Transcript Disparities in Treated Prevalence among Medicaid

Lecture III: Data collection and analysis
Mihail Samnaliev, PhD
Senior Health Economist
Children's Hospital Boston
_______________
The speaker for this session has reported NO FINANCIAL RELATIONSHIPS with a
commercial entity producing healthcare-related products and/or services
1
Outline



Data sources
Data analysis
Economic models
2
What cost data do you need?
Q1: Where do we get the data?
Q2 : Is it good quality?
Q3: Does it capture all resource utilization?
Sources of data
1. Health care utilization and cost data at BCH
2. External sources of health care utilization / costs
3. Collecting data using surveys
3
What data do we need?

Focus on types of resources /costs expected to change
• Intervention, Health care costs, non-health costs, etc.

Depends on the perspective
• Provider, HC system, payer, societal
• Determines type of data and time to obtain it
• Different perspectives may lead to different
results/conclusions
4
1. BCH data
1.1 Alliance

The hospital’s departmental costs (including overhead) are
assigned to individual (billable) procedures and activities
based on relative value units (RVU)

Data contains:
•
•
Charges, costs , overhead costs
Other data fields: MRN, Date of admission and discharge, whether
emergency or/and inpatient, Department, service description, etc.
5
1. BCH data
MRN
Admit Discharg Type SERV DEPT Dept CPT/ CPT/HCP Billed DIRECT INDIRE TOTA
e
DATE NAME Group HCP CS
Units COST
CT
COST
CS NAME
COST
XXY 11/15/ 11/16/20 Eme 11/15 diagno
2013
08 rgen /2008 stic
cy
radiolo
gy
XXY 11/15/ 11/16/20 Eme 11/15 diagno
2013
08 rgen /2008 stic
cy
radiolo
gy
XXY 11/15/ PREDNISOLONE
2013 ORAL
Diagn 710
ositc 20
Servic
es
Diagn 710
ositc 20
Servic
es
Phar
macy
CHEST
X-RAY
1
$43
$53
$96
CHEST
X-RAY
1
$41
$32
$74
$14
$4
$14
6
1. BCH data
1.2 Clinical/utilization data from the specific departments
• Alliance does not contain professional fees
• Differences in costs need to be adjusted for background
differences, which may not be available in Alliance
7
Charges, payments and costs
Evaluations should be based on:
• costs (societal/ HC system perspective)
• payments (payer’s perspective)
$100
$90
$80
$70
$60
Charges
$50
Payments
$40
Costs
$30
If only charges are available, they should
be converted to costs, preferably using
department-specific cost-charge ratios
$20
$10
$0
$
Costs are on average ~65% of charges, but
varying between 20% to 90% depending
on department, procedure, etc.
8
2. External sources of health utilization/costs

Administrative, e.g. claims data

National Surveys

Databases of published economic evaluations

Other datasets (e.g. The KID’s inpatient dataset, PHIS)
9
2.1 Claims data

Widely used

Large n; rich datasets

Captures all (most) health care utilization (not just inpatient)
10
2.1 Claims data




Created for billing purposes; not tailored to answer specific
research hypotheses
Limitations similar to all observational studies (e.g.
selection bias, missing data, inadequate control groups)
Does not include important information that may confound
with outcomes (e.g. results from a severity)
Coding limitations (e.g. certain conditions may be
misdiagnosed /underdiagnosed, because of payment
incentives; diagnosed prevalence ≠ true prevalence)
11
2.1 Claims data
• Typically need to request from each payer
• Medicaid data: from MassHealth or CMS Max files (more
•
than one state)
Takes time to obtain permission and data
12
2.1 Claims data: APCD
• All payers claims database (APCD)
• Combines all private payers and Medicaid
http://www.mass.gov/chia/researcher/hcf-dataresources/apcd/
• Application -> review - > approval
• Need to specify what files (medical, dental, pharmacy,
provider, eligibility) and variables you need
13
2. National Surveys
Medical Expenditures Panel Survey (MEPS)





Nationally representative (about 30,000 respondents/year)
Detailed data on annual health care utilization and
expenditures
Diagnostic information
Employment, earnings, income
EQ-5D and SF-12 questions (earlier versions)
14
Example: Using MEPS to estimate costs
associated with a particular condition

Economic burden associated with eating disorders (ED)
and comorbidities:

Health expenditures over 12 months

•
•
Individuals with ED vs general population
ED and comorbidities vs ED only
Employment, earned income, family income
•
•
Individuals with ED vs general population
ED and comorbidities vs ED only
15
2.3. Databases of published health economic
evaluations
Cost-effectiveness analysis registry (Tufts)
https://research.tufts-nemc.org/cear4/
UK’s NHS Economic Evaluation Database
Includes cost-benefit analyses, cost-utility analyses, and
cost-effectiveness analyses.
http://www.crd.york.ac.uk/CRDWeb/ResultsPage.asp
The Paediatric Economic Database Evaluation (PEDE)
http://pede.ccb.sickkids.ca/pede/database.jsp
16
3. Data collection using surveys
• As an alternative way to measure health utilization
• Ask respondents about health care utilization events
• Monetize utilization (e.g. using average costs by setting)
• Non-health costs
• Recall period can be longer (e.g. 12 months) for major
events, but shorter otherwise
• Whenever possible use validated surveys
17
3. Example: Asthma related health care
events (Akinbami et al. 2012)
Core outcomes
• Asthma-specific hospitalizations, ED visits, outpatient
visits, medication use,
Supplemental outcomes
•
•
•
•
•
Primary care: scheduled
Primary care: unscheduled
Respiratory healthcare use
School absence
Work presenteeism and absenteeism
18
Examples
“During the past 12 months, how many times did [you/your child]
visit an emergency room or urgent care because of asthma?
NHIS 1997-present
“In the past 3 months, [have you/your child] taken prescription
asthma medicine using an inhaler?” IF YES
“In the past 3 months, what medications did [you/your child] take
by inhaler” Mark all that apply INSERT LIST OF MEDICATIONS
BFRSS Asthma Call-back Survey 2005-present
19
Common Non-Health care costs
Lost productivity
Lower rates of employment
Lower earnings due to fewer hours at work
Reduced productivity while at work
Missed school days
Linked with school performance
May impact future earnings,
Dependence on welfare programs
Special education placement
20
Common Non-Health care costs
Time costs
Even if not working, time is valued by individuals
Labor theory suggests tradeoff between labor and leisure
Caregiver costs
Formal caregivers
Informal (e.g. family members)
Travel costs
Commute, parking, tolls
21
Measurement of lost produtivity
Work Productivity and Activity Impairment Questionnaire (WPAI )
Work Limitations Questionnaire (WLQ)
The Health and Work Performance Questionnaire (HPQ)
Health and Work Questionnaire
Endicott Work Productivity Scale
The Health and Labor Questionnaire
Source: Prasad M, Wahlqvist P, Shikiar R, Shih YC. A review of self-report instruments measuring healthrelated work productivity: a patient-reported outcomes perspective. Pharmacoeconomics. 2004;22:225–244.
22
Example: WPAI: GH
Are you currently employed (working for pay)? _ NO_ YES
During the past seven days, how many hours did you miss from work
because of your health problems? “
Consider only how much health problems affected
productivity while you were working
Health problems
had no effect on my
work
0
1
2
3
4
5
6
7
8
9
10
Health problems
completely prevented
me from working
Source: Work Productivity and Activity Impairment Questionnaire: General Health V2.0 (WPAI:GH)
http://www.reillyassociates.net/WPAI_General.html
WPAI outcomes are expressed as impairment percentages
23
Travel costs
What mode of transportation did you take to come to BCH?
Bus or train  How much did you pay one way to BCH? _Dollars
Taxi
 How much did you pay one way to BCH? __Dollars
Car  How many miles did you drive one way to BCH? __Miles
Do you have to pay for parking upon leaving?
Yes □ No
24
Aside: Often need to interpolate
100
90
80
70
60
50
Treatment
Clinical endpoints (as before)
40
Placebo
HRQL
30
Health care events/utilization
20
Non-health care costs (e.g. lost work days)
10
0
0
6
12
18
24
25
Outline




Sources data
Analysis
Economic models
Planning an evaluation
26
Accounting for inflation
(1) List each resource utilization and (2) multiply each service
/procedure by the same unit costs (e.g. 2014)
E.g. a patient had
•
•
•
1 X Ray in 2010 which cost $50
2 X-Rays in 2014 which cost $60 each
Total = 3*60 = $180
Eliminates effects of inflation and other price fluctuations
Feasible when we have data for each resource that is used.
Not feasible for total medical expenditures
27
Accounting for inflation
II. Inflation indices, e.g CPI –U from the BLS
28
Inflation

Example:
• Sept 2009 index=398.682
• Sept 2010 index=410.327
• Inflation = (410.327-398.682) / 398.682 = 2.9%
• $1000 (2009 dollars) =$1029 (2010 dollars)

Most indexes calculated monthly, some once every 2
months.
29
Discounting

Reflects time preferences. Particularly important in future
projections with differential timing in costs and benefits/effects

$100 today preferred to $100 tomorrow (regardless of
inflation)

Discounting is applied to both costs and effects (e.g. QALYs,
symptom days, etc.)

Recommended rate 3%, but in sensitivity analyses vary 0 to
7% (U.S. Panel on Cost-Effectiveness in Health and Medicine 1996)
30
Distribution of cost data

Skewed

Many zeroes (e.g., most pts not hospitalized within past year)

Greater variability than clinical outcomes

Outliers
31
Distribution of cost data
32
Distribution of cost data
Inpatient care
33
Descriptive analyses

Average, sd, median, Q1-Q3, min max

Average costs often most meaningful to decision makers
•
•
Expected TME per patient
Used in cost-effectiveness analysis (medians may lead to biased
conclusions about cost-effectiveness)
34
Regression analyses to estimate
incremental costs and effects

Ordinary least squares (OLS)

OLS on transformed costs (e.g., log Cost)

Generalized linear models (GLM)

Longitudinal evaluations
•
E.g., interrupted time series analyses
35
Regression Analysis: OLS



Often used as a starting point /naïve model
Unbiased estimates if sample size is large
Concerns with influential outlier observations
36
Regression Analysis: Log transformations

E(ln(y)/x)=Xβ

Problem when many $0 (generates missing data)

•
•
IF not too many $0, replace 0 with a small cost, e.g. $1
Criticized as arbitrary
Lack of normality may still persist, although transformed
costs typically look ‘more normal’
37
Regression Analysis: Log transformations

What do you do with estimates based on logged costs?
“Congress does not appropriate logged dollars”

Can not take the antilog

One solution is the Dual smearing estimator to convert back
to dollars
Naihua Duan Smearing Estimate: A Nonparametric Retransformation Method,
Journal of the American Statistical Association, Vol., 78, No. 3838. (Sep., 1983), pp.
605-610.
38
Regression Analysis: GLM
GLM
g (E(y/x))=Xβ ,
g is the link function

Specify a link function between E(y/x) and Xβ ( log link is
typically used for costs)

Distribution (gamma typically appropriate for costs)
OLS on logged costs:
GLM with a log link :
E(ln(y)/x)=Xβ
ln(E(y/x))=Xβ
39
Example: Total ED cost
40
Example: Total ED cost
Estimate


Pvalue
Method
Incremental cost
1. Linear model (OLS)
-$101
0.013
2. Log Cost (Duan
retransformation)
-$191
<0.001
3. GLM
-$113
0.006
Why does Duan produce such different results? Possibly because of
heteroscedasticity (White test p<0.001; Null: errors are homoscedastic)
Often report results from more than one models
41
Two part models
Inpatient care
42
Two-part models

Often there is a large % clustered at 0
•
•



COSTS: e.g. 5% of the study population was hospitalized in the past
year ; therefore, 95% of the inpatient costs are $0
UTILITY, e.g., Half or the respondents had utility of 1, the rest < 1.
Part 1; Estimate probability of utilization/utility<1
Part 2 Costs among utilizers/patients with U<1
Considered good practice
• more transparent, circumvents problems with extremely
uneven distribution of the aggregate data
43
Aside: Why do costs vary so much

Large variability in costs across geographic areas
• Different and many inputs
• Prevailing wages/cost of living depend on region
• Market imperfections, monopolies
• Methodologically difficult to accurately estimate costs
• Total costs combine multiple resource use
• Variability driven by complex treatments/medical
conditions
44
Modeling for economic evaluations

Randomization (e.g. CEA added to RCT)

Observational (more common in health economics)

Modeling can be viewed as an alternative /supplemental
45
Economic models
Allow assessment of broad and long term economic impacts:
To extrapolate beyond the duration of a trial
To project to other populations
Useful for complex decision processes
To combine all available data
• E.g., BCH costs with HRQL (from other studies)
To synthesize published studies
46
Economic models
Randomized controlled trials (RCTs) may not include all
relevant comparator interventions
 The duration of follow-up in RCTs may be limited
 Patient populations in RCTs may not be reflective of
plan populations
 Safety data may be limited, or from disparate sources
 Healthcare cost impacts may not be generalizable
across payers
 Allow formal assessment of uncertainty
(AMCP 2009)

47
Example of a simple decision tree model
Assume: $0 costs for no comorbidities
$1000 /year for pts with diabetes
48
Hypothetical example
INT
Year 1
Year 2
Total
CONTR
Year 1
Year 2
Total
No
Diabetes
Heart disease Costs
1000
600
360
200
320
200
320
$400,000
$640,000
$1,040,000
1000
400
160
300
420
300
420
$600,000
$840,000
$1,440,000
49
Example of a simple decision tree model
50
Economic models: more information
More general about the use of models
NICE
http://www.nice.org.uk/media/b52/a7/tamethodsguideupdatedjune2008.pdf
The AMCP Format for Formulary Submissions Version 3.1
http://www.amcp.org/practice-resources/amcp-format-formulary-submisions.pdf
Specific recommendations:
ISPOR
http://www.ispor.org/workpaper/practices_index.asp
51
A checklist for economic evaluations
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Who are the users of the evaluation (hospital, researchers, government, a
particular payer)?
Evidence base (e.g. published economic evaluations in the field)
Determine the type of economic evaluation that will be appropriate
Scoping: comparators, population, subgroups
Research design: observational, added to RCT or economic model?
Define the perspective/s of analysis
Determine the analytic time frame
Define all outcomes and how they will be measured
What categories of health and non-health costs will be included?
Start planning the data collection: available datasets that you may need to
request or collect own data (e.g. surveys)
Write a brief research proposal (e.g. outline of background, aims, data,
proposed economic analyses, statistical tests, timeline)
Research team: PI, health economist, statistician, project manager, etc.
Clearly define roles and responsibilities; schedule periodic meetings;
Time frame: Can everyone commit in terms of time, effort, funding?
IRB and patient consent: include economic endpoints
52