20150818AVU to Staff 76.doc

AVISTA CORPORATION
RESPONSE TO REQUEST FOR INFORMATION

JURISDICTION: IDAHO DATE PREPARED: 08/12/2015
CASE NO.: AVU-E-15-05/AVU-G-15-01 WITNESS: Tara Knox
REQUESTER: IPUC RESPONDER: Tara Knox/Curt Puckett(DNV GL)
TYPE: Production Request DEPARTMENT: State & Federal Regulation
REQUEST NO.: Staff - 076 TELEPHONE: [phone]

REQUEST:

Table 11 of the Electric Load Research Study presented in Ms. Knox' workpapers is a sample design. According to the accompanying text (page 14), this sampling plan is based on the ratio model. Please answer the following:
a. Were samples (within each strata) selected randomly? If so, why was a Model Based Statistical Sampling (MBSS) methodology used instead of a design-based methodology?
b. Were MBSS results compared to results of a design-based analysis? If so, how did system load factor estimates and class load factor, coincident peak, and non-coincident peak estimates compare?

RESPONSE:

Table 11 in Ms. Knox’ workpapers is a summary of the assumptions used in the original sample design and includes an indication of the “expected” relative precision anticipated from the original design. The table shows the state, rate code(s), class, error ratio assumption, planned sample size, and the expected relative precision. The original sample sizes were selected to yield an anticipated relative precision of ±10% at the 90% level of confidence for “key” variables of interest. Please note that there are 8,760 individual hourly demands that could be used in the design stage and we typically select a sample size that is expected to yield good precision for the majority of hours, e.g., 75% of all hours, and for the specific hours typically used in the rate allocation process, e.g., coincident peak demand, class peak demand, etc.

Yes, the samples were selected randomly for each stratum. Model-Based Statistical Sampling (MBSS) is standard practice in the load research industry and is documented in the Association of Edison Illuminating Companies (AEIC) Load Research Manual. The MBSS theory effectively uses “strong stratification” to construct efficient stratification strategies for the application of stratified ratio estimation analyses to produce effective results. MBSS methods are not model-dependent and are effectively design-based since the resulting precision does not rely on the model.

The MBSS approach builds a statistical model, much like a regression model, to predict the residual variance of each element in the population. Then, the model, together with the population data, is used to construct the sample design.

A key point is the following: The model is used to guide the sample design but it is not used to analyze the sample data. As in conventional survey sampling, the analysis of the sample data is based on the actual sample design. This means that there is no need to defend the assumed model when the analyses are presented. Moreover, the results of the analysis will not be biased if the assumed model is wrong. We emphasize this point because in the 70’s and 80’s Richard Royall developed another theory of model-based sampling. Royall’s approach was strongly criticized because the estimators could be biased if the assumed model was wrong. Royall’s approach has nothing to do with MBSS but some people have confused the two methods.

b. MBSS results were not compared to other design based, e.g., weighted mean-per-unit estimation, methods in this study. In other work, including an EPRI monograph, MBSS results were found to be consistently more efficient, i.e., unbiased with improved statistical precision, when compared to other estimators. This is one of the reasons that stratified ratio estimation has become “standard practice” in many load research projects and is taught at the AEIC Advanced Applications of Load Research seminar as the seminal analysis approach.

Page 2 of 2

1