UAB Pulmonary board review study design and statistical principles

Pulmonary Board Review:Study Design and Statistical Principles

Terry Shaneyfelt, MD, MPHUAB Division of General Internal Medicine

@EBMTeacher

EBMTeacher.com

UABEBMcourseYouTube logo image credit (CC0): http://commons.wikimedia.org/wiki/File:Youtube.svg

Topics to be covered Diagnostic testing

Choosing a test and interpreting studies

Randomized controlled trialsRandomization, power, types 1 and 2 errors,

outcome measures

Observational studiesStudy design, RR, OR

ScreeningOutcomes and biases

Diagnostic Testing

Diagnostic Testing Choosing a test

Sensitivity, specificity, likelihood ratiosSpPin and SnNout

Interpreting the results of a diagnostic test studyPositive and negative predictive values

What is the role of testing?

35yo F with 2% pretest probability

48yo M with 27% pretest probability

75yo M with 99% pretest probability

Rule in with high pretest probabilityRule out with low pretest probability

Which test would you choose to rule in disease?

Sensitivity Specificity LR + LR -

Test A 95% 80% 4.75 0.06

Test B 90% 90% 9 0.11

Test C 70% 95% 14 0.32

Which test would you choose to rule out disease?

Rule in with specific testsRule out with sensitive tests

SpPin ( A specific test, if positive, rules in disease in a high risk person)

SnNout (A sensitive test, if negative, rules out disease in a low risk person)

Alternatively: Choose test with highest positive LR (to rule in) and/or lowest negative LR (to rule out)

Sensitivity Specificity LR + LR -

Test A 95% 80% 4.75 0.06

Test B 90% 90% 9 0.11

Test C 70% 95% 14 0.32

To learn more watch my other videos

Sensitivity: http://bit.ly/1FOlqry

Specificity: http://bit.ly/1IYjv2A

LR: http://bit.ly/1JOofZz

CTA High Pretest Probability

IntermediateLow Pretest Probability

Positive Predictive Value

(%)96 92 58

Negative Predictive Value

(%)60 89 96

Adapted from Table 5 PIOPED II (NEJM 2006;354:2317)

In patients with high pretest probability what is the chance that the patient has a PE with a negative

CTA?

Sens 83%Spec 96%

Predictive value varies with prevalence

To learn more watch my other videos

PPV: http://bit.ly/1HDXWpm

NPV: http://bit.ly/1Faegbq

Randomized ControlledTrials

RCTs Study design

Randomization

Power, type 2 error, and sample size

p-values and type I error

Outcome measuresRRR, ARR, HR, NNT

Image from PrevMedFellow (CC A SA license): http://commons.wikimedia.org/wiki/File:Flowchart_of_Phases_of_Parallel_Randomized_Trial_-_Modified_from_CONSORT_2010.png

Control

What do you think is the greatest risk of bias in a therapy study?

A. Failure to randomize

B. Failure to conceal allocation

C. Failure to blind participants and study personnel

D. Failure to use intention to treat analysis

E. Failure to treat groups equally except for the intervention

2 Reasons:

1. Reduces selection bias

2. Equally distributes prognostic factors (both known and unknown)

Why Is Randomization So Important?

The validity of a clinical trial depends on treated & control patients being prognostically equal, other than

the intervention being tested

TRUTHDifference No difference

Study Conclusion

Difference

No difference

Beta/ Type II error

Alpha/ Type I error

We estimated that with enrollment of 1130 subjects, the study would have 90% power to show a significant difference between the two groups in the time to the first acute exacerbation of COPD, assuming that 50% of the participants in the control group and 40% in the azithromycin group would have an acute exacerbation, that the rate of nonadherence would be 20%, and that 6% of participants would die or be lost to follow-up during the study, with a two-sided type I error of 0.05.

Azithromycin for Prevention of Exacerbations of COPD NEJM 2011;365:689

Power

Power (greater the desired power the greater the sample size)

Estimated difference between groups (smaller the difference the greater the sample size)

Type 1 error rate (usually 0.05 but the smaller the greater the sample size)

Variability in the measurements made within each comparison group (greater the variability the greater the sample size)

Sample size is affected by…

Time to event data is often displayed as a Kaplan-Meier curve

From N Engl J Med 2011; 365:689-698

Azithromycin(# of events)

Placebo(# of events)

Hazard Ratio(95% CI)

p-value

Hospitalization related to COPD

156 2000.82

(0.64 - 1.07)0.15

ED or urgent care visit

199 2570.81

(0.63 - 1.04)0.09

Unscheduled office visit

1202 13450.85

(0.74 - 0.98)0.02

Adapted from Table 2 from NEJM 2011; 365:689

What does a HR of 0.82 mean?

What does the confidence interval mean?

Where do p-values come from?

Statistical Approach to Compare 2 Groups

Calculate:1. Main effect2. Variance in main

effect

State a null hypothesis (the main effect is 0)

Calculate the test statistic to

determine p value

Calculate the 95% confidence interval around the main

effect

New Drug

Placebo

What does the 95% CI 0.64-1.07 imply?

Statistical Tests Mathematical formulas that produce test

statistics to assess the likelihood that chance (or sampling error) accounts for the results observed in the study

Many different tests. Choice depends on several factors:

Type of data (continuous, dichotomous, etc) Distribution of data (normally distributed or not) Study design (# of groups, etc)

TRUTHDifference No difference

Study Conclusion

Difference

No difference

Beta/ Type II error

Alpha/ Type I error

2 errors can be made with hypothesis testing

P-value

Probability that the results seen (or one more extreme) could have occurred by chance alone

○ Assuming that there is in fact no difference between groups (null hypothesis)

Cannot tell you if there is bias in a study

Does not indicate clinical significance

To learn more watch my other videos

NNT: http://bit.ly/1F4xONy

RRR: http://bit.ly/1Fyef3F

Observational Studies

Observational Studies

Study design

Measures of effectOR and RR

Overview of Study Designs

Used to determineprognosis & harm

Outcome Measure: RR

Population

Exposed

Outcome

No outcome

Unexposed

Outcome

No outcome

Time

Cohort Study

Establish incidence (risk) directly

Multiple outcomes

Study of rare exposures

Strengths Weaknesses

Not good for rare diseases

Not good for diseases that take a long time to develop

Can’t study multiple exposures

Cohort: Strengths & Weaknesses

The incidence of pulmonary embolism in the COPD cohort was 1.37 per 10,000 persons/year and in the non-COPD cohort was 0.35 per 10,000 persons/year.

Multiple ways to express riskIncidenceRisk difference (attributable risk)Relative risk (risk ratio)

Interpreting RRRR = 1 (no association)RR > 1 (increased risk of disease)RR < 1 (decreased risk of disease)

http://bit.ly/1dtFFhV

Outcome Measure: OR

Exposed

Diseased

(Cases)

Non-Diseased

(Controls)

Non- Exposed

Exposed

Non-Exposed

Time

Case-Control Study

Good for diseases with long latency

Good for rare diseases

Can determine multiple exposures

Faster results

Strengths Weaknesses

Can’t establish estimate of risk directly nor determine prevalence

Can only study one disease

More prone to bias

Case-Control: Strengths & Weaknesses

Low-dose glucocorticoid use (prednisolone daily dose equivalent 5 mg) carried a twofold increased risk of PE (OR, 1.8; 95% CI, 1.3-2.4), whereas a 10-fold increased risk was observed for the highest dose of glucocorticoids (prednisolone 30 mg) (OR, 9.6; 95% CI, 4.3-20.5). The authors are incorrect in the statements of risk. Do you know why?

Can only determine relative frequency of exposure among cases and controlsOdds ratio

Interpreting OROR = 1 (no difference of exposure)OR > 1 (frequency of exposure higher among cases)OR < 1 (frequency of exposure lower among cases)

http://bit.ly/1HHm2Nd

Screening

Screening

Prevalence vs incidence screens

Outcomes of screening studies

BiasesLead timeLength timeOverdiagnosis

Identification of disease or a risk factor in asymptomatic individuals

Screening

Biologic Onset Outcomes

Clinical diagnosis

Screen detection

Fundamental Principles of Screening-1

3 prerequisites:▪ Disease must have a great enough burden of

suffering

▪ Screening test can identify disease earlier than usual

▪ Earlier therapy leads to better outcomes

Fundamental Principles of Screening-2

Target disorders are relatively rare (low prevalence)

Must screen large numbers of people Most positive tests are false positives

Risks of screening tend to be rare but apply to all

Benefits accrue only to a few

Disease Prevalence is Low LOW low predictive value

Sensitivity=95%Specificity=95%

LR=19

Prevalence Predictive value

10% 67%

1% 16%

0.1% 1.8%Watch Predictive Value Estimates From Studies Can Be Misleading

http://youtu.be/3zq82uiGS3o

When choosing a test for a screening program you want the test to be….?

1. Highly sensitive

2. Highly specific

Do you find more cases of disease on the first round of screening or subsequent rounds?

Dx

Dx

Dx

Dx

Dx

Dx

Dx

Dx

Dx

Dx

Dx

Dx

1 2 3Round of screening

Number of cases newly detected

5 3 2

What is the appropriate outcome of a screening study?

A. Survival?

B. Mortality?

C. Disease detected?

Solving lead time bias problem: compare age-specific mortality between screened and unscreened. Not survival! Count from date of randomization

Adapted from Clinical Epidemiology The Essentials 3rd edition

Dx †

Dx †

Unscreened

Screened but early Rx ineffective

Ons

et

Early

Usu

al

Lead Time Bias

From cancer.gov

Compare outcomes via RCT with a control group and a group offered screening

Count all outcomes regardless of method of detection

Avoiding Length Time Bias

Is PSA screening causing prostate cancer ???

From cancer.org

PSA approved

Thank You