SUPPLEMENTAL MATERIALS FOR
“VALUE CHOICES AND AMERICAN
PUBLIC OPINION”
William G. Jacoby
Michigan State University
February 2006
[email protected]
http://polisci.msu.edu/jacoby/
This report provides supplemental information about some of the empirical analyses reported in
“Value Choices and American Public Opinion.” The first section focuses on the determinants of
intransitive choices among values. It explains why value triads are used as the units of analysis in the
logistic regression reported in the paper and replicates the analysis using individual MIS respondents as
the units. The second section of the report discusses the definition of the “equality” value employed in the
MIS interview schedule and explains why this is probably not an undue limitation on the analysis. Next,
the third section briefly considers the implications of the very limited value set available in the MIS. The
fourth section of the report addresses concerns about the generalizability of the findings due to the unique
nature of the Multi-Investigator Study. This section also introduces some data that replicate the basic
findings from the MIS regarding transitivity and reported in “Value Choices and American Public
Opinion.” The fifth section provides the specific coefficients and the variance-covariance matrix of the
estimates for the conditional effects effects model that tests for framing effects in the impact of value
choices on attitudes toward government spending. Finally, the sixth section replicates the heteroskedastic
regression analysis of the impact of value choices on issue attitudes for three additional issues.
DETERMINANTS OF INTRANSITIVE VALUE CHOICES
In “Value Choices and American Public Opinion,” the determinants of intransitive value choices
are examined using value triads as the units of analysis. This means that the logistic regression equation
reported in Table 3 is based upon four observations for each MIS survey respondent (i.e. for each of the
four value triads generated by that person’s pairwise choices). The dependent variable is a dichotomy,
scored zero if a person’s responses to a given triad are transitive, and scored one if the responses to that
triad are intransitive. An alternative strategy would be to use individual survey respondents as the units of
analysis. In that case, the dependent variable would be a dichotomy, scored zero if all of a person’s
pairwise value choices are transitive (when they are assembled into triads) and scored one if any subset of
the choices produces an intransitive triad.
Why Use Triads as the Units Rather Than Individuals?
The approach using triads as units of analysis has an important advantage over that using
individual survey respondents: Terms for the separate values can be included in the equation, in order to
see whether the presence of a particular value in a triad affects the probability of an intransitivity
involving that triad. This cannot be done if the analysis is conducted using individual survey respondents
as the units. The problem is that including a term for an individual value, or for a particular value triad,
leads to perfect classification on the dichotomous dependent variable; this, in turn, makes it impossible to
calculate maximum likelihood estimates of the coefficients for those variables. In other words, consider
an independent variable for the value “liberty.” It would be scored zero if all of a person’s pairwise
choices involving that value are transitive, and one if any of the choices involving liberty are intransitive
with respect to two other values. In this scenario, anytime the independent variable is scored one, the
dependent variable is also one, by definition. This problem is exactly the reason that the triad-level
analysis is used in the paper, rather than an analysis of individual survey respondents.
Replication Using Individuals as Units of Analysis
Of course, it is still possible to obtain individual-level ML estimates using the other, remaining
determinants of intransitive choices (i.e., apart from the values, themselves). Table S1 presents the results
from such an analysis, and they are fully consistent with those reported in the paper. The dependent
variable is the dichotomous indicator for the presence of intransitivity in a survey respondent’s choices.
The independent variables are the same ones used in Table 3, in “Value Choices and American Public
Opinion.” Just as in the latter case, the overall fit of the individual-level equation is statistically
significant (χ2 = 49.49, with a probability value of 0.000) but it does not really provide a very accurate
description of the data: The pseudo-R2 is quite close to zero, at 0.037. Once again, this suggests that much
of the observed intransitivity is due to measurement error in the pairwise choices.
Turning to the separate independent variables, the estimation problem described above precludes
any terms for specific values. Therefore, the only variable representing the ambivalence hypothesis is
2
ideological self-placement. The coefficient for this variable is quite small (-0.050) and it is not
statistically different from zero. Just as in the original analysis, self-identified liberals are not reliably
more likely to exhibit intransitive value choices. In contrast, the coefficients for political knowledge,
family income, and education beyond high school are all significantly different from zero in the expected
direction (negative), indicating that all of these factors do reduce the probability that a person will express
intransitive pairwise choices among the four values. The remaining three variables have nonsignificant
coefficients showing that less-educated people, African Americans, and women are neither more nor less
likely to have intransitive choices. This is exactly what occurs in Table 3 of the paper. Thus, the analysis
using MIS survey respondents as units produces results that are identical to those when value triads are
employed.
THE DEFINITION OF “EQUALITY” USED IN THE MIS
One could argue that the definition provided in the MIS interview schedule for the value
“equality” is too specific to be used in an analysis of general value choices: “By equality, we mean
narrowing the gap in wealth and power between the rich and the poor.” Admittedly, it would be better if
this value were defined in terms of equal opportunity. However, I do not think the wording used in the
MIS survey constitutes a fatal flaw in the data, for several reasons.
First, there is ample precedent in the literature for discussing equality in terms of economic and
social outcomes. For example, Rasinski (1987) measures individual feelings about equality through
reactions to statements like “In America, everyone should be treated equally because we are all human
beings” and “Those who are well off in this country should help those who are less fortunate.” Similarly,
Feldman and Steenbergen (2001a; 2001b) use statements like “If wealth were more equal in this country
we would have many fewer problems” and “More equality of income would allow most people to live
better” to define their measure of egalitarianism. Thus, several prominent scholars have addressed the
value of equality in economic terms that are quite similar to the question wording employed in the MIS.
3
Therefore, I would argue that the current analysis can be regarded as building more or less directly upon
their work.
Second, even though the MIS respondents were instructed to think in terms of “rich” and “poor”,
their value choices should be, at the very least, related to broader conceptions of equality. For example,
group-based cleavages almost always have an economic component, in which a disadvantaged subset of
society (whether it be African Americans, homosexuals, Catholics, or immigrants) is confronted by a
wealthier, more advantaged group (whites, heterosexuals, Protestants, or native-born Americans). Thus,
phrasing a question about equality in terms of rich and poor actually transcends virtually all specific
forms of inequality that have occurred in American society.
Third, and in a similar vein, a number of researchers have pointed out that Americans’ judgments
about wealth and poverty are inextricably tied to ideas about race (e.g., Peffley, Hurwitz, Sniderman
1997; Lieberman 1998; Gilens 1999; Avery and Peffley 2003). So, one would suspect that the MIS
respondents’ choices about equality are similarly affected by their racial feelings. And, while certainly not
denying the importance of “equality” as a value involved in all group struggles, racial issues are definitely
the most salient context within which it has been raised as a concern for American society and politics.
Thus, it seems likely that the responses regarding equality reflect relatively broad feelings— at least
encompassing race— even if the MIS respondents did not have to acknowledge this openly or perhaps
even realize it, themselves.
Fourth, the MIS data do provide some relatively direct empirical evidence on the latter point. In a
separate battery of questions, respondents were asked to judge the importance of “racial harmony and
equality” against “free speech,” “religious values,” and “pride for country.” The rankings given to “racial
harmony and equality” are positively and significantly correlated with the importance rankings for
“equality”— that is, the data used in this analysis (Pearson’s r = 0.193). The MIS respondents were also
asked whether they agreed or disagreed with the statement “We should give up the goal of racial equality
because blacks and whites are so very different.” Again, the responses to this item (coded on a five-point
4
scale) are correlated with the importance rankings for “equality” (r = 0.160). Still another MIS question
asked respondents to rate the importance for the country of “... the way black people are treated.”
Answers to this item (scored on a ten-point scale) are also correlated to the importance ratings for
“equality” (r = 0.145).
Note that the preceding correlations probably do not reflect the MIS respondents’ generalized
affective feelings toward African Americans: The correlation between the equality importance scores and
feeling thermometer ratings of “Blacks as a group” is much smaller and not statistically significant, at
only 0.004. Instead, it does appear to be the case that the equality rankings used in the current manuscript
tap feelings about racial, as well as economic, inequality in American society. Thus, the wording used in
the MIS interview schedule for the definition for equality probably does not constitute an excessive
constraint on the ways that the respondents thought about, or answered, the value questions that followed
it.
THE LIMITED SET OF VALUES IN THE MIS
The Multi-Investigator Study only asks respondents about their choices between four political
values. This is potentially problematic because human life and social interaction, of course, involve many
more considerations about desirable and undesirable conditions beyond liberty, equality, economic
security, and social order. For example, Rokeach’s seminal study (1973) involved eighteen terminal
values (along with eighteen more instrumental values). And, Shalom Schwartz’s work often examines
several dozen distinct values (e.g., Schwartz 1996). Thus, it is important to emphasize that the limited set
of values used in “Value Choices in American Public Opinion” does not represent an exhaustive set of the
values that are salient in American political culture.
There is a somewhat technical issue involving the number of values that must be considered. A
major objective of “Value Choices in American Public Opinion” is to test for the existence of hierarchical
structure in value choices; this requires comparisons between pairs of values. But, the number of pairs
increases much more rapidly than the number of values: Four values result in six pairs, but five values
5
would entail ten pairs, six values would require fifteen pairs, and so on. If a large number of values were
used, the necessary pairwise survey items would probably result in respondent fatigue that would, in turn,
compromise the reliability of not only the value choices, but also of responses to the remaining items in
the MIS interview schedule (e.g., Schuman and Presser 1996, pp. 50-51; Tourangeau, Rips, Rasinski
2000, pp. 228-229). Using four values, therefore, seems like a reasonable compromise. On the one hand,
it enables direct empirical tests of transitivity in the value choices. On the other hand, the number of items
is not excessive, so it retains the cooperation and interest of the MIS respondents.
A closely related concern involves the omission of two values from this study that have received
a great deal of attention from political scientists, perhaps due to their frequent inclusion in the interview
schedules for the National Election Studies: Individualism and moral traditionalism. Of course, there is
really nothing I can do about this, since the MIS data already exist. But, I do not believe that the lack of
information about individualism and moral traditionalism poses any serious problem for the present
analysis. From an intuitive perspective, any feelings the MIS respondents may have about individualism
are probably correlated to their importance rankings for liberty. Similarly, feelings about moral
traditionalism are probably correlated with the perceived importance of social order. So, I suspect that the
MIS items cover much of the same ground, even if some specific values are not explicitly invoked in the
survey items.
The preceding intuitive argument can be recast with a more theoretical foundation. Schwartz
(e.g., 1992; 1994) argues that particular values can be subsumed within a smaller number of relatively
broad value types. People should respond similarly to specific values that fall within a single type, and
differently to values that arise within different types (with the exact degree of difference determined by
how close the types are to each other).
Schwartz’s empirical analysis shows that three of the four values from the MIS fall within
different types: Liberty (Schwartz actually uses “Freedom” which I assume to be a synonym to
“Liberty”), equality, and social order correspond to Schwartz’s “Self-direction,” “Universalism,” and
6
“Security” types, respectively. Schwartz does not examine economic security, but that value seems to fall
within his “Power” type (which also subsumes “wealth”). Schwartz’s work also does not consider the
values that are at issue here, individualism and moral traditionalism (nor did Rokeach’s earlier work).
Nevertheless, individualism would almost certainly fall within the “Self-direction” type, like liberty (or
freedom), since this type also contains the values “independent,” “choosing own goals,” and “self
respect.” And, moral traditionalism probably coincides with the “Security” value type, since the latter also
includes “family security,” “sense of belonging,” “clean,” and “reciprocation of favors” as well as social
order. So, once again, we are probably safe in concluding that responses to liberty and social order
probably reflect the kinds of responses that would have been obtained if the MIS had elicited choices with
respect to individualism and moral traditionalism, respectively.
GENERALIZABILITY OF FINDINGS
One could, perhaps, argue that the evidence from the MIS data, by itself, is not sufficient to
support the conclusions reported in “Value Choices and American Public Opinion.” At the very least,
secondary confirmation of the basic findings regarding transitivity and hierarchical structure in individual
value choices would be very useful for establishing the theoretical importance of the findings. This is
clearly an important concern that needs to be addressed. And, in response, I would emphasize several
points.
First, the unique nature of the data is precisely the reason that they are so critical to this analysis. I
conducted an extensive search of the literature in psychology, sociology, and political science and found
that the basic assumption (i.e., transitivity) underlying the hypothesized hierarchical structure of
individual value systems has never been subjected to a rigorous empirical test. And, the MultiInvestigator Study is the only survey that contains the necessary items for doing so. Therefore, this dataset
provides an unparalleled opportunity to test a hypothesis that has, heretofore, simply been treated as an
assumption. It would definitely be desirable to assess transitivity in choices across a larger set of values.
But, I believe that even the somewhat limited test that is possible with the MIS data is much better than
7
no test at all. Given the prominence that values have been accorded in recent research on mass political
behavior, I would argue that it is critical to place relevant evidence— like that which can be obtained
from the MIS data— before the scholarly community.
It is also important to emphasize the high quality of this particular dataset. The MIS is based upon
a national sample. Hence, the external validity of the data is optimized, especially in comparison to the
college students and local/convenience samples that have been used in most prior psychological research
on values. Furthermore, the study was carried out by a well-known and highly-regarded survey facility
(the University of California-Berkeley Survey Research Center). Therefore, we can be quite certain that
the technical aspects of the data collection (e.g., drawing the sample, contacting respondents, conducting
the interviews, coding the data, etc.) have been handled in an appropriate manner. In short, there is
nothing whatsoever to suggest that the information from the MIS is any more idiosyncratic or problematic
than the data obtained from other sources such as the University of Chicago’s NORC or the University of
Michigan’s SRC.
There also is some preliminary empirical evidence that provides additional support for the
findings from the MIS data. Recently, Knowledge Networks conducted two internet surveys that
contained items measuring value choices. The values, themselves, were intentionally modeled after those
included in the MIS. The first survey, designated the “Political Opinion Survey” by Knowledge Networks
and conducted in Fall 2002, presented respondents with six values— liberty, equality, economic security,
social order, morality, and patriotism. The values were given brief definitions, similar to those used in the
MIS items. Then, respondents were asked “If you absolutely had to choose, which value would you say is
most important?” After they made a choice, respondents were then asked “Now, of the values that remain
which is the most important?” And, the latter question was repeated until only one value (i.e., the leastpreferred) remained unselected.
Now, this question format does not allow for direct assessment of transitivity. But, if individual
value choices were “truly” intransitive, then we would expect that people would not be able to complete
8
the full ranking task (and, note that skipping a question is particularly easy in an internet survey, where
there is no direct social pressure to respond from a human interviewer). Nevertheless, 98.13% of the
respondents (n = 7,533) who started the series of value questions provided a fully-ranked set of value
preferences. Undoubtedly, some of the respondents are simply complying with the request for answers
and, thus, “masking” intransitivities that might occur under a different question format. Nevertheless,
such a high completion rate should not occur in this question series if people are experiencing widespread
difficulties in making choices between values. A more complete analysis of these data is provided in
Jacoby and Sniderman (2006).
The second Knowledge Networks survey was conducted during Summer 2005. It was supported
by the program for Time-Sharing Experiments in the Social Sciences and is designated “TESS 37,
Jacoby” by Knowledge Networks. In this study, five values were presented to, and defined for, the
respondents: Liberty, equality, economic security, social order, and morality. The values were then
presented in all possible sets of three. For each such “triad” of values, the respondents were asked to
indicate the one they thought was most important and the one they thought was least important. This
information can be used to extract the pairwise choices from within each triad. And, across the full set of
value triads (with five values, there are ten triads), this produces three replications of each pairwise
choice. Taking the dominant selection for each pair of values (i.e., the value that was chosen two or three
times in each pair of values), 87.52% of the respondents (n = 649) provided fully transitive pairwise
choices across the full set of five values. Note that the Knowledge Networks data show a much lower
level of intransitivity than the MIS data. That is probably due to the replicated pairwise choices, which
reduce the amount of measurement error in the former dataset. A preliminary report on these data is
provided in Jacoby (2006).
Taken together, the evidence from these two Knowledge Networks internet surveys shows that
the basic finding from the MIS data (i.e., largely transitive value choices) is generalizable. Once again,
people seem to have very little difficulty expressing their preferences among different values. And, the
9
more recent data involve larger sets of values (i.e., six values for the 2002 survey and five values in the
2005 survey), so this should also help to alleviate any lingering concerns about the inclusion of only four
values in the MIS data.
COEFFICIENT ESTIMATES FOR THE HETEROSKEDASTIC REGRESSION MODEL
The section of “Value Choices and American Public Opinion” titled “Value Choices and Issue
Attitudes” discusses a heteroskedastic regression analysis in which the effects of the independent
variables on the mean level of the dependent variable (attitudes toward government spending, in this case)
are allowed to vary across three experimental conditions (corresponding to different frames of the
government spending issue). The parameters for the mean and variance functions of the model are
estimated simultaneously, using an approach described in Greene (1997, pp. 558-567). The resultant
equation is a bit unwieldy, since it includes six explanatory variables, two dummy variables, and
multiplicative terms between each of the explanatory variables and each of the dummies. For this reason,
Table 4 in “Value Chocies and American Public Opinion” summarizes the results by presenting
conditional effects for the three experimental groups.
These conditional effects are calculated by summing the appropriate coefficients and
multiplicative terms. The standard errors of the conditional effects are obtained using the simple
procedure described by Friedrich (1982). Specifically, consider a situation where we are interested in the
conditional effect of a variable designated X 1 in each of three experimental conditions. The latter are
indexed by two dummy variables, designated D 2 and D3, and the equation includes multiplicative terms
D2 X1 and D3 X1 . The estimated coefficient for X 1 is b1, and the coefficients for the multiplicative terms
are b2, and b3, respectively. The general formula for the conditional effect of X1 is calculated as follows:
CEX ' b1 % b2 D2 % b3 D3
1
Of course, the specific value of the conditional effect will vary because the two dummy variables’ values
will change from zero to one under the different experimental conditions. The general formula for the
sampling variance for the conditional effect of X 1 is obtained from the following equation:
10
2
2
2
2
2
S 2 ' Sb1 % D2 Sb2 % D3 Sb3 % 2 D2 Sb ,b % 2 D3 Sb ,b
1
2
1
3
The standard error of the conditional effect is the square root of the preceding quantity. Note that it is not
necessary to include the covariance between the two dummy variables in the preceding equation, since
they will never both be equal to one at the same time
As a specific example, we can calculate the conditional effects for liberty-versus-equality choices
(coincidentally, designated X1 in Tables S2 and S3) under the three issue frames. The frame for “crime,
the environment, and public education” is the reference category. Hence, both dummy variables (N and M
in Tables S2 and S3) are equal to zero. The conditional effect for this issue frame is, therefore, just equal
to the coefficient for X1 or 0.106 (from Table S2). Similarly, the sampling variance of the conditional
effect is just the sampling variance for this coefficient, or 0.0007 (from Table S3); the standard error is the
square root of this quantity, or 0.027 (approximately— the actual calculations for these and the other
conditional standard errors used information out to more than four decimal places). In the “services for
the needy” frame, N is equal to 1 and M is zero. Therefore, the conditional effect is the sum of the
coefficients for X1 and X1N in Table S2, or:
0.106 + -0.010 = 0.096.
The sampling variance for the conditional effect is obtained by summing the sampling variances for these
two coefficients and two times the covariance of the coefficients (all from Table S3) or:
0.0007 + 0.0016 + (2 × -0.0007) = 0.0009
The square root of this quantity is the standard error, or 0.029 (approximately). Finally, the conditional
effect for the “services for minorities” frame is obtained by summing the coefficients for X 1 and X 1M from
Table S2:
0.106 + 0.017 = 0.123
Again, the sampling variance for the conditional effect is obtained by summing the sampling variances for
these two coefficients and two times the covariance of the coefficients (all from Table S3):
0.0007 + 0.0015 + (2 × -0.0007) = 0.0008
11
The square root of the preceding quantity gives the conditional standard error: 0.028 (again,
approximately). Of course, similar calculations are carried out to obtain the conditional effects and
conditional standard errors for the remaining explanatory variables, X 2 through X6.
REPLICATING THE ANALYSIS WITH ADDITIONAL ISSUE ATTITUDES
The analysis in “Value Choices and American Public Opinion” tests the hypothesis that value
choices function as an influence on issue attitudes. However, the empirical analysis in the manuscript is
limited to attitudes on a single issue— government spending. This is justified for two reasons: First,
government spending is a highly salient issue in American politics that is also known to be susceptible to
issue framing effects. Second, including additional issues would lengthen an already-long manuscript.
Nevertheless, the basic argument definitely should apply to value influences on other policy issues, as
well.
This section of the report replicates the heteroskedastic regression analysis for three other issues:
Women’s rights; affirmative action; and tolerance of free speech groups. Most of the non-value
independent variables that are hypothesized to influence issue positions are identical across the three
issues (i.e., party identification, ideological self-placement, race, and family income). Similarly, the two
variables hypothesized to affect the disturbance variance remain the same— the number of tied value
rankings and the political knowledge index.
But, the question wordings from the MIS interview schedule are quite different in each case. And,
the measures for two of the issues also include experimental manipulations. Therefore, the latter will
require additional dummy variables and multiplicative terms to test for variability in effects. Furthermore,
slightly different subsets of value choices are relevant to each of these issues. For these reasons, the
results are discussed separately for each issue. The exact results differ across the three analyses. But, the
substantive conclusions relevant to the manuscript are the same in each case: Value choices affect issue
attitudes, and non-ordered choices do not seem to inhibit the process.
12
Women’s Rights
Attitudes toward women’s rights are measured using a summated rating scale based upon three
items from the MIS interview schedule. Respondents were asked, “How important are women’s rights to
you personally?”, “How important is it to you that the federal government does what you think is best on
issues related to women's rights?”, and “When people think about women's rights they have different
reactions. Some think it is bad and some think it is good. How about you?” Responses to all three items
were scored on a five-point scale, where smaller values indicate greater support for women’s rights and
larger values indicate less support. The reliability of this scale is very good, with an Alpha coefficient of
0.752. Therefore, the individual scale scores will be regarded as a general measure of personal support for
women’s rights.
The value choices hypothesized to influence attitudes on women’s rights are identical to those
used for government spending. Once again, the conflict between liberty and equality permeates American
political culture, so it is certainly relevant to questions about equal rights for a group that has been
economically disadvantaged in the past. And, the latter point also suggests that choices with respect to
economic security might also have an impact on attitudes.
The empirical estimates from the heteroskedastic regression analysis are shown in Table S4. The
first panel of the table shows the maximum likelihood estimates for the influences on issue positions. For
present purposes, the most important result is the significant, positive, coefficient for liberty-equality
choices. Given the coding of the variables, this means that people who believe equality is more important
than liberty are also more likely to support women’s rights. The coefficient for economic security is
positive, indicating that people who support economic security are more opposed to women’s rights.
While this contradicts the hypothesized direction of influence, the coefficient is not statistically larger
from zero, so the effect is not reliably different from sampling error.
Among the remaining independent variables, party identification and ideology both have positive
effects. And, since these variables and the liberty-equality variable are all measured on seven-point scales,
13
it is possible to compare the magnitudes of the influences. Liberal-conservative ideology has the strongest
effect, with a coefficient of 0.146. Party identification comes next, with a coefficient of 0.088. The
coefficient for liberty-equality choices is smaller still, at 0.054. Note, however, that the latter is not
significantly different from the coefficient for party identification. Turning to the demographic controls,
family income has a very small coefficient which is not statistically significant. And, African Americans
are significantly more likely than others to support women’s rights.
The second panel of Table S4 shows the hypothesized influences on the disturbance variance in
the model of influences on women’s rights attitudes. Both coefficients are relatively small, and neither
one has the sign that would be anticipated by prior hypotheses. So, the error variance is somewhat smaller
when a person gives a larger number of tied value rankings, and the variance is slightly larger at higher
levels of political knowledge. However, neither of the coefficients for these variables are significantly
different from zero. Therefore, once sampling variation is taken into account, value conflict and political
sophistication really have no discernible effect on the clarity of the process through which the
independent variables impinge on attitudes toward women’s rights.
Affirmative Action
Attitudes toward affirmative action are measured using responses to a single item in the MIS
interview schedule. However, there were three experimental manipulations applied to this question. The
introductory sentence, read to all respondents, is as follows: “Next, I’d like to ask your opinion on federal
laws requiring affirmative action programs for women and minorities provided there are no rigid quotas.”
Next, the MIS respondents were assigned randomly to four conditions, which determined whether “most
citizens,” “the Supreme Court,” “the President,” of “Congress” appeared in the next sentence of the
survey item: “As you may know (the randomly-determined reference group) now favor(s) federal laws
requiring affirmative action programs for women and minorities.” At this point, respondents were, again,
assigned to three randomly-determined groups. For one group, the next sentence stated “(the randomlydetermined reference group) favor(s) such laws because (they/he/it) think(s) these laws are needed in
14
order to balance out the effects of past discrimination against women and minorities.” A second group
was read the following sentence, instead: “One argument is that such laws are needed in order to balance
out the effects of past discrimination against women and minorities.” And, a third group of respondents
was given no rationale for affirmative action programs. Finally, all respondents were asked “How about
you? Do you favor or oppose federal laws requiring affirmative action?” And then, “Would you say you
strongly (favor/oppose) or only somewhat (favor/oppose) these laws?” Responses were scored on a fivepoint scale, with smaller values indicating greater support for affirmative action.
This analysis will not use the first manipulation, varying the reference group that was mentioned
to the respondents. Preliminary analyses showed that varying the external sources of support (i.e., the
reference group mentioned in the question) for affirmative action programs has no effect whatsoever on
the ways that values impinge on individual attitudes toward those programs. Similarly, the second
manipulation will be dichotomized so that it only differentiates between respondents who received a
rationale for affirmative action programs and those who did not.
Turning to the independent variables, the only values measure included in this analysis is the
signed difference in rank scores for liberty and equality. These are the only two values that are directly
related to the common rhetoric that surrounds the affirmative action issue in American political discourse.
And, again, the scores for the other two values are completely uncorrelated with affirmative action
attitudes.
Along with the other independent variables, a dummy variable is included for the question
manipulation on the dependent variable. It is coded zero for MIS respondents who received no rationale
for affirmative action, and coded one for those who were given a rationale. This dummy variable is also
used to create multiplicative terms, allowing the effects of all other independent variables to differ across
the two experimental conditions.
Table S5 summarizes the results from the heteroskedastic regression analysis of attitudes toward
affirmative action programs. The first panel of the table shows the conditional effects of the independent
15
variables on respondents’ positions regarding affirmative action, within each experimental group (i.e.,
those who received a rationale and those who did not). Once again, the most important result in the
present context is the significant positive impact of individual choices between liberty and equality.
People who believe liberty is more important are less supportive of affirmative action programs than are
those who believe equality is more important. The size of this effect does not change significantly across
the two experimental conditions.
Among the other independent variables, ideology and race both show influences that do not vary
across the experimental conditions. As one would expect, conservatives are significantly less supportive
of affirmative action, while African Americans are significantly more supportive. Family income does not
have a significant effect under either condition. Finally, party identification does have a significant effect,
but only when a rationale is provided for affirmative action programs.
On this issue, value choices between liberty and equality have an effect that is comparable to that
of personal ideology. The conditional effects of these two variables are not significantly different from
each other within either of the experimental conditions. And, when a rationale is provided, partisanship
exhibits a similarly strong effect on attitudes.
Turning to the second panel of Table S5, the signs of the coefficients are, at least, consistent with
prior hypotheses: People with tied ranks for liberty and equality exhibit slightly larger error variance,
while those at higher levels of political knowledge exhibit slightly smaller error variance. Again,
however, both of the coefficients are quite small and neither one is statistically different from zero. So, as
was the case with the previous issue, neither value conflict nor political sophistication have much effect
on the disturbance variance.
Tolerance of Free Speech Groups
Tolerance of free speech groups is measured using a three-item scale. The MIS respondents were
first randomly-assigned to three conditions, corresponding to different groups: “Members of the Ku Klux
16
Klan,” “People who burn the American flag,” and “homosexual or gay activists.” They were then
presented with the following statements, with the appropriate group placed into the blank space:
How much do you agree or disagree with the following statement? _____ should be allowed to
make a speech in your city.
Suppose _____ wanted to have a weekly television program on a community access cable
television station in your area? How would you feel about that? Would you be strongly in favor,
somewhat in favor, somewhat opposed, or strongly opposed?
Suppose _____ wanted to hold a rally in your area? How would you feel about that? Would you
be strongly in favor of their right to hold the rally, somewhat in favor, somewhat opposed, or
strongly opposed?
Responses to each of the preceding questions were recorded on five-point scales, with smaller values
indicating greater tolerance for the respective groups. Response scores are summed to form a single scale,
which has a reliability of 0.888 (Cronbach’s Alpha).
A single heteroskedastic regression equation is used in this analysis, with dummy variables and
multiplicative terms included in order to allow the independent variables’ effects to differ, depending
upon the target group. This analysis only includes one variable for value preferences: The signed
difference between the rank scores for liberty and for social order. This new measure is used because it
taps the two values that come most directly into conflict in questions about tolerance for nonconformist
groups within American society. In contrast, equality (at least, as it is defined in the MIS) and economic
security do not seem to be immediately relevant to tolerance and civil liberties issues. Therefore, choices
regarding the latter two values are omitted from this analysis.
The results are shown in Table S6. The first panel shows the conditional effects of the
independent variables separately for each of the three free speech groups. Precisely as expected, the effect
of individual choices between liberty and social order is negative, in all three cases. Given the coding of
the variables, this means that people who favor liberty are more tolerant of nonconformist groups while
those who favor social order are less tolerant. Note, however, that conclusions regarding the statistical
significance of this effect are a bit ambiguous. The conditional effect is significantly different from zero
(0.05 level, one-sided test) for tolerance of the Ku Klux Klan. But, the null hypothesis cannot be rejected
17
(using the same decision criterion) for flag burners or gay activists. At the same time, however, the
difference between the three conditional effects is not statistically significant. So, it is equally reasonable
to conclude that the impact of value choices on feelings about the rights of the latter two groups is no
different than it is for tolerance of the Ku Klux Klan. This seemingly contradictory evidence is somewhat
frustrating, but it merely reflects the presence of statistical uncertainty in the estimates. Perhaps the
“safest” conclusion is that the data do not directly contradict the hypothesis that choices between liberty
and social order affect individual tolerance of free speech groups.
The effects of the other independent variables also vary quite a bit across the three groups. For
example, party identification shows a significant negative effect for the Ku Klux Klan (i.e., Republicans
are more tolerant, Democrats less tolerant), a negative but nonsignificant effect for flag burners, and a
nonsignificant positive effect for gay activists. For ideology, the direction of influence is constant (i.e.,
the effect is always positive, meaning that conservatives are less tolerant than liberals), but the effect is
statistically different from zero only for flag burners and gay activists. Income shows a consistently
negative effect— that is, people with higher family incomes are always more tolerant. And, finally,
African Americans are significantly less tolerant of Ku Klux Klan members and gay activists; they are not
significantly different from others with respect to flag burners.
The second panel of Table S6 shows the results for the disturbance variance from the equation
predicting tolerance of nonconformist groups. Once again, the coefficient for tied value rankings is very
small and nonsignificant. Thus, value conflict does not affect the degree to which the independent
variables exert their respective impacts on feelings of tolerance. In contrast, the coefficient for the
political knowledge scale is significant and positive. While this is contrary to the general hypothesis that
political sophistication facilitates the translation process from influences to attitudes, it is definitely not an
unreasonable result. The influence of political knowledge probably stems from ambivalent feelings that
only emerge among people who recognize the potential problems inherent in allowing disliked groups to
18
publicize their own positions. But, it is important to emphasize that this ambivalence is not directly
related to choices between the two most relevant values, liberty and social order.
Summary of Findings for Issue Replications
The heteroskedastic regression analyses of the preceding three issues provide general
confirmation for the findings reported in “Value Choices and American Public Opinion,” particularly
those directly related to values. The results presented here show that individual choices between relevant
values do influence personal issue attitudes. However, the magnitude of this effect varies quite a bit from
one issue to the next, and the impact is never substantially greater than that from other well-known
determinants of political attitudes, such as party identification and ideology. So, just as stated in the
manuscript, itself, value choices exert an important, but not overwhelming, effect on the development of
issue attitudes. The results also show that individual value conflict— operationalized here as the inability
or unwillingness to provide a clear ranking of the relevant values— does not inhibit the process by which
value choices, symbolic predispositions, and other background factors impinge on issue attitudes. Once
again, this mirrors precisely the findings reported in the “Value Choices and American Public Opinion.”
The one finding that does diverge from the results reported in the paper involves the impact of
political knowledge on the disturbance variance: Here, the knowledge scale never shows a significant,
negative effect, as it does in the case of attitudes toward government spending. This may well raise
serious questions about the generalizability of sophistication effects. And, the positive coefficient in the
analysis of tolerance toward nonconformist groups also suggests that sophistication may well induce
ambivalence, precisely in the manner hypothesized by Alvarez and Brehm (2002). But, once again, it is
important to emphasize that this effect occurs separately from individual value choices. Hence, it is not
the central focus of attention in this study and definitely does not contradict the major findings reported in
“Value Choices and American Public Opinion.”
19
REFERENCES
Alvarez, R. Michael and John Brehm. (2002) Hard Choices, Easy Answers. Princeton, NJ: Princeton
University Press.
Avery, James M. and Mark Peffley. (2003) “Race Matters: The Impact of News Coverage of Welfare
Reform on Public Opinion.” In Sanford F. Schram, Joe Soss, Richard C. Fording (Editors), Race and
the Politics of Welfare Reform. Ann Arbor, MI: University of Michigan Press.
Feldman, Stanley and Marco Steenbergen. (2001a) “The Humanitarian Foundation of Public Support for
Social Welfare.” American Journal of Political Science 45(3): 658-677
Feldman, Stanley and Marco Steenbergen. (2001b) “Social Welfare Attitudes and the Humanitarian
Sensibility.” In James H. Kuklinski (Editor), Citizens and Politics: Perspectives from Political
Psychology. Cambridge, UK: Cambridge University Press.
Friedrich, Robert J. (1982) “In Defense of Multiplicative Terms in Multiple Regression Models.”
American Journal of Political Science 26: 797-833.
Gilens, Martin. (1999) Why Americans Hate Welfare. Chicago: University of Chicago Press.
Greene, William H. (1997) Econometric Analysis (Third Edition). Upper Saddle River, NJ: Prentice-Hall.
Jacoby, William G. (2006) “Testing for Hierarchical Structure and Priming Effects Among Individual
Value Choices.” Unpublished paper, available at http://polisci.msu.edu/jacoby/.
Jacoby, William G. and Paul M. Sniderman. (2006) “The Structure of Value Choices in the American
Public.” Paper presented at the 2006 Annual Meetings of the Southern Political Science Association.
Lieberman, Robert C. (1998) Shifting the Color Line: Race and the American Welfare State. Cambridge,
MA: Harvard University Press.
Peffley, Mark; Jon Hurwitz; Paul M. Sniderman. (1997) “Racial Stereotypes and Whites’ Political Views
of Blacks in the Context of Welfare and Crime.” American Journal of Political Science 41: 30-60.
Rasinski, Kenneth A. (1987) “What’s Fair is Fair— Or is it? Value Differences Underlying Public Views
About Social Justice.” Journal of Personality and Social Psychology 53: 201-211.
Tourangeau, Roger; Lance J. Rips; Kenneth Rasinski. (2000) The Psychology of Survey Response.
Cambridge, UK: Cambridge University Press.
Table S1: Logistic regression showing influences on the probability that an individual will
exhibit intransitive preferential choices among values.
ML Coefficient
Estimate (Standard
Error in parentheses)
Liberal-conservative
identification
Odds
Ratio
-0.050
(0.041)
0.951
Political
knowledge
-0.142*
(0.055)
0.868
Family
income
-0.049*
(0.030)
0.952
Less than high
school education
0.0176
(0.240)
1.018
Education beyond
high school
-0.558*
(0.163)
0.572
AfricanAmerican
-0.133
(0.240)
0.876
Female
respondent
-0.244
(0.143)
0.784
Constant
0.023
Likelihood-ratio
Chi-square (7 df)
49.490
Probability for
Chi-Square
0.000
PseudoR2 value
0.037
* Coefficient is statistically smaller than zero, 0.05 level, directional hypothesis test.
Data Source: 1994 Multi-Investigator Study. The units of analysis are individual survey respondents,
so the number of observations is 1294.
Table S2: Maximum likelihood coefficient estimates and standard errors for the mean function
of the heteroskedastic regression model predicting attitudes toward government
spending. Terms within parentheses in the leftmost column are the labels used to
represent the variables in Table S3.
Coefficient
Standard Error
Needy People (N)
0.133
0.209
Minority Groups (M)
0.144
0.223
Liberty-Equality (X1)
0.106
0.027
Liberty-Equality × Needy (X 1N)
-0.010
0.039
Liberty-Equality × Minority (X 1M)
0.017
0.039
Economic Security (X2)
0.000
0.043
Economic Security × Needy (X 2N)
-0.032
0.064
Economic Security × Minority (X 2M)
0.015
0.062
Party Identification (X3)
0.058
0.042
Party Identification × Needy (X 3N)
0.085
0.034
Party Identification × Minority (X 3M)
0.066
0.036
Ideology (X4)
0.124
0.035
Ideology × Needy (X4N)
0.003
0.008
Ideology × Minority (X4M)
0.028
0.035
Family Income (X5)
0.018
0.016
Family Income × Needy (X5N)
0.106
0.023
Family Income × Minority (X 5M)
0.002
0.023
African-American (X6)
-0.521
0.160
African-American × Needy (X6N)
-0.086
0.226
African-American × Minority (X 6M)
0.153
0.239
Intercept ((X0)
2.036
0.191
Dummy Variables for Issue Frame
Value Choices
Symbolic Predispositions
Demographic Factors
Data Source: 1994 Multi-Investigator Study
Table S3: Variance-covariance matrix of estimates from the mean function of the heteroskedastic regression model predicting attitudes
toward government spending. Variable labels are listed in leftmost column of Table S2.
N
M
X1
X1N
X1M
X2
X2N
X2M
X3
X3N
X3M
X4
X4N
X4M
X5
X5N
X5M
X6
X6N
X6M
X0
N
M
X1
X1N
X1M
X2
X2N
X2M
X3
X3N
X3M
.0438
.0199
.0002
-.0004
-.0002
.0035
-.0077
-.0035
.0022
-.0042
-.002
.0001
-.0001
.0001
.0010
-.0021
-.0010
.0057
-.0134
-.0057
-.0210
.0496
.0001
-.0002
.0002
.0034
-.0034
-.0066
.0012
-.0018
-.0026
.0008
.0001
-.0030
.0011
-.0010
-.0023
.0052
-.0055
-.0104
-.0219
.0007
-.0007
-.0007
.0002
-.0002
-.0002
-.0001
.0001
.0001
.0000
-.0000
.0000
.0000
-.0000
-.0000
.0004
-.0004
-.0004
-.0003
.0016
.0007
-.0003
.0006
.0003
.0002
-.0002
-.0002
-.0000
-.0000
.0000
-.0000
-.0000
.0000
-.0004
.0012
.0004
.0002
.0015
-.0003
.0003
.0005
.0001
-.000
-.0002
-.0000
.0000
-.0002
-.0000
.0000
-.0000
-.0004
.0004
.0013
.0003
.0018
-.0018
-.0018
-.0000
-.0000
-.0000
-.0001
.0000
.0000
.0001
-.0001
-.0001
-.0003
.0003
.0004
-.0032
.0041
.0018
-.0002
.0001
.0000
-.0000
.0000
-.0000
-.0001
.0001
.0001
.0003
-.0000
-.0003
.0038
.0038
-.0000
.0000
.0002
.0000
.0000
-.0002
-.0001
.0001
.0001
.0003
-.0003
-.0005
.0034
.0018
-.0006
-.0006
.0008
-.0002
.0002
-.0000
.0000
.0001
.0010
-.0007
-.0013
-.0055
.001
.0006
.0000
.0000
-.0000
.0001
-.0001
-.0001
-.0009
.0020
.0009
.0019
.0013
.0002
-.0000
-.0005
.0001
-.0001
-.0001
-.0010
.0009
.0019
.0015
Data Source: 1994 Multi-Investigator Study
Table S3 (Continued): Variance-covariance matrix of estimates from the mean function of the heteroskedastic regression model
predicting attitudes toward government spending. Variable labels are listed in leftmost column of Table S2.
N
M
X1
X1N
X1M
X2
X2N
X2M
X3
X3N
X3M
X4
X4N
X4M
X5
X5N
X5M
X6
X6N
X6M
X0
X4
X4N
X4M
X5
X5N
X5M
X6
X6N
X6M
X0
.0012
-.0002
-.0004
.0000
-.0000
-.0000
-.0001
.0002
-.0002
-.0044
.0000
-.0000
-.0000
.0000
-.0000
-.0000
-.0000
.0001
.0009
.0012
-.0000
-.0000
.0000
.0002
-.0000
-.0005
.0008
.0002
-.0002
-.0002
.0001
-.0001
-.0001
-.0012
.0005
.0002
-.0001
.0002
.0001
.0011
.0005
-.0001
.0001
.0002
.0010
.0258
-.0257
-.0257
-.0054
.0513
.0257
.0047
.0573
.0064
.0366
Data Source: 1994 Multi-Investigator Study
Table S4: Heteroskedastic regression showing influences on attitudes toward women’s rights.
A. Factors affecting each respondent’s position on the women’s rights issue.
Maximum Likelihood
Regression Coefficient
Standard
Error
Values
Differential Importance
of Liberty and Equality
0.054*
0.021
Importance Ranking for
Economic Security
0.053
0.034
Party
Identification
0.088*
0.019
Ideological
Self-Placement
0.146*
0.021
-0.008
0.012
AfricanAmerican
-0.374*
0.117
Intercept
1.641
Likelihood ratio Chi-Square
176.913
Chi-Square Probability (8 df)
0.000
R2
0.125
Symbolic Predispositions
Sociodemographic Factors
Family
Income
* Coefficient is statistically different from zero at the 0.05 level (directional test).
Note: Data source is the 1994 Multi-Investigator Study.
Table S4: Heteroskedastic regression showing influences on attitudes toward women’s rights
(Continued).
B. Factors affecting the variance of the regression error term (specifically, the log of the error
variance is specified to be a linear function of the variables included in the following table).
Maximum Likelihood
Regression Coefficient
Standard
Error
-0.087
0.055
Political
Knowledge
0.046
0.029
Intercept
0.184
R2 for predicting error variance
0.121
Value Ambivalence
Number of Tied Scores for Liberty,
Equality, and Economic Security
Political Sophistication
Note: Neither coefficient in Table S2B is significantly different from zero.
Table S5: Heteroskedastic regression showing influences on attitudes toward affirmative
action programs.
A. Factors affecting each respondent’s position on the affirmative action issue.
Experimental Condition
Received rationale for
affirmative action
programs
No rationale for
affirmative action
programs
Observed probability
for difference
between coefficients
0.201*
(0.060)
0.215*
(0.059)
0.866
Party
Identification
0.204*
(0.061)
0.078
(0.058)
0.128
Ideological
Self-Placement
0.188*
(0.062)
0.239*
(0.063)
0.568
-0.042
(0.036)
-0.016
(0.035)
0.606
AfricanAmerican
-0.764*
(0.344)
-1.137*
(0.343)
0.441
Intercept
2.471
2.752
0.750
Likelihood ratio Chi-Square
143.992
Chi-Square Probability (13 df)
0.000
R2
0.200
Values
Differential Importance
of Liberty and Equality
Symbolic Predispositions
Sociodemographic Factors
Family
Income
* Coefficient is statistically different from zero at the 0.05 level (directional test).
Note: Cell entries are conditional effects, calculated from the maximum likelihood coefficient estimates
obtained from heteroskedastic regression. Standard errors (in parentheses) are obtained using
procedures described by Friedrich (1982). None of the conditional effects are significantly
different across the experimental conditions. Data source is the 1994 Multi-Investigator Study.
Table S5: Heteroskedastic regression showing influences on attitudes toward affirmative
action programs (Continued).
B. Factors affecting the variance of the regression error term (specifically, the log of the error
variance is specified to be a linear function of the variables included in the following table).
Maximum Likelihood
Regression Coefficient
Standard
Error
Value Ambivalence
Tied Score for
Liberty and Equality
0.010
0.198
Political
Knowledge
-0.013
0.043
Intercept
1.142
R2 for predicting error variance
0.202
Political Sophistication
Note: Neither coefficient in Table S3B is statistically different from zero (0.05 level, one-sided test).
Table S6: Heteroskedastic regression showing influences on tolerance toward nonconformist
groups.
A. Factors affecting each respondent’s position on the tolerance scale.
Free Speech Group
Ku Klux
Klan Members
People Who
Burn the
American Flag
Homosexual
and Gay
Activists
Observed probability
for difference
between coefficients
-0.144*
(0.052)
-0.069
(0.053)
-0.033
(0.048)
0.287
-0.113*
(0.052)
-0.054
(0.054)
0.037
(0.051)
0.113
0.057
(0.058)
0.214*
(0.059)
0.343*
(0.056)
0.002**
Family
Income
-0.144*
(0.035)
-0.224*
(0.034)
-0.231
(0.032)
0.144
AfricanAmerican
0.768*
(0.299)
-0.099
(0.357)
0.542
(0.310)
0.167
Intercept
5.722
5.237
3.232
Likelihood ratio χ 2
285.021
χ2 Probability (19 df)
0.000
R2
0.196
Values
Differential Importance
of Liberty and Social Order
Symbolic Predispositions
Party
Identification
Ideological
Self-Placement
Demographic Factors
0.000**
* Coefficient is statistically different from zero at the 0.05 level (directional test).
**
Chi-square test of the null hypothesis that conditional effects are identical across the framing
conditions is rejected at the 0.05 level.
Note: Cell entries are conditional effects, calculated from the maximum likelihood coefficient estimates
obtained from the heteroskedastic regression. Standard errors (in parentheses) are obtained using
procedures described by Friedrich (1982). Data source is the 1994 Multi-Investigator Study.
Table S6: Heteroskedastic regression showing influences on tolerance toward nonconformist
groups (Continued).
B. Factors affecting the variance of the regression error term (specifically, the log of the error
variance is specified to be a linear function of the variables included in the following table).
Maximum Likelihood
Regression Coefficient
Standard
Error
Value Ambivalence
Tied Score for Liberty
and Social Order
0.026
0.147
0.075*
0.031
Political Sophistication
Political
Knowledge
Intercept
0.973
R2 for predicting error variance
0.196
* Coefficient is statistically different from zero at the 0.05 level, nondirectional test.