Predicting outcomes following cognitive behaviour therapy in child anxiety disorders: the inﬂuence of genetic, demographic and clinical information

,


Introduction
Gene-environment interaction (G 9 E) in the context of psychological disorders is predominantly studied within a diathesis stress framework, which proposes that individuals carrying genetic vulnerabilities are disproportionately likely to be adversely impacted by an environmental stressor.However, this model posits a truncated form of G 9 E, focusing exclusively on adversity and negative outcomes, neglecting positive environments and adaptive outcomes.An alternative framework, the differential susceptibility hypothesis, addresses this issue (Belsky & Pluess, 2009).Individuals considered 'vulnerable' (strongly affected by adversity), may also benefit most from supportive environments.Thus, individual differences in developmental plasticity may result in genetic influences that act in a 'for better and for worse manner' (Belsky & Pluess, 2009).Derived from this framework is the concept of vantage sensitivity, which proposes that individuals will vary (for genetic and other reasons) in the extent to which they gain benefit from positive and enriching environmental influences (Pluess & Belsky, 2012).
One example of vantage sensitivity could be response to psychological intervention.For example, while many people experience positive outcomes of Cognitive Behaviour Therapy (CBT) for mood and anxiety disorders, a large minority (35%-45%) retains significant impairments.Very few studies have investigated the source of this individual variation in treatment response despite the potential for stratified medicine and improved outcomes.Biological measures, for example genetic or physiological factors, have rarely been investigated as the source of individual differences in response to psychological interventions (Pluess & Belsky, 2012).Within the child anxiety treatment field, clinical and demographic risk factors (such as age, gender, pretreatment severity, comorbid disorders) have proven to be modest and somewhat inconsistent predictors of treatment response.This inconsistency may in part arise from relatively small sample sizes and differences in assessment of outcome.One approach posits that the most likely source of predictors of treatment response may be the origins of the disorder; that is, 'cause should inform cure' (Uher, 2008).Genetic variants therefore represent plausible predictors of psychotherapy response.Testing for an interaction between a therapeutic intervention and a genetic variant represents a special case of G 9 E (Uher, 2011) and provides an investigation of the vantage sensitivity concept.In a therapeutic G 9 E study, the environment is positive and predictable, allowing for prospective analysis.
At present, there is modest preliminary evidence that G 9 E interactions are relevant not only in the development but also the remission of psychopathology.One method by which predictive power can be potentially enhanced is through the use of analytic techniques that aggregate across multiple polymorphisms and/or genes into a single predictive parameter or risk score.In the present study, we present a first preliminary test of the combined predictive value of both genetic variants and measures of demographic and clinical factors within a therapeutic G 9 E design.With an end point focus and hence greater clinical utility, our outcome was nonremission from the primary anxiety diagnosis following CBT in 384 children with anxiety disorders.Our aim here was not to provide a definitive method for determining risk scores, but to use this approach to illustrate one way in which G 9 E data could be repurposed to potentially provide clinical utility.Our goal was to create a risk index in which scores of 0, 1 or 2 (to allow ease of use for clinicians) were allocated for each clinical and genetic variable such that those scoring high on this scale would be the individuals least likely to benefit from current CBT protocols, providing the opportunity to offer an enhanced treatment from the outset.Although at present it is expensive and impractical for clinicians to obtain DNA samples from clients as part of routine practice, in the future this may become a more viable option given the momentum of the field of individualised medicine.

Method
Participants Three hundred and eighty-four children aged 6-13 years who met DSM IV criteria (American Psychiatric Association, 1994) for a primary diagnosis of an anxiety disorder were selected.Participants were excluded from the analyses if any of the genetic, demographic or clinical variables (see below) were missing or if they had significant physical/intellectual impairment, psychoses and concurrent treatment (N = 186, initial sample was 570). 2 Subjects were recruited from four trials at the Centre for Emotional Health at Macquarie University, Sydney and from two trials at the Berkshire Child Anxiety Clinic, University of Reading (see online Appendix S1 for further information on treatment format).

Measures
Child diagnosis.Diagnoses were made at pretreat- ment and follow-up with the Anxiety Disorders Interview Schedule for DSM-IV, Parent and Child versions (Silverman & Albano, 1996).Diagnoses and Clinical Severity Ratings (CSR; 0-8) were assigned by graduate or clinical psychologists based on composite parent/child report.Where the child met diagnostic criteria and received a CSR of 4 or more, diagnosis was assigned.Children were allocated a primary diagnosis (most interfering) as well as comorbid diagnoses (see online Appendix S2 for further information).
Ethnicity.Child ethnicity was based on reports of grandparent ancestry.As we wanted to develop a risk index of broad clinical utility, we included all individuals in our analyses regardless of ethnicity.The percentage of participants within each ethnic subgroup was: white European (62.8%);African or Caribbean (0.3%); Asian (1.3%); Arab and Middle Eastern (1.6%); Mixed (7.6%) and Ancestry unknown and missing data (26.6%).

Procedure
Ethical approval was received from Human Ethics and Biosafety Committees at both sites.Parents provided informed consent, children provided assent.Buccal swabs were collected either at the clinic or through the post.Families provided data at pre-, post-(N = 373; 97.1%) and follow-up (N = 384; 100%).The follow-up point differed across trials: 52 were assessed at 3 months, 307 at 6 months and 25 at 12 months.

Statistical analysis
We classified treatment outcome on the absence/ presence of the primary anxiety disorder.We focused on the outcome at follow-up as this was where we saw the strongest effect in our previous studies (Eley et al., 2012;Lester et al., 2012).To aid in computing a cumulative risk score for nonremission (primary anxiety disorders still present at follow-up), all predictor variables were di-or trichotimised.We focused on the alleles associated with risk for poor treatment outcome, the L allele for 5HTTLPR and the C allele for NGF rs6330.A recessive model was used for 5HTTLPR (i.e., SS coded as 0; SL/LL coded as 1), and an additive model for NGF rs6330 (TT coded as -1; CT coded as 0 and CC coded as 1 [For the purposes of calculating the risk score in later analyses, TT was scored as 0, CT was scored as 1 and CC was scored as 2]) in line with previous analyses (Eley et al., 2012).The remaining variables were coded as follows: age (0 = lower 50%, 1 = upper 50% of distribution); gender (0 = male, 1 = female); pretreatment anxiety severity (0 = CSR of 4-6, 1 = CSR of 7-8); comorbid mood disorders (0 = absence, 1 = presence); comorbid externalising disorders (0 = absence, 1 = presence) and parental psychopathology (0 = neither parent affected, 1 = 1 parent affected, 2 = both parents affected).
We tested the predictive performance of the risk index using a cross-validation technique (Hastie, Tibshirani, & Friedman, 2009).3We performed five rounds of cross-validation using a repeated random subsampling validation method.On each round, the sample was partitioned into a training set (80% of the sample) and a validation set (20% of the sample).For each round of cross-validation, multiple linear4 regression analyses with robust standard errors were modelled in the training set to provide parameter estimates (unstandardised beta coefficients) for each predictor of treatment outcome (coded 0 = absence of primary anxiety diagnosis, i.e. remission, and 10 = presence of primary anxiety diagnosis, nonremission).5 Risk scores were then calculated for each individual in the validation set using two methods: (a) by assigning a score for each predictor variable from the unstandardised beta coefficient estimated in the training set and (b) by assigning a score for each predictor using a mean unstandardised beta coefficient computed across the five training sets and then rounded to the nearest integer.This latter approach is appealing as by averaging parameter estimates across training sets, variability is reduced.Furthermore, while rounding parameter estimates may somewhat reduce precision, it may increase clinical utility through its simplicity.In both approaches, the risk score took the following form: Predictive accuracy was assessed in each validation set by testing the extent to which the risk score6 correctly classified remission and nonremission.Area under the curve (AUC) values were computed using nonparametric receiver operating characteristic analyses.Finally, we report exploratory analyses that investigate the extent to which the rounded risk score was associated with treatment outcome in the entire sample (N = 384).

Descriptive statistics
Table 1 provides descriptive data and test statistics for the genetic, demographic and clinical predictors comparing training and validation sets for each of the five cross-validation partitions.Within each cross-validation partition, the training and validation sets did not differ significantly on any variables reported.More broadly, rates of remission did not differ significantly between the Reading and Sydney sites, v 2 (1) = 0.81, p = .37.The rates of remission for the entire sample were 60.9% (n = 234).ipants were at a significantly greater risk of retaining their primary anxiety diagnosis at follow-up.Gender, primary anxiety severity, comorbid mood disorders and comorbid externalising disorders also significantly predicted treatment outcome, albeit not consistently across the five training sets.Where these variables were significant predictors, poor outcome was associated with being female, greater anxiety severity pretreatment and the presence of comorbid mood and externalising disorders.Parental psychopathology, 5HTTLPR genotype and age did not make a significant contribution to the prediction of treatment outcome in these analyses.However, the purpose of this step was not to determine statistical significance per se but, instead, to estimate effect sizes for each predictor, irrespective of significance level, that could then be carried forward to calculate a risk score in the validation sets.

Predicting treatment outcome in the training set
Calculating and testing the predictive fit of a risk score in the validation set As described above, we calculated two risk score variants: sample specific and rounded risk scores.To test how well the risk scores predicted outcome, we conducted logistic regression analyses with robust standard errors with risk score entered as a continuous predictor of presence (1) or absence (0) of primary anxiety disorder at follow-up.Table 3 (top panel) reports descriptive statistics for the sample specific risk score, odds ratios, p values and area under the curve (AUC) values with associated confidence intervals.
There was some variability in the predictive performance of the sample specific risk score across the five validation sets, with two of the five models providing statistically significant prediction (p < .05) of treatment outcome in the validation data set.However, as anticipated, all models reported odds ratios exceeding 1 (range 1.32-1.69),indicating that as risk score increased, the odds of nonremission (i.e., still having primary anxiety diagnosis) increased.For example, an odds ratio of 1.69 was identified in validation set 2: this indicates that the odds of nonremission increased by 1.69 times for every single point increment on the risk score.For the sample specific risk score, AUC7 ranged from 0.62 to 0.68 with a mean value of 0.65.
Table 3 (bottom panel) reports summary statistics for the cross-validation rounded risk score.Four of the five models attained statistical significance.All models reported odds ratios exceeding 1 (range 1.38-1.70),again indicating that as the rounded risk score increased, the odds of nonremission also increased.For example, an OR of 1.70 was reported in validation set 2 indicating that the odds of nonremission increased by 1.7 times for every point increment on the risk score.AUC ranged from 0.65 to 0.69, with a mean AUC of 0.66.The rounded risk score and sample specific risk score therefore discriminated between remission and nonremission to a similar extent.Age and parental caseness had rounded scores of zero indicating that they did not contribute to prediction of outcome over and above the other variables in the model.As a result of the zero scores, age and parental caseness do not contribute to the risk index.

Characterising treatment outcome as a function of risk score in the entire sample
As a final step, we performed exploratory analyses to investigate the extent to which the rounded risk score was associated with treatment outcome in the entire sample.From a statistical perspective, this has the obvious limitation that it is not an independent sample from that in which the risk score was developed.However, these exploratory analyses were undertaken solely to demonstrate the potential clinical utility of the risk score approach for predicting treatment outcome.Figure 1 depicts the distribution of risk scores in remitters and nonremitters.The distribution of risk scores was shifted to the right for nonremission compared with remission with a mean risk score of 3.64 (SD 1.45) and 2.87 (SD 1.30), respectively, t(382) = À5.44,p < .001,d = 0.57. Figure 2 reports the percentage of participants with each risk score who retained their primary anxiety at follow-up.Given the low number of participants for some scores within the scale, we divided the sample into the following four risk categories: 0-2, 3, 4 or 5-8.
Nonremission increased in a linear manner with each risk category with just 23% of those scoring 0-2 retaining their primary anxiety disorder at follow-up compared with 62.3% of those scoring 5-8.Risk score significantly predicted treatment outcome at follow-up (OR = 1.51 [95% CI: 1.28-1.77],p < .001).AUC was also commensurate to that seen in the validation sets (0.66).8 (see online Appendix S4 for further information on sensitivity and specificity results.)

Discussion
This paper presents a novel risk score approach to combining genetic, demographic and clinical data to predict nonremission to psychological therapy for child anxiety disorders.In doing so, we illustrate one clinically informative method of repurposing G 9 E data that may assist in the identification of individuals who may require an enhanced treatment package.Our combined clinical trial data set also permitted greater clarification regarding the importance of clinical factors in predicting remission of anxiety diagnoses.We found that greater pretreatment severity, comorbid mood disorders, comorbid externalising disorders and female gender were all significantly associated with poorer remission rates in at least one of our 5 validation sets when controlling for other variables (including genetic factors).In line with our hypothesis, the risk score combining genetic and demographic/clinical factors had an odds ratio significantly greater than 1 in all analyses.This indicates that the odds of nonremission increased for every one-point increment on the risk score.The risk index showed moderate predictive ability, with a mean AUC of .65 and .66 using the sample specific and rounded risk scores respectively.Finally, in the entire sample, the risk score was significantly higher in nonremitters than remitters and those with scores in the highest risk band (5-8) were almost three times as likely not to remit following treatment as those in the lowest band (0-2).

Implications: predictors of treatment outcome
Previous research examining nongenetic predictors of treatment outcome for child anxiety has been limited by low power, resulting in inconsistent findings.In our larger sample, we were able to examine the role of demographic and clinical factors whilst also taking into account preliminary genetic findings.As previously demonstrated (see Rapee, Schniering, & Hudson, 2009), higher pretreatment severity of child anxiety predicted poorer outcome, even after accounting for both 5HTTLPR and NGF rs6330 genotypes.In line with some (Liber et  Finally, we also demonstrated that female gender was associated with poor treatment outcome.The majority of studies comparing treatment outcome for girls and boys have shown that gender does not moderate treatment outcome (Rapee et al., 2009).The current findings cannot be explained by increased severity or comorbidity as the effect was evident after controlling for both.It is possible that factors associated with increased anxiety in girls also increase the likelihood of nonremission.This finding requires replication and further investigation.
Of note, neither age nor parental psychopathology were significant predictors of outcome, although our sample was restricted to children 13 years and under.Previous evidence regarding the role of parental psychopathology in treatment outcome following CBT for child anxiety is mixed (Rapee et al., 2009).In our analyses, parental psychopathology was not significant in any of the five validation sets, and the rounded mean beta coefficient was zero.If entered alone, parental psychopathology did predict outcome in this sample, but this effect was no longer present once other variables were included.However, we also note that our measure of parental psychopathology relied solely on a self-report questionnaire which may have led to underreporting of parent symptoms, thus reducing the strength of this variable as a unique predictor of child outcome (Bogels & Brechman-Toussaint, 2006).Measurement of parental psychopathology using structured diagnostic interviews may result in a more accurate picture of symptoms.

Implications: risk index for poor treatment outcome
Our risk index created from genetic, demographic and clinical factors was moderately good at predicting remission at follow-up.These results provide early promise that combining genetic information with clinical and demographic risk factors may be an informative approach to predicting individual differences in response to psychological treatment.Our data also support for the idea that genetic factors (over and above demographic and clinical variables) moderate response to a positive environmental influence, namely psychological therapy (Pluess & Belsky, 2012).Our risk index had similar predictive ability regardless of whether we used the sample specific or rounded beta coefficients.The next step is clearly to replicate these results, and to extend them, taking into account other potential predictors of nonremission, including additional genetic polymorphisms.What is more encouraging is that those in the highest band of our risk index were almost three times as likely not to remit following treatment as those in the lowest band, suggesting that the approach is worth pursuing further.

Limitations
We note a number of limitations.First, our sample remains small for a G 9 E study, even though it is the largest of its kind to date.Second, we recruited from six trials across two sites.However, not only does this make the data a closer representation of reality than utilising data from a single trial, it also means that our findings rose above the noise our sampling strategy inevitably incurred.Third, we focus here on the follow-up data, as our previous papers indicate that the effects of genotype occurred at this time-point (Eley et al., 2012).We note, however, that the follow-up time-point was not consistent across participants, which may have influenced our findings.We have previously discussed that we believe the late emergence of a significant genetic effect may be explained by continued practising of techniques learned during CBT.It is possible that what our risk index measures is the likelihood of a child continuing to benefit from treatment once it has ended.Fourth, this study focuses exclusively on two candidate genes.We anticipate in the future that multiple genes (of very small effect) will be identified to predict treatment outcome.The field of therapygenetics is in its infancy and future work will be better placed to use polygenic risk scores.Finally, although our sample size allowed a greater number of children with mood disorders to be identified compared with smaller studies, there remained a low prevalence of mood disorders (8.9%), resulting in large standard errors around their odds ratios.Replication of these findings will be important.
In conclusion, we have illustrated that combining genetic, demographic and clinical data into a risk score is an informative means of repurposing G 9 E data.Our data provide preliminary evidence that children with a high number of genetic, demographic and clinical risks are least likely to benefit from standard CBT.If replicated, a risk index using both genetic and clinical information could be applied in a clinical context to help decide whether a child is likely to benefit from standard CBT alone or whether enhanced treatment is required.

Table 1
Table 2 reports unstandardised beta coefficients for each predictor variable, and p values.We also report mean unstandardised beta coefficients calculated across the five cross-validation partitions, and rounded beta values based on these mean beta coefficients.Genetic, demographic and clinical predictors in combination significantly predicted treatment outcome, with all model p values <.05.The proportion of the variance in treatment outcome accounted for by the model (R 2 ) approximated 8%.The most consistent predictor of treatment outcome was NGF rs6330 genotype, which significantly predicted treatment outcome in all five data sets (p values from .008 to .03).With each extra C allele, partic-Sample characteristics for the entire sample and test statistics comparing training and validation sets for each cross-validation sample b Mean (SD); c N; d Mean (SD); % in severe primary anxiety severity category (CSR 7-8).Analyses: e Test statistics compare frequencies of a binary variable (e.g., young versus old age group; severe versus moderate primary anxiety severity).Bold values indicate beta values significant at p > .05

Table 3
Model fit statistics and logistic regression analyses using the sample specific and rounded risk score as a predictor of treatment outcome in the five validation sets

of remission category (remission/non- remission) Risk score
al., Distribution of risk scores for participants with and without their primary anxiety disorder diagnosis remaining at follow-up © 2013 The Authors.Journal of Child Psychology and Psychiatry published by John Wiley & Sons Ltd on behalf of Association for Child and Adolescent Mental Health.