Stats test stuff

• Nominal scale (level of measurement of data)


  Categories eg (blue eyes, brown eyes or green eyes)

  But hardly measuring people with this scale - not determining who’s better or worse (or more or less). Just putting them in groups
• Ordinal scale


  Ordering or ranking (eg from first to last or highest to lowest)

  Often used w data where people rank their feelings or attitudes

  but provides no info about diffs in scores (how much better is the first person than the last?)
• Interval scale


  Differences of numbers (not just order of them) are meaningful

  Eg IQ tests use interval scale - can order them based off intelligence AND look at diffs (Ayanna with IQ 140 is as much more intelligent than Ash with IQ 120 as she is then Juan with IQ 100)

  But can’t talk ratios as there’s no meaningful 0: can’t say IQ 140 is twice as more as IQ 70
• Ratio scale


  Highest form of measurement: both intervals between scores and ratios between scores are meaningful. Because ratio scale has an absolute starting point (0)

  Eg height, weight, time

  Make sense that someone 6 feet tall is twice as tall as someone who’s 3 feet tall
• Null hypothesis

  (Ho). We use statistical test to calculate probability of null hypothesis being correct

  If stat test reveals it has a low chance of being correct then we can accept our experimental hypothesis (which would now be called alternate hypothesis- H1)
• Accept or reject null

  We usually reject null if it has 5% or less probability of being correct

  P < 0.05 (with the smaller or equal to symbol). Would then be ‘significant’ (in psych means - unlikely to be due to chance)
• can we prove anything?

  Even with ‘significant’ results can’t be certain that null isn’t correct (5% chance it is) so we say ‘suggest’ not ‘prove’ a hypothesis
• Type 1 error (1eniant)


  Incorrectly reject the null and accept the experimental

  Eg if we tossed a coin 1000 times and it landed on heads all times then we’d be inclined to believe there is only heads on the coin - this would be a type 1 error if we suggested our results significant
• How to reduce chance of type 1 error


  set a stricter significance level. Might be 0.001 (1 in 1000) or even stricter. Risk of type 1 would then be negligible (v v small)
• type 2 errors

  When we accept null hypothesis incorrectly

  Eg of p was 0.06 then would have to reject experimental (which still has 96% chance of being correct)

  So if testing for an important medicine would be more important not to make this error
• Stat test for correlation?

  Spearman’s Rho
• Stat test for independent groups?

  Mann Whitney
• Stat test for repeated measures?

  Wilcoxon (T-test)