Wednesday, April 29, 2009

Part III: Have you thought about your tracker lately?

Last month, we issued the second of three installments of newsletters centered on identifying opportunities to drive more “bang for the research buck” on tracking studies and included some suggestions on how to execute that process. In this third and final installment on the subject, we will outline some of the practices required on preserving historical trends of customer satisfaction data and bridging the potential gap between results obtained from the previous research supplier to the new research firm commissioned to execute a better or more updated tracking program.

The key objective in managing these changes to an existing tracking study is to preserve historical trends that have been obtained in previous waves of research by mitigating the risks involved in this change. If a risk mitigation plan is not deployed, new results may not be comparable to previous ones, and the time, money, and resources spent on historical trends are wasted.

With that end in mind, the mitigation plan must be designed to identify all sources of variance (characteristics of the data and collection methods that may prevent newer data from being comparable to past data) and, one by one, eliminate or control for each source of variance to the extent possible. In the end, if there are still significant differences between past data and new data, an algorithm must be built to equate historical data to new data and enable comparability to historical trends.

The sources of variance in tracking study migration include:
  1. Method of data collection
  2. Survey items’ wording, scaling, and sequence/position in questionnaire
  3. Quality and other relevant behaviors of interviewers
  4. Changes that result from implementation of programs or other organization-related policies and processes that impact customers’ perceptions of and satisfaction with the organization
  5. Changes in the marketplace that impact customers’ perceptions of the organization

The first three above can be considered “error variance sources” that could be eliminated. The last two sources should be considered “market-related variance sources” which cannot be eliminated but can be accounted for and controlled.

Maintaining the same data collection method and using the exact same questionnaire are key to mitigating risk and preserving historical trends because consistency in method of data collection and surveys can eliminate the first two sources of error variance. However, if the exact same surveys and exact same method are used by different suppliers, then finding significant differences between historical and new satisfaction scores can only be attributable to differences in interviewers (error variance source # 3), market-related variance sources notwithstanding.

In order to control for market-related sources of variance, we recommend conducting waves of a given tracking study in parallel – that is continue to allow the research supplier responsible for historical trends to collect data and have the new research supplier collect data in the exact same format and with the exact same population. If market-related sources of variance exist during the timing of this parallel data collection process, then those variances sources will have an equal effect on data collected by both suppliers. Therefore, these variance sources will be held constant and thus controlled for in comparing the data collected from one supplier to the other.

In terms of parallel testing outlined above, we recommend running a given tracking study in parallel for at least three months. While this may result in increased study costs, it bears the benefit of preventing the loss of historical trends, which are usually far more costly. To reduce these costs, the number of interviews administered by one of the suppliers need not be as many as the number of interviews from the other supplier.

Therefore, if all other study-related sources of error variance are eliminated and market-related sources of variance are held constant, and if significant differences are found in the parallel test, an in-depth analysis of those data is required. This analysis entails statistical testing of the scores or levels of satisfaction reported as well as the variance of each survey item’s data. While industry standard for significance testing is set at the 95% confidence level, in this case we recommend that the confidence level be reduced to 90% or even lower.

So, statistically test the differences between all attribute scores from both studies completed in parallel. Any significant differences between attribute scores should be examined and a complete EDA (Exploratory Data Analysis) should be completed on that study. However, it will also be necessary to test for differences in sample composition. This sample composition test should be completed prior to the samples being finalized, but interviewing techniques and completion rates may impact respondent composition between the studies.

At the completion of the three month parallel tracking study, a complete technical report containing any significant differences found between the two studies should be provided with complete explanations (ANCOVAs) for the differences. In the case of an attribute score being significantly different between the two studies, the historical data can be statistically adjusted.