Term
True: False
Data are more likely to be normally distributed when each value is dependent upon one or a few factors |
|
Definition
False:
To be normally (gaussian) distributed, data points must be dependent upon a large number of factors (blood pressure is a good example). |
|
|
Term
| How many standard deviations from the mean is equivalent to a Z score of 2? |
|
Definition
|
|
Term
| If a distribution is skewed, what measure is used in place of mean? Standard deviation? |
|
Definition
Median (n+1/2)
Interquartile range (25%-75%) |
|
|
Term
| How many degrees of freedom are there in a sample of 100 people from a population of 10,000? |
|
Definition
99 (n-1)
An easy way to remember DF is to think about how many data points are "free to vary." In this case, if you know the sample of 100 people has a total salary of $32,000,000, and you know how much money each of 99 of the people makes, then the last individual's salary is not free to vary. |
|
|
Term
| Provide a conceptual definition of "standard error of the mean" |
|
Definition
| SEM is the standard deviation of a distribution of sample means. Remember, even if a population is not normally distributed, the plot of means will ALWAYS be normally distributed. |
|
|
Term
| How is standard deviation different from standard error? |
|
Definition
| SEM is an estimate of the precision of a sample as an estimate of the population mean, while standard deviation is a measure of variability around a mean. |
|
|
Term
What does it mean to examine the ratio of "explainable:unexplainable" variation in a test of significance between two groups?
Which test compares the "Mean Squares" between groups (explainable) to that within groups (unexplainable)? |
|
Definition
"Explainable variation" arises from differences b/w treatment groups and "unexplainable variation" is attribute to individual differences. If a treatment caused a significant difference, then the ratio of "explainable:unexplainable" variation should be large.
ANOVA |
|
|
Term
| What three assumptions must hold true in order to use an ANOVA appropriately? |
|
Definition
1) Samples must be independent and randomly selected
2) Normally distributed populations
3) Equal variances b/w groups (i.e. the degree of underlying variability within the population is consistent- homogeneity of variance |
|
|
Term
| What test is most appropriate to use when testing differences b/w groups on a nominal scale? Why is it sometimes important to addend this test with something called the "Yates correction"? |
|
Definition
Chi Square
If the Df is 1 [ (r-1 * c-1) ], then it is important to account for this using the Yates correction. |
|
|
Term
An experimenter collects data on the number of motorcycle accidents causing deaths vs. recoveries b/w groups wairing a standard helmet and a super-duper helmet.
What type of test would you use to compare these groups and why? |
|
Definition
These are nominal, categorical data.
Use a Chi-Square test to see if there is a statistically significant difference and follow up with a yate's correction since DF= 1 |
|
|
Term
| You are conducting a small study comparing the efficacy of a new surgical technique to stop peptic ulcer-related bleeding, and you want to compare it to the older standard. What test would you use to compare the treatments in two groups in you had a small sample size? Why use this test in particular. |
|
Definition
Fischer exact test (2X2 table where nominal value <5 for a given element).
P = R1!R2!S1!S2!/ N!a!b!c!d! (formula may not be important)
Chi-square is the other test that would be used to compare nominal variables, but it relies on the assumption that a normal distribution can be used in place of a binomial distribution as an approximation, and this becomes invalid when sample sizes are small. |
|
|
Term
| When drawing data form a population that is not normally distributed, what test can you use in place of ANOVA or T-tests? What about if the data are paired, as in the case of baseline cholesterol and cholesterol 3-months post treatment? |
|
Definition
1) Mann-whitney rank sum test (significant difference b/w 2 measures)
2) Wilcoxon signed-rank test (significant difference b/w unlimited paired data points). |
|
|
Term
| Why use Kruskall-Wallis test instead of a Mann-Whitney? |
|
Definition
| If you are making a nonparametric comparison with more than 2 groups- (DF= K-1) |
|
|
Term
| What statistical test would you use to test whether or not there is an association between the temperature on a given day and the number of asthmatic-related doctor's visits, provided that the data are normally distributed? What about if the data are skewed? |
|
Definition
1) Pearson correlation to find "r," and then a T-test using that r value (DF= n-2).
2) Can do a spearman rank-order correlation |
|
|
Term
| You are conducting a study in which you want to determine the utility of looking at smoking habits in terms of # of pack years, as a predictor of early-onset lung cancer. What statistical test would you plan to use? |
|
Definition
The independent variable in this case in continuous (# of pack years), but the outcome is nominal (cancer vs. no cancer).
You are trying to determine "predictive value," so you need to use a regression, and since the outcome variable is categorical, you should use a Logistic Regression |
|
|
Term
| You are conducting a study in which you want to determine the utility of looking at smoking habits in terms # of pack years, as a predictor of survival time, post-quitting. What statistical test would you plan to use? |
|
Definition
| Both the independent and outcome variables are continuous, and their is only one independent variable, so you should use a linear regression (as opposed to a multiple regression for multiple, continuous, independent variables). |
|
|
Term
| When evaluating new treatments, three criteria should apply to testing. What are they? |
|
Definition
1) The test should be evaluated among the type of patients who would actually receive a test in clinical practice
2) The test should be compared to an appropriate gold standard
3) The gold standard should be performed in all the patients being studied regardless of the result of the test being evaluated
**New tests form a 2x2 contingency table** |
|
|
Term
True:False
Sensitivity and Specificity are measures that vary with disease prevalence (i.e. proportion of population with a disease) |
|
Definition
False!
PPV and NPV vary with disease prevalence. Sensitivity and Specificity are fixed values for a given test |
|
|
Term
| In introducing a new diagnostic test, you wish to compare it to the gold standard. You first want to determine of people suffering from a given disease, how likely is it that your test will reflect that? |
|
Definition
Specificity.
Remember, Specificity and Sensitivity are fixed for a given test assume a priori knowledge of the actual disease state, while PPV and NPV vary with disease prevalence and are focused on a positive or negative test result.
In this case, you know patients do NOT have the disease, and you want to know what proportion of tests you run on this group of patients will come back negative. This is an example of SPECIFICITY (D/ D+B) |
|
|
Term
| When determining the diagnostic value of a new test that has more than 2 possible outcomes, what statistical test should you use? How does this test relate to sensitivity and specificity? |
|
Definition
Likelihood Ratio!
Remember, >10 means the disease is likely and <0.1 means the disease is very unlikely.
When only two test outcomes,
LR+= sensitivity/ (1-specificity)
LR -= (1-sensitivity)/specificity |
|
|
Term
How does Bayes Theorem relate to Likelihood Ratios?
If the probability of a patient having lung cancer is 33% and a test comes back with "High probability," which is equivalent to a LR of 15, how likely are they to have cancer? |
|
Definition
1) BT= Prior odds X Bayes Factor= Final Odds
For a diagnostic test
Pre-test odds X LR (of a given test)= post-test odds
2) 33% probability is equivalent to a 1:2 odds
so, using BT, (1/2) X 15= 15/2, which is then translated to 15/(15+2)= 0.88 or an 88% chance of having the disease after taking the test. |
|
|
Term
| Why are diagnostic tests most useful in patients with intermediate pretest probabilities? |
|
Definition
If the pre-test probabilities are either very high or very low, the test result will have much of an effect on the post-test probability.
For example, if pre-test odds are 1/99, a positive test-result (for example, high LR of having a heart attack) will only make the post-test odds 5/99, which is still extremely low and uninformative. |
|
|
Term
| What are 4 common errors of estimation? |
|
Definition
1)Availability: a tendency to over-estimate the frequency of vivid events and under-estimate the frequency of mundane events (SARS in the media)
Conjunction Fallacy: Overestimation of a rare situation based on extraneous factors (asbestosis b/c a person works in basements, even though asthma is FAR more common).
Support Fallacy: more detailed descriptions may lead to overestimation (describing every minutia of a heart attack increasing the pretest probability)
Overemphasis of positive findings or underestimation of negative findings |
|
|
Term
| When SARS was accentuated heavily in the media, physicians tended to confuse many common cold cases with possible SARS outbreaks. What error of estimation is this an example of? |
|
Definition
|
|
Term
| A young, politically-inclined college student graduates from college and starts her first job. Is it more likely that she will be a bank teller, or a bank teller who also participates in election-related activities? |
|
Definition
Bank teller!
Answering the second option is an example of Conjunctive Fallacy. |
|
|
Term
| What are the characteristics of a good clinical question? |
|
Definition
FINER
feasible
interesting
novel
ethical
relevant |
|
|
Term
| What are the major characteristics of each phase in diagnostic research trials (i.e. phase 1 vs. 2 vs. 3 vs. 4 |
|
Definition
Phase 1) Does test have potential to detect differences b/w ill and well
Phase 2) What is the cutoff for a particular marker in the test (i.e. upper left corner of ROC, which is plotted as sensitivity vs. 1- specificity)
Phase 3) Gold standard comparison in target patients
Phase 4) Does the test represent a significant improvement? |
|
|
Term
| What are the four structural elements included in an appropriately framed clinical question? |
|
Definition
PICO
P) A specifically defined patient Population or Problem
I) an Intervention to study
3) A Comparison intervention to compare the study with
4) A precisely defined Outcome of interest |
|
|
Term
| What is the difference b/w internal validity and external validity in a study? |
|
Definition
Internal- "noise" created by confounding variables is minimized.
External- Study is generalizable to a greater population of people outside of the study |
|
|
Term
| What is the difference b/w inclusion and exclusion criteria in selecting participants for a study? |
|
Definition
Inclusion Criteria: the target population that is accessible to the researcher (i.e. men and woman aged 50 or older with diagnosed A-fib)
Exclusion Criteria: patients who meet the inclusion criteria, but are not suitable subjects for some other pre-determined reason (an example might be a patient who is the right age and has the right condition for study, but who has had previous heart conditions) |
|
|
Term
| You are studying a population of 1000 people with atrial-fibrillation in Pittsburgh and you want to study 100 of them. You generate a list of 100 random numbers between 1 and 1000 to choose your subjects. What kind of sampling is this? |
|
Definition
| Simple Random Sampling (probability sampling): using some technique to randomly include and exclude appropriate patients in the study |
|
|
Term
You are studying a population of 1000 people with atrial-fibrillation in Pittsburgh and you want to study 100 of them. You want to study race as a covarying factor in your study, so you break your population up by race and then generate two lists of 50 random numbers, between 1 and the # of people in a particular racial group (1 list for each group), to select from those subgroups. You end up with 50 African american men and woman, and 50 caucasian men and women for your study.
What type of sampling is this? |
|
Definition
| Stratified random sampling (probability sampling): artificially tweaking the simple random sampling to include a population of particular interest in appropriate numbers. Done by stratifying the population, then randomly selecting from the strata |
|
|
Term
| You want to study patients with A-fib across U.S hospitals. You randomly select 5 centers from around the country and then randomly select 500 patients from each center. What type of sampling is this? |
|
Definition
| Cluster Sampling (probability sampling): uses natural groupings of test subjects |
|
|
Term
| To save money and time in your study on A-fib in patients living in Pittsburgh, you decide to include the first 500 people who agree and who fulfill the inclusion and exclusion criteria of your study. What type of sampling is this? |
|
Definition
| Consecutive (non-probability): involves recruitment of all acceptable subjects |
|
|
Term
| You are at a conference for patients suffering for A-fib and you are trying to conduct a study on the subject at your home institution. Ethical issues aside, you recruit 100 of them for your study, thinking that it is your ideal population. What type of sampling is this an example of? |
|
Definition
| Convenience (non-probability)- involves recruiting the most readily available subjects |
|
|
Term
1) What type of error have you committed when you conclude that taking the drug Ceprotex leads to different outcomes from the gold standard when it does not?
2) What about when you determine Ceprotex and the standard lead to the same outcome when there is actually a difference? |
|
Definition
1) Type I Error: incorrectly rejects H0 when treatments are equal
2) Type II Error: incorrectly accepts H0 when treatments are not equal |
|
|
Term
True or False:
An α (alpha) value represents the acceptable chance of committing a type 1 error. |
|
Definition
True!
α (alpha): the acceptable chance of being wrong by finding a difference where none actually exists
β (beta): the acceptable chance of being wrong by not finding a difference where one really exists |
|
|
Term
| In your study, you set β (beta)= 0.2 to a achieve a power of _____? |
|
Definition
Power= β - 1, so P= 1- 0.2
80% power, meaning that you are 80% certain that if there is a difference, you will detect it (i.e. 20% chance of committing a type 2 error) |
|
|
Term
| You are comparing a new, very expensive drug to the gold standard. Why might you be concerned of making a type 1 error when making inferences from your data? What about a type 2 error? |
|
Definition
If you made a type 1 error, that would cause people to switch to a more expensive medication when it has not been proven to improve their health.
If you made a type 2 error, you might be ignoring a significant therapeutic benefit of the novel treatment. |
|
|
Term
| What is the non-centrality parameter and how does it relate to the "magnitude of effect size"? |
|
Definition
Non-centrality parameter (ψ): the ratio of the Magnitude of Effect (δ) relative to the variability within a population (population deviation σ)
ψ= δ/σ
The larger the value of ψ, the higher the power of the study and the more likely a study is to detect a difference.
Recall, Magnitude of Effect (δ) is the magnitude of difference that the test is designed to detect by limiting the minimal range of acceptable effect |
|
|
Term
| What are the 3 determinants of statistical power in a study? |
|
Definition
1) Non-centrality parameter (ψ)- maximize ψ to maximize power
2) Type 1 error rate (α)- lower α, the lower the power
3) Sample size- maximize to maximize power |
|
|
Term
| How many subjects do you need to detect an absolute magnitude of effect of 10 mgHg between two blood-pressure medications, assuming a standard deviation of 20 mgHg, and alpha level of 0.05 and 80% power? |
|
Definition
Recall: n= 16/ψ^2 & ψ= δ/σ
so... n= 16/( δ/σ)^2
n= 16/ (10/20)^2 = 16/0.25 = 64 people! |
|
|
Term
| What is allocation concealment and why is it important? |
|
Definition
| Allocation concealment is the process of keeping those who are allocating subjects unaware of the different groups. It is important because it prevents experimenters by being biased in their subject assignment. |
|
|
Term
| What is the difference b/w single, double and triple blind experiments |
|
Definition
In single blind studies, only the subject is unaware of their allocation (which drug they are taking)
In double blind studies, both the patient and the investigator are unaware of the allocation of the subject (the drug administrator might treat patients differently depending on the drug they are taking)
In triple blind studies, the patient, investigator, and the assessor of the data don’t know which group the subjects have been assigned to (the neurologist interpreting the results of the treatment might be biased) |
|
|
Term
| Why is block randomization sometimes used in subject allocation? |
|
Definition
to make sure that groups are balanced!
Example) ABBA, or BBAA (there are 6 possible permutations of these 2-group blocks). Assign a random number generator to pick block permutations randomly |
|
|
Term
| Why would you decide to choose stratified random sampling? What does it have to do with prognostic factors? |
|
Definition
To avoid prognostic factors! For example, you might think that age and sex would effect your outcomes, and decide to stratify your groups based upon those categories as a result.
Remember, Prognostic factors are variables that are suspected as confounders |
|
|
Term
What is relative risk reduction (RRR) and how does it relate to experimental event rate (EER) and control event rate (CER)?
What is one limitation of RRR? |
|
Definition
RRR= CER-EER/CER
RRR= 0.5 would mean that compared to control, the experimental group is 50% less likely to get the disease.
RRR doesn't take "magnitude" into account! |
|
|
Term
| What is the absolute risk reduction (ARR) and how does it relate to number needed to treat (NNT)? |
|
Definition
ARR= ICER - EERI
NNT= 1/ARR or 1/(CER-EER)
for example, an ARR of 0.074 means that out of 100 people treated, 7.4 fewer people in the drug group suffered from a disease, and that approximately 14 people need to be treated with the drug to prevent 1 case of the disease. |
|
|
Term
| What are some limitations about using p-values exclusively to interpret results? |
|
Definition
1) The "p-value fallacy": Assuming that the null hypothesis has this probability of being true given the results, when it is the opposite
2) only show statistical significance (not practicality, magnitude)
3) Can be manipulated by changing sample sizes |
|
|
Term
1) Why use confidence intervals instead of just using P-values?
2) How do you interpret them? |
|
Definition
1) Confidence intervals provide information about severity and magnitude of effects.
Recall, A 95% confidence interval can be interpreted as “95% of the time the test results will fall in this interval, and only 5% of the time would the results lie outside this interval.”
The width of the confidence interval is the precision of the estimate of the difference between two studies
2) If a confidence interval for an absolute difference includes 0, we should accept the null hypothesis, since 95% of the time, there may be no absolute difference between treatments
If the confidence interval is for a likelihood ratio, then accept the null hypothesis if the interval includes 1 |
|
|
Term
| You are following a group of 20 smokers over a period of 10 years to see if they develop emphysema. What type of study design is this? What if you decided to look back through the medical records of a population of 100 smokers? |
|
Definition
Both are Cohort designs. The first option is a prospective cohort study, while the second is retrospective cohort.
Since you are measuring the incidence of the disease, this is a "descriptive cohort study." Remember, an analytical cohort study would focus on a risk factor and have an experimental and control group. |
|
|
Term
| What is the difference between an internal control and an external control in a cohort study? |
|
Definition
Internal comes from the "inception cohort" (the group of people recruited).
External comes from outside the inception cohort. |
|
|
Term
| What is the relative risk (RR) (also called "risk ratio") in a cohort study? What would an RR of 5 mean? |
|
Definition
Using a 2 X 2 contingency table, the RR is the ratio of the incidence of the outcome among subjects "exposed" to the incidence of the outcome among those "not exposed".
It usually translates to RR= [a/(a+b)] / [c/(c+d)]
An RR of 5 would mean that the incidence of a disease is 5X more likely in a group "exposed" than "not exposed" |
|
|
Term
| In a prospective cohort study comparing the incidence of asthma between people who smoke and those who don't, you discover an RR of 5, 95% CI (0.8, 7.2). What would you draw from this study? |
|
Definition
| The RR is high (would suggest 5x greater likelihood of disease onset after exposure), but the CI includes 1 and the results are therefore inconclusive, statistically speaking. |
|
|
Term
| What is the major difference b/w case-control studies and cohort designs? What is Anamnesis? |
|
Definition
In a cohort design, you do not know the outcome in advance, while is a case-control, you are comparing a group exhibiting a particular outcome, to a second group that does not (control).
Case-control studies are retrospective, while cohort studies can be either prospective or retrospective.
"Anamnesis" refers to collecting information about a subjects life that could be relevant to the outcome of interest. |
|
|
Term
| What is the difference b/w Odds Ratios (OR) and Relative Risk Ratios? What is the same about their interpretation? |
|
Definition
1) Odds Ratios are reflective, while RR is predictive
2) ORs can be used in cohort and case-control, but RR only in cohort
3) In "case-control," OR gives the odds that a diseased person was exposed to a particular risk factor.
In "Cohort," OR gives the odds that the exposed will develop the disease
4) OR= (a/c)/(b/d) or ad/bc
RR= [a/(a+b)] / [c/(c+d)]
** If either CI includes 1, they are insignificant** |
|
|
Term
| Provide 5 justification for why a cohort design might be preferred over a case-control design? |
|
Definition
1) Sampling Bias: choosing cases that are not representative of the general population of interest
2) Survivor Bias: only survivors can be polled retroactively
3) Differential Recall Bias: cases with a condition are more aware of their risk factors
4) Cohort designs establish sequences of events, while case-control studies can merely distinguish exposure
5) Cohort studies can measure an infinite number of variables |
|
|
Term
| Why might you choose a case-control design over a cohort study? |
|
Definition
1) Case-control studies are less expensive
2) Cohort studies are impractical for studying rare diseases or diseases requiring many years to appear
3) Cohort studies are more prone to confounding |
|
|
Term
| What is a "Forest Plot" in a systematic quantitative review article? |
|
Definition
Forest Plot: a set of graphics from many different studies that form a quantitative summary of results
The results are weighted and combined to form new, more comprehensive data for the systematic review |
|
|
Term
| Why would you use Cochran's Q Test? How does it relate to consistency (I^2)? |
|
Definition
1) It is a way to measure statistical heterogeneity b/w studies in a systematic review article
This value is compared to the Chi-squared value with k-1 degrees of freedom, where k is the number of studies
When the number of studies is few, Q has little power to detect heterogeneity (many studies could show a significant Q despite low levels of heterogeneity)
2) I2 can be defined as the percentage of variability among studies that is due to statistical heterogeneity rather than chance
If I2>50%, there are high levels of statistical heterogeneity |
|
|
Term
| Why might you use a "random effects" model for pooling results from different studies versus a "Fixed-effects model"? |
|
Definition
If there is statistical heterogeneity.
Remember, both models are used to account for differential weighting of studies being pooled
Fixed Effects Model: assumes that all studies in a meta-analysis are estimating the same effect, so any variability is due to random variation that would be eliminated if all studies were infinitely large
The Random Effects model assumes each study is estimating a different effect and variability is accounted for when calculating the overall effect. |
|
|
Term
| How and why does someone conduct "sensitivity analysis" in a systematic review? |
|
Definition
Sensitivity analysis varies the way the systematic review is done to check the accuracy of the conclusion
One way is to vary the inclusion/exclusion criteria
Another way is to vary between random and fixed effects models |
|
|
Term
| What are 3 examples of Bias in systematic review articles? How do you detect bias in a review? |
|
Definition
1) Publication Bias is only incorporating those studies that have been published, since they will generally be studies with statistically significant results, and ignoring the unpublished studies can result in ignoring vital, yet statistically insignificant data
2) Citation Bias results when studies are found by looking at the sources from a particular paper
3) Biased inclusion criteria can occur when the reviewer knowingly or unknowingly allows certain studies over others into the review
A funnel plot can detect certain forms of bias by plotting the effect estimates of individual studies in a review against the studies’ sample sizes or other accuracy variable
(Larger studies give more precise effects, and smaller studies give a wider effect, so if there is no bias in the publication or identification of studies, the funnel plot should resemble a funnel. Asymmetry indicates bias) |
|
|
Term
| What are the 3 popular types of Fixed Effects Methods for pooling results in systematic reviews and when is each one employed? |
|
Definition
1) Inverse Variance- for continuous data
2) Mantel-Haenszel- best for small # of studies or small # of people within each study
3) Peto- used when one of the cells in a 2 X 2 table is "0" |
|
|
Term
| What are two popular types of Random Effects Methods for pooling data in systematic reviews? |
|
Definition
Remember, REMs are used because they take into account, the variance between studies as well as within each study (good when there is heterogeneity)
1) DerSimonian
2) Laird |
|
|
Term
| How can "censoring" occur in survival analysis? |
|
Definition
When there are unknown survival times in a study.
1) sometimes the patient has not experienced the event by the end of the study
2) lost during followup
3) May experience other events that compromise the results of the event of interest |
|
|
Term
| What is the difference b/w right censoring and left censoring? |
|
Definition
Right censoring- the event takes place beyond the end of the follow-up
Left censoring- the event takes place before the beginning of the follow-up period
** Interval censoring occurs when study subjects come and go during the follow-up period** |
|
|
Term
| Why is "informative censoring" a problem is Kaplan-meier risk analysis |
|
Definition
| Censoring is related to prognosis, and therefore invalidates the statistical procedures normally applied. Uninformative censoring involved a small number of patients who are censored because of reasons not having to do with prognosis |
|
|
Term
| Why is the kaplan-meier method important for survival analysis? |
|
Definition
| It take into account censoring (especially "right censoring"). You can see censoring in the "tick" marks on the curve (i.e. people that were lost from the study). |
|
|
