PhD Course in Biostatisticsuser.medunigraz.at/andrea.berghold/phd/PHD_Course... · 2013. 5. 1. · Berghold, IMI, MUG PhD Course in Biostatistics Univ.-Prof. DI Dr. Andrea Berghold

Berghold, IMI, MUG

PhD Course inBiostatistics

Univ.-Prof. DI Dr. Andrea Berghold

Institute for Medical Informatics, Statisticsand Documentation

Medical University of Graz

[email protected]

Berghold, IMI, MUG

Content

• Introduction to Medical Statistics

• Study designs in medical research

• Exploring and summarizing data

• Populations and samples

• Statements of probability and confidence intervals

• Drawing inferences from data - Hypothesis testing

• Estimating and comparing means

• Proportions and chi-square tests

• Correlation and regression

• Diagnostic tools

• Methods for analysing survival data

Berghold, IMI, MUG

Literature

• Martin Bland: An Introduction to Medical Statistics. 3rd ed. Oxford University Press, 2000.

• Douglas Altman: Practical Statistics for Medical Research. Chapman & Hall.

• Aviva Petrie and Caroline Sabin: Medical Statistics at a Glance. Blackwell Science, 2000

• …

Berghold, IMI, MUG

NEJM June 2001: Methods Section of Full-Length Original Articles (by article, in column inches)

Statistical Methods - medical literature

Statistical Methods All methods Percentage

4.6 35.7 12.9 %

7.9 53.6 14.7 %

12.2 51.6 23.6 %

7.3 36.8 19.8 %

32.0 177.7 18.0 %

Berghold, IMI, MUG

In the same issue the following statistical methods were mentioned:

Statistical Methods - medical literature

Bonferroni method

Chi-square test for independence

Chi-square test for goodness-of-fit

Confidence intervals

Cox proportional hazards models

Cumulative mortality

Fisher's exact test

Intention-to-treat analysis

Interim analysis

Kaplan-Meier survival curves

Logistic regression

Logrank test

Mantel-Haenszel adjusted relative risks

Noninferiority testing

Odds ratio

Power Analysis

P-values

Randomization

Relative risk reduction

Repeated measures ANOVA

Sample size estimation

Spearman correlation

t-tests

Wilcoxon test

Berghold, IMI, MUG

Is it worth to struggle with statistics?

Bad statistics leads to bad research,

and bad research is unethical

Altman (1982)

Statistics

Berghold, IMI, MUG

• Design of studies- How do I get adequate data?

• Data analysis using statistical methods- What do I do with the data?

• Critical appraisal- How do I interpret study results?

Biostatistics - Medical Statistics

Berghold, IMI, MUG

Study

interpret

analyse data

collect data

plan study

Berghold, IMI, MUG

1. Stating the problem

• Major objective of the study -determine relevant variables und factors

• Search the literature, discussion with experts

Study

2. Designing the study

• Study design, sample size calculation etc.

• Statistical analysis plan

• Study protocol

Berghold, IMI, MUG

A Study

3. Collecting data

• Collecting data and plausibility checks

4. Data analysis

• Graphs and summary statistics

• Statistical inference

5. Interpretation of results and conclusions

• Discussion of new information

Berghold, IMI, MUG

Some questions which should be answered in advance:

Stating the problem

• What is the major objective of the study?

• Is the question clearly defined?

• Is it also relevant?

Berghold, IMI, MUG

1. Are there differences in the one-year rate of restenosis usingstents or PTA of stenosis of arteria iliaca?

2. Does a betablocker decrease all-cause mortality in patientswith chronic heart failure?

3. Have cancer patients who have anemia a worse prognosisthan patients without anemia?

4. Which method should be used for training of laparascopicsurgery?

5. …

Examples

Berghold, IMI, MUG

• Primary variable, endpoint1. rate of restenosis;2. all-cause mortality;3. 5 year disease-specific survival;4. number of stitches per minute; …

• Factors1. none2. stage (NYHA class); 3. anemia, size of tumour, lymph nodes;4. method, playing an instrument; ...

• Other factorsAge, sex, smoking ....

Variables

Berghold, IMI, MUG

• Random error

• inter- and intraindividual variability

• Systematic error - Bias

• Selection bias

• Assessment bias

• Information bias

• …

Try to avoid bias and reduce random error!

Errors

Berghold, IMI, MUG

Types of studies

• Observational studies

• Experimental studies

Berghold, IMI, MUG

Types of studies

Main types of studiesin medical research

Observational studies Experimental studies

Cross-sectionalstudiy

case-controlstudy

cohortstudy Clinical trial Laboratory

experiments

Berghold, IMI, MUG

Observational Studies

Berghold, IMI, MUG

Cross Sectional Study

Populationsubjects

selected forstudy

with outcome

without outcome

Onset of study Time

no direction of inquiry

Berghold, IMI, MUG

Example

disease – asthma

Exposure

totalboys girls

yes 344= a

221= b

565= (a+b)

no 4885= c

4787= d

9672= (c+d)

total 5229= (a+c)

5008= (b+d) 10237

prevalence 344 / 5229= 0,066

221 / 5008= 0,044

565 / 10237= 0,055

OR = = = 1.53a / b 344 / 221c / d 4885 / 4787

Berghold, IMI, MUG

Case-Control study

cases

controls

exposed

unexposed

exposed

unexposed

Onset of studyTime

Direction of inquiry

Berghold, IMI, MUG

Example

Exposure(during

childhood)

Disease- Melanoma

totalcases controls

no sun protection

303= a

290= b

593

sun protection 99= c

132= d

231

total 402 422 824

OR = = = 1.39

95% confidence interval: [1.02; 1.89]

a / c 303 / 99b / d 290 / 132

Berghold, IMI, MUG

Odds

The Odds of a probability P is defined by

It is the chance, that an event happens.

Example:

P = 0.5 : an event will happen with a probability of 50%

Odds(P) = 0.5/0.5 = 1 (chance of 1:1)

P = 0.8

Odds(P) = 0.8/0.2 = 4 (chance of 4:1)

Odds (P) = P1-P

Berghold, IMI, MUG

Odds Ratio

ExposureDisease

yes(cases)

no(controls)

exposed a b

not exposed c d

OR = =a / c adb / d bc

OR =Chance, that case was exposed

Chance, that control was exposed

Berghold, IMI, MUG

Cohort Study

PopulationCohort

selected forstudy

exposed(subjects)

unexposed(controls)

with outcome

without outcome

with outcome

without outcome

Onset of study Time

Direction of inquiry

Berghold, IMI, MUG

Examples of cohort studies

Onset of study Exposure

Prognostic study time of diagnosisor

start of therapyprognostic factors

e.g. influence of anemia on survival

Epidemiological study"Start" of

observation risk factorse.g. Framingham study

Berghold, IMI, MUG

Example

Association between cigarette smoking and incidence of stroke in a cohort of 118 539 women (age 30-55) – follow-up 8 years

Exposure No. of cases of stroke Person-years Incidence

(per 100 000 person-years)

Smoker 139 280141 49.6

Ex-smoker 65 232712 27.9

Never-smoked 70 395594 17.7

RR = = 2.8

95% confidence interval RR: [2.1; 3.7]

139 / 28014170 / 395594

Berghold, IMI, MUG

Relative Risk

ExposureDisease

totalyes no

exposed a b a+b

not exposed c d c+d

RR =a / (a+b)c / (c+d)

RR =Incidence rate of exposedIncidence rate of not-exposed

Berghold, IMI, MUG

Experimental Studies

Berghold, IMI, MUG

Comparison of the efficacy of different drugs, therapies, vaccinesetc. after controlling for confounders (e.g. age, sex, stage ofdisease, …).

Clinical trial

Aim:

Observed differences in success rates between treatment groups can exclusively be put down to the fact that differences are caused by the efficacy of the

different treatments.

Berghold, IMI, MUG

Statistical issues

• The efficacy and safety of treatments have to be judged against a background of biological variability

• In designing studies, two main points have to be kept in mind:

• the effect of bias

• the effect of chance

Berghold, IMI, MUG

Focus

• Comparative trials:

• Interested in treatment effect and treatment comparisons

• Concurrent control group• Investigate a new experimental intervention versus placebo or a

“standard” intervention• compare two alternative commonly-used interventions with each

other• Study the result of adding an additional agent to a standard regimen• Compare different doses or intensities of an intervention

• Pre-defined study objective

Berghold, IMI, MUG

Design techniques to avoid bias

• Randomization

• Blinding

„The most important design techniques for avoiding bias in clinical trials are blinding and randomisation.“ (ICH E9: Statistical Principles in Clinical Trials)

Berghold, IMI, MUG

Randomization

• To allocate treatments to subjects in a trial at random (using coins, dice, random number tables or generators)

• Allocation concealment

• Neither the subject nor the investigator knows ahead of time what treatment the subject will receive

• Benefits:

• Eliminates assignment basis – avoids selection bias

• Tends to produce comparable groups

• Statistical basis for a valid treatment comparison

Berghold, IMI, MUG

20 patients will be allocated at random to two groupsPatients:1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20We throw the die once for each patient:odd number: group Aeven nuumber: group B

Group A: Group B:

Result: ?2

1

?5

2

?3

, 3

?

, 4, 5, 6 , 7, 8, 910

,, 11

, 1213

,, 14, 15, 16, 17

18,, 19

, 20

Randomization

Berghold, IMI, MUG

A B

1 32 54 678910

Berghold, IMI, MUG

Restricted randomization

• Disadvantages of simple randomization:

• No guarantee of equal or approximately equal sample size in each treatment group at any stage of the trial

• With n = 20 on two treatments A and B, the probability of a 12:8 split or worse is approximately 0.19

• No protection against long runs of one treatment• Subject characteristics may change over time

• Restricted randomization:

Permuted blocks (Matts & Lachin)Biased coin (Efron)Urn design (Wei)Big Stick (Soares & Wu)…

Berghold, IMI, MUG

Pat. allocation therapy

1 A Radiatio 2 A Radiatio3 B Rad.+ Chem.4 B Rad.+ Chem.

5 A Radiatio6 B Rad.+ Chem.7 A Radiatio8 B Rad.+ Chem.

9 B Rad.+ Chem.10 A Radiatio11 A Radiatio12 B Rad.+ Chem.

13 B Rad.+ Chem.14 B Rad.+ Chem.15 A Radiatio16 A Radiatio.... .... ......

randomization listblock randomization:1: AABB2: ABAB3: ABBA4: BABA5: BAAB6: BBAA

n! 4!n1! n2! 2! 2!

= = 6

Randomization list(only at study coordinating centreand not for the researcher)

Randomization

Berghold, IMI, MUG

Stratified randomization

• Balance treatment groups with respect to prognostic factors

• For large studies, randomization “tends” to give balance

• For smaller studies a better guarantee may be needed

• Common factors used for stratification - e.g. clinical centre, age, sex, disease severity

• Define strata – e.g. Age: < 40, 40-60, > 60;Sex: M, F (3 x 2 strata)

• Randomization is performed within each stratum and is usually blocked

• Rule of thumb – use as few stratification factors as possible

Berghold, IMI, MUG

Randomized controlled trials

The trial carried out by the Medical Research Council (MRC, 1948) to test the efficacy of streptomycin for the treatment of pulmonary tuberculosis is generally considered to be the first randomized experiment in medicine.

target population: patients with progressive bilateral pulmonary tuberculosis (bacterially proven), aged 15-30 years

107 patients in 3 centers were allocated by a series of random numbers drawn up for each sex at each centre.

Berghold, IMI, MUG

Implementation

• Sequenced sealed envelopes

• Phone call / fax to trial coordination centre

• Interactive Voice Response Systems

• Internet-based Systems (e.g. Randomizer for Clinical Trials)

Berghold, IMI, MUG

Randomizer for Clinical Trials

Berghold, IMI, MUG

Online Randomization

www.randomizer.at

Berghold, IMI, MUG

Blinding - Masking

• To limit the occurrence of bias in the conduct and interpretation of the trial (in the care, the assessment of endpoints, the attitude of subjects to treatments etc.)

• Double-blind: neither subject nor investigator/staff are aware of the treatment received

• placebo, “double dummy”, masked vials

• blinding may not be possible• surgical versus medical intervention• one intervention has obvious side-effect

• Outcome assessed by masked observer

• Single-blind

• Open-label trial

Berghold, IMI, MUG

Randomized controlled trials

• Choice of target population

Selection of patients: Definition of target population using inclusion and exclusion criteria

• Trial Design

• Parallel – Design

• Cross-Over – Design

Berghold, IMI, MUG

Parallel - Design

Elig

ible

an

d w

illig

ing

subj

ects

Con

trol

Ran

dom

izat

ion

Ass

essm

ent

Test

Scr

eeni

ng

Pop

ulat

ion

Berghold, IMI, MUG

Cross – Over - Design

Ran

dom

izat

ion

Ass

essm

ent

Pop

ulat

ion

Scr

eeni

ng

Ass

essm

net

Con

trol

Con

trol

Elig

ible

an

d w

illiin

g su

bjec

ts

Test

Test

Berghold, IMI, MUG

• The statistical analysis has to be defined before the study is carried out

• Statistical analysis plan (SAP)

• Population used for analysis:

• All-Randomized patients – Intention-to-treat analysis

• On-treatment patients – Per-protocol analysis

• Safety population

Statistical analysis

Berghold, IMI, MUG

Intention-to-Treat

all randomized patients must be included in the analysis -

they have to be included in the group they were randomised to, independent of what happened after randomization.

Berghold, IMI, MUG

Intention-to-Treat (ITT) Analysis

Randomization

Treatment A Treatment B

Treatment Aper protocol

Treatmentwithdrawal

Treatment Bper protocol

Treatmentwithdrawal

Intention-to-Treat: 1+2 vs 3+4Per-Protocol (PP): 1 vs 3

1 2 3 4

Berghold, IMI, MUG

Illustration

Propanolol Atenolol Placebo

ITT – Analysis 7.6% 8.7% 11.6%

PP - Analysis 3.4% 2.6% 11.2%

Withdrawal 15.9% 17.6% 12.5%

Percentage of patients who died within 6 weeks after heart infarction (Wilcox et. al.)

Berghold, IMI, MUG

Efficacy and Effectiveness

Efficacyeffect under optimal conditions

All patients are included in the analysis, who were treated per protocol.

Per-Protocol Analysis

Effectivenesseffect under „real“ conditions.

All patients are included in the analysis, who were included in the study (Withdrawal, changing treatment etc.).

Intention-to-treat Analysis

Berghold, IMI, MUG

Main points RCTs

• Randomization – concealed allocation

• Blinding – double blind study

• Minimal loss in follow up

• Intention to treat Analyse

• Carry out specified analysis

Berghold, IMI, MUG

Laboratory Experiments

Exactly the same principles apply to laboratory experiments on

animals or on biological specimens as for clinical trials

• Stricter control of extraneous factors is possible

• Effect of uncertainties is minimized – use of control group,

randomization, replication

• Principles of randomization is often not well understood

• Using genetically similar animals – little biological variablity