Manipulations checks are postexperimental measures aiming at “ensuring that an experiment actually has been conducted (i.e., that the Independent Variable has been effectively manipulated)” (Sansone et al. Reference Sansone, Morf and Panter2003, p. 244). They typically take the form of comprehension questions immediately following the experimental treatment to check that subjects paid attention and understood the treatment (Kane and Barabas Reference Kane and Barabas2019). The inclusion of manipulation checks enters standards of best practices in experimental political science (Mutz and Pemantle Reference Mutz and Pemantle2015). They are particularly important to avoid Type II error – i.e. the false negative conclusion that the research hypothesis is wrong while it is actually true – in case of null results.

A widespread practice is to exclude participants who failed manipulation checks in order to limit the analyses to subjects who understood the experimental prompt (Aronow et al. Reference Aronow, Baron and Pinson2019). However, some studies have warned that this may bias the analyses (Aronow et al. Reference Aronow, Baron and Pinson2019; Berinsky et al. Reference Berinsky, Margolis and Sances2014; see also Montgomery et al. Reference Montgomery, Nyhan and Torres2018). Berinsky et al. (Reference Berinsky, Margolis and Sances2014) showed that individual responses to screeners correlate with a range of personal characteristics. Dropping inattentive subjects may distort the sample to certain “races, ages, and levels of education”. More problematically, Aronow et al. (Reference Aronow, Baron and Pinson2019) demonstrated that this may bias the estimation of causal effects by creating an asymmetry between experimental arms.

This study offers a new illustration on how dropping subjects who failed manipulation checks may bias experimental results. Aronow et al. (Reference Aronow, Baron and Pinson2019) presented an illustrative experiment in which dropping subjects lead to underestimating the effect size of the treatment. The present study presents another experimental case in which dropping subjects increases the risk of Type I error when testing a hypothesis of interest – i.e. drawing a false positive conclusion that confirms the research hypothesis while it is actually wrong.

The experiment was conducted in an online survey filled during April 2019 by nationally representative samples from Denmark, France, Germany, Italy, Spain, and the Netherlands. A total of 3949 subjects participated in the experiment, based on the economic threat manipulation from Stenner (Reference Stenner2005). Subjects were randomly assigned to one of two short fictional news stories about the national economic context, respectively, depicting an improving situation (prosperity) or a worsening situation (decline) ^{Footnote 1} . The initial purpose of the experiment was to test whether subjects express more nostalgia after the decline treatment compared to the prosperity treatment.

Following the treatment, subjects answered a manipulation check question: “According to the news story, the national economic situation is: Worsening/Stable/Improving”. Only 63% of subjects provided the correct answer regarding their experimental treatment – i.e. “Improving” in the prosperity treatment and “Worsening” in the decline treatment. More problematically, subjects who failed did apparently not respond at random, as shown in Figure 1. 30% of subjects declared that, according to the news story, the economic situation was improving, 27% that it was stable and 43% that it was worsening.

Figure 1 Share of subjects by experimental treatment and by response to the manipulation check question (with 95% Confidence Interval).

Prior to the experiment, subjects answered a question about their own perception of the economic situation: “Would you say that the economic situation now is better or worse to how it was 5 years ago?”. Subjects responded with a 11-point scale from 0 (“Much worse”) to 10 (“Much better”). Results from a two-way ANOVA indicate that responses to this question are significantly related to responses to the manipulation check, F(2,3731) = 55.79, p < .0001. As shown in Figure 2, the more favorably subjects perceived the economy, the less they responded that the news story depicted an economic decline and the more they responded that the story depicted economic stability ^{Footnote 2} . This means that the manipulation check actually captured some subjects’ preexisting beliefs about the economic situation.

Figure 2 Share of responses to the manipulation check question depending on the perception of the national economy prior to the experiment (with 95% Confidence Interval).

Now, what happens if we drop subjects who failed the manipulation check? Figure 3 presents the average perception of the economy prior to the experiment for subjects of the prosperity versus decline treatments. When including all subjects, there is no significant difference across treatments in the perception of the national economy prior to the survey experiment, t(3752) = 0.6062, p = .5444. This is what we expect from randomization: the treatment is independent from the subjects characteristics prior to the experiment. In contrast, when excluding subjects who failed the manipulation check, there is a significant difference across treatments in the perception of the national economy prior to the survey experiment, t(2366) =4.5688, p < .0001. It is impossible that the experimental treatment had a causal effect on responses to a question asked earlier in the survey. Thus, this reflects a selection effect emerging from dropping subjects who failed the manipulation check.

Figure 3 Average perception of the national economy prior to the experiment depending on the experimental treatment (with 95% Confidence Interval).

Suppose that we want to test the effect of the treatment on a dependent variable. After the treatment, the subjects’ level of nostalgia was assessed. Subjects indicated on a 5-point scale from “strongly disagree” to “strongly agree” to what extent they agreed that “the society used to be a much better place”. As shown by regression models in Table 1, when dropping subjects who failed the manipulation check (model 1), one would conclude that the decline treatment had a significant positive effect on nostalgia. Yet, this effect is reduced when controlling for the preexisting perception of the economy (model 2). Since it is impossible to measure all potential preexisting characteristics of subjects selected through the manipulation check, the best option is to avoid dropping subjects (model 3). This decision is conservative: it is likely to greatly increase the noise in the data and reduce effect sizes – with increased risk of Type II error. In model 3, the effect of the treatment is no longer significant ^{Footnote 3} . Nonetheless, this is the only way to ensure that, if some treatment effect is observed, it is of genuinely causal nature.

Table 1 Results from linear regression models of the level of nostalgia

Standard errors in parentheses.

* p < 0.05, ** p < 0.01, *** p < 0.001.

The present study does not advocate against the inclusion of posttreatment manipulation checks. These can be informative tools – especially in the development phase of experiments – to assess the degree of attention and comprehension of the treatment in the given type of sample. In the present study, the manipulation check reveals that a very large fraction of subjects were inattentive. One possibility is that the pretreatment question about the subjects’ perception of the national economy induced subjects to disregard the content of the experimental vignette. This would explain the high rate of failure in the responses to the manipulation check and their close correlation with subjects’ preexisting beliefs about the economy. To test fort this, it would be necessary to have a control group for whom the initial question was not included. However, this limitation does not affect the overall conclusion that exclusion based on posttreatment manipulation checks must be avoided.

What are then the options available to researchers? A first option highlighted by the literature is to include pretreatment questions to gauge subjects’ attentiveness. A common tool is instructional manipulation checks – or “screeners”. Screeners are similar to classic survey questions but ask participants to ignore the standard response format and instead provide a confirmation that they have read the instruction Berinsky et al. (Reference Berinsky, Margolis and Sances2014); Oppenheimer et al. (Reference Oppenheimer, Meyvis and Davidenko2009). One disadvantage is that screeners may induce subjects to think that researchers want to trap them, which alters their responses to subsequent questions (Hauser and Schwarz Reference Hauser and Schwarz2015). Alternatively, Kane et al. (Reference Kane, Velez and Barabas2020) recently proposed ready-to-use “mock vignettes”. A mock vignette mimics common king of descriptive content in political science experiments but appear before the researcher’s treatment. All subjects read the same vignette and must then answer factual questions about it, allowing to check for subjects’ attentiveness.

The latter tools come with the cost of sacrificing survey space. Another alternative is to rely on timers as a proxy to identify inattentive subjects. Various studies highlight that subjects with short response times are generally less attentive (see for instance Börger Reference Börger2016; Wood et al. Reference Wood, Harms, Lowman and DeSimone2017) ^{Footnote 4} . Read et al. (Reference Read, Wolters and Berinsky2021) designed a method to identify inattentive subjects based on multiple question timers. Their method does not induce posttreatment selection bias when computed on question timers before the treatment. Besides, it allows to identify slow but nonetheless inattentive subjects.

Depending on the space available in survey, researchers may use these methods to perform analyses on sub-sample(s) of attentive subjects, in order to limit the risk of Type II error without inducing posttreatment bias. These measures of attentiveness may correlate with politically relevant variable, such as age, race, and education (see Berinsky et al. Reference Berinsky, Margolis and Sances2014; Kane et al. Reference Kane, Velez and Barabas2020). Thus, restricting analyses to attentive subjects comes with the risk of drawing conclusions that are not representative of the population. To mitigate this risk, the best practice should be to report estimates of treatment effects based on both the overall sample and subsample(s) of attentive subjects.