Experiments can help scholars to explain how individuals’ identities shape their political behavior. We aim to draw attention to risks identity scholars face when placing identity-related covariates in an experimental design. Helpful work exists on this practice (e.g., Acharya, Blackwell and Sen Reference Acharya, Blackwell and Sen2016; Blackwell Reference Blackwell2013; Montgomery, Nyhan and Torres Reference Montgomery, Nyhan and Torres2018), but does not speak to the unique setting of identity politics, wherein key covariates can exert consequential priming effects.

Best practices for analyzing experiments warn against conditioning results on posttreatment covariates. As Acharya, Blackwell and Sen (Reference Acharya, Blackwell and Sen2016) explain, “conditioning on a posttreatment variable changes the quantity of interest from an overall average treatment effect to a direct effect of the treatment net the posttreatment variable” (514). Montgomery, Nyhan and Torres (Reference Montgomery, Nyhan and Torres2018) recommend measuring covariates before the treatment using a panel or “before the experimental manipulation during a single survey” (773). They state that “even variables that seem likely to remain fixed when measured after treatment, such as measures of racial or partisan identification, can be affected by treatments”.Footnote ¹

Measuring covariates posttreatment can indeed create analytical problems. However, experimentalists should not always measure identities pretreatment either. Instead, researchers must base this decision on case-specific theory regarding the relationship between the treatment and measure of identity, and an explicit trade-off of the risks of posttreatment bias and priming effects.

Priming occurs when a consideration becomes accessible and receives extra weight when forming subsequent evaluations (e.g., Mendelberg Reference Mendelberg2008). Countless studies document the ease of priming and its consequences. Transue (Reference Transue2007) finds asking Americans about their national identity substantially influences their support for a tax increase (see also Sniderman, Hagendoorn and Prior Reference Sniderman, Hagendoorn and Prior2004). Morris, Carranza and Fox (Reference Morris, Carranza and Fox2008) ask respondents for their political identification either at the beginning or at the end of a survey. The former design leads Republicans to choose higher-risk investment preferences and Democrats to choose lower-risk preferences. Klar (Reference Klar2013) asked respondents to consider their partisanship before evaluating policies, finding that even the weakest partisan primes significantly change policy evaluations.

Identity primes can even contaminate attitudes that seem unrelated to identity-based interests. Williams et al. (Reference Williams, Turkheimer, Magee and Guterbock2008) randomly asked respondents a question about their race either before or after completing a survey on public health attitudes. The prime influenced their responses about health and hygiene.Footnote ² Benjamin, Choi and Strickland (Reference Benjamin, Choi and Strickland2010) randomly assigned respondents to complete a questionnaire about either their linguistic/immigration history or their cable subscriptions. Asian subjects who received the linguistic/immigration questions displayed significantly different reward-seeking behaviors, as did African-American subjects in a replication.

Identification comprises both group membership (e.g., I am a woman) and the strength of that identification (e.g., My gender is important to me). Reports of group membership are highly stable and we know of no experiments that change a respondent’s reported self-categorization regarding their gender, racial, or religious self-classification (i.e., from one category to another). While certain contexts can persuade Independents to report a partisan leaning (e.g., Craig and Richeson Reference Craig and Richeson2014; Klar and Krupnikov Reference Klar and Krupnikov2016) and the intensity with which partisans identify with their group is malleable (e.g., Abrams et al. Reference Abrams, Wetherell, Cochrane, Hogg and Turner1990; Lupu Reference Lupu2013), we know of no treatments that lead partisans to change their party preference (e.g., from Democrat to Republican or vice versa). This is consistent with the tremendous stability of partisan classifications (Green, Palmquist and Schickler Reference Green, Palmquist and Schickler2002). It is thus crucial to consider the specific type of identity measure when considering where to place identity-related covariates. If the treatment influences the measure, then posttreatment indicators bias estimates of the intended causal effect. Pretreatment measures, on the other hand, may change the definition of the causal parameter being estimated from the effect of the treatment when identity is non-salient to the effect when it is salient, which may not be the effect the experiment is interested in detecting.

One solution is to simply place covariates in a distinct wave using panel data. However, this is infeasible in many settings [e.g., exit polls (Klar Reference Klar2013), rallies (McClendon Reference McClendon2014), or public events (Harrison and Michelson Reference Harrison and Michelson2016)]. Panels are also prohibitively expensive: a two-wave survey of 1000 Americans is approximately three times more costly than a one-wave sample due to attrition, re-contact, and re-incentivizing respondents. Another proposed solution is to use a list of “buffer” items intended to dilute priming effects. However, primes can persist beyond such buffers (Druckman and Chong Reference Druckman and Chong2010). Finally, it is possible to simply check for posttreatment bias by regressing posttreatment measures on indicators for each treatment (e.g., Mummolo Reference Mummolo2016). Yet, this modified form of a balance test also poses its own risks with respect to Type I error (see Mutz, Pemantle and Pham Reference Mutz, Pemantle and Pham2019).

We recommend that identity scholars explicitly base their design choice on case-specific theory concerning the relationship between the treatment and the measurement of the identity. Social identity approaches, for instance, can be helpful in considering cases when an identity may be affected by a treatment. Placing identity questions on the posttest is problematic when the treatment threatens the “value” or “distinctiveness” of the identity, as this can impact the self-categorization, and behavior, of high and low identifiers in offsetting ways (e.g., Branscombe et al. Reference Branscombe, Ellemers, Spears, Doosje, Ellemers, Spears and Doosje1999). However, if the treatment does not target the value of an identity, and that identity has no demonstrated history of instability, then it is defensible to place it on the posttest. Table 1 provides a schematic for thinking through this process. It is up to the researcher to theorize, based on existing literature and, if possible, pretests, whether (i) the covariate is susceptible to being influenced by the treatment and (ii) the covariate might contaminate the study via priming. Depending on each consideration, identity scholars should carefully decide whether covariates should appear pre- or posttreatment.

Table 1 A Framework for Measurement Order Decisions

Demographic and social identities can be consequential in any setting. Identity scholars must consider the causal order of variables and the risks of each form of bias, rather than following conventional wisdom that might not apply to them.