What to use anova or regression model

[AdityaKir]

Hello,

I have the below data structure.

Dress_ID	Style	Price	Rating	Size	Season	NeckLine	SleeveLength	waiseline	Total Sales
1006032852	Sexy	Low	4.6	M	Summer	o-neck	sleevless	empire	2114
1212192089	Casual	Low	0	L	Summer	o-neck	Petal	natural	2160
1190380701	vintage	High	0	L	Automn	o-neck	full	natural	2206
966005983	Brief	Average	4.6	L	Spring	o-neck	full	natural	2252
876339541	cute	Low	4.5	M	Summer	o-neck	butterfly	natural	2298
1068332458	bohemian	Low	0	M	Summer	v-neck	sleevless	empire	2344
1220707172	Casual	Average	0	XL	Summer	o-neck	full	null	2390
1219677488	Novelty	Average	0	free	Automn	o-neck	short	natural	2436
1113094204	Flare	Average	0	free	Spring	v-neck	short	empire	2482
985292672	bohemian	Low	0	free	Summer	v-neck	sleevless	natural	2528
1117293701	party	Average	5	free	Summer	o-neck	full	natural	2574
898481530	Flare	Average	0	free	Spring	v-neck	short	null	2620

I have to find out the factors which are effecting the Total Sales variable. Since most of the variables are categorical or rather all are, is using Anova a better option then going for Regression?

Regards,

Aditya

[Reeza · Accepted solution]

Regression. ANOVA primarily deals with one variable at a time where regression does not and you can consider interaction terms in a regression model. 

Make sure you understand the assumptions for your regression model and how categorical values are handled. 

[SteveDenham]

I would agree with @Reeza, and go with categorical regression using PROC GLM.  Hopefully, there is much more data than this, as I see at least 18 different levels across the categorical variables, and only 12 results, so finding a least squares solution may be impossible for this small sample.

Steve Denham

[Ksharp]

I would not agree with Reeza.
Your data did not conform to Normal data,so not suited for ANOVA.But I would try Log-Linear Model.

data have;
infile cards expandtabs truncover;
input Dress_ID :$20.    Style :$20. Price :$20. Rating $ Size :$20.
Season  :$20. NeckLine  :$20. SleeveLength  :$20. waiseline :$20. TotalSales ;
cards;
1006032852  Sexy    Low 4.6 M   Summer  o-neck  sleevless empire   2114
1212192089  Casual  Low 0   L   Summer  o-neck  Petal   natural 2160
1190380701  vintage High    0   L   Automn  o-neck  full    natural 2206
966005983   Brief   Average 4.6 L   Spring  o-neck  full    natural 2252
876339541   cute    Low 4.5 M   Summer  o-neck  butterfly natural 2298
1068332458  bohemian    Low 0   M   Summer  v-neck  sleevless empire 2344
1220707172  Casual  Average 0   XL  Summer  o-neck  full    null 2390
1219677488  Novelty Average 0   free    Automn  o-neck  short   natural 2436
1113094204  Flare   Average 0   free    Spring  v-neck  short   empire  2482
985292672   bohemian    Low 0   free    Summer  v-neck  sleevless natural 2528
1117293701  party   Average 5   free    Summer  o-neck  full    natural 2574
898481530   Flare   Average 0   free    Spring  v-neck  short   null    2620
;
run;
proc catmod data=have;
weight totalsales;
model Dress_ID*Style*Price*Rating*Size*Season*NeckLine*SleeveLength*waiseline
=_response_
/ noparm pred=freq;
loglin Dress_ID Style Price Rating Size Season NeckLine SleeveLength waiseline;
quit;






Your data actually is count data,so you could try Poisson Regression,But I would doubt the result,
Since You have so many levels for each variable, especially for Dress_ID.
Which mean you have N Dress_ID and you have N obs, I could doubt if it could be convergent.

proc genmod data=have;
class Dress_ID Style Price Rating Size Season NeckLine SleeveLength waiseline;
model totalsales= Dress_ID Style Price Rating 
Size Season NeckLine SleeveLength waiseline / dist=poisson link=log;
run;

[Reeza]

@Ksharp I didn't recommend ANOVA 😉

[Ksharp]

Sure. My bad .

[Ksharp]

I suggest you remove Dress_ID from model if you don't take it as a influence variable.