Mind the (explanatory) gap - California State University, Fullerton

Daniel R. Cavagnaro
California State University Fullerton
Gabriel Aranovich
University of California San Francisco and SF VA Medical Center
• Psychiatry has relied on a diagnostic system based on
presenting signs and symptoms, with the result that “current
definitions do not adequately reflect relevant neurobiological
and behavioral systems.” (Cuthbert & Insel, 2013)
• NIMH RDoC program created to “develop new ways of
classifying mental disorders based on behavioral dimensions
and neurobiological measures.” (NIMH website)
• Use tools from game theory, behavioral economics, and machine
learning to “characterize mental dysfunction in terms of
aberrant computations” (Montague et al., 2012)
• Aims:
• Refine diagnostic system
• Derive cognitive phenotypes that will structure the search for genomic and
neural correlates.
• Mental disorders under investigation:
• Obsessive Compulsive Disorder (OCD)
• Hoarding Disorder (HD)
• Compare groups in terms of “dimensions” of risk preference, as
measured by estimated parameters of Cumulative Prospect
Theory (CPT)
• Risk aversion
• Probability weighting (curvature and elevation)
• Presence of obsessions and/or compulsions.
• Obsessions: recurrent and persistent thoughts, urges, or images that are
experienced as intrusive and unwanted.
• Compulsions: repetitive behaviors or mental acts that an individual feels
driven to perform in response to an obsession or according to rules that
must be applied rigidly.
• Prominent model of OCD holds that OCD involves excessive risk
aversion and intolerance to uncertainty
• Experimental results using self-report measures and gambling
tasks have been mixed (Morein-Zamir et al. 2014)
• “Cognitive theorists have postulated that people who develop OCD
overestimate the likelihood and severity of danger…and are unable to
tolerate uncertainty and ambiguity that could possibly signal threat.”
--Barlow et al. (2002):
• Persistent difficulty discarding or parting with possessions,
regardless of their actual value.
• Perceived need to save items and distress associated with
discarding them.
• Lawrence et al. (2006) found “Link between hoarding and risky
behavior on the Iowa Gambling Task.” OCD did not differ from
controls in that study.
• “Decision-making has been suggested to be impaired in compulsive
hoarding. Hoarding participants may be risk averse, as demonstrated by
their avoidance of discarding because of hypothetical future need.”
-- Grisham et al. (2010):
• “A limited understanding of cognitive factors associated with compulsive
hoarding may contribute to the modest treatment outcomes for this disorder.”
--Grisham et al. (2010)
• Overlap between neural regions associated with risk computation
and dysfunction in OCD and HD. (Hsu et al. 2009; Cavedini et al.
2002; Preuschoff et al. 2008; Admon et al., 2007)
• Some experts suspect that the literature is off, and that OCD/HD are
both actually less risk averse than healthy controls
• Three components of the CPT model
• Value function
• “Classical” notion of risk aversion (e.g., Arrow-Pratt)
• Probability weighting function (PWF)
• Subjective overweighting or underweighting of probabilities relative to
their normative values, capturing differential treatment of uncertainty.
• Choice function
• Behavioral consistency
• E.g., logistic choice rule
• All gambles in our experiment will be over the same three
possible outcomes: $25, $350, and $1000
• For example:
Which do you prefer?
Gamble A
0.2
$25
0.4
$350
Gamble B
0.4
$1000
0.1
0.6
0.3
$25
$350
$1000
or
• So, without loss of generality, we assume
• u($25) = 0
• u($350) = v
• where 0<v<1
• u($1000) = 1
• Prelec’s (1998) 1- and 2- parameter forms:
• r controls curvature
• r < 1 implies inverse-s shape;
• r > 1 means s-shape
• s controls elevation
• s < 1 implies systematic overweighting of probabilities
• s > 1 implies systematic underweighting of probabilities
• We assume a logistic choice function with one “behavioral
consistency” parameter, b.
𝑝 𝑒𝑟𝑟𝑜𝑟 =
1
1 + 𝑒 𝑏|𝑢
𝑔1 −𝑢 𝑔2 |
• Higher values of b imply more consistent choice behavior.
• b = 0 implies 50/50 chance of choosing each option on each
trial.
• Risky choice task to estimate parameters of the probability weighting
function at the individual and group levels in OCD, HD, and healthy
control samples.
• Sample: 28 HC, 29 OCD, 29 HD
• 50 choices between 3-outcome gambles on $10, $350, and $1000
• Adaptive Design Optimization (Cavagnaro et al., 2013)
• Decision stimuli were tailored to each subject to be maximally diagnostic for
estimating parameter of the CPT model.
• Hypotheses:
• Group differences in v, r, s parameters of the CPT model;
• OCD+HD more risk averse than controls -> greater v parameter in
OCD+HD; greater s parameter?
• OCD+HD closer to EU behavior -> more linear PWF?
• Using ADO, the within-subject estimates of each parameter
were very precise:
• Contrary to expectations, v was parameter lower for OCD+HD
 less risk averse.
• Group differences in r parameter estimates as well.
• Healthy controls display “classic” inverse s-shaped PWF.
• OCD, and to greater extent, HD, display s-shaped PWF.
• S-shaped PWF would predict gambling aversion among
OCD+HD.
• Interesting because it has been suggested that Pathological Gambling
(PG) is an “obsessive-compulsive spectrum” disorder, since PG symptoms
can appear obsessive-compulsive.
• Suggests PWF may help identify subgroups of patients with different
cognitive phenotypes.
• We tested whether parameter estimates related to symptom severity
• YBOCS – predominant self-report measure of OCD signs/symptoms
• In OCD sample, both v and r parameter values predicted compulsion
subscale scores (multiple regression, r parameter coef. = -7.13
(p<0.001), v parameter coef. = -7.02 (p<0.05)) ;
• No relationship with obsession subscale scores.
• UHSS – widely used hoarding severity scale.
• In HD sample, r parameter predicted UHSS score (r parameter coef. =
8.96, p<0.05).
• Hence, r parameter value may differentiate OCD from HD; consistent with
recent literature highlighting differences that culminated in creation of distinct
category for HD in DSM-V.
• We tested whether parameter estimates related to symptom
severity
• Beck Depression Inventory – predominant diagnostic questionnaire for
depression.
• Greater risk aversion predicted lower depression score (v parameter
coef. = -14.78, p<0.01) .
• Prior experimental evidence for relationship between risk aversion
and depression is mixed (Smoski et al, 2008; Chapman et al., 2007).
• All of the reported analyses were based on the 1-parameter
PWF, which does not account for elevation.
• Although the elevation parameter (s) in the PWF improves
model-fit within subjects, it interacts with the other parameters
of the model, making them all more difficult to interpret.
• Correlation between v1 and v2 was 0.36
• Correlation between r1 and r2 was 0.21
• Correlation between v2 and s2 was -0.44
• Significant relationships and group-differences in risk aversion
(v) and curvature (r) disappear when the elevation parameter is
included.
• Hsu et al. (2009) J. Neuroscience
• "We focused on the one-parameter version of the Prelec function, as it fits
only slightly worse than the two-parameter version, and permits simpler
cross-subject comparison (since each subject’s curvature is expressed by
one parameter rather than two)."
• ADO successfully converged to best fitting PWF parameter
estimates at the individual level.
• We found that, OCD & HD patients were less risk averse than
healthy controls. , contrary to a widely-held view in the
literature
• Healthy control group exhibited “classic” PWF shape.
• OCD+HD parameter estimates suggest S-shaped PWF.
• S-shape predicts aversion to gambling and absence of “certainty effect”;
has implications for obsessive-compulsive spectrum disorder hypothesis.
• Parameter estimates were correlated with clinical symptom
scales, suggesting that cognitive phenotypes defined by
parameters may be related to processes fundamental to OCD
& HD.
• r parameter correlations with clinical scales support hypothesis that OCD
and HD are distinct disorders.
• Taken together, results argue for the value of the computational
approach to psychiatry.