Summary Module 1 - IPPCR Video and Handout …...Summary Module 1 Laura Lee Johnson Fall 2014 IPPCR...

Summary Module 1

Laura Lee JohnsonFall 2014

[email protected]

Disclaimer• This presentation reflects the views of the

author(s) and should not be construed to represent FDA’s views or policies

OutlineReview• Odds Ratio and Relative Risk• 2 Graphs• Interactions (Effect Modification) and

Confounding• BMJ hypothesis testing examples

Final Exam and CertificateRegistered 2014-2015 IPPCR Participants

• 75% or higher on the final exam • Posted online after last lecture of the course• Multiple choice format• 50 to 60 questions• Open book / notes• Roughly 3 weeks to electronically submit

answers• Exam is instantly graded• Print and save your certificate of completion• You can only take the final exam one time

Review• Videos should be fixed

– If not, follow emailed advice from Daniel McAnally

• Slides• Chapters (if have the book)• ?Study groups?

Module 1• Your question comes first

– It needs to matter• How do you decide which participants to

study?• How generalizable should the sample be?• Balancing internal vs external validity• Feasibility and implementation• Outcomes of interest, how measuring

Measurement• Measurement of constructs

– Validity, reliability, sensitivity to change, scale, feasibility

• Questionnaire development– Self report– Data collection methods– Mixed (qualitative and quantitative)

methods• Problems with measures• Problems with measurement

Take Home: What you need for N

• What difference is scientifically important in units– 0.01 inches?– 10 mm Hg in systolic blood pressure?

• How variable are the measurements (accuracy)? – Pilot!– Plastic ruler, Micrometer, Caliper

Define the Survival or Time to Event Outcome Variable

• What is the event? – What is the y-axis?– Mother-Infant HIV transmission; testing

positive; death• Where is the time origin?

– Infant birth; time of diagnosis; time of first treatment

• What is the time scale?– Weeks; months; years; seconds

• How is time at which the event ‘occurs’defined?

Using Potentially Available Data

• Structured and unstructured data– Natural language processing

• Developing questions and resources to answer those questions– US Medicare Data– Similar data sets in other countries and

health care settings

Two Types of Research Studies• Observational

– Goal is to observe and collect data on characteristics of interest without influencing the participant, environment or disease course

• Interventional (or Experimental)– Researcher deliberately influences course of

events and investigates effects of an intervention on a carefully selected population of subjects

– Experimental studies done on human subjects are referred to as clinical trials or clinical studies

Epidemiology• Study of the distribution and determinants

of disease and injury in human, animal, plant, or other populations

– [Human] disease does not occur at random

– [Human] disease has causal and preventive factors that can be identified through systematic investigation of different populations or subgroups of individuals within a population• Hennekens and Buring, 1987

Observational Studies

• May be the only information outside of the laboratory

• Fundamental limitation of observational studies– Distinguish associations– CANNOT inherently determine causation

Randomization and Random Samples

• Main goal of random sampling– Obtain representative sample of the population– Results based on the sample reflect population

as a whole• Main goal of randomization

– Balance known and unknown baseline characteristics across intervention groups

– Both groups are similar in all aspects, except for the assigned intervention

– This allows us to make causal inference (link the results directly to the interventions)

Randomization• Randomization, blinding/masking, intent to

treat (ITT) analyses go hand in hand in helping to mitigate– Selection (in to a study arm) bias– Investigator bias– Participant response bias

• What is randomization• Why, whom/what, and how to randomize• Stratify on site in multi-site studies

Sample Size and Power Impacted by…

• Difference (effect) to be detected (δ)• Variation in the outcome (σ2)• Significance level (α)

– One-tailed vs. two-tailed tests• Equal/unequal arms• Superiority or equivalence or non-inferiority

• Much more (e.g. drop-outs)

Common Survival Models• Kaplan Meier (KM)

– One way to estimate survival or time to event– Nice, simple, can compute by hand – Can add stratification factors– No sensible interpretation for competing risks– Cannot evaluate covariates like Cox model– Kaplan Meier is a workhorse

• Cox models are today’s workhorse– Covariates, competing risks, etc– Test proportional hazards assumption

Truncation and Censoring• Independence is key• Truncation is about entering the study

– Right: Event has occurred (e.g. cancer registry)

– Left: Have the event and fall out of view before they can enter to be counted

• Censoring is about leaving the study– Right: Incomplete follow-up (common)– Left: Observed time > survival time

18

Questions• Why should we not use Correlation? What

problems may arise if we use it in place of Regression.

• I would also like to know which test has more power one-sided or two-sided and if we have a very small sample size e.g in some rare disease can we rely on one-sided test?

An Aside: Correlation

• Range: -1 to 1• Test is correlation is ≠ 0• With N=1000, easy to have highly

significant (p<0.001) correlation = 0.05– Statistically significant that is– No where CLOSE to meaningfully

different from 0• Partial Correlation Coefficient

Do Not Use Correlation.Use Regression

• Some fields: Correlation still popular– Partial regression coefficients

• High correlation is > 0.8 (in absolute value). Maybe 0.7

• Never believe a p-value from a correlation test

• Regression coefficients are more meaningful

More Hypothesis Testing Reminders

Source of Image: Effect Size FAQs by Paul Ellis http://tinyurl.com/pn2dt68Lydia Flynn has a nice video also explaining type I and II errors at

https://www.youtube.com/watch?v=Dsa9ly4OSBk

Unknown Truth and the DataTruth

DataH0 Correct HA Correct

Decide H0

“fail to reject H0”

1- αTrue Negative

βFalse Negative

Decide HA

“reject H0”α

False Positive1- β

True Positiveα = significance level

1- β = power

P-value Interpretation Reminders

• Measure of the strength of evidence in the data that the null is not true

• A random variable whose value lies between 0 and 1

• NOT the probability that the null hypothesis is true.

Take Home Hypothesis Testing

• Turn questions into hypotheses• Failing to reject the null hypothesis DOES

NOT mean that the null is true• Every test has assumptions

– A statistician can check all the assumptions– If the data does not meet the assumptions

there are non-parametric versions of tests (see text)• Non-parametric tests also have

assumptions

Confidence Intervals (CI)

• Gives us some idea of the size of the difference between two groups and the direction of the difference

• In practice use Bootstrap

Little Diagnostic Testing Lingo

• Positive Predictive Value (PPV)– Probability diseased given POSITIVE

test result• Negative Predictive Value (NPV)

– Probability NOT diseased given NEGATIVE test result

• Predictive values depend on disease prevalence

Analysis Follows Design

Questions → Hypotheses → Experimental Design → Samples →Data → Analyses → Conclusions

Meta-Analysis• Formulate the question• Define eligibility criteria• Identify studies and data abstraction• Analysis• Reporting and interpreting results

Research is Hard• Measure many things• Measure each thing many different ways• Measure each of those VERY accurately

– Often– Do not lose any data– Same way every time

• You cannot know what you do not measure• Clinical research

– People participate– Coming modules discuss additional aspects

ethical, legal – Not to be missed!

Study Design Taxonomy• Intervention vs. Observational• Longitudinal vs. Cross-sectional• Prospective vs. Retrospective• Blinded/Masked or Not Blinded/Masked

– Single-blind, Double blind, Unblinded• Randomized vs. Non-Randomized

Ideal Study - Gold Standard• Treatment / control• Parallel groups• Superiority• Prospective• Double blind / masked• Randomized

Observational Studies• Case Reports• Case Series• Cross-Sectional or Prevalence Surveys• Case-Control Study• Cohort Study (longitudinal)• Natural History Studies• Ecological Studies (data on population rather

than individual level)

Types of Randomized Studies• Parallel Group – classic• Sequential Trials – physical sciences• Group Sequential trials – classic• Cross-over – intervention washout• Factorial Designs – independence • Adaptive Designs – gaining popularity• Enriched Enrollment – regression to the

mean• Cluster Randomized Designs

Intervention Based Research Spectrum

• Epidemiology• Quasi-experimental• Pre-clinical studies• Phase 0• Phase I• Early/Late Phase II• Phase III• Phase IV• Dissemination and Implementation• Comparative or Cost Effectiveness

Quasi Experimental, One/Single Arm, or Non-Randomized Experimental Studies

• No control group– Early in investigation

• Concurrent control “group”– Treatment assignment not by

randomization• Historically controlled

– Missing data– Poor data– Non-comparability of groups

Non-Randomized Randomized

• Can ONLY show Association

• You will never know all possible confounders!

• Can show Association AND Causality

• Well done non-adaptive randomization → unknown confounders should not create problems

Time and Other Elements• Time is our favorite confounder in

uncontrolled studies– Differential time participating is an issue– Differential drop-outs– Time in an environment, age, season

• Social support– Meeting in a group may have an impact– Talking to someone, empathy, may impact

• Exercise– Exercise helps cardiovascular risk factors– Exercise helps stress

Masking/Blinding• Less common in non-randomized studies,

but can mask outcome assessors as to hypothesis

• Specify whom to be masked, why, how, and to what

• Assess effectiveness of masking • Specify criteria for unmasking, whom to be

unmasked • Mask determination of outcome so that

reviewers are unaware of treatment assignment; provide information on "need to know" basis

Reproducible Measurements:Regardless of Study Design

• Well defined cohort• Exclusion and inclusion criteria• Study conduct• Outcomes• Study data, analyses

Studies ≠ Gold• True for randomized and non-randomized

studies– Volunteer bias– Inclusion/exclusion criteria– Measures– Artificial interventions/treatment

definitions

Nuances• Many different ways to approach a question• Usually the different answers you hear are

because of certain different assumptions the speakers have made

• The speakers likely agree with each other• Rigor of human clinical studies is applicable

to animal and laboratory studies

Conclusion• Many needs• Many questions• Research spectrum and continuum• Rationale way of thinking, creating

knowledge and evaluating evidence

Conclusion• Do a good job every time at every stage• Think about the situation from multiple

points of view• Ethics and statistics are your friends

If you do not question more once you leave here, if you do not pass on the knowledge

and encourage questioning and finding answers in a structured way: we failed

Remember• It is about the people• Patients, participants, subjects

• Jerry Sachs– Communication– Understanding– Empathy

Extra Material:Introduction to Odds Ratio (OR) and Relative Risk (RR)

Thanks to Steve Simon http://www.pmean.com/01/oddsratio.html

Statistics for Binary Data (1)• Let p = probability of an event

– 30-day survival in a study of septic patients– Proportion of TB cases that are MDR

• Relative risk (RR): p1/p2– Good for prospective studies– RR not valid in retrospective case-control

studies, biased because probability of being a case is enriched by design

50

Statistics for Binary Data (2)• Odds = p/1-p

– Used in logistic regression, especially in case-control studies when RR cannot be used

– Example: for 6-sided die with rolls {1,2,3,4,5,6}

• Odds of rolling a 3 is 1/5; compare to Prob(3)=1/6.

• Odds of rolling an even number is 3/3=1• Odds are different from probability• Comparing odds (odds ratio) different than RR

(p1/p2).51

Comparing Groups with OR • Odds ratio (OR): p1/(1-p1) ÷ p2/(1-p2)

– OR: valid way to compare groups in all case-control studies

– In retrospective cc studies, cannot compare p1 vs p2 directly, i.e. compute RR

• Fisher’s exact test will tell you if odds is significantly different groups (OR significantly different than 1)

• For rare events OR ≈ p1/p2 = RR• For non-rare events OR should not be

considered the same as RR

52

Contingency Table

Character-istic/

Exposure

Presence of Disease

TotalNumber with

Disease

Number without Disease

Present a b a + bAbsent c d c + dTotal a + c b + d N

53

Odds• If probability of diseased person being

exposed is a/(a+c), odds are:

caa1ca

a

Odds and Odds Ratio• Odds of exposure in cases: a/c• Odds of exposure in controls: b/d

Odds Ratio (OR) = [a/c] / [b/d] = [ad] / [bc]

Why we like (need) odds ratios:• For case-control studies cant calculate

probability or odds of disease for any group• But Odds Ratio for disease comparing

exposed vs unexposed = OR for exposure comparing cases to controls= ad/bc (magic)

55

Relative Risk (RR)• Risk in exposed [A/(A+B)] divided by risk in

unexposed [C/(C+D)]• Simple comparison between two groups• RR = 1 no difference in risk between the

groups

• But not used in case-control studies unless…..

Rare Disease, OR, RR• A is small compared to B

– All with exposure, # with disease vs. # without

• C is small compared to D– All without exposure, # w/ dx vs. # w/o dx

• Odds ratio estimates the relative risk well– OR is always further from unity– OR overestimates the magnitude of

protective or harmful association

Why We Debate the Interpretation

• Group A has 25% chance of death• Group B has 50% chance of death• Group B has it twice as bad! (Relative Risk)• Less intuitive to say Group B’s odds are 3x

bigger than A’s; but accurate– Actually, sadly, someone will see the odds

ratio=3 and say that you are 3 times more likely to die in Group B. Which is False. They might say the odds are 3 times higher, which will be misinterpreted.

58

Uglier• 25% mortality from current weird flu• New mutation• 75% mortality• Relative risk of 3, odds ratio of 9• Risk difference has an even different

interpretation

Risk: Difference vs. Ratio• Difference in the absolute risks

– Attributable risk– Excess risk attributable to exposure

• Relative Risk (RR)– Ratio of two absolute risks

• Hazard Ratio (HR)– Ratio between predicted risk of an event

for member of A and that of a member of B, holding everything else constant

• Is ratio the best to talk to people?

Difference vs. Ratio• Invasive breast cancer WHI (JAMA 288[3]:321-

33)• Increase observed estrogen+progestin group

– Difference in risk• 38 vs 30 per 10 000 person years

– Hazard Ratio (HR)• 26%

• Is your personal risk 26%? No• 8 more invasive breast cancers per 10 000

person years? Yes

Ratios and Differences• Odds ratios are common

– Thanks, Logistic regression• Many better understand risk ratios and risk

differences• Feasible for risk ratio and odds ratio to

decrease and risk difference to increase– So what do you believe?– 2008 Stat Med Brumbeck and Berg

Odds Ratios and Relative Risk (Risk Ratios)

• Both compare the relative likelihood of an even occurring between two distinct groups

• Both have limitations• While relative risk (RR) seems more intuitive,

sometimes it is unclear which RR we are comparing

Relative Risk• For every problem Relative Risk can be

computed two possible ways– Risk of death, Risk of survival– Does an intervention increase the

probability of breast feeding success, or decrease the probability of breast feeding failure

• Small relative change in the probability of one event’s occurrence is usually associated with a large relative change in the even not occurring

2 Examples (Steve Simon @ pmean for details)

• Physician cardiac catheterization recommendations for patients with chest pain (watched videos)

• OR is 0.57 or 1.74– Authors report (Schulman et al 1999)

physicians make different recommendations for male patients than for female patients

DataNo Cath Cath Total

Male 34 (9.4%) 326 (90.6%) 360Female 55 (15.3%) 305 (84.7%) 360Total 89 631 720

• Schwartz et al said OR overstated the effect– Relative Risk only 0.93 (reciprocal 1.07)

• But is it appropriate to look at 90.6% vs 84.7%?• Comparing rates for recommending a less

aggressive intervention (9.4% vs 15.3%) – Relative Risk 1.63, reciprocal 0.61

ExampleContinued Stopped Total

Treatment 19 (37.3%) 32 (62.7%) 51Control 5 (8.8%) 52 (91.2) 57Total 24 84 108

• Look at 3mo, did breastfeeding (BF) continue with intervention? Or Stopped?

• ‘Failure’ RR=0.69 (recip 1.45)– OR=0.16 (recip 6.2)

• ‘Successful’ breastfeeding leads to different # s• (do Survival analysis)

Odds Ratio have this issue?• OR is not dependent on focusing on one

event’s occurrence or the failure to occur, one is the reciprocal of the other

• But they have other issues

What is Important• Which event matters? Likely they both do• Some say absolute changes in risk matter

more– Who cares if you triple your risk of a rare

outcome [well….] 3*0.00000001 is basically 0

– 10% change in a common outcome is HUGE

• No wonder they say we lie with statistics

Changes• 10 fold increase in lung cancer death• 2 fold increase in risk of death from heart

disease• But heart disease kills a lot more people in

general• This is why some people discuss number

needed to treat (NNT) and number needed to harm (NNH)

NNT (from pmean)• Daily low dose aspirin for a year (genpop):

NNT=102– One fewer stroke on average for what?

• Rates may not be homogenous over time, so careful with the person years

• Giving prophylaxis antibiotic after a dog bite: NNT=16– For every 16 dog bites treated with

antibiotics we see one fewer infection on average

For Vaccines• See lots of NNT, NNH and the balance

between• Numbers make ‘sense’ to people but they

have strong assumptions

Use OR When• Case Control Designs (except maybe rare

events)• Need covariate adjustment for confounders,

etc– It is feasible to adjust a relative risk but

tricky

• But be careful when prevalence is not rare– OR can get extreme values

PLEASE DO NOT LIE WITH GRAPHSEVEN PRETTY ONES

Graphs and Tables• A picture is worth a thousand t-tests• Vertical (Y) axis can be misleading

Like the Washington Post Weather, Though

EFFECT MODIFICATION AND CONFOUNDING

Different Variables May Be• Effect Modifier(s)• Potential Confounder(s) • Other things

• If measured these are usually “covariates” in the statistical model

Effect Modification• Interaction• Synergy

– Could be larger or smaller• Association between outcome and another

variable (e.g. intervention) is modified by different levels of a third variable

Smoking, Asbestos, andLung Cancer

• Smoking (alone) ↑ risk of lung cancer by A• Asbestos exposure (alone) ↑ risk of lung

cancer by B• Smoking AND having asbestos exposure ↑

risk of lung cancer by MORE/LESS than A+B

Effect Modification• A Short Introduction to Epidemiology

– Neil Pearce chapter (2005)• The phrase effect modification, defined for

different professions– Biostatisticians, public health workers,

physicians, lawyers, biologists, epidemiologists,….

Confounding• Two or more variables• Known or unknown to the researchers• Confounded when their effects on a common

response variable or outcome are mixed together

• Association between an exposure and outcome is misestimated due to the failure to account for a third factor (the confounder)

Consider• Association observed between carrying

matches in your pocket and lung cancer– Carrying matches causes lung cancerOR– Association between carrying matches

and lung cancer is result of confounding by another unmeasured variable associated with both

(Pam Shaw, CTR Course 2013)

Coffee and Pancreatic Cancer

Coffee Drinking

Pancreatic Cancer

Coffee and Smoking

Coffee Drinking

Smoking

Confounding Example• Relationship between coffee and pancreatic

cancer, BUT• Smoking is a known risk factor for pancreatic

cancer• Smoking is associated with coffee drinking

– Coffee drinking is associated with smoking

• Smoking is not a result of coffee drinking

Coffee and Pancreatic Cancer

Coffee Drinking

Smoking

Pancreatic Cancer

What is Confounding• If an association is observed between coffee

drinking and pancreatic cancer– Coffee actually causes pancreatic cancer,

or

Coffee Causes Pancreatic Cancer

Coffee Drinking

Pancreatic Cancer

What is Confounding• If an association is observed between coffee

drinking and pancreatic cancer– Coffee actually causes pancreatic cancer,

or– The coffee drinking and pancreatic cancer

association is the result of confounding by cigarette smoking

Smoking is a Confounder: Coffee does NOT cause Pancreatic CA

Coffee Drinking

Smoking

Pancreatic Cancer

How to Handle Confounding• Identify potential confounders

– MEASURE THEM!– In the data analysis use

• Stratification, or• Adjustment (add the variable to the

model)• Fear the unknown!

More to Confounding? Yes!

• Residual confounding– Poor measure of the confounder

• Physical activity– Even when we put the confounder as

measured in the model, not really explaining the effect of real physical activity in the model

• Example– Ever Smoked yes/no; pack years

Randomization = No Confounders! Wrong!

• Side note• Randomization helps protect against

confounding• Does not prevent confounding• Non-random drop-out or attrition• Patients testing substance

– And then dropping out, or taking more of item

Confounding and Effect Modification

John Powers 35 March 2014 IPPCR https://ippcr.nihtraining.com/lecture_detail.php?lecture_id=210&year=2013

TEST YOUR KNOWLEDGE ABOUT P-VALUES AND CONFIDENCE INTERVALS

Randomized Trial to Determine Effects of a Low Glycemic Index Diet in Pregnancy

• Population: at risk women/fetuses– Women in second pregnancy– Previously delivered infant weighing

greater than 4000g• Outcomes of interest

– Birth weight of infants (mean) • Primary outcome

– Incidence of infant macrosomia (yes/no)– Gestational weight gain– Maternal glucose intolerance

BMJ 2012;345:e5605 Low glycaemic index diet in pregnancy to prevent macrosomia (ROLO study): randomised control trial

Statistical Testing Information• Independent samples t test of primary

outcome of birth weight• Two-tailed hypothesis testing• Critical level of significance (allowable type I

error) 0.05 (5%)• Superiority trial


Which of the following statements, if any, are true?

A. The P value provides a direct statement about the size of the difference between groups in the mean gestational weight gain

B. The P value provides a direct statement about the directions of the difference between groups in the mean gestational weight gain

C. The P value provides a dichotomous test of significance of the statistical hypothesis

D. The 95% confidence interval provides a dichotomous test of significance of the statistical hypothesis

BMJ 2013; 346:f321217 May 2013 Philp Sedgwick


A. The alternative hypothesis states that, in the population sampled, treatment with the low glycemic diet is inferior or superior to placebo with regard to the secondary (gestational weight gain) endpoint

B. It can be inferred that the null hypothesis was not true

BMJ 2014; 348:g355730 May 2014 Philp Sedgwick


A. In the population, it can be inferred that no difference exists between low glycemic diet and the control treatment in mean birth weight

B. The lack of significance between treatment groups in birth weight could have been the result of a type II error

C. The alternative hypothesis of the statistical hypothesis test for gestational weight gain is true

BMJ 2014; 349:g47511 August 2014 Philp Sedgwick


A. A type I error would have occurred if no difference between treatment groups in mean gestational weight gain had existed in the population

B. The type I error rate for each statistical hypothesis test was 5%

C. The type I error rate for the multiple statistical hypothesis tests was 5%


Interventions and Proportion Birth Weight > 4000g

• Interventions– Low glycemic diet from early pregnancy

• 189 of 372 infants (51%)– No dietary intervention

• 199 of 387 infants (51%)• P = 0.88


Outcome Low Glycemic Diet (n=372*)

Mean (95% CI)

Control Group (n=387)

Mean (95% CI)Birth weight (g) 4034

(3982, 4086)4006

(3956, 4056)Gestational weight gain (kg)

12.2(11.8, 12.6)

13.7(13.2, 14.2)

*If interested, read the article and later discussion pieces for more on the study. Good lessons are discussed. Not

a perfect study.


A. The difference in birth weight between treatment groups was not significant at the 5% level because the 95% confidence intervals for the two groups overlapped

B. The difference in gestational weight gain between treatment groups was significant at the 5% level because the 95% confidence intervals for the two groups did not overlap


Outcome Low Glycemic Diet (n=372)

Mean (95% CI)



(3982, 4086)4006


12.2(11.8, 12.6)

13.7(13.2, 14.2)

Interventions and Birth Weight Results• Interventions

– Low glycemic diet from early pregnancy• Mean birth weight 4034g• Standard deviation (SD) 510

– No dietary intervention• Mean birth weight 4006g, SD = 497

• Mean difference in birth weight = 28.6g• 95% CI = (-45.6 to 102.8)• P = 0.449• “absence of evidence is not evidence of

absence”BMJ 2012;345:e5605

Low glycaemic index diet in pregnancy to prevent macrosomia (ROLO study): randomised control trial


A. The difference in birth weight between treatment groups was not significant at the 5% level because the 95% confidence intervals for the two groups overlapped

B. The difference in gestational weight gain between treatment groups was significant at the 5% level because the 95% confidence intervals for the two groups did not overlap



Mean (95% CI)



(3982, 4086)4006


12.2(11.8, 12.6)

13.7(13.2, 14.2)

Interventions and Gestational Weight Gain Results

• Interventions– Low glycemic diet from early pregnancy

• Mean = 12.2 kg (4.4)– No dietary intervention

• Mean = 13.7 kg (4.9)• Mean difference = -1.3 kg• 95% CI = (-2.4, -0.2)• P = 0.017



Mean (95% CI)



(3982, 4086)4006


12.2(11.8, 12.6)

13.7(13.2, 14.2)

Which statement best describes the information provided by a 95% confidence

interval (CI) for mean gestational weight gain?A. 95% of sample participants in the diet group

achieved a weight gain between 11.8 and 12.6 kgB. 95% of the population would achieve a weight gain

between 11.8 kg and 12.6 kg if they received the low glycemic diet

C. There is a probability of 0.95 that the population mean gestational weight gain would be between 11.8 and 12.6 kg

D. There is a probability of 0.95 that the sample mean weight gain for the diet group was between 11.8 and 12.6 kg

BMJ 2012; 344:e31479 May 2012 Philp Sedgwick


A. Statistical hypothesis testing based on a critical level of significance is a dichotomous test

B. The P value provides a direct statement about the direction of a difference between treatment groups in mean birth weights

C. The P value is the probability that the alternative hypothesis was true

BMJ 2014; 349:g455011 July 2014 Philp Sedgwick

Good Luck! Enjoy the Other Modules!

• Send video and course questions to Daniel McAnally

• Send content questions to the discussion boards/speakers

Summary Module 1 - IPPCR Video and Handout …...Summary Module 1 Laura Lee Johnson Fall 2014 IPPCR...

Documents

Transcript of Summary Module 1 - IPPCR Video and Handout …...Summary Module 1 Laura Lee Johnson Fall 2014 IPPCR...