Clinical Research Oriented Workshop (CROW) Meeting: Nov 17, 2011

Present: Kairn Kelley, Amanda Kennedy, Rodger Kessler, Ben Littenberg, Connie van Eeghen

1. Start Up: Amanda – goal for today is to become fluent in “survival analysis” and why the article (Ridker, Rosuvastatin…, 2008) raises flags about supporting the treatment it advises.

2. Ben: Has been reading “Made to Stick” – a book that indicates that numbers alone are rarely enough to change behavior; the “story” needs to be present too. (Relates back to H. Pylori as the cause of ulcers, with treatment by antibiotics, translated to practice by personal demonstration rather than RCTs.) The interpretation of survival analysis studies is based on the following:

a. Approach like any hypothesis testing exercise, with time involved

b. Two variable outcome: will there be an event and when will it happen?

c. Analytical rules

i. One event (even a non-fatal event) removes the subject from the study

ii. Death for any reason is a cause for removal as well

d. The curves illustrated in the article (placebo and experimental drug) show the cumulative incidence to time “t,” or the rate of events for the entire population up to any time “t” conditional on being at risk in the time period (“time to event” analysis)

e. Kochs figured out how to smooth the step-wise curve; Kaplan-Meier created a log rank, non-parametric test. It is an omnibus test: it tests all comparisons but doesn’t say where any differences lie, just whether they are present.

f. The p value indicates whether the cumulative survival rate over time is (or is not) different.

g. This analysis gets a little more complicated when subjects are recruited over time (they don’t all start at the same time); survival for a “short time” subject isn’t as meaningful as for a “long time” subject. Calendar time survival shows incoming and outgoing subjects throughout the study:

i. Survivors who are short timers are “censored out” of the denominators and numerators. Elapsed time survival “pulls back” new recruits; they don’t have an effect on later time periods because they haven’t been in the study long enough:

ii. K-M calculates the cumulative effect of the probability for each time period within the study

h. Is the graph in the article convincing? Reasons that the difference between 2 groups could exist:

i. Randomness, which is affected by:

1. Sample size

2. Measure of variance

3. Effect size

4. And is measured by p value

ii. Bias – systematic error, sampling bias, or fraud – in general, we never know for sure

iii. Real effect – the difference is a reflection of the truth – and we’ll never know, again

i. Precision: decreases over time, as the number of survivors decreases over time. This is why the “number at risk” should be included below the graph for each time interval.

j. Questions to ask about a study:

i. Is it worth it (to prescribe a drug) if the drug costs $1000/year, which for every 20 out of 100 will have unwelcome effects and 1 will be saved a critical event that would cost ~$50K? Is the effect size worth it?

ii. Is the biology convincing?

iii. Was it funded by a neutral source? Some judgment is required here.

3. Next Workshop Meeting(s): Thursday, 12:30 p.m. – 2:00 p.m., at Given Courtyard Level 4

a. Nov 24: Cancelled – Thanksgiving

b. Dec 1: Connie – data presentation or R03 draft (no Rodger)

c. Dec 8: (no Ben)

d. Dec 15: (no Ben)

e. Dec 22: (no Amanda, Kairn)

f. Dec 29: UVM closed

g. Jan 5: Kairn - update

h. Future agenda to consider:

i. Ben: budgeting exercise for grant applications

ii. Rodger: Mixed methods article; article on Behavior’s Influence on Medical Conditions (unpublished); drug company funding

iii. Amanda: presentation and interpretation of data in articles

iv. Future: Review of different types of journal articles (lit review, case study, original article, letter to editor…), when each is appropriate, tips on planning/writing (Abby)

Recorder: C. van Eeghen