Longitudinal Workforce Analysis using Routinely Collected Data: Challenges and Possibilities Shereen Hussein, BSc MSc PhD Kings College London.

Longitudinal Workforce Analysis using Routinely

Collected Data:Challenges and Possibilities

Shereen Hussein, BSc MSc PhD

King’s College London

Longitudinal Analysis

General advantages

• Can control for individual heterogeneity

• Subject serve as own control

• Between-subject variation excluded from error

• Can better assess causality than cross-sectional data

General challenges

• Conventional statistical methods require independence between observations

• Longitudinal data are likely to violate this assumption

• Missing data due to attrition

• Data availability

29/5/2012 2

Workforce Data Example: NMDS-SC

• Structure• Design• Coverage• Time span• Type of information collected• Data collection and archiving• size

29/5/2012 3

NMDS-SC data structureSocial care providers in England

Complete NMDS-SC returns

Aggregate information on the workforce Detailed information on

all or some individual workers

Providers’ Database

workers’ DatabaseLinkable

29/5/2012 4

NMDS-SC longitudinal analysis: potential

• Data coverage• Wide range of providers and individual workers’

information• Sector specific- uniqueness• Hierarchical structure• Workforce development and business sustainability• Timely

– Demographics, austerity, unemployment• Economics

– Care costs, including turnover costs– Pay

• Linkable to local data characteristics

29/5/2012 5

Challenges in NMDS-SC longitudinal analysis

• No sampling framework

• No regular intervals for data collection

• Irregularities in data completion by different providers

• Additions/alterations of variables and fields

• Cumulative nature and consequences on data size and structure

• Archiving 29/5/2012 6

Challenges in NMDS-SC longitudinal analysis- continued

Computational• Data size

– Innovation in system design and architecture

• Accumulative property– Scalability of the system

• Changes in data fields• Variable additions and

omissions• Data over-ride and

archiving– Software and hardware

issues

Methodological

• Unusual patterns of follow-up– Censoring

• Variability in the database over time

• Unbalanced cohort design• Missing data

– Update frequency– Attrition– True exit

• Other methodological issues

29/5/2012 7

Providers’ level longitudinal mapping

• From December 2007 to March 2011• Linked 18 separate databases on the providers’

level• Each has records from 13,095 to 25,266

421,671 valid records included in the construction

• Number of updates ranged from 0 to 18 per provider

• Continuous process, more records added every 3 months

29/5/2012 8

Meta-data analysis: providers with different number of events

29/5/2012 9

Specific example 1: Providers with 18 updates

29/5/2012 10

29/5/2012 11

Specific example 2: Providers with 2 updates

Density distribution plot of providers with at least 2 updates during the period December 2007 to March 2011

29/5/2012 12

density distributions of number of days elapsed between two updated providers’ events

29/5/2012 13

Simple example using providers’ database:

workforce stability over time• Longitudinal changes in care workers’

turnover and vacancy rates over time – From January 2008 to January 2010

• Changes in reasons for leaving the sector, identified by employers– Differentiating between those with improved

(reduced) turnover rates and those with worse (increased) turnover rates

29/5/2012 14

Pre analysis

• Selecting and constructing providers’ panel– Including those with at least two updates

within +/- 3 months of T1 and T2 – 2953 providers with mean coverage duration of

602d

• Investigate sample representation

• Data quality checks

• Data manipulation/imputation

29/5/2012 15

Some findings: changes in turnover rates

29/5/2012 16

Reason for leaving and turnover rate changes

29/5/2012 17

Analysis expansion: next steps

• Consider changes over a longer period of time• Examine other providers’ characteristics• Different take on panel inclusion criteria • Link to individual workers’ longitudinal

databases to examine relations with detailed workforce structure– Pay, qualifications, profile etc.

• Build economic elements within analyses models, e.g. specific-turnover costs, within the longitudinal model

29/5/2012 18

Workers’ level longitudinal analysis

• A much larger database– Same period of time- over 11M records

• Providers not required to complete information for ‘all’ workers– Structural/design missing data– True missing data

• Linkage issues – more data fields required for identification and linkage

• Considerably large number of variables and fields– Careful planning; analysis-tailored data retrieval

• Changes in database– Amendments, new variables etc.– Programming intensive and demanding models (may not be

replicable for different databases)29/5/2012 19

29/5/2012 20

Issues to consider• Suitability of models

– Longitudinal structure– Competing risks

• Measurement window– Late entry into risk sets

• Use proxies, other variables in the dataset• Adopt suitable approach/model

– Censoring (LHS and RHS)• Assumptions

– Guided by:• Sector-specific knowledge• Intelligence from other variables in the data

29/5/2012 21

Current longitudinal researchWatch this space!!

• Workforce mobility within the sector

• Occupation durations

• Characteristic-specific probabilities of exiting or remaining in the sector

• Characteristic-specific probabilities of moving employer within the sector or having multiple jobs

• Career pathways within the sector

29/5/2012 22

Acknowledgments

• Thanks to the Department of Health for funding this work

• Thanks to Skills for Care for providing the data on regular basis

• Thanks to Analytical Research Ltd for their technical and quantitative support

29/5/2012 23

Further information

• [email protected]

• 02078481669

• See:

• http://www.kcl.ac.uk/sspp/departments/sshm/scwru/res/knowledge/nmdslong.aspx

29/5/2012 24