Exploratory network analysis of large social science questionnaires

Abstract

There are now many large surveys of individuals that include questions covering a wide range of behaviours. We investigate longitudinal data from the Add Health survey of adolescents in the US. We describe how structural inference for (dynamic) Bayesian networks can be used to explore relationships between variables in such data and present this information in an interpretable format for subject-matter practitioners. Surveys such as this often have a large sample-size, which, whilst increasing the precision of inference, may mean that the posterior distribution over Bayesian networks (or graphs) is concentrated on disparate graphs. In such situations, the standard MC3 sampler converges very slowly to the posterior distribution. Instead, we use a Gibbs sampler (1), which moves more freely through graph space. We present and discuss the resulting Bayesian network, focusing on depression, and provide estimates of how di erent variables a ect the probability of depression via the overall probabilistic structure given by the Bayesian network.

INTRODUCTION

Hypotheses of multifactorial causes of symptoms and outcomes play an important role in the social sciences and in public health. Regression-based approaches are widely-used in these elds to explore such hypotheses. A great deal of insight can be gained through such approaches, but it is sometimes overly constraining to x a particular quantity as the dependent variable, especially if the goal is to explore the possibility of unexpected relationships between the data. Instead, we can consider a number of variables on an equal footing, and study the possibility of unexpected relationships in the data.

Graphical models provide a statistical framework within which the relationship between variables can be studied. These models enable complex multivariate distributions to be decomposed into simpler local distributions. This can reveal a great deal about the relationships between the variables, as well as provide a statistical and computationally tractable description of their (often large) joint distribution. The decomposition is formed by the conditional independence structure, which can be represented by a graph. The use of graphs helps to make the interpretation of the model simpler. In this paper, we focus on the structure of the model, as given by the graph. We aim to make inference about this using statistical model selection. The structure of the model suggests how the di erent components of the system interact, which may be helpful in understanding the system as a whole. These methods have been widely adopted in molecular biology ( 2, 3 ), and have been used in some areas of medical sciences ( 4 ).

Consideration of unexpected relationships between factors requires datasets that incorporate a wide range of topics. Such data is now widely available for representative samples of populations in many countries, and for many sub-groups of interest. Many of these datasets are derived from surveys that are general in scope, and are not collected to study any one particular question. For example, in the US, the health of the whole population is representatively sampled annually for the Behavioral Risk Factor Surveillance System (BRFSS) survey, and the Add Health study, which we use here, followed a cohort of young people from 1994 until 2008. Data from both of these have been used in scores of studies, but these commonly focus on one speci c aspect, often using the data to evaluate existing hypotheses. Given the wide scope inherent in the design of these studies and the large samples available in many cases, it is possible to broaden the scope of the analysis by considering richer structures. In this paper, we discuss the potential that such a more explorative approach yields. We do not seek to make conclusive causal claims, but instead suggest that a broader approach may uncover important aspects that have been neglected.

Our focus will be on depression among adolescents in the US, drawing on data from the National Longitudinal Study of Adolescent Health (Add Health). It is estimated that around 1{6% of adolescents each year are a ected by depression ( 5, 6 ). The e ects of depression in this age-group are wide-ranging ( 7 ), and include the stigma associated with poor mental health more generally ( 8 ). There is considerable evidence that there are a wide range of causal factors for depression amongst adolescents, spanning biological, psychological and social domains. Understanding these causal factors and separating them from the consequences of depression has been recognised as an important aim ( 9 ). Some of the relevant causal factors may interact and the approach taken here accounts for this. The remainder of this paper is organised as follows. We rst introduce the AddHealth dataset and describe the Bayesian network framework. Inference for Bayesian networks is performed using Markov Chain Monte Carlo (MCMC), but the large sample size of the dataset we consider makes achieving convergence di cult because the posterior distribution may be concentrated on disparate graphs, and so we describe an alternative sampler that has superior properties in this situation. Whilst the PC-algorithm ( 10, 11 ) has properties that often make it attractive in such contexts, we found that the results in this situation were not robust (see Discussion). We then present and discuss the results for the Add Health dataset. 2 2.1

MATERIALS AND

METHODS

Add Health

The data that we use are drawn from the National Longitudinal Study of Adolescent Health (Add Health) that explores health-related behavior of adolescents ( 12 ) in the US. The questionnaire contains over 2000 questions that cover many aspects of adolescent behaviours and attitudes. We consider the representative sample of adolescents from Waves I and II of the in-home section, and the parental questionnaire from Wave I of the study. The analysis we perform is not feasible when the data is not complete (see Discussion), and so individuals with missing data were removed from the study. Removing incomplete samples leaves 5975 individuals in the study.

Our measure of depression is a self-assessed scale based upon the Centre for Epidemiologic Studies Depression Scale (CES-D) ( 13 ). Two questions from the 20-item scale are omitted from AddHealth, and two are modi ed, and so we scale the score given by the available questions ( 14 ). A Receiver Operating Characteristic (ROC) analysis showed that thresholds of 24 for females and 22 for males provided the best agreement with clinical assessments of depression ( 15 ). We use this threshold to create a binary indicator of depression status.

Many of the remainder of the variables that we consider (Table 1) are drawn from the risk factors described in the depression literature, and the mental health literature more generally. A recent review ( 8 ) described a wide range of factors that are associated with poor mental health in young people, including gender, poverty, violence and the absence of social networks in the local neighbourhood. The quality of relationships with parents is also thought to be important, especially with the mother ( 16 ), as are parental alcohol problems ( 17 ) and parental discord ( 16 ). The individual's use of alcohol, drugs, smoking and HIV/AIDS are all also associated with depression ( 18, 19 ). Physical exercise has been proposed in some studies as a useful intervention for the management of depression, but many of these studies have been deemed to be poor quality ( 20 ). 2.2

Bayesian Networks

Our study uses Bayesian networks to explore the relationships between variables in the Add Health study. Bayesian networks are a particular type of graphical model that enable classes of probability distributions to be speci ed using a directed acyclic graph (DAG). A Bayesian network G is represented using a DAG with vertices V = (V1; : : : ; Vp), and directed edges E V V . The vertices correspond to the components of a random vector X = [X1; : : : ; Xp]T , subsets of which will be denoted by XA for sets A f1; : : : ; pg. For 1 i; j p, we de ne the parents Gj of each node Vj to be the subset of vertices V such that Vi 2 Gj , (Vi; Vj ) 2 E. Specifying the parents of the vertices determines the edges E of the graph G. We denote by G the space of all possible directed acyclic graphs with p vertices. We will use XGi to refer to the random variables that are parents of Xi in the graph G.

The graph speci es that the joint distribution for X, with parameters = ( 1; : : : ; p), can be written as a product of conditional distributions p(Xi j XGi ; i), given the variables XGi corresponding to the parents of Xi in the graph.

p(X j G; ) = p Y p(Xi j XGi ; i) i=1 We will need to be able to evaluate the marginal likelihood p(X j G) easily, and so we consider only a conjugate analysis in which the conditional distributions p(Xi j XGi ; i) are multinomial, with Dirichlet priors p( i) for each i. In this case, the marginal likelihood can be evaluated analytically. Suppose each Xi takes one of ri values, and de ne qi as the number of levels of the sample space of XGi , each element of which we call a con guration. For each con guration j of XGi , let Nijk be the number of observations in which Xi takes value k. We assume the Dirichlet priors for each i, each with hyperparameters Ni0jk, are independent. We de ne Nij = Prki=1 Nijk and Ni0j = Prki=1 Ni0jk, and the local score p(Xi j XGi ) to be

qi p(Xi j XGi ) = Y j=1

(Ni0j ) Yri (Nij + Ni0j ) k=1 (Nijk + Ni0jk) : (Ni0jk) The marginal likelihood can be shown to equal the product p(X j G) = Qp

i=1 p(Xi j XGi ) of these local scores ( 21 ). 2.3

Structural inference for Bayesian

Networks We aim to make inference about the DAG G, given data X and so our interest focuses on the posterior distribution Pr(G j X) on Bayesian networks. Under the assumptions we have made, this can be written in terms of the marginal likelihood p(X j G), and a prior (G) for the Bayesian network structure.

Pr(G j X) /

p (G) Y p(Xi j XGi ) i=1 The priors (G) can be chosen to encode domain information ( 3 ). For the analyses in this paper, we choose an improper prior (G) / 1 that is at across the space of graphs.

The posterior distribution Pr(G j X) is di cult to evaluate, because cardinality of G grows superexponentially in p. This motivates the use of approximations to Pr(G j X), which are usually based on Markov chain Monte Carlo (MCMC). 2.4

Approximate inference for Bayesian

Networks The standard form of MCMC that is used for structural inference for Bayesian networks is MC3 ( 22 ). This is a Metropolis-Hastings sampler that explores G by proposing to add or remove a single edge from the current graph G. This sampler works surprisingly well in many situations, but if the posterior distribution is not unimodal, the local moves may fail to explore the space fully because the sampler may become `trapped' in one mode. This issue becomes more severe as the sample size increases because the posterior distribution becomes more concentrated. A natural approach in such situations is to use the PC-algorithm ( 10, 11 ), which has been shown to be asymptotically consistent ( 23 ), but we found in this case that the results were not robust (see Discussion).

Our analyses in this paper were performed using a Gibbs sampler ( 1 ), which we found to converge rapidly to its equilibrium state. A nave Gibbs sampler for structural inference that proposes single-edge additions and removals can easily be constructed, but this sampler o ers no advantages over the analogous MC3. This nave scheme, however, can be improved by `blocking' together a number of components, and sampling from their joint conditional distribution. In theory, any group of components can be taken as a block, but sampling from their joint conditional distribution needs to be possible and, ideally, computationally quick.

For Bayesian networks, the most natural blocks are those consisting of parent sets G1; : : : ; Gp. This is natural because the marginal likelihood p(X j G) for a graph G factorises across vertices into conditionals p(Xj j XGj ) and these conditionals depend on the parent set of the vertex. Therefore, since any graph G 2 G can be speci ed by a vector G = (G1; : : : ; Gp) of parent sets, the posterior distribution on Bayesian networks G 2 G can be written as functions of G1; : : : Gp in the following way.

Pr(G1; : : : ; Gp j X) / p (G1; : : : ; Gp) Y p(Xi j XGi ) i=1 In the following, we will denote subsets of the vector G = (G1; : : : ; Gp) by GA = fGk : k 2 Ag, and the subset given by the complement AC = f1; : : : ; pg n A of a set A will be denoted by G A = fGk : k 2 AC g. In particular, the complete graph can be speci ed by G = (G1; : : : Gp) = (Gi; G i) for any i 2 f1; : : : ; pg. To be able to construct a Gibbs sampler using parent sets, we need to nd their conditional distribution, given the other parent sets G j = fG1; : : : ; Gj 1; Gj+1; : : : ; Gpg. Parent sets Gj for which G = (Gj ; G j ) is cyclic will have no probability mass in the conditional distribution. Let Kj? be the set of parent sets Gj such that G = (Gj ; G j ) is acyclic. The conditional posterior distribution of Gj is multinomial, with weights given by the posterior distribution of G = (Gj ; G j ). When the cardinality of Kj? is constrained (for example, by restricting the maximum number of parents of each node) the conditional posterior distribution for Gj 2 Kj? can be evaluated 1.0 ●● 0.8

● n20.6 ● uR00..42 ●● ●●●●●●●●●●● ●●●●● ●● ●●●●●

●●●●●●● ● ●● ● 0.0 ●●●●●

MC3 sampler ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●

Gibbs sampler ● ●●●●●

●●●● 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Run 1 We can improve the speed of convergence of this sampler by allowing pairs of parent sets to be sampled together. At each step of the Gibbs sampler we conditionally sample pairs of parent sets (Gj1 ; Gj2 ), given the remainder of the graph G fj1;j2g. Parent sets G fj1;j2g such that G = (Gj1 ; Gj2 ; G fj1;j2g) is cyclic have no probability mass in the conditional distribution. Let Kj?1;j2 be the set of pairs of parent sets (Gj1 ; Gj2 ) such that G = (Gj1 ; Gj2 ; G fj1;j2g) is acyclic. For (Gj1 ; Gj2 ) 2 Kj?1;j2 , the conditional posterior distribution is multinomial, by analogy with ( 1 ), with weights given by posterior distribution of G = (Gj1 ; Gj2 ; G fj1;j2g).

Pr(Gj1 ; Gj2 j G fj1;j2g; X) = P

Pr(Gj1 ; Gj2 ; G fj1;j2g j X) (Gj1 ;Gj2 )2Kj?1;j2 Pr(Gj1 ; Gj2 ; G fj1;j2g j X) Similarly, sets of three parent sets can be conditionally sampled. Full technical details are presented in ( 1 ).

Depressed (time point 1) Didn't present to doctor ( 2 )

Female ● Didn't present to doctor ( 1 )

Good health ( 2 )

Good health ( 1 )

Victim of violence ( 2 ) Strong academically ( 2 )

Drug user ( 1 ) ● ●

● ● ● ● ●● ● ●●● ●● ●●● ●●

● The variables that we consider are detailed in Table 1. As is common when using graphical models ( 24 ), all of these variables were grouped, initially into `Background', `Wave I' and `Wave II', and then rened into whether the question asked about the longor short-term, as shown in Table 2. These groups dene constraints on the Bayesian networks that are considered. Speci cally, no edges can be directed backwards through the groups. Edges, however, are allowed within groups. For example, no edge is allowed to be directed into `Gender', and no edge can pass backwards in time, for example, from Depression at Wave II to Depression at Wave I. Additionally, no edge can pass from a short-term variable to a longterm variable, for example, from Depressed at Wave I to Have HIV/AIDS at Wave I.

We precomputed the local scores, and then drew 100,000 samples (the rst half of which were discarded as burn-in) using the Gibbs sampler (Section 2.3), which took 30 minutes (on a single core of a cluster computer). The graph space was constrained such that no node had more than 3 parents, to ensure Equation 1 could be evaluated.

We ran 5 independent samplers, with disparate initial states. This enables a simple test of convergence to be performed that compares the posterior edge probabilities obtained from each of the independent runs ( 25 ). The agreement between runs can be examined graphically by plotting the edge probabilities against Been expelled ( 2 )

Victim of violence ( 2 )

Family bereavement ( 1 ) Parents unhappy together ( 1 ) Been expelled ( 1 )

Been suspended ( 2 )

Seen shooting ( 2 ) Family bereavement ( 2 ) Hisp/Latino

Family poor ( 1 )

Black/Af Am

Parent drinks ( 1 ) Householder smokes ( 1 )

Live with father ( 1 )

Experiences prejudice ( 2 ) Live with father ( 2 )

Live with mother ( 1 ) Live with mother ( 2 )

Experiences prejudice ( 1 )

White Parents aid decisions ( 1 )

Mother warm/loving ( 1 ) Mother warm/loving ( 2 )

Alcohol ( 1 )

Age Victim of violence ( 1 )

Strong academically ( 1 ) Good health ( 1 )

Exercises ( 1 )

Have HIV/AIDS ( 2 ) Strong academically ( 2 )

In physical fights ( 2 )

In physical fights ( 1 )

Learning disability

Talks to neighbours ( 1 ) Severely injured ( 1 )

Talks to neighbours ( 2 )

Exercises ( 2 ) Depressed ( 2 )

Severely injured ( 2 ) Skips school ( 2 )

Depressed ( 1 ) Drug user ( 2 )

Didn't present to doctor ( 2 )

Been suspended ( 1 ) Seen shooting ( 1 )

Female

Skips school ( 1 )

Drug user ( 1 ) Smoker ( 1 )

Didn't present to doctor ( 1 ) Alcohol ( 2 )

Smoker ( 2 )

Good health ( 2 )

Asian/Pac Isl.

Am Ind/Nat Am Have HIV/AIDS ( 1 )

Other race each other (Figure 1). Mean Pearson correlation coefcients between edge probabilities from pairs of runs were 0:9999 0:0002 (standard deviation) for the Gibbs sampler and 0:6322 0:0477 for MC3. The agreement between the independent runs of the Gibbs sampler gave us con dence in our results, in contrast to the large disagreements between MC3 runs. In addition, cumulative edge probability plots for each edge showed regular excursions around the mean ( 26 ), and a numerical diagnostic ( 27 ) monitoring the number edges in the sampled graph also clearly suggested that su cient samples had been drawn (R^ 1:0).

The samples drawn using MCMC allow the posterior distribution of Bayesian networks to be approximated. In particular, the samples can be used to estimate the posterior edge probability P (ejX) with e 2 E. Figure 2 displays all edges with posterior probability of at least 0.5.

Our focus is on depression, the parents of which in Figure 2 we observe are \Didn't present to doctor" and \Gender". It important, however, to note that the model does not say that these are the only factors that are important. For example, \Drug user" at Wave I is related to depression through \Didn't present to doctor" at Wave I and II (Figure 2).

This is shown in Figure 3, which gives the conditional probability of being depressed at Wave 2 when a particular variable is set to a speci c value. We see that general health, violence, academic performance and drug use all a ect the conditional probability of depression at Wave II. Note that to compute this probability, links from the parents of the variable in which we `intervene' are removed; this is equivalent to the `do-operator' in the terminology of Pearl ( 28 ). The analysis reveals the interaction between the many aspects of life that have an impact on depression. The connection between the depression and its two parents in Figure 2 have been previously discussed in the literature. The importance of gender in depression is particularly extensively documented in the literature ( 8 ). The connection to a failure in seeking medical care even when the individual thinks they should has also been discussed in the literature, often in terms of poor accessibility of health care services for young people ( 29, 8 ). Several decades of research have revealed the complex causation of depression in young people, as suggested by this study ( 8 ). 4

DISCUSSION

There is a large amount of information held in large social science questionnaires. In this paper we have examined a graphical model approach to inferring structure amongst the variables in such questionnaires. In contrast to the standard regression-based approaches, a graphical model approach forgoes the need to specify a particular variable as the response. Instead, a more comprehensive estimate of the entire structure of the underlying system can be obtained. Regression approaches posit a particular conditional-independence structure, while graphical approaches allow consideration of more general structures.

The limitations of this study include those of all similar studies using observational data that are collected for multiple audiences. These forms of data, including the longitudinal data used here, do not permit strong causal conclusions to be drawn. In particular there may be important variables that we have not included in the analysis. However, the results are consistent with studies that have used other research approaches including experimental designs. The connection between an individual not seeking medical care when they think they should and depression supports current practice guidance in the UK ( 30 ) where there is an emphasis on providing access to health care through the school system rather than expecting young people to seek health care themselves. Not seeking medical care despite believing it should be sought is a complex factor because it captures both barriers to getting medical care within the individual, such as lacking motivation to seek care, and barriers within the individual's environment, such as poor access to care. This may mean that the variable encapsulates a number of di erent characteristics related to depression, and thus may form a `marker' for depression. However, the use of a form of the question \Has there been any time over the past year when you thought you should get medical care, but you did not?" as a screening question in di erent contexts needs further consideration. This method of analysis clari es the complexity of depression and suggests why when using traditional methods of analysis it can be di cult to clarify whether or not factors, such as experiences in the family, in the wider community and at school, impact on the experience of depression for young people. It may also suggest why interventions for prevention of depression have not yet been demonstrated to be cost e ective ( 31 ).

We performed structural inference for the Bayesian network using a Gibbs sampler ( 1 ), because MC3 did not mix in a reasonable time. We have also found ( 1 ) this algorithm to be superior to the REV sampler ( 32 ), and it has the advantage of avoiding the need to consider an order prior as required by order MCMC methods ( 33, 34 ), which induces a bias that can only be corrected exactly by NP-hard computation of a correction factor.

An alternative to the MCMC method used here is the PC-algorithm ( 10, 11 ). This method is computationally e cient and is asymptotically consistent. However, to test whether the sample size available here is su cient to reach the asymptotic regime, we applied the PC-algorithm (without constraints) to 10 di erent subsamples, each containing 90% of the data. We found that these results di ered signi cantly, with a mean 84 in structural Hamming distance between the pairs of completed partially directed acyclic graphs (CPDAGs) given for the subsamples.

We used a Multinomial-Dirichlet model for the local conditional distributions, which yields a closed-form marginal likelihood. This model posits an entirely general discrete distribution, allowing its form to be guided by the data. However, the number of parameters in the local distributions for this model increases exponentially with the number of parents, which may mean that overly-sparse models are preferred. This is problematic when the sample size of the available data is small, because models with many parameters cannot be assessed adequately without a large dataset. The large sample size of the dataset used here minimises this issue, but it would nonetheless be worthwhile to consider more compact parameterisations. However, estimating such models ( 35 ) signi cantly increases the complexity of the model space, which makes such an approach computationally challenging in this setting. For this paper, we removed samples with missing data. It is possible to handle missing data formally, for example by using structural EM ( 36 ), and similarly consider latent variables (e.g. shared genetics driving both child and parent behaviour). However, at present, doing so whilst robustly exploring large model spaces remains an open challenge. Tackling these computational and inferential issues is a key area for future research.

Experiences prejudice

In physical ghts

Didn't present to doc- 2 tor Severely injured 3

Have HIV/AIDS

Seen shooting

Mother warm/loving Been suspended

Been expelled Good health Talks to neighbours Age

Live with mother Live with father Smoker Drinks alcohol

Exercises

Depressed r Question 2 Interviewer, please con rm that R's sex is (male) female. (BIO SEX) 2 Are you of Hispanic or Latino origin? (H1GI4) 2 What is your race? [White] You may give more than one answer (H1GI6A) 2 What is your race? [Black or African American] You may give more than one answer (H1GI6B) 2 What is your race? [American Indian or Native American] You may give more than one answer (H1GI6C) 2 What is your race? [Asian or Paci c Islander] You may give more than one answer (H1GI6D) 2 What is your race? [Other] You may give more than one answer (H1GI6E) 4 [If SCHOOL YEAR:] During this school year [If SUMMER:] During the 19941995 school year how many times HAVE YOU SKIPPED/DID YOU SKIP school for a full day without an excuse? (H1ED2; H2ED2) 3 [If SCHOOL YEAR:] Students at your school are prejudiced [If SUMMER:] Last year, the students at your school were prejudiced. (H1ED21; H2ED17) 4 In the past 12 months, how often did you get into a serious physical ght? (H1DS5; H2FV16) Has there been any time over the past year when you thought you should get medical care, but you did not? (H1GH26; H2GH28) Which of these best describes your worst injury during the past year? (H1GH54; H2GH47) 2 Have you ever been told by a doctor or a nurse that you had... HIV/AIDS (H1CO16D; H2CO19D) 3 During the past 12 months, how often did each of the following things happen?

You saw someone shoot or stab another person. (H1FV1; H2FV1) 4 Most of the time, your mother is warm and loving toward you. (H1PF1; H2PF1) 2 Have you ever received an out-of-school suspension from school? (H1ED7;

H2ED3) 2 Have you ever been expelled from school? (H1ED9; H2ED5) 3 In general, how is your health? Would you say... (H1GH1; H2GH1) 2 In the past month, you have stopped on the street to talk with someone who lives in your neighborhood? (H1NB2; H2NB2) 5 Age at interview, computed from date of birth, and date of interview (Constructed from IYEAR, IMONTH, IDAY, H1GI1Y, H1GI1M) 2 Indicator variable (Constructed from H1HR3A-T; H2HR4A-Q) 2 Indicator variable (Constructed from H1HR3A-T; H2HR4A-Q) 4 Frequency of smoking (Constructed from H1TO1/2/5; H2TO1/5) 4 Frequency and amount of drinking alcohol (Constructed from H1TO12/15/18;

H2TO15/19/22) 3 Amount of exercise (Constructed from H1DA4/5/6; H2DA4-6) 2 Rescaled CES-D, following ( 14 ) (Constructed from H1FS1-18; H2FS1-18)

Victim of violence Family bereavement 2 3

Strong academically 4

Drug user 2 Family poor 5 Parents unhappy to- 4 gether Parent drinks 4 Householder smokes 3

Has learning disability

Parents aid decisions

2 5

Indicator variable (Constructed from H1FV2-6; (H2FV2-5) Number of bereavements (Constructed from H1NM2/F2, H1FP24A1-5; H2NM4/F4, H2FP28A1-3) Quartiles (Constructed from H1ED11-4; H2ED7-10) Indicator variable (Constructed from H1TO30/34/37/41; H2TO44/50/54/58) Census Bureau measure of poverty (Constructed from H1HR2/3/7/8, PA55) (Parent asked.) Do you and your partner argue/talk of separating? (Constructed from PB19/20) (Parent asked.) Number/frequency of drinks (Constructed from PA61/2) (Parent asked.) Either parent or others in household smokes (Constructed from PA63/4) (Parent asked.) Does (he/ she) have a speci c learning disability, such as di culties with attention, dyslexia, or some other reading, spelling, writing, or math disability? (PC38) (Parent asked.) How often would it be true for you to make each of the following statements about fchild's nameg? fChild's nameg and you make decisions about (his/ her) life together. (PC34B)

Background

Female Age Hisp/Latino White Black/Af Am Am Ind/Nat Am Asian/Pac Isl.

Other race Has learning dis.

Wave I Long-term Skips school Experiences prejudice In physical ghts Didn't pres. to doctor Severely injured Have HIV/AIDS Seen shooting Mother warm/loving Been suspended Been expelled Good health Alcohol Victim of violence Family bereavement Strong academically Drug user Family poor Parents unhappy togth. Parents aid decisions Wave I Short-term Househol. smokes Smoker Live with mother Live with father Parent drinks Talks neighbours Exercises Depressed

Wave II Long-term Seen shooting Alcohol Drug user

Mother warm/loving Have HIV/AIDS Family bereavement Experiences prejudice Been expelled Been suspended Victim of violence In physical ghts Strong academically Didn't pres. to doctor Skips school Severely injured Good health

Wave II Short-term

Smoker Live with mother Live with father Talks neighbours Exercises Depressed

References

[1] Goudie , R. J. B. and Mukherjee , S. M. ( 2011 ). An E cient Gibbs Sampler for Structural Inference in Bayesian Networks . CRiSM Working Paper 11- 21 (Dept. of Statistics, University of Warwick).

[2] Friedman , N. ( 2004 ) Science, 303 , 5659 , 799{ 805 .

[3] Mukherjee , S. and Speed , T. P. ( 2008 ) Proc Natl Acad Sci USA , 105 , 38 , 14313{ 14318 .

[4] Acid , S. , de Campos , L. M. , Fernandez-Luna , J. M. , Rodr guez , S., Rodr guez , J. M. and Salcedo , J. L. ( 2004 ) Artif Intell in Med, 30 , 3 , 215{ 232 .

[5] Costello , E. J. , Mustillo , S. , Erkanli , A. , Keeler , G. and Angold , A. ( 2003 ) Arch Gen Psych, 60 , 8 , 837{ 844 .

[6] Costello , E. J. , Erkanli , A. and Angold , A. ( 2006 ) J Child Psychol Psych , 47 , 12 , 1263{ 1271 .

[7] Thapar , A. , Collishaw , S. , Potter , R. and Thapar , A. K. ( 2010 ) Br Med J , 340 , c209 .

[8] Patel , V. , Flisher , A. J. , Hetrick , S. and McGorry , P. ( 2007 ) Lancet, 369 , 9569 , 1302{ 1313 .

[9] Barnett , P. A. and Gotlib , I. H. ( 1988 ) Psych Bull, 104 , 1 , 97{ 126 .

[10] Spirtes , P. , Glymour , C. and Scheines , R. ( 2000 ) Causation , Prediction, and Search (The MIT Press, Cambridge, MA).

[11] Korb , K. B. and Nicholson , A. E. ( 2011 ) Bayesian Arti cial Intelligence (CRC Press, Boca Raton , FL).

[12] Harris , K. M. , Halpern , C. T. , Whitsel , E. A. , Hussey , J. , Tabor , J. , Entzel , P. and Udry , J. R. ( 2009 ) The National Longitudinal Study of Adolescent Health: Research Design .

[13] Radlo , L. ( 1977 ) App Psych Meas, 1 , 3 , 385{ 401 .

[14] Goodman , E. ( 1999 ) Am J Pub Health , 89 , 10 , 1522{ 1528 .

[15] Roberts , R. E. , Lewinsohn , P. M. and Seeley , J. R. ( 1991 ) J Am Acad Child Adolesc Psych , 30 , 1 , 58{ 66 .

[16] Holt , S. , Buckley , H. and Whelan , S. ( 2008 ) Child Abuse & Neglect, 32 , 8 , 797{ 810 .

[17] Obot , I. S. and Anthony , J. C. ( 2004 ) J Child Adolesc Subst Abuse , 13 , 4 , 83{ 96 .

[18] Brown , R. A. , Lewinsohn , P. M. , Seeley , J. R. and Wagner , E. F. ( 1996 ) J Am Acad Child Adolesc Psych , 35 , 12 , 1602{ 1610 .

[19] Battles , H. B. and Wiener , L. S. ( 2002 ) J Adoles Health , 30 , 3 , 161{ 168 .

[20] Larun , L. , Nordheim , L. V. , Ekeland , E. , Hagen , K. B. and Heian , F. ( 2006 ) Cochrane Database Syst Rev, 3 , CD004691 .

[21] Heckerman , D. , Geiger , D. and Chickering , D. M. ( 1995 ) Mach Learn, 20 , 197{ 243 .

[22] Madigan , D. and York, J. C. ( 1995 ) Int Stat Rev , 63 , 2 , 215{ 232 .

[23] Kalisch , M. and Buhlmann, P. ( 2007 ) J Mach Learn Res , 8 , 613{ 636 .

[24] Cox , D. and Wermuth , N. ( 1996 ) Multivariate Dependencies Models, Analysis and Interpretation (Chapman & Hall, London).

[25] Robert , C. P. and Casella , G. ( 2004 ) Monte Carlo Statistical Methods (Springer, New York).

[26] Yu , B. and Mykland , P. ( 1998 ) Statistics and Computing, 8 , 3 , 275{ 286 .

[27] Gelman , A. and Rubin , D. B. ( 1992 ) Statistical Science, 7 , 4 , 457{ 472 .

[28] Pearl , J. ( 2000 ) Causality: Models, Reasoning , and Inference (Cambridge University Press, New York).

[29] Rickwood , D. J. , Deane , F. P. and Wilson, C. J. ( 2007 ) Med J Aus, 187 , 7 Suppl, S35{ S39 .

[30] National Institute for Health and Clinical Excellence ( 2005 ) Depression in Children and Young People (NICE , London).

[31] Merry , S. N. ( 2007 ) Curr Opin Psych, 20 , 4 , 325{ 329 .

[32] Grzegorczyk , M. and Husmeier , D. ( 2008 ) Mach Learn, 71 , 2 - 3 , 265 { 305 .

[33] Ellis , B. and Wong , W. H. ( 2008 ) J Am Stat Assoc , 103 , 482 , 778{ 789 .

[34] Friedman , N. and Koller , D. ( 2003 ) Mach Learn, 50 , 1 - 2 , 95 { 125 .

[35] Friedman , N. and Goldszmidt , M. ( 1996 ) In Proc. Twelfth Conference on Uncertainty in Arti cial Intelligence (UAI-96) (Morgan Kaufmann Publishers Inc.), 252 { 260 .

[36] Friedman , N. ( 1998 ) In Proc. Fourteenth Conference on Uncertainty in Arti cial Intelligence (UAI-98) (Morgan Kaufmann Publishers Inc.), 129 { 138 .