Final Flashcards
Procedure Selection Bias
When healthy patients are given one treatment over another which results in the inflation of good qualities towards the treatment
Fix: randomization
Post Entry Exclusion
When the exclusion criteria is changed after the data has been seen
Fix: no post hoc changes
Selective Loss of Data
Caused when there is data missing resulting in a difference in the population. Effects vary depending on what the mechanism of missing is.
Assessment Bias
When the physician overestimates or underestimates a condition based upon the exposure or group of the individual
Fix: mask
Ascertainment Bias
When one group is sampled more or in greater detail relating to their illness or treatment group
Fix: mask
Single Mask
Patient doesn’t know treatment
Double Masking
Patient and be physician do not know the status of the individual
Triple mask
Patient doctor and DMC do not know the groups. This works best for objective testing of the stopping point so as not to favor one group over another.
Concurrent Trials
Eliminates the time effect of the study that would be found in historical controls. Eliminates the differences in selection criteria and time treatment
Active Follow Up
Follow up patient who drop out to make sure that there is no mechanism of why one group quits more.
Drop In
Patients who should have been in the control are in the treatment (p)
Drop Out
Patients who should have been in the treatment group are given the placebo
SS calculation win drop in and drop out
N=N*/[(1-p-q)^2(1-r)]
Cross Over Design
Repeated measure longitudinal study where the patient will receive both treatments. Subject to bias if there is a crossover effect.
Binary Cross Over Null
p12-p21=0
This is the off diagonal of the patients who showed a preference to one treatment over another
McNemars Test of Cross Over
Binomial(.5, n10+n01)
The null is that the preferential response is random with a frequency of 50%
Factorial Design
Only test that can test interaction effects the interaction effect can be removed afterwards unless one reverses another.
Incomplete Factorial
There is no placebo. Used to test is a combination of drugs is better than each of the mono therapies.
Null: one is better than the combination
Alt: the combination is better than both of the other two
Adaptive Design
Allows for a change in the study design, statistics, sample size, treatment, hypothesis, power, and end points
Very flexible
Sequential Analysis
SS is not fixed at start but while data is collected till a predefined stopping rule is reached. Lowers the over all cost
Can be done for futility, safety, superiority, or cost reasons
Conditional Power
It is the probability of getting statistically significant results given the trend in data
Drop the Loser Design
The inferior treatment group is dropped at a interim analysis and the population groups are changed to allow for new testing. Power is determined after second stage
Interim Analysis
Data dependent stopping or continuation. Check to see if null can be accepted or rejected. Also allows stopping for new literature
Continuation alleys for increase in precision and power
Type I Error
The probability of rejecting a null that is true. Overall type one error should be kept below .05
aT=1-(1-a1)(1-a2)…. Or aT=1-(1-a)^n
Stopping Rule
If the test statistic at that point falls into the boundary of acceptance or failure then the trial is stopped and the parameters are estimated.
Pocock p-value adjustment
Assumes a fixed p value at each location which means there is a high starting p-value and a lower ending one. Typically stops trials early and bad since the study cannot reject a significant hypothesis at the end
OBrien Fleming P-value adjustment
Most used but with the most amount of rules. Required the amount of IA to be specified beforehand and that there is equal spacing in the IA. Ending p-value is .05
Alpha spending function
The p-value is a function of the amount of information collected at each point
a(t) where t=n/N or t=d/D
a(0)=0 and a(t)=a
Non-inferiority Trials
Null is that the new treatment is inferior (uC-uT>a)
One sided test where the entire CI must be (-a,0)
Margin of Indifference
The allotted difference in the new treatment and the old treatment if the new treatment has some sort of advantage
Longitudinal Study
Patients are blocks with high correlation with their measurements. Repeated measures with high correlation. Use GLMM.
GLMM
Has both fixed and random coefficients for subject specific differences. Error is no longer independent for the subjects. The study parameter regression coefficients assumed same over the entire population and the coefficients of the person fixed for that person
Estimated with the generalized least squares
Case Symmetric Compound
Variances are the same but different from the covariances which are the same. Unrealistic
Unstructured Covariance
Requires most amount of data to fit the most parameters. Most liberal
Auto regressive
Makes the measurements closer together similar but drop off rapidly as the values get further apart.
AIC
Want it to be lower and is a good check for model fit
Goodness of Fit
Add the two -2loglikelihoods and test it with a chi square with two degrees of freedom
Missing Data
Can use GLMM when there is missing data for MCAR or MAR
GLM: Continuous Data
Use a normal distribution with g(u)=u
GLM: Binomial Data
Use a proportion in a logistic regression
g(u)=log(u/(1-u)) logic
GLM: Count Data
Use a poison regression with g(u)=log(u)
GLM: positive continuous data
Do a Gamma model with a gamma distribution and a link function g(u)=log(u)
GEE
Generalized estimating equations when the Covariance is unknown
Fixed Effect Meta Analysis
Mathematical assumption that the studies are looking at the same treatment effect
Random Effects Meta Analysis
That there is a difference in the treatment effect over the studies. Good for when there is statistical heterogeneity in the studies (look at Forest plot)
Selection Bias
It is when the sample is not representative of the target population resulting in lower external validity
Fix: Random sampling