According to Nielsen:
Electronic metering technology is at the heart of the Nielsen ratings process. Our tools capture not only what channel is being watched, but also who is watching and when, including “time-shifted” viewing.
Nielsen’s TV families represent a cross-section of representative homes throughout the U.S. Their viewing is measured by our TV meters and Local People Meters which capture information on what’s being viewed and when and, in the major U.S. markets, specifically who and how many are watching. Additionally, we collect more than two million paper diaries from across the country each year during “sweeps.”
The method that the asker calls into question is Nielsen's traditional diary based ratings calculations. (Although note that even the automated data collection is only permitted in households that have been invited by Nielsen). Put simply, the diary ratings work like this: A call center places calls to households in target market areas and ask for permission to send a "ratings diary". If the household agrees, they will be sent one diary for each working television in the household (up to five), and each diary will contain a cash (the amount varies) as an incentive to complete the diary. The notion that "only certain types of people" would take time to complete the diaries is strictly opinion, but it isn't a far stretch to think that perhaps more retired or unemployed people would find this arrangement ideal.
The diaries usually contain fourteen pages (although if the household has digital video recording hardware two additional pages are included, allowing for the previously mentioned 'time-shifted' viewing to be recorded). Each page can hold twelve hours worth of viewing.
I found this picture of an actual diary page by searching Google images:
As you can see, the household is expected to provide the following information for every fifteen minutes during the survey period (of seven days):
I. Whether or not the tv is on
II. The call letters of the station, or the name
III. The channel number
IV. The name of the program
V. Who is viewing the program
The household can indicate an above response by drawing a vertical line under the information that they wish to "carry down" to lower grids.
While theoretically perfect in it's design, the diaries are rarely filled out with 100% accuracy. So how are the books interpreted? An office in Venice, FL collects all of the diaries via US mail, and specially trained temporary employees enter the data into computer terminals. If data on the diaries conflicts (basic example: what if the channel number does not match the supplied call letters?) a stringent set of guidelines is used to come up with the appropriate data, and sometimes the data cannot be used. The information entered into the terminals is double checked by separate employees who's only job is to verify. The Nielsen company claims that this results in over 99% accuracy.
One personal criticism of the system as it exists now is this: sometimes households misinterpret how to fill out the diaries. As simple as they seem, some people struggle with them. If the employee receives a filled-out diary with information that seems "sketchy" at best (for example: a user mistakes the "tv on" column for the "tv off" column and indicates that the tv is always on), the procedure is to enter the data regardless (in some occasions). The method is scientific (they can't prove the user was confused, what if they really DID mean to do this?), but I'm unconvinced they've really taken into account the margin of human error on the household's end.
The truth is that nobody can provide a definite answer in regards to whether or not the ratings are 100% accurate, the numbers are just too high. In my opinion they're as accurate as they're going to get until Nielsen adopts a strictly electronic form of television viewership monitoring.