Chi Square Test. Paul E. Johnson Department of Political Science. 2 Center for Research Methods and Data Analysis, University of Kansas

Transcription

1 Chi Square Test Paul E. Johnson Department of Political Science 2 Center for Research Methods and Data Analysis, University of Kansas 2011

2 What is this Presentation? Chi Square Distribution Find Things to Compare With That

3 The χ 2 Distribution Describe accumulated deviations Example: Add up ν (Greek letter nu ) Standardized Normal Observations S ν = Z Z Z 2 ν For each ν, we expect a different probability distribution (obvious!)

4 probability of S Chi Square Probability Densities ν = 1 ν = 2 ν = 3 ν = 6 ν = 10 ν = Observed sum (S)

5 Leads to a natural one tailed test Many models have a predicted number and deviations in the form of a difference b/t observed and predicted The sum of squared deviations can be scaled onto the χ 2 distribution

6 Density Critical value of Χ 2 = Shaded: Prob(Χ 2 > 43.77)= A Chi Square Variable (ν = 30)

7 Check a Larger ν Density Critical value of Χ 2 = Shaded: Prob(Χ 2 > 341.4)= A Chi Square Variable (ν = 300)

8 Moments Expected Value of S: ν Variance of S: 2ν Standard Deviation of S: 2ν

9 ν has a Nickname ν is sometimes known as degrees of freedom. Conceptually, it is the number of unrestricted squares added together.

10 How To Use For Hypothesis Tests Wrestle some observed problem into a format that resembles a sum of squares Get a statistician to figure out a transformation that can be used to compare it with a χ 2.

11 Consider a table that is prepared in accordance with the Iron Law of Crosstabs Hair Juggling brown red yellow Totals great OK none Totals

12 You Make a Prediction For Each Cell Hair Juggling brown red yellow Totals great OK none Totals

13 Now Some Statistical Hocus Pocus Consider any Cell, row i and column j Here s my first guess. Square the mistakes: (Obs ij Pred ij ) 2 (1) Add across all cells (Obs ij Pred ij ) 2 (2) cells That s not quite right yet, because it is not a Normal variable it is not Standardized either

14 But You Can See Where This is Heading, Can t you? Suppose cells (Obs ij Pred ij ) 2 is very small. Then your predictions are pretty good. Don t reject them. Suppose cells (Obs ij Pred ij ) 2 is HUGE. Your predictions were bad. Reject the model on which you based them. We just need a way to scale this sum of squares.

15 They Call it Pearson s χ 2 Test Karl Pearson claimed that if we sum this quantity across cells (Obs ij Pred ij ) 2 Pred ij (3) we can compare the result against a χ 2 distribution. The Pearson Chi Square Statistic (suppose we call that X 2 ) X 2 = #rows i=1 #columns j=1 (Obs ij Pred ij ) 2 Pred ij (4) Pedantic. The number X 2 is not the same as the distribution χ 2. X 2 is some number Karl Pearson figured out that has a distribution similar to χ 2 under some conditions.

16 Start Approximating Pearson showed that X χ 2 v, but only asymptotically as the N of cases used to calculate Pred j1 tends to infinity. and only if the true probability that a case will land in row i, column j is greater than 0 What s the correct value of ν? A little bit tricky. Depends on how you calculate Pred ij

17 Why so Vague about ν (Here s my best guess) The ν depends on the amount of information you used from the data to make your predictions If you just put in your personal wild guesses, then use ν = r c If you make a silly prediction that each cell should have an equal proportion of all observed cases, then you are only using the N of cases to make your prediction, so ν = r c 1 If you do the more-or-less standard identically distributed columns prediction, then you use ν = (r 1)(c 1)

18 Standard Story about Identical Column Proportions The null hypothesis is that all columns are samples from identical random processes. Multinomial random variable assigns outcomes to row 1, 2, 3 with probability (p 1, p 2, p 3 ). Note: If only 2 rows, then we have a Binomial distribution (coin flips). We don t say what the p i might be, only that they the same for each column.

19 Standard Story (cont.) Use the observed proportions as estimates of (p 1, p 2 p 3 ). Suppose we call them ˆp 1, ˆp 2, ˆp 3. If each column were drawn from that multinomial distribution, we should predict each column s cells as a reflection of those probabilities. For a column witht j cases, we predict the cell counts are (T j ˆp 1, T j ˆp 2, T j ˆp 3 ), so call those predictions (Pred j1, Pred j2, Pred j3 ) (5) By estimating ˆp i from the data, we use up one degree of freedom for each row. By using T j from each column, we use up a degree of freedom for each column. Hence, the correct reference value for the X statistic is χ ν where ν = (r 1)(c 1)

20 Crosstabs I o p t i o n s ( width = 55) l i b r a r y ( gmodels ) C r o s s T a b l e ( i n f e r t $ e d u c a t i o n, i n f e r t $ induced, expected = TRUE, format = SPSS, chisq = T, f i s h e r = T, mcnemar = T) C e l l C o n t e n t s Count Expected V a l u e s Chi square c o n t r i b u t i o n Row P e r c e n t Column P e r c e n t T o t a l P e r c e n t Total O b s e r v a t i o n s i n Table : 248 i n f e r t $ i n d u c e d i n f e r t $ e d u c a t i o n Row T o t a l 0 5yrs % % % % % % % % % % 6 11yrs

21 Crosstabs II % % % % % % % % % % 12+ y r s % % % % % % % % % % Column T o t a l % % % S t a t i s t i c s f o r A l l Table F a c t o r s Pearson ' s Chi squared t e s t Chi 2 = d. f. = 4 p = McNemar ' s Chi squared t e s t

22 Crosstabs III Chi 2 = d. f. = 3 p = e 27 F i s h e r ' s Exact Test f o r Count Data A l t e r n a t i v e h y p o t h e s i s : t w o. s i d e d p = Minimum expected frequency : C e l l s with Expected Frequency < 5 : 2 of 9 ( %)