Real-time mortality monitoring in England andWales

Pia Hardelid

Statistics Unit, HPA Centre for Infections

19 May 2010

1 / 26

Outline

Background

Mortality data flow

Rapid system for monitoring weekly mortality

Results from mortality monitoring

Retrospective attribution of weekly deaths to various causes

2 / 26

Background

Mortality is a key indicator of the severity of important healththreats such as pandemic influenza, or extreme cold or heat

Timely monitoring of mortality is therefore important torapidly assess public health impact

In response to H1N1 pandemic, HPA developed new system ofage-specific mortality monitoring system using weekly datareceived directly from General Registry Office (GRO).

3 / 26

Reporting structure for deaths

4 / 26

Mortality data provision 2009-2010

Prior arrival of pandemic influenza H1N1, only all-cause,all-age mortality data by date of registration was available,with a two week delay.

From April 2009, all deaths reported from registry offices toHome Office IPS using an online system (RON) sent to HPAdaily with one day’s delay

At end of April 2009 60% of deaths registered in RON, by 1stJuly 2009 100% of deaths on RON

Receive counts of death registered by age, registration districtand day of death

5 / 26

Rapid system for monitoring weekly mortality

Data monitored weekly on Wednesday (i.e. all registrations bythe Tuesday).

This allows deaths occurring during the weekend (and bankholiday Mondays) to be registered.

Daily deaths collated by ISO week and age group (<1, 1-4,5-14, 15-24, 25-44, 45-64, 65+), corrected for reporting delayand compared against baseline

Weekly report produced

Procedure largely automated using STATA, R and LATEX

6 / 26

Estimating baseline mortality

Deaths registered between 1999 and 2008 by age and date ofdeath used to estimate baseline

Poisson regression models with cyclical terms (similar tomethod suggested by Serfling, 1960) fitted:

ln(yi ) = β0 + β1(weeki ) + β2(sin 2π(weeki )52 ) + β3(cos 2π(weeki )

52 ) + εi

7 / 26

Options for excluding previous periods of high mortality

Exclude years with influenza epidemics (suggested by Serfling)

Exclude deaths during summers and winters

Exclude weeks with high influenza activity (previous HPAmethod)

Downweight observations according to size of Anscomberesiduals (suggested by Farrington et al, 1996):

weighti =

{r−2Ai if | rAi |> 1

1 otherwise

8 / 26

Calculation of baseline and predicition intervals

Downweighting of residuals and model refitting was carriedout twice

Standard errors were rescaled if overdispersion present

99% prediction intervals estimated (Farrington et al, 1996):

p99i = (y2/3i + 2.58 ∗ (

4 y1/3i9 ∗ (ϑ+ yiε

2i ))1/2)3/2

Any observed death count > prediction interval is inexceedance

9 / 26

Reporting delays

Delays vary greatly by age

In the short term they also depend on the day of the week ofdeath (weekend effect) and holidays.

Deaths either get reported within about 10 days or they fallinto a group that can take many months (coroners’ inquests).

10 / 26

Reporting delay distribution

0 10 20 30 40 50

20

40

60

80

10

0

Delay (weeks)

% d

ea

ths r

ep

ort

ed

65+

45−64

<1

5−14

25−44

1−4

15−24

11 / 26

Dealing with reporting delay in observed weekly mortalitydata

Observed delay distribution by age group is calculated weeklyevery Wednesday with reference to deaths occurring in week 1being reported by Tuesday week 2. All deaths assumed tohave been reported by 2 years

Adjust observed deaths using observed delay distribution

Upper prediction limit of baseline also adjusted for uncertaintydue to reporting delay

12 / 26

Mortality baseline

●

●

●

●

●●●●●

●●

●●

●

●●●●●●●●●

●●●●●

●

●●

●●●●●●●

●

●●

●●●●

●●

●

●

●

●

●

●

●

●

●

●

●●

●●

●

●●●●

●●

●●●●●●●●●●●●●●●●●●●●●●●●

●

●●

●●●●●●●

●

●

●

●●●

●●

●●

●●

●

●●

●●●●●●●●

●●●●●●

●●

●

●●●●●●●

●●●●●●●

●●●●●●

●

●

●

●

●●

●

●

●

●●●●

●

●●●●●●

●●●●

●●●●●●●●

●

●

●

●●

●●●

●●●

●

●

●●●●●●

●

●

●●●●

●

●●●

●●●

●●●●●●●●

●●●●

●●●●●

●●

●●

●

●

●

●●●

●●

●●●

●

●●●

●

●●

●●

●●●

●

●●

●

●●

●●●●

●●

●●●●●●●●●●

●

●●●●●●

●

●●●●●●

●●

●●●●●●

●●●●

●

●

●

●

●

●●

●●●●●●●

●

●●●

●●●●●●

●●

●●

●●

●

●

●

●

●●●●●●●●●●●●●●

●

●

●●●

●

●●●

●●●●●●●●●

●●

●●

●

●●●●●●

●

●

●

●●

●

●

●

●●●●●●●●

●●

●●

●●●●●●●●

●

●

●●●●●

●●●

●

●●●●●●

●●●●●●●●

●●●●●●

●

●

●●●●●●

●●●●●

●

●●

●

●

●●●

●

●●●

●

●●●●

●●

●●●

●●●

●●

●

●●●●●●●●●●●●●

●●●●●●●●

●●●●

●●●●

●

●

●

●

Time

We

ekly

de

ath

s

80

00

10

00

01

20

00

14

00

01

60

00

18

00

0

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008

● ObservedModel fit: baseline99% upper limit

Observed weekly deaths and baseline mortality 1999−2008

13 / 26

Findings from mortality monitoring2009/2010

14 / 26

● ● ●

●●

●

●●

●

●

●

●● ●

● ●●

●

●

●

●

●

●●

●

●

●●

●● ● ● ● ● ●

Week

Estim

ate

d s

win

e flu

ca

se

s

04

00

00

80

00

0

23−2009 27−2009 31−2009 35−2009 39−2009 43−2009 47−2009 51−2009 2−2010

−5

51

52

5

Te

mp

era

ture

(C

°)

●

TempCases

All age mortality, temperature and estimated H1N1 casesJun 2009−Jan 2010

● ●

● ● ●

● ●●

●●

●● ●

● ●

●●

●

●●

●

● ●● ● ●

●●

●

●●

●

●

●

●

Week

De

ath

s

80

00

10

00

0

23−2009 27−2009 31−2009 35−2009 39−2009 43−2009 47−2009 51−2009 2−2010

● DeathsPredicted99% UL

15 / 26

Observed mortality by age group: children 1-14 years(highest H1N1v incidence)

●●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

Week

Nu

mb

er

of

de

ath

s

05

10

15

20

25

30

18−2009 26−2009 34−2009 42−2009 50−2009

1 to 4 years

●●

●

●

●

●

●

●●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

Week

Nu

mb

er

of

de

ath

s

05

10

15

20

25

30

18−2009 26−2009 34−2009 42−2009 50−2009

5 to 14 years

16 / 26

Mortality by age group: 65-74 years

●

●

●

●

●

●●

●

●

● ●

●

● ●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

Week

Num

ber

of d

eath

s

1300

1400

1500

1600

1700

1800

1900

2000

18−2009 26−2009 34−2009 42−2009 50−2009

All regions, 65 to 74 years

●

●

●

●

●

●

● ●

●

●

● ●

●

●

●

● ●

●

●●

●

●

● ●●

●

●

●

●

●●

●

●●

●●

●

●

●

●

Week

Num

ber

of d

eath

s

2500

3000

3500

4000

18−2009 26−2009 34−2009 42−2009 50−2009

All regions, 75 to 84 years

17 / 26

Mortality by age group: 85+ years

●●

● ●

●

● ●

●

● ●

●●

●

●●

●

●

●●

● ●

●

●

●

●

● ● ●

●

● ●

●●

●

●

●

●

●

●

●

Week

Nu

mb

er

of

de

ath

s

30

00

35

00

40

00

45

00

50

00

18−2009 22−2009 26−2009 30−2009 34−2009 38−2009 42−2009 46−2009 50−2009 1−2010

All regions, 85+ years

18 / 26

Further work

Daily model (more useful during heatwaves)

Use of CUSUM, to detect smaller but sustained shifts frombaseline

Evaluate differences between using date of death vs. date ofregistration in models

Attribution of excess at end of season

19 / 26

Retrospective attribution of weekly deaths to variouscauses

Not possible to specify cause of excess mortality using currentmonitoring system

Parliamentary questions and queries from Chief MedicalOfficer following high mortality observed during 2008/9 winter

20 / 26

Data sources

Weekly mortality data by date of death from Office forNational Statistics 1999-2008

Central England Temperature data (daily mean, minimum,maximum) converted to weekly time series

Counts of positive isolates from laboratory surveillance forinfluenza A, B and RSV (note: no denominators)

21 / 26

Suggested method

Negative binomial GLMs with identity link function fitted withall-cause, all-age mortality by week of death for whole countryas dependent variable

Year (as categorical variable) temporal trend

Temperature fitted as linear spline, with 4 d.f (pre-specifiedknots to indicate temperature above and below whichmortality increases)

Counts of positive flu A, flu B and RSV specimens fitted aslinear terms with lags up to three weeks and interaction termswith year

Week number fitted as cubic spline with 9 d.f. (unexplainedseasonality)

22 / 26

Results - model fit

●

●

●

●

●

●●●

●

●●

●●

●

●

●

●●●●●●●

●●●●●

●

●●

●●

●●●●●

●

●●

●●●●

●●

●

●

●

●

●

●

●

●

●

●

●●

●●

●

●●●●

●●

●●●●●●●●

●●●●●●

●●●●●●●

●●●

●

●●

●●

●●●●●

●

●

●

●●●

●●

●

●

●●

●

●●

●●●●●●●

●

●●●●

●●

●

●

●

●●

●●

●●●

●●●●●●●

●●

●●●●

●

●

●

●

●

●

●

●

●

●●●●

●

●

●●●

●●

●●●●

●

●●●●●●●

●

●

●

●●

●●●

●●●

●

●

●●●●

●●

●

●

●

●●●

●

●●●

●●

●

●●●●●●●

●

●●●●

●●●●●

●●

●●

●

●

●

●

●●

●●

●

●●

●

●●●

●

●●

●●

●●●

●

●●

●

●●

●●●●

●●

●●●●●●●

●

●●

●

●●●●

●●

●

●●●

●

●●

●●

●

●●●

●●

●●

●●

●

●

●

●

●

●●

●●●●●●

●

●

●●●

●●

●●●●

●●

●●

●●

●

●

●

●

●●●●●●●●

●●●

●

●●

●

●

●●

●

●

●●●

●●●

●●●●●●●

●

●●

●

●●●●●●

●

●

●

●

●

●

●

●

●●●●●●●●

●

●

●●

●●●●●

●●●

●

●

●●●●●

●●

●

●

●

●●●●●

●

●●●●●●●●●

●

●●

●

●

●

●●●●●●

●●●●●

●

●●

●

●

●●●

●

●

●

●

●

●●●●

●●

●●●

●●●

●●

●

●●●●

●●●●●●●

●●

●●●●●

●●●

●●●●

●

●●●

●

●

●

●

Week

Nu

mb

er

of

de

ath

s

80

00

10

00

01

20

00

14

00

01

60

00

18

00

0

1−1999 1−2000 1−2001 1−2002 1−2003 1−2004 1−2005 1−2006 1−2007 1−2008

● Observed number of deathsModel prediction

Observed and predicted weekly mortality 1999−2008

23 / 26

Deaths due to influenza/extreme temperature 2000

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51

Week

Nu

mb

er

of

de

ath

s

05

00

01

00

00

15

00

02000

UnexplainedInfluenza A+BExtreme temperature

Observed deaths

24 / 26

Comments

Work in progress!

Useful in attributing deaths at end of season, not real-time

Currently no mortality data available between December 2008and April 2009

Not age-specific

25 / 26

Selected references

Farrington CP, Andrews NJ, Beale AD, Catchpole MA: Astatistical algorithm for the early detection of outbreaks ofinfectious disease. J Roy Stat Soc Stat Soc. 1996. 159.547–563

Rocklov, J. & Forsberg, B. The effect of temperature onmortality in Stockholm 1998–2003: a study of lag structuresand heatwave effects. Scand J Public Health. 2008. 36.516–523

RE. Serfling, Methods for current statistical analysis of excesspneumonia-influenza deaths. Publ Hlth Rep. 1963. 78494–506

Pitman, R. J. et al. Assessing the burden of influenza andother respiratory infections in England and Wales. J. Infect.2007. 54. 530–538

26 / 26