1/57
Vocabulary practice flashcards covering bivariate correlation, linear and multiple regression diagnostics, ANOVA, and data transformations from the Exam 2 Study Guide.
what is line of best fit? what makes it fit best
A line that is closest to the most data points on a scatterplot, minimizes prediction errors/residuals/sum of squared errors, and must pass through points (0,a) and (X,Y).
Covariance
The statistical basis for correlation, reflecting the degree to which 2 variables vary together; it is positive when both scores are above the mean, and negative when one score is above and the other is below the mean.
Adjusted r
A statistic used to obtain a more unbiased estimate of the population correlation coefficient by penalizing degrees of freedom; the difference between r and adjusted r is substantial for small samples and negligible for large samples.
when to use Point-biserial correlation? interpretation
A correlation appropriate when one variable is truly dichotomous and the other is approximately continuous, interpreted in the same manner as a Pearson correlation.
Interpretation: A ——- correlation revealed that there were no sex differences in physical health, r(df)_pb=___, p= __.
Spearman correlation (rs) definition, interpretation, and drawbacks
A correlation coefficient used when both variables are ordinal (ranked data)
interpretation: a ____ correlation revealed that people who rank their own health as good will also rank their health as better than their same sex peers, rs(df)=__, p=__.
its drawbacks include the inability to calculate confidence intervals or standard errors for small samples, and that data cannot be normally distributed.
Kendall's Tau and interpretation
A rank-based correlation calculated from the number of inversions and reversals in rankings; preferred over Spearman correlation for small datasets containing a large number of tied ranks, better estimate of correlation in the population than spearman
Interpretation: People who rank their own health as good also tend to rank their health as better than same age/sex peers, Kendall’s tau=__ , p= __. no df and no effect size here
describe the components of a regression equation
Y (hat) = bx +a
Y hat is the predicted value of Y
b is slope of the regression line (amount of difference in Y hat associated with a 1 unit difference in X)
a is the intercept (value if Y hat when X=0)
X is the value of the predictor variable
goal is to find the line of best fit
What are the standardized regression coefficients?
the difference in 1 unit in X or Y represents a difference of 1SD (unitless regression estimate calculated when data are converted into z scores)
b becomes Beta if the data are standardized
r is the slope if variables are standardized
Standardized regression coefficient (β)
A unitless regression estimate calculated after converting data into z-scores, where a 1-unit difference in X or Y represents a difference of 1 SD.
what affects accuracy of prediction?
sum of squares (sum of squared deviations from the predicted point)
standard error of the estimate (sd of Y predicted from X)
residuals/error of variance= (SE of estimate) squared
df = N-2 ( lose 1 for a, lose 1 for b)
inverse relationship between r and the standard error of the estimate
r squared
Standard error of the estimate
The standard deviation of Y predicted from X, calculated as dfSum of Squares; it has an inverse relationship with the correlation coefficient r.
Coefficient of determination (r2)
A measure of predictive variability that indicates the variance in the dependent variable explained by the regression model; does not imply causation; like eta squared for correlations
describe the assumptions for correlation and regression
homogeneity of variance in arrays (variance of Y for each value of X is constant)
normality of arrays (pop values of Y corresponding to any value of X are normally distributed around Y hat)
Homogeneity of variance in arrays
An assumption of bivariate and multiple regression stating that the variance of Y across each specific value of X remains constant.
Normality of arrays
The regression assumption specifying that the population values of Y corresponding to any particular value of X are normally distributed around Y^.
null hypothesis for correlation and regression
H0: ρ = 0
means that there is no relationship between X and Y in the population
assumptions/factors that affect correlation
range restrictions: looking at smaller parts of the data, results in variance when one or both variables are low, results in lowered correlation
heterogenous subsamples: potential 3rd variable problem, overall correlation might be artificially inflated/deflated because of 2 subsamples
covariance
r (pearson product-moment correlation coefficient)
steps to remember for correlation
graph variables first
make sure there is variability
make sure variables are normally distributed
are there outliers? do they have undue influence on the results?
benefits of multiple regression
can explain more variance in the DV
can compare predictors
correlation matrix as it relates to multiple regression
the slope of each predictor represents its contribution to the DV
must graph variables, correlate them, then run the regression
can detect indicators of suppression by comparing correlation to regression coefficients
Collinearity
The presence of relatively high interrelationships among predictor variables in a multiple regression model, which can be a problem because of non-uniqueness and difficulty in explaining unique variance.
Multiple correlation coefficient (R)
A statistic representing the overall linear correlation between the criterion variable (DV) and all predictor variables (IVs) combined.
R squared
amount of variance in DV accounted for by all the IVs
different from R because R is how all variables correlate together (IVs and DV)
regression analysis interpretation
Interpretation: controlling for age (Beta= -.05, p = .426) and self rated health (Beta= .11, p= .063), people who have completed more education score significantly better on the BDS than people with less education (Beta = .22, p<.001). Age, education, and self rated health account for 8% of the variance in BDS scores.
pick standardized slope that best describes overall for last, control for non-sig vars then focus on how R squared is the IVs account for DV
How are regression coefficients and R squared tested for statistical significance?
b and Beta: linear transformations of each other, so if one is significant, the other is
SE: less SE, more likely to be sig
Use t test
R squared: rejecting the null if sig predicts at better than chance levels, F explains if it is different than 0
distribution assumptions of multiple regression
multivariate, normal distribution (departures from normal are okay bc tests are robust)
conditional distributions of Y (value of Y for specific value of X) are not independently distributed (moderate departures okay)
issues surrounding sample size for multiple regression
At least 10 people per IV
N should be bigger than p by at least 50, by at least 40
use power as a basis
MORE IS BETTER
what makes b and Beta partial regression coefficients?
unique effect of predictor on DV when other IVs are partialled out
Partial correlation definition and interpretation
The correlation between two variables after the influence of one or more other variables has been controlled for and removed from both X and Y.
recognize it bc there’s a column of control variables (variables taken out of the model)
Interpretation: A partial correlation revealed that age and stress were unrelated when negative mood was completely removed, r(df)= , p= __.
Semipartial correlation (Part correlation) definition and interpretation
The correlation between the dependent variable and a predictor variable when other predictors have been partialled out of only that single predictor variable.
recognize by three columns at the end with zero-order, partial, and part(look at part for r)
interpretation: A ____ correlation revealed that age and stress were unrelated when negative mood was removed from age r(df)= p= .
Suppressor variable
A predictor variable that increases the predictive variability of another predictor; indicated when its regression coefficient is larger in magnitude or opposite in sign compared to its simple bivariate correlation.
statistics for regression diagnostics and 2 sentences to describe them
Cook’s D, residual, studentized residual, centered leverage values
if all of them are small, no problem with outliers
Tolerance and VIF: if VIF is larger than tolerance, suggests potential problem among predictors because they are overlapping more than they should be (multicollinearity)
There doesn’t appear to be a problem with outliers. The results between tolerance and VIF suggest that there is a problem among predictors because they are overlapping more than they should be.
Cook's Distance (Cook's D)
A regression diagnostic statistic measuring the influence of an individual data point on the model; a value greater than 1 indicates a problematic case.
residual
measure of distance, smaller is better
Studentized residual
A t-test on the magnitude of each residual value used as a diagnostic tool; statistical significance indicates a potentially problematic outlier.
centered leverage values
measure of leverage; smaller is better
Tolerance
A collinearity diagnostic indicating the degree to which one independent variable can be predicted by other independent variables; values near 0 indicate severe collinearity problems.
Variance Inflation Factor (VIF)
The reciprocal of tolerance (Tolerance1) that measures predictor overlap; values significantly larger than tolerance indicate multicollinearity among predictors.
All subsets regression
program chooses best predictors based on your specifications, can include irrelevant variables but good at looking at all predictors
Backward elimination and drawbacks
A regression model selection procedure that begins with all predictors in the model and iteratively eliminates the variables that contribute the least based on partial or semipartial correlations.
capitalizes on chance, can drop good predictors that might not fit this model
Stepwise regression
An automated regression method that starts with no predictors and enters variables sequentially based on the highest semipartial correlation and predefined entry criteria, used primarily when no guiding theory exists.
Listwise deletion and drawback
A missing data handling method where an entire case is excluded from analysis if it is missing a value on any variable in the model, which can result in a substantially lower sample size (N).
Pairwise deletion and drawback
A missing data approach where all available data are used to calculate each correlation separately, which can lead to an unrepresentative correlation matrix.
benefits of Analysis of Variance (ANOVA)
compare more than 2 means
compare 2+ IVs simultaneously
more common in experimental designs
categorical IVs and 1 continuous DV (regression all continuous)
conceptually same as regression but all IVs must bc categorical
assumptions of ANOVA
homogeneity of variance
normality of DV scores
independence of observations'
null hyp is that all means are equal
logic of ANOVA
compares variance estimates and sum of squares total is the SS treatment plus SS error
interpretation of ANOVA with 1 IV
The three groups of sex control do not have the same health ratings, F (2, 291) = 4.63, p= .01, Eta squared= .03
OR
Sex control is related to health ratings F (2, 291) = 4.63, p= .01, Eta squared= .03
how are ANOVA df determined?
total: N-1
treat: k-1 ( k is # of treatments)
df error= df total - df treat
assumptions of ANOVA sample sizes
assumed to be equal, and if not equal, the larger group gets weighted more in SS treat calculation
Eta squared (η2)
An effect size statistic in ANOVA representing the proportion of total variance explained by the treatment (SStotalSStreat); displayed as R2 in SPSS output.
Omega squared (ω2)
A conservative, less-biased ANOVA effect size calculated as SStotal+MSerrorSStreat−(k−1)MSerror; reported as Adjusted R2 in SPSS output.
what components are necessary to calculate power in ANOVA
N and effect size
phi = phi prime sqrt(n)
phi prime is effect size
power= 1- table entry
Logarithmic transformation
A nonlinear data transformation used to correct positively skewed distributions by changing interval distances between points, which cannot be applied if the data contain zero (0).
Square root transformation
A nonlinear transformation applied to reduce skewness specifically in count data, which can successfully be computed when values contain zero (0).
Reciprocal transformation
A nonlinear transformation used when extreme scores exist in the positive tail to reduce their disproportionate influence on the model.
Arcsine transformation
A nonlinear transformation applied to leptokurtic distributions to stretch out both tails relative to the center.
Trimmed sample
A method to reduce outlier influence by symmetrically removing an identical fixed percentage of extreme scores from both tails of the distribution.
interpretation for pearson correlation
interpretation: A ___ correlation revealed that people who score high on var X will also tend to score high on var Y, r(df)= ___, p =___. (positive example)