Final Exam Flashcards
How to calc variance and std dev
- Find diff b/n ea value and the mean; square this difference
- Add up all the squared diff and divide by total # of values
- Std Dev= Square root of variance
FINER
- Feasibility
- Interest
- Novel
- Ethical
- Relevant
PICOT
- Population
- Intervention
- Control
- Outcomes
- Time
Categorical Variable (2 types)
- Nominal (no order) v ordinal (high/med/low)
- Use percentages and frequencies for central tendency
Skewed Distributions
- Neg/L - more values on high end
- Pos/R - more value on low end
**if skewed use range and IQR for measure variance NOT variance and SD
Central Limit Theorem and Confidence Intervals
- If we take repeated random samples from our population, calculate each sample mean , and plot out those sample means, then:
- Mean of sample means = population mean
- Std error=std dev of sample means (std error=sample SD/square root of sample size)
- Confidence Intervals- range of numbers that population mean will likely be within x% probability given observed sample mean and size
- 95% CI is approx sample mean +/- 2SE
- 95% CI is wider than a 90% CI
What does p-value mean?
- Probability of observing difference this extreme or more given null hypothesis is true
- If value is sufficiently small, then reject null hypothesis and accept alternative hypothesis
NOT the probability that the null is true OR that 1-p represents probability that alternate is true
3 Ways to Determine Stat Sig
- P-value below cutoff? (.05 normally)
- Test statistic exceeds the critical value
- Confidence interval of desired probability EXCLUDE 0 (group difference) or 1 (group ratio)
- If 2 groups are the same (null is true) then the diff b/n there means would be 0 and the ratio between their means would be 1
Parametric v Non-parametric
- Parametric assume dependent variable is normally distributed so…
- Takes advantage of known properties of a distribution , allows for efficiency (less subjects, detection of smaller effect size ), allows for effect estimation ( i.e. confidence interval of group effect )
- Null Hypo- no diff b/n means of groups
- Non-parametric - dependent variable is not normally distributed or too few observations to assume so…
- Based on ranks (observations ordered from high to low - ties receive avg ranks)
- Null Hypo- no diff b/n dist of ranks b/n groups
- Advantages: Less requirements , useful for dealing with outliers , intuitive , useful for certain categorical data
- Disadvantages: less efficient , hypothesis testing over effect estimation , too many rank ties problematic
Chi Square
dependent and independent are categorical
non-para equivalent is Fisher Exact
T test
cont dependent variable and 2 category indep variable + equal variance b/n groups
non-para equivalent is Mann Whitney
Paired t Test
same as t test but observations/dependent variables can be paired
non-para equivalent is Wilcoxon Signed Rank Test
ANOVA
cont dependent variable and 3+ category independent variable
non-para equivalent is Kruskal Wallis Test
Pearson Correlation
normally distributed/cont dependent and independent variable
non-para equivalent is Spearman Correlation
Linear v Logistic Regression
- Linear - cont dep var
- Output = can determine stat sig of ea independent var b/c gives you coefficient (pos or neg) and CI for ea
- CI should not include 0 b/c group difference
- Logistic- dichotomous dep var
- Output= given in odds ratio (which can then be converted to probability) and CI for ea independent var
- CI should not include 1 b/c group ratio
Deductive v Inductive Reasoning
Deductive reasoning (does pt fit pattern?) -start w disease
Inductive reasoning (what pattern does this pt fit?) - diff diagnosis
Accuracy
a+d/total
(how many were true either way?)
Sensitivity
a/a+c
(how many people w/ disease did it pick up?)
-wanted for screening test (catch all)
Specificity
d/ b+d
(how many people w/o disease also tested neg?)
-wanted for confirmatory test
Pos Predictive Value
a/a+b
how often is pos test right?
Neg Predictive Value
d/c+d
how often is neg test right?
Receiver Operating Curve
- Sensitivity v 1-specificity
- Higher area under curve = better distinguishing ability
Pos and Neg Likelihood Ratios
- Pos LR - sensitivity/ 1-specificity
- Neg LR - 1-sensitvity /specificity
Bayes Theorem
Pre-test odd of disease X LR of test result = post-test odds of disease
- Estimate pre-test probability
- Convert to odds
- Mult by LR of test result that comes back = post-test odds
- Convert back to probability to make clinical decision
- Probability = odd/ 1+ odds
Case Control v Cohort
Case Control
- ID pts w/ certain disease then determine exposure status
- Observational
- ALWAYS RETRO
- Adv- quick, cheap, good for rare diseases, can access must factors associated w/ outcome
- Disadv- cannot calc relative risk b/c do not know prevalence, recall bias, confounders, not causal
Cohort
- ID people who were exposed —> follow to see if they develop disease
- Observational
- Often PROSPECTIVE
- Adv- less/no recall bias, can access association of mult factors, stronger support for causality if prospective, CAN CALC REL RISK
- Disadv- need more resources and time, ppl lost to follow-up, confounders, still may not be able to claim cause and effect
Relative Risk
- Ratio of probability of developing disease if exposed v not exposed
- Based on incidence of disease given exposure status
- Risk among exposed/risk among unexposed
- =1 then same risk regardless of exposure
- x>1 then x fold inc in risk if exposed
- 1/y <1 then y fold dec in risk if exposed
Odds Ratio for Case Control
- Odds of exposure among people with disease vs. odds of exposure among people without disease
- Because start w/ people w/ disease and look for exposure or not
- Odds for cases/odds for controls
- OR=1 no relationship
- OR > 1 there is increased odds of exposure for cases compared to controls
- OR < 1 there is decreased odds of exposure for cases compared to controls
Odds Ratio for Cohort
- Odds of developing disease among exposed vs odds of developing disease among nonexposed
- Because start w/ people w/ exposure and look for disease or not
- Odds for exposed/odds for non-exposed
- OR= 1 no relationship
- OR > 1 there is increased odds of developing the disease for exposed compared to not exposed
- OR < 1 there is decreased odds of developing the disease for exposed compared to not exposed
Censoring
- unknown survival time for subset of subjects (when event occurs b/n follow-ups
- Types: Right (after follow up begins), left (before follow up begins), interval
- Informative censoring: Censoring is related to outcome or systematic
Kaplan Meier Method
Probability of survival at any given time = prob survived up to that time X prob surviving through that time
- Curve (time v survival probability)
- Median survival time = time when 50% survival OR 50% experience outcome
- Censoring denoted by vertical hash marks
Log Rank
- Log Rank - compares 2 survival curves - get p value
- Null hypothesis: There is no difference in probability of an event at any given point between the 2 groups
- Assumptions:
- Uninformative censoring
- Survival probability the same regardless of recruitment time points
- Events happened at time specified - Limitations:
- Can only say whether difference exists, not magnitude/range of difference
- Allows comparison with respect to one factor only
- Survival curves can’t cross
Cox Proportional Hazard Model
- Regression that predicts probability of event occurring (hazard) at specific time for one group over another given mult variables
- Hazard Ratio (similar to RR)
- HR =1 no inc or dec in risk
- HR <1 dec risk
- HR > 1 inc risk
ITT v Per Protocol
- Intention to Treat - analyze according to which group they are assigned to regardless of compliance
- Effectiveness analysis - real world conditions
- Preserve random allocation
- Per Protocol - analyze by actual treatment received
- Efficacy analysis - ideal conditions
- Better for safety
Sample Size Calc
- Power- probability of detecting diff in groups if one exists (1-B)
- Effect size- amount of diff b/n groups you wish to detect
- If higher power, smaller effect size, smaller alpha (tolerance for false pos) …need HIGHER sample size
2 RTC Errors
- Type 1 error (a): Rejecting null hypothesis when study arms are not different
- Usually set at 0.05 or 5%
- Type 2 error (B): Accepting null hypothesis when the study arms are different
- Usually set at 0.20 or 20%
CER
EER
(control event rate)- # in control group w/ outcome
(experimental event rate) - # in experimental group w/ outcome
Rel Risk Reduction v Absolute Risk Reduction
- Rel Risk Red (RRR)= (CER-EER)/CER
- *can exaggerate really small differences**
- Absolute Risk Reduction (ARR)= CER-EER
NNT v NNH
- NNT- number needed to treat - # needed to treat in order to see one positive outcome
- =1/ARR (in primary/good outcome)
- NNH - number needed to harm - # needed to treat before harmful event occurs
- =1/ARR (in safety or bad outcome)
CER v ICER
- CER- (cost effectiveness ratio)- cost/# bad outcome prevented OR # good outcomes promoted
- Compare new/protocol to no action at all
- ICER - (incremental cost effectiveness ratio) - (Costnew – Costcurrent)/(Effectnew – Effectcurrent)
- Compare new to current method
Forest Plot
Systemic Review
- Shows CI of ea study used plus box for ea study
- *Box width shows relative contribution of that study to overall summary stat
- Diamond = combined stat
- *Stat significant if it does not pass the vertical line of no difference
Funnel Plot
Systemic Review
- Used to evaluate publication bias (positive trials more likely to be published)
- Plot summary stat v sample size
- Want cluster around combined outcome stat AND equal distribution on both sides
