When planning studies for the usability lab, sooner or later, the question gets asked, “How many users to do we need to test?” Depending on the goals of the study, and whom you ask, you’ll get answers ranging from 5 to 30. Most experts agree that testing more than that is not the best use of your limited usability budget, since each additional test participant costs money to recruit, test, and compensate.
In the world of online user experience research, a similar question comes up: “How many respondents do we need to consider the survey complete?” In the online realm, additional survey respondents are expensive not so much in terms of money, but in terms of time. How long can you wait for more and more people to complete the survey?
How many responses you’ll need really depends on what you plan to do with the data. What are the main goals of the project? Do you intend to use the data to inform design decisions (e.g., for a site re-design effort) or do you intend to use the data as benchmarks/metrics and compare it to some other data set (e.g., data collected in previous rounds or future rounds of research)? Related to the overarching goals of the research project are the analyses you’d like to have done on the data. Are you interested in analyzing click-stream data or are you strictly interested in survey responses? Will you want to “slice and dice” the data numerous ways to see how different demographic groups respond or how certain survey responses relate to other survey responses?
If you desire complex analyses, analysis of click-stream data, or multiple cross-tabulations of the various survey questions, then we recommend a minimum of 3,000 responses. This number of responses makes allowances for the large amount of variation that we see in site visitor behavior and helps to prevent any particular sub-group of respondents (e.g., first-time visitors) from being too small for meaningful analysis.
If you know at the outset that you are not interested in analyzing click-stream data and that your desired analysis of the survey responses does not involve complex or multiple cross-tabs, then answering the question of “How many responses do I need?” really boils down to two different scenarios:
- You need to compare this data set to another data set (e.g., from past or future rounds of research). If you will eventually have more than one data set AND you want to answer questions such as, “Did success increase from Round 1 to Round 2?” then we’d recommend gathering as much data as time would allow. If you intend to compare the data across data sets OR if you just don’t know whether or not you’ll need to compare the data at some point in the future, then we usually recommend a minimum of 1000 responses. Gathering 1000 or more sessions for a survey gives you greater flexibility in terms of how you might use the data in the future. One thousand sessions provide a sufficiently narrow margin of error (± 2.6% at a 90% confidence level) that you can draw conclusions about apparent differences between the two data sets and trust that those conclusions are reliable.
- You primarily are running the survey for qualitative purposes (for example, in order to inform design decisions, gather verbatim feedback, discover usability issues, etc.). If you know that you are not going to need to make numerical comparisons between two data sets, then you can feel reasonably comfortable with fewer sessions. The fewer sessions you gather, the wider your margin of error becomes. For example, at 90% confidence, here’s how the margin of error looks for various sample sizes less than 1000:
As you can see, with 400 survey responses, your margin of error is somewhere close to ±4%. You can also see from the table that the relationship between number of sessions and margin of error is not linear, and there is a point of diminishing returns. If your research goals fall in this second category, you need to consider how wide a margin of error you feel comfortable with and balance that with how long you have available to let the survey run.
So, just as with lab-based studies, the answer to “How many responses do we need?” varies depending on the goals of your study. In general, though, it should fall between a few hundred and a few thousand.