3
$\begingroup$

Let's assume that I have a dataset which gives me the two variables: height ($x_1$), daily calorie intake ($x_2$) and weight ($y$) of a person. In this dataset, we assume I have a large enough number ($n$) of examples in my dataset to make it statistically accurate.

Now, I can do a regression with multiple variables (in this case: 2), to obtain a prediction of a person's weight based on their height and calorie intake. It's rather easy: If I want to find out the predicted weight of a person, I need simply insert the value $x_1$ and $x_2$ into my fitted equation to obtain $y$.

However, I am wondering how I can do something a bit different:

Given a height $x_1$ and a daily calorie intake $x_2$, what is the probability of a person having a weight greater than or less than a certain value, or having a weight in a certain range.

For example: Assume I have $2000$ calories intake per day and a body height of $180 \text{ cm}$, what is the probability that the person weighs between $80\text{ kg}$ and $90\text{ kg}$.

Is such an estimation possible? And if so, how can I do it (analytically or numerically)? Is there any way for me to bound the error that this estimate will have?

$\endgroup$
1
  • $\begingroup$ How about estimating the distribution of residuals and reading the answer from that? $\endgroup$ Nov 4 '14 at 9:57
3
$\begingroup$

If you're taking a frequentist approach, it sounds you're trying to compute the distribution that's used in a prediction interval.

You can do that with a regression, but you do rely on the distributional assumption more than for other forms of inference.

The usual way is to construct a version of a pivotal quantity, but there are some differences from the usual definition of pivotal quantity because instead of trying to get a pivot for a parameter, you need something that will work for a future observation.

In regression, if $y_f$ is that future observation and $\hat{y}_f=x_f\hat{\beta}$ its predicted value, then $Q=\frac{y_f-\hat{y}_f}{s_f}$ has a t-distribution with $n-p-1$ d.f. where $p$ is the number of predictors and $s_{f}$ is the standard error of the prediction.

The formula for $\text{Var}(y_f-\hat{y}_f)$ is often given in derivations of the prediction interval; $s_f$ is the square root of its sample estimate. If $x_f$ is the row-vector of predictors (including the constant) for the future observation, $y_f$, then

$\text{Var}(y_f-\hat{y}_f) = \sigma^2(1+x_f (X'X)^{-1} x_f')$

To obtain $s_f$, replace $\sigma^2$ by $s^2$ in the above formula and take the square root.

You can make probability statements about $Q$, and from those, derive statements about $y_f$ in terms of $\hat{y}_f$ and $s_f$ that should have the appropriate coverage if the assumptions hold.

Packages that will give one-sided prediction intervals can be used to help with the sort of computation you need, but it's easy enough to do directly.

$\endgroup$
2
  • $\begingroup$ Dear Glen_b, could you maybe define what y_f^hat is? I admit I am somewhat of a beginner when it comes to statistics (regressions are not exactly high end), and I am not familiar with the notation. How can I calculate it? $\endgroup$ Nov 5 '14 at 10:49
  • 1
    $\begingroup$ I've made a brief addition to the answer. Hopefully that's clearer now. $\endgroup$
    – Glen_b
    Nov 5 '14 at 10:57
2
$\begingroup$

You can do it easily if you know the distribution of the error terms (or make an assumption). In case you describe, normality could be reasonable. Then $Y \sim N(\beta_1 x_1 + \beta_2 x_2, \sigma^2)$. Lets say that $\beta_1=0.4$, $\beta_2=0.004$ and $\sigma^2 = 100$. Then for the height and intake you describe, the weight is distributed as $Y \sim N(80, 10^2)$, and hence probability that his weight is between 80-90 kilos is roughly $\Phi(1) - \Phi(0)\approx 0.34$). But be advised that there are multiple assumptions which should be checked so that this inference makes sense, such as constant variance and the normality or errors.

$\endgroup$
2
  • $\begingroup$ Thank you for your comment! There is just one thing I don't quite understand: By your definition, the Normal distribution is centered around the output of the function I compute. This, so far, I understand. Edit: Sorry, hit enter by accident. Now, am I correct in assuming that I estimate sigma^2 by summing over y_i - y'_i / n-p-1? (with n = training examples, p = parameters, in this case 2, y_i known (correct) outputs and y'_i the output of my function) Second: Is there any way for me to give an error bound to this? i.e.: A person is 80-90kilos with 0.34 +/- c? Now, $\endgroup$ Nov 4 '14 at 11:32
  • $\begingroup$ I think error bounds on probability estimates makes little sense: the lack of certainty is already encoded in the statement. $\endgroup$
    – psarka
    Nov 5 '14 at 14:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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