-square, beta values, and p-values, is provided in the study. The refined full-text review yielded 28 acceptable studies comprising the final sample for the meta-analysis, including 25 published and three unpublished studies. Figure 1 depicts the procedures for the literature search and inclusion.

4.2 Coding procedures

I then extracted and documented information from the acceptable studies. Two categories of data were coded in the synthesis: data indicating the effect sizes and data for the moderators.

Referring to established conventions and suggestions from Ringquist (2013), this study used r-based effect sizes to represent the passive representation-symbolic representation effect from original studies. When the correlation coefficient r or the standardized regression coefficient β were not provided, the following equation was used to calculate the r-based effect sizes: $r=\sqrt{t^2/\left(t^2+\mathit{df}\right)}$ (or, $r=\sqrt{Z^2/n}$ if using maximum likelihood models).

The t-scores, if not reported, were calculated via the parameter estimates and standard errors. Following the lead of previous meta-analyses published in top public administration journals (e.g., Lu, 2018; Zhang, Li, & Yang, 2022), when original studies identify statistically significant parameter estimates using asterisks or p-values only, I set the t-score equal to the value of t at the symbol threshold and given degrees of freedom. Insignificant coefficients with no additional information were coded as an effect size of zero.

For the studies employing tests of means or analysis of variance (ANOVA) only, group difference-based effect sizes (Cohen's d) were calculated and then converted into r-based effect sizes. For the studies estimating probit or logit models, I first recorded the odds-based effect sizes and then transformed them into r-based effect sizes. For experimental studies performing both mean comparisons and regressions, the regression results were adopted. Further, Fisher's r-to-Z transformation was conducted to remedy the slight downward bias of the r estimate. The formulas of transformation are as follows: $Z_r=0.5\ln \left[\left(1+r\right)/\left(1-r\right)\right]$, with variance $V\left[Z_r\right]=1/\left(n-3\right)$.

Drawing on the suggestion of Ringquist (2013, p. 73), when an original study reports multiple effect sizes, mostly due to different model specifications or sample restrictions, all relevant effect sizes were coded to maintain within-study variation. Finally, a total of 286 effect sizes were drawn from 28 primary studies. Several articles, however, did include multiple independent samples within a single research (mainly from different periods or countries). Therefore, 32 independent samples were identified from the 28 studies, with separate effect sizes distinguished by year or country (Baniamin & Jamil, 2023; Benton, 2020).

In light of the potential sources of variations for the passive-symbolic representation link, four substantive, one methodological, and two contextual indicators were generated for the meta-regression. Details of the coding rules for the moderators are included in Table 1. Table A1 presents an overview of the distribution of the primary study descriptors for all 28 studies.

5.1 Population effect size

A random-effects model was employed to pool the 286 effect sizes given its more applicable assumptions and greater external validity (Hedges & Vevea, 1998; Ringquist, 2013), which was also supported by effect-size heterogeneity tests. A Q-test was first conducted to test the null hypothesis that variance in observed effect sizes is solely caused by sampling error. Since the Q statistic was 2153.97 with a p-value smaller than 0.001, the null hypothesis was rejected. Moreover, the I² statistic of 86.8% indicates a fairly high heterogeneity across effect sizes. Both tests justified the adoption of the random-effects model for integrating effect sizes, and implied that meta-regression is needed to untangle the potentially patterned variations.

Notably, most primary studies generated more than one effect size; the effect sizes in this meta-analysis are thus not independent but adhere to a nested structure. Therefore, a three-level random-effects model was fitted to better capture these dependencies while pooling the effect sizes (Cheung, 2014),² using the R package metafor. Based on the model, the weighted average effect size in Fisher's z was calculated as 0.062 (t = 3.96, p < 0.001), with a 95% confidence interval of [0.031, 0.093]. As a suboptimal strategy, I aggregated the effect sizes by sample and re-ran the conventional random-effects analysis, yielding a population effect size of 0.060 (t = 3.48, p < 0.01), with a 95% confidence interval of [0.025, 0.095]. Table 2 presents the meta-analysis results at the effect and sample levels. Despite the small magnitude of correlation, the consistent results in both cases suggest a positive association between passive representation and symbolic representation. That is, the results empirically support the significance of symbolic representation, suggesting that passive bureaucratic representation can result in overall favorable outcomes.

5.2 Moderator analysis: Meta-regression

A meta-regression analysis was then implemented to examine whether the variations in the collected effect sizes could be explained by the moderators previously identified. Table 3 provides the results of the meta-regression analysis, with random effects assigned to the sample level as the grouping factor. A cluster-robust test was also conducted using a sandwich-type estimator to generate robust variance estimations of the model coefficients. Model 1 includes all the identified moderators. Model 2 introduces the source of the study to check if publication bias exists, which is elaborated upon in the next section of this article.

Additionally, subgroup analysis based on the above moderators was conducted through multi-level meta-analysis as an additional reference (see Table A2). The average effect sizes from all subcategories were found to be positive, though they were not equally significant in a statistical sense.

As shown in both regression models, the symbolic effects of managerial representation appear less robust than that of frontline-level representation (β = −0.13, p < 0.05). This result implies that given the effect sizes detected by existing research designs, stronger responses from citizens are found to the presence of bureaucratic representatives at the street level rather than the managerial positions. A significant effect was also found on the methodological moderator. The observed passive-symbolic representation link turned out to be weaker in observational studies than in the studies that adopted experimental designs (β = −0.092, p < 0.1). With a sub-group aggregate effect estimate of 0.082 (see Table A2), the experimental studies, which contribute 127 effect sizes from 12 samples, reported a more substantial symbolic effect of passive representation than their observational counterparts.

Results for the rest of the moderators suggest no significant impact of these factors on the observed passive-symbolic representation linkage. Regardless of whether the passive representation is operationalized as aggregate proportions of representative bureaucrats or one-to-one identity congruence, it does not significantly moderate the salience of symbolic representation. At the same time, there is no evidence that the representativeness of race or gender resulted in greater symbolic effects than other types of identities that are involved in the current sample studies, such as profession, immigrant status or local identity. In addition, whether the representation outcomes were measured by substantive behaviors and policy outcomes on the client side, or their attitudes and perceptual changes, makes no significant difference in terms of the symbolic effects.³ Moreover, symbolic representation may offer invaluable insight even in policy contexts that do not typically emphasize representation issues (like environmental protection or emergency management). Finally, the research location itself does not seem to matter with regard to untangling the myth of symbolic representation. Given the insignificant coefficients, the observed strength of symbolic representation does not vary much among the US and non-US countries.

5.3 Publication bias

Publication bias occurs when the probability of a study getting published is affected by its results (Rothstein et al., 2005). Given the potential threat of a distorted sample to the validity of effect estimates, tests for publication bias are necessary for meta-analysis. In the present study, I first drew a funnel plot to visually check the distribution of observed effect sizes (see Figure 2), by which the value of all 286 effects (the X-axis) and their standard error (the Y-axis) are displayed. With no publication bias, the data points in such a plot should form an approximately symmetrical, upside-down funnel. It may be noted that, for the most part, the figure points toward rough symmetry and does not present a pattern that would be evident of public bias (Zhang, Chen, et al., 2022). However, several effect sizes fell out of the ideal limits, and apparent outliers remain, suggesting the need to examine the observed funnel symmetry statistically.

I then conducted Egger's regression test (Egger et al., 1997),⁴ with the results listed in Table 4. As noted in the table, the estimate of the intercept does not significantly differ from zero at a 95% confidence level (p > 0.05), which provides insufficient support for funnel plot asymmetry or the existence of publication bias. An additional test that takes the sample sizes of studies into account by using the square root of the sample size as a measure of precision (Zhang, Li, & Yang, 2022)⁵ generated results that were nearly identical to those of Egger's test. Finally, as shown in the meta-regression model 2, which includes the indicator of published study, the corresponding coefficient was proven insignificant. Taken together, given a high degree of effect size heterogeneity and the results of multiple tests, publication bias is not a major threat in the present meta-analysis.