Correlational Data Flashcards
What is the problem with classic experimental designs in clinical neuropsychology?
We often can’t randomly assign people to groups; instead, we often take groups as they are, we cannot manipulate them –> gives us less control.
Which three correlational designs are there?
- Mediators and moderators
- Logistic Regression
- Outliers in correlational designs
Explain what moderators are and how they work
A moderator changes (strenghtens/weakens) an effect between the IV and DV; to assess effects on the IV on the DV, you need to investigate the interaction terms of the moderator x IV on the DV.
Example: what is the influence of NP-impairments on daily life functioning? Moderator: compensatory skills.
Explain what mediators are and how they work.
Mediators are variables that explain an effect between the IV and DV; when controlling for this variable, there is no significant effect between the IV and DV anymore.
Example: the effect of NP-impairment on impairment in daily life functioning exists through school underachievement.
when do you use logistic regression and how does it work?
You use logistic regression when the DV is categorical; you use the ‘odds ratio’ instead of B to analyze the effect of the IV on DV.
How do you interpret the odds ratio and which categorizations are used?
Odds ratio is the proportionate change in odds (=that the outcome occurs) given a one unit increase in the predictor.
Categorizations:
OR = 1: the predictor does not affect the odds of outcome
OR > 1: the predictor is associated with higher odds of outcome (=positive effect)
OR < 1: the predictor is associated with lower odds of outcome (= negative effect)
What does an OR of 0.5 say?
This means that a 1 point increase in, for example, a questionnaire is associated with a 50 times decrease in the chance of having an impairment (= outcome measure). The OR cannot be below 0!!
What is the reason that objective measures of ADHD don’t resemble subjective measures of ADHD?
They both measure different aspects of cognition; objective tests measure optimal performance (under “perfect” conditions), while subjective tests measure typical performance (daily life functioning).
What is multinominal logistic regression and how does it work?
Multinominal logistic regression is an extension of logistic regression; you use it when you have more than 2 groups (as is the case in binary LR).
In using MLR, you always work with a “reference category”, since you have more than 2 groups. This works as follows: using self-evaluation of cognitive functioning, you want to predict if someone has one cognitive impairment relative to no cognitive impairment.
Or: if someone has more than one cognitive impairment relative to no cognitive impairment.
What are the 3 assumptions of logistic regression?
- Linearity between the continuous predictors + the logit of the outcome variable
- Independence of errors: cases should not be related
- Multicollinearity of predictors
–> fewer assumptions than in linear regression!
How can feigned ADHD be detected with logistic regression?
By using questionnaire scores (such as the CAARS) to predict in which group someone belongs (HC, ADHD, naive and instructed simulators).
Where could you use hierarchical logistic regression for?
To see the added value of different predictors when some of these are already in the model + find out the incremental validity of adding possible new measurements to a already existent set of measuremnts
What are the 4 key statistical values in logistic regression?
- P-values > model significance
- B-values > the effects of the predictors on the criterion (aka outcome)
- Odds of effect of predictor on outcome (odds ratio)
- R-square > % of explained variance
Why are outliers in correlational analysis important in CNP?
Since we often work with small samples (w/ patients for example, often only small samples available), outliers can cause lots of bias; these are therefore important to consider.
What is the definition of an outlier?
An observation that seems inconsistent with the other data.
Which two kinds of correlations are there?
- Pearson correlation = used for linear relationship between x and y
- Spearman correlation (rho) = pearson correlation on ranked data; non-parametric
What are the 6 different trends with outliers and how do Pearson/Spearman handle this?
- Linear trend: both Pearson/Spearman can be applied
- Monotonic trend: it preserves the given order/it never increases or it never decreases > both Pearson (when linear) and Spearman can be applied
- One outlier: causes overestimation of the association > pearson more vulnerable than spearman
- Several outliers: pearson = distorted, spearman = not distorted
- Split data (aka no true association): both give false estimations, but pearson more strongly than spearman
- Masking: there is a true association, but this is masked by outliers in Pearson but not in Spearman correlation.
How can you detect outliers + how do you handle them?
Through plotting the data visually in a scatterplot and then inspect possible outliers and their influence.
The best way to handle outliers is through being transparant in how you deal with them; u can use non-parametric statstics, form subgroups, exlude them, as long as you mention what you do.