For Steers without conditions on the target respondent, a sample size of 500 provides a ~4% confidence interval (with a confidence level of 95%). As a quick example assume that we get 500 responses from a Yes/No steer with a result of 80% Yes - this would mean that we would be 95% confident that if we asked the full population the same question, the result would lie in the interval ~76-84%. The extreme values of the confidence intervals tend to occur when a survey returns 50% Yes/No, which shows it can be more difficult to strictly model the behaviour of a panel which is split down the middle.
For steers with stricter conditions, such as a combination of (age, gender, previous purchasing behaviour), a sample size of 200 provides a ~7% confidence interval (with 95% confidence level). On the other hand, decreasing the sample size in this case is also countered with decreasing the overall population - if we restrict to looking at only Male respondents, the total panel will be about 50% smaller. Decreasing the total population (by restricting by age, gender etc.) does reduce the confidence interval (so that it is closer to a confidence interval of ~6.5%, say).
Reducing the sample size does indeed increase the width of the confidence interval, but the relationship is not linear! Doubling or halving the sample size does not double or halve the confidence interval. See the graph below:
Once the sample size gets below ~100, the confidence level can start to become too wide to make sensible conclusions.
The graph also shows that increasing the sample size beyond 200 has only a small effect on the confidence interval. A smaller sample size is also favourable in the case of a highly specific consumer profile because it leads to quicker survey completion. If a highly specific combination of demographics is required then it can take some time to allow panelists to see the survey and respond. If there are many such surveys then the overall time taken is increased.
As we pay consumers per response, a 500 sample size survey costs 150% more than a 200 sample size survey but the effect measured only increases in resolution by ~2%.