Political science has devoted an increasing amount of attention to the study of minority groups, with topics ranging from the behavior of voters (Tate Reference Tate1991; Gay Reference Gay2001; Barreto, Segura and Woods Reference Barreto, Segura and Woods2004; Barreto Reference Barreto2007; Hajnal Reference Hajnal2010) to candidates (Shah Reference Shah2014; Juenke and Shah Reference Juenke and Shah2015; Fraga Reference Fraga2018; Atsusaka Reference Atsusaka2021). Previously, a paucity of data on the racial and ethnic identities of voters or elected officials necessitated the use of aggregate data to infer individual-level behavior. But with recent statistical innovations, researchers are able to study individual-level, rather than group-level, political outcomes by imputing group categories. Namely, Bayesian Improved Surname Geocoding (BISG), which uses Bayes’ Rule to combine geographic and surname information, has become a popular approach due to its computational efficiency. The development of various machine learning models for group classification has further facilitated the growth of research on topics such as racial disparities in turnout (Fraga, Reference Fraga2016b, Reference Fraga2018), public opinion (Menshikova and Tubergen Reference Menshikova and van Tubergen2022), campaign contributions (Grumbach and Sahn Reference Grumbach and Sahn2020), and descriptive representation (Enos Reference Enos2016; Schwemmer and Jungkunz Reference Schwemmer and Jungkunz2019).
However, an analysis of various prediction methods reveals unique limitations. First, under certain conditions, Bayesian predictions that rely on geographic information can lead to biased estimates in downstream analyses (Argyle and Barber Reference Argyle and Barber2024). Studies using BISG to generate measures of outcome variables, for example, can produce systematic measurement errors if prediction errors are correlated with covariates in a downstream regression model. When independent variables are geographic in nature – as they often are in studies of descriptive representation and minority turnout – biased and overconfident inferences are a possible consequence. Second, an alternative Bayesian approach that solely relies on surnames tends to generate imbalanced accuracy, with particularly high error rates among Blacks. Third, machine learning techniques, particularly image-based models, improve accuracy for racial categories but have difficulty distinguishing Hispanics from non-Hispanic whites. We propose an alternative approach that improves upon existing classification methods while sidestepping issues that arise due to the use of geography-dependent prediction methods. Specifically, we combine surname-based Bayesian estimation with image-based predictions generated by a convolutional neural network model. Using predicted probabilities from both methods as features, the proposed classification approach not only mitigates geography-driven biases but also yields a higher out-of-sample accuracy rate than alternatives.
An empirical validation study using data on over 16,000 US local elected officials shows that the proposed method results in lower error rates than other techniques. We also find that the proposed method has an overall classification success rate of 0.93 or greater across the four main ethnoracial categories in the Census – the strongest performance among all prediction methods evaluated. Furthermore, we find that our method reduces bias in settings involving geographic predictors.
Our approach contributes to the growing literature on ethnoracial prediction that has revealed impressive accuracy for machine learning approaches. Comparing the performance of BISG against a number of machine learning models, Decter-Frain (Reference DeLuca and Curiel2022) finds that “even a simple switch from BISG to multinomial logistic regression can substantially improve individual predictions” (p. 4, see also Wong et al. Reference Wong, Zaiane, Davis and Yasui2020; Cheng et al. Reference Cheng, Gallegos, Ouyang, Goldin and Ho2023). Similarly, Argyle and Barber (Reference Argyle and Barber2024) incorporate BISG predictions into a random forest model to reduce misclassification bias. Our proposed method contributes to this increasing use of machine learning tools by incorporating rich information embedded in images. We conclude with a discussion of future directions, ethical and practical considerations, and recommendations for applied researchers. We also provide step-by-step instructions in online materials to facilitate its use.
An Overview of Existing Methods
Possessing individual-level race and ethnicity information is crucial for modeling turnout and vote choice, among other outcomes. However, publicly available data sets often lack this information, and scholars have to rely on a variety of prediction tools to construct proxies. Table 1 highlights the diversity in existing methods commonly adopted in the literature in terms of both statistical and practical approaches.
Table 1. Assessment of existing prediction methods and the proposed hybrid approach. The ‘hybrid method’ is used as shorthand to describe the proposed prediction method introduced in this paper.
a Based on a literature review of published articles in social science. See Appendix B for details.
b 1 = High; 6 = Low. Based on the overall classification rate identified in Section ‘Applications’.
c Refers to biases when predictions are used as model outcomes. See Section ‘The Bayesian Approach’ for details.
d Costs may be incurred from data labelling. See Section ‘Practical Considerations’ for further discussions.
We first provide an overview of existing methods, which can largely be divided into two statistical approaches: Bayesian imputation and machine-learning techniques. We then introduce our proposed method, highlighted in the bottom row of Table 1. We demonstrate its use via empirical application and detail its benefits and limitations in comparison to other established methods.
Prior to proceeding, we note the fundamental differences between the two statistical approaches. Bayesian imputation relies on unsupervised learning, and thus, does not require labels or classifications. By contrast, machine-learning approaches are supervised methods that require a sufficient amount of labelled data.Footnote 1
The Bayesian Approach
The Bayesian approach has been the most popular method due to its parsimony and usability, especially with the introduction of the R package wru (Imai and Khanna Reference Imai and Khanna2016). To quantify its popularity, we reviewed over ninety recently published articles in the social sciences that utilized racial prediction methods and found that 60 percent of the articles implemented either Bayesian Surname-Only (BSO) or Bayesian Improved Surname Geocoding (BISG) methods.Footnote 2
BSO infers race using the racial distribution of surnames from American Decennial Census Surname Files. While this method performs well for Hispanics/Latinos because surnames are a strong predictor of ethnic self-identification, it is less capable of distinguishing Blacks from non-Hispanic whites. For example, using Florida voter registration records, Imai and Khanna (Reference Imai and Khanna2016) find false negative rates to be above 0.83 among Blacks and false positive rates to be above 0.52 among non-Hispanic whites for the surname-only Bayesian prediction.
BISG improves upon BSO by combining information from the Census surname list with data on local area demographics (Elliott et al. Reference Elliott, Morrison, Fremont, McCaffrey, Pantoja and Lurie2009; Imai and Khanna Reference Imai and Khanna2016). In practice, the Bayesian approach provides posterior probabilities across ethnoracial categories recorded by the Census – e.g., Asian, Black, Hispanic/Latino, white, and Other – for each individual.Footnote 3 BISG produces fairly accurate predictions of individuals’ self-reported ethnoracial identification, although accuracy varies by group. For example, Elliott et al. (Reference Elliott, Morrison, Fremont, McCaffrey, Pantoja and Lurie2009) estimate a weighted average correlation coefficient of 0.76 between predictions and self-reported data, which is 41 percent more accurate than BSO. Using the Florida voter file, Imai and Khanna (Reference Imai and Khanna2016) report a similar accuracy rate, as well as improved estimates of aggregate-level turnout rates for different racial and ethnic groups, when compared to standard statistical methods.
In recent years, several approaches have aimed to improve the standard BISG methodology. Namely, a fully Bayesian Improved Surname Geocoding (fBISG) improves upon BISG by both accounting for Census measurement error and incorporating additional data on first and middle names taken from voter files of six Southern states that offer data on racial self-identification (Imai, Olivella and Rosenman Reference Imai, Olivella and Rosenman2022). Similarly, zipWRU extends BISG by drawing on ZIP code Tabulation Area (ZCTA) (DeLuca and Curiel Reference DeLuca and Curiel2022), as opposed to the larger units of aggregation that are commonly used in this research. Although the Bayesian approach is technically applicable to contexts beyond the USA, its adoption has surged within the USA, propelled by the availability of Census data as well as the convenience of R packages such as wru.
Despite its utility, the risks of systematic measurement error, due to the use of geographically-informed predictions, might become more prevalent as research on race and politics expands to answer a wider set of questions. To provide a running example, consider a typical question from the descriptive representation literature such as how many minority legislators are elected to a given office. In the USA, minority candidates face challenges in running for office, or getting elected, even when their districts have large co-ethnic populations (Swain Reference Swain1995; Canon and Posner Reference Canon and Posner1999). For example, prior to 2014, Ferguson, MO had never elected a Black mayor, and only two Black members had ever served on the council despite the fact that 67 percent of the city is African American (Eligon Reference Eligon2015). But if a researcher were to use BISG to estimate the racial diversity of the local government, it would disproportionately assign a high predicted probability to the Black category for all elected officials on the council due to the city’s large Black population. In fact, BISG classifies 36 percent of candidates who ran for local elections in Ferguson as Black, when the true proportion is only 5 percent.Footnote 4 This example demonstrates how geography-informed predictions can understate the extent of racial inequality in local politics.
This issue arises in research that analyzes ethnoracial predictions as a function of geographic factors. Of course, using a predicted measure as a dependent variable would not introduce bias if the measurement error only adds stochastic variation (Berry and Feldman Reference Berry and Stanley Feldman1985; Hausman Reference Hausman2001). However, when measurement error is correlated with independent variables, biased estimates are a consequence. First, non-stochastic measurement error occurs when the observed outcome variable systematically measures some other variable(s) in addition to the true outcome, causing the observed outcome to vary on the basis of the other variable(s). If BISG is used to predict outcomes, this is a possible issue, as this outcome measure will also capture systematic variation due to the ethnoracial composition of each location. Second, if a downstream regression model includes a predictor that covaries with the measurement error, this can lead to biased estimates even after adjusting for name and location (Argyle and Barber Reference Argyle and Barber2024; McCartan et al. Reference McCartan, Goldin, Ho and Imai2023).Footnote 5
Applications involving geographic variables are common in political science research that examines the relationship between candidate ethnicity and district racial composition, or between district racial composition and minority turnout. For example, Hankinson and Magazinnik (Reference Hankinson and Magazinnik2023) employ BISG to estimate the race of California city councillors, Grumbach and Sahn (Reference Grumbach and Sahn2020) and Alvarez, Katz and Kim (Reference Alvarez, Katz and Seo-young Silvia Kim2020) use BISG for political donors, Abott and Magazinnik (Reference Abott and Magazinnik2020) and Kogan, Lavertu and Peskowitz (Reference Kogan, Lavertu and Peskowitz2021) for school board members, and Conroy and Green (Reference Conroy and Green2020) and Shah and Davis (Reference Shah and Davis2017) for prospective candidates – all of which include geography-specific independent variables in downstream analyses.
In Online Appendix D, we provide a simulation result that documents the potential magnitude and direction of bias when BISG is inappropriately used (that is, in cases where geographic variables are included as independent variables and there is systematic measurement error). Specifically, we simulate a downstream model where BISG predictions of legislator ethnorace are regressed on a geographic predictor. We find that the bias, while varying in magnitude, is generally upward when BISG is used and downward when surname-only predictions are used.
Machine Learning Approaches
Political scientists and social scientists more broadly have increasingly adopted machine-learning algorithms. Based on our literature review, text-based machine learning methods constitute 27 percent of published articles examined, which is comparable to the rate of BSO’s use. Namely, many empirical papers employ Long Short-Term Memory Models (LSTM) that exploit the sequence of letters in a name. Trained on a diverse set of training data such as voter files and Wikipedia, LSTM has shown higher accuracy than conventional BISG, particularly among Blacks and Asians (Sood and Laohaprapanon Reference Sood and Laohaprapanon2018). LSTM has been implemented to examine racial disparities in academic institutions (Marschke et al. Reference Marschke, Nunez, Weinberg and Yu2018), public opinion among Twitter users (Menshikova and Tubergen Reference Menshikova and van Tubergen2022), and vulnerability to natural disasters (Messager et al. Reference Messager, Ettinger, Murphy-Williams and Levin2021).
Automatic estimation of demographic attributes (for example, age, gender, and race) from images has become another topic of growing interest and controversy, with many potential applications. Our literature review illustrates this growing demand, as we find that 5 percent of articles, all published since 2019, have leveraged an image-based machine learning algorithm. Current research relies primarily on convolutional neural network (CNN) models, which have demonstrated strong performance on challenging prediction tasks. In essence, these deep learning algorithms take an input image, assign importance to various latent features, and classify demographic attributes based on this information. CNNs capture spatial dependencies in an image through the application of filters of varying levels of complexity that are learned by the algorithm. As examples of how CNNs have been used in recent work, Schwemmer and Jungkunz (Reference Schwemmer and Jungkunz2019) investigate descriptive representation in digital platforms by annotating the gender and ethnoracial identities of TED Talk speakers using Kairos diversity recognition software, and Lyu et al. (Reference Lyu, Fan, Xiong, Komisarchik and Luo2021) employ DeepFace API to predict the ethnic and racial identities of Twitter users when examining their opinion toward racially motivated hate crimes in the USA.
However, one of the shortcomings of the image-based approach is that it is less able to distinguish between Hispanics/Latinos and non-Hispanic whites, presumably because it relies on skin tone as a feature to predict ethnoracial identity, and Latinos are a multi-racial community. This sensitivity is, in part, due to the quality of the data set that the model is trained on. In the realm of deep learning, having a substantial amount of data in terms of both quantity as well as ethnoracial diversity is crucial for effectively training CNNs. However, the availability of such data sets has been lacking due to the challenges involved in annotating such data (Abdulwahid Reference Abdulwahid2023). Because ethnorace cannot be quantitatively measured, human annotators must establish ground truth values (Sulaiman and Kocher Reference Sulaiman and Kocher2022).
A growing number of training databases have been introduced to close this gap. For example, the FairFace data set is comprised of over 108,000 images balanced across seven different ethnoracial groups (Karkkainen and Joo Reference Karkkainen and Joo2021). Other similar data sets include VGG-Face2, which contains 2.6 million images with annotations for six different ethnoracial groups, and the VGG-Face2 Mivia Ethnicity Recognition (VMER) data set, which improves upon VGG-Face2 by obtaining ethnoracial classifications from three independent annotators belonging to different groups (Greco et al. Reference Greco, Percannella, Vento and Vigilante2020).
Supervised models trained on these data sets with a more granular set of ethnoracial categories could improve estimation methods. Aggregating ethnoracial groups with distinct cultures can limit research on categories not covered by existing classification schemes that rely on the Census (see footnote 3). For example, Middle Eastern and North African (MENA) Americans are categorized as white in the Census, but racial self-identification among MENA individuals typically does not reflect this (Maghbouleh, Schachter and Flores Reference Maghbouleh, Schachter and Flores2022). Similar research on Asian American identity also finds distinct political behaviors and self-identification patterns between South and East Asian Americans (Yamashita Reference Yamashita2022). While conventional Bayesian approaches use five ethnoracial categories, certain pre-trained models (like FairFace and VGG-Face) distinguish between Middle Eastern and white individuals and Southeast Asian and East Asian individuals. This finer-grained classification could be helpful for researchers studying political behavior within these subgroups.
Furthermore, the fact that the image-based approach relies on perceptual classifications as ground truth may serve as a complement to the Bayesian approach, which is typically evaluated using self-reported data. Studies of ethnic politics have long been interested in the interaction between self-identification and perceived identification of others (Habyarimana et al. Reference Habyarimana, Humphreys, Posner and Weinstein2007; Harris and Findley Reference Harris and Findley2014). This difference may be useful for research that examines two “ingredients” of identity and how they might affect political outcomes such as representation.
Proposed Method: Hybrid Approach
We propose a general supervised learning approach that improves prediction accuracy while mitigating the downstream bias observed with the previous methods. That is, we use probabilities obtained from both BSO and image-based CNN as features in a simple multinomial logistic model and generate predictions of individual ethnorace. The use of BSO predictions not only improves prediction accuracy among Hispanics – a group where image-based methods perform poorly – but also mitigates the potential bias in downstream analyses embedded in other Bayesian methods that rely on geographic information. The use of image-based predictions also improves accuracy for Blacks, a group where the Bayesian approach performs poorly. By allowing the strengths of one to compensate for the weaknesses of the other, we show that this hybrid approach is able to not only reduce bias in downstream analyses but also increase the accuracy of the individual race/ethnicity prediction.
The proposed approach proceeds as follows. First, given data on names, we obtain predicted probabilities for each individual using BSO, which can be done using publicly available tools such as wru. Second, we collect images online using full names as search engine keywords. We then use a pre-trained CNN model to generate predicted probabilities of racial and ethnic categories based on the collected images. Third, we train a multinomial logistic model using these two sets of predicted probabilities as features to generate predictions.
In the following sections, we demonstrate the proposed methodology using large-scale data on local elected officials in the USA.Footnote 6 We show that, compared to other prediction techniques used in the literature, the proposed classification approach not only mitigates geography-driven biases but also yields a higher out-of-sample accuracy rate. We then discuss in detail the potential risks and comparable benefits associated with the hybrid approach.
Applications
To demonstrate the proposed methodology, we use a large-scale database of local elected officials. At the time of this study, the data set contained over 16,000 unique elected city councillors and mayors in the USA between 1950 and 2021, representing over 800 unique cities across the country. Importantly, it includes full names as well as ethnoracial identities of the elected officials classified as white, Black, Asian, Hispanic/Latino, or other. The data set also includes district identifiers, election year, and party affiliation (if available).Footnote 7
Using this data set, we validate our method in two ways. First, we use the individual-level ethnoracial information in the data set as ground truth and compare the performance of the proposed approach, as well as other commonly used prediction techniques. Second, we test for bias in downstream analysis by conducting two descriptive analyses examining minority representation in government offices. That is, we assess how much estimates computed using our proposed approach differ from those that rely on the ground truth classification.
Procedures
We now document the implementation of our method. First, we use wru to predict the ethnoracial identities of all legislators in our data set using the surname-only Bayesian method (BSO) and calculate the posterior probabilities across five ethnoracial classifications: non-Hispanic white, non-Hispanic Black, non-Hispanic Asian, Hispanic/Latino, and Other.Footnote 8
Second, we run a pre-trained CNN model to predict individual ethnoracial categories using publicly available images. To collect facial images associated with each name, we use the first and last names of elected officials as keywords and search for the associated images on a public search engine.Footnote 9 To predict the ethnoracial classifications of the collected images, we use a CNN model pre-trained on the FairFace data set. As discussed, the FairFace data set was constructed with the goal of mitigating racial bias in existing public face data sets and includes samples that are equally distributed among seven different ethnicity groups, annotated by at least three workers for each image (Karkkainen and Joo Reference Karkkainen and Joo2021). Studies find that a neural network trained on FairFace data sets generalizes better than the same model trained with other pre-existing training data sets such as UTKFace and LFWA+ (Greco et al. Reference Greco, Percannella, Vento and Vigilante2020).Footnote 10
We emphasize that our method does not use facial images tied to particular individuals, but a collection of individual faces associated with a given name. Collecting photos of each elected official and generating predictions based only on those images is possible but we chose an alternative approach for two reasons. First, we attempt to mitigate the ethical concerns around the academic use of personally identifiable photos by carrying out a procedure where images are collected, machine classification is performed, and images are subsequently deleted.Footnote 11 Second, independent of ethical considerations, collecting photos for every individual is difficult or often impossible. Through an automatic collection of publicly available images on a large scale, we demonstrate that our method is able to increase efficiency while maintaining high performance, even when the images may not be of a particular individual in the data set.
Using the pre-trained weights from FairFace, we generate predictions for the following classifications: white, Black, East Asian, Indian, Southeast Asian, Middle Eastern, and Hispanic/Latino. To be consistent with the ethnic and racial classifications of the ground truth data – and the categorizations used in Bayesian approaches – we combine the Middle Eastern category with white and all Asian categories into a single Asian category. While we choose to aggregate the ethnoracial categories as such in order to evaluate the prediction performance across different methods, we encourage researchers to utilize more granular categorization in their research.
Lastly, we train a multinomial logistic model using the ground truth ethnoracial classifications as a dependent variable and the following predictions as features: 1) the predicted probabilities of five ethnoracial categories obtained from BSO and 2) the predicted probabilities of four ethnoracial categories obtained from the image-based predictions. We use 70 percent of the data as a training set and test its performance on the remaining 30 percent of the data set.
Prediction accuracy
We first validate the prediction accuracy of the proposed method. To do so, we compare the ethnoracial predictions from the following methods against the ground truth: 1) a surname-only Bayesian approach (BSO); 2) a Bayesian Improved Surname and Geocoding (BISG); 3) a fully Bayesian Improved Surname and Geocoding (fBISG); 4) a full name-based LSTM;Footnote 12 5) image-only predictions based on FairFace; and, finally, 6) the proposed hybrid method that combines predictions 1) and 5).
We calculate the overall error rate, which represents the proportion of officials whose ethnoracial classification is incorrect, as well as the false positive (Type I error) and false negative (Type II error) rates. The results are shown in Table 2. We find that our proposed method reduces the overall error rate to 11.2 percent compared to 18.1 and 18.8 percent from BSO and the image-based prediction, respectively. The results also show that the proposed method is more accurate than fBISG (14.6 percent).
Table 2. Overall classification error, Type I error and Type II error rates for each ethnoracial category across six prediction methods. The lowest error rate in each row is in bold. Results are based on out-of-sample data (
$N$
=4,976)
As expected, BSO yields high false negative rates among Blacks as well as high false positive rates among whites but performs well among Hispanics/Latinos.Footnote 13 Conversely, the image-based prediction yields high false negative rates among Hispanics/Latinos but shows a significant improvement in predicting the identity of Black legislators when compared to the surname-only Bayesian approach.Footnote 14 Importantly, the proposed method shows a significant improvement over the two prediction methods alone; specifically, the false negative rates among Blacks are reduced to their lowest values compared to the two methods alone. Among Hispanics/Latinos, the proposed method decreases the false negative rate by a factor of 6 compared to the image-based method. The proposed method also improves predictions among Asian Americans, yielding the lowest false positive rates when compared to all other methods.
Table 2 evaluates the performance of each method based on the plurality-based ethnorace assignment, which obscures the uncertainty embedded in the distribution of probabilities. To address this concern, we provide two additional diagnostics: tetrahedron diagrams and receiver-operated curve (ROC) analyses.Footnote 15 First, the tetrahedron diagrams in Figure 1 display the distribution of predicted probabilities across four ethnoracial categories from each prediction method. A tetrahedron diagram is a three-dimensional plot that takes a pyramid form with four vertex corners. Given that there are four primary ethnoracial categories generated by each method, these diagrams are useful for inspecting the composition of the probabilities and capturing the uncertainty associated with them.Footnote 16
Figure 1. Tetrahedron diagrams of predicted probabilities across four ethnoracial categories (A = Asian, B = Black, H = Hispanic/Latino, and W = white).
Each data point in the diagram represents an individual
$i$
with colours indicating the magnitude of the predicted probabilities allocated to each category, from low (blue) to high (red). For example, the red data points near the corner H(ispanic) represent individuals whose predicted probabilities of being Hispanic/Latino are close to 1. Conversely, the blue data points in the center represent individuals whose predicted probabilities are rather evenly distributed across the four categories, indicating high prediction uncertainty. The diagrams on the top row display the composition of predicted probabilities only for those individuals whose highest predicted probability correctly predicts their actual ethnoracial category (that is, true positives and true negatives). The diagrams on the bottom are limited to those individuals whose highest predicted probabilities incorrectly predict their ethnoracial category (that is, false positives and false negatives).Footnote
17
Both BISG and fBISG show a relatively higher level of uncertainty as compared to BSO, which is unsurprising as the high probability assigned to the white category gets distributed to other ethnoracial categories once the city’s racial composition is accounted for. All models, however, assign high probabilities to incorrect ethnoracial categories quite frequently, as illustrated by the concentration of red points in the bottom diagrams. One exception is the image-based method: the reduction of red points along the corners of A(sian) and B(lack) illustrates that when the officials are incorrectly predicted to be of these categories, the predicted probabilities associated with these ethnoracial categories tend to be low. This is also reflected in the hybrid method that incorporates image-based predictions. Nonetheless, we find that the magnitude of predicted probabilities does not necessarily correlate with whether or not the prediction can be made correctly based on the plurality-based ethnorace assignment.
Next, we examine the ROC curve for each prediction method, as shown in Figure 2. Rather than classifying individuals on the basis of their greatest predicted probability, ROC curves display the true positive rate against the false positive rate for a number of different classification thresholds between 0 and 1. The performance of a prediction model is evaluated using the area under the ROC curve, the AUC, which summarizes the overall classification success.
Figure 2. ROC curves across different ethnoracial prediction methods. The AUC scores for each method are shown within each plot.
The results show that our method yields the highest AUC score – 0.93 or greater – across all racial categories: BSO alone does not perform well among Blacks and whites, while the image-based method alone does not perform well among Hispanics/Latinos. But combined, the proposed hybrid method is able to reach a higher level of accuracy. With the proposed method, it is possible to reduce the false positive rate among Hispanics/Latinos to 0.05 while maintaining the true positive rate above 0.91. This means that the combined method correctly classifies over 91 percent of Latinos, while only misclassifying 5 percent of non-Latinos as Hispanics/Latinos. This is a significant improvement over the image-based method alone, which generates a false positive rate of 0.42 if it were to maintain the same true positive rate as the combined method. Similarly, the proposed method allows for a reduction of the false positive rate among Blacks to 0.10 while maintaining the true positive rate above 0.80. BSO alone would generate a false positive rate of 0.24 if it were to maintain the same true positive rate.
Evaluating downstream bias
In this section, we assess how well the proposed method describes actual patterns of minority representation in local politics. We do so by examining a popular research question in the descriptive representation literature: the relationship between local ethnoracial composition and the success of co-ethnic candidates. Studies have long theorized that larger co-ethnic populations increase the likelihood of descriptive representation (Shingles Reference Shingles1981; Lublin Reference Lublin1997; Ocampo Reference Ocampo2018; Atsusaka Reference Atsusaka2021). In addition to being the central inquiry in representation studies, this research question views the ethnoracial identity of elected officials as an outcome. Consequently, many empirical studies employ regression models that predict the ethnoracial makeup of politicians while considering factors such as the size of the co-ethnic population and other district-specific covariates (Fraga Reference Fraga2016a; Grumbach and Sahn Reference Grumbach and Sahn2020; Komisarchik and White Reference Komisarchik and White2022). This research question therefore offers an ideal empirical setting for evaluating downstream bias when predicted ethnorace is used as an outcome.
Before testing bias in downstream analyses, we first depict how descriptive representation has evolved over time within city councils. Figure 3 shows loess curves fitted through the proportion of Asian, Black, and Hispanic elected officials over time. The solid blue curve indicates an estimate based on the ground truth, which we use as a benchmark to evaluate the remaining six curves that are based on predictions.Footnote 18
Figure 3. The proportion of non-white officials over time predicted with different ethnoracial category prediction methods.
Examining the ground truth trend (blue line), we find that local government has experienced an increase in the share of Black and Hispanic legislators since the 1950s, but Asian representation shows stagnation until the 2000s and has experienced a slight uptick in recent years. However, the degree to which a researcher can observe such a pattern depends on which prediction method is used to measure the outcome. We observe that the over-time trend estimated with the proposed method most closely tracks the benchmark across all ethnoracial categories. In particular, all other methods diminish the overall increasing trend in Black representation, suggesting that the ability of Black voters to achieve descriptive representation in local government has remained stagnant over the past several decades. Furthermore, BISG and fBISG generally overestimate minority representation, especially during the earlier years when actual minority representation tends to be low. This may be due to the fact that the Census Surname File is only available for 2000 and 2010; thus, the ethnoracial identities of legislators prior to 2000 are predicted with lower accuracy. In comparison, the ability of the hybrid method to produce a fairly accurate result is noteworthy, given that it largely relies on recently collected images.
We now test for bias in the downstream analysis by estimating the relationship between minority population size and ethnoracial diversity in local government. The goal here is to investigate whether the proposed prediction, when used as a dependent variable, can reduce bias in a model that includes a geography-specific factor as an independent variable. We estimate a multinomial logistic model that regresses the ethnoracial category of an elected official
$i$
in state
$s$
and decade
$t$
(
${Y_{i,t}}$
) on
$k$
population share in her city. Formally, for
$k \in K = \{ $
Asian, Black, Hispanic
$\} $
,
$$ln\Big({{{Pr({Y_i} = k)}\over{Pr({Y_i} = White)}}\Big) = \alpha + {\beta _A}Asia{n_{i,t}} + {\beta _B}Blac{k_{i,t}} + {\beta _H}Hispani{c_{i,t}} + \theta {{\bf{X}}_{i,t}} + {\delta _s} + {\gamma _t}}$$
where
$Asia{n_{i,t}},Blac{k_{i,t}}$
and
$Hispani{c_{i,t}}$
represent Asian, Black, and Hispanic/Latino city population, respectively,
${{\bf{X}}_{i,t}}$
represents a collection of covariates (for example, incumbency and office title), and
${\delta _s}$
and
${\gamma _t}$
denote state and decade fixed effects, respectively. We use the white category as the baseline category. The coefficients
${\hat \beta _A}$
,
${\hat \beta _B}$
, and
${\hat \beta _H}$
capture the adjusted correlation between the racial/ethnic composition of the population and the composition of the city council.
We run this model six times, each time with a different
${Y_{i,t}}$
but with the same set of predictors. We first use the ground truth ethnoracial classification for
${Y_{i,t}}$
and use the
${\hat \beta _k}$
obtained in this model as a benchmark (‘benchmark model’). We then create
${Y_{i,t}}$
using each of the six prediction methods to run the same model (‘comparison model’) and compare the
${\hat \beta _k}$
against the benchmark estimate.
We choose this particular model in order to reflect the most commonly employed modelling approach in similar research. For example, Fraga (Reference Fraga2016a) uses a generalized estimating equation with election year indicators and within-district clusters. Both Komisarchik and White (Reference Komisarchik and White2022) and Grumbach and Sahn (Reference Grumbach and Sahn2020) use a difference-in-difference design including year and geography fixed effects.Footnote 19 In Appendix H, we also run alternative specifications and show that the relative RMSE is most stable when the model’s outcome is predicted with the proposed hybrid method.
Furthermore, because the multinomial method only allows for a nominal outcome variable, studies typically assign each individual the racial category with the highest predicted probability. This plurality-based method has become a common approach in studies. Indeed, our literature review shows that almost one-third of political science studies involving Bayesian racial prediction use the plurality-based method in their downstream analyses.Footnote 20 While we follow this general approach in our application, we recognize that this approach does not account for the uncertainties embedded in the entire distribution of predicted probabilities (Clark, Curiel and Steelman Reference Clark, John, Curiel, Tyler and Steelman2021; McCartan et al. Reference McCartan, Goldin, Ho and Imai2023). In Appendix H, we present three additional downstream analyses: 1) a Dirichlet regression that takes into account the entire predicted probability distribution as an outcome; 2) a weighted model using the predicted probability as analytic weights; and 3) a thresholding approach that limits the data to individuals whose predicted probability exceeds a predetermined threshold.Footnote 21 We find that the results are substantively similar across approaches.Footnote 22
Figure 4 displays
${\hat \beta _k}$
and 95 percent confidence intervals (CIs) from the six comparison models separately for each outcome category
$k$
. The black horizontal lines and the grey shaded areas represent
${\hat \beta _k}$
and the 95 percent CIs from the benchmark model. Similar to the simulation results in Appendix D, the estimates using geography-dependent predictions BISG and fBISG are consistently higher, while the estimates obtained from BSO and the image-based method are consistently lower than the benchmark estimate. By contrast, estimates from the hybrid method are closer to the ‘benchmark’ estimate in virtually all cases.
Figure 4. Multinomial logistic regression results. The solid black line and shaded grey area represent estimates and 95 percent CIs from the benchmark model. See Appendix H for full regression results.
Substantively, the benchmark model predicts that a Black population shift from 1 percent (10th percentile of the data) to 42 percent (90th percentile) increases the probability of electing a Black official by 32 percentage points (pp). This ‘marginal effect’ decreases to 22pp under the hybrid method but increases to 48pp and 38pp under BISG and fBISG, respectively, and decreases to 2pp under BSO. The results are even more striking for Hispanic representation, where the estimated coefficient of the hybrid method is lower than that of the benchmark only by 0.0063: the benchmark model predicts that a Hispanic population shift from 2 percent (10th percentile of the data) to 42 percent (90th percentile) increases the probability of electing a Hispanic elected official by 13pp. This predicted probability is nearly identical under the hybrid method (0.57pp higher) while it increases to 17pp and 18pp under BISG and fBISG, respectively. Although the results are noisier, similar patterns are observed for Asian representation.
Assumptions
Comparing across six estimation methods in a sample of over 16,000 local elected officials, we find that the hybrid approach yields concrete advantages over alternatives with respect to predictive accuracy and bias. However, using a supervised machine learning approach with images as additional information requires additional theoretical and statistical assumptions. We address each of these assumptions below.
Image data generating process
Our proposed method relies on a collection of publicly available images associated with individual names, which assumes that these images are representative. We assess potential violations of this assumption by exploring three aspects of the data generating process for images: availability, relevance, and diversity.Footnote 23
First, we assess whether image availability for a given name may influence accuracy. If the total number of available images (for example, popularity of names) varies significantly by racial groups and if availability is correlated with misclassification errors, the image-based predictions would yield lower and imbalanced accuracy. To evaluate this concern, we collect up to 50 recent images per individual, remove those that do not include faces, and exclude exact duplicates. For a given name, we are able to gather 19, 21, 22, and 24 images on average for Asian, Black, Hispanic, and white individuals, respectively, finding comparable levels of availability across racial groups. Further, we find that fewer available images generate higher accuracy for minority groups, while the reverse is true among non-Hispanic whites, implying that having more images does not necessarily improve accuracy. Therefore, we rely on the five most recent images rather than using all 50 images collected, which reduces the overall classification error rate from 0.199 to 0.188, as reported in Table 2.
Second, we assess whether variation in image relevance biases accuracy. If the collected images are irrelevant and if such irrelevance is correlated with prediction errors, the accuracy rate may be reduced and unbalanced across groups. We define relevance as the order in which the images appear in the search engine, with images that appear earlier in order as more relevant. We find a mild positive correlation between accuracy and image relevance: the earlier the images appear, the more likely the predictions from those images are correct. To minimize the classification error due to relevance (or a lack thereof), we take the average of the probabilities weighted by the relevance when we aggregate the predictions from five images for a given name.Footnote 24
Third, we assess the diversity of images. If a search engine returns facial images unrelated to the name itself, the collected images may add further noise to the prediction. While we cannot assess this directly by identifying the ‘true’ racial and ethnic identification of each image, we evaluate this issue indirectly by comparing the dispersion of predicted probabilities generated from images against those generated from BSO. If the dispersion is higher for the image-based predictions, we treat the images as having more diverse ethnoracial identities relative to the true distribution of race for a given name in the USA, which may impact the prediction accuracy. We find that the dispersion pattern of image-based predictions is similar to BSO: when the image-based predictions assign lower predicted probabilities across all racial categories, so does BSO. Importantly, among names where BSO is highly uncertain (for example, when BSO assigns 50 percent to one and 50 percent to another racial category), the image-based predictions are more aligned with the ground truth.
Despite the superior performance on metrics such as accuracy and downstream statistical bias, scholars should always be mindful of systematic measurement error. For example, if racially mixed districts tend to elect people who look racially ambiguous, the image-based predictions may also be correlated with the district’s racial composition. In Appendix F, we put this to the test and find no evidence of this non-linear expectation.
Neural network architecture
While users may choose a CNN model of their choice, we choose the FairFace architecture, as it is trained on one of the most “complete data sets currently available for ethnicity recognition” (Greco et al. Reference Greco, Percannella, Vento and Vigilante2020). Across a wide range of image data sets, the model trained on the FairFace data set increases accuracy for various racial groups, whereas preserving the architecture and training with other data sets leads to asymmetric accuracy (Karkkainen and Joo Reference Karkkainen and Joo2021). Another advantage of using FairFace is that its pre-trained model is publicly available, making it easy for users to implement the model without having to build their own. However, as a robustness check, we also generate predictions using the VGG-Face architecture (Parkhi, Vedaldi and Zisserman Reference Parkhi, Vedaldi and Zisserman2015) and find substantively similar, but slightly lower, predictive performance (see Appendix F). New training data sets and more refined deep network architectures are currently being developed, however, and future research may explore these other models to further improve on the prediction accuracy presented in this paper.Footnote 25
Supervised learning model
Although more sophisticated machine learning algorithms for multi-class classification exist, we choose a multinomial logistic model because of its ability to generate fairly accurate predictions despite its simplicity. We find that estimating a super-learner ensemble using LASSO, Random Forest, and eXtreme Gradient Boosting, as base learners generate error rates that are similar to those from the multinomial logistic regression model proposed in the main text.Footnote 26 Similarly, Decter-Frain Reference Decter-Frain(2022) finds that a multinomial logistic regression improves upon BISG just as well as tree-based prediction models (for example, Bayesian Additive Regression Trees).
Concept of interest
To date, the use of racial predictions in social science has predominantly relied on administrative data such as Census and voter files, focusing primarily on self-reported identity as the ground truth. However, racial identity is comprised of a complex dynamic involving self-identification and ascription (Huddy Reference Huddy2001; Lee Reference Lee2008), a dimension which prior studies have largely overlooked.
As elaborated in Appendix E, the validation of our method using a local elections data set heavily relies on the ethnoracial identity of elites perceived by annotators. Therefore, by operationalizing perceived identity as the ground truth, our method also focuses on an additional aspect of identities – ascription – that is worthy of further exploration. This perspective is especially relevant in studies of elites, where candidates and elected officials often strategically use different identities to signal their proximity to various constituents (Sriram and Grindlife Reference Sriram and Grindlife2017; Kreiss, Lawrence and McGregor Reference Kreiss, Lawrence and McGregor2020; Janusz Reference Janusz2021; Burnett and Kogan Reference Burnett and Kogan2022). For example, former governor of South Carolina, Nikki Haley, listed her race as white on her 2001 voter registration form, even though she publicly identifies herself as Asian or Indian American (Fraga Reference Fraga2016a). Similarly, Stephanie Murphy, the first Vietnamese-American woman to be elected to Congress – who consistently presented herself as Asian – would receive a ‘white’ classification under BSO due to her English-sounding name. This discrepancy is most pronounced among Black politicians whose names often obscure their public portrayal. For example, both Representative Hakeem Jeffries (D-NY) and Senator Tim Scott (R-SC) are predicted as Black in the FairFace model, while BSO predicts them as white. Thus, using perceived identity as the ground truth may offer a more accurate reflection of both how politicians present themselves in public as well as how voters perceive politicians.
Still, integrating two distinct definitions of racial identity into a unified classification outcome implies that this method may not be optimal for studies aiming to isolate one particular facet of racial identity. For example, research on substantive representation among minority legislators critically hinges on the assumption that these legislators self-identify as members of racial or ethnic minority communities (Broockman Reference Broockman2013; White, Nathan and Faller Reference White, Nathan and Faller2015). Therefore, while our method introduces new dimensions of identity to the imputation literature, the selection of a prediction model should carefully account for the particular concept of identity that the researcher intends to explore.
Practical Considerations
Our proposed method stands as one among numerous existing predictive tools, each accompanied by unique advantages and disadvantages. In this section, we explore both the benefits and limitations of the proposed method vis-à-vis other commonly used techniques in terms of applicability and practicality.
Applicability
The proposed method broadens the set of prediction tools that can be used in the USA. With the recent implementation of differential privacy in the U.S. Census, racial composition data at a granular geographic level (for example, block) may no longer be reliable for future research (U.S. Census Bureau 2023). Newer approaches, such as those explored here, may be promising tools to address these challenges.
The proposed hybrid method may also be useful in contexts outside the USA, where micro-level data are difficult to collect. Across the globe, detailed data on the ethnic or racial composition of the population are scarce, while large collections of names (for example, phone books and voter records) are ubiquitous (Harris Reference Harris2015). A United Nations review of data collection on race and ethnicity demonstrates a dearth of ethnicity data, particularly in Africa, where over half of the surveyed countries did not report data on ethnic backgrounds (UNDC 2003), and in Europe, where some countries prohibit collecting data on race and ethnicity (Escafré-Dublet and Simon Reference Escafré-Dublet and Simon2011; Simon Reference Simon2013). In such instances, names and/or images are often the only publicly available resources to generate predictions and answer pressing questions related to group disparities in government. In fact, recent studies have adopted various name-based algorithms to predict race and ethnicity in countries including UK (Webber Reference Webber2007; Kandt and Longley Reference Kandt and Longley2018), the Netherlands (Bodewes, Agyemang, and Kunst Reference Bodewes, Agyemang and Kunst2019), Scotland (Lakha, Gorman, and Mateos Reference Lakha, Gorman and Mateos2011), France (Mazières and Roth Reference Mazières and Roth2018), Kenya (Harris Reference Harris2015), and Brazil (Monasterio Reference Monasterio2017), to name a few.Footnote 27 In Appendix I, we demonstrate the applicability of our proposed method by applying it to members of the UK Parliament.
Cost and Time
In terms of cost, the Bayesian approach relies on Census data available to the public, as well as names included in voter files. The proposed method uses publicly available images, as well as an open-source CNN model, to increase precision without incurring significant data collection costs. However, training the multinomial model requires annotating true ethnorace for a portion of the desired data set. While researchers may do so themselves at zero cost, crowdsourcing the annotation may generate the training set much more quickly at a relatively low cost.Footnote 28
In terms of implementation time, the Bayesian method takes less than one minute to predict 1,000 names using the fBISG method via wru whereas FairFace predicts 1,000 images in just under three minutes. Additionally, the hybrid method requires extra time to collect images: our data collection record shows that it takes about thirty seconds to collect the twenty most recent images while simultaneously checking for whether faces are present, as well as removing duplicates. However, as discussed under the Section ‘Assumptions’, using as few as five images per name continues to maintain superior accuracy rates over other methods. Hence, it is possible to increase speed with only a marginal loss in accuracy. To facilitate the use of the hybrid method, we also provide step-by-step instructions and open-source Python scripts.
Ethical Guidelines
Even though the proposed method is geared towards public figures, a group that has historically been subject to weaker human subjects protections, scholars should always be mindful of ethical concerns regarding the use of administrative and public data for research purposes. On the one hand, using potentially sensitive information, such as names and addresses in BISG/fBISG (and images in convolutional neural networks), can lead to accuracy gains. However, such sensitive information can be exposed in a data breach, putting human subjects at risk. On the other hand, maximizing human subjects’ protection at all costs may lead to sub-optimal accuracy in domains where accuracy is crucial. For example, one could redact any personally identifiable information from administrative data sets. However, this approach would render predictions impossible and could have detrimental consequences when using such statistics in domains like public policy, redistricting, or other settings that demand accurate estimates of group behaviors.
The balance between accuracy and human subjects protection is a decision that ought to involve various stakeholders, including institutional review boards (IRBs), researchers, and communities under study. In ongoing research, we find that the general public views research involving ethnoracial classifications of elites as more ethically permissible than research classifying individual voters, and that accurate research designs are seen as more ethically permissible than those presenting a less accurate view of racial inequality (Lee and Velez Reference Lee and Velez2023a). While low-stakes research without direct policy implications may entail a stronger justification for the use of administrative information or public images, research that directly informs public policy and litigation might require a different approach, given the critical importance of accurate descriptive estimates.
Whether the data are collected by administrative units (for example, names and addresses) or publicly available (for example, images), stringent data security plans are necessary. We strictly followed data collection procedures approved by the IRB. In Appendix A, we describe our data collection procedure and how confidentiality was maximized by using anonymized identifiers and deleting images immediately after predictions were generated. Researchers interested in applying the proposed method to a more general sample must make explicit the ethical concerns and weigh possible benefits against potential social or political costs in advance. They should also abide by applicable laws in their country governing the use of biometric information by researchers.
Discussion
The stakes of accurately estimating populations are high. Policies may be justified on the basis of descriptive statistics about protected classes. Litigation involving voting rights depends on the best available data about joint distributions between race and voting behavior. Redistricting depends on precise counts of populations. Naturally, the importance of obtaining estimates that faithfully represent reality privileges the need for accuracy when addressing questions of inequality.
By combining the predictive power from the surname-only Bayesian method with an image-based prediction model, our proposed method increases predictive accuracy compared to other techniques widely used in the literature and may minimize bias due to geography-induced systematic measurement error. Our method yields an overall classification success of 0.93 or greater across the four main ethnoracial categories – the highest among all other prediction methods evaluated. It also recovers point estimates that are closest to benchmarks based on hand-coded data. While we choose particular models to generate features in our application (that is, FairFace for image-based and BSO for surname-based predictions), we leave the choice of particular CNN models or alternative name-based methods other than BSO to the discretion of researchers.Footnote 29
The selection of prediction methods involves balancing theoretical, statistical, and practical considerations. The hybrid method excels in overall accuracy, independence from geolocation data, and applicability to contexts where Census data may be unavailable but may demand more computational resources. Furthermore, the proposed method proves effective in mitigating bias in downstream analyses, a concern that arises in contexts where predictions are used as outcomes that correlate with geography-specific covariates. Nonetheless, geography-dependent prediction methods maintain high accuracy rates and remain valuable outside of these scenarios.Footnote 30
Researchers may find this method useful for addressing a variety of questions related to ethnic politics, including minority candidate emergence, patterns of descriptive representation, campaign finance, and coalition-building for ethnic minority candidates. Moreover, as the proposed method utilizes both perceptual and self-prescribed indicators of identity, it opens avenues to examine the interplay between the two ingredients of identity and how they might affect political outcomes at the level of individuals and institutions.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0007123424000772.
Data availability statement
Replication data for this paper can be found at https://doi.org/10.7910/DVN/KRTVHD. Python scripts to implement the hybrid method can be found at https://github.com/DianaDaInLee/PredEthnorace.
Acknowledgements
This study was approved by IRB at Columbia University (IRB-AAAT8619). The data collection was supported by the Russell Sage Foundation. We would like to thank Yuki Atsusaka, James Bisbee, Naoki Egami, Donald P. Green, Macartan N. Humphreys, Santiago Olivella, Michael Peress, Chris Warshaw, Justin de Benedictis-Kessner, Dan Thompson, and participants at PolMeth XXXIX for helpful comments.
Financial support
The data collection was supported by the Russell Sage Foundation.
Competing interests
None.
Open access
