THEORY AND HYPOTHESES

We draw our hypotheses from two distinct literatures, one on colorism and the other on descriptive representation, and we discuss them serially below. We also discuss a possible refinement to the theory of descriptive representation—that it operates with greater force amongst politically mobilized social groups.

STUDY DESIGN

Parliamentary and state assembly elections in India are low-information exercises (Chandra Reference Chandra2004). As a result, candidate posters and images are strong cueing instruments.Footnote ¹⁵ In India, voters find out about candidate appearance through campaign rallies and meetings, as well as through election materials, including billboards, posters, pamphlets, badges, television appearances, and newspaper ads. In fact, for a substantial portion of the electorate, those who are still unable to read, candidate appearance may hold even greater importance in terms of informing voting preferences. We test these colorism, descriptive representation, and mobilization hypotheses using a survey experiment.

We chose to proceed with a survey experiment to mitigate biases in the estimated effect of candidate skin-tone preferences on electoral choices associated with reverse causation and incidental association.Footnote ¹⁶ Such techniques have become commonplace in political science because of their ability to reveal causal effects, their ease of implementation, and the fact that they avoid many of the pitfalls of cross-sectional data (Gaines, Kuklinski, and Quirk Reference Gaines, Kuklinski and Quirk2006). We also chose a survey experiment because such an approach would allow us to avoid social desirability bias, which often arises when sensitive topics like race, ethnicity, or religion are involved.Footnote ¹⁷ Though skin color bias is non-racial in India, individuals might seek to hide their skin color biases if asked directly about them.

We conducted this survey experiment in DelhiFootnote ¹⁸ in the lead up to the 2013 state assembly elections.Footnote ¹⁹ To create our sample of respondents, we first randomly selected three polling stations from South Delhi, a relatively affluent area, and three from East Delhi, a relatively poor area. Next, we randomly selected 120 individuals from the electoral roll for each polling station. Sampling in this manner allowed us to capture a random distribution of the population across the six polling stations.Footnote ²⁰

To manipulate candidate skin color we used a file photograph of an Indian manFootnote ²¹ who had a wheatish complexion. We then used software to create a fair and dark version of this same photograph, as seen below (Figure 1).

Figure 1.

Experimental manipulation of candidate skin color

We used all three of these photographs in conjunction with our survey, first randomly sampling the electorate from the Delhi electoral rolls, as described above, and then ensuring that the particular photograph each respondent saw was randomly assigned, with one-third of the respondents from each constituency being assigned to each photograph. For each polling station, we used six interviewersFootnote ²² and each interviewer randomly received a list of 20 voters,Footnote ²³ with each polling station being assigned two lists for each treatment photograph.

During survey administration, interviewers asked respondents to look at a photograph of an individual who was “thinking about running for office in [their] district.” Interviewers told respondents that the individual in the photograph had not chosen a party affiliation yet, something that is common in India. Interviewers then read each respondent this statement purportedly from the candidate: “Corruption is damaging everything in our society. I promise honesty and transparency while serving you.”Footnote ²⁴ Our survey did not reveal the candidate's caste, either directly or more indirectly through the candidate's last name. Then, after respondents had seen the potential candidate's photograph and been presented his statement, interviewers asked respondents if they would vote for this individual. Our survey also asked for each respondent’s basic demographic information, including caste/subcaste, income, age, and educational background. During the survey, our interviewers discreetly categorized the skin color of each respondent as “fair,” “wheatish,” or “dark,” the very same categories that most Indians routinely use to identify skin color.Footnote ²⁵

DATA ANALYSIS

We use regression analysis to measure skin color preference. Our dependent variable is an indicator that the respondent would support the candidate he or she was shown.Footnote ²⁶ Regression analysis permits us to control for respondent characteristics that were not balanced perfectly across treatment groups. Appendix Table A1 shows some statistically significant imbalances in respondent characteristics across treatment groups. To the extent that any of these characteristics induce greater support for candidates in general, failing to control for these imbalanced characteristics would bias estimates of the relative treatment effects.Footnote ²⁷

We run logit regressions of two forms. First, to test for color preferences within any given sample, we run:

(1) $$Z = {{\rm \beta} _0} + {{\rm \beta} _{FC}}FC + {{\rm \beta} _{DC}}DC + {\bf \gamma} ^{\prime}{\bf x} + e;$$

where Z is the latent propensity to vote for the candidate, and e is the logistically distributed error. FC and DC are dummy variables indicating that the respondent was shown a fair or a dark candidate respectively (the wheatish candidate is the excluded category), and x is a vector of respondent characteristics. These characteristics include dummies capturing the respondent's caste (Other Backward Castes are the excluded category), class (middle class is the excluded category), complexion (wheatish is the excluded category), sex, and literacy status, as well as age. When β _FC is positive and statistically significant, this indicates a preference for a fair over a wheatish candidate.Footnote ²⁸ Similarly, β _DC < 0 indicates a preference for a wheatish over a dark candidate. A preference for fair over dark candidates is indicated by β _FC  − β _DC > 0, which we test post-estimation using the delta method. Colorism (Hypothesis 1) predicts three statistically significant pairwise comparisons β _FC > 0, β _DC < 0, and β _FC  − β _DC > 0 in any given subgroup of respondents.

Our descriptive representation and mobilization hypotheses concern differences in preferences across subgroups of respondents. We examine them by running regressions with interactions of the form:

(2) $$Z = {{\rm \beta} _0} + \left( {{{\rm \beta} _{FC}} + {{\rm \delta} _{FC}}{\,\ast\,}{D_G}} \right)FC + \left( {{{\rm \beta} _{DC}} + {{\rm \delta} _{DC}}{\,\ast\,}{D_G}} \right)DC + {\bf \gamma} ^{\prime}{\bf x} + e;$$

where D _G is a dummy variable indicating that the respondent belongs to some group (G) hypothesized to have different color preferences. When this group consists of voters who are Dalit, poor, or dark skinned, descriptive representation (Hypothesis 2) predicts that δ _DC > 0 and δ _FCR  − δ _DCR < 0: respondents belonging to darker groups have stronger preferences for (or weaker biases against) dark relative to wheatish or fair candidates respectively, than do respondents who do not belong to darker groups. Similarly, mobilization (Hypothesis 3) predicts that δ _DC > 0 and δ _FCR  − δ _DCR < 0 when this group consists of voters who are Dalit or of voters who are poor.

Table 1 presents the coefficients from three regressions estimated on the full sample. The first is the baseline model (Specification 1), which tests for colorism in the aggregate. The second and third are interaction models (Specification 2) which we introduce to test the descriptive representation and mobilization hypotheses. In addition to testing the significance of the differences between groups in color preference of respondents using these interaction models, we also examine the significance of color preferences within groups by running the baseline model on subsamples of respondents from particular groups. To reveal the magnitude of color preference, we present not just the logit coefficients, but also the marginal effects of candidate skin tone. Specifically, Table 2 presents the differences in the probabilities that a respondent with average characteristics for the subgroup would vote for candidates of different skin colors (columns 3–5). To provide a sense of the relative electoral advantage conferred by these marginal effects, Table 2 also provides the fraction of respondents in each subgroup willing to support whichever candidate they were shown (column 2). The marginal effects in the top row of Table 2 are calculated post-estimation from regression 1 in Table 1. The remaining marginal effects in Table 2 come from analogous regressions run on subsamples of respondents, whose coefficients are not displayed in the interest of brevity.

Table 1.

Logit regression estimates (full sample)

Note: Dependent variable is the respondent's willingness to vote for the candidate. ***p < .01, **p < .05, *p < .1. First regression follows Specification (1). Second and third regressions follow Specification (2).

Table 2.

Differences in approval rates by candidate skin tone pairs (marginal effects)

Note: Dependent variable is the respondent's willingness to vote for the candidate. ***p < .01, **p < .05, *p < .1.

^a The marginal effects in each row are calculated following a logit regression of Specification (1), but using the indicated subsample. Where the subsample all belongs to one caste, class, or complexion group, the unidentifiable dummy variables are dropped.

^b A respondent is a member of a darker group if they are Dalit, poor, or dark-skinned.

^c A respondent is a member of a politically mobilized group if they are Dalit or poor.

Our full sample, baseline regression (Table 1, first regression) indicates a marginally significant preference for fair over wheatish candidates in the sample as a whole (β _FC > 0), but no preference between wheatish and dark candidates. There is also a marginally significant preference (not shown) for fair over dark candidates (β _FC  − β _DC > 0). These coefficients translate into substantial marginal effects, with fair candidates enjoying roughly an 8 percentage-point higher support rate than wheatish or dark candidates (Table 2, top row, columns 3 and 5). Given that 23.5% of the candidates in the full sample are supported (Table 2, top row, column 1), this indicates that support for the fair candidate is roughly one-third higher than support for the other candidates. This suggests, consistent with colorism (Hypothesis 1), that fair skin tone can be a major electoral advantage. This preference for fair candidates appears to be fairly ubiquitous, though not always statistically significant: at a point estimate, the fair candidate is preferred to the dark candidate in every subsample save for among the poor, and to the wheatish candidate in every subsample (Table 2, rows 2–10).

Nevertheless, the results are not fully consistent with strict colorism (Hypothesis 1). Specifically, contrary to the predictions of colorism, there is no significant preference for wheatish over dark candidates in the full sample (Table 1, regression 1; Table 2, row 1) or in any subsample (Table 2, rows 2–10). Interestingly, this is at variance with the significant evidence of preference for wheatish over dark skin tone in other social contexts such as marriage.

The marginal effects in rows 2–10 of Table 2 suggest that this exception to colorism occurs because preferences for wheatish over dark candidates are heterogeneous across social groups, and where they are significant, actually favor the dark candidate. For example, among Dalit and poor respondents the dark candidate receives support rates over 15 percentage points higher than the wheatish candidate.Footnote ²⁹ Given overall support rates in these two groups of 20% and 28.9%, this implies that wheatish complexion carries a substantial penalty in these two groups. Conversely, upper caste and middle class respondents display statistically insignificant, but noticeable preferences for wheatish over dark candidates.

The results so far establish a fairly ubiquitous preference for fair candidates, but heterogeneous rankings of the wheatish over the dark candidate across social groups. This motivates us to more carefully examine differences in color preferences between respondent groups to see whether they are consistent with expectations if descriptive representation (Hypothesis 2) and mobilization (Hypothesis 3) are important. As we argued earlier, descriptive representation could influence skin color preference amongst Dalit and poor voters, if these groups are darker skinned. Table 3 confirms that they are. It shows that, in the first-order stochastic dominance sense, Dalit respondents are darker than non-Dalit respondents, and poor respondents are darker than non-poor respondents. Thus, in addition to dark respondents, we consider poor and Dalit respondents to be members of “darker groups.”

Table 3.

The darker groups

Note: Analysis of the distribution of caste, class, and complexion in the full sample of 599 respondents.

To see whether descriptive representation can help to make sense of respondents’ preference orderings, the second regression in Table 1 tests the significance of differences in color preference between respondents who are poor, Dalit, or dark-skinned, and those who are neither. While none of these differences in color preference are statistically significant, rows 11 and 12 of Table 2 suggest differences in color preference between respondents who do and do not belong to those darker groups. The results among respondents who do not belong to this darker group are all consistent (at a point estimate) with strict colorism, with the fair candidate significantly preferred to the dark one. Conversely, those in the darker groups prefer the fair to the wheatish candidate, but (at a point estimate) prefer the dark over the wheatish candidate. These results are consistent with darker groups preferring descriptive representation. Together with the statistical insignificance of differences in color preference between groups, this suggests some, muted role for descriptive representation.

Finally, we cut the data in two ways to see whether the effects of descriptive representation on color preferences are limited to politically mobilized categories (Hypothesis 3). This could account for the mixed evidence on Hypothesis 2. First, in the final row of Table 2, we look at color preference among the politically mobilized (poor or Dalit) respondents and find significant preferences for the fair and dark candidates over the wheatish candidate. Second, the third regression in Table 1 uses an interaction model that allows treatment effects to vary depending upon whether the respondent is Dalit, poor, or dark (or some combination of the three). This allows us to test for color preferences in the eight non-overlapping sub-samples provided in Table 4.Footnote ³⁰ We note, from the interaction term coefficients in Table 1, that the Dalit respondents have significantly stronger preferences for the fair and dark candidates over the wheatish candidate than do otherwise identical but non-Dalit respondents. Similarly, using the delta method and these same coefficients, we find that poor respondents have significantly (p = .090) less preference for fair over dark candidates than non-poor candidates (δ _FCR  − δ _DCR  = −0.107 − 0.745 < 0). Thus, preference for dark candidates is stronger amongst respondents belonging to the politically mobilized groups—Dalits and the poor. In contrast, being dark, if anything, reduces preference for dark over fair or wheatish candidates (δ _FCR  − δ _DCR > 0, but is insignificant).

Table 4.

Color preference within well-defined subgroups

Note: Dependent variable is the respondent's willingness to vote for the candidate. ***p < .01, **p < .05, *p < .1. Differences in latent propensity (Z) are calculated post-estimation from the third regression in Table 1, which uses Specification (2).

Table 4 provides post-estimation tests of color preferences within our eight sub-samples based on this same regression. For analytical simplicity, it presents preferences using combinations of coefficients, rather than marginal effects. The table provides further indications that color preferences depend significantly on whether the respondent is Dalit and/or poor. Specifically, preference for the fair over the dark candidate is limited to respondents who are neither Dalit nor poor. Even dark respondents who are not Dalit or poor prefer the fair to the dark candidate. Dalit respondents prefer the fair to the wheatish candidate, regardless of whether they are poor or dark. The only group to discriminate between the wheatish and dark candidates is those who are both Dalit and poor (but not dark). This group prefers the dark candidate to the wheatish one. All of the above suggest that absent mobilization on the basis of skin color, being dark is not enough to produce the impulse for descriptive representation and overcome colorism.Footnote ³¹ There is therefore support for Hypothesis 2 only among the politically mobilized categories, which confirms Hypothesis 3.

DISCUSSION AND IMPLICATIONS

Our survey experiment shows that evidence for colorism, a general preference for fairer skin over darker skin, is weak in electoral politics in India. This is despite fair skin tone being venerated and dark complexion stigmatized in many aspects of private and public life, including marriage, religion, cinema, and television.

We find that there is weak support for colorism in electoral politics in India not because people do not discriminate based upon skin tone in electoral contexts, but because skin color preferences are heterogeneous. To be sure, there is substantial support for candidates of fair skin tone across all subsets within our sample. But this expected preference is matched, at least among some groups, by relatively strong support for candidates with dark complexion. This unexpected support for dark-skinned candidates comes from groups that have been marginalized within Indian society, namely Dalits and the poor. When we look at individuals who fit into both of these categories, Dalits who are poor themselves, we find this effect to be particularly pronounced. Indeed, among this subset of our sample, dark skin tone is preferred to wheatish skin color in electoral contexts. Our findings have important implications.

But first, what accounts for the seemingly bimodal support pattern we see among Dalits and the poor? While our main hypothesis, driven by colorism, predicts a preference for fair skin tone over both wheatish and dark complexions, our two alternative hypotheses, both motivated by descriptive representation theory, do allow for heterogeneity of skin color preferences. Even within a subgroup like the poor, colorism might drive some individuals to choose a fair candidate, while others who have been mobilized along class lines might see dark skin tone as a class cue and support a dark candidate. The latter behavior could reflect respondents’ desire to have their economic needs represented by someone who they perceive to be similar to them. Taken together, our results suggest that the impulse to prefer fair skin tone can coexist with a desire to seek descriptive representation from a candidate with dark skin color.

Our work suggests that political mobilization can be important in the process of making skin color politically salient. Among darker groups, only those that have been politically mobilized express similar support for dark and fair candidates. Individuals with dark complexion alone, an identity that has not been politically mobilized in India, express preferences that are consistent with colorism. It is most interesting that Dalits and the poor appear to treat dark skin color as an implicit marker for caste and socioeconomic status even though these groups are not mobilized through the use of overt skin-color-based appeals. This behavior is explicable, however, as our data also indicate that members of these groups are darker, on average, than their counterparts from higher caste and socioeconomic backgrounds. Thus, if they desire descriptive representation and have little means by which to determine which candidate is most “like them” then expressing a preference for dark skin color is one way to proceed.

This desire for descriptive representation among Dalits and the poor is likely rooted in both their recognition by the state and their mobilization by political parties as Dalits and as poor citizens. Skin color then becomes salient for these groups because it is connected to their politically mobilized identities. Dark complexioned voters who share the same skin color with the dark candidate, by contrast, show little support for him. Their inability to identify with the dark candidate is rooted in the fact that skin color on its own does not constitute a political identity; it is neither recognized by the state, nor mobilized by political parties, despite the discrimination that many dark-complexioned individuals face in India on a regular basis.

It follows from the heterogeneity of preferences described above, that fair candidates enjoy a skin-color-related advantage, while dark skin tone is not a disadvantage for a candidate running in a constituency that has either a large Dalit or poor population. Both in Delhi and in India, Dalits and the poor form a substantial portion of the electorate. In a democratic context where numbers matter and where the poor turn out to vote, support for dark candidates among the poor is not a trivial finding. It suggests that dark skinned candidates can get elected to office despite the evidence of colorism we find among certain subsets of the population. Our analysis of the Delhi state legislature and national parliament confirms that there are a substantial number of dark-complexioned candidates in elected office in India.

This descriptive development is not purely symbolic. We believe it also has a substantive implication. We think that the election of dark politicians to office associates dark skin color with power and prestige, and as a result, gradually, it may reduce the longstanding stigma associated with dark skin tone. Social science research has begun to point to similar political-office-related effects for Dalits, another group trapped in a historically stigmatized identity (Chauchard Reference Chauchard2014; Hardtmann Reference Hardtmann2009).

Finally, we note that this is an exploratory study, the first of its kind in India. Our goal was to probe the degree to which candidate skin color influences voter choices in a non-racial society. Our results therefore highlight important questions for future research. In particular, we note the substantial overlap between caste, socioeconomic status, and skin color categories. Future studies will require larger samples to determine more conclusively which of these collinear factors drive voter preferences. We are also left wondering whether the bias against Dalit candidates that numerous scholars have highlighted is driven by caste alone, or whether the fact that Dalits, on average, have darker skin color than upper caste individuals also contributes to it. Finally, we wonder about the degree to which our results travel. Does the bias in favor of fair skin manifest similarly or differently in rural and urban areas, in North and South India, and for male and female candidates? We are currently gathering data that should allow us to address some of these questions in future publications.