Ben. I’ve been enjoying these posts. Thanks to Maxim for sharing his sub stack and now getting me addicted to quite a few of these posts.
While you wish to discuss rates and prediction, couldn’t the causal problem that you described in your previous post and continuing here, be simplified to the following setting: the language learning problem is one where the inference is almost entirely dependent on one feature to which the predictor has access, and the depression problem to one where the prediction is really dependent on a lot of features but the predictor simply does not have access to some of those features (lazy doctors, silent patients, biased drug companies, and possibly even too many features such as what someone is eats everyday, could also contribute to lack of such data collection). So, it’s hard to know, if such features may be strongly correlated with the outcome. Absent such data, one may view this as a forced sparse prediction problem. The real issue is that while many of these could perhaps be controlled in the lab setting, it’s impossible to control in the real world unless doctors tell their patients to restrict everything else or notate to evaluate the efficacy of the drug. Curious to see where you got with this. :)
Ben. As someone who growing up in India, learned more than three languages, I do want to share that learning a language does not necessarily translate to speaking it well. This happens all the time in India; for example with southern Indians learning Hindi, but never traveling to the north where it is spoken the most, and hence never using it until much later in life when such an opportunity presents. Then they stumble because they may technically know the grammar but they never spoke it at ease. I also see it with my kid who is taking AP Spanish lit this year as a senior in high school, and can easily decode the language while reading or listening but speaking Spanish isn’t all that easy for her. The way you set up your problem seems to suggest that learning a language implies knowledge of the language enough to speak but I don’t believe that is essentially the case unless there is the other factor or feature of immersion. So, there may be more than one feature here . I don’t mean to drive too deep into this as I understand that you were trying to make distinction between two rather distinctly different inference problems. But there may well be hidden variables in every prediction problem but some a lot more than others. :)
Hi Lalitha, I 100% agree. What I hoped to say is that *proficiency* couldn't happen without the language. Of course, as you detail, people can take language lessons for years and never be able to converse with native speakers. It's more that the *positive* outcome can only be explained by the intervention. The *negative* outcome, as you describe, has many potential explanations.
Ben. I’ve been enjoying these posts. Thanks to Maxim for sharing his sub stack and now getting me addicted to quite a few of these posts.
While you wish to discuss rates and prediction, couldn’t the causal problem that you described in your previous post and continuing here, be simplified to the following setting: the language learning problem is one where the inference is almost entirely dependent on one feature to which the predictor has access, and the depression problem to one where the prediction is really dependent on a lot of features but the predictor simply does not have access to some of those features (lazy doctors, silent patients, biased drug companies, and possibly even too many features such as what someone is eats everyday, could also contribute to lack of such data collection). So, it’s hard to know, if such features may be strongly correlated with the outcome. Absent such data, one may view this as a forced sparse prediction problem. The real issue is that while many of these could perhaps be controlled in the lab setting, it’s impossible to control in the real world unless doctors tell their patients to restrict everything else or notate to evaluate the efficacy of the drug. Curious to see where you got with this. :)
Interesting. What's the single feature you have in mind for language learning?
Perhaps the number of years of language immersion measured by perhaps years of learning formally or related quantity.
Ben. As someone who growing up in India, learned more than three languages, I do want to share that learning a language does not necessarily translate to speaking it well. This happens all the time in India; for example with southern Indians learning Hindi, but never traveling to the north where it is spoken the most, and hence never using it until much later in life when such an opportunity presents. Then they stumble because they may technically know the grammar but they never spoke it at ease. I also see it with my kid who is taking AP Spanish lit this year as a senior in high school, and can easily decode the language while reading or listening but speaking Spanish isn’t all that easy for her. The way you set up your problem seems to suggest that learning a language implies knowledge of the language enough to speak but I don’t believe that is essentially the case unless there is the other factor or feature of immersion. So, there may be more than one feature here . I don’t mean to drive too deep into this as I understand that you were trying to make distinction between two rather distinctly different inference problems. But there may well be hidden variables in every prediction problem but some a lot more than others. :)
Hi Lalitha, I 100% agree. What I hoped to say is that *proficiency* couldn't happen without the language. Of course, as you detail, people can take language lessons for years and never be able to converse with native speakers. It's more that the *positive* outcome can only be explained by the intervention. The *negative* outcome, as you describe, has many potential explanations.
Fair enough. Works clearly in this null vs positive hypothesis case.