Inference. Again, while not conclusive by itself, this data suggests that the Existing Site has experienced a twelve percent (12%) decrease vs. expectation during the Post Period due to localized factors specific to the Existing Site trade area, and not due to DMA-wide variables. Multi-State Disclosure Document Control No. 040114 Exhibit E to Procedures for Resolving DisputesRelating to the Development ofNew RestaurantsDevelopment Agreement #000000 Xxxxxxxxx Xxxxxxx, Xxxxx STEP II: If there appears to be an impact on the Existing Site that is due to factors within its trade area, rather than due to broader, DMA-wide trends (Example 2 above), identify all significant factors that may have contributed to this impact in addition to the New Restaurant. These could include but may not be limited to the following:

Inference. Many of the Existing Site’s customers will be drawn to the traffic generators in the new center, and allocate some of their limited eating out dollars to whatever restaurant choices are available when they are there, since it is convenient. All or a substantial portion of the 8% drop in sales can reasonably be attributed to the new center. Multi-State Disclosure Document Control No. 040114 Exhibit E to Procedures for Resolving DisputesRelating to the Development ofNew RestaurantsDevelopment Agreement #000000 Xxxxxxxxx Xxxxxxx, Xxxxx Example 4: Same as Example 3, except that a new EPL Restaurant opens at the new Power Center 6 months after the last anchor tenant opens. Existing Site’s sales then decline further, to a 12% overall decline vs. before the Power Center opened, as follows: Existing Site’s Avg. Monthly Sales: Amount Cumulative % Change 12 months prior to New Power Center opening $ 83,333 — First 6 months New Power Center open $ 76,666 (8 %) Next 12 months New Restaurant Open $ 73,333 (12 %) Inference: It would appear that the majority of the decline is due to the existence of the new Power Center (8%), and that only about 4% is due to the New Restaurant (= 12%-8%).

Inference. (except for the intercept) must be eliminated to Bayesian inference is based on the posterior of the model. In terms of the GLMM representation of the model we obtain Yp p(βunp, βpen|y) ∝ L(y, βunp, βpen) p(βpen|τ 2)

Inference. As discussed in Section 2.1.1, we optimize the negative log marginal likelihood. For the implementation this means we need a likelihood, a prior and data to perform the inference.

Inference. To define optimal estimation, we consider a loss function. The loss function is the loss (or cost, risk) that occurs when the correct label is t, but the estimated output by the mathematical model is tˆ. It is denoted as a non-negative function l(t, tˆ). Examples include quadratic loss l2(t, tˆ) = (t − tˆ)2 and 0-1 loss l0(t, tˆ) = |t − tˆ|0 = 1(for t − tˆ 0), |t − tˆ|0 = 0(for t − tˆ = 0). Once the loss function is chosen, the optimal prediction tˆ(x) is chosen to minimize the generalization loss. To learn the model pD(t|x) earned for a given data set D as close as possible to the true predictive distribution p(t|x) from a given training set D, machine learning is performed by first selecting a group of parametric probabilistic models, also called hypothesis classes, and then learning the parameters of the model to fit D.

Inference. Many of the Existing Site's customers will be drawn to the traffic generators in the new center, and allocate some of their limited eating out dollars to whatever restaurant choices are available when they are there, since it is convenient. All or a substantial portion of the 8% drop in sales can reasonably be attributed to the new center. Example 4: Same as Example 3, except that a new EPL Restaurant opens at the new Power Center 6 months after the last anchor tenant opens. Existing Site's sales then decline further, to a 12% overall decline vs. before the Power Center opened, as follows: Existing Site’s Avg. Monthly Sales: Amount Cumulative % Change 12 months prior to New Power Center opening $ 83,333 — First 6 months New Power Center open $ 76,666 (8 )% Next 12 months New Restaurant Open $ 73,333 (12 )% Inference: It would appear that the majority of the decline is due to the existence of the new Power Center (8%), and that only about 4% is due to the New Restaurant (= 12%- 8%).

Inference the relation can be identified and classified through inference via other existing relations (e.g. two events linked to two different timexes can be ordered by means of the comparison of the values of timexes only). •

Inference. The above methodological aspects differ in terms ofscope and complexity. For instance, accuracy assessment addresses almost the whole statistical production process: from collecting data through data processing to data analysis. Almost all stages of statistical production process have something to do with accuracy. Some issues are however more Big Data specific. Dealing withchanges indata sources is a good example. Being technology driven, Big Data sources change more rapidly over time (certainly compared to administrative data) and NSIs need to adapt accordingly. If NSIs do not adapt to such changes, variation over time could be attributed not only in variation in the object of interest, but also in the variation in data sources as well. Such a component in overall variation is missing in survey-based statistics. Literature is abundant with Big Data good practices. However, most of those good practices are one-time-only research work. For instance, a group of expert take a Big Data source, process data and come up with inference about the company’s business model gap, or gaps in city management, etc. The researchers do not care about data source stability or the ability to link the data with other data sources in the long run. Production of Big Data based statistics on a regular base is quite different. For this, NSIs -for instance- need to deal with changing data sources and with a lot of restriction on the access to that data. WP5 clearly demonstrated that Big Data sources are not readily available. Despite the fact that no WP has reached a full statistical production cycle yet, there are already many insights on the use of Big Data for official statistics. The accumulated experience obtained so far is used to derive steps on how to embed Big Data in official statistics (see 3. Conclusions). Looking back we can make an association with sample surveys. Despite being the foundation of nearly all statistical production process nowadays, sample surveys have not been accepted quickly by social sciences. In the beginning doing a survey has been ridiculously compared with a “cartographer to map only one square in ten on his grid.”(Ayrton, 2017). In the remainder of this report, the above mentioned 11 methodological issues are described one by one, followed by conclusions.