Misuse Confidence

7/28/2019 Misuse Confidence

1/33

Perinatal Epidemiology Workshop

Russell S. Kirby, PhD, MS, FACE

Universi ty of Wisco nsin Medical School

As of July 2002:

Department of Maternal and Child HealthSchool o f Publ ic Health

Univers i ty of A labama-Birm ingham


2/33

Objectives

Focus on components of infant mortality

Review the theoretical basis for conventionally usedmeasures (neonatal, postneonatal)

Evaluate the utility of these measures with population-based data

Consider alternative methodologies foroperationalizing these measures

Propose study designs to test alternative measures inrelation to conventional measures


3/33

TOP TEN LIST

TEN BEST WAYS TO MISUSECONFIDENCE INTERVALS

With apologies to David Letterman, and thanks for editorial assistance to

Elizabeth Kirby and for their insights to the following Internet contributors:

Patrick Remington, University of Wisconsin Medical School

Robert Meyer, N.C. Birth Defects Monitoring Program, State Center for Health

Statistics

N.C. Division of Public Health

Richard Miller, Wisconsin Bureau of Health Information

Russel Rickard, Colorado Responds to Children with Special Health Needs

Kim Hauser, University of South Florida

R.S. Kirby, March 2002


4/33

"If the confidence interval is verytight, the case for causation isstrengthened...."

Submitted without attribution as a quotation from a

manuscript under review by

Patrick Remington

University of Wisconsin Medical School

February 24, 2002


5/33

Being a statistician means never having to

say you're wrong. Don't be afraid to use

those 100% CIs so that you can state with

authority:

"I'm 100% sure that the true populationparameter lies between zero and infinity!"

Number 10

Top Ten List: Ten Best Ways to Misuse Confidence Intervals

Say It With Total Confidence


6/33

Whether you calculated the CIs or not, it isnt

necessary or desirable to include them in your

publication.

A general statement in the text to the effect that

all statistical values were significant should

be sufficient for all but lay audiences.

Number 9

Out of Sight, Out of Mind



7/33

Assess statistical significance through visual

comparison of confidence intervals.

For example: which of the following confidence

intervals is larger?

OR >-------


8/33

Number 7

Smoke and Mirrors


On a graph, overly wide confidence bands can be

adjusted by plotting the data on an arithmetic

scale and the CIs on a logarithmic scale.

Your friends will be amazed at how tight those CIs

become.


9/33

Remember that any results would have been

statistically significant if only your sample size

was large enough. This should not ruin otherwise

good science.

To compensate for small sample sizes, adjust your

CIs so that the null value is always

excluded. Don't be afraid to use 15% CIs if

necessary, if that will help support your

hypothesis.

Number 6

How Big is Your N?



10/33

Never include your point estimate within the

confidence limits.

When questioned, blame it on the computer

program.

Number 5

To Really Foul Things Up . . .



11/33

Number 4

Will 99 and a Half Do?


95% CIs are trite and commonplace. Be

creative Try reporting results such as the2log(67.45%) CIs and see what exciting results

you get.

No one will understand it, but they dontunderstand the 95 or 99% CIs either.


12/33

Selectively quote the results concerning one

confidence limit. For example, when the estimate

of the confidence interval is 0 to 70, perhaps for

the association between watching late night

reruns and sleep disorders:

This risk factor decreases the risk of the

outcome up to 70%.

Number 3

Sound Bytes Are Best



13/33

To demonstrate statistical significance for a

comparison of rates or proportions, visually examine

the two confidence intervals.

Try comparing the lower confidence bound of the

smaller value with the upper confidence bound of the

larger value. If that still doesnt work, try fitting ever

narrowing confidence bands (e.g. the 15% CIs as inNumber 6 above) until the confidence limits no longer

overlap.

Number 2

Why Infer When You Can Guess?



14/33

The best way to do misuse confidence

intervals is to confuse statistical significancewith true, substantive significance.

An observed statistical difference begs the

question So what? which is too seldomasked.

Number 1

Is That All, Folks?



15/33

If you pay strict attention to these

suggested methods for using confidence

intervals, you too can misuse confidence

intervals with confidence.

-- R. S. Kirby

March 25, 2002



16/33

Brief Summary for Those Who Are Knitting, Doin

Crossword Puzzles, or Discerning the Geometric

Pattern in the Carpeting

Too often in public health, we use measures and present data in

familiar forms because thats the way weve always done it.

The concepts underlying measures of neonatal and postneonatal

mortality have their origins in sociological theory from the period

1880-1940. Periodically our indicators and outcome measures need to be

reviewed for contemporary validity (especially construct validity).

Some alternatives (for infant mortality classifications) include more

careful consideration of underlying causes of death, more

sophisticated record linkages, and developmental measures.

This is fertile ground for innovative, multidisciplinary research.


17/33

Resource

Kirby, Russell S., "Neonatal and

Postneonatal Mortality: Useful Concepts or

Outdated Constructs?", Journal ofPerinatology, 13,6 (November-December

1993), 433-441.


18/33

Historical Context

Neonatal and postneonatal mortality developed as

as demographic measures

Proxies for general categories of cause of infant

death

Endogenous causes (perinatal, congenital and

immediate newborn period)

Exogenous causes (nonperinatal infection, injury,homicide, other external causes)


19/33

Conventional Wisdom

Deaths occurring early in the first year of life aremore likely associated with endogenous causes

Deaths after the early part of infancy are more

likely associated with endogenous causes (socio-economic or standard of living)

Since the 1940s, we have continued to defineneonatal mortality as deaths in the first 27 days of

life, and postneonatal mortality as deaths in thebalance of the first year of life


20/33

Whats Happened Since 1940

Secular decline in infant mortality rates

Since the 1970s, a more rapid decline in neonatalmortality rates

Prior to about 1985, much of this decline resulted inpostponement of death into the postneonatal period

More recently, technological advances and improvements inneonatal resuscitation and NICU management have led to greaterrelative declines in neonatal mortality

In the 1990s, some health promotion programs have beensuccessful in targeting SIDS, a condition that occurs primarily inthe postneonatal period


21/33

Neonatal and Postneonatal Mortality Rates

United States, 1960 - 1999

26.024.7

20.0

16.1

12.610.6

9.2 7.6 7.1

0

5

10

15

20

25

30

1960 1965 1970 1975 1980 1985 1990 1995 1999

Infant Neonatal Postneonatal

Rate per 1,000 live births

Source: National Center for Health Statistics, final mortality data

Prepared by March of Dimes Perinatal Data Center, 2002


22/33

Methods for the Useful Constructs Study

Information on deaths from linked birth and infantdeath certificates

Underlying causes of death classified into

endogenous and exogenous causes Mortality rates compared within birth weight

strata between endogenous and neonatal andbetween exogenous and postneonatal mortality

rates


23/33

The Bottom Line

At least in the 1980s, neonatal and endogenous mortalitymeasures are fairly synonymous

Postneonatal mortality is no longer a useful proxy forexogenous infant mortality

This renders the conventional classification suspect forcurrent purposes

New on the scene are advances in perinatal-infant medicineand nursing management, the new ICD-10 classification fordiseases, and more readily available data resources throughrecord linkage, chart review, and other data acquisitionmethods


24/33

Future Studies?


25/33

Three Alternative Formulations:

One

Use conventional definitions, but calculate

neonatal and postneonatal mortality rates bases on

corrected gestational age

Thus, a term neonate who dies during the first monthafter birth is a neonatal death

A 32 week infant who dies prior to the 12th week of

life (40-32=8 + 4 weeks of conventional neonatal

period) is considered a neonatal death


26/33


Two

Include fetal deaths in the analysis, regarding all

of these events as non-postneonatal.

Conduct the analysis conventionally, or use the

corrected gestational age definition


27/33


Three

Incorporate linked birth-infant death certificate and infant

hospital discharge records into the analysis

Define neonatal and postneonatal in relation to initial

continuous hospital stay. Example: a baby born at one hospital who is transferred to an

NICU at another hospital and is discharged at 9 weeks has a

neonatal period lasting 63 days.

Another example: a healthy newborn is discharged on the

second postpartum day, and has a neonatal period lasting 2 days


28/33


Three (continued)

Use the person-days method for calculating rates.

Thus, the neonatal mortality rate is the number of

neonatal deaths according to this definition,

divided by the total number of initial newbornstay days. The postneonatal mortality rate is the

number of non-neonatal infant deaths, divided by

the balance of lived postneonatal days.


29/33

However beautiful the strategy, you

should occasionally look at the

results.

-- Sir Winston Churchill


30/33

Critique of Alternatives


31/33


Corrected gestational age will only work when

gestation is measured accurately. That is a topic

for at least another workshop if not an entire

MCH epidemiology conferencebut allowing forsome error at early gestational ages its worth a try

All of the problems with birth certificate and

infant death certificate data quality are at least an

order of magnitude worse with fetal death records


32/33


Hospital discharge databases are becoming more widelyavailable. However, in most states these files are not linked inthe manner necessary to support the proposed analysis.

Most public health professionals and demographers are

unfamiliar with rates expressed as person-time measures

A validated measure for neonatal and postneonatal mortalitycould serve as the basis for more careful multivariate analyses,

but first its validity and reliability must be independently

verified For the time being, it is likely we are stuck with the same old

same old.


33/33

Questions?

Im interested in hearing from you!

I can be reached by e-mail at

[email protected]

414-219-5610, FAX 414-219-5201

Ill provide forwarding information atthese locations after July 2002
mailto:[email protected]:[email protected]:[email protected]:[email protected]

Date post:	03-Apr-2018
Category:	Documents
Upload:	minervastanciu
View:	224 times
Download:	0 times

Misuse Confidence

Documents