topic Re: How to analyse factorial design with an outcome having two types of variance? in Discussions

How to analyse factorial design with an outcome having two types of variance?

gustavjung — Thu, 08 Jun 2023 21:03:33 GMT

Hello

I have a dataset from a 2-level factorial design with 2 factors.

The dataset contains purchase amount per each user.

Each row is a unique user. Purchase with 0 value means user didn't purchase. A and B are treatments.

I want to calculate the effect of treatments A and B on average purchase per user (sum of purchase in USD divided by number of users).

However, I assume we need to account for two types of variance.

First – the variance of the purchase rate, that is, did the user order or not. Second – the variance in average purchase value.

Can you suggest the appropriate method to do such an analysis in JMP?

I have attached the dataset.

Re: How to analyse factorial design with an outcome having two types of variance?

Mark_Bailey — Thu, 08 Apr 2021 13:29:28 GMT

You only have 265 purchases out of 15,771 observations. OLS regression is inappropriate in this case. It fails miserably and the model assumptions are grossly violated. I transformed the purchase amount to a binary response Purchased = { no, yes } and then modeled it with logistic regression. The diagnostics look bad. The model never predicted Purchased = yes. I tried Poisson log-linear model and it worked, but demonstrated very little evidence for an effect of A or B.

I attached your data table with my additions.

Re: How to analyse factorial design with an outcome having two types of variance?

gustavjung — Fri, 09 Apr 2021 19:06:27 GMT

Thank you very much!

I do understand that we can use logistic regression for binomial data. But what if we had a positive effect of A or B on purchased (Yes, No) and the outcome of interest was an average purchase amount per user. This outcome is different because this is a product of binary metric (Yes, No) and an average purchased amount of orders (with purchased = yes). We can have more success purchases, but low average purchased amount of orders, as a result we will get lower average purchase amount per user. In OFAT tests for this outcome we can use non-parametric Mann-Whitney U Test. But what about cases whet we want to do a regression analysis?