Summary and conclusions

Two different sets of call logs were examined. The call outcomes, ring time, call duration, and dial time were different between the two call logs but not significantly so. It is almost impossible to determine the result of dialing a number from the number alone. It is only possible to determine the probability of a particular outcome.

The key observations from these analyses are these:

▪When calling known numbers, the majority of calls are either connected or not answered. Other typical reasons for a call not connecting are because the line is busy (3 to 6% of calls) or because a fax/modem was detected (up to 3% of calls). When calling known numbers, other reasons for failure are uncommon, normally less than 1% of calls.

▪When the sample includes random digit dialing, a percentage of the calls will be returned as unobtainable numbers or call failures (cause unknown). For the Project A this was around 10% of dialed calls but in practice this will depend on the ratio of random to known numbers in the sample.

▪The percentage of no answer calls is mostly dependent on the no answer time-out and the type of sample being dialed, but is likely to be around 25% to 35% of dialed calls when a time-out of 25 seconds is used. At the minimum time-out (15 seconds) the percentage of no answer calls may be very high (for example, over 50% of all calls).

▪For a given set of randomly distributed sample records, the call outcome distribution is fairly constant over time. However, residential surveys may exhibit an increase in the connect rate from afternoon into evening as increasing numbers of respondents are at home.

▪The ring time distribution of calls is relatively constant. Most calls are normally answered around six to ten seconds or 18 to 22 seconds. For some sample, possibly depending on geographical location, a number of calls may be answered within one second.

▪Most calls are relatively short. 30% to 40% of calls are likely to be less than eight seconds, and most calls (more than 95%) are less than 90 seconds in length.

▪Most projects show a similar call duration distribution. The variation is normally between eight and 60 seconds. For some projects, the call duration distribution decays linearly with increasing call length, while for other projects it seems to follow a normal distribution curve.

▪Considering that most calls are either connected or not answered and the call duration and no answer time-out are generally at least 20 seconds, that is, normally significantly longer than the dial time, the variation in dial time is relatively insignificant.

The conclusions drawn from the analysis of all six projects and illustrated here by projects A and B are as follows.

▪Busy numbers and failed calls were detected within the dial time which is generally fairly short (less than three seconds). Fax/modems were relatively uncommon (under 3% of calls) and 80% of calls to fax/modems are answered in about seven seconds. This means that busy numbers, failed numbers, and fax/modems are relatively insignificant to the predictive dialing algorithm as they are not very common, and since they are detected relatively promptly it means a replacement call will be dialed quickly.

▪More significant are no answer calls and connected calls. The no answer calls are significant because it takes a relatively long time to detect that a call is not going to be answered, and if the no answer time-out is decreased then a high proportion of the calls will not be answered. Connected calls are significant because the call duration for each one is normally relatively short so another call will need to be dialed to replace it.

The simplest predictive dialing algorithm could just take account of the ratio of connected calls to dialed calls and dial one extra call for every call that is expected to fail (no answer, busy, fax/modem, failed number).

The next simplest algorithm would also take account of dialing extra calls due to connected calls disconnecting. It should be possible to estimate the rate of call disconnects based on the number of connected calls and the average call length.

As the ratio of call outcomes, the ring time distribution, and the call duration distribution appear to be relatively consistent over time (provided that the sample is evenly distributed), the most effective style of predictive algorithm is one that uses this information. It should be possible to predict the probability of an in-dial call timing out based on how long it has been ringing and the probability of a connected call terminating based on how long it has been connected. If these probabilities reach a certain threshold then a replacement call would be dialed. The threshold would be determined by the confidence in the ring time and call duration distributions and how acceptable is the risk of making a silent call. This is the option that is provided by the Predictive dialing algorithms.

A more advanced predictive algorithm could take account of differences in geographical location as can be determined by the phone number. This is because the ring time distribution shows a slight variation, and the dial time, although relatively short, often shows quite a strong correlation to geographical location.

A better analysis of calls could be performed if call.log information were combined with interview call dispositions as this would enable a more detailed breakdown of the connected calls (for example, answering machine, refusal, completed interview). This should be possible from the UNICOM Intelligence Interviewer sample history table which includes information on both dialer and interviewer call outcomes. Logs could also be checked to see whether call outcomes have any relationship to any of the other information that is held for a sample record (age, gender, or occupation, for instance). For example, it might be possible to determine that a sample record with a certain combination of sample values, dialed at a certain time of the day, is more likely to result in a specific call outcome.