GMS Formula Analysis QRESEARCH 2005 09 Feb 2006

  • Slides: 28
Download presentation
GMS Formula Analysis QRESEARCH 2005 09 Feb 2006 Julia Hippisley-Cox Jon Ford Ian Trimble

GMS Formula Analysis QRESEARCH 2005 09 Feb 2006 Julia Hippisley-Cox Jon Ford Ian Trimble

Aims of presentation – Brief overview of methods – Present key results from analysis

Aims of presentation – Brief overview of methods – Present key results from analysis – Comparison of models – Hand over to Jon Ford

Overall aim of the analysis • To derive a regression model linking workload to

Overall aim of the analysis • To derive a regression model linking workload to patient and practice characteristics • To inform revision of the funding formula for essential and additional services

Sampling: Practices Practice inclusion criteria for analysis ØEngland Wales only Ø At least 1000

Sampling: Practices Practice inclusion criteria for analysis ØEngland Wales only Ø At least 1000 patients ØAt least 2 consultations/person-year ØComplete data for period in question Ø Decided not to sample proportionately by region

Patient inclusion criteria • Patient level analysis • Study period 01 April 2003 -

Patient inclusion criteria • Patient level analysis • Study period 01 April 2003 - 31 March 2004 • Included if registered at any point during study period • Included temporary residents, new patients and patients who died • Person days denominator for rates

Principal outcome • Number of consultations (GP + nurse) in study year • Regardless

Principal outcome • Number of consultations (GP + nurse) in study year • Regardless of setting • Excluding community/district nurses

Patient level variables • • Sex Ageband: standard as in Carr Hill Registration period

Patient level variables • • Sex Ageband: standard as in Carr Hill Registration period (6+ months; <6 or new) Temporary patients (yes/no) New GMS diseases (yes/no for each) Townsend score/IMDS % white/non white

Practice level variables • • List size Number of GP principals Townsend score Rurality

Practice level variables • • List size Number of GP principals Townsend score Rurality White/non white Mean prevalence of QOF diseases Region

Patient level analysis • Variables included at patient or at practice level • Both

Patient level analysis • Variables included at patient or at practice level • Both person years and registered population were used

QRESEARCH practices Compared with UK average – Similar size – Similar distribution – Similar

QRESEARCH practices Compared with UK average – Similar size – Similar distribution – Similar prevalence – Similar age-sex – Comparable consultation rate LARGE Representative sample Results generalisable

Results: study practices • 454 practices in England Wales • 3. 8 million patients

Results: study practices • 454 practices in England Wales • 3. 8 million patients registered at any point in study year • 33, 727 deaths • 319, 435 new patients • 97, 239 temporary residents

Summary of comparison QRESEARCH practices • Slightly bigger • More in East Midlands and

Summary of comparison QRESEARCH practices • Slightly bigger • More in East Midlands and fewer in London • Otherwise similar w. r. t. age-sex and disease prevalence

Prevalence of diabetes in patients over 15

Prevalence of diabetes in patients over 15

Consultation rates by age and sex

Consultation rates by age and sex

Models • We fitted a series of ‘a priori’ statistical models specified in our

Models • We fitted a series of ‘a priori’ statistical models specified in our protocol and then were asked to fit additional ones • ‘a priori’ models included patient level assigned data where available (eg QOF diseases, Townsend score) • Supplementary models included practice level data (QOF disease prevalence, mean Townsend score)

Results: A priori Model 7 bi (person years denominator) Consultation rates: – Registered for

Results: A priori Model 7 bi (person years denominator) Consultation rates: – Registered for 6+ months = baseline – Registered for < 6 months = 72% higher rate – Temporary residents = 86% higher rate • Person years controls for length of registration period • patients registered within 6 months before start of study year or during study year have a 72% higher consultation rate compared to long standing patients

A priori model: Townsend score Fairly flat gradient with deprivation (Quintile 5 is deprived)

A priori model: Townsend score Fairly flat gradient with deprivation (Quintile 5 is deprived) – Quintile 1 = baseline – Quintile 2 = 0. 4% higher – Quintile 3 = 1. 4% higher – Quintile 4 = 4. 1% higher – Quintile 5 = 6. 1% higher

A priori model: Rurality and ethnicity Urban areas = baseline Rural areas = 1.

A priori model: Rurality and ethnicity Urban areas = baseline Rural areas = 1. 7% higher Ethnicity: 99 -100% white = baseline 97 -98. 9% white = 0. 5% lower 90 -96. 9% white = 4. 1% lower < 90% white = 11. 6% lower

A priori model: QOF diseases For all diseases, people with the disease had higher

A priori model: QOF diseases For all diseases, people with the disease had higher consultation rates compared to those without the disease e. g. CHD = 38% higher Diabetes = 54% higher Asthma = 63% higher

A priori model: practice variables List size: 2. 2% lower rate for each additional

A priori model: practice variables List size: 2. 2% lower rate for each additional thousand patients for a given number of GP principals (head count not wte) 1. 4% higher rates for each additional GP principal for a given list size

Process • Undertook patient level modelling • Then asked to do practice level modelling

Process • Undertook patient level modelling • Then asked to do practice level modelling for implementation • Concerns about how well practice level models can be applied at patient level • Results were counter-intuitive (Ecological fallacy)

Ecological fallacy • Applying practice level variables to a patient population produces spurious and

Ecological fallacy • Applying practice level variables to a patient population produces spurious and counter-intuitive results • Well described statistical phenomenon • Practice level models don’t work

Additional model : (practice level data) Inclusion of all QOF disease prevalence values together

Additional model : (practice level data) Inclusion of all QOF disease prevalence values together in models showed some negative associations: e. g. CHD = 4. 7% lower rate Thyroid disease = 1. 1% lower rate both for a 1% increase in practice prevalence.

Additional model: Townsend score Inclusion of mean practice Townsend score showed a negative association:

Additional model: Townsend score Inclusion of mean practice Townsend score showed a negative association: Consultation rates were 2. 9% lower for a 1 unit increase in mean practice Townsend score

FRG review • Requested additional patient level model WITHOUT prevalence (model 18) • Key

FRG review • Requested additional patient level model WITHOUT prevalence (model 18) • Key comparison then is patient level with and without prevalence

Explanatory power Akaike Information criterion • AIC statistical test for explanatory power • Lower

Explanatory power Akaike Information criterion • AIC statistical test for explanatory power • Lower values indicator better models • Absolute value increases with sample size • Relative difference more important

AIC results Both models patient level, person years denominator, age-sex, rurality, ethnicity Model 7

AIC results Both models patient level, person years denominator, age-sex, rurality, ethnicity Model 7 b AIC = 16, 415, 351 – Townsend quintile – Prevalence – No region Model 18 – Townsend continuous – No prevalence – Region AIC = 16, 763, 190

Summary • Person years adjustment give better fit for new registrations/TRs • Patient level

Summary • Person years adjustment give better fit for new registrations/TRs • Patient level analyses produce intuitively acceptable results • Practice level analyses counterintuitive results likely to lead to controversy (ecological fallacy) • Comparisons between patient level models with and without prevalence are presented for Plenary’s consideration