Planned VS Post-Hoc Contrasts and Exam Details Flashcards
RECAP
- ANOVA is used to find if IV effects performance on a DV (ie. finding a needle in a haystack; WHERE specifically is the effect? between who?)
CONTRAST MEANINGS
- they’re like co-ordinates; here’s a needle (0, -1, 0, 0), while here there is none (0, 0, -2, -1)
- they enable you to hone in and ask detailed questions about result patterns
CONTRAST CLASSES
PLANNED/A PRIORI/PLANNED COMPARISONS: - designed before experimenter sees data UNPLANNED/POST-HOC/A POSTERIORI: - devised after experimenter sees data - difs must be much larger here than in planned; it is possible to find STATSIG if planned BUT always non-STATSIG if they only thought of looking at effect after
CC: PROCEDURES
PLANNED: - calculated via previous work UNPLANNED: - Newman/Keul's - Duncan's Multiple Range - Tukey's - Dunnett's - Bonferroni's - Scheffe's
CC-P: SCHEFFE’S
VIRTUES:
- easy calculations
- if Scheffe is STATSIG, so are the others, making them unnecessary
CC-P-S: PROCEDURE
- Select contrast weights and do as you would in a planned contrast.
- Use Scheffe’s code.
- Ignore Sig. column if results.
- Calculate Scheffe criterion (ie. 5xF(5, 12)
- Find F in ELE table (ie. 5x3.106=15.53)
- Calc F-ratio = STATSIG if computed value is >/= of Schaffe’s criterion (ie. 9.600<15.53 = NOT STATSIG! at post-hoc check)
CC-P-S: SCHEFFE’S CRITERION
(k-1) x Tabled value of F(k-1, error df)
- k = number of levels in examined factor
- error df = degrees of freedom of denominator of original ANOVA
CC-P-S: SCHEFFE’S CODE
UNIANOVA score BY group /contrast(group)=special(-1, 0, 1, 0, 0, 0) /METHOD = SSTYPE(3) /INTERCEPT = INCLUDE /CRITERIA = ALPHA(.05) /DESIGN = group
CONTRAST PARADOX
- EG contrast (9.600) is STATSIG if planned as comparison in advance (9.600>p.009) BUT not STATSIG if comparison was not planned (9.600<15.53)
- THIS MAKES SENSE! strange but a result can be both STATSIG/not STATSIG
CP: NULL HYPOTHESIS DISTRIBUTION
NHD = distribution of outcomes to be expected if there are no genuine differences between the groups
IMAGINE…
- if EG scores were imaginary (but groups the same), SPSS may show (say) F (5, 12) = 1.3
- if this were repeated 18 more times, say now its F (5, 12) = .690; now repeat x10,000
- most of x10,000 will be close to 1; 500 of them all will exceed 3.106; this is NHD
CP-NHD: EXPLANATION
- if results of real exp were randomly sampled from outcome pop (aka. no genuine effect) then only 5% chance of f-ratio exceeding 3.106
- BUT if dif is genuine, results wouldn’t be sampled from this pop but rather from the one skewed right, SO high f-values (>3.106) is much more likely
CP-NHD: IMPLICATIONS
- if exp has f-ratio of 4.640, outcome can be either:
1. no genuine dif between 6 groups; f-ratio = chance of rare outcome (<5% of random selection)
2. genuine dif between 6 groups; working from wrong start assumption assuming no dif; f-ratio = could easily reappear on bulk of repetitions - 1 forces assumption of rarity, so sensible to reject it; this means rejecting NHD and accepting experimental hypothesis (EH)
STATISTICAL ERRORS
- classical testing has inbuilt possibility of two error types in decision process: TYPE 1: - no dif pop sampled - f-ratio = higher via chance - reject NHD; accept EH TYPE 2: - genuine dif pop sampled - f-ratio = lower via chance - accept NHD; reject EH
SE: IMPLICATIONS
- T1 risked every time NHD rejected; built into testing procedure (one 5% criterion test = 5% T1 chance; twenty tests = higher T1 chance; at least once conclusion = misleading)
PLANNED/POST-HOC RELATIONSHIPS
- correction for both requires raised bar via using more stringent statistical criterion PLANNED: - single test; 5% T1 - comfortably low POST-HOC: - interesting pattern hunt; 5-15 tests - inflated T1 chance