Monthly Archives: October 2018

A small food survey

This post is about a small side project: A (non-representative) food survey. Which foods are (un-)popular? Do food preferences correlate with traveling activities? Are there different “food types”? And many more.

The idea

During lunch, naturally, we often talk about food. About different tastes, likes and dislikes. How (un-)popular are certain foods, are there different “types” of eaters, or do travelling destinations correlate with food preferences? To shed some light onto these questions, because we did not find accessible studies, and with the tools of web programming and statistics at hand, we decided to start a fun side project.

Now we present you a small food survey! Since we do not select a random sample, the survey does not allow representative statements. Its findings only hold for the group of participants, not for Germany and certainly not on a global scale. Nevertheless, we hope that both we and you enjoy our survey and find some interesting connections.

Click here to start the survey (in German only)!

Additional infos

After going through three medium-short pages of questions you will see some preliminary results, and how your answers compare to those of others. Moreover:

  • Of course, all data are fully anonymous and will not be used for any commercial purpose.
  • Throughout the course of the survey, we will update this blog post to keep you informed.
  • Please feel free to share this survey as you wish.

If there are any questions or comments, please send us a message. Enjoy!

dqsample: A bias-free alternative to base::sample()

For many tasks in statistics and data science it is useful to create a random sample or permutation of a data set. Within R the function base::sample() is used for this task. Unfortunately this function uses a slightly biased algorithm for creating random integers within a given range. Most recently this issue has been discussed in a thread on R-devel, which is also the motivation of the dqsample package. Currently dqsample is not on CRAN, but it can be installed via drat:

Example for the bias

When sampling many random integers the density of odd and even numbers should be roughly equal and constant. However, this is not the case with base::sample:

plot of chunk base

Or with slightly different parameters:

plot of chunk base-oszi

This particular example for the bias was found by Duncan Murdoch.

In dqsample the algorithm suggested by Daniel Lemire (2018, <arXiv:1805.1094>) is used. With this algorithm there is no observable bias between odd and even numbers:

plot of chunk dqsample

Where does the bias come from?

Internally the base::sample() function needs uniformly distributed random integers in an half-open range [0, n). In order to do so, R uses random floating point numbers that are uniformly distributed in [0, 1), multiplies by n and truncates the result to the next smaller integer. This method would be fine, if the random numbers used as starting point would be real numbers in the mathematical sense. However, this is not the case here.

The default random-number generator in R is a 32 bit version of the Mersenne-Twister. It produces random integers uniformly distributed in [0, 2^32), which are then divided by 2^32 to produce doubles in [0, 1). We can now invert the procedure described above to see how many integers are mapped to a certain result. For example, we could simulate rolling ten dice using sample(6, 10, replace = TRUE). Since 2^32 is not a multiple of six, the distribution cannot be completely even:

We see that both one and four are very slightly less likely than the other numbers. This effect gets much more pronounced as the number of items increases from which one can choose. For example, we can use the m from above to see how that uneven distribution of odd and even numbers came about:

Here we see that while only two integers map to any odd number, there are three integers mapped to the even numbers. This pattern shifts half way through the possible results, making the odd numbers more likely, leading to the first image displayed above. As one goes away from m, these pattern shifts occur more rapidly, leading to the oscillatory behaviour seen in the second image. As one moves further away from m, these oscillations happen so rapidly, that a density plot of odd and even numbers looks constant, but the bias is still there. For example, for m - 2^20 one such pattern shift happens between 982 and 983:

Below this point, even numbers are more likely than odd numbers. After this point, the pattern is reversed.

Conclusion

The algorithm used by base::sample() is biased with non-negligible effects when sampling from large data sets. The dqsample package provides an unbiased algorithm for the most common cases. It can be used as a drop-in replacement for the functionality provided by base::sample().