Using the NNHS versus the LEHD
& NHC to Assess Whether
Nursing Home Staff Turnover
Affects Resident Outcomes
Sally C. Stearns1
Laura P. D’Arcy1
Daria Pelech2
1The University of North Carolina at Chapel Hill
2Duke University
UNC Institute on Aging
September 22, 2009
Supported by the National Institute on Aging and the Demography and Economics of Aging Research
(DEAR) Program at the Carolina Population Center (Grant 5-P30-AG024376)
Facilitated by the National Center for Health Statistics and the Triangle Census Research Data Center
Disclaimer
This research was carried out at the Triangle
Census Bureau Research Data Center
facility. The results and conclusions of the
paper are those of the authors and do not
indicate concurrence by the Census Bureau.
These results have been screened to avoid
revealing confidential data.
Overview
Turnover among nursing home staff problematic
High annual rates for nursing assistants (68% to 170%)
High costs to facilities
May compromise quality of care
Evidence on effect of turnover on outcomes
Mixed or inconclusive results
Most studies:
Don’t address endogeneity of turnover and outcome
Use small/non-representative samples
Use aggregated facility data
Research Question (Pilot)
What is the effect of facility-level turnover among certified nursing
assistant (CNA) staff on resident-level outcomes?
Real dearth of information nursing home staff turnover data
Pilot study conducted at RDC used 2004 National Nursing Home
Survey
Merged facility and area data with resident surveys
Good methods
Facility fixed effects
Proposed instrumental variables for endogeneity of turnover
But turnover data are single point in time (not annual)
per facility
Conceptual Model (1)
Area Economic
Indicators Turnover or Other Facility
- Employment Churning Characteristics
- Housing Value
Resident Resident Outcomes (Bad)
Characteristics - Hospital Use
- Sociodemographic - ER Use
- Medical/clinical - Ulcers
- Functional - Pain
- Falls
- Any of the Above
Conceptual Model (2)
Area Economic
Indicators Turnover or Other Facility
- Employment Churning Characteristics
- Housing Value
Resident Resident Outcomes (Bad)
Characteristics - Hospital Use
- Sociodemographic - ER Use
- Medical/clinical - Ulcers
- Functional - Pain
- Falls
- Any of the Above
Empirical Model: Pilot
Turnover=f(Facility characteristics, area IV)
Estimated using single year facility-level
observations
Bad Outcomes=f(Turnover, resident
characteristics, other facility characteristics)
Single year multiple resident-level observations
per facility for cross sectional pilot study
Area Instruments:
Pilot & Proposed Study
County unemployment
Median home value
Median income
Percent housing units vacant
NA hourly mean wage
Food/beverage server hourly mean wage
HHI total certified beds
Data: Pilot Study
2004 National Nursing Home Survey
Started with1,140 facilities and 13,425 residents
Needed to work at Triangle Census Research to access file
created by NCHS
Can not merge public use versions of facility &
resident surveys
Exclusions (age<65 or missing data) resulted in a analysis
file of 9,279 residents at 981 facilities
Range of 1 to 12 residents per facility
Turnover Measures: Pilot
Two measures:
Turnover among certified nursing assistants
(CNAs) in the past three months (annualized)
Average over all residents: 52%
Proportion of CNAs on staff for less than one year
Average over all residents: 37%
Outcome Measures: Pilot
Resident-level observations of:
Hospital Admission in past 90 days (7%)
ED visits in past 90 days (8%)
Any pressure ulcer (10%)
Fell in past 30 days (16%)
Fell in past 31-180 days (28%)
Any pain in past 7 days (25%)
Any negative health outcome above (55%)
Methods: Pilot
Linear probability models
Facilitates FE and IV estimation
OK if reasonable variance in dependent variables
Adjusted for survey weights and clustering
Three types of models estimated:
Naïve LPM
Facility Fixed Effects
Facility Fixed Effects – Instrumental Variables
Results: Pilot Study
Any Bad Outcome (mean of 0.55)
FE are arguably the best estimates:
Increase in CNA turnover of 0.1 associated with 0.0025
increase in likelihood of bad outcome
Increase in proportion of CNAs at facility less than one year
of 0.1 associated with 0.0094 increase in likelihood of bad
outcome
Summary: Pilot
FE estimates show modest effect of turnover or low retention on
bad outcomes
Other observed facility characteristics had comparable effects
High occupancy or lack of care plan increased bad outcomes
For-profit status or offering fully paid health insurance for the
CNA’s family decreased bad outcomes
Effects were strongest for “any pain” outcome
IV estimates larger, but:
Weak instruments
Cross-sectional area instruments can not explain within-facility
variation in resident outcomes
Policy Implications: Pilot
Interventions to reduce CNA turnover are likely
beneficial and may reduce cost, but other observed
and unobserved facility characteristics may have as
great of an effect on resident outcomes
Comprehensive programs to ensure quality
administration and oversight at facilities may be
required to jointly reduce CNA turnover and improve
resident outcomes
Limitations: Pilot Study
Have not:
Allowed for non-linear effects of turnover or low retention
Controlled for staffing levels (though is picked up in fixed effects,
so estimation is quasi-reduced form)
Can not distinguish between turnover once in many positions
versus lots of turnover in a few positions
Cross-sectional data
IV correction may not work due to:
Weak instruments
Intrinsic problem that cross-sectional IVs can not explain within-
facility variation in outcomes
Research Question (Revised)
What is the effect of facility (establishment)
churning on facility-level resident outcomes?
Proposed Study: Merge Quality Workforce
Indicator (turnover) data with Nursing Home
Compare
Longitudinal facility-level panel will:
Facilitate IV approach
Provide within-facility variation in turnover over time
But lots of limitations, so is it worth it?
Proposed Study
Nursing Home Compare (NHC)
www.medicare.gov/nhcompare/
Annual facility-level records since 2003 of facility
characteristics, inspection results, residents, staff and
ratings
Would enable annual panel from 2003-2008 for up to
17,000 nursing homes (~15,000 free-standing??)
Quarterly Workforce Indicators (QWI)
Generated from Longitudinal Employment Household Data
(LEHD)
Provides measure of turnover for all employees at a firm
But only available for approximately 30 states
Currently available through 200? (at least 2004)
Empirical Model:
Proposed Study
Turnover=f(Facility characteristics, area IV)
Estimated using panel of annual facility-level
observations
Bad Outcomes=f(Turnover, resident
characteristics, other facility characteristics)
Facility-level observations for proposed
longitudinal study
Proposed Study Challenges
1. Limitations to turnover measure from QWI
Cannot distinguish employees or turnover by
position (e.g., nurses vs CNAs vs gardeners)
Establishment (facility) level measures available
only through a multiple imputation process
2. Merging NHC and imputed turnover
Can not get employer identification number (EIN)
for NHC facilities
Need to merge by name & address
1a. QWI Turnover Measure
QWI uniquely identifies:
Firm (SEIN)
Establishment (SEINUNIT)
Provides firm-level turnover measure
= turnover at time t for firm k
FA is # of full quarter accessions
FS is # of full quarter separations
F is average full quarter employment
1b. QWI Turnover Measure
Need to use multiple imputation to get establishment
(facility) turnover
Process developed by John Abowd at Cornell
Generates most likely establishment for each employee
based on distance, employee distribution within firm,
employee work history, and period of establishment
existence
Imputation validated in Minnesota (which associates
establishments & employees) and appears to work for
99.5% of employers.
2. Linking NHC Data to QWI
Nursing home is equivalent to establishment
(SEINUNIT), but EIN not available
Name, address, zipcode available; in theory can
get Medicare provider number or ***possibly***
even the EIN from Centers for Medicare &
Medicaid services
Two possible paths for linkage (but both have
problems)
Via the Business Register Bridge (BRB)
*MAYBE* via the Geocoded Address List (GAL)
Proposed Study Worth It?
Even if match does not work, arguably
valuable to Census & other researchers to
know that linkage is not currently feasible
If linkage works sufficiently well, then:
Valuable to Census/researchers to know
matching for other studies feasible
Longitudinal panel of annual observations on
facility turnover and aggregated resident
outcomes would enable strong FE and IV
estimation of relationship