Introduction to Analysis of Variance (ANOVA) Part 1

Transcription

1 Introducton to Analss of Varance (ANOVA) Part 1 Sngle factor The logc of Analss of Varance Is the varance explaned b the model >> than the resdual varance In regresson models Varance explaned b regresson model vs unexplaned varance In ANOVA models Varance explaned b Factors >> than unexplaned varance In common language s the varablt among treatments greater than varablt wthn treatments 1

2 ANOVA vs regresson One factor ANOVA: 1 contnuous response varable and 1 categorcal predctor varable (factor) Compare wth regresson: 1 contnuous response varable and 1 contnuous predctor varable Ams Measure relatve contrbuton of dfferent sources of varaton (factors or combnaton of factors) to total varaton n response varable Test hpotheses about group (treatment) populaton means for response varable

3 Termnolog Factor (predctor varable): usuall desgnated factor A number of levels/groups/treatments = p Number of replcates wthn each group n Each observaton: Data laout Factor level (group) 1 Replcates j j... j n n... n Sample means 1 Populaton means 1 Grand mean estmates 3

4 Tpes of predctors (factors) Fxed factor: all levels or groups of nterest are used n stud conclusons are restrcted to those groups Random factor: random sample of all groups of nterest are used n stud tpcall ndvdual groups are not of nterest conclusons extrapolate to all possble groups Lnear model Lnear model for 1 factor ANOVA: j = + + j where overall populaton mean effect of th treatment or group ( - ) j random or unexplaned error (varaton not explaned b treatment effects) 4

5 Compare wth regresson model = x + ntercept s replaced b slope s replaced b (treatment effect): predctor varable s categorcal rather than contnuous stll measures effect of predctor varable Datoms & heav metals Effect of heav metals on speces dverst of datoms n streams n Colorado Response varable: speces dverst of datoms Predctor varable: heav metal level categorcal wth 4 groups (background, low, medum, hgh) Replcates are statons 5

6 H 0 : 1 = = = Null hpothess No dfference between populaton group (treatment) means Mean speces dverst of datoms s same for 4 heav metals levels H 0 - fxed factor No effects of specfc groups (treatments) H 0 : 1 = = = = 0 where = - No effect of 4 heav metal levels on datom speces dverst Inference s onl to these 4 heav metals 6

7 Streams and datoms Does datom dverst var b stream? H 0 - random factor No varaton among means of all possble groups (treatments) H 0 : A = 0 : / 1 =0 where groups =1 to N (streams) are chosen randoml Test: No varaton n datom speces dverst between randoml chosen streams Inference s to all streams (wthn??? Regon) sampled b N number of streams 7

8 Basc assumpton of ANOVA (sngle factor) 1 = = = = = where = populaton varance of dependent varable ( ) n each group (ths s the wthn group varaton) Each group (or treatment) populaton has smlar varance homogenet of varance assumpton Parttonng varaton Varaton n response varable parttoned nto: varaton explaned b dfference among groups (or treatments) varaton not explaned (resdual varaton, wthn group) 8

9 Regresson: Analss of varance n Y ( ) Total varaton (Sum of Squares) n Y ( ) Varaton n Y explaned b regresson (SS Regresson ) ( ) Varaton n Y unexplaned b regresson (SS Resdual ) ˆ ( ) ( ) ( ˆ ) Y } } ( ˆ ) } ( ) ( ˆ ) least squares regresson lne x x X 9

10 ANOVA SS Total j ( ) SS Between groups + SS Wthn groups (Resdual) n( ) ( j ) Parttonng the Varance Group 10

11 Parttonng the Varance Group Parttonng the Varance ( ) ( ) ( ) j j 1 1 ) ( j ( ) 3 Wthn Groups Between Groups 1 3 Group 11

12 Parttonng the Varance ( j ) n ( ) ( j ) Wthn Group Group Parttonng the Varance ( j ) n ( ) ( j Between Groups n = 4 (n ths example) ) 1 ( ) Between Groups (n = 4) Group 1

13 Mean squares Average sum-of-squared devatons Degrees of freedom: number of components mnus 1 df total [pn-1] = df groups [p-1] + df resdual [p(n-1)] Mean square s a varance: SS dvded b df Source SS df MS Groups n ( ) p-1 Resdual ANOVA table ( ) j p(n-1) n ( ( ) ( p 1) j p( n 1) ) Total ( ) j pn-1 13

14 Treatments (= groups) explan nothng, e. SS Groups equals zero Replcate Group1 Group Group3 Group Mean Grand mean = 16.0 Treatments (= groups) explan everthng, e. SS Resdual equals zero Replcate Group1 Group Group3 Group Mean Grand mean =

15 Testng ANOVA H 0 Remember: Lnear model for 1 factor ANOVA: j = + + j and u u, where can be or All populaton group means the same 1 = = = a = Fxed factor: H 0 : 1 = = = = 0 Means that there s no varablt across a fxed set of group means (lmted nference) Random factor (A): H 0 : A = 0 Means that there s no varablt across all possble group means (broad nference) ANOVA table Source SS df MS F Groups n ( ) p-1 n ( ) MS g /MS res ( p 1) Resdual ( ) j p(n-1) ( j ) p( n 1) Total ( ) j pn-1 15

16 F-rato statstc F-rato statstc s rato of sample varances (.e. mean squares) Probablt dstrbuton of F-rato known dfferent dstrbutons dependng on df of varances If homogenet of varances holds, F- rato follows F dstrbuton F dstrbuton P(F) 3, 4 df F 16

17 Expected mean squares If factor s fxed and homogenet of varance assumpton holds: MS Groups estmates MS Resdual estmates n ( ) p 1 F rato = Ms groups MS Resdual Testng H 0 - fxed factor If H 0 s true: all s = 0 MS Groups and MS Resdual both estmate so F-rato 1 If H 0 s false: at least one 0 MS Groups estmates + treatment effects so F-rato > 1 MS Groups n ( ) p 1 MS Resdual F rato = Ms groups MS Resdual 17

18 If factor s fxed and homogenet of varance assumpton holds: MS Groups MS Resdual Expected n ( ) p 1 n Calculated ( ( p 1) ( ) j p( n 1) ) F rato = Ms groups MS Resdual Expected mean squares (random factor) If factor s random and homogenet of varance assumpton holds: MS Groups estmates MS Resdual estmates n A F rato = Ms groups MS Resdual 18

19 Testng H 0 - random factor If H 0 s true: A = 0 MS Groups and MS Resdual both estmate so F-rato 1 If H 0 s false: A > 0 MS Groups estmates plus added varance due to groups or treatments so F-rato > 1 MS Groups n A MS Resdual F rato = Ms groups MS Resdual If factor s random and homogenet of varance assumpton holds: MS Groups MS Resdual Expected n A n Calculated ( ( p 1) ( ) j p( n 1) ) F rato = Ms groups MS Resdual 19