5 ms·

It's interesting to see that even on HN people don't know the difference between correlation and causation.

hkt 6y ago
I'd guess there's not much cause for it among computer science people. I've noticed this in the past: empirical disciplines are better at this stuff. Still, I can't quite work out if a biology degree would cause somebody to understand it better..
- oehtXRwMkIs 6y ago
  In my experience, most empirical majors beat you over the head with the chanting of "correlation does not imply causation" pretty much every term. It's too bad that it's one of those things that's not naturally intuitive to people.
- notafraudster 6y ago
  I mean one of the two dominant frameworks for causal inference is Judea Pearl's directed acyclic graphs / do-calculus, and Pearl is surely a computer scientist.
Dayshine 6y ago
Have you read the paper? They adjust for damn near everything including reverse causation. What confounders do you believe the UK biobank data can't account for?
- oehtXRwMkIs 6y ago
  It doesn't matter if you eliminate 10 million confounds. No matter how good your correlation study is, it cannot establish a causation. To do that you need random assignment at the very least.
  - ncmncm 6y ago
    ... and even that gives you the wrong answer if your diagnostic specificity is faulty. This last is why we keep seeing muddled reports that "antidepressants don't work". Diagnosis of "depression" is a snakepit; the only known way to discern which among at least six conditions you might have, all labeled "depression", is to see what medications work, if any. Imagine trying to conduct a plausible RCT for that. "Gold standard"? Ha.
    - oehtXRwMkIs 6y ago
      Did you reply to the wrong comment?
      - ncmncm 6y ago
        I replied to "...it cannot establish a causation. To do that you need random assignment at the very least."
        oehtXRwMkIs 6y ago
        Oh, your quoting of "gold standard" really threw me off.
  - 0d9eooo 6y ago
    10 million confounds is a lot to eliminate. Randomization only works because it decreases the probability of a spurious association existing. At some point the number of confounds eliminated would probably be more worthwhile. I'd definitely pay more attention to a study that eliminated 10 million confounds than most randomized controlled trials. I agree a randomized controlled trial would be nice now. But even that is potentially fraught. For example, someone else mentioned the possibility that glucosamine reduces joint pain, which increases mobility, which increases longevity. Randomizing wouldn't really control for that sort of scenario. And that's not even getting into preregistration, meta-analysis etc.
- tmoertel 6y ago
  Adjusting for the wrong confounders is one of the ways to create spurious correlations. For example, consider the causal system X -> Z <- Y in which X and Y are random coin tosses (heads=0, tails=1) and Z is their sum. Even though X and Y independent, they will be correlated if you control for Z. If you want to see it yourself, let's simulate this system in R: # Simulate 1000 trials in which X and Y are random coin tosses (T or F) # and Z is their sum. > X <- rnorm(1000) < 0.5 > Y <- rnorm(1000) < 0.5 > Z <- X + Y # As expected, X and Y are not correlated: > summary(lm(Y ~ X)) Call: lm(formula = Y ~ X) Residuals: Min 1Q Median 3Q Max -0.6947 -0.6893 0.3053 0.3108 0.3108 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 0.694704 0.025816 26.910 <2e-16 *** XTRUE -0.005455 0.031330 -0.174 0.862 <<< No correlation. --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 0.4625 on 998 degrees of freedom Multiple R-squared: 3.038e-05, Adjusted R-squared: -0.0009716 F-statistic: 0.03032 on 1 and 998 DF, p-value: 0.8618 # However, if we include Z as a possible confounder, we create a # spurious correlation between X and Y: > summary(lm(Y ~ X + Z)) Call: lm(formula = Y ~ X + Z) Residuals: Min 1Q Median 3Q Max -2.45e-14 -8.40e-18 2.04e-17 2.04e-17 3.59e-15 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) -3.370e-16 6.292e-17 -5.356e+00 1.05e-07 *** XTRUE -1.000e+00 8.241e-17 -1.213e+16 < 2e-16 *** <<< Oops! Z 1.000e+00 5.873e-17 1.703e+16 < 2e-16 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 8.582e-16 on 997 degrees of freedom Multiple R-squared: 1, Adjusted R-squared: 1 F-statistic: 1.449e+32 on 2 and 997 DF, p-value: < 2.2e-16
  - crdrost 6y ago
    I do not know why you are simulating this system. Given random x, y, controlling for z = x + y trivially generates a perfect anticorrelation as x = z – y. Like, I get that you want to make a point, just seems weird to have a numerical solution in 5 lines of code to a problem that was already theoretically solved when you wrote down the problem statement and therefore required exactly 0 equations...? Also like the theoretical solution clarifies that this happens for arbitrary probability distributions equally, it is not specific to Bernoulli random variables.

hkt 6y ago

I'd guess there's not much cause for it among computer science people. I've noticed this in the past: empirical disciplines are better at this stuff. Still, I can't quite work out if a biology degree would cause somebody to understand it better..

oehtXRwMkIs 6y ago
In my experience, most empirical majors beat you over the head with the chanting of "correlation does not imply causation" pretty much every term. It's too bad that it's one of those things that's not naturally intuitive to people.
notafraudster 6y ago
I mean one of the two dominant frameworks for causal inference is Judea Pearl's directed acyclic graphs / do-calculus, and Pearl is surely a computer scientist.

Dayshine 6y ago

Have you read the paper? They adjust for damn near everything including reverse causation. What confounders do you believe the UK biobank data can't account for?

oehtXRwMkIs 6y ago
It doesn't matter if you eliminate 10 million confounds. No matter how good your correlation study is, it cannot establish a causation. To do that you need random assignment at the very least.
- ncmncm 6y ago
  ... and even that gives you the wrong answer if your diagnostic specificity is faulty. This last is why we keep seeing muddled reports that "antidepressants don't work". Diagnosis of "depression" is a snakepit; the only known way to discern which among at least six conditions you might have, all labeled "depression", is to see what medications work, if any. Imagine trying to conduct a plausible RCT for that. "Gold standard"? Ha.
  - oehtXRwMkIs 6y ago
    Did you reply to the wrong comment?
    - ncmncm 6y ago
      I replied to "...it cannot establish a causation. To do that you need random assignment at the very least."
      - oehtXRwMkIs 6y ago
        Oh, your quoting of "gold standard" really threw me off.
- 0d9eooo 6y ago
  10 million confounds is a lot to eliminate. Randomization only works because it decreases the probability of a spurious association existing. At some point the number of confounds eliminated would probably be more worthwhile. I'd definitely pay more attention to a study that eliminated 10 million confounds than most randomized controlled trials. I agree a randomized controlled trial would be nice now. But even that is potentially fraught. For example, someone else mentioned the possibility that glucosamine reduces joint pain, which increases mobility, which increases longevity. Randomizing wouldn't really control for that sort of scenario. And that's not even getting into preregistration, meta-analysis etc.
tmoertel 6y ago
Adjusting for the wrong confounders is one of the ways to create spurious correlations. For example, consider the causal system X -> Z <- Y in which X and Y are random coin tosses (heads=0, tails=1) and Z is their sum. Even though X and Y independent, they will be correlated if you control for Z. If you want to see it yourself, let's simulate this system in R: # Simulate 1000 trials in which X and Y are random coin tosses (T or F) # and Z is their sum. > X <- rnorm(1000) < 0.5 > Y <- rnorm(1000) < 0.5 > Z <- X + Y # As expected, X and Y are not correlated: > summary(lm(Y ~ X)) Call: lm(formula = Y ~ X) Residuals: Min 1Q Median 3Q Max -0.6947 -0.6893 0.3053 0.3108 0.3108 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 0.694704 0.025816 26.910 <2e-16 *** XTRUE -0.005455 0.031330 -0.174 0.862 <<< No correlation. --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 0.4625 on 998 degrees of freedom Multiple R-squared: 3.038e-05, Adjusted R-squared: -0.0009716 F-statistic: 0.03032 on 1 and 998 DF, p-value: 0.8618 # However, if we include Z as a possible confounder, we create a # spurious correlation between X and Y: > summary(lm(Y ~ X + Z)) Call: lm(formula = Y ~ X + Z) Residuals: Min 1Q Median 3Q Max -2.45e-14 -8.40e-18 2.04e-17 2.04e-17 3.59e-15 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) -3.370e-16 6.292e-17 -5.356e+00 1.05e-07 *** XTRUE -1.000e+00 8.241e-17 -1.213e+16 < 2e-16 *** <<< Oops! Z 1.000e+00 5.873e-17 1.703e+16 < 2e-16 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 8.582e-16 on 997 degrees of freedom Multiple R-squared: 1, Adjusted R-squared: 1 F-statistic: 1.449e+32 on 2 and 997 DF, p-value: < 2.2e-16
- crdrost 6y ago
  I do not know why you are simulating this system. Given random x, y, controlling for z = x + y trivially generates a perfect anticorrelation as x = z – y. Like, I get that you want to make a point, just seems weird to have a numerical solution in 5 lines of code to a problem that was already theoretically solved when you wrote down the problem statement and therefore required exactly 0 equations...? Also like the theoretical solution clarifies that this happens for arbitrary probability distributions equally, it is not specific to Bernoulli random variables.