Seeking Alpha (again), page-52

Lol ... yes Briking I am confident that the Alpha exposure critique is fundamentally flawed. But no-one should make investment decisions on anything I say. My commentary does not involve a money back guarantee.

But just for the fun of it let me drill down on one of the attacks in the article.

What is a least squares mean (LSM)?

According to Alpha exposure:

“Least squares mean is a fancy calculation to be used on a patient population so small. When properly used, it controls for a covariate such as center of enrollment, patient age, etc. However, an impact of this is that it alters the weight of each data point. With only 16 patients in each arm in the trial, every patient's result is important and should be counted equally.”

Correct. Except for the obvious contradiction. According to Alpha exposure a LSM is for when patient populations are small. But not apparently in this case when with only 16 patients each should be counted equally.

Now take a description from google.

“This is a mean estimated from a linear model. In contrast, a raw or arithmetic mean is a simple average of your values, using no model. Least squares means are adjusted for other terms in the model (like covariates), and are less sensitive to missing data. Theoretically, they are better estimates of the true population mean.”
http://dawg.utk.edu/glossary/g_least_squares_means.htm

The key message is a model, adjusting for covariates and a better estimate.

Here is how Alpha Exposure provides a worked example of the concept.


We can note the following problems.
First there is no model – he simply shows a mean calculation that is very silly.
Second there are no covariates and no method for adding them.
Third it produces a terrible estimate that is clearly wrong.

So what are the correct LSM estimates for the data used in the example provided by Alpha Deal? Using STATA and an Anova model we produce:



The LSM are shown in the “margins” output. This is just another term that is commonly used to describe them.

What does this analysis show?

First we have estimated a model - to be precise an Analysis of Variance. Second we could easily add covariates to this model if we wanted to. Third it produces excellent estimates. The LSM estimate for the first group is 1. For the second group it is 9. Exactly as we think they should be.

There is one further rather compelling argument in favour of this approach. STATA is used by thousands of health researchers around the world. It is one of the most widely used statistical packages out there. Calculating LSMs following a linear analysis is bread and butter work. If we believe Alpha exposure there are lot of people who are going to have egg on the face because everyone has been doing this incorrectly for years.

Or alternatively Alpha exposure does not have the foggiest clue what he is talking about.

One last thing.

I agree 100% with the company that the Seeking alpha the article was distorted. I hope MSB come out punching.

But I think there is an obligation on MSB to grapple with this problem of communicating complex trial results to lay audiences. I don’t think they are there yet. This example exposes a gap between researchers who are skilled at knowing how to communicate in publications and at conferences and company IR types who put out positive marketing. There does need to be something in between.