Machine Learning Applied to NHS Electronic Staff Records Identifies Key Areas of Focus for Staff Retention

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Machine learning applied to NHS electronic staff records identifies key areas of focus for staff retention

Comments to the Authors:

Limitations of the model: The limitations of the model that may question the practical applicability to HR decision-making are also not discussed in detail, as modest predictive performance (AUC 0.65) is admitted.

Generalisability Concern: Because it is applied to one NHS Trust the analysis has a major limitation in its generalisability. Multi-site data might improve robustness and increased applicability.

The Insights generated by the Predictive are Shallow: As much as the analysis establishes the main predictors such as the length of service and the distance to work, it does not examine the effects of interaction between factors, which could provide deeper insights.

Inadequate Theoretical Approach: The paper would be enhanced by having a more theoretical approach to the staff retention models to offer some context to the findings in the larger body of management of human resources.

Misuse of Temporal Trends: The independent snapshot mentalities of analysing annual data overlook the potential benefits of fish within the longitudinal means of analysing the changes in retention over time.

Lack of Practical Interventions: Although there is the identification of risk factors, the study fails to go further ahead and list retention strategies that are practical and actionable based on findings of the model.

Inappropriate Use of First-Person Pronouns: The frequent use of "we" and "our" throughout the paper is unsuitable for academic business research writing and should be replaced with objective, third-person phrasing.

Author Response

admsci-3739163_revision1

Machine learning applied to NHS electronic staff records identifies key areas of focus for staff retention

We thank the reviewers for their comments and insights, and have incorporated the feedback and hope that we have improved the manuscript. Our detailed responses are set out below point-by-point. All references to line numbers are for the ‘track changes on’ version of the document.

Comments to the Authors and responses

Based on dataset derived from Electronic Staff Records at Ashford & St Peter’s NHS Foundation Trust, using a machine learning approach, the article identifies a few predictors of staff retention. The main result of the text is not the approach itself (contrary to declaration of the Authors), nor the list of predictors identified, but a proof of the fruitfulness of this procedure using a relatively small data set.

The article is a well-written report from an empirical study (teaching and using AI to a dataset) and presentation of the results obtained, which would be a good starting point for a consultancy report. As an article for a scientific journal, however, it has several weaknesses that must be eliminated before its publication.

1. The entire text does not contain any references to the theory of fluctuations, neither in the introductory part nor in the discussion of the results. Such conduct disqualifies the article as a scientific text, which must indicate how the obtained result (and not just the methodology) complements scientific knowledge in a given field.

Response: The reviewer is of course correct that the theory of fluctuations could play a valuable part in this sort of analysis. We had considered this initially, for example how compensation, job satisfaction, external factors (i.e. the jobs market), life events and so on might feed into ‘staff departures’ as a stochastic process. Fluctuation theory would also have the advantage of allowing for tipping point thresholds. An example of this lies in our findings that in some classes of employees, such tipping points may be more prevalent than others (i.e. our finding that for medics, short-staffing which may be caused by prior departures is much more likely to trigger further departures).

Unfortunately, electronic staff records alone do not really allow for this type of modelling, especially as we did not have the unique staff identifiers to track single employees over time. We have expanded the Introduction to address fluctuation theory and other models (including references), added some comments on tipping points, and highlighted the potential for these methods of research in limitations / future work [lines 55-74 and 346 in the track changes version]. We have also added additional material on theory relevant to the topic, especially around extrinsic hygiene factors and intrinsic motivation factors [lines 43-47].

2.   There are also no references to other examples of analogously used AI methodology, although they have been postulated in the literature and described at least in terms of their assumptions for many years see: (Strohmeier, 2007) in "Human Resource Management Review"

Response: We agree that our manuscript focused mainly on employment issues in the NHS and not the wider field of AI methodologies in HR, including e-HRM. This has now been included [lines 70-74 and 316]. The suggested reference is an excellent article and we have included it, and thank the reviewer for bringing it to our attention.

3. It should be clearly stated that the adopted procedure does not identify prospective predictors, but only excludes from the previously established list of variables those that do not provide such predictions. This remark is important because it indicates that the usefulness of the procedure is based on the list of variables that are constructed a priori on the basis of existing data (the selection of variables for analysis in this study was therefore random and not dictated by any analysis other than intuitive)

Response: we agree with the general point that the reviewer makes, but this is an unavoidable limitation, that our list of variables was constructed a priori on the basis of existing data. Collecting new data would have necessitated a completely different approach, and we believe that a pilot study has value in demonstrating the types of data that might be informative. We have added an explicit acknowledgement that our variables were predetermined by the availability in the NHS electronic staff records [line 326 in the track changes version]. The ability to bring in additional variables will be improved in future as the NHS is currently undergoing pre-procurement of new
e-HRM systems, see:

 https://www.nhsbsa.nhs.uk/future-nhs-workforce-solution-transformation-programme

4. The discussion of how to use retention predictors in individual cases requires further discussion. The analogy („In this study we show which factors contribute most to increased probabilities of staff departures. The work is not intended to suggest that machine learning can be used to predict individual reasons for staff leaving. In this sense, the algorithm is similar to car insurance models: it can identify factors that increase the likelihood or risk of an event, but will not perform well in forecasting outcomes for specific individuals.”) although accurate, it is misleading. In the case of car insurance, the reaction to a high value of a prospective predictor is an increase in the price of insurance for an individual who belongs to the group indicated by the predictor (discouraging them from insuring with that company), while the company must introduced some actions to influence employees who are already employed and not to discourage them from belonging to the company (rejecting job candidates from a given type of group).

Response: On the point that the analogy of insurance is not perfect, we agree, but we are only trying to make the point that a model that cannot accurately predict individual outcomes can still have value in managing overall behaviours. We do agree however that the reaction is not exclusively one of price (salary) but of other changes employers can make in the NHS. The other reviewer made similar points, and we have extended our comments throughout the Discussion. As this is a pilot study of a single Trust, we would also comment that providing specific and immediate recommendations for the NHS would probably be over-reach, and we hope that our revised Discussion strikes the right balance.

5. From a formal perspective, the article should be constructed differently in the discussion of results and conclusions. Currently, the discussion of results concerns basically the predictive value of specific variables. If an extensive discussion of the utility value of prospective predictors does not appear in the introduction, it must appear in the discussion of results. This means a reference to the literature from the e-HRM area and to the literature from the fluctuation theory area. Such a reference must also appear in the conclusions and an emphasis on the actual value of the text, i.e. the description of the identification of such predictors based on a small set of data. The limitations of both the study, its methodology and conclusions must be described in the conclusions.

Response: These are all helpful suggestions which have some overlap with other comments including those of the other reviewer. We have explicitly included references to e-HRM, we have added to the utility aspects by expanding on measures for retention, and we have emphasised the limitations of the work further, including in the Conclusion.

6. The description in the abstract is incorrect („Conclusions: These findings highlight the multifactorial nature of staff retention and suggest the potential of local administrative data to improve workforce planning, for example through hyperlocal recruitment strategies. Our analysis offers actionable insights for NHS managers seeking to stabilise staffing levels and reduce attrition through targeted interventions beyond pay and tenure.”) – proposals for actions towards employees at risk of turnover are not included in the Conclusions, and the suggestion to recruit employees “hyperlocally” – although potentially correct, is usually absurd in practice (other solutions should be proposed as well and included in the discussion)

Response: We agree completely, but hope that by answering the previous points (especially adding suggested measures to retain staff) the Abstract will be more in-line with the work that has been presented. In particular, we have noted that policies should take more account of staff subtype (rather than being one-size-fits-all), and that solutions such as managing staff in urban and suburban areas be given higher priority by line managers. We have also slightly rephrased the abstract [line 15] to emphasise that this work shows the potential for local administrative data (as it is a pilot study only).

To sum up, the reviewer believes that the article is important and interesting, but before publication it requires embedding it in previous research from two areas of scientific research, which it belongs to: research on turnover and e-HRM (research on prospective predictors).

Response: We thank the reviewer for the comments; like all manuscripts, we believe it is much improved from a reading by 'fresh eyes' and hope that our changes have moved the research forward. Especially we agree that the value here is demonstrating that the small dataset used was still able to yield insight, and that this could be a springboard for future and more comprehensive works across multiple NHS Trusts, ideally within a Secure Data Environment where more data can be shared.

Comments to the Authors and responses

Limitations of the model: The limitations of the model that may question the practical applicability to HR decision-making are also not discussed in detail, as modest predictive performance (AUC 0.65) is admitted.

Response: We agree with the reviewer that the AUC demonstrates weak individual predictive ability, just as a car insurance model predicts individual accidents very weakly. We still think that the insights can show which classes or characteristics are associated with departure risk, but have added additional text [lines 305-306] to emphasise further that we do not claim predictive capabilities for our work. We had also noted that future work should require integration of other data sources using staff specific identifiers that we did not have access to (e.g. using employee unique identifiers to connect staff records to other data, such as survey responses). Indeed, such additional inclusions would take us closer to being able to model turnover using fluctuation theory, as suggested by the other reviewer.

Generalisability Concern: Because it is applied to one NHS Trust the analysis has a major limitation in its generalisability. Multi-site data might improve robustness and increased applicability.

Response: We agree, and have amended the text to emphasise this limitation explicitly [lines 325 and 339-341].

The Insights generated by the Predictive are Shallow: As much as the analysis establishes the main predictors such as the length of service and the distance to work, it does not examine the effects of interaction between factors, which could provide deeper insights.

Response: Our data access limited the amount of longitudinal analysis we could conduct, which prevented a comprehensive mixed-effects analysis model including interaction effects. We agree that this would be an excellent area for future research, and this is explicitly stated as a future work suggestions [line 343].

Inadequate Theoretical Approach: The paper would be enhanced by having a more theoretical approach to the staff retention models to offer some context to the findings in the larger body of management of human resources.

Response: The other reviewer made similar points, especially around acknowledging the theory around fluctuation models and also in acknowledging the literature around e-HRM. We have made a number of textual changes in the Introduction and Discussion to acknowledge alternative approaches [lines 55-74 and 345]. We did not want to go too far in discussing these, as our work is more of a practical demonstration of potential; if the reviewer feels that a more extensive discussion of the theory is warranted, we can certainly add this, but hope that the paragraphs and references added will be sufficient.

Misuse of Temporal Trends: The independent snapshot mentalities of analysing annual data overlook the potential benefits of fish within the longitudinal means of analysing the changes in retention over time.

Response: We acknowledge the reviewer’s desire to see proper longitudinal analysis, and not ‘point-in-time’ prediction. Unfortunately, we were limited in conducting comprehensive longitudinal analysis, as the de-identified data prevented us from tracking staff individually over time. Such an approach would ideally require a Secure Data Environment and an embedded collaboration with the NHS. The practicalities of securing access and funding are that such initiatives are more likely when a pilot or limited study has demonstrated some initial potential, hence this work. We are at the pilot stage, but hope we can progress this approach to a major work programme in the future. We have made this more clear in the Conclusions section [line 341].

Lack of Practical Interventions: Although there is the identification of risk factors, the study fails to go further ahead and list retention strategies that are practical and actionable based on findings of the model.

Response: We agree with the reviewer, and on reflection our manuscript did not include practical applications of changes to be made. We have consulted with the Workforce team at ASPH NHS Trust and added a number of potential suggestions throughout the Discussion and thank the reviewer for pointing this out. We have also added links to current discussions on NHS policies (for example the        10 Year Health Plan for England: https://www.gov.uk/government/publications/10-year-health-plan-for-england-fit-for-the-future). These are valuable improvements to our manuscript and we should have included them at the outset. We also note however that this is a pilot study of a single Trust, so whilst we have highlighted potential actions, we do not want to overreach in our recommendations. We hope that our revised manuscript has the right balance.

Inappropriate Use of First-Person Pronouns: The frequent use of "we" and "our" throughout the paper is unsuitable for academic business research writing and should be replaced with objective, third-person phrasing.

Response: We have noted the reviewer’s preference for third person phrasing – this is a convention that is changing over time in the literature, and we additionally hope that our manuscript will find a readership beyond academic business research writing (for example NHS HR managers and policymakers). We would like to make the writing as accessible as possible for the widest possible audience.

If the editor feels that this is inconsistent with the journal standard, we will of course rewrite the document to use the third person throughout, but if the editor is content we would prefer to leave the phrasing “as is”.

 

Reviewer 2 Report

Comments and Suggestions for Authors

Based on dataset derived from Electronic Staff Records at Ashford & St Peter’s NHS Foundation Trust, using a machine learning approach, the article identifies a few predictors of staff retention. The main result of the text is not the approach itself (contrary to declaration of the Authors), nor the list of predictors identified, but a proof of the fruitfulness of this procedure using a relatively small data set.

The article is a well-written report from an empirical study (teaching and using AI to a dataset) and presentation of the results obtained, which would be a good starting point for a consultancy report. As an article for a scientific journal, however, it has several weaknesses that must be eliminated before its publication.

1. The entire text does not contain any references to the theory of fluctuations, neither in the introductory part nor in the discussion of the results. Such conduct disqualifies the article as a scientific text, which must indicate how the obtained result (and not just the methodology) complements scientific knowledge in a given field.
2. There are also no references to other examples of analogously used AI methodology, although they have been postulated in the literature and described at least in terms of their assumptions for many years see: (Strohmeier, 2007) in "Human Resource Management Review"
3. It should be clearly stated that the adopted procedure does not identify prospective predictors, but only excludes from the previously established list of variables those that do not provide such predictions. This remark is important because it indicates that the usefulness of the procedure is based on the list of variables that are constructed a priori on the basis of existing data (the selection of variables for analysis in this study was therefore random and not dictated by any analysis other than intuitive)
4. The discussion of how to use retention predictors in individual cases requires further discussion. The analogy („In this study we show which factors contribute most to increased probabilities of staff departures. The work is not intended to suggest that machine learning can be used to predict individual reasons for staff leaving. In this sense, the algorithm is similar to car insurance models: it can identify factors that increase the likelihood or risk of an event, but will not perform well in forecasting outcomes for specific individuals.”) although accurate, it is misleading. In the case of car insurance, the reaction to a high value of a prospective predictor is an increase in the price of insurance for an individual who belongs to the group indicated by the predictor (discouraging them from insuring with that company), while the company must introduced some actions to influence employees who are already employed and not to discourage them from belonging to the company (rejecting job candidates from a given type of group).
5. From a formal perspective, the article should be constructed differently in the discussion of results and conclusions. Currently, the discussion of results concerns basically the predictive value of specific variables. If an extensive discussion of the utility value of prospective predictors does not appear in the introduction, it must appear in the discussion of results. This means a reference to the literature from the e-HRM area and to the literature from the fluctuation theory area. Such a reference must also appear in the conclusions and an emphasis on the actual value of the text, i.e. the description of the identification of such predictors based on a small set of data. The limitations of both the study, its methodology and conclusions must be described in the conclusions.
6. The description in the abstract is incorrect („Conclusions: These findings highlight the multifactorial nature of staff retention and suggest the potential of local administrative data to improve workforce planning, for example through hyperlocal recruitment strategies. Our analysis offers actionable insights for NHS managers seeking to stabilise staffing levels and reduce attrition through targeted interventions beyond pay and tenure.”) – proposals for actions towards employees at risk of turnover are not included in the Conclusions, and the suggestion to recruit employees “hyperlocally” – although potentially correct, is usually absurd in practice (other solutions should be proposed as well and included in the discussion)

To sum up, the reviewer believes that the article is important and interesting, but before publication it requires embedding it in previous research from two areas of scientific research, which it belongs to: research on turnover and e-HRM (research on prospective predictors).

Author Response

admsci-3739163_revision1

Machine learning applied to NHS electronic staff records identifies key areas of focus for staff retention

We thank the reviewers for their comments and insights, and have incorporated the feedback and hope that we have improved the manuscript. Our detailed responses are set out below point-by-point. All references to line numbers are for the ‘track changes on’ version of the document.

Comments to the Authors and responses

Based on dataset derived from Electronic Staff Records at Ashford & St Peter’s NHS Foundation Trust, using a machine learning approach, the article identifies a few predictors of staff retention. The main result of the text is not the approach itself (contrary to declaration of the Authors), nor the list of predictors identified, but a proof of the fruitfulness of this procedure using a relatively small data set.

The article is a well-written report from an empirical study (teaching and using AI to a dataset) and presentation of the results obtained, which would be a good starting point for a consultancy report. As an article for a scientific journal, however, it has several weaknesses that must be eliminated before its publication.

1. The entire text does not contain any references to the theory of fluctuations, neither in the introductory part nor in the discussion of the results. Such conduct disqualifies the article as a scientific text, which must indicate how the obtained result (and not just the methodology) complements scientific knowledge in a given field.

Response: The reviewer is of course correct that the theory of fluctuations could play a valuable part in this sort of analysis. We had considered this initially, for example how compensation, job satisfaction, external factors (i.e. the jobs market), life events and so on might feed into ‘staff departures’ as a stochastic process. Fluctuation theory would also have the advantage of allowing for tipping point thresholds. An example of this lies in our findings that in some classes of employees, such tipping points may be more prevalent than others (i.e. our finding that for medics, short-staffing which may be caused by prior departures is much more likely to trigger further departures).

Unfortunately, electronic staff records alone do not really allow for this type of modelling, especially as we did not have the unique staff identifiers to track single employees over time. We have expanded the Introduction to address fluctuation theory and other models (including references), added some comments on tipping points, and highlighted the potential for these methods of research in limitations / future work [lines 55-74 and 346 in the track changes version]. We have also added additional material on theory relevant to the topic, especially around extrinsic hygiene factors and intrinsic motivation factors [lines 43-47].

2.   There are also no references to other examples of analogously used AI methodology, although they have been postulated in the literature and described at least in terms of their assumptions for many years see: (Strohmeier, 2007) in "Human Resource Management Review"

Response: We agree that our manuscript focused mainly on employment issues in the NHS and not the wider field of AI methodologies in HR, including e-HRM. This has now been included [lines 70-74 and 316]. The suggested reference is an excellent article and we have included it, and thank the reviewer for bringing it to our attention.

3. It should be clearly stated that the adopted procedure does not identify prospective predictors, but only excludes from the previously established list of variables those that do not provide such predictions. This remark is important because it indicates that the usefulness of the procedure is based on the list of variables that are constructed a priori on the basis of existing data (the selection of variables for analysis in this study was therefore random and not dictated by any analysis other than intuitive)

Response: we agree with the general point that the reviewer makes, but this is an unavoidable limitation, that our list of variables was constructed a priori on the basis of existing data. Collecting new data would have necessitated a completely different approach, and we believe that a pilot study has value in demonstrating the types of data that might be informative. We have added an explicit acknowledgement that our variables were predetermined by the availability in the NHS electronic staff records [line 326 in the track changes version]. The ability to bring in additional variables will be improved in future as the NHS is currently undergoing pre-procurement of new
e-HRM systems, see:

 https://www.nhsbsa.nhs.uk/future-nhs-workforce-solution-transformation-programme

4. The discussion of how to use retention predictors in individual cases requires further discussion. The analogy („In this study we show which factors contribute most to increased probabilities of staff departures. The work is not intended to suggest that machine learning can be used to predict individual reasons for staff leaving. In this sense, the algorithm is similar to car insurance models: it can identify factors that increase the likelihood or risk of an event, but will not perform well in forecasting outcomes for specific individuals.”) although accurate, it is misleading. In the case of car insurance, the reaction to a high value of a prospective predictor is an increase in the price of insurance for an individual who belongs to the group indicated by the predictor (discouraging them from insuring with that company), while the company must introduced some actions to influence employees who are already employed and not to discourage them from belonging to the company (rejecting job candidates from a given type of group).

Response: On the point that the analogy of insurance is not perfect, we agree, but we are only trying to make the point that a model that cannot accurately predict individual outcomes can still have value in managing overall behaviours. We do agree however that the reaction is not exclusively one of price (salary) but of other changes employers can make in the NHS. The other reviewer made similar points, and we have extended our comments throughout the Discussion. As this is a pilot study of a single Trust, we would also comment that providing specific and immediate recommendations for the NHS would probably be over-reach, and we hope that our revised Discussion strikes the right balance.

5. From a formal perspective, the article should be constructed differently in the discussion of results and conclusions. Currently, the discussion of results concerns basically the predictive value of specific variables. If an extensive discussion of the utility value of prospective predictors does not appear in the introduction, it must appear in the discussion of results. This means a reference to the literature from the e-HRM area and to the literature from the fluctuation theory area. Such a reference must also appear in the conclusions and an emphasis on the actual value of the text, i.e. the description of the identification of such predictors based on a small set of data. The limitations of both the study, its methodology and conclusions must be described in the conclusions.

Response: These are all helpful suggestions which have some overlap with other comments including those of the other reviewer. We have explicitly included references to e-HRM, we have added to the utility aspects by expanding on measures for retention, and we have emphasised the limitations of the work further, including in the Conclusion.

6. The description in the abstract is incorrect („Conclusions: These findings highlight the multifactorial nature of staff retention and suggest the potential of local administrative data to improve workforce planning, for example through hyperlocal recruitment strategies. Our analysis offers actionable insights for NHS managers seeking to stabilise staffing levels and reduce attrition through targeted interventions beyond pay and tenure.”) – proposals for actions towards employees at risk of turnover are not included in the Conclusions, and the suggestion to recruit employees “hyperlocally” – although potentially correct, is usually absurd in practice (other solutions should be proposed as well and included in the discussion)

Response: We agree completely, but hope that by answering the previous points (especially adding suggested measures to retain staff) the Abstract will be more in-line with the work that has been presented. In particular, we have noted that policies should take more account of staff subtype (rather than being one-size-fits-all), and that solutions such as managing staff in urban and suburban areas be given higher priority by line managers. We have also slightly rephrased the abstract [line 15] to emphasise that this work shows the potential for local administrative data (as it is a pilot study only).

To sum up, the reviewer believes that the article is important and interesting, but before publication it requires embedding it in previous research from two areas of scientific research, which it belongs to: research on turnover and e-HRM (research on prospective predictors).

Response: We thank the reviewer for the comments; like all manuscripts, we believe it is much improved from a reading by 'fresh eyes' and hope that our changes have moved the research forward. Especially we agree that the value here is demonstrating that the small dataset used was still able to yield insight, and that this could be a springboard for future and more comprehensive works across multiple NHS Trusts, ideally within a Secure Data Environment where more data can be shared.

Comments to the Authors and responses

Limitations of the model: The limitations of the model that may question the practical applicability to HR decision-making are also not discussed in detail, as modest predictive performance (AUC 0.65) is admitted.

Response: We agree with the reviewer that the AUC demonstrates weak individual predictive ability, just as a car insurance model predicts individual accidents very weakly. We still think that the insights can show which classes or characteristics are associated with departure risk, but have added additional text [lines 305-306] to emphasise further that we do not claim predictive capabilities for our work. We had also noted that future work should require integration of other data sources using staff specific identifiers that we did not have access to (e.g. using employee unique identifiers to connect staff records to other data, such as survey responses). Indeed, such additional inclusions would take us closer to being able to model turnover using fluctuation theory, as suggested by the other reviewer.

Generalisability Concern: Because it is applied to one NHS Trust the analysis has a major limitation in its generalisability. Multi-site data might improve robustness and increased applicability.

Response: We agree, and have amended the text to emphasise this limitation explicitly [lines 325 and 339-341].

The Insights generated by the Predictive are Shallow: As much as the analysis establishes the main predictors such as the length of service and the distance to work, it does not examine the effects of interaction between factors, which could provide deeper insights.

Response: Our data access limited the amount of longitudinal analysis we could conduct, which prevented a comprehensive mixed-effects analysis model including interaction effects. We agree that this would be an excellent area for future research, and this is explicitly stated as a future work suggestions [line 343].

Inadequate Theoretical Approach: The paper would be enhanced by having a more theoretical approach to the staff retention models to offer some context to the findings in the larger body of management of human resources.

Response: The other reviewer made similar points, especially around acknowledging the theory around fluctuation models and also in acknowledging the literature around e-HRM. We have made a number of textual changes in the Introduction and Discussion to acknowledge alternative approaches [lines 55-74 and 345]. We did not want to go too far in discussing these, as our work is more of a practical demonstration of potential; if the reviewer feels that a more extensive discussion of the theory is warranted, we can certainly add this, but hope that the paragraphs and references added will be sufficient.

Misuse of Temporal Trends: The independent snapshot mentalities of analysing annual data overlook the potential benefits of fish within the longitudinal means of analysing the changes in retention over time.

Response: We acknowledge the reviewer’s desire to see proper longitudinal analysis, and not ‘point-in-time’ prediction. Unfortunately, we were limited in conducting comprehensive longitudinal analysis, as the de-identified data prevented us from tracking staff individually over time. Such an approach would ideally require a Secure Data Environment and an embedded collaboration with the NHS. The practicalities of securing access and funding are that such initiatives are more likely when a pilot or limited study has demonstrated some initial potential, hence this work. We are at the pilot stage, but hope we can progress this approach to a major work programme in the future. We have made this more clear in the Conclusions section [line 341].

Lack of Practical Interventions: Although there is the identification of risk factors, the study fails to go further ahead and list retention strategies that are practical and actionable based on findings of the model.

Response: We agree with the reviewer, and on reflection our manuscript did not include practical applications of changes to be made. We have consulted with the Workforce team at ASPH NHS Trust and added a number of potential suggestions throughout the Discussion and thank the reviewer for pointing this out. We have also added links to current discussions on NHS policies (for example the        10 Year Health Plan for England: https://www.gov.uk/government/publications/10-year-health-plan-for-england-fit-for-the-future). These are valuable improvements to our manuscript and we should have included them at the outset. We also note however that this is a pilot study of a single Trust, so whilst we have highlighted potential actions, we do not want to overreach in our recommendations. We hope that our revised manuscript has the right balance.

Inappropriate Use of First-Person Pronouns: The frequent use of "we" and "our" throughout the paper is unsuitable for academic business research writing and should be replaced with objective, third-person phrasing.

Response: We have noted the reviewer’s preference for third person phrasing – this is a convention that is changing over time in the literature, and we additionally hope that our manuscript will find a readership beyond academic business research writing (for example NHS HR managers and policymakers). We would like to make the writing as accessible as possible for the widest possible audience.

If the editor feels that this is inconsistent with the journal standard, we will of course rewrite the document to use the third person throughout, but if the editor is content we would prefer to leave the phrasing “as is”.

 

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

For an article - not a research report - you should deep a litttle bit more in the background - check fluctuation theories for the point which could give you an arguments for the predictor and for disscussion.

Author Response

For an article - not a research report - you should deep a little bit more in the background - check fluctuation theories for the point which could give you an arguments for the predictor and for discussion.

Response: 

We acknowledge that Reviewer 2 has a focus on fluctuation theory, and have amended the document to extend the references to this field and link our analyses to this approach [lines 257-258, 274-277 and 316-318]. Whilst we certainly agree it is helpful to acknowledge fluctuation theory, a full discussion or implementation is beyond the scope of this manuscript (because we do not have the data on e.g. team dynamics, stochastic variables as time series), and this is also flagged as a limitation.

Lines 257-258:

Issues such as these can receive less attention than headline pay levels, but can nonetheless contribute to cumulative dissatisfaction. 

Lines 274-277:

ASPH targets 10 days sickness absence over a rolling 12 month period as a trigger for formal sickness management, typical within NHS Trusts. The data analysed here indicate that the increased probability of departure occurs sooner than this point, consistent with the fluctuation theory stance on departures not being single decisions, but an accumulation of micro-decisions driven by stress over time. 

Lines 316-318:

Multi-method approaches to data collection such as the integration of staff surveys or other sources could improve model performance via a comprehensive e-HRM approach (incorporating greater user of fluctuation theory, for example to model the contribution from changes in environmental instability), but were not possible here