An observation bias is a particular variety of bias introduced into scientific studies when the method of observation used causes the results to be skewed in some nonrandom manner, leading to inaccurate results. Although this can occur in any variety of study, correlational studies, in which there is no independent variable and all information is from observation, tend to be the most vulnerable.
Observation bias can occur when the group being studied is not representational of the group as a whole. An example would be giving a school a rating based upon testing the children in the "special class", making a judgement abut the effects of video games in general on children based entirely on Grand Theft Auto, extrapolating a national opinion on gay marriage from a poll taken entirely in Alabama, or estimating national literacy rates from a written survey. Non-representative samples can still grant useful information, but the information can only be generalized to those who are represented by the sample rather than the population as a whole. (A poll conducted from random households in Alabama could be used to generalize to all of Alabama, but not all of humanity.) "Non-representational group" is the most common form of observation bias.
This sort of bias is not easy to spot. It may not be immediately obvious that most psychological experiments have the gotcha that they can only draw conclusions over the realm of "people willing to volunteer or get paid for psychology experiments", which is not a truly representative sample of people as a whole. (Involuntary experiments are not vulnerable in this way, but medical ethics prevent such experiments from occurring.) Similarly, there may be other ways that, for any given experiment, everything being tested on fits into some subcategory that the entire population does not fit into.
A reporting bias should be considered as a particular subtype of observation bias, but one so prominent and common it is considered a bias of its own sort. Therefore, full discussion for it should remain in its node.
Biases of poor subject selection are not the only things that can cause observation bias. Another popular cause is when the equipment being used for observation is itself inaccurate, or changes what is being observed. Programmers occasionally meet debuggers that alter the running environment of a program enough that a bug, otherwise consistent, won't occur while running the debugger. Other such problems occur in quantum physics, where observing the state of a quark changes its state, and only partial information can be determined. (This isn't truly an observation bias, as it's a well-documented phenomenon and is expected.)
A similar form of observation bias is when one result is easier to observe than another, or two results can be confused with one another. A rather abstract example would be an event that occurs for only a split second versus one that takes much longer, or perhaps nothing at all occurring; it would be easy to overlook the quick event unless special gadgetry is being used to ensure everything gets watched.
An extremely nasty bit about an observation bias is that, like all bias, it is consistent. Therefore, unless one is actively looking for it, it can be hard to spot when it occurs. The only way to ensure that a bias is not taking place is by testing the initial hypothesis multiple times, using a different method each time. If some experiments give different answers, then one or more of the experiments was invalid, and it's a bit of a puzzle to figure out which ones had the problem...