Statistics/Research Flashcards

Question

Parametric regression anaylsis. What Why (goals) Assumptions

Answer 1

A regression analysis makes it possible to infer or predict a variable based on one or more other variables. I.e. what variables impact of salary (i.e. age, education, years working, gender etc.) Dependent variable = criterion (salary) Independent variable = predictors (age/education etc.) Used to achieve two goals: 1. Measurement of the influence of one or more variables on another variable. 2. Prediction of a variable by one or more other variables. Assumptions: 1. Constant variance for different values of dependent variable 2. Dependent variable is a linear combination of the independent variables 3. In theory low/no measurement error 4. No high collinearity (where two or more independent/predictor variables are closely related to one another)

Answer 2

Simple linear regression - dependent variable is a continuous variable -one independent/predictor variable Multiple linear regression - dependent variable is a continuous varibale - multiple independent variables

Answer 3

Simple logistic regression * Dependent variable is a categorical variable * One independent/predictor variable Multiple logistic regression * Dependent variable is a categorical variable * Multiple independent variables

Answer 4

Internal validity - when an observed difference between groups are related to the intervention tested in the trial and not to flaws in study design. The internal validity of a clinical trial is directly related to appropriate design, conduction, and reporting of the study. - ensure confounders have been addressed - Bias minimised The two main threats to internal validity are bias and random error. Bias hereby refers to a systematic error that leads to a systematic deviation of the results from the truth due to flaws in the design, conduction, or reporting of the trial. Typical sources of bias are flaws in collection, statistical analysis, or interpretation of study data. Consequently, the true difference between study groups may be either under- or overestimated.

Answer 5

the extent to which the results of an RCT can be generalized into clinical practice and the general population. internal validity of a study is the prerequisite of its external validity since incorrect data due to missing internal validity can, per se, not be applied to the general population. Even if internal validity is assumed, insufficient external validity may reduce the clinical relevance of an RCT.

Answer 6

Flaws in collection, statistical analysis, or interpretation of study data. Causes an under or overestimation of the true difference between study groups. Selection bias Recall bias Funding bias Reporting bias Detection bias Exclusion bias Attrition bias Observer bias Analytical bias

Answer 7

When the study group is not fully representative of the population intending to be studied. (Higher education, English speaking patient's more likely to join a clinical trial). Can also occur when recruiters selectively enrol paitents based on what the next treatment allocation is likely to be. (overcome using allocation concealment and radnomly vary block size and use large block sizes)

Answer 8

When people remember things incorrectly. Often an issue with case-control trials

Answer 9

If the studied is sponsored more likely (even subconsciously) to please sponsor. (overcome by doing studies without sponsorship, or registering clearly the outcome measures and SAP prior to commecning study thourgh trial databse and published protocol)

Answer 10

More likely to publish a clinical trial that shows a positive finding. (over come by registering clinical trials in a trials database prospectively, and journals publishing both positive and negative research)

Answer 11

When the act of looking for an effect erroneously gives the impression of a large effect. e.g. doctors are more likely to look for diabetes in obese patients than thinner patients. Can give you a skewed perception. (overcome with double blinding)

Answer 12

Excluding patients results in a non generalisable cohort (excluding non-english speakers) (can be overcome by not having overly restricting exclusion criteria)

Answer 13

Patients drop out - participants may withdraw consent, may be uncontactable, participants violate the study protocol or decline to continue treatment. (Can be overcome with ITT rather than PP analysis.)

Answer 14

Researcher believes they know the outcome. (overcome with blinding)

Answer 15

Alter analysis to ensure you get the answer you expect. (overcome with blinidng of statistician, publishing protocol and registrering clinical trial prior to commencing).

Answer 16

A variable, other than the independent variable, that may affect the dependent variable. Distortion in the measure of association due to a relationship between the study factor (exposure) and another variable which also influences the outcome. Example in a study looking at embryo transfer catheters and live birth rate. Female age will be a confounder.

Answer 17

Design phase: - randomisation (best way to control for both known and unknown confounders) - exclusion criteria (e.g. if smoking is a confounder then restrict the study to nonsmokers, limits generalisability of study) - matching (each pair of persons enrolled in a study are similar for one or a few characteristics) - stratification (association between exposure and outcome is examined within different levels (strata) of the confounding variable. Analysis phase: - stratified analysis - multivariable analysis

Answer 18

Survival time analysis is a group of statistical methods in which the variable studied is the time until an event occurs. Doesn’t need to have anything to do with actual survival. Examples: 1. Kaplan-Meier curve – tells you how likely it is that a something will last longer than a certain point in time. I.e. could be used for age of first child (can’t compare independent samples) 2. Log rank test – compares the distribution of the time until an occurs of two or more independent samples. (so can be used for hypothesis testing) 3. Cox regression (Cox proportional hazards model) – used to determine the influence of different variables on survival time. Could be used for outcome Time to pregnancy. Note – censoring of data very important for survival time analysis (i.e. if the event doesn’t occur in the study timeframe this data is censored)

Answer 19

Diagnostic tests can produce a binary result (yes/no) or a number on a continuous scale. EG DVT diagnosis ultrasound doppler - yes/no, D-Dimer is a continuous scale. Diagnostic tests that produced continuous results are often dichotomised as the outcome of interest is generally binary. Done by setting a threshold where the test is 'positive' or 'negative' (D-Dimer 500) Setting the correct threshold can be assisted through the use of ROC curves. ROC curve: First the sensitivities and specificities for different values of a continuous test measure are tabulated. This results in a list of various test values and the corresponding sensitivity and specificity of the test at that value. The graphical ROC curve is then produced by plotting sensitivity (TPR) on the y-axis and 1-specificity (FPR) on the x axis. A ROC curve that follows the diagonal line y=x produces true positives at the same rate as false positive (coin toss). Area under the curve (AUC) of a ROC is a global measure of the ability of a test to discriminate whether a specific condition is present or not. AUC 0.5 = no discriminatory ability (coin toss - true positive the same as false positive rate), AUC 1.0 is perfect discrimination. When selecting the optimal threshold you need to consider aims of the test. If you give equal weight to sensitivity and specificity then you would use the point at the top left most corner of the ROC curve (Youden Index).

Answer 20

- an independent risk factor for the outcome of interest. - associated with the study factor. -not caused by the study factor For a variable to be a confounder - - has to have an association with the dependent variable (female age has an association with live birth rate) - has to have an unequal distribution between the exposed and unexposed cohorts (the group with one embryo transfer catheter has a high mean age than the other) - can't be a causal link in the chain between exposre and outcome

Answer 21

Decide on type of study (superiority or non-inferiority). Decide on the desired effect size (the smaller the effect size the larger the required sample). Decide on the power measurement (1-Beta error) - usually at least 80%. Choose significance level (alpha value) usually 0.05 (95% CI).

Answer 22

Likelihood of testing positive in those with the disease Sn = TP/(TP+FN) Good screening test has high sensitivity. SNOUT = sensitivity rules out disease. Useful for excluding disease as the false neg rate is low in a highly sensitive test. I.e. FFN for PTB - if negative highly likely to be correct as false neg rate very low. So can be reassured won't labour. Independent of prevalence.

Answer 23

Likelihood of testing negative in those without disease. Sp = TN/(TN+FP) High specificity tests can be used for definitive diagnosis. SPIN = specificity rules in a disease. Useful for diagnosis as the false positive rate is low in a highly specific test (i.e. pregnancy test is a highly specific test - if positive the patient is almost certainly pregnant) Independent to prevalence.

Answer 24

Clinical relevance of a test PPV = the ability of a test to detect the presence of disease. How many people with a positive test have the disease. PPV = TP / (TP+FP) Directly related to prevalence. As the prevalence increases the PPV gets higher. (not good for rare diseases).

Answer 25

NPV: The ability of a test to detect the absence of disease. The number with a negative test who don't have the disease. NPV = TN/ (TN+FN) Inversely related to prevalence. As the prevalence of a disease increases the NPV falls.

Answer 26

Incidence is a measure of disease risk Number of new cases of a disease over a specific time period divided by no. persons at risk (often then multipled by 1000). Written as "400 cases per 1000 population per year" Every person in the demoninator has the ability to be in the numerator. (i.e. people with uterus removed can't be in the denominator for the incidence of uterine cancer)

Answer 27

Measure of disease burden Number of affected persons in the population/the number of all persons in that population. Prevalence = incidence x duration (determined by death rate and cure rate) Generally the shorter the duration of an illness the lower the prevalence. Good tool for distributing health resource

Answer 28

Risk - chance of a person getting a condition in certain period of time - need defined population, defined period of time and number of new cases (during that period of time). (basically the same as incidence). Rate - reported as cases per person years at risk. Odds - number of events divided by the number of non events.

Answer 29

Main measures of association in cohort studies. Also used in RCTs. Expresses how many times more (or less) likely an exposed person develops an outcome relative to an unexposed person. Relative risk = the risk in the exposed group divided by the risk in the unexposed group. RR = incidence of outcome with exposure / incidence of outcome without exposure. Interpretation: RR >1 = increased risk of outcome RR = 1 = no risk of outcome RR <1 = reduced risk of outcome I.e SGA Normal weight total Smoked 19 139 158 Did not smoke 53 798 851 Relative risk = risk of SGA in those who smoked / risk of SGA in those who didn’t smoke RR = 19/158 / 53/851 = 0.12/.062 = 1.9 Smoking during pregnancy nearly doubles your risk of SGA

Answer 30

Main measure of association in case-control studies. OR = odds that a case was exposed / odds that a control was exposed. OR = A/C / B/D or A x D / B x C Odds = Probability/ (1-probability) Equation different that RR as you don't use 'totals' so not using total exposed or unexposed anywhere in the equation How many times more likely the odds of finding an exposure in someone with disease is compared to finding the exposure in someone without the disease. Interpretation: OR >1 = increased frequency of exposure among cases OR = 1 = no change in frequency of exposure OR <1 = decreased frequency of exposure Lung cancer No Smoker 87 147 No 201 508 OR = 87/201 / 147/508 .43 / .29 = 1.49

Answer 31

Application to proceed with phase 1 trial must have: - Animal study data and toxicity (side effects that cause great harm) data -Manufacturing information -Clinical protocols (study plans) for studies to be conducted -Data from any prior human research Phase 1 (cohort) - 20-50 participants. - sometimes randomsied. - goal is to identify 1.safety 2. correct route and dose of medication 3. Preliminary data on response to treatment. Phase 2 -<100 - most are RCTs - goal to identify effectiveness compared to current treatment - safety profile and side effects still closely monitored. Phase 3 (Pivotol study) - RCTs involving hundreds to thousands of patients - control group either placebo or standard therapy. - intervention - new treatment. - used to assess safety and efficacy. FDA then approve use Phase IV trial = continues to study side effects and risks of a new treatment.

Answer 32

Cross-over design. RCTs where study groups “cross-over” part way through. Pros: Permits within person comparison Potential saving in sample size (usually half the sample size) Allows assessment of patient preference Cons: Condition has to be suitable (chronic and stable) Treatment effect has to be reversible and temporary Issues with period effect and carryover effect In infertility trials they are problematic because outcome is usually live birth and so that participant would exit study before crossover and this can bias treatment effect

Answer 33

Research summary that addresses a focused clinical question in a structured, reproducible manner. - Often (but not always) accompanied by a meta-analysis.

Answer 34

statistical pooling or aggregation of results from different studies providing a single estimate of effect.

Answer 35

Conventional – uses study level data – aggregate data from each study are combined Network – technique for comparing three or more interventions simultaneously in a single analysis by combining both direct and indirect evidence across a network of studies Individual participant data (IPD) – when data from every individual enrolled in each study are combined.

Answer 36

1. Credibility of the methods of the systematic review (quality of the conduct of the SR) 2. Degree of confidence in the estimates that the evidence warrants (quality

Answer 37

- Sensible clinical question - Exhaustive and comprehensive literature search - Demonstrates the reproducibility of the selection and assessment of studies - Presents results in a useful manner - Did the review address confidence in the estimates of effect

Answer 38

- Study design - Risk of bias - Precision and consistency of results - Whether results directly apply to the patient of interest - Likelihood of reporting bias

Answer 39

1. Formulate the question 2. Define the eligibility criteria for included studies a. PICO b. Study design (on RCTs or other studies) 3. Develop a priori hypotheses to explain heterogeneity 4. Conduct the search 5. Screen titles and abstracts 6. Review full text of possibly eligible studies 7. Assess risk of bias 7b. assess for trustworthiness 8. Abstract data 9. Meta-analysis (if being performed) a. Generate summary estimates and CIs b. Look for explanations of heterogeneity c. Rate confidence in estimates of effect

Answer 40

- Summary of all the available evidence - Include a greater range of patients - Can enhance confidence in results - If performed with a meta-analysis o Can improve precision of estimates o If inconsistency among studies – can explore why Dis - Produce estimates that are as reliable as the studies summarised - If the studies included are heterogeneous and at high risk of bias the results may be misleading. - Not all SR and MA will follow rigorous design and may produce biased results

Answer 41

- Graphical way of depicting results of a meta-analysis - Point estimate (mean) of each study represented by a square o Size of square proportional to weight of study - Confidence interval presented as a horizontal line. - Combined summary effect/pooled estimate, typically represented by a diamond o Lateral points of diamond representing the confidence or credible interval - Vertical line representing no effect is also plotted. o If CI overlap this line, it demonstrates that at the given level of confidence their effect sizes do not differ from no effect for the individual study. o If the points of the diamond overlap the line of no effect the overall meta-analysed result cannot be said to differ from no effect at the given level of confidence.

Answer 42

- A scatter plot that compares the precision and results of individual studies o How close the estimated intervention effect size is to the true effect size - Commonly used in meta-analysis to visually detect publication bias - Precision of the estimated intervention effect increases with the size of the study o Small study effect estimates will typically scatter more widely at the bottom of the graph, with the spread narrowing among larger studies as they are more precise and closer to the true effect. - Should only be used when at least 10 studies are included in a meta-analysis, because the power of the tests is lower when there are fewer studies. - Asymmetrical funnell plot usually indicates bias, indicates whether the summary estimate is likely over or underestimating the intervention effect. - Possible reasons for asymmetrical funnel plot o Non reporting bias o Poor methodological quality leading to exaggerated effects o True heterogeneity o Artefactual o Chance (usually when small number of studies

Answer 43

Grading of recommendations, assessment, development and evaluation. Includes; Risk of bias Imprecision Heterogeneity Publication bias Indirectness Separeted into High Moderate Low Very low

Answer 44

RCTs always start with high confidence. Observational studies always start with low confidence. Factors that can modify these ratings: Decreased confidence with – high risk of bias, inconsistency, imprecision, indirectness, concerns about publication bias Increase in confidence is uncommon but occurs primarily in observational studies when the effect size is large.

Answer 45

Risk of bias Many different ways to assess this, should look at trial design, conduct and reporting Cochrane risk of bias tool can be used: Random sequence generation (selection bias) Allocation concealment (selection bias) Blinding of participants and personnel (detection bias) Incomplete outcome data (attrition bias) Selective reporting (reporting bias) Other bias

Answer 46

Heterogeneity in study outcomes. Large differences in point estimates or Cis that do not overlap suggest that random error is an unlikely explanation of the different results. Can perform statistical evaluation of heterogeneity 1. Cochrane Q test a. Null hypothesis = underlying effect is the same in each of the studies (i.e. the RR in study 1 is the same as in study 2/3/4) i. A low P values means that random error is an unlikely explanation for the differences in results, thus decreasing confidence in a single summary estimate. (i.e. you want a high P value) 2. I2 statistic a. focuses on the magnitude of variability rather than statistical significance. b. I2 of 0% suggests chance explains variability in point estimates and a single summary estimate can be used. c. As I2 increases we become progressively less comfortable with unexplained variability in results. Heterogeneity is study outcomes can be attempted to be explained using subgroup analysis and test of interaction. Or through meta-regression. Often remains unexplained, and high inconsistency/heterogeneity would decrease confidence in evidence.

Answer 47

1. Systematic error (bias) 2. Random error (relates to precision)

Answer 48

- When sample size and number of events are small = imprecise - When sample size and number of events are large = precise Assessed using confidence intervals (the range of values within which the true effect plausibly lies)

Answer 49

Optimal evidence for decision making; 1. Research that directly compared the interventions in which we are interested 2. Evaluated the population in which we are interested 3. Measured outcomes important to patients If population, intervention or outcome in studies differs from those of interest, the evidence can be viewed as indirect.

Answer 50

1. Positive study 3x more likely to be reported than one showing no effect 2. When authors or study sponsors selectively report specific outcomes or analyses the term selective outcome reporting bias is used. 3. Can be identified in met-analysis by the use of the funnell plot with a gap in the plot indicating that studies may have been done and not published. 4. Can be overcome – by obtaining unpublished studies – looking at prospective trial registration

Answer 51

Preferred reporting items for systematic reviews and meta-analyses

Answer 52

Any kind of variability among studies in a systematic review may be termed heterogeneity. - Clinical heterogeneity/diversity = Variability in the participants, interventions and outcomes studied (PICOs for studies are different) - Methodological diversity/heterogeneity = variability in study design and risk of bias - Statistical heterogeneity = variability in the intervention effects being evaluated in the different studies, this is a consequence of clinical or methodological diversity, or both, among the studies. Statistical heterogeneity manifests itself in the observed intervention effects being more different from each other than one would expect due to random error (chance) alone. Clinical variation will lead to statisical heterogeneity if the intervention effect is affected by the factors that vary across studies; most obviously, the specific interventions or patient characteristics. Differences between studies in terms of methodological factors, such as use of blinding and concealment of allocation, or if there are differences between studies in the way the outcomes are defined and measured, may be expected to lead to differences in the observed intervention effects. Statistical heterogeneity associated solely with methodological diversity would indicate the studies suffer from different degrees of bias. Meta-analysis should only be considered when a group of studies is sufficiently homogeneous in terms of PICO to provide a meaningful summary. It is often appropriate to take a broader perspective in a meta-analysis than in a single clinical trial.

Answer 53

Random-effect model - accounts for variability both within and between studies, suitable when there is significant heterogeneity. Meta-regression - extends a meta-analysis by examining the relationship between study-level covariates and the effect size. Useful for exploring sources of heterogeneity. Sensitivity analysis - assess the robustness of the meta-analysis results by repeating the analysis using different assumptions or excluding certain studies. Can help determine if the results are sensitive to particular studies or assumptions. Subgroup analysis - analyses subgroups of studies that share similar characteristics to see if the effect size differs across these subgroups. Useful for exploring sources of heterogeneity.

Answer 54

Cochran's Q test - a statistical test to detecet presence of heterogeneity - provides a p-value indicating whether observed heterogeneity is greater than expected by chance. I2 statistic - describes the percentage of variability in effect estimates that is due to heterogeneity rather than sampling error. 25%, 50% 75% indicate low, moderate and high heterogeneity.

Answer 55

Artificial intelligence – Umbrella term - broad field of computer science focused on creating systems capable of performing tasks that would typically require human intelligence (includes deep learning and machine learning) Machine learning – developing algorithms that allow computers to learn from and make predictions or decisions based on data, instead of being programmed for every task, ML systems improve their performance as they are exposed to more data. Process of learning from experience without explicit programming Logistic regression is an example of machine learning. Deep learning – subset of machine learning that works on principles of human neural networks with many layers. Uses artificial neural networks to extract, process and predict information by learning from examples.

Statistics/Research Flashcards

(79 cards)