Exam 4 Lecture 6 Flashcards
Clinical trials- the gold standard
A PROSPECTIVE, LONGITUDINAL research study designed to answer biomedical or behavioral questions in humans
Clinical trials usually are testing whether and/or how well a new intervention (medication, surgery, behavioral therapy) works… and whether there are risks to the interventions.
It is RIGOROUSLY designed to control for lots of things and be really precise in DESIGN & VARIABLES= HIGHLY RELIABLE RESULTS
- The FDA and other governing/professional bodies use Clinical Trial data to decide whether or not to approve/recommend a new intervention
Intervention-> treatment aka something that disrupts/stops/improves a problem
Clinical trials- the gold standard
A PHASE 1 trial tests an experimental drug or device on a small group of people (around 20 to 80) to judge its safety, including any side effects, and to test the amount (dosage)
A PHASE 2 trial includes more people (around 100 to 300) to help determine whether a drug is effective. This phase aims to obtain preliminary data on whether the drug or device woks in people who have a certain disease or condition. These trials also continue to examine safety, including short-term side effects.
A PHASE 3 trial gathers additional information from several hundred to a few thousand people about safety and effectiveness, studying different populations and different dosages, and comparing the intervention with other drugs or treatment approaches. If the FDA agrees that the trial results support the intervention’s use for a particular health condition, it will approve the experimental drug or device.
A PHASE 4 trial place after the FDA approves the drug or device. The treatment’s effectiveness and safety are monitored in large, diverse populations. Sometimes, side effects may not become clear until more people have used the drug or device over a longer period of time.
PHASE 3 TRIAL IS WHAT IS MOST COMMONLY TALKED ABOUT.
Clinical trials- the gold standard
- Very expensive/resource intensive, and time consuming
- Nothing is perfect
- Humans designed it but often stem from many meetings and lots of thought!
- Nothing is perfect for long
When people drop out of a longitudinal study- ATTRITION
If those who drop out are different than those who don’t-> BIAS
INTENT-TO-TREAT ANALYSES helps manage this (considered “best practice”)
Other things that make clinical trials more ‘RIGOROUS’
- Double blind
- Placebo controlled but…
- Quality control is managed throughout to avoid STUDY DRIFT
- Randomization
Side point: There is such a thing as too much power
The danger of a sample size that is too small is real and the problems are pretty obvious
- Are those few truly representative of the masses?
- Less data = less opportunity to find evidence
- Insufficient evidence doesn’t mean the evidence doesn’t exist
The danger of a sample size that is too large is also real. The problems are less obvious.
- To be ‘overpowered’ means that the likelihood of a false-negative is extremely low but it also means you can make a very rare thing seem relatively common… or very small differences seem statistically relevant
Too much power… we all know… is no good
As power increases, the ability to detect an effect increases
- That’s good except what if we test 1 million cannabis users and 1 million non-users and compare resting HR
Users average HR= 68.39 +/- 0.25 bpm
Non-users average HR= 68.32 +/- 0.81 bpm
- Statistically, this effect may be SIGNIFICANT, but is it MEANINGFUL?
Would you go to CNN and report that cannabis is bad for your heart because it increases resting heart rate by 0.7 bpm- even if it was ‘statistically significant’ (p<.05)?
Sadly, this happens all the time because people (even in the media) lack DATA LITERACY
Big data can…
Detect tiny differences.
Find rare values.
Which is good for finding needles in haystacks…
BUT
also good for
Making tiny differences
and tiny changes
‘STATISTICALLY SIGNIFICANT’ (p<.05)
Which then captures the media’s attention
And they love to
Make minor blips into giant trends ($$$$)
Reviews & Meta-analyses
A review is an article that summarizes related ‘primary’ sources (primary= the articles that actually studied something).
- Since every study is a bit different, a good review is a summary of the overarching ideas, especially ideas that are consistently observed across different types of studies.
- If 5 studies show that exercise is good for diabetes, a review will highlight for whom, what doses, when/how, and also identify what still is unknown.
A meta-analysis is an article that statistically combines data from primary sources.
- “Pools” data to create more precise statistical estimates of effects
- Often uses ODDS RATIOS (like with logistic regressions, but fancier)
IV is continuous and DV is categorical
Regression, logistic
Likelihood of developing heart disease based on endurance exercise history
Logistic regression
Models the likelihood (probability) of a categorical outcome (DV) based on a continuous (IV)
- Likelihood of cirrhosis of the liver (you either have it or you don’t) based on how much you drink
- Likelihood of getting approved for a new CC (they either give you one or they don’t) based on your credit record
This analysis asks whether an independent variable is related to the likelihood of an event taking place?
- Above examples, DV= how much you drink/your credit record
Result is expressed as an ODDS RATIO.
- Probability of success (or “yes”)/probability of failure (or “no”)
- 0.8/0.2= odds are 4 to 1 for success
Analysis also measures “goodness of fit”
Logistic Regression
To be significant (to reject the null)- the odds (central tendency) and confidence intervals (variance) cannot cross the 1.0 vertical dashed lines.
- 1.0 means the odds are equal for both outcomes.
- Further away from vertical dashed line= greater likelihood that variable (IV) is related to the outcome (DV)
- Also related to lower Type 1 (false-positive)
To left means lower odds of resulting in the outcome
To right means higher odds of resulting in the outcome
Sometimes, meta-analyses talk about ‘effect sizes’
Hedge’s g
- Size of the intervention effect (continuous DV) in each study is pooled and statistically compared to 0 using a z-statistic (z=9.48, p<.001)
- Then, they use a Qw statistic to determine if the differences between studies was due to sampling error. - Non-significance means differences were just due to sample differences, no other major source of differences are evident.
- Then, a funnel plot is drawn with central tendency (g) on x-axis and variance (standard error) on y-axis.
- Looks at efforts reported in each paper, with some estimate of study size/precision
- Bigger/more powerful studies show up on top, smaller, lower-powered studies tend towards the bottom
- When all individual studies fall within the ‘funnel’, the risk of bias is low/study results are reasonably similar