Title: Rohlings Interpretive Method: Use of MetaAnalytic Procedures for Single Case Data Analysis
1Rohlings Interpretive Method Use of
Meta-Analytic Procedures for Single Case Data
Analysis
- Martin L. Rohling, Ph.D. University of South
Alabama L. Stephen Miller, Ph.D. University of
Georgia
2Introduction to the RIM Process
- Flexible battery (multiple measure) use
- Is the most frequently cited model of assessment
among neuropsychologists. - Only 15 of neuropsychologists use a fixed
battery (Sweet, et al, 2000, TCN) - Regarding the suitability, practicality, and
usefulness of any fixed battery - We know of no batteries that fully satisfy these
criteria. - (Lezak, Howieson Loring 2004,
Neuropsychological Assessment, 4th Edition, p
648.)
3Advantages of Flexible Battery Approaches
- Dynamic.
- Cover 1 or many domains.
- Flexible, can be adapted for each patient.
- Can oversample domains.
- Well suited as a hypothesis-driven approach.
4Potential Problems - Flexible Battery Approaches
- Inflated error rates.
- Multicollinearity.
- Weighting decision problems.
- Unknown veracity/reliability of sets of tasks.
- Human judgment errors.
5Human Judgment Errors(Wedding Faust, 1989, ACN)
- Hindsight bias.
- Confirmatory bias.
- Overreliance on salient data.
- Under-utilization of base rates.
- Failure to take into account co-variation.
6RIM Potential
- Judgment errors can threaten reliability and
validity of multiple measure test batteries. - RIM was designed to reduce these effects.
- Based on meta-analytic techniques.
- Uses a linear combination of scores placed on a
common metric.
7RIM Potential
- A strategy that produces summary results
analogous to those generated in a fixed-battery
approach (e.g., HII, GNDS, AIR). - Takes advantage of psychometric properties of
same metric data e.g. T-Scores.
8Todays Intent
- Present a set of procedures that allows for a
quantitatively-based comparison of an overall
battery of measures. - Non-specific to battery measures themselves.
- Can be used for any individual patient.
- Demonstrate importance and practicality of use
of established statistical indices. - (e.g., alpha, beta, effect size).
9Todays Intent (contd)
- Present a data format for any set of measures to
be inspected at - Global level (OTBM)
- Domain level (DTBM)
- Test measure level (ITBM)
- Present a series of calculations to assist in the
generation of these indices. - Present Steps in conjunction with clinical
judgment from an informed position.
10RIM Categories
- Symptom Validity (SV) Tests.
- Emotional / Personality (EP) Measures.
- Estimated Premorbid General Ability (EPGA).
- Test Battery Means.
- Overall (OTBM), Domain (DTBM), Instrument
(ITBM). - Cognitive Domains
- VC, PO, EF, AML, VML, AW, PS
- Non-Cognitive Domains.
- PM, LA, SP
11Sample RIM Summary Table
12Sample RIM Graphic Display
13Brief of RIM Steps
- There are 24 steps to the RIM process
- 17 calculation steps
- Advice on design of the battery
- Calculation of summary statistics
- Generation of graphic displays
- 7 interpretative steps.
- Detail a systematic procedure for use of the
statistical summary table and graphic displays
to - Assess and verify summary data.
- Identify strengths/limitations of current data.
- Obtain a reliable diagnosis.
- Develop tx plans based on sound judgments.
- We briefly review each step in just a moment.
14Support for the RIM Process
- Rational support/reasoning Reduce clinical
judgment errors. - The RIM is a Process, not a program.
- A way of formulizing your thinking and
interpretation of your data. - This is operationalizing what you already do.
15Support for the RIM ProcessSpecific Advantages
- Psychometric properties at level with fixed,
co-normed batteries, without their limitations. - Flexibility of test selection.
- Flexibility of theoretical view of cognition
(domain structure)
16Support for the RIM ProcessSpecific Advantages
- Quantitatively support your conclusions and
interpretations - Statistical evaluation
- Measure of confidence in findings
- Measure of limitations of findings
- Ability to present data at different levels of
interpretation - Greater defensibility
17The RIM as a Procedure of Specific Steps
18RIM Steps 1-4 Summary Data
- Design administer battery.
- Use well standardized recently normed tests.
- Estimate premorbid general ability.
- Use Reading (WTAR), Regression (OPIE-III),
academic records (rank, SAT, ACT). - Convert test scores to a common metric.
- We recommend T scores, but z or SS OK too.
- Assign scores to domains.
- Factor analysis to support assignment (Tulsky et
al., 2003)
19RIM Steps 5-8 Summary Data
- Calculate domain M, sd, n.
- Calculate test battery means (TBM).
- Overall TBM All scores, large N high power.
- Domain TBM Avoids domain over weighting.
- (e.g., attention memory).
- Instrument TBM One score per norm sample.
- Calculate p for heterogeneity.
- Have you put apples oranges together?
- Determine categories of impairment.
- Recommend using of Heaton et al. (2003).
20RIM Steps 9-12 Summary Data
- Determine of test impaired.
- Analogous to Halstead Impairment Index
- scores below cutoff / total of scores
- Calculate ES for all summary stats.
- Use Cohens d (Me Mc) / SD pooled
- Calculate confidence interval for stats.
- 90 CI 1.645 x SEM
- Upper limit of performance for impair.
- Look for overlap between 90 CI of EPGA (lower)
Summary Stats (upper)
21RIM Steps 13-17 Summary Data
- Conduct one-sample t tests.
- Use EPGA as reference point
- Conduct a between-subjects ANOVA.
- Looking for strengths weaknesses
- Conduct power analyses.
- Only needed for those NS differences
- Sort scores for visual inspection.
- Graphically display summary statistics.
22RIM Steps 18-20 Interpretation
- Assess battery validity.
- Examine the Symptom Validity scores.
- Caution in accepting low power results.
- Look at heterogeneity of summary stats.
- Normative sample unrepresentative of patient.
- Scores assigned to wrong domain.
- Inconsistent performance on construct measures.
- Examine influence of psychopathology.
- Examine scores for heterogeneity.
- Check OTBM, DTBM, ITBM impairment.
23RIM Steps 21-24 Interpretation
- Examine strengths/weaknesses looking at
- Confidence intervals overlap.
- Results from one-sample t tests.
- Results of ANOVA.
- TI show differences otherwise not evident.
- Determine if pattern existed premorbidly.
- Examine non-cognitive domains.
- Psychomotor, Lang/Aphasia, Sensory Percept
- Explore Type II errors need more tests?
- Examine sorted T-scores
- Look for patterns missed by summary stats.
24RIM Sample Case 1 Obvious TBI
- Age 37
- Handed Left
- Race Euro-American
- Sex Female
- Ed 14 years
- Occup Nursing
- Marital Sep. 10 yrs
- Living Camper in parents backyard
- Reason for Referral
- TBI in head-on boat accident. Propeller hit pt in
right parietal-occipital lobe (LOC 7 days GCS
3). Eval. to determine capacity for medical
financial decisions, parenting skills,
occupational prognosis, disability status.
Significant emotional, behavioral, occupational,
and social problems pre-TBI.
25RIM Sample Case 1 Obvious TBI
26RIM Sample Case 1 Obvious TBI
27TBI Dose Response CurvesDikmen ESs Meyers T
Scores
28Return to Work Study OTBMs for 4 Groups of TBI
Survivors
29RIM Sample Case 2 Obvious TBI Normal
Distribution of T Scores
30RIM Sample Case 2 Subtle Diabetes
- Reason for Referral
- 2 yrs dangerous work habits. Eval to see if
atrial fib Type II diabetes impairs cognition.
Hospitalized TIA-like Sx. Admitted to problems
for 20 yrs, cardiac dysrhythmia bradycardia,
pacemaker, blood sugar difficult to manage,
family Hx of heart disease diabetes.
- Age 55
- Handed Right
- Race Euro-American
- Sex Male
- Ed 13 years
- Occup Mechanic
- Marital Married 20 yr
- Living at home w/wife
31RIM Sample Case 2 Subtle Diabetes
32RIM Sample Case 2 Subtle Diabetes
33RIM Sample Case 2 Subtle Diabetes Normal
Distribution of T Scores
34RIM Critiques Concern 1
- The method of calculating the standard deviations
(SDs) for summary statistics and domain scores is
incorrect. - Since many of the remaining steps of the RIM
depend on the use of these SDs, this error is
magnified in the subsequent steps. - SDs statistically can not exceed 9.99 and are
more likely to be around 6.4
35Response 1 RIM Means 4 Large Datasets
36Response 1 Inter-Individual Ms SDs
37Response 1 RIM SDs 4 Large Datasets
38Response 1 Intra-Individual Ms SDs
39RIM Critiques Concern 2
- More false-positives then clinical judgment.
- Palmer et al. (2004) expressed concern that
- We failed to distinguish statistical from
clinical significance. - This failure is a critical error that precludes
the prudent clinician from using the RIM.
40Response 2 RIM vs. Manual Detecting Differences
Overall
41Response 2 RIM vs. Manual Detecting
Differences ES
42Response 2 RIM vs. ManualDetecting Differences
Scaled Scores
43RIM Critiques Concern 3
- Clinicians who use the RIM will
- Idiosyncratically assign scores to cognitive
domains. - This will result in low inter-rater reliability
in analysis diagnosis.
44RIM Critiques Concern 4
- Scores on domains are unit weighted, which
introduces error. - Willson Reynolds (2004) said scores load on
multiple domains. Assignment to domains weights
depend on - Battery of tests administered.
- Patients whose test scores are being examined.
45Response 4 Cross-Validation Unit Wts
- Conducted 4 multiple reg. on 457 pts WAIS-R.
- Split sample in ½ - assess shrinkage.
- Regressed patients verbal subtests onto PIQ.
- Generated ideal weights for the 1st ½ of sample.
- Used wts to predict PIQs in the 2nd ½ of sample.
- Pre-PIQs regressed on actual PIQs 2nd ½ sample.
- Also, generated weights for the 2nd ½ of sample.
- Use Pre-PIQs regress on actual PIQs 1st ½
sample. - Repeated, except performance subtests predict VIQ
- split sample ½ generate same statistics as
before.
46Response 4 Cross-Validation Unit Wts
- Purpose of these procedures
- How much variance in wts. is sample specific.
- Amount of shrinkage using cross-validated wts.
- Shrinkage error compared to error introduced by
using unit wts vs. ideal wts. - Results 98 of the variance accounted for with
unit wts. Compared to ideal weights. - Support use of unit wts. Rather than ideal wts.
- Also see, Dawes, R. M. (1979). The robust beauty
of improper linear models in decision making.
American Psychologist, 34, 571-582.
47RIM Critiques Concern 5
- Multiple measures used to generate composite
scores - Results in less accurate estimates of the
cognitive domains.
48Response 5 Estimate FSIQ Using Scaled Score
Meanss
49RIM Critiques Concern 6
- A general ability factor is used to represent
premorbid functioning for all domains. - This not supported by the literature.
- This results in inaccurate conclusions regarding
degree of impairment suffered by a patient in
each cognitive domains assessed.
50RIM Critiques Concern 7
- Norms used come from samples that are of
undocumented comparability. - Furthermore, even when norms used were generated
from different but comparable samples, their
format prohibits ready comparisons.
51Response 7 Split-Half Reliability
- Analyze Dataset 2 - OTBMs from 42 DVs.
- Individuals data split into two sets
- 21 test variables for each OTBM (1 2).
- 2 independent OTBMs created for patient.
- Split DVs intentionally - separated so that no
normative sample was included in both OTBMs.
52Response 7 Split-Half Reliability
- Results r .81, 66 of variance accounted
- Slope of the regression line was .82 (SE .027)
- Intercept 9.2 (SE 1.20).
- Mean OTBM-1 45.0 (SD 7.3).
- Mean OTBM-2 43.6 (SD 7.2).
- Results simulate worse case scenario.
- used an entirely different set of norms.
- Est. test-retest r for OTBM 42 DVs increased r
.82 - .88 (Spearman-Brown correction).
53Response 7 Split-Half Reliability
- No overlap in normative samples.
- Worst-case condition, generally administer
instruments (e.g., WAIS-III) with OTBMs generated
from co-normed variables. - Meyers Rohling test-retest reliability of .86.
- When different norms used, often gave same
instruments (e.g., AVLT or RCFT). - Our simulation, no instruments included in OTBM-1
included in OTBM-2. - Heaton et al.s (2001) - schizophrenic pts.
- Obtained a test-retest reliability of .97.
- Comparing 2 identical batteries, not worst-case.
54RIM Critiques Concern 8
- The RIM will result in an undue inflation of
clinicians confidence. - Such overconfidence results in more error in a
interpretation, not less.
55RIM vs. Tulsky et al. (2003) Case 1
56RIM vs. Tulsky et al. (2003) Case 2
57Summary of the Rohling Interpretive Method of
Statistical Analysis of Neuropsychological Data
58Summary of RIM Steps
- 24 total steps to the RIM process
- 17 calculation steps
- Battery Design
- Calculation of summary statistics
- Generation of graphic displays
- 7 interpretative steps.
- Use of summary table and graphic displays to
- Assess and verify summary data
- Identify strengths/limitations of current data
- Obtain a reliable diagnosis
- Develop tx plans based on clinical judgments.
59Summary of RIM Advantages
- Formulize thinking interpretation of data
- Operationalize what you already do.
- Reduce judgment errors thru RIM Process.
- Take advantage of psychometric properties at
level with fixed, co-normed batteries. - Allows flexibility of test selection.
- Allows flexibility of theoretical view of
cognition (e.g., domain structure)
60Summary of RIM Advantages contd
- Gives Quantitative support for your conclusions
and interpretations - Statistical evaluation
- Measure of confidence in findings
- Measure of limitations of findings
- Ability to present data at different levels of
interpretation - Equals greater defensibility
61Our RIM Cautions/Concerns
- Does not replace clinical judgment, rather,
informs clinical judgment. - This still means CJ errors are possible.
- Susceptibility T-Scores to distrib. deviance
- Process, not program
- Pre-morbid ability estimates
- Domain selection, test placement
62Published Research Findings Using the RIM
- 1) RIM vs. HRB
- 2) Variance Accounted for by SVT
- 3) Effect of Depression on NP Results
- 4) Prediction of Employment after TBI
63RIM of HRB OTBM vs. HII
- Heaton et al.s (1991) HRB norms for OTBM
- T Score (M50, sd10)
- OTBM r with HII -.79
- (p lt .0001)
- 62 variance account.
- Over predicts low
- Under predicts high
64RIM of HRB OTBM vs. GNDS
- OTBM r with GNDS -.87
- 76 variance acc.
- OTBM neither under nor over predicts across range
of GNDS - Intercept impairment is T Score 46.0
- Reitan Wolfson (GNDS 29)
65RIM of HRB OTBMs Relationship to Global Indices
66RIM of HRB Diagnostic Classification Using the
HII
67RIM of HRB Cross-Validation of RIM using HRB in
2 Samples
- Regressed Dikmen Meyers TBI data
- Generated a predicted HII for pts in OK dataset.
- Correlation actual predicted HII .95
- Sense .60, Spec .77, PPV .78, NPV .59
- Overall Correct Classification 71
- Predicted HII from MSBs OTBM more accurate
indicator of impairment than actual HII.
68Factor Loadings of Domain Scores
69Means SDs of Composite Scores
70 Mean z Score on Objective Tests
- Small differences between Gen. Normal Gen.
Neuro on NPT. - No differences between Exag Normal Exag Neuro
on NPT. - Deficits for Exag Neuro were more modest than for
Exag Normals on SVT. - Interaction between Validity Neuro Status.
71 Mean z Score Self-Report
- No differences between Gen Neuro Exag Neuro on
Memory Complaints. - No differences between Gen Exag Neuro on
Psychiatric Sx. - Deficits for Exag Normal on the Psych Sx Memory
Complaints latter is larger. - Interaction between Validity Neuro Status.
72Depression Study Reference
- Rohling, M. L., Green, P., Allen, L. M.,
Iverson, G. L. (2002). Depressive symptoms and
neurocognitive test scores in patients passing
symptom validity tests. Archives of Clinical
Neuropsychology, 17, 205-222.
73Mood Group Assignment
- Patients classified into 2 subgroups
- From entire sample, 420 passed all SVTs
- Sample split based on BDI
- Low-Depressed 25ile on BDI (lt 10)
- n 178, M 6 (3)
- High-Depressed 75ile on BDI (gt 25)
- n 187, M 31 (6)
74Depression Study Participants
- All 365 patients referred for evaluation for
compensation-related purposes - All diagnostic groups included
- 53 Head injury referrals
- 22 Medical referrals
- 14 Psychiatric referrals
- 11 Other neurological
- Age 42 (11) Ed 13 (3) Sex 64 males
Non-English 18 Handedness 9 Left
75Results Mood Validity Status
SVT Status
Mood BDI
Genuine
Exaggerating
175 (48)
Depress 75ile
186 (52)
NonDep 25ile
266 (74)
95 (26)
76Results Sample Split by Validity
77Effect of Mood Depends on Effort
- Exaggerating patients accounted for
- 39 of High-Dep group
- 14 of Low-Dep group
- Mood Effort used as IVs and Cognition DV
- Effect for effort, no effect for mood
- However, when Memory Complaints DV
- Effects for both effort and mood
- Also, when other Emotion/Personality DV
- Effects for both effort and mood
78Effect of Mood Depends on Effort
- When both mood groups were included in
regression analysis, as predicted - Memory ratings related to mood
- (r .60 p lt .0001)
- Mood not correlated with cognition
- (r .10 p gt .10)
- Memory ratings not related to cognition
- (r .13, p .06)
79Mood Replication
- Gervais pain sample findings (n 177)
- Exaggerating patients accounted for
- 55 of High-Dep 33 of Low-Dep group
- Memory ratings related to mood (r .55)
- Mood not correlated with cognition (r .06)
- Memory ratings related to cognition (r .15)
- Group means correlated with Greens .94
- all patient (High-D, Low-D, Gen, Exag).
80Effect if Pain on OTBM
81Effect if Pain on OTBM
82Return to Work after Injury
- Three main hypotheses using MSB-RIM
- OTBM will predict return to work level
- Cognitive domain that will be most predictive
will be executive function - Adding the Patient Competency Rating Scale will
improve work prediction - PCRS is by Prigatano (1985)
83Return to Work ANOVA of OTBM
84Logistic Regression Using OTBM
85Return to Work Summary
- OTBM differences between groups.
- Disabled /Unemployed not able to separate.
- Below/At Previous not able to separate.
- Collapsed groups result in 71 correct
- above base rate of 52 correct.
86Return to Work Domain Analysis
- Executive function not the most predictive.
- Most of variance carried by Perceptual
Organization Working Memory - Using Cognitive Domains
- OTBM increases Correct from 71 to 74
- Incremental validity of PCRS very low.
- 7 of the variance
87Return to Work Domain Analysis
- By including premorbid variables, increases
diagnostic accuracy most helpful being - Premorbid IQ, level of occupation, education
- Including acute measures increases accuracy most
helpful being - LOC group
- Time since injury
88Depression Study Conclusions
- Memory complaints not synonymous with impairment
in compensation sample - Findings replicated
- Effort accounts for more variance in self-ratings
of cognition objective performance than mood - Findings replicated
89Rohlings Interpretive Method Use of
Meta-Analytic Procedures for Single Case Data
Analysis
- Martin L. Rohling
- L. Stephen Miller
- Questions Comments Welcome!