7
$\begingroup$

Apologies if the question is unsuitable for this site. Please direct me to the appropriate place, and I will take this down.

I am a statistician, and I have been struggling to find a meaning to the existence of statistics as a discipline in today's world where everyone cares about big black box models applied to big datasets. Statistics traditionally has been based on generative models, assumed some structure in the data and has developed methods to extract structure and do inference.

However, today people just care about prediction. Nobody cares about inference, and perhaps rightly so, because inference always necessitates a generative framework. The models we study today are extremely complicated and it's not clear if there is any hope for theory.

Time and again, we have heard statements like statistics is the least important part of data science. It is kind of painful to hear this as I have been trained as a statistician, and I wonder what is the way forward.

Do you struggle with this? What are your views on this? What advice would you give to a budding statistician given the trends you observe today?

$\endgroup$
7
  • 1
    $\begingroup$ I agree @utobi. $\endgroup$ Mar 16, 2023 at 5:52
  • 1
    $\begingroup$ But I feel it is opinion-based but won't be haste to vote to close it for the same. $\endgroup$ Mar 16, 2023 at 5:54
  • 1
    $\begingroup$ Mods would see it. In any case, I have posted it in the chat too. @utobi. Flag might be used for the purpose to allure mods' attention. $\endgroup$ Mar 16, 2023 at 6:34
  • 2
    $\begingroup$ Although I am new to this channel, I find the question important to many readers. I think it is worth to keep it open. $\endgroup$ Mar 16, 2023 at 6:43
  • 2
    $\begingroup$ A question can be important in its own merit, but that doesn't necessarily deter it from being an opinion based question @RomkeBontekoe. $\endgroup$ Mar 16, 2023 at 7:45

2 Answers 2

8
$\begingroup$

Statisticians often work in some form of consultation. As you said, many people need to validate a certain hypothesis, in science, medicine, etc. A statistician can analyze the data, but more importantly set up the proper experimental design. Much of science is based on the ability to do certain inferences. As long as there is a need for science there will be a need for statisticians to assist the scientists in that process.

There is also the educational part to statistics. People who work in data analysis often need help from people who understand statistics better. You may argue that the algorithms do not require a lot of math, and while that might be true to an extent, those algorithms are impossible to understand without a background in math. It is always easy to look at the final answer and not how one reached that answer. Many of the libraries are used by people who have very limited statistical background. If you want to help those people, and educate them in how those libraries work, or possibly edit them for specific purposes, then you will need a strong statistics background.

$\endgroup$
2
  • $\begingroup$ +1 to both your paragraphs. Re the 1st one: studies in medicine and psychology get more and more complicated, to the point that more and more specialized statistical understanding is necessary to analyze them and not go on a wild goose chase. The recent "replicability crisis" in psychology highlights what happens when psychologists think they can do their analysis themselves. $\endgroup$ Mar 16, 2023 at 7:22
  • $\begingroup$ Re the 2nd paragraph: I have seen too many Machine Learners with no understanding of variability (which I would argue is something statisticians have a better innate feeling for) blindly trust their models. Thinking that just because we got a prediction, we don't need to understand the statistics around it is close to voodoo science. (But then, I'm active here, so what else would I say?) $\endgroup$ Mar 16, 2023 at 7:23
3
$\begingroup$

The role of the statistician continues to be essential. I would say that there is a shortage of people with good statistical thinking. Testing hypotheses is required in any scientific rigorous decision process.

Novel health interventions (such as drugs against cancer) could not be developed without the support of a statistician in clinical trials.

Big tech companies also conduct A/B testing whenever they want to implement a new feature. A/B testing is a clinical trial to evaluate whether a feature improves a performance metric.

More recently, the literature of prediction models (whether based on black-boxes or classical regression methods) has advocated that clinical trials should be conducted to evaluate the impact on clinical outcomes of the implementation of these predictive models in a clinical setting. So, ML scientists interact with statisticians to design trials and test hypotheses.

In the end of the day, science only moves forward with formulation of hypotheses, data collection and test of hypotheses.

$\endgroup$
3
  • 2
    $\begingroup$ "Big tech companies also conduct A/B testing" I have a question about this. Is this problem truly so complicated that they need statisticians to solve it? There are online applications where you simply input the sample sizes for A/B, and it automatically computes it for you. Am I missing something here? $\endgroup$ Mar 16, 2023 at 5:20
  • $\begingroup$ Indeed, performing a frequentist A/B test is often a t-test/Chi-square test. You could say the same about most of the clinical trials when no covariates are considered. However, how about the sample size? More recently, Bayesian A/B testing has been often discussed as an option because it might reduce significantly the sample size. Which priors should be chosen? Furthermore, how about the experimental design? Can we consider a factorial design to test two features and the impact of their interaction? Is selection bias an issue? How about regression to the mean? $\endgroup$ Mar 16, 2023 at 5:42
  • $\begingroup$ @NicolasBourbaki Aside from Márcio's points above, I would imagine that the harder part of running an A/B test or survey is generalising from your sample to the population of interest (and designing the data collection step so that this is feasible and efficient). $\endgroup$
    – mkt
    Mar 16, 2023 at 10:35

Not the answer you're looking for? Browse other questions tagged or ask your own question.