Prostitution Arrest Spatial Forecasting in an Era of Increasing Decriminalization

Round 1

Reviewer 1 Report

This manuscript predicted prostitution hotspots using machine learning models. It has some interesting information, but there are still some issues need to be addressed. My concerns are as followings:

1.     The title of this article emphasizes decriminalization, but the model and criminogenic factors have nothing to do with it. In other words, this study can be conducted in any city, regardless of how it defines the culpability of both parties to prostitution.

2.    The paper adopted two models, and the authors should compare the results of the two models and tell the reader which is the best one.

3.    The factors selected by the author that affect prostitution are not different from other types of crimes. It is better to find some specific ones for prostitution crimes.

4.    The optimal size of the selected hexagon needs to be more justified, since the modifiable areal unit problem is a serious analytical issue for research using aggregated spatial data.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper presents the results of modeling prostitution arrests in Chicago during 2009-2016 using two models: logistic regression and artificial neural network (ANN).

I find the paper easy to follow, with a clear presentation/discussion of data. However, it pertains to several limitations in many essential aspects, including methodology and result discussion. Because of these significant limitations, I find that the paper is not having an acceptable quality for publication and recommend rejection or reconsideration after a major revision.

My main concerns are:

(1)    The models used in this paper are not new. I agree that their applications and implications for modeling prostitution arrests might be exciting and provide insights into the phenomenon. However, the discussion in both the methodology and the result sections is unsatisfactory. It doesn’t provide the readers with a clear understanding of (a) why these models should be used, (b) how they work on predicting the arrests, (c) what are the important parameters to play with, and (d) why they are reliable/trustable for the tasks.

 Specifically:

 1.1. Why both logistic and ANN should be presented? These are very different modeling methods. Thus, what are the differences in the results provided? The author briefly presented the set of significant variables for predictions as an outcome of the logistic model in Table 2. Are these different compared with variables of high weights suggested by the ANN? What if the ANN model suggests a different set of variables?

1.2. For the logistic model, what are the residuals? Is there any spatial pattern that exists within the residuals? Although it is not presented in the paper, I think it is very likely that spatial autocorrelation exists within the residuals. If yes, how should these spatial elements be treated to improve performance? I don’t think that the ANN model results have been analyzed thoroughly.

1.3. For the ANN model:

o   I am particularly bothered because the discussion on how different variables were used and weighted within the ANN model was entirely overlooked. The author even stated in lines 435-436 that “The neural network does not provide any specific information about the individual impact of each independent variable on the model.” which is incorrect. The weights used for each variable in an ANN need to be fully understood and discussed so the results can be convincingly presented and considered valid.

o   Why should a threshold of 3 for high-low classification be chosen? I think this is a critical threshold and highly impacts the prediction results presented here. The author(s) mentioned that different threshold values were tested. And I think it is vital to present the result of this sensitivity test so readers can scrutinize the performance of the ANN model.

o   Analyzing the ANN model outcome using only a high-low classification limits our understanding of the actual prediction values. Although the result reported in Figure 2 shows a high portion of “correct prediction”, how well the model predicts the raw data (count or percentage) that was shown in Figure 1 (before the smoothed density surface was generated) remains questionable. For example, the models predicted a hexagon having a high number of arrests, meaning it has more than 3 arrests. However, it is unclear to me if it had a little over 3 (4 or 5) or a lot over 3 (10 or 15) arrests. I don’t think the ANN model results have been analyzed thoroughly.

 (2)    The author(s) suggested a strong connection between commercial sex sellers, thus arrests between 2009-2016, to sex trafficking victimization. This also seems to be a strong motivation for understanding the patterns of arrests. However, this aspect hasn’t been considered in the modeling task itself. Readers were neither presented with a related background of sex trafficking behavior in the study area.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The manuscript can be accepted in present form.

Reviewer 2 Report

No further comments. Thank you for the responses.