Simple -> Multiple Regression Flashcards
SIMPLE REGRESSION
R^2 - GOODNESS OF FIT
- for simple regression, R^2 = square of correlation coefficient
- reflects variance accounted for in data by best fit line
- takes values between 0 (0%)/1 (100%)
- frequently expressed as percentage > decimal
- high values show good fit; low values show poor fit
LOW R^2 VALUES
- R^2 = 0
- 0% = randomly scattered points; no clear relation between X/Y
- implies that a bets-fit line will be v poor description of data
- aka. good best fit line = high proportion of variance explained
HIGH R^2 VALUES
- R^2 = 1 (impossible!)
- 100% = points lie directly on line; perfect relation between X/Y
- implies best fit line = v good description of data
- aka. moderate best fit line = less variance explained
HIGH R^2 VALUES
- R^2 = 1 (impossible!)
- 100% = points lie directly on line; perfect relation between X/Y
- implies best fit line = v good description of data
SIGNIFICANCE TESTS
- reported in SPSS output
SIMPLE REGRESSION - t-test; established if model describes statsig proportion of variance in data
MULTIPLE REGRESSION - uses ANOVA to discover if proportion of variance in data explained by model = statsig
MULTIPLE REGRESSION
R^2 - GOODNESS OF FIT
- R^2 will get larger every time another IV (regressor/predictor) is added to model
- new regressors may only provide small improvement in amount of variance in data explained by model
- need to establish “added value” of each additional regressor in predicting DV
EFFECTIVENESS (VS EFFICIENCY)
- all possible contributory causes (IVs); maximises R^2
- explains largest possible proportion of variance in DV/outcome
- maximises R^2 (ie. maximises proportion of variance explained by model)
EFFICIENCY (VS EFFICTIVENESS)
- only includes most important variables (IVs)
- gives largest step increase in R^2ADJ
- maximises increase in R^2ADJ upon adding another regressor (ie. if new regressor doesn’t add much to variance explained it isn’t worth adding)
EFFECTIVENESS SCALE
0-25%
- v poor; likely unacceptable
25-50%
- poor BUT may be acceptable
50-75%
- good
75-90%
- very good
90%
- likely there’s something wrong w/analysis
ARE REGRESSORS STATIG ASSOCIATED W/DV?
- ANOVA test checks if model (as whole) has statsig relation w/DV
- part of predictive value of each regressor may be shared by 1/+ of other regressors in model
- aka. model must be considered as whole (ie. all IVs together)
- read off ANOVA table in SPSS output; report as in ANOVA (ie. F (3,12) = 4.33; p = .028)
INDIVIDUAL REGRESSORS x DV
- SPSS output entitled COEFFICIENTS
- column headed “un-standardised coefficients - B” gives regression coefficient for each regressor variable (IV)
- units of coefficient = same as for regressor (IV)
- all other variables must be held CONSTANT
REGRESSOR W/GREATEST EFFECT ON DV
- units for each regression coefficient = dif aka. we must standardise them to compare
- column headed “standardised coefficients - beta”
- can compare beta weights for each regressor variable to compare effects of each of DVs
- larger beta weight indicates stronger effect of regressor on DV values
ARE REGRESSOR RELATIONS W/DV STATSIG?
- assessed using t-test
- check values in column headed t/sig
- if regression coefficient = negative -> t-value will also be negative (doesn’t matter about sign; t size = important)