Trends in Public Stigma of Mental Illness in the US, 1996-2018

Key Points Question What changes in the prejudice and discrimination attached to mental illness have occurred in the past 2 decades? Findings In this survey study of 4129 adults in the US, survey data from 1996 to 2006 showed improvements in public beliefs about the causes of schizophrenia and alcohol dependence, and data from a 2018 survey noted decreased rejection for depression. Changes in mental illness stigma appeared to be largely associated with age and generational shifts. Meaning Results of this study suggest a decrease in the stigma regarding depression; however, increases and stabilized attributions regarding the other disorders may need to be addressed.

A post-hoc power analysis suggests that the study is adequately powered (>80%) to detect a change in proportions of 0.06 or greater across any two GSS survey modules at an alpha of 0.05. This corresponds to a Cohen's h value of 0.12, which is characterized as a very small effect size. Even within vignette conditions, where the n is smaller, we are powered at 80% to detect a change in proportions of 0.10 or greater across survey modules. This corresponds to a Cohen's h of 0.25, or a small to moderate effect size.
The Vignette Strategy. The replicable sections of the NSS employ a vignette strategy to examine public knowledge of and response to mental illness and substance abuse. A vignette strategy avoids identifying the nature of the problem to allow for data collection on knowledge, recognition, and labeling among respondents. 2 Vignettes depicting individuals with problems of living have now become standard methodological tools for social science and health researchers (see review 9 ). The vignette is presented to respondents as an enhanced or "a more elaborate stimulus" 10 to elicit normative attitudes, beliefs and predispositions to behavior towards a hypothetical person showing symptoms and behaviors consistent with a professional evaluation. Vignettes provide a concrete stimulus that helps to standardize the information that is presented to respondents and at the same time minimizes the abstract or nonspecific nature on which attitudinal questions, especially regarding mental illness, are often based. 11 They give priority to a set of specific circumstances, and, unlike statements about "a person with a mental illness," avoid evoking multiple unknown images. In studying sensitive topics, vignettes give respondents some distance by focusing on a hypothetical person, without making heavy demands on concentration. 11 Respondents were randomly assigned to one vignette, wherein the gender, race/ethnicity, and education level of the vignette character also varied randomly. Cases focused on the recent appearance of symptoms to optimize the measurement of MH literacy. Vignettes were both read aloud and handed to respondents in written form; and were followed by sets of questions on MH literacy and public stigma. The items measuring MH literacy and stigma were designed to assess respondents' problem recognition, their assessment of the underlying causes, their treatment endorsements, and their preferences for social distance. 23 for specific mental disorders. The original vignettes included depression, schizophrenia, alcohol abuse, drug dependence (dropped in later module due to a confound of a second stigmatizing status, i.e., suspicion of stealing). Finally, a control condition of "daily troubles," an individual described with day-to-day problems but reaching no DSM diagnostic criteria, was included as a control in all years. In all vignettes, the gender (male/female), race (White, African American, Hispanic), and education (8th grade, high school, college) of the vignette character is randomly varied. Respondents were randomly assigned to a single vignette, were read the vignette by the interviewer, and were given a card with the vignette printed on it. Name] has started to drink more than his/her usual amount of alcohol. In fact, he/she has noticed that he/she needs to drink twice as much as he/she used to to get the same effect. Several times, he/she has tried to cut down, or stop drinking, but he/she can't. Each time he/she has tried to cut down, he/she became very agitated, sweaty and he/she couldn't sleep, so he/she took another drink. His/Her family has complained that he/she is often hung-over, and has become unreliablemaking plans one day, and canceling them the next.

Measurement.
Respondents were read the randomly assigned vignette, given a card with the vignette printed on it, and asked questions in three broad areas, which we refer to as attributions, treatment endorsement, and stigma (see eTable 1 for a detailed description of each item). Other core items on the GSS were used to construct covariates.

Attributions.
Respondents were asked how likely it is that the person in the vignette is experiencing "a mental illness" and/or "the normal ups and downs of life," as well as how likely the situation might be caused by "a genetic or inherited problem," "a chemical imbalance in the brain," "his or her own bad character," "God's will," and/or "the way he or she was raised." Questions were not mutually exclusive, and respondents could endorse multiple attributions. Responses of "very likely" and "somewhat likely" were coded 1; "not very likely," "not at all likely," and "do not know" were coded 0. Analyses were run again with responses of "do not know" coded as missing as well as including controls for the vignette character's race, gender, and education, and substantively similar results were obtained. A biomedical conception measure was coded 1 if the respondent labeled the problem as mental illness and attributed cause to either a chemical imbalance or a genetic problem; it was coded 0 otherwise.

Treatment endorsement.
Respondents were asked whether the person described in the vignette should seek consultation with or treatment by "a general medical doctor," "a psychiatrist," "a mental hospital," and/or "prescription medications." Responses were coded 1 if the respondent said "yes" and 0 if they said "no" or "do not know." Stigma. Two sets of measures, for social distance and for perceptions of dangerousness, were used. The first asked respondents how willing they would be to have the person described in the vignette work closely with them on a job; live next door; spend an evening socializing; marry into the family; and as a friend. Respondents were also asked how willing they would be to live near a group home that serves the person described in the vignette. Responses of "definitely unwilling" and "probably unwilling" were coded 1 (i.e., stigmatizing) and responses of "probably willing," "definitely willing," and "do not know" were coded 0. The second measure asked respondents how likely is it that the person in the vignette would "do something violent toward other people" and/or "do something violent toward him/herself." Responses of "very likely" and "somewhat likely" were coded 1; responses of "not very likely," "not at all likely," and "do not know" were coded 0.
Covariates. Respondents' age (in years), sex (coded 1 for female, 0 for male), education (coded 1 for at least a high school degree, and 0 otherwise), and race (code 1 for white, 0 for other) were included as controls. In 1996, the mean age of respondents was 43 years (SD=16); 53% were female, 31% completed more than a high school degree, and 81% were white. In 2006, the mean age was 45 years (SD=17); 55% were female, 36% completed more than a high school degree, and 71% were white. In 2018, the mean age was 46 years (SD = 18), 51% were female, 41% completed more than a high school degree, and 74% were white. These profiles are broadly consistent with Census Bureau data.

Statistical analyses.
Our statistical analyses were divided into two main parts: one examining basic trends in Americans' views on mental illness and one examining age, period, and cohort processes.
Basic trends. We evaluated changes across years in moral and biomedical attributions, endorsement of treatment, perceptions of dangerousness, and preferences for social distance by comparing unadjusted percentages obtained from the 1996, 2006, and 2018 waves of the GSS ( Figure 1 and eTables 2a-f). To adjust for possible demographic shifts between survey years, we fit logistic regression models for each outcome and for each vignette condition with controls for respondents' age, sex, education, and race, and the sex, education, and race of the person described in the vignette. We then computed the difference in the predicted probabilities for a given outcome (e.g., mental illness) between 1996 and 2006, between 2006 and 2018, and over the entire time period holding the control variables at their means for each individual sample; these are referred to as discrete change coefficients and are presented graphically, as dots, in Figure 2 (see eTables 3a-f for the raw estimates and test statistics). Variance estimates were computed using the delta method. In supplementary analyses, we fit expanded models that included interactions between survey year and indicators of respondents' gender, educational attainment, and race/ethnicity.
Results from these models, which are summarized graphically in Appendix Figure A1, allow for inferences about subgroup specific time trends.
Age-period-cohort analyses. The substantive question that we seek to answer-has stigma toward mental illness changed in the US?-requires attention to age (A), period (P), and cohort (C) effects. It is well-known that disaggregation of these three dimensions is difficult due to their perfect linear relationship. [13][14][15][16] In the demographic and sociological literatures, the traditional age-period-cohort model includes age, period, and cohort as three independent variables in a statistical equation, implying that cohort effects could operate independently of age and period effects. This definition of cohort effects is arbitrary and problematic because, as Ryder (1965) 5 argues, cohort effects only occur when period effects are differential depending on the age group. This critique, which is not new, 17,18 has recently led to renewed questions about the validity of the traditional APC accounting framework. 19,20 Luo and Hodges 23 have also demonstrated that statistically, even if one is willing to accept the traditional framework's definition of cohort effects as independent and additive, the cohort effects estimated in such models are in fact a mix of age and period effects and their interactions.
In our analyses, we apply Luo and Hodges's new age-period-cohort-interaction (APC-I) model 23 to investigate the unique contribution of cohort membership to overall trends in stigma. Conceptually, the APC-I model is distinct from previous APC methods-including all forms of the classical APC accounting model-in that it defines cohort effects as the differential effects of social change (i.e., period effects) by age. Luo and Hodges argue that this conceptualization-which does not require the presence of additive cohort effects, and thus poses no challenges with respect to model identification-is better aligned with theoretical accounts of the conditions under which cohort effects occur. By explicitly modeling cohort effects as age-period interactions, the APC-I framework emphasizes the dependence of age, period, and cohort effects, as Ryder (1965) 24 originally proposed. Substantively, the cohort effect estimated by the APC-I model can be interpreted as the unique deviations associated with cohort membership in the outcome from the expected rate or score based on age and period main effects.
The basic APC-I model can be written as: where g is the link function; ( ) is the expected value of the outcome, Y, for the ith age group in the jth time period; is the mean difference from the global mean associated with the ith age category; is the mean difference from associated with the jth period; and ( ) is the interaction of the ith age group and jth period group, corresponding to the effect of the kth cohort. Under this setup, the effect of one cohort includes the multiple age-by-period interaction terms ij(k) that lie on the same diagonal in a table with ages in rows and periods in columns. We expanded Eq. (1) slightly to include sociodemographic variables (educational attainment, gender, and race) characterizing the respondent and the individual described in the vignette. Sum-to-zero effect coding is used throughout, following recommendations from Aiken and West 21 and Jaccard and Turrisi. 22 This means that all estimates have the same reference group-the next lower level in the hierarchy of main effects and interactions.
To evaluate the fit of the fully specified APC-I model, and to compare it to other candidate models (e.g., age effects only or period effects only), we used the simple three step procedure recommended by Luo and Hodges: 4 1. Perform a global deviance test. In our APC analyses, we start by asking whether there is variation in the outcome (i.e., preferences for social and interactional distance) associated with cohort membership that cannot be explained by age and period main effects. To answer this question, we fit an ANOVA model that included age main effects, period main effects, and their interactions. We then tested the variation attributable to the age-by-period interaction, with (a − 1)(p − 1) degrees of freedom. A significant global test result, which we obtained (F = 2.57; p = .005), indicates that cohort effects may be present. Note that a significant global test does not characterize cohort effects, nor is it a sufficient condition for the existence of cohort effects. In addition to the global deviance test, we also fit a series of auxiliary models, beginning with a model that included only age effects, and then working our way up to our fully specified APC-I parametrization, with age and period main effects, and an age-by-period interaction. We present fit statistics (Akaike's information criteria) for all models in Table S4. Lower values indicate better model fit (i.e., less information loss relative to the "true" model that generated the observed data). 2. Perform deviation magnitude tests. In the second step, we ask whether membership in a specific cohort matters after accounting for age and period main effects. If the deviance test rejects the null hypothesis, we can conclude that membership in that specific cohort has unique effects on the outcome variable. Our analyses produced significant deviation magnitude tests for the 1937-1946 and 1987-2000 birth cohorts.
The p-values were 0.004 (F = 3.429) and 0.023 (F = 3.186), respectively. See Table S5 for the full set of results by cohort. 3. Perform average deviation tests. For each cohort that significantly deviated from age and period main effects based on the deviation magnitude tests from Step 2, we computed the average of the age-byperiod interaction terms contained in that cohort and used a t test to examine the average of that cohortspecific deviation. These averages and associated t tests can be used to assess differences between cohorts in terms of their mean deviation from the age and period main effects. Table S6 of the  supplementary appendix contains the results and Table S7 provides the relevant age and period main effects.
Our dependent variable in the APC analyses is a summative scale measuring respondents' preference for social and interactional distance. The scale was created by summing respondents' answers to the six social distance items described above, where 1 = definitely willing, 2 = probably willing, 3 = probably unwilling, and 4 = definitely unwilling. We then divided these sums by the total number of items the respondent answered. The reliability coefficient for the resulting scale was 0.852. Exploratory factor analyses confirmed that a single factor was sufficient to characterize respondents' preferences for social distance.  Note: AIC provides the Akaike information criterion for each of the candidate models.
Lower values indicate better model fit (i.e., less information loss relative to the "true" model that generated the data). All models include controls for the respondent's gender, educational attainment, and race, as well as the gender, educational attainment, and race of the person described in the vignette. See the technical appendix for more details.