Back to all Advice

Writing your paper: conclusions section

Posted 7/24/2023

The conclusion is your paper’s second most important section, after the abstract. This is where you most concretely state what you have contributed: the whole point of the paper. Many expert readers first skim a paper's conclusions to identify the takeaways, then go back to read relevant parts of the paper to find the support for those ideas. And if they don’t see clear ideas or contributions in this section, they are likely to simply quit reading altogether.

It may be stress-inducing to imagine someone reading your paper and moving on because your writing doesn’t interest them. On the other hand, once you know your reader’s motivations, it is clear how to serve them. With this in mind, I recommend the following structure for your conclusions:

1. Restate: What are the key ideas you presented?

2. Summarize: What studies and analyses did you describe in this paper, and what did you learn?

3. Broader implications: How should we understand the field of study now, in light of the insights from this work?

Use your opening paragraph to restate key ideas you presented. Focus on the big picture. Write something to make your reader care about your work and findings. Draw a connection back to the problems and questions you raised in the introduction, and explain how you resolved them. Re-define acronyms or key terms, for the reader who is reading this section first before reading the paper, or who has simply forgotten some earlier jargon. Ideally, avoid acronyms completely, unless you have given an acronym to a new procedure you propose, in which case it makes sense to restate it.

Next, summarize your findings. Depending on the document’s length and complexity, this may be a single paragraph or a group of paragraphs that each summarize one aspect of the study. Don’t repeat all the detail from the paper, but try to restate some key results to keep the text from sounding vague. Avoid introducing completely new ideas or assertions—if an idea is important enough that you feel compelled to write about it here, it should have been introduced earlier in the paper. This section of your paper is a gathering of key results from above, and the reader can go back to learn the details if needed.

Many authors feel compelled to list the limitations of their study in their final paragraph, to balance all the previous assertions they made. This comes from noble instincts, but is a mistake. Listing limitations in an isolated paragraph neglects the important task of reconciling the insights versus the limitations. More importantly, this makes the paper finish on a sour note, with these last-minute concerns freshest in the reader’s mind and superseding the prior positive contributions.

Instead, finish by discussing the broader implications of the work. Address limitations and unresolved issues, but in the context of explaining the work’s value, rather than with a tone of undercutting your prior statements' contributions. Your message should be, “yes, there are limitations to the approach I used, and some related unresolved questions, but the findings are still valuable because…” This way, the reader leaves with a satisfying and balanced understanding of the work’s potential and its limits.


The following examples illustrate the above structure. These are short, to concisely illustrate the structure, but conclusions can be longer. I preceded each paragraph below with a comment to indicate its scope. Add labels like this to your own draft paper’s paragraphs, to check whether you stay on topic within each paragraph.

Costa, R., Wang, C., and Baker, J. W. (2022). “Integrating Place Attachment into Housing Recovery Simulations to Estimate Population Losses.” Natural Hazards Review, 23(4), 04022021. DOI

[Restate the study and approaches used]

This study integrated place attachment considerations into housing recovery simulations. Place attachment was used as a surrogate for willingness to rebuild. We identified households with low place attachment and whose housing recovery process is expected to be the most challenging. Our premise is that households with low place attachment are less willing to take on debt and wait extended periods to restore their livelihoods. We introduced a classification algorithm that combines the synthetic minority oversampling technique and adaptive boosting (SMOTEBoost) to estimate household place attachment from data from the American Housing Survey. We used the place attachment estimates to study postearthquake decisions of households. We introduced a housing recovery simulation framework to estimate repair costs and housing recovery time for single-family buildings. We combined the place attachment, repair cost, and repair time results to estimate population losses. The place attachment assessment and the housing recovery simulations are decoupled. Thus, the place attachment assumptions can be revised without rerunning the computationally expensive housing recovery simulation. Although we focused on postearthquake decisions, the SMOTEBoost algorithm can be used to assess place attachment and investigate postdisaster decisions after other types of extreme events, such as hurricanes and floods.

[Description of case study]

The application of the framework was demonstrated in a case study of the potential population loss in San Francisco during the recovery from hypothetical earthquakes on the San Andreas Fault. The case study quantified housing repair costs (relative to household income), time to secure funding, and building repair time for 124,563 single-family households in San Francisco. The potential population loss was investigated under different scenarios. The results indicated that low-income renters occupying older buildings are the most prone to moving away after a disaster.

[Discussion of general implications]

The framework presented in this study addresses the concern with the loss of populations with low place attachment which has emerged recently in studies of the regional impacts of earthquakes (Johnson et al. 2020). Previous studies ignored the influence of place attachment or assumed a priori which demographic groups are most prone to moving away after a disaster. As a consequence, existing approaches provide limited insight into the demographic groups expected to struggle and perhaps move away during postearthquake recovery. The framework in this paper is based on a review of studies of previous disasters. It employs data from the American Housing Survey which are publicly available for multiple locations in the US. It is empirically based, can be employed in multiple regions, and is more nuanced in determining the demographic groups most prone to residential mobility. The framework can be incorporated in predisaster studies to estimate population losses using what-if scenarios (Johnson et al. 2020) and to evaluate the benefits of taking actions to improve neighborhood cohesion (Lee and Otellini 2016). Some challenges remain in the application of the proposed framework, as highlighted in the Discussion section. Nonetheless, it offers a more robust procedure that can replace semiheuristic approaches and can help formalize the simulation of housing recovery.

Bowers, C., Serafin, K. A., and Baker, J. W. (2022). “A Performance-Based Approach to Quantify Atmospheric River Flood Risk.” Natural Hazards and Earth System Sciences, 22(4), 1371–1393. DOI

[Restate the objective and approaches used. Note the re-definition of the PARRA acronym introduced earlier in the paper.]

This paper introduced the Performance-based Atmospheric River Risk Analysis (PARRA) framework to quantify AR-induced flood risk. The framework captures the physical processes connecting atmospheric forcings, hydrologic impacts, and economic consequences of AR-driven fluvial flooding. Using a performance-based engineering paradigm, this approach offers several benefits. It quantifies the uncertainty surrounding the physical processes by following a deliberate, ordered simulation procedure. It connects multiple physical processes in sequence by constructing a chain of discrete component models that link together at defined pinch points. Pinch points in the model chain serve to facilitate intercompatibility across different disciplines and to better understand the complexity of the hazard and risk.

[Recap of the proposed model]

Section 3 discussed the fit and calibration of five individual component models: precipitation estimation, hydrologic routing, inundation modeling, depth–damage relationships, and loss estimation. We demonstrated the uncertainty quantification capabilities and the modularity of the component models through case studies of historic AR events affecting the lower Russian River in Sonoma County. We performed step-by-step comparisons between each of these component models and ground-truth data from the case study AR events to show how the differences between the observed and simulated values produced new insights into what drove certain events toward extreme consequences and not others.

[Recap of the presented example, including a few specific results]

In Section 4, we ran a fully probabilistic simulation of a damaging February 2019 AR event using only the observed AR characteristics and antecedent soil moisture as input and examined both the probabilistic range and the spatial distribution of the predicted losses. We also used the PARRA framework to generate a first-of-its-kind loss exceedance curve for the lower Russian River to understand the full spectrum of potential loss events rather than a single scenario event or a long-term annual average. We quantified the reduction in flood risk from a hypothetical mitigation decision: when 200 homes were elevated above the 100-year flood elevation, the average annual loss was reduced by half, and the average benefit for events with return periods of 100 years or longer was found to be USD 50–75 million per AR. Section 5 highlighted additional nuances about implementation and validation for potential users of the PARRA framework to consider.

[Discussion of general implications]

Although the case study showed examples of the specific insights that can be gained from implementing the component models for a community risk assessment, the theory and scientific merit of the PARRA framework stand on their own, independent of the specific benefits and tradeoffs inherent in any local implementation. We have proposed a new method for the structured assessment of AR-driven flood risk that is physically based, modular, probabilistic, and prospective. The PARRA framework is ideally suited to performing a forward-looking evaluation of potential impacts for events outside of the historic record, or events that have not yet occurred but could in an evolving climate. It can similarly be used to estimate changes in future flood risk due to land use shifts, population change, and more. The framework presented here has been shown to work in a real-world implementation and has the potential to greatly expand our understanding of the risks associated with AR-induced flooding.


Use the following link to register to receive very occasional updates about new offerings on this page. I will not share your information with anyone.

Go to the registration form