Michael l dennis ph d chestnut health systems bloomington il
This presentation is the property of its rightful owner.
Sponsored Links
1 / 36

Michael L. Dennis, Ph.D. Chestnut Health Systems, Bloomington, IL PowerPoint PPT Presentation


  • 86 Views
  • Uploaded on
  • Presentation posted in: General

Using a cluster analysis based case-mix solution to facilitate the evaluation and development of adolescent substance abuse treatment programs. Michael L. Dennis, Ph.D. Chestnut Health Systems, Bloomington, IL. Objectives.

Download Presentation

Michael L. Dennis, Ph.D. Chestnut Health Systems, Bloomington, IL

An Image/Link below is provided (as is) to download presentation

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.


- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -

Presentation Transcript


Michael l dennis ph d chestnut health systems bloomington il

Using a cluster analysis based case-mix solution to facilitate the evaluation and development of adolescent substance abuse treatment programs.

Michael L. Dennis, Ph.D.

Chestnut Health Systems, Bloomington, IL


Objectives

Objectives

  • Identification of Clients with similar presenting pathology based on a cluster analysis of the GAIN’s core psychiatric and behavior scales.

  • Demonstration of how the “case-mix” of these subgroups impacts program averages.

  • Illustration of how psychiatric case mix groups can be used to aid program evaluation and planning within or across program evaluation.


Global appraisal of individual needs gain

Global Appraisal of Individual Needs (GAIN)

  • A standardized bio-psycho-social that integrates clinical and research assessment for diagnosis, placement, treatment planning, process measures, outcome monitoring, and economic evaluation.

  • Core sections include cognitive assessment, background/access, substance use, physical health, risk behaviors, mental health, environment, legal, vocational, staff ratings

  • Over 100 scales/indices, with alpha over .9 on main scales and over .7 on subscales

  • Test retest data suggest reliability of items/scales over .7

  • Self reported use consistent with urine, salvia, and collateral reports (Kappa of .81 or more)

  • Predicts blind diagnosis of co-occurring psychiatric disorders including ADHD (kappa = 1.00), Mood Disorders (kappa = 0.85), Conduct Disorder or Oppositional Defiant Disorder (kappa = 0.82), Adjustment Disorder (kappa = 0.69), and No other diagnosis (kappa = 0.91)


Factor structure and cluster analysis based on 2968 clients from 61 treatment units

Factor Structure and Cluster Analysis based on 2968 Clients from 61 Treatment Units

Farmington, CT

Chicago, IL

New York, NY

Peoria, IL

Philadelphia, PA

Oakland, CA

Bloomington, IL

Baltimore, MD

Cantonsville, MD

Maryville, IL

Los Angeles, CA

Shiprock, NM

Phoenix/Tempe, AZ

Tucson, AZ

St. Petersburg, FL

Miami, FL

Adolescent Outpatient/IOP

Adolescent Inpatient/Therapeutic Community

Adult Outpatient/IOP/OP Methadone Treatment

Adult Inpatient/Therapeutic Community


Hypothesized structure of the gain s psychopathology measures

Internal Mental Distress

Crime and Violence

Behavioral Complexity

Hypothesized Structure of the GAIN’s Psychopathology Measures

* Main scales have alpha over .85, subscales over .7


Michael l dennis ph d chestnut health systems bloomington il

rs

.64

.55

SIIY

.80

.51

.71

SA Problems

SAIY

.78

.88

SDIY

.54

ri

.74

SSI

.67

.73

.60

DSI

.82

.27

.52

Internal

HSTI

.77

.88

.78

ASI

.68

.47

.23

TSI

General

re

Severity

.60

HII

.71

.62

.51

.83

.91

IAI

External

.68

.46

CDI

.50

.39

rv

.54

GCTI

.62

.63

.25

PCI

.79

.62

Crime/Violence

.79

ICI

.74

.55

DCI

Confirmatory Factor Analysis (CFA)

Comparative Fit Index: .974

Root Mean Square Error of Approximation: 0.079

Invariant vs Variant Across

Age and Level of Care

Comparative Fit Index: .97 vs .98

Parsimony Ratio: .80 vs .70

CFI x PR: .78 vs .68

Root Mean Square Error

of Approximation: .04 vs .04


Creating cluster code types

Creating Cluster Code Types

  • The overall severity and four core dimensions were used to create 7 code types with Ward’s minimum distance cluster analysis.

  • Total and four dimensional scores triaged into low, medium and high based on +/- .5 standard deviations from the mean

  • Code types labeled most common group as:

    • High, medium or low overall severity on total score

    • Labeled in order from highest to lowest severity dimension

    • Lines // used to separate those in high/ medium/ low severity on each each of four dimensions

    • Sample size

  • Discriminate Function Analysis for Classifying New Cases (Kappa =.82)


7 cluster code types

Code Type (A,B,C..)

High to Low Severity order

Hi / Med / Low range divided by //

7 Cluster Code Types

High G.,

CV, BC ID, SP//

(N=214)

Low A.

8%

//CV, ID, BC, SP

(N=545)

High F.

ID, BC, SP, CV//

19%

(N=336)

12%

Low B.

SP/ID/ CV, BC

(N=370)

High E.

13%

CV, BC, SP/ ID/

(N=429)

15%

Med. C.

Med. D.

/BC, CV/ID, SP

(N=467)

SP/ BC, ID/ CV

16%

(N=471)

17%


General severity by code type

General Severity by Code Type


Substance problem sp by code type

Substance Problem (SP) by Code Type


Internal distress id by code type

Internal Distress (ID) by Code Type


Behavior complexity bc by code type

Behavior Complexity (BC) by Code Type


Behavior complexity cv by code type

Behavior Complexity (CV) by Code Type


Case mix by age and level of care

Case Mix by Age and Level of Care


Michael l dennis ph d chestnut health systems bloomington il

ATM

Adolescent Treatment Model

Program Sites

New York, NY

Baltimore, MD

Oakland, CA

Bloomington, IL

Cantonsville, MD

Shiprock, NM

Tempe, AZ

Los Angeles, CA

Tucson, AZ

1999

1998

Miami, FL

Sponsored By:

Center for Substance Abuse Treatment (CSAT),

Substance Abuse and Mental Health Services Administration (SAMHSA),

U.S. Department of Health and Human Services (DHHS)


Atm involved the full range of code types

ATM involved the full range of Code Types


Evaluating cluster code types

Evaluating Cluster Code Types

  • Severity should go up with level of care (LOC) – one of the most commonly used case mix variables.

  • The cluster code type should do better than LOC in terms of:

    • Maximizing individual differences between cluster subgroups

    • Minimizing individual indifference by LOC within cluster subgroups

  • The cluster code types should help to predict differential response patterns to treatment


Case mix severity goes up with level of care

PCM Index Score (Weighted Average)

Case Mix Severity Goes up With Level of Care

G-High

CV,BC,ID,SP//

100%

90%

F-High

ID,BC,SP/CV/--

80%

E-High

70%

CV,BC,SP/ID/

60%

D-Mod

SP/BC,ID/CV

50%

C-Mod

40%

BC/CV,ID/SP

30%

B-Low

/SP,ID/CV,BC

20%

A-Low

10%

//CV,ID,BC,SP

0%

PCM Index Score

Early Intervention

OP/IOP

LTR

STR


Level of care is related to average severity

Individual Differences explained by LOC quantified with Cohen’s effect size f

Level of Care Is Related to “Average” Severity

4.0

OP (n=553)

3.0

LTR (n=373)

2.0

STR (n=573)

1.0

Z-score

0.0

-1.0

-2.0

-3.0

-4.0

Total Score

(f=0.4)

ID. Internal

(f=0.29)

Distress

Complexity

(f=0.28)

SP. Substance

Problem

(f=0.26)

(f=0.14)

CV.

BC Behavior

Crime/Violence


However cluster subgroups are more distinct from each other

+561%

+310%

+338%

+85%

+750%

Cohen’s effect size f increased by 85% to 750%

However Cluster Subgroups are More Distinct From Each Other

A-Low

//CV,ID,BC,SP

4.0

(n=208)

3.0

B-Low

/SP,ID/CV,BC

2.0

(n=101)

C-Mod

1.0

BC/CV,ID/SP

Z-score

(n=286)

0.0

D-Mod

-1.0

SP/BC,ID/CV

(n=252)

-2.0

E-High

-3.0

CV,BC,SP/ID/

(n=281)

-4.0

F-High

ID,BC,SP/CV/--

Total Score

(f=1.75)

(n=180)

ID. Internal

(f=1.19)

Distress

Complexity

(f=1.85)

SP. Substance

Problem

(f=0.48)

BC Behavior

(f=1.19)

Violence

CV.Crime

G-High

CV,BC,ID,SP//

(n=191)


A low cv id bc sp

Once we account for subgroup, LOC differences are gone and Cohen’s effect size f goes down

A-Low //CV,ID,BC,SP

4.0

OP (n=114)

3.0

LTR (n=59)

2.0

STR (n=35)

1.0

Z-score

0.0

-1.0

-2.0

-3.0

-4.0

Total Score

(f=0.05)

ID. Internal

(f=0.11)

Distress

Complexity

(f=0.16)

SP. Substance

(f=0.04)

Problem

BC Behavior

(f=0.04)

CV.

Crime/Violence


B low sp id cv bc

B-Low /SP,ID/CV,BC

4.0

OP (n=38)

3.0

LTR (n=23)

2.0

STR (n=40)

1.0

Z-score

0.0

-1.0

-2.0

-3.0

-4.0

Total Score

(f=0.08)

ID. Internal

(f=0.06)

Distress

Complexity

(f=0.02)

SP. Substance

(f=0.12)

Problem

BC Behavior

(f=0.09)

CV.

Crime/Violence


C mod bc cv id sp

C-Mod BC/CV,ID/SP

4.0

OP (n=138)

3.0

LTR (n=82)

2.0

STR (n=66)

1.0

Z-score

0.0

-1.0

-2.0

-3.0

-4.0

Total Score

(f=0.18)

ID. Internal

(f=0.22)

Distress

Complexity

(f=0.13)

SP. Substance

(f=0.13)

Problem

BC Behavior

(f=0.09)

CV.

Crime/Violence


D mod sp bc id cv

D-Mod SP/BC,ID/CV

4.0

OP (n=78)

3.0

LTR (n=57)

2.0

STR (n=117)

1.0

Z-score

0.0

-1.0

-2.0

-3.0

-4.0

Total Score

(f=0.17)

ID. Internal

Distress

(f=0.14)

Complexity

(f=0.1)

SP. Substance

Problem

(f=0.18)

BC Behavior

(f=0.1)

CV.

Crime/Violence


E high cv bc sp id

E-High CV,BC,SP/ID/

4.0

OP (n=103)

3.0

LTR (n=50)

2.0

STR (n=128)

1.0

Z-score

0.0

-1.0

-2.0

-3.0

-4.0

Total Score

(f=0.13)

ID. Internal

(f=0.14)

Distress

Complexity

(f=0.08)

SP. Substance

(f=0.22)

Problem

BC Behavior

(f=0.08)

CV.

Crime/Violence


F high id bc sp cv

F-High ID,BC,SP/CV/

4.0

OP (n=43)

3.0

LTR (n=44)

2.0

STR (n=93)

1.0

Z-score

0.0

-1.0

-2.0

-3.0

-4.0

Total Score

(f=0.06)

ID. Internal

(f=0.05)

Distress

Complexity

(f=0.06)

SP. Substance

Problem

(f=0.18)

(f=0.08)

CV.

BC Behavior

Crime/Violence


G high cv bc id sp

G-High CV,BC,ID,SP//

4.0

OP (n=39)

3.0

LTR (n=58)

2.0

STR (n=94)

1.0

Z-score

0.0

-1.0

-2.0

-3.0

-4.0

Total Score

(f=0.15)

Complexity

(f=0.13)

SP. Substance

Problem

(f=0.28)

(f=0.06)

Distress (f=0.1)

CV.

ID. Internal

BC Behavior

Crime/Violence


Cluster subgroups significantly reduces the individual differences associated with level of care

Cluster Subgroups Significantly Reduces the Individual Differences Associated with Level of Care

A-Low

//CV,ID,BC,SP

100.0%

(n=208)

80.0%

B-Low /SP,ID/CV,BC

60.0%

(n=101)

40.0%

C-Mod BC/CV,ID/SP

20.0%

(n=286)

Change in LOC Effect Size f

0.0%

D-Mod SP/BC,ID/CV

-20.0%

(n=252)

-40.0%

-60.0%

E-High CV,BC,SP/ID/

(n=281)

-80.0%

-100.0%

F-High

ID,BC,SP/CV/--

Total Score

(n=180)

ID. Internal

Distress

Complexity

SP. Substance

Problem

BC Behavior

CV.

Crime/Violence

G-High

CV,BC,ID,SP//

(n=191)


Outpatient by cluster types

Differentiates initial severity, and differences in response

Outpatient by Cluster Types


Long term residential by cluster types

Can identify subgroups (E, B) that are a higher risk of relapse or having other problems

Long Term Residential by Cluster Types


Short term residential by cluster types

Short Term Residential by Cluster Types

Different levels of care/programs may do well (A,F,G) or have problems (B,C,D, E) with different subgroups


C mod bc cv id sp by loc

However this is still quasi-experimental and the adjustments are often imperfect

For a Given Subtype, it can identify when a particular level of care (or program) appears to do better.

C-Mod BC/CV,ID/SP by LOC


Conclusions

Conclusions

  • Clustering people based on presenting problems appears to work better than level of care for describing initial case mix but is also correlated with it.

  • Clinical subtype clusters can help to identify subgroups for which a program works well and/or where continuing care or other services may be needed.

  • Within a clinical subtype, comparisons of level of care (programs, services etc) could be used to guide placement decisions and/or identify promising areas for experimentation.


Contact information

Contact Information

Michael L. Dennis, Ph.D.

Lighthouse Institute, Chestnut Health Systems

720 West Chestnut, Bloomington, IL 61701

Phone: (309) 827-6026, Fax: (309) 829-4661

E-Mail: [email protected]

A copy of these slides will be posted at: www.chestnut.org/li/posters


Errata

Errata

The following additional slide was presented by the discussant, Dr. Mark Fishman, to show how case mix varied at the program level even within level of care.


Case mix by level of care atm program

Early Intervention at the low end

Also demonstrates that Level of Care is only a rough proxy of case mix

STR/LTR dominates high end

Case Mix by Level of Care/ATM program


  • Login