## Overview

Sampling from websites and APIs can be tricky: limits to server load (e.g., retrieval limits) or snowballing effects (e.g., seed of 100 users, sample 100 of their peers and obtain all of their peersâ€™ consumption patterns for 50 weeks). In addition to the minimum sample size necessary to satisfy statistical power requirements, an important consideration is therefore the *technically feasible sample size*. That is, the sample size that can be obtained from a website or API while considering resource constraints.

## Formula

\[N = \frac{req \times S}{r \times freq}\]whereby

- $N$ = sample size (i.e., number of instances of an entity to extract data from),
- $req$ = retrieval limit (maximum number of requests per time unit, allowed for each scraper or authenticated API user),
- $S$ = number of scrapers used (e.g., computers with separate IP addresses, or authenticated users of an API),
- $r$ = number of URL calls to make to obtain data for each instance, and
- $freq$ = the desired sampling frequency for each entity per time unit.

Convert all input parameters to the same time unit (e.g., the retrieval limit may initially be expressed in fifteen-minute intervals but needs to match the desired sampling frequency, which may be expressed in hours)!

## Example

Suppose you wish to know the technically feasible sample size for collecting data from an online social network. In other words, you want to solve for $N$.

The input parameters are:

- $req$ = 5 requests per second = 5 x 60 x 60 requests per hour (18,000)
- $S$ = 1 scraper, authenticated via the service’s API
- $r$ = 2 (the scraper needs to visit two URLs: one to obtain users' meta data, and one to obtain suers' usage history)
- $freq$ = Each user should be vsisited at least once every fifteen minutes (once every 15 minutes = 4 times per hour).