Appsumo reveals its A/B testing secret: only 1 out of 8 tests produce results | Hacker News Reader

Appsumo reveals its A/B testing secret: only 1 out of 8 tests produce results

visualwebsiteoptimizer.com

44 pointsparaschopra26 comments

This is roughly my experience, too -- take a look at how many null results ("no significant difference") I rack up: http://www.bingocardcreator.com/abingo/results

To add insult to injury, half of the remainder were significant... in the wrong direction.

(Edit to add: null results aren't failures, though. What is the Edison quote: you now know one more thing that didn't work.)

Yup, same experience with me too (while observing customers' tests and doing my own A/B tests). Noah provided great counter-examples. I mostly publish only positive A/B test results as case studies but this doesn't mean every A/B test produces spectacular results.

I wouldn't call them failed tests. In science a null result mean that you didn't learn anything but in this case who cares if p is 0.6? It is still more likely to make you money.

I think precisely the point of not being able to reject the null hypothesis is that you can't be sure it's more likely to make you more money; otherwise you could have required a lower confidence level and still rejected the null hypothesis.

The thing about non-statistically-significant tests is you can't, in fact, say "it is still more likely to make you money." You can say that the B set did in fact make you more money this time, but you can't rule out that being luck of the draw. It might even have a (very small) bias AGAINST making you money, but that bias got lost in the noise. That's why it's important to discard non-statistically-significant tests (or expand the sample size if you _really_ want to know...)

Testing involves paying a price in example counts in order to buy analytical power at some unknown rate. Insignificant results arise when either (1) people react with a lot of variance and that power/example rate is small or (2) the effect you're trying to analyze is so tiny that you need a great deal of power to see it at all.

Unfortunately, it's difficult to determine the difference between those two cases. If many of your tests are failing to be significant, it could be that you're simply never investing enough to get the power you need: your users are somewhat inflexible to the design changes you're making.

It could also be that many of your changes simply are pretty insignificant and by the time you wait a full year gathering simply infinite numbers of impressions, you'll find it was just a waste of time.

Again, practice statistics with a great deal of self-awareness. They're only meant to inform.

These aren't failures, if the A/B test shows no impact or that the change is a negative, then both of those add value in telling you what to concentrate on.

Before people did A/B test, it was generally assumed on Facebook ads that the title was the second most important factor on CTR so people spent a lot of time tuning title copy.

After lots of people did A/B testing it turns out that title copy has almost no impact, people switched their title to Chinese (for a non-chinese speaking audience) without seeing any change. That "no significant difference" means that thousands of man hours can be spent elsewhere rather than copywriting ad titles.

I'd be wary of concluding too much from other people's A/B tests. Yes, they serve as a good starting point but doing A/B tests yourself is still imperative. Case in point: Noah found out headlines didn't matter to his audience. But for many of our customers, a small change in headline does produce significant results.

Sure, I wasn't saying that the headline results were generalizable, but "no significant difference" A/B testing results can tell you to stop wasting time in an area.

A lot of these "no results" are just bad statistics. When a test shows no result it does not mean there is no significant difference. It means that the test cannot reliably tell given the available data. Technically, this is because as you decrease the probability of a type I error (lower the p-value) you increase the probability of a type II error. To lower the probability of a type II error, and hence have null results actually mean there is no significant difference, you needs lots of hits. Under assumptions that are reasonable for most e-commerce sites (e.g. 5% conversion, you want to distinguish changes of 5%) you need about 100'000 hits. Very few people have this much data, and hence their inferences are flawed. You can do a lot better than A/B testing with better maths. More here:

http://untyped.com/untyping/2011/02/11/stop-ab-testing-and-m...

I assume people who have run these "no results" A/B tests only gave up after enough traffic to establish no significant result was found given pre-determined X power and Y statistical significance cutoff. We have a calculator here to help determine what is the maximum number of visitors you need before giving up on a test: http://visualwebsiteoptimizer.com/ab-split-test-duration/

That's a pretty big assumption to make. I've seen people doing A/B testing w/ a few thousand hits max, which is not enough for most scenarios.

What's the big assumption? We are using math here, not heuristics. There is no thumb rule like thousands and hundreds of visitors needed to get significant results.

I was questioning this assumption of yours: "I assume people who have run these "no results" A/B tests only gave up after enough traffic to establish no significant result was found given pre-determined X power and Y statistical significance cutoff".

I am familiar with the maths for A/B testing. The figure of approximately 100'000 hits comes from "Controlled experiments on the web: survey and practical guide". As I stated, this assumes 5% conversion rate and a few other things. Here's the quote:

"If, however, you were only looking for 5% change in conversion rate (not revenue), a lower variability OEC [Overall Evaluation Criteria] based on point 3.b can be used. Purchase, a conversion event, is modeled as a Bernoulli trial with p = 0.05 being the probability of a purchase. The standard deviation of a Bernoulli is √p(1− p) and thus you will need less than 122,000 users to achieve the desired power based on 16 ∗ (0.05 · (1−0.05))/(0.05 · 0.05)^2."

(The actual value is 121'600.)

122,000 seemed absurdly large to me so I just looked up the paper you reference [1]. It looks like there is a calculation error in the paper.

The formula they use is n = 16 σ^2 / Δ^2. σ is the standard deviation, and Δ is the size of the difference. Thus in this problem, Δ = 0.05 and their formula gives n = 16 * (0.05 * (1-0.05)) / 0.05^2 = 304. This is much more in line with what you get from using a 2 sample proportion test (with H_a: p_1 =/= p_2), ~440 in each group [2].

But maybe I misunderstand their formula.

[1] http://exp-platform.com/hippo_long.aspx

[2] http://statpages.org/proppowr.html

edit: fixed Greek letters and added final comment.

Interesting -- I'll check it out. (My working day has almost finished, so this thread will probably be dead before I get back it.)

I think you misunderstand what they're doing with the formula. Delta is the size of the difference you want to detect. In this case, they're saying they want to detect a 5% (relative) change in a 5% (absolute) conversion rate. So σ^2 = 0.05 . (1-0.05), but Δ is 5% of 5% or 0.0025 and the denominator needs to be 0.0025^2.

If you go to your reference [2] and enter the numbers 0.05, 80, 0.05, 0.0525, 1.0, you'll see that they come up with a sample size of about 122k in each group (so 244k in both together).

(The figure of 304 or 440 is what you would get if you wanted to detect an absolute change of 5% in the conversion rate: going from 5% to 0% or to 10%.)

You're right, reading the paper that way, they are trying to detect a change in convergence rate from 5% to 5.25%, I was confused by the use of % in two separate contexts in the same sentence. That being said, I think this is not a good argument against A/B testing.

Fair enough, it would take a very large sample (122K is close enough) to detect a change from 5% to 5.25%. Being concerned about a change that small seems really silly unless 0.0025 * N visitors * revenue per user is a big enough number to be concerned with. I contend it won't be unless either

(1) N visitors is very large or

(2) revenue per user is very large.

If (1) is true, then testing on 122K users is not a big deal. If (2) is true you probably want to have a much more targeted approach, like someone doing sales.

Fourteen relative changes of 5% will double your conversion rate. Or halve it. The cost of an A/B test is small enough that if it took, say, a sample size of 1000 then it would be well worth A/B testing changes that might make a 5% relative difference. On the other hand, if it takes a sample of 122k then indeed you might well decide not to bother -- e.g., because it might be impossible. Which is why "it takes 122k rather than 1k to tell with any confidence" is interesting.

(Rough numbers: suppose you get 1000 visitors per day and convert at 5%, and suppose each conversion is worth $10 to you. Then you're bringing in about $180k/year from them, and a relative change of 5% in that is about $9k. Seems worth doing a modestly-sized A/B test for, but if it takes 4 months then you might reasonably decide to spend your effort elsewhere. (Or, of course, not: the actual cost of doing the test is rather small. But a lot can change over 4 months.)

Every test produces results. Only 1 in 8 produce positive results.

Knowing that 2 different approaches are approximately the same is a great help in design... It eliminates that 'Which way is better?' worry and lets you design it the way that looks best, instead.

Yep, it is true that even insignificant results can help you concentrating on how to make it look better. But insignificant results are disappointment if your goal is to increase conversion rate.

They might also want to test a link to an FAQ or Details page. I was trying to see if this was only for iPhone or Android but there's really only one option : give us your email.

Good point. It's mostly for web or client (Mac / Windows). Might do deals on phone apps in the future. http://appsumo.com/faq

I'm not sure these examples are entirely failed tests - rather they teach what is working and what isn't.

That's really true. It's funny that some of the changes we weren't expecting to make much of an impact have had the biggest results.

Negative results are results too.

Reply on news.ycombinator.com