charleston presentation v4

Upload: phongvhp2728

Post on 06-Apr-2018

224 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/2/2019 Charleston Presentation v4

    1/35

    Click to edit Master subtitle style

    Propensity Score Matching to Evaluate theImpact of Substance Abuse Treatment

    Services for Child Welfare Clients

    Richard Barth, Claire Gibbons, & Shenyang GuoJordan Institute for FamiliesSchools of Social Work (1) and Public Health (2)

    University of North Carolina

    Presented to the Annual Grantees MeetingSubstance Abuse Policy Research ProgramCharleston, South CarolinaDecember 2, 2004

    PSM analyses were funded by the Robert Wood Johnson Foundations Substance Abuse PolicyResearch Program NSCAW data used to illustrate PSM were collected under funding by theAdministration on Children, Youth, and Families of the U.S. Department of Health and HumanServices. Findings do not represent the official position or policies of the U.S. DHHS. Results arenot quotable in print. Contact [email protected].

    mailto:[email protected]:[email protected]
  • 8/2/2019 Charleston Presentation v4

    2/35

    Todays Presentation

    Introduction to the research questions

    Brief explanation of the data sources and

    reasons to use PSM to answer the researchquestions.

    Description of post PSM analyses

    Discussion of the findings

  • 8/2/2019 Charleston Presentation v4

    3/35

    Background for the Study

    Next to poverty, substance abuse is considered mostcentral reason for CWS involvement Linda GordonHeroes of Their Own Lives (1800s)

    ACF Building Common Ground (1999)

    SAMSHA and ACF National Resource Center on ChildWelfare and Substance Abuse (2003)

    Largest population of children who receive child

    abuse reports remain at home, but almost all CWSand SAT research is about placement into out-of-home care

  • 8/2/2019 Charleston Presentation v4

    4/35

    Research Questions for this Sub-study

    Does participation in SAT result in an increasein child safety, as reflected in a reduction inthe rate of subsequent child abuse reports

    Are there other mechanisms operating that help toexplain the relationship between SAT andsubsequent child abuse reports?

  • 8/2/2019 Charleston Presentation v4

    5/35

    First National Random Sample Study Of CWS

    Extended Research Team includes:

    Research Triangle Institute

    University of North Carolina

    Caliber Associates

    San Diego Childrens Hospital, CASRC

    CSRD, Pitt Medical Center

    Duke Medical Center

    National Data Archive on Child Abuse andNeglect, Cornell

    92 Local Child Welfare Agencies Admin. For Children and Families

    Children and Families

  • 8/2/2019 Charleston Presentation v4

    6/35

    Click to edit Master subtitle style

    NSCAW Cohort at Baseline

    Total6,231

    Enter throughinvestigation

    5,504

    No services1,725

    Ongoing services3,779

    In home2,312

    Out-of-home1467

    Other gateways600

    Long-termfoster care

    727

  • 8/2/2019 Charleston Presentation v4

    7/35

    Click to edit Master subtitle style

    NSCAW Child Sampling is Complex

    Stratified: 8 Certainty States and remainder (28) states

    92 PSUs (basically, county agencies)

    Over Sampled on the basis of:

    Children/Families Receiving Services Infants

    Sexually Abused Children

    Not Sampled on the basis of:

    Substantiated Reports (cases are included whether abuse allegations aresubstantiated or not)

    ALL

    ANA

    LYSES

    AREWEIGHTED

  • 8/2/2019 Charleston Presentation v4

    8/35

    Data Collection TimelineTarget population: Childreninvolved in investigationsclosed between October 1, 1999and December 31, 2000

    W

    ave1:BaselineNov,1999

    Apr,2001

    Wav

    e2:12Month

    Follow-upO

    ct,2

    Wave3:

    18MonthFoll

    ow-

    1999 2000 2001 2002 2003 2004

    Wave4:36Mont

  • 8/2/2019 Charleston Presentation v4

    9/35

    Achieved Samples

    Wave 1 (Baseline) Current caregiver : (CPS) 70% (OYFC) 73%

    Child welfare worker: (CPS) 86% (OYFC) 80%

    Children: (CPS) 66% (OYFC) 70%

    Wave 2 (12-Months) Current caregiver: (CPS) 83% (OYFC) 89%

    Child welfare worker: (CPS) 84% (OYFC) 85%

    Wave 3 (18-Months) Current caregiver : (CPS) 85% (OYFC) 87% Child welfare worker: (CPS) 94% (OYFC) 95%

    Children: (CPS) 82% (OYFC) 84%

    Differential

    attrition wasvery modestand adjustedwith W3weights

  • 8/2/2019 Charleston Presentation v4

    10/35

    Counterfactuals and Propensity ScoreMatching

    Theory of Counterfactuals The fact is that some people who could use substance

    abuse treatment (SAT) receive it and some do not. The key assumption of the counterfactual framework is

    that individuals selected into treatment andnontreatment groups have potential outcomes in both

    states: the one in which they are observed and the onein which they are not observed (Winship & Morgan,1999).

  • 8/2/2019 Charleston Presentation v4

    11/35

    Counterfactuals and Propensity ScoreMatching II

    Counterfactuals cannot be observedwe can onlycreate an estimate of them

    These estimates are generally viewed as biased unlessgroups are randomized

    Yet, some interventions, like SAT, are considered sovital that randomization is highly improbable

    PSM is one correction strategy that endeavors toreduce selection biases and permit comparisonsbetween outcomes of observed and unobservedtreatment when randomization has not been achieved

  • 8/2/2019 Charleston Presentation v4

    12/35

    Types of Selection Bias Likely at Playin this Study

    Self selection (clients influencing the service theyreceive),

    Initial attendance

    Quitting treatment

    Bureaucratic selection (staff in service programs makedecisions based on agency guidelines or more implicitdecision rules [e.g., the availability of treatment]),

    Geographic selection (some services are only availablein some areas).

  • 8/2/2019 Charleston Presentation v4

    13/35

    Two Types of Selection Bias EspeciallyGermane to SAT and CWS Research

    Creaming: Best cases are selected intoprograms because most disorganized clientelecannot manage program participation

    Souring (triaging): Most difficult cases aretriaged into services--sometimes required by

    law to participateso that the treatment groupmay be likely to look worse at the outset andafter treatment

  • 8/2/2019 Charleston Presentation v4

    14/35

    PSM in a Nutshell

    Employs a predicted probability of groupmembershipe.g., treatment vs. control group--based on observed predictors, usually obtained

    from logistic regression to create acounterfactual group

    Propensity scores may be used for matching and

    as covariates in secondary analysis Time to first re-reportevent history analysis

  • 8/2/2019 Charleston Presentation v4

    15/35

    Creating the Propensity Score:

    Dependent Variable AOD service receipt

    Stayed overnight in AOD treatment program Clinic or doctor

    CWW reports CG received treatment

    CG reported currently receiving treatment

  • 8/2/2019 Charleston Presentation v4

    16/35

    Creating the Propensity Score:

    Independent Variables Marital status

    Education

    Employment Poverty

    Case status

    Child race/ethnicity Child age

    Caregiver age

    Trouble paying for basic necessities

    CG mental health CG arrest/jail time

    Prior AOD treatment

    Maltreatment type

  • 8/2/2019 Charleston Presentation v4

    17/35

    Independent Variables (contd)

    Need for treatment

    Risk assessment

    CIDI-SF Screen, Dependence

    CG report of need

  • 8/2/2019 Charleston Presentation v4

    18/35

    Data & Study Sample

    NSCAW data from two waves: baseline (1999-2000) andthe 18-month follow-up.

    The sample for this study was limited to families where:

    children lived at home (n=4034)

    primary caregivers were female (n=3670)

    the primary caregiver had at least one indicator of a substanceabuse problem (n=1472)

    there was a non-missing value of AOD service receipt and anon-missing value of time to first re-report (n=1074)

    Of these children, 276 (26%) formed the treatmentgroup, and 798 (74%) comparison group.

  • 8/2/2019 Charleston Presentation v4

    19/35

    Selecting the Best Model

    12 matching schemes (using Mahalanobis distance and nearestneighbor matching) were used with different tolerances andwith and without including the propensity score as a covariate. Differences in KS survivor functions were all in the same direction;

    that is, the treated group has a faster rate of re-report than the

    nontreated group.

    Did not use Heckmans difference-in-difference because theresults are not amenable to event history analysis

    Nearest neighbor matching within a caliper of of thestandard deviation of the propensity score Retained the most cases

    Had the fewest remaining statistical differences between items

    Had broad common supportarea.

  • 8/2/2019 Charleston Presentation v4

    20/35

    0

    2

    4

    6

    0

    2

    4

    6

    0 .5 1

    Nontreated

    Treated

    Density

    Predicted ProbabilityGraphs by aodserv

    Histogram of Predicted Probabilities of Using SATBy Treatment Group Before Matching

  • 8/2/2019 Charleston Presentation v4

    21/35

    Sample Comparison: Demographics

    Before matching(n=1074)

    After matching(n=276)

    Child race p=.02 p=.50

    Child age p

  • 8/2/2019 Charleston Presentation v4

    22/35

    Sample Comparison: Caregiver Risks and Need

    Before matching(n=1074)

    After matching(n=276)

    Trouble paying for basicnecessities

    p=.05 p=.90

    Mental health problems p

  • 8/2/2019 Charleston Presentation v4

    23/35

    Research Question

    Whether or not children of caregivers receivingsubstance abuse services are living in a safeenvironment?

    Operationalized as: Does substance abuse treatment for

    caregivers affect the risk of child

    maltreatment re-reports by caregiver and family?

    by caregiver alone?

  • 8/2/2019 Charleston Presentation v4

    24/35

    Kaplan Meier and Cox Regression

    Dependent variable(s) Time to first re-report (18 month follow-up) by:

    Caregiver or relative

    Caregiver alone

    Independent variables for Cox Regression AOD service receipt

    Child age

    CG age

    Prior child welfare services

    Open child welfare case Family cumulative risk

    Urban/rural

  • 8/2/2019 Charleston Presentation v4

    25/35

    Findings

    EHA of the best PSM resample shows that thegroup difference on survivor functions arestatistically significant with recipients of SAThaving a greater hazard of re-reports than thenontreatment comparison group (roughly 20%vs. 10% during the 18-months).

    Time to re-report by caregivers or relatives (p < .02)

    Time to re-report by caregivers (p < .05)

  • 8/2/2019 Charleston Presentation v4

    26/35

    Percent Children RemainingWith No Re-report on Caregivers

    Substance Abuse Sample (n=1,074) Resampled with PSM (n=276)

    The proportion of children re-reported within 18-months was higher for SAT

    after the PSM correction (19 % vs. 10%) but not before (18% vs. 18%)

    0 . 0 0

    0 . 2 5

    0 . 5 0

    0 . 7 5

    1 . 0 0

    d u r 2

    0 . 0 2 . 5 5 . 0 7 . 5 1 0 . 0 1 2 . 5 1 5 . 0 1 7 . 5 2 0 . 0

    S T R A T A : A O D S E R V _ = 0 C e n s o r e d A O D S E R V _ = 0 A O D S E R V _ = 1 C e n s o r e d A O D S E R V _ = 1

    0 . 0 0

    0 . 2 5

    0 . 5 0

    0 . 7 5

    1 . 0 0

    d u r m

    0 . 0 2 . 5 5 . 0 7 . 5 1 0 . 0 1 2 . 5 1 5 . 0 1 7 . 5 2 0 . 0

    S T R A T A : a o d s e r v _ = 0 C e n s o r e d a o d s e r v _ = 0 a o d s e r v _ = 1 C e n s o r e d a o d s e r v _ = 1

  • 8/2/2019 Charleston Presentation v4

    27/35

    Estimated Hazard Ratios:Cox Regression based on the Resample

    ________________________________________________________

    Variable Re-report Re-reportRelative or CG Primary CG

    SA Service Receipt (No)

    Yes 2.02 * 2.03 *

    Child Age (11+)

    0-2 .60 .58

    3-5 .76 .72

    6-10 1.44 1.49

    Caregiver Age (35 or Older)

    < 35 1.51 1.66

    Prior Child Welfare Service(No)

    Yes 2.44 ** 2.51 **

    Open Child Welfare Case (No)

    Yes .77 .74Family Cumulative Risk (High)

    Low .68 .75

    Moderate .85 .89

    Residence (Nonurban)

    Urban 1.9 1.93

    ________________________________________________________

    ** p

  • 8/2/2019 Charleston Presentation v4

    28/35

    Limitations of PSM^

    Large samples are required

    The sample size did drop, but is still substantial (unweighted n=276)

    Group overlap must be substantial before matching

    There were significant differences between the groups prior to PSM

    About half of the SAT cases were excluded from the final models

    Hidden bias may remain because matching only controls for observedvariables (to the extent that they areperfectly measured)

    We based our variable selection on a thorough review of predictorsof service use

    Although we have 14 indicators of the need for substance abuse

    services, this information is still notperfectly measured^Shadish, Cook, & Campbell, 2002

  • 8/2/2019 Charleston Presentation v4

    29/35

    Conclusions

    Caregivers with allegations of child abuse andneglect against them, and with substance abuse

    problems, often remain at home following the CWSinvestigation.

    Children whose female caregivers receive substanceabuse treatment are more likely to receive asubsequent child maltreatment report in thefollowing 18-months: Whether the source of the report includes all relatives or

    is restricted to the caregiver, and Whether or not the selection bias into SAT is controlled

    for using PSM methods.

  • 8/2/2019 Charleston Presentation v4

    30/35

    Possible Mechanisms

    The higher rate of re-reports may be attributable tosome of the following mechanisms, each of which wastested and found not to differbetween the SAT andnontreatment group:

    The type of maltreatment at re-report

    The average CIDI-SF score for alcohol and drug use, acrossthe two waves

    Childrens behavior on the CBCL (ages 2 and up)

    Childrens reports (ages 11 and up) with regard to harsh orsevere parenting

  • 8/2/2019 Charleston Presentation v4

    31/35

    Alternate Explanations for Findings

    PSM did not match for important unobserved covariates Services may not have been used in full

    Substance abuse services may interfere with parentaladequacy

    Focus is on parents recovery not childs welfare

    Time and effort for SAT can be burdensome to parent

    Services may result in greater surveillance which resultsin more observed behaviors that might place children atrisk, thus more reports

    But there were also more placements into foster care (p < .05).

  • 8/2/2019 Charleston Presentation v4

    32/35

    Implications

    Children whose caregivers receive SAT may be saferif the early re-reports were for failure to participate insubstance abuse treatment and prevented abuse orneglect

    Children whose caregivers receive SAT may be lesssafe if the re-reports followed abuse or neglect Caregivers who are involved with SAT may increase

    their likelihood of remaining involved with CWS, atleast for 18-months

    Much more needs to be known about the impact ofSAT on child welfare caregivers and children This study suggests that it is not unambiguously beneficial

    for reducing maltreatment and preserving families

  • 8/2/2019 Charleston Presentation v4

    33/35

    Partial References

    Gregoire, K. A., & Schultz, D. J. (2001). Substance-abusing child welfare parents: Treatment

    and child placement outcomes. Child Welfare, 80(4), 433-452.

    Leathers, S. J. (2002). Parental visiting and family reunification: Could inclusive practicemake a difference? Child Welfare, 81(4), 595-616.

    Rosenbaum, P. R., & Rubin, D. B. (1985a). Constructing a control group using multivariatematched sampling methods that incorporate the propensity score.American Statistician, 39,

    33-38.Semidei, J., Radel, L. F., & Nolan, C. (2001). Substance abuse and child welfare: Clearlinkages and promising responses. Child Welfare, 80(2), 109-128.

    Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002).Experimental and quasi-experimentaldesigns for generalized causal inference. Boston: Houghton Mifflin.

    Sosin, M. R. (2002). Outcomes and sample selection: The case of a homelessness and

    substance abuse intervention.British Journal of Mathematical and Statistical Psychology, 55,63-91.

    U.S. Department of Health and Human Services. (1999, April).Blending perspectives andbuilding common ground. A report to Congress on substance abuse and child protection.Retrieved May 23, 2000, http://aspe.hhs.gov/hsp/subabuse99/subabuse.htm

    http://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htm
  • 8/2/2019 Charleston Presentation v4

    34/35

    PSM and Project Resources

    This Presentation

    Substance Abuse and Child Welfare Services: Research Update andNeeds (2003)

    Introduction to PSM: A New Device for Program Evaluation (2004)

    Substance Abuse among Caregivers Involved with Child WelfareServices: Prevalence and Identification by Child Welfare Workers(2004) http://sswnt5.sowo.unc.edu/VRC/Lectures/index.htm

    Substance Abuse Needs and Services for Families Involved in the Child(brief project description and link to other project materials) http://sswnt5.sowo.unc.edu/JordanIF/jif_map/projects_family_nonframe.cfm

    http://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://aspe.hhs.gov/hsp/subabuse99/subabuse.htmhttp://sswnt5.sowo.unc.edu/VRC/Lectures/index.htmhttp://sswnt5.sowo.unc.edu/JordanIF/jif_map/projects_family_nonframe.cfmhttp://sswnt5.sowo.unc.edu/JordanIF/jif_map/projects_family_nonframe.cfmhttp://sswnt5.sowo.unc.edu/JordanIF/jif_map/projects_family_nonframe.cfmhttp://sswnt5.sowo.unc.edu/JordanIF/jif_map/projects_family_nonframe.cfmhttp://sswnt5.sowo.unc.edu/VRC/Lectures/index.htm
  • 8/2/2019 Charleston Presentation v4

    35/35

    Thank you very much

    Questions?

    ?