eliciting value judgements about health inequality aversion: testing for framing effects
TRANSCRIPT
1
HESG Sheffield Jan 2014 Paper A061
Work in progress – not for citation.
Eliciting value judgements about health inequality aversion: testing for framing effects
Shehzad Ali1*; Richard Cookson
1; Aki Tsuchiya
2; Miqdad Asaria
1; Ruth Helstrip
1
* Corresponding author: [email protected]
1. Centre for Health Economics, University of York
2. School of Health and Related Research, and Department of Economics, University of Sheffield
Abstract
Questionnaire-based estimates of health inequality aversion are potentially vulnerable to framing
effects. This study tests for four framing effects and a sample selection effect related to framing in the
broad sense of cognition and information processing: (1) small versus unrealistically large health
inequality reductions; (2) population-level versus individual-level descriptions of inequality
reductions; (3) concrete versus abstract scenarios; (4) online versus discussion mode of
administration; and (5) “academic versus non-academic” sample. Twenty nine respondents
participated in a discussion group meeting, and a separate convenience sample of 156 respondents
completed an online questionnaire. In line with previous studies we found that between 20% and 61%
of respondents in different conditions expressed extreme inequality aversion that violates
monotonicity, and between 3% and 20% expressed zero inequality aversion. We found small but non-
significant effects of (1), (2) and (4), and substantial and significant effects of (3) and (5): a higher
proportion of respondents expressed zero health inequality aversion in the concrete scenario, and a
lower proportion expressed extreme health inequality aversion in the academic sample despite
expressing similar or more egalitarian social attitudes to the welfare state and income redistribution.
Keywords Health inequality, inequality aversion, value judgments, framing effects, distributional cost-
effectiveness
Acknowledgements
This study was funded by the National Institute for Health Research Health Services and Delivery Research Programme (project number 11/2004/39). We are grateful to Paul Toner for his
input, and to everyone who participated in our discussion groups or responded to our online survey.
Our errors and opinions are our own and do not reflect those of the HS&DR Programme, NIHR, NHS or the Department of Health.
2
1. Introduction
1.1 Background
Health economists have developed questionnaire instruments to measure how much people
care about unfair health inequality (“equity”) relative to overall health (“efficiency”) and
methods to analyse the data building on the social welfare function (SWF; see for example
Abásolo, Tsuchiya, 2004, 2013; Dolan, Tsuchiya, 2011). The resulting estimates of health
inequality aversion can be used in distributional cost-effectiveness analysis (DCEA) to help
decision makers evaluate trade-offs between improving total health and reducing unfair
health inequality (Asaria et al. 2013). However, these estimates of health inequality aversion
are likely to be influenced by framing effects which have not yet been fully documented.
1.2 Aim
This study aims to tests for four potential framing effects and a sample selection effect related
to framing in the broad sense of cognition and information processing: (1) small instead of
unrealistically large health inequality reductions; (2) population-level instead of individual-
level descriptions of health inequality reductions; (3) concrete instead of abstract intervention
scenarios; (4) online instead of face to face mode of administration; and (5) “academic versus
non-academic” respondents. Each is explained below.
1.3 The five framing effects
(1) Small versus unrealistically large health inequality reductions: Empirical studies that
elicit health inequality aversion parameters typically use hypothetical scenarios that involve
unrealistically large changes in health (Abásolo, Tsuchiya, 2004, 2013; Dolan, Tsuchiya,
2011). For example, respondents may be asked to imagine a government programme that
will extend average life expectancy by several years, with one social group gaining more
years than another. General population average health gains of this size are unrealistic in the
short run, and unlikely to be achievable even with a massive and sustained “once in
generation” programme of cross-government social, political and economic reform. In
practice, the public policy alternatives actually considered by social decision makers deliver
much smaller average health benefits to the population. For example, a case study of
different ways of promoting uptake of the NHS Bowel Cancer Screening Programme among
disadvantaged groups estimated incremental gains in general population average life
expectancy of only a few hours (Asaria et al. 2013). This is because most people do not have
bowel cancer and will gain nothing from the screening programme, but a few people will gain
3
many years of life. It is not known whether the level of inequality aversion found for large
health gains is applicable to health gains of a smaller and more realistic size. We hypothesise
that using small health gains may lead to smaller inequality aversion relative to using large
health gains. Our intuition is that health inequality is a less familiar concept than total health,
perceived at a less fine-grained level of detail. So respondents may see small reductions in
health inequality as worthless while still seeing small total health gains as worthwhile.
(2) Population-level versus individual-level descriptions of health inequality reductions:
Studies typically present health benefits in terms of the average change to individuals (e.g. 2
years per person). However, health benefits to a population can also be expressed in terms of
total gains to the group (e.g. 2,000,000 person-years across a million people). We
hypothesise that when average benefits are small – for example, a few hours per person –
then framing health benefits in terms of population totals may lead to larger inequality
aversion than using average health benefits per person. Our intuition is that using population
totals may make the reductions in health inequality seem larger and more worthwhile.
(3) Concrete versus abstract intervention scenarios: Typical scenarios used in previous
studies do not refer to the cause of the existing health inequality, or the policy mechanism
through which health inequality is to be reduced. Abstract scenarios are used so that
respondents do not bring their own unobserved “baggage” to the interpretation of the
scenarios. Arguably, however, respondents will always fill in the missing gaps in the
scenario to come to their decision, and abstract scenarios may be more susceptible to this
problem than more tightly described concrete scenarios. Furthermore, it is not known how
far people’s “abstract” views about health justice are transferable to their “context-specific”
views about health justice in specific policy contexts. When more concrete scenarios are
given, and people are encouraged to think more realistically rather than abstractly, this may
have an effect (in either direction) on the level of inequality aversion they support.
(4) Online versus face to face mode of administration: Typical surveys of equality aversion
and the social value of health have been conducted face to face, but this is a time consuming
costly undertaking. Increasingly, there is an interest in using online surveys to elicit social
values of health (Schwappach, 2003; Shah et al, 2012; Brazier et al, 2013; Linley, Hughes,
2013; Norman et al, 2013; Skedgel et al, forthcoming), partly because of the high speed,
convenience and low cost of conducting the surveys. However, because questions on social
4
values require careful deliberation, and are not topics that people are familiar with, online
surveys are less likely to reflect considered opinions. On the other hand, face to face
administration may encourage people to give “socially desirable” responses, which may lead
to higher levels of inequality aversion (DeMaio 1984). We therefore hypothesise that using
online surveys may lead to smaller inequality aversion relative to face to face administration.
(5) “Academic versus non-academic” background: A large proportion of respondents in
previous studies – often more than half – express extreme aversion to health inequality that
violates monotonicity (Abásolo, Tsuchiya, 2004, 2013). One possible explanation is that
thinking about ethical trade-offs between improving health and reducing health inequality is
cognitively demanding. Respondents may reduce this cognitive burden by treating health
inequality as a “sacred value” (Tetlock et al. 2000, Tetlock 2003) and employing the simple
decision making heuristic “always choose the more egalitarian programme”. Respondents
with certain backgrounds such as researchers or technocrats may be more comfortable
handling cognitively demanding trade-off tasks, and less likely to adopt a “sacred value”
approach. Strictly speaking this is a sample selection effect, rather than a framing effect,
though it relates to framing in the broad sense of how respondents understand and process the
cognitive tasks they are faced with. Here, the term "academic" refers to individuals whose
background we judge as likely to be related to the academic profession given our recruitment
strategy and the respondent’s self-reported profession, including the phrases “health
economist”, “economist”, researcher”, “lecturer”, “assistant professor”, “analyst”, “scientist”
and “student”.
2. Methods
2.1 Questionnaire
The basic questionnaire instrument used in this study (see Appendix) is adapted from
Abásolo, Tsuchiya (2004) and Dolan, Tsuchiya (2011). It starts off by presenting an
inequality in health across two groups. The respondent is then asked to look at a pair of
alternative government programmes (A and B) which produce the same amount of total
health benefit across the groups but benefit the two groups differently: programme B offers a
reduction in health inequality compared with programme A. After the respondent makes a
choice, they are then presented with more pairs of programmes. Programme A remains the
same, but the health gain to the worse off group in programme B becomes smaller and
smaller – and so both the reduction in health inequality and the gain in total health offered by
5
programme B become smaller and smaller. The aim of the exercise is to record whether or
not the respondent is inequality averse, and if so, by how much. The degree of inequality
aversion is captured by observing the point at which a respondent is indifferent between the
two programmes, one offering more total health and the other offering reduced health
inequality. A specific social welfare function is assumed, fitted to the outcomes of the two
health programmes at indifference, and solved for the inequality aversion parameter.
In this study, the two groups used are “the richest fifth”, who on average live 74 years in full
health, and “the poorest fifth”, who on average live 62 years in full health. Both groups are
made up of around 10 million individuals, and it is assumed that the remaining three fifths of
the population are not affected by any of the hypothetical health programmes. There are four
questions, and each question consists of five pairs of health programmes, A and B. In the first
question, the first pair presents programme A giving a 7-year gain in life in full health to the
richest fifth and a 3-year gain in life in full health to the poorest fifth; and programme B
giving a 3-year gain to the richest fifth, and a 7-year gain to the poorest fifth. In the
subsequent four pairs, the health gains in programme A (7 years to the richest; 3 years to the
poorest) and the health gain to the richest in programme B (3 years) are fixed, while the
health gain to the poorest in programme B decreases gradually from 6 years to 3 years. At
each pair, respondents are asked to choose between A and B, or indicate indifference.
Appendix Figure A1 reproduces the first question. This first question is the “large-average”
presentation, which corresponds to the format used in a few earlier studies. All subsequent
questions use the same background inequality across the richest fifth and the poorest fifth,
and the same ratios of health gains (but scaled down proportionately). In the second question,
the health gains are small and measured in hours per head. In the first pair, programme A
gives a 7-hour gain in life in full health to the richest fifth and a 3-hour gain to the poorest
fifth, while programme B gives a 3-hour gain to the richest fifth and a 7-hour gain to the
poorest fifth. The gain for the poorest group under programme B at the fifth pair is 3 hours.
Thus this represents the “small-average” presentation.
The third question represents the “small-population” presentation. In the first pair,
programme A gives a 7,000-person-year gain in life in full health to the richest fifth as a
group and a 3,000-person-year gain in life in full health to the poorest fifth as a group, while
programme B gives a 3,000-person-year gain to the richest fifth and a 7,000-person-year gain
6
to the poorest fifth. The gain for the poorest group under programme B at the fifth pair is
3,000 person-years. Thus, this represents the “small-population” presentation. Note that,
since each quintile is assumed to consist of 10 million individuals, 7,000 person-years
amounts to 0.0007 years per head, which is equivalent to 6.132 hours per head. We use 7
hours in the second question instead of 6.132 hours in order to maintain the 7:3 ratio of
benefits across the richest and the poorest.
The fourth question (used only in face to face mode of administration) introduces a more
“concrete presentation” to the health inequality and the health programme, using a topic taken
from Asaria et al (2013). A presentation on the different take up rates of bowel cancer
screening by income groups is given, followed by a description of two health programmes on
reminders to participate in screening: one that sends impersonal reminders to all eligible
individuals, and another that sends personalised and GP-endorsed reminder letters to
individuals in deprived areas who have a lower take-up rate. The former programme will
benefit the richest fifth more, while the latter programme will benefit the poorest fifth more.
Besides the concrete context, this question is exactly the same as the third question.
The whole questionnaire is available from the corresponding author on request.
2.2 Data collection
There are two samples. One is a “citizens’ panel” sample, where members of the public were
invited to participate in an event involving: presentations by facilitators; facilitated
discussions in groups of five or six; individual completion of the questionnaire; sharing the
responses within the group; and opportunities to change the questionnaire responses. The
citizens’ panel looked at a number of tasks including the four questions reported in this paper.
The participants for the citizens’ panel were recruited in two ways. First, there were
advertisements in a monthly free local magazine (Your Local Link) distributed to all homes
across York (35 postcode sectors) for two issues (July and August, 2013). Second, 810
leaflets were distributed door to door in 10 of the most deprived streets in York. Those
interested were asked to contact the research team with their name, contact details, age, sex
and postcode. A quota was set by four age groups and sex so that each age/sex group had a
capacity for three to four participants, including one from a postcode with higher deprivation.
7
Those with university academic/research jobs were excluded. A payment of £70 was offered
per participant for the 5-hour session on a Saturday in a city centre location in York.
The second sample is an “online” sample. Select sections of the citizens’ panel
questionnaire were posted online with an internet host called Smart-Survey. This included
the first three questions above (large-average; small-average; small-population). The survey
was publicised on the website of the Centre for Health Economics at the University of York,
on social media, the York Local Link magazine above, and the jiscmail mailing list for health
economists, inviting people to access the web survey. Respondents could complete the
survey anonymously, or leave their name and e-mail address to receive a copy of the findings.
No remuneration was offered for taking part.
Research ethics approval was obtained from the University of York Health Sciences Research
Governance Committee.
2.3 Analysis
Descriptive statistics of the two samples are reported separately in terms of key demographic
variables and answers to key political attitude questions. The online sample is further broken
up into “academics” and “non-academics” sub-samples based on self-reported background.
Each of the main questions allows us to distinguish five different types of preferences, which
correspond to five different principles of health justice. At each pair, respondents have three
choices: programme A, programme B, or indifference. As a shorthand, we represent these
three choices using the characters A, B and =, respectively. We label respondents who prefer
programme A in all pairs as “pro-rich” (AAAAA), and respondents who are indifferent
between the two programmes in the first pair but choose A subsequently as “health
maximisers” (=AAAA). Collectively, we label these first two types as “non-egalitarian”.
Our third type of preference is the “trader” or “weighted prioritarian” (BXXXA), who
chooses B in the first pair, then switches to A at some point (indicated by the XXX in the
middle – see below), and chooses A in the final pair. The term “weighted prioritarian” comes
from the philosophy literature and means people who give priority to the worse off but not
exclusive priority: they still give some weight to gains for the better off. They are thus
willing to make trade-offs between improving total health and reducing health inequality, but
8
will not violate monotonicity which refers to a condition that any increase in individual health
will result in an increase in social welfare. Their preferences can therefore be represented by
a social welfare function with a finite inequality aversion parameter. Strictly speaking, a
respondent who switches to A only in the final pair (BBBBA) might be “leximin” rather than
“weighted prioritarian”, i.e. they give almost exclusive priority to the health of the worse off
group, but are willing to use a second principle such as health maximisation as a “tie-breaker”.
We can however identify respondents who give fully exclusive priority to the health of the
worst off – these are respondents who choose B in the first four pairs but become indifferent
in the final pair. We label these respondents maximin (BBBB=). This preference type can be
represented by the limit of a standard social welfare function as the inequality aversion
parameter tends to infinity. Finally, we label respondents who prefer programme B in the
final pair as “strict egalitarians” (BBBBB). This type of preferences violates monotonicity
and so cannot be represented using standard monotonicity-respecting social welfare functions.
Collectively, we label these last two types as “strong egalitarians”.
Within the “weighted prioritarian” type (BXXXA), we can distinguish seven distinct
response categories by breaking up the XXX in the middle, depending on the point at which
the respondent switches to programme A – (BAAAA), (B=AAA), (BBAAA), (BB=AA),
(BBBAA), (BBB=A), (BBBBA). We can associate each of these with an estimated trade off
point at which the respondent switches to programme A, in terms of the number of units of
health benefit to the poorest fifth in programme B. The first of these (BAAAA) represents a
trade off point of 6.5 (since the respondent switches at some point between 7 and 6 units of
health benefit). The second sub-category (B=AAAA) represents a trade off point of 6 (since
the respondent is indifferent at 6 units of health benefit), and so on down in half units to the
seventh sub-category (BBBBA) which represents a trade off point of 3.5. Similarly, the “pro-
rich” and “health maximisation” categories represent trade off points of >7 and 7,
respectively, and the “maximin” and “strict egalitarian” categories represent trade off points
of 3 and <3, respectively. We thus obtain 11 separate response categories, each associated
with a numerical trade off point ranging from >7 to <3, which can be ranked in order from
the least egalitarian (>7) to the most egalitarian (<3).
By question and by sample, seven distributions (viz. four for the citizens’ panel sample, three
for the online sample) of responses across the five principles of justice defined above are
summarised. Furthermore, the degree of inequality aversion implied by the median
9
respondent is calculated for each distribution. Following previous work, this is done by
fitting a social welfare function through the two outcomes where the median respondent is
indifferent, and then solving for the inequality aversion parameter. A symmetric social
welfare function with constant elasticity of substitution (CES) is used:
[
]
where W is the level of health related social welfare; Hi (≥ 0) is the level of years of life in
full health for population group i, where i = R for the richest fifth and i = P for the poorest
fifth; and r (≥ -1, ≠ 0) is the inequality aversion parameter. When a respondent is indifferent
between two programmes A and B, these two outcomes are on the same social welfare
contour. At the mid-point of A and B, the gradient of the contour is:
( )
[( ( ) ( )) ⁄
( ( ) ( )) ⁄]
which is also (approximately) equal to:
( ) ( )
( ) ( )
From this relationship, a unique value of r can be obtained, either mathematically (Abásolo,
Tsuchiya, 2004) or by using the solve tool in a spread sheet (Dolan, Tsuchiya, 2011). Once
the value of r is obtained, the weight given to a marginal improvement in the health of the
poorest fifth relative to the health of the richest fifth can be calculated from the marginal rate
of substitution:
[ ]
To explore the four hypotheses set out above, combinations of questions are used. To
examine (1) small versus unrealistically large health inequality reductions, the first (large-
individual) and second (small-individual) questions are used, within each of the two samples
(citizens’ panel and online non-academic). For (2) population-level versus individual-level
descriptions of health inequality reductions, the second (small-individual) and third (small-
population) questions are used, within each sample (citizens’ panel and online non-academic).
To explore (3) concrete versus abstract intervention scenarios, the third (small-population)
and fourth (concrete) questions are used, within the citizens’ panel sample. To examine (4)
online versus discussion group mode of administration, the first, second, and third questions
from the citizens’ panel sample are compared with the corresponding questions in the online
non-academic sample. And finally, to explore (5) academic vs non-academic respondents the
10
three questions available in the online survey are compared across the two sub-samples. In
each case, the cumulative distribution across the 11 ordered response categories (from less to
more egalitarian) are compared using the Wilcoxon rank-sum test. In addition, the
proportions of non-egalitarian (pro-rich and health maximising) and strong egalitarian
(maximin and strict egalitarian) responses are compared using the chi-square test.
3. Results
3.1 The sample
Table 1: Descriptive statistics of the discussion group and on-line survey respondents
Discussion
group (N = 29)
Online group:
non-academic
(N = 83)
Online group:
academic
(N = 46)
Baseline Statistic n Statistic n Statistic n
Male (%) 37.9% 11 32.5% 27 32.6% 15
Age (%)
Under 18 0.0% 0 0.0% 0 2.2% 1
18-34 27.6% 8 18.1% 15 39.1% 18
35-49 20.7% 6 15.7% 13 39.1% 18
50-64 31.0% 9 42.2% 35 17.4% 8
65+ 20.7% 6 24.1% 20 2.2% 1
Mean deprivation quintile (mean)
(1 = most deprived; 5 = most affluent) 3.21 29 3.17 83 3.39 46
Social attitude statements* (mean)
(1 = strong agree; 5 = strong disagree)
The creation of the welfare state is one
of Britain's proudest achievements. 1.41 29 1.36 82 1.37 46
Government should redistribute
income from the better-off to those
who are less well off.
3.00 29 2.05 82 2.07 46
*1 suggests most egalitarian and 5 suggests most non-egalitarian
The first two samples have similar age, gender and deprivation characteristics, though the
academic online sample is younger and slightly more affluent than the other two samples.
Respondents in both online groups are slightly more likely to have egalitarian social attitudes
than respondents in the discussion group. A full list of the stated occupations and how they
were coded as “academic” and “non-academic” is available from the authors.
3.2 The distribution of switching point by question and by sample
Figure 1 represents the distribution of responses across the five principles of health justice,
inferred from the switching points by question and by sample. The first panel is for the
11
citizens’ panel sample and the second and third for the online sample, distinguishing
respondents with “academic” and “non-academic” backgrounds respectively.
Figure 1: Inferred principles of health justice by question and sample design * **
* Pro-rich (AAAAA); health maximiser (=AAAA); weighted prioritarian (BXXXA); Maximiner (BBBB=); Strict
egalitarian (BBBBB)
** The vertical line indicates the location of the median respondent
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
O3: Small-populationquestion
O2: Small-averagequestion
O1: Large-averagequestion
Percentage of respondents
Online mode: academic
Pro-rich
Health maximiser
WeightedprioritarianMaximiner
Strict egalitarian
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
D4: Concretequestion
D3: Small-populationquestion
D2: Small-averagequestion
D1: Large-averagequestion
Percentage of respondents
Discussion mode
Pro-rich
Health maximiser
Weightedprioritarian
Maximiner
Strict egalitarian
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
O3: Small-populationquestion
O2: Small-averagequestion
O1: Large-averagequestion
Percentage of respondents
Online mode: non-academic
Pro-rich
Health maximiser
WeightedprioritarianMaximiner
Strict egalitarian
12
Each stacked bar indicates the proportion of responses ranging from “pro-rich” on the left end
to “strict egalitarian” on the right end. As can be seen, the median respondent in the
discussion sample is “strong egalitarian” (i.e. “maximin” or “strict egalitarian”) in the first
three questions. By contrast, the median respondent in the two online samples are always
“weighted prioritarian” in these same three questions. However, the median respondent in
the discussion group sample shifts to “weighted prioritarian” in the concrete question (this
question was not used on the online sample).
Table 2 summarises the calculated inequality aversion parameter r for the median respondent,
and the associated implied weight for a marginal change in the health of the poorest fifth
compared with the health of the richest fifth. Appendix Table A1 summarises this calculation
for each response category.
Table 2: Implied inequality aversion parameter for the median respondent, using constant elasticity of substitution (CES) social welfare function*
Discussion mode Online mode
(non-academic)
Question Median
response
Inequality aversion
parameter
Implied
weight
Median
response
Inequality aversion
parameter
Implied
weight
Large-average question BBBBB Undefined Undefined BBBBA 9.9 6.8
Small-average question BBBB= Undefined Undefined BBBBA 10.8 8.0
Small-population
question BBBBB Undefined Undefined BBBBA 10.8 8.0
Concrete question BBBBA 10.8 8.0 n/a n/a n/a
* The parameter is undefined in the case of “maximin” and “strict egalitarian” response types.
** In all other cases the median respondent trade-off point is 3.5 (BBBBA). For mathematical reasons, this
corresponds to a slightly higher inequality aversion parameter for small gains compared with large gains.
Table 3 shows the results of statistical tests. Responses were more egalitarian for the large-
average compared with the small-average question in both discussion and online samples,
though tests did not quite reach statistical significance. Responses were slightly more “non-
egalitarian” under the small-average compared with the small-population frames in both
samples, but the p-values were far from significant. Discussion group respondents were
significantly more likely to be “non-egalitarian” under the concrete than the abstract framing,
and also significantly less egalitarian in rank. Discussion group respondents were more likely
than online non-academic respondents to be “strong egalitarian”, though the p-values were
far from significant. Finally, online respondents with an academic background were
substantially and significantly less likely to express extreme inequality aversion.
13
Table 3: Non-parametric statistical tests: all respondents
Comparison: discussion (D) versus online non-academic (O) mode of administration
Difference in % that are strong egalitarian
*
Difference in % that are non-egalitarians
*
Wilcoxon rank-sum equality test on unmatched data
ǂ
Large-average question (D1 – O1) +15.0% (p =0.176) +0.9% (p = 0.808) z =0.837; p = 0.403
Small-average question (D2 – O2) +13.1% (p =0.266) +4.2% (p = 0.535) z = 0.654; p = 0.513
Small-population question (D3 – O3) +13.2% (p =0.264) -2.5% (p = 0.657) z = 1.044; p = 0.296
Comparison: within discussion group Difference in % that
are strong egalitarian*
Difference in % that
are non-egalitarians*
Wilcoxon signed-rank
equality test on matched dataǁ
Large vs small-average question (D1 – D2) +4.3% (p = 0.748) -8.4% (p = 0.246) z = 1.935; p = 0.053
Small-average vs small population question (D2 – D3) 0% (p = 1.00) +8.0% (p = 0.297) z = -0.976; p = 0.329
Large vs small population question (D1 – D3) +4.3% (p = 0.748) -0.4% (p = 0.935) z = - 1.731; p = 0.083
Small-population vs concrete question (D3 – D4) +16% (p = 0.258) -16% (p = 0.082) z = 2.590; p = 0.010
Comparison: within online non-academic group Difference in % that are strong egalitarian
*
Difference in % that are non-egalitarians
*
Wilcoxon signed-rank equality test on matched data
ǁ
Large vs small-average question (O1 – O2) +2.46% (p =0.773) -5.1% (p =0.167) z = 1.915; p = 0.056
Small-average vs small population question (O2 – O3) +0.1% (p =0.991) +1.4% (p =0.767) z = 0.311; p = 0.756
Large vs small population question (O1 – O3) +2.6% (p =0.765) -3.8% (p =0.281) z = 0.979; p = 0.328
Non-academics (Na) versus academics (A) in the online group
Difference in % that are strong egalitarian
*
Difference in % that are non-egalitarians
*
Wilcoxon rank-sum equality test on unmatched data
ǂ
Large-average question (Na1 – A1) +23.8% (p = 0.012) -6.6% (p = 0.114) z = 2.930; p = 0.003
Small-average question (Na2 – A2) +29.2% (p = 0.004) -6.9% (p = 0.283) z = 2.506; p = 0.012
Small-population question (Na3 – A3) +26.8% (p = 0.006) -1.05% (p = 0.838) z = 2.489; p = 0.013
* Strong egalitarians = maximiner or strict egalitarian; Non-egalitarians = pro-rich or health maximisers. p-values provided for test of equality of proportions.
* Negative sign indicates less egalitarian (or non-egalitarian in the third column) in the first group compared to the second group, and vice versa.
ǂ Null hypothesis: two independent samples (i.e., unmatched data) are from populations with the same distribution
ǁ Null hypothesis: both distributions on the matched sample are the same
14
Figure 2 gives a visual representation of the rank-sum test, using the example of the large-
average question comparing the discussion versus online non-academic samples. Along the
horizontal axis are the 11 ranked points at which respondents can be indifferent between the
two programmes. The vertical axis represents the cumulative proportion of respondents who
have switched to the less egalitarian programme A by that point. The stronger is the
inequality aversion, the later is the point at which a respondent switches to A, and thus the
lower is the cumulative curve. In Figure 2, the cumulative curves are similar up to 4 but then
the online group rises more rapidly showing a smaller proportion of respondents giving the
“strong egalitarian” responses of 3 (“maximin”) and <3 (“strict egalitarian).
Figure 2: Cumulative distribution of trade-off points for the large-average question, by mode
of administration (D1 vs O1)* **
* The trade off point represents the point at which the respondent would be indifferent between the less
egalitarian programme A, in terms of the gain in years of life in full health to the poorest fifth in Programme B
** In terms of our five principles of health justice, < 3 is "strict egalitarian", 3 is "maximin", 3.5 through to 6.5
are "weighted prioritarian", 7 is "health maximisation" and 7+ is "pro-rich".
Figure 3 shows a cumulative rank comparison within the discussion group sample, comparing
questions three and four i.e. abstract and concrete versions of the small-population
presentation. This time the curves start to diverge early on, with a higher proportion of <3
(“pro-rich”) and 3 (“health maximiser”) responses under the concrete question frame.
Wilcoxon rank sum test: p = 0.403
0.0
0.2
0.4
0.6
0.8
1.0
Cum
ula
tive
resp
onse
pro
port
ion
>7(PR)
7(HM)
6.5(WP)
6(WP)
5.5(WP)
5(WP)
4.5(WP)
4(WP)
3.5(WP)
3(MM)
<3(SE)
Trade off point in years (inferred principle)
Discussion Online (non-academic)
15
Figure 3: Cumulative distribution of trade-off points for the small population vs concrete
question for the discussion group (D3 vs D4)
4. Discussion
A number of studies have quantified societal values in health, including preferences for
severity, end of life, burden of illness, and inequality aversion. While these studies elicit and
report public preferences, most do not have methodological explorations of framing effects
that may bias the final results. This paper contributes towards filling this gap in the context
of health inequality aversion, by examining four possible framing effects: (1) small instead of
unrealistically large health inequality reductions; (2) population-level instead of individual-
level descriptions of health inequality reductions; (3) concrete instead of abstract intervention
scenarios; and (4) online instead of face to face mode of administration. We also explore a
sample selection issue related to framing in the broad sense of how people understand and
process cognitive tasks: (5) respondents with an academic versus non-academic background.
We find evidence of an anti-egalitarian concrete scenario effect (3) and an anti-extreme-
egalitarian academic sample selection effect (5). Within the discussion sample the proportion
expressing zero inequality aversion rose substantially in the concrete scenario, and within the
online sample the proportion expressing extreme inequality aversion was substantially higher
among respondents with an academic background. There is less support for our hypothesised
anti-egalitarian small health gains effect (1), our hypothesised pro-egalitarian population
perspective effect (2), and our hypothesised pro-egalitarian group discussion effect (4).
Wilcoxon rank sum test: p = 0.010
0.0
0.2
0.4
0.6
0.8
1.0
Cum
ula
tive
resp
onse
pro
port
ion
>7(PR)
7(HM)
6.5(WP)
6(WP)
5.5(WP)
5(WP)
4.5(WP)
4(WP)
3.5(WP)
3(MM)
<3(SE)
Trade off point in years (inferred principle)
Abstract small-pop Concrete small-pop
16
However, in each case the pattern of findings was in the expected direction and it is possible
the samples were under-powered to detect small effects of this kind – especially for (1) which
had p-values that were very nearly significant at the 5% level.
From a methodological perspective it is reassuring that we found no substantial and
significant framing effects of using small rather than unrealistically large health gains (1), or
using population total rather than average presentations (2), or using an online rather than
discussion group mode of administration (focusing on the non-academic sample) (4).
However, the two substantial and significant effects we did find are cause for concern and
suggest that further research is needed to refine the value elicitation methods in this area
before findings can be used to inform decision making.
One possible explanation for the concrete scenario effect (3) is that the intervention in
question (a reminder letter to promote uptake of bowel cancer screening) was an “agentic”
intervention to promote individual health behaviour change, rather than a “structural”
intervention to alter the social determinants of health. “Agentic” interventions may make
concepts of individual preferences and responsibility for health behaviour more salient in
respondents’ minds, thus reducing aversion to health inequality. If so, health inequality
aversion parameters elicited from abstract scenarios may over-estimate public concern for
reducing health inequality through “agentic” interventions.
The “academic” sample selection effect (5) cannot plausibly be attributed to anti-egalitarian
social attitudes among those coded here as “academics”. The two social attitude questions
reported in Table 1, on attitudes to the welfare state and income redistribution, reveal that the
two online sub-samples had very similar attitudes to each other (the same mean responses to
two decimal places) and the “academic” sub-sample was slightly more egalitarian than the
discussion group sample. Yet the academic sub-sample was substantially less likely to
express extreme health inequality aversion than respondents in either of the other two
(sub)samples. A possible explanation is that some people with a specific type of training or
professional background are more comfortable than the broader general public in handling
cognitively demanding tasks involving trade-offs between competing social values, and are
less likely to stick religiously to the simple decision making heuristic, “always choose the
more egalitarian option”. And we may be observing that the “academic” sub-group has a
higher proportion of such individuals. If so, there may be a case for developing creative new
17
ways of helping members of the general public think through these trade-off problems
without getting “locked in” to simple decision heuristics.
In the analysis deriving the inequality aversion parameter, we have assumed that symmetry
holds and that public views on unfair health inequality can be captured by a single parameter
(r), regardless of the source of health inequality. An interesting thought experiment is to
imagine how respondents might have responded if they were told that the richest fifth had
worse health. In that case, would they support prioritising the health of the richest fifth, as a
symmetric health related social welfare function would require? Or would they be less
egalitarian, which would be in line with an asymmetric health related social welfare function.
More generally, would their degree of inequality aversion change if they were presented with
health inequality trade-offs between different social groups – for instance, by parental income,
ethnicity, mental illness, illicit drug use, or some other social factor? If so, a single parameter
social welfare function may be unable to provide a full representation of public views on
health justice. The study does not test the symmetry assumption, or other aspects of public
views on health justice, so the possibility remains that the inequality aversion parameters and
implied weights are misspecified. However, the relative magnitudes and the analysis of the
11 response categories are not affected by this.
Furthermore, the visual aids used show health gains, and not the baseline or final distributions
of health. In this respect, this study may be criticised for favouring gain egalitarianism over
outcome egalitarianism (Tsuchiya, Dolan, 2009). However, since identical framing with a
fixed ratio of gains has been maintained across the four questions, we expect this effect to
cancel out in the comparisons across pairs of questions that we carried out.
In order to develop estimates of public aversion to health inequality that are sufficiently
reliable for decision makers to use as benchmarks in distributional cost-effectiveness analysis,
we need a better understanding of the nature and extent of framing effects and the psychology
of how people understand and process questions about complex social trade-offs.
18
References
Abásolo, I., Tsuchiya, A, (2004) Exploring Social Welfare Functions and violation of
monotonicity: An example from inequalities in health. Journal of Health Economics 23(2):
313-329
Abásolo I, Tsuchiya A, (2013) Is more health always better for society? Exploring public
preferences that violate monotonicity, Theory and Decision, 74: 539-5
Asaria, M, Griffin, S, Cookson, R, Whyte, S, Tappenden, P. (2013). Distributional cost-
effectiveness analysis of health care programmes. CHE Research Paper 91. Centre for
Health Economics, University of York, York.
Brazier J, Rowen D, Mukuria C, Whyte S, Keetharuth A, Hole AR, Tsuchiya A, Shackley P.
(2013) Eliciting societal preferences for burden of illness, therapeutic improvement and end
of life for value based pricing: a report of the main survey. EEPRU Research Report 01/13
DeMaio TJ: Social desirability and survey measurement: a review. In Surveying subjective
phenomena. Volume 2. Edited by Turner C, Martin E. New York: Russell Sage; 1984::257-
282.
Dolan, P., Tsuchiya, A. (2011), Determining the parameters in a social welfare functions
using stated preference data: an application to health, Applied Economics, 43(18): 2241-2250
Linley WG, Hughes DA (2013), Societal views on NICE, Cancer Drug Fund and Value-
Based Pricing criteria for prioritising medicines: A cross-sectional survey of 4118 adults in
Great Britain, Health Economics, 22(8): 948-964
Norman R, Hall J, Street D, Viney R. (2013) Efficiency and equity: a stated preference
approach. Health Economics 22:568-581
Schwappach D (2003), Does it matter who you are or what you gain? An experimental study
of preferences for resource allocation, Health Economics. 12(4):255-262
Shah K, Tsuchiya A, Hole AR, Wailoo A (2012), Valuing health at the end of life: A stated
preference discrete choice experiment, NICE Decision Support Unit Report, Dec 2012
Skedgel CD, Wailoo AJ, Akehurst RL (forthcoming) Choosing vs allocating: Discrete choice
experiments and constant-sum paired comparisons for the elicitation of societal preferences,
Health Expectations, doi: 10.1111/hex.12098
Tsuchiya, A., Dolan, P. (2009), Equality of what in health?: Distinguishing between outcome
egalitarianism and gain egalitarianism, Health Economics, 18: 147-159
Tetlock, P. E. (2003). Thinking about the unthinkable: Coping with secular encroachments on
sacred values. Trends in Cognitive Science, 7, 320–324.
Tetlock, P. E., Kristel, O., Elson, B., Green, M., & Lerner, J. (2000). The psychology of the
unthinkable: Taboo trade-offs, forbidden base rates, and heretical counterfactuals. Journal of
Personality and Social Psychology, 78, 853–870.
19
APPENDICES
Appendix Table A1: Calculation table to estimate the inequality aversion parameter of a symmetric social welfare function with CES
for the large-average question (source: Abasolo and Tsuchiya 2004)
Types of preference
(choice behaviour)
Trade-off
point
LE* of the rich
(Prog A)
LE of the rich
(Prog B)
LE of the
poor (Prog A)
LE of the
poor (Prog B)
Inequality
aversion
parameter
Implied
weight
1 Pro-rich (AAAAA) >7 81 77 65 Undefined [**] [**]
2 Health maximiser
(=AAAA) 7.0 81 77 65 69 -1.00 1.00
3 Weighted prioritarian
(BAAAA) 6.5 81 77 65 68.5 -0.21 1.15
4 Weighted prioritarian
(BIAAA) 6.0 81 77 65 68 0.67 1.34
5 Weighted prioritarian
(BBAAA) 5.5 81 77 65 67.5 1.67 1.60
6 Weighted prioritarian
(BBEAA) 5.0 81 77 65 67 2.86 1.98
7 Weighted prioritarian
(BBBAA) 4.5 81 77 65 66.5 4.34 2.57
8 Weighted prioritarian
(BBBEA) 4.0 81 77 65 66 6.40 3.70
9 Weighted prioritarian
(BBBBA) 3.5 81 77 65 65.5 9.87 6.85
10 Maximin (BBBB=) 3.0 81 77 65 65 Undefined Undefined
11 Strict egalitarian
(BBBBB) <3 81 77 65 Undefined Undefined Undefined
* LE: Life expectancy
** Assuming symmetry, the implied preference is inequality seeking.