Chi-Square Test Mon, Apr 19 th, 2004
Chi-Square ( 2 ) wAre 2 categorical variables related (correlated) or independent of each other? wCompares # in categories that would be expected by chance (E) to # in categories actually observed (O) wNull hyp (Ho) – no relationship between 2 variables (they’re independent of ea.other) wAlternate (Ha) – the 2 variables are related (not independent)
2 Formula w 2 = [(fo – fe) 2] fe So we’ll compare observed and expected frequencies for each cell in the table…
Example wIs age ( 30) related to preference for analog/digital watches? DigitalAnalogUndecided Under Over
Step 1: Compute Marginals DigitalAnalogUndecided Under Over Marginals are the row and column totals:
Step 2: Compute Expected Frequencies (f E ) wf E = (Column marginal * Row marginal ) / N For people under 30: fe (digital) = 100 * 140 / 200 = 70 fe (analog) = 80*140 / 200 = 56 fe (undec) = 20*140 / 200 = 14 Over 30: fe (digital) = 100*60 / 200 = 30 fe (analog) = 80*60 / 200 = 24 fe (undec) = 20*60 / 200 = 6
Step 3: Compute X 2 wFind difference (residual) betw observed & expected for each cell (fo – fe) wSquare those differences wDivide squared differences by fe wSum the results
Summary FoFeFo-fe(fo-fe) 2 (fo-fe) 2 / fe <30 D <30 A <30 U >30 D >30A >30 U
(cont.) wLast step: Add up (fo-fe) 2 / fe w 2 = = wStep 4: Compare to 2 critical with df = (# columns – 1) (# rows – 1) wHere df = (2-1)(3-1) = 2 df, =.05, critical = 5.99
Hypothesis Test wIf 2 observed > 2 critical, reject Ho –Reject Ho conclude there is a relationship between the 2 variables –Here, > 5.99, reject Ho, there is a relationship between age & watch preference
In SPSS wAnalyze Descriptive Stats Crosstabs –Choose whichever variable you’d like for ‘row variable’ and the other for ‘column variable’ –Click “Statistics” button, and check chi- squared option –Click “Cells” Button, choose “expected count”
SPSS (cont.) wOutput – look for “Pearson chi-sq” and “Asymp Sig” column gives significance value for chi-sq test: –If “Asymp Sig” value is <.05 (alpha), reject Ho wNote there is an option for clustered graphing, read this example in the lab