Thank you very much for your comments, they have helped with clarity and improved the manuscript. For your convenience, we have also included a PDF version of the response to both reviewers together in-line with the reviewers’ comments.

We have added the following to the first paragraph to underscore the importance of uncertainty affecting decisions,

“Further, the effects of endogenous model uncertainties on model error may be different from their effects on the ranking of alternatives, and therefore on decision making. This difference has been largely understudied and is the focus of this paper.”

Lines 102-105: The management alternatives were developed and described in greater detail in Mautner et al. (2020). We have changed the text as follows,

“Four management alternatives designed to increase groundwater recharge within the basin while avoiding flooding are drawn from Mautner et al. (2020). The alternatives were chosen based on conversations with local practitioners and previous modeling efforts. The alternatives are: the implementation of spatially distributed infiltration basins, demand management through repair of leaks in the water supply network, injection of treated wastewater at existing wastewater treatment plants, and the status quo historical alternative.”

Lines 107-108: The management objectives are drawn from the previous study. However, the calculation of the objectives was modified to avoid outlier values that would occur when parameter combinations led to high quantities of model error. For example, dry cells are represented in the model as large negative values for groundwater head and thus lead to unrealistic values for energy use based on the elevation of groundwater head. These outlier values of the objectives led to difficulties in characterizing the sensitivity of those objectives to the parameters. Thus, a minimum value for groundwater head was set for each well based on the model cell and layer depth. We have modified the text as follows,

“The management objectives evaluated are drawn from Mautner et al. (2020), modified to avoid outlier values that would occur when parameter combinations led to high quantities of model error that would affect the sensitivity analysis.”

Line 166: Correct, it should read,

“While perfect monitoring and representation is the ideal, in reality, simplifying assumptions must be calibrated to create models that can better inform policy and management.”

Lines 168-170: We agree, it is more clear as follows,

“Changes in the observations used to evaluate error can lead to differences in the behavioral parameter sets that are chosen as the best performing simulations.”

Lines 255-265: We have attached the supplemental figures here and will include with the manuscript.

Fig. 5-6 & discussion: We have added the following text at the beginning of Section 3.2 following the first sentence introducing Figure 5,

“The value of δ can range from 0, indicating that the output is independent of the parameter in question, to 1. There is not a standard value for δ that is considered to be highly sensitive because parameter sensitivities should be evaluated in relation to each other and in the context of each case study. Based on the sensitivity values for this system, we consider a δ of roughly 0.2 and above to be highly sensitive.”

Thank you for your detailed comments, we have responded to each point below. Excerpts from the manuscript are shown italicized in quotations with changes in bold. For your convenience, we have also included a PDF version of the response to both reviewers together in-line with the reviewers’ comments.

1. We agree, the text has been amended as follows,

“While the impacts of external uncertainties, such as climate and population growth, have been widely studied, there is limited understanding of how decision support is altered by endogenous uncertainties arising from model parameters and observations used for calibration.”

1. We agree, the text has been amended as follows,

“Error metrics (i.e., squared residuals of piezometric head) are not necessarily controlled by the same parameters as the head based objectives needed for decision-making.”

1. Yes, this is correct. Updated text:

“In hydrogeologic models, endogenous uncertainty is contributed by model parameters describing natural and human components of the system, and the set of historical observations used to calibrate or constrain the parameters.”

1. Upon consideration, the sentence beginning “Etter et al. (2018)…” has been removed in the revision.
2. We agree that this could be more clear, we have added the following text to provide this context,

“In conflict with the previous two objectives, certain parts of the city lie in areas that are affected by seasonal flooding resulting from medium-term groundwater mounding, which is particularly damaging in urban areas. To take into account these possible negative effects from increasing groundwater head within the valley, the urban flood risk (F) is the sum of the urban area in cells with groundwater mounding (a) divided by the total urban area in the model (A) during the time periods (t) in the last year of the model period.”

1. The model would take between 1 and 30 minutes, with most on the order of 5-8 minutes depending on the processor itself and the parameter combinations. We have added the following line to the text,

“A single model run is on the order of 5 minutes, depending on the combination of parameters and the processor speed.”

1. This comment may have been cropped, but we interpret it to refer to lack of clarity on the number of observations used and how that may affect the results. We used all the available observations, which ranged from 1 to 29 annual observations during the 30 year model period, depending on the well. The text has updated as follows,

“Well observations vary from 1 to 29 data points per well over the 30 year model period, with a maximum of one observation per year.”

[Comment numbers skip from 7 to 12]

1. We agree, the text has been amended as follows,

“The sensitivity of the management objectives across the parameter space with respect to both management alternatives and cluster behavioral parameter sets indicates the variability of uncertainty with respect to individual physical model parameters.”

1. The equation for delta and the accompanying text description have been added as follows,

“ δ_i = 1/2 E_Xi [ ∫|dμ_Y - dμ_Y|Xi|] ,

where the moment independent sensitivity indicator of parameter X_i with respect to the output Y (δ_i) represents the normalized expected shift in the distribution of Y, as a function of μ_Y and μ_Y|Xi , the unconditional and conditional measures of Y, respectively. In this study, the parameters (X_i) are the 33 parameters shown in Table 1 and the output (Y) are the three management objectives described in Equations 1-3.”

We have also updated the objective equations (Equations 1-3), so that they read Y_E , Y_W , and Y_F .

1. We have rewritten the paragraph in question to be more clear as follows:

“We evaluate the model results to visualize changes in ranking according to three types of comparisons using sets of heatmaps that summarize the ranking of the alternatives across all the parameter sets. The comparisons evaluated are: (1) all three objectives across the observation cluster behavioral sets; (2) all three objectives for observation cluster behavioral set C-12345 across the range of the alluvial hydraulic conductivity parameter [HK2]; and (3) the water quality objective (Y_W) across the observation cluster behavioral sets and parameter HK2, simultaneously.

In all three comparisons, the first step is to rank the alternatives in relation to each other according to the objective(s) from lowest (1) to highest (4) in each individual parameter set. Then, the ranking data for all the parameter sets in each comparison are summarized as follows:

Three objectives across observation cluster behavioral sets - The count of rankings for each alternative. Each column (cluster) in each objective row will sum to 5,000.

Three objectives for C-12345 across parameter HK2 - The 5,000 sample set is separated into ten bins along the parameter value range from Table 1. The ranking count in each bin is divided by the total number of parameter samples in each bin to allow for direct comparison across all bins. This is necessary because the distribution of behavioral parameters can be non-uniform. Each column (parameter value bin) in each objective row will sum to 100%, or null if there are no parameter sets in that bin.

Y_W across the observation cluster behavioral sets and HK2, simultaneously - Same as in the previous comparison, but for only the water quality objective. This is repeated for the remaining observation cluster behavioral sets (C-00001 to C-00005). Each column (parameter value bin) in each cluster row will sum to 100%, or null if there are no parameter sets in that bin.”

1. This value is the sum of squared weighted residuals, meaning that it is in units of m² and is expected to be large because there are a total of 8,181 observations. Missing simulated values for an observation are by default represented in the model as -1E6, which can be a result of a dry cell or an incomplete model run based on incompatible parameter values. Therefore, the squared residual would be on the order of 1E12, leading to the highest values seen in Figure 4. We have added the units of m² to the caption to clarify this point.
2. We agree that this will help improve the analysis. The 95% confidence interval for each sensitivity value has been added as a red line over each bar and the caption was updated to indicate this change, a higher resolution image has been included in the Supplement. The confidence intervals have also been included in the relevant supplemental figures. In general, we find that the differences in parameter sensitivity discussed in the text are significant (i.e., 95% intervals do not overlap), which supports the convergence of the sensitivity analysis as recommended by the reviewer. The following text was added to the caption,

“The bootstrapped 95% confidence interval for each sensitivity value is shown as a red line.”

1. You are correct, this has been fixed as follows,

“ δ sensitivity of the error metric and three management objectives (rows) according to the 5,000 filtered samples for the 8 model parameters (columns) with the largest differences in sensitivity between clusters for the historical management alternative.”