Quantifying the relationship between US residential mobility and fire service call volume

Abstract

Purpose

The COVID-19 pandemic dramatically affected the fire service: stay-at-home orders and potential exposure hazards disrupted standard fire service operations and incident patterns. The ability to predict incident volume during such disruptions is crucial for dynamic and efficient staff allocation planning. This work proposes a model to quantify the relationship between the increase in “residential mobility” (i.e. time spent at home) due to COVID-19 and fire and emergency medical services (EMS) call volume at the onset of the pandemic (February – May 2020). Understanding this relationship is beneficial should mobility disruptions of this scale occur again.

Design/methodology/approach

The analysis was run on 56 fire departments that subscribe to the National Fire Operations Reporting System (NFORS). This platform enables fire departments to report and visualize operational data. The model consists of a Bayesian hierarchical model. Text comments reported by first responders were also analyzed to provide additional context for the types of incidents that drive the model’s results.

Findings

Overall, a 1% increase in residential mobility (i.e. time spent at home) was associated with a 1.43% and 0.46% drop in EMS and fire call volume, respectively. Around 89% and 21% of departments had a significant decrease in EMS and fire call volume, respectively, as time spent at home increased.

Originality/value

A few papers have investigated the impact of COVID-19 on fire incidents in a few locations, but none have covered an extensive number of fire departments. Additionally, no studies have investigated the relationship between mobility and fire department call volumes.

Citation

Franqueville, J.I., Scott, J.G. and Ezekoye, O.A. (2024), "Quantifying the relationship between US residential mobility and fire service call volume", International Journal of Emergency Services, Vol. 13 No. 3, pp. 285-303. https://doi.org/10.1108/IJES-04-2024-0024

1. Introduction

The COVID-19 pandemic presented a sudden and unprecedented challenge to the fire service. For one, medical emergencies involving a contagious disease present an exposure hazard to first responders. Furthermore, the pandemic led to dramatic changes in mobility patterns due to stay-at-home orders, increased remote work, and aversion to public gatherings. Fire departments increasingly rely on data to inform planning decisions such as staff allocation, firefighter training, and public education efforts (Phillips and Hare, 2017). Quantifying how COVID-19 has affected fire service call volume can be valuable for fire departments to quickly plan for staff allocation in the event of future potential disruptions or public health crises.

The focus of this paper is to quantify the relationship between changes in mobility due to the pandemic and fire and EMS incident call volume for fire departments subscribing to the National Fire Operations Reporting System (NFORS). The analysis used two datasets: the Google Community Mobility Reports (Google, 2022) and fire and EMS data from NFORS provided by the International Public Safety Data Institute (IPSDI) (IPSDI, 2024). A Bayesian hierarchical model, which included a Gaussian process framework, was used in the analysis. The aim of this study is to help inform resource allocation decisions within the fire service in times of mobility disruptions caused by unusual events such as a pandemic. The analysis is limited to the initial stage of the pandemic (February to May 2020).

2. Literature review

Few papers have investigated the impact of COVID-19 on fire incidents. Suzuki and Manzello (2022a) found that the monthly number of residential fires at the onset of the pandemic in Tokyo, San Francisco, London, and New York City did not significantly change compared to the average number of incidents in previous years. They also found that cooking fires increased in San Francisco and Tokyo but not in New York City. Suzuki and Manzello (2022b) also found that the weekly (as opposed to monthly) number of residential fires increased in London at the very beginning of lockdown periods when comparing fire counts to previous years. Hollerbach and Jahnke (2021) highlighted the need for more research on the impact of the pandemic on the fire service to help inform firefighter training. Poulter et al. (2021) found that the number of active forest fires in the southwestern United States declined during the pandemic compared to previous years. Haas et al. (2021) examined the relationship between the ratio of the number of non-EMS (including fire) calls to EMS calls and various socioeconomic factors, as well as COVID-19 cases. They found a negative relationship between COVID-19 case volume and the ratio of the number of non-EMS to EMS calls. However, they found that fire department and socioeconomic factors were stronger predictors. Studies suggest that residential fires are correlated with human mobility (Ahrens and Maheshwari, 2013; Spearpoint and Hopkin, 2020; Wuschke et al., 2013); in regular times, residential fires are highest around dinnertime and are generally high in the afternoon and at night when people are home. According to NFPA (the National Fire Protection Association) (Ahrens and Maheshwari, 2013), cooking equipment is the leading cause of home fires in the United States, followed by heating equipment (such as central heating units, space heaters, and fireplaces), arson, and electrical distribution/lighting equipment. Because of restaurant closures and reluctance to eat in proximity to other people, it may be reasonable to expect cooking fire counts to have increased over the pandemic. However, most fires are handled without help from the fire service because small fires can easily be put out without assistance. Increased time spent at home may have enabled early detection and intervention for some fires and lowered the number of reported fires.

Beyond fire incidents, the fire service also typically responds to EMS calls. EMS calls have been affected by the pandemic in several respects. First, the literature shows that EMS calls reporting respiratory symptoms increased in some places during the pandemic. For example, this was the case in New York City (Prezant et al., 2020), western Pennsylvania (Satty et al., 2021), and Massachusetts (Goldberg et al., 2021). However, several other types of calls decreased as a secondary effect of the pandemic. For example, in Massachusetts, the number of calls for cardiac and stroke emergencies decreased after March 2020 (Goldberg et al., 2021). This decrease may be attributed to the general perception that seeking care may increase the risk of COVID-19 infection. Other types of calls increased. For example, Friedman et al. (2021) found that overdose-related calls increased in the United States. They noted that the increase in overdose calls was inversely correlated with general mobility. In terms of the overall number of EMS calls, some areas, such as New York City (Prezant et al., 2020), Copenhagen (Jensen et al., 2021), Paris (Lapostolle et al., 2020), and Catalonia (Rudilosso et al., 2020), experienced an increase in calls. However, in other areas, such as Canada (Ferron et al., 2021; Grunau et al., 2021), England (Thornton, 2020), Finland (Laukkanen et al., 2021), the United States (Hartnett et al., 2020; Lerner et al., 2020; Satty et al., 2021; Xie et al., 2021), and Japan (Igarashi et al., 2021), the overall volume of EMS calls or emergency services usage decreased.

To the authors’ knowledge, the impact of changing mobility patterns on fire incident frequency has not been studied, and few studies have examined the impact on EMS incident frequency (Friedman et al., 2021; Sharma et al., 2021). Yet, mobility data can be useful for dynamically predicting call volume. Companies such as Apple, Google, and SafeGraph (Apple, 2022; Google, 2022; SafeGraph, 2022) started making mobility data available to inform public health decisions related to the COVID-19 pandemic. These data usually consist of several indicators for mobility change; for example, the Google Community Mobility Reports report percentage change in mobility relative to before the pandemic in five categories: retail and recreation, supermarkets and pharmacies, parks, public transport, workplaces, and residential. These datasets have been used in several other fields, for example, in criminology (Halford et al., 2020; Nivette et al., 2021) and epidemiology (Badr et al., 2020; Grantz et al., 2020).

3. Methodology for data processing

Before building the model, data from the two sources used (NFORS and Google, further described in Section 3.1) were processed. The steps outlined below were followed to process the data:

• (1)

  Converting Google mobility data from daily to weekly counts

• (2)

  Pre-processing incident data

• (3)

  Calculating baselines for the incident data

• (4)

  Calculating the percentage change in EMS and fire incidents at the onset of the pandemic compared to before the pandemic using the baselines

• (5)

  Matching the processed incident data to the mobility data by location

• (6)

  Choosing a mobility category to run the final model

These steps are outlined in more detail below, starting with an explanation of the data sources used.

3.1 Data sources

NFORS Analytics (IPSDI, 2024) is a platform that enables fire departments to report and visualize operational data. The names of the fire departments subscribing to this platform are not shared in this paper to provide anonymity. Many variables are included in NFORS, such as timestamps related to each incident, socioeconomic quantities extracted from census data associated with the address of the incident, weather information tied to the incident time and location, and many others. In this study, the variables used were fire department, incident date, comments, and category. The fire department and date variables are self-explanatory. The comments variable is an unstructured text field used to describe the incident. The “category” variable can be “EMS”, “FIRE”, or “OTHER”, although only the “FIRE” and “EMS” incident types were used. “FIRE” incidents can be outdoor and indoor fires, including fire alarms. “EMS” incidents include a variety of incidents such as cardiac, stroke, trauma, vehicle crash, etc.

The Google Community Mobility Reports (Google) are generated using data from users who have agreed to provide location history for their Google account. For the United States, the Google Community Mobility Reports consist of county-level mobility trends in the following categories: retail and recreation, supermarkets and pharmacies, parks, public transport, workplaces, and residential. Although the term “mobility” suggests movement, the Google data is a measure of the change in time spent in the various categories compared to before the pandemic. For each day and each county, the mobility change for the categories is expressed as a percentage change from a baseline. For example, a positive percentage change in residential mobility means that people stayed at home more compared to before the pandemic. Google calculated the baseline by taking the median value for the corresponding day of the week during the five weeks from January 3 – February 6, 2020.

The first step was to process the NFORS and Google mobility data. The goal was to obtain a dataset with the following columns for each department and incident type (EMS or fire): date, percentage change from baseline for incidents, and percentage change in mobility for the various mobility categories. The incident types “EMS” and “FIRE” were considered separately, given that resource allocation depends on incident type.

3.2 Processing mobility data

As mentioned, the Google mobility data were already expressed as a “percentage difference from baseline”; therefore, calculating a baseline was not necessary. The Google mobility data are available for each county in the United States and each day; however, weekly data rather than daily data was preferred. The main advantage of working with weekly data over daily data is that the amount of data is greatly reduced, which made the primary model used in this work (see section 4) computationally manageable. Therefore, the only modification made to these data was to express them as a weekly “change form baseline” rather than daily. A 7-day moving average was calculated for each county and each mobility category. The weekly mobility change was taken to be the value of the moving average on Mondays to represent the weekly change in mobility. Therefore, the dates present in the “date” column of the final dataset are all Mondays, but the corresponding data are averages of the respective Mondays and the six previous days.

3.3 Initial processing of incident data

Before calculating the baseline for incident data, the incident data were processed by removing incident counts where two or more consecutive days had zero incidents, given that this is highly unlikely. There were 53 departments that had EMS data and 56 that had fire data. Note that the analysis of EMS incidents included three fewer departments because they did not report EMS data in 2020. The states represented in the data for fire incidents were Colorado, New Mexico, Virginia, Maryland, Ohio, Washington, Massachusetts, Florida, Indiana, California, Minnesota, Arizona, Oregon, Kentucky, Tennessee, New York, Kansas, Illinois, Arkansas, Connecticut, Missouri, New Jersey, and District of Columbia. The states represented for EMS incidents were the same except for Missouri, Connecticut, and New Jersey due to the three departments that were excluded. The departments varied in size; Figure 1 shows the average daily number of incidents for all 56 departments ranked from largest to smallest (calculated from 2019 data).

It should be noted that while there is no reason to believe that the COVID-19 virus directly influenced fire incident counts, this is not the case for EMS incidents. For some areas, such as New York City, spikes in EMS calls were observed during the pandemic due to the increased number of respiratory calls from people likely experiencing COVID-19 symptoms. It was preferable to decouple the effect of mobility and COVID-19 cases on EMS incidents to study the relationship between mobility and EMS incidents rather than both mobility and COVID-19 cases on EMS incidents. Therefore, EMS calls that (likely) reported COVID-19 symptoms were removed from the dataset using notes written by first responders in the comment field of the NFORS data. If comments for an incident contained keywords associated with COVID-19, the incident was removed. Nonetheless, the final model was run both on EMS data with and without COVID-19 calls removed. Figure 2 shows the number of daily COVID-19 incidents identified across all departments, the total daily number of EMS incidents, and the difference between the two. At the aggregated level, the daily total number of EMS incidents dropped by at most around 30% following the start of the pandemic.

3.4 Baseline calculation for incident data

After the initial incident data processing, a baseline was calculated to express the change in incidents in terms of a percentage change compared to 2019, in a similar manner to the Google mobility data. Fire data are somewhat seasonal, so it is essential to consider the month in addition to the day of the week to calculate a baseline. The baseline could have been calculated by taking a simple median or mean number of incidents for each fire department and incident type (EMS or fire) for each month and day of the week in 2019. However, there are at most (assuming no missing days) five data points per month for each day of the week, which would have led to noisy baseline estimates. Instead, a regression model was used to predict the number of incidents for each department and each day of each month (the “baseline”). A regression model provides a function that describes the relationship between independent variables (here, variables describing the day and month) and a dependent variable (here, the number of incidents). Two regression approaches could have been used. In the first approach, the data of all fire departments could have been pooled together, but this method does not consider heterogeneity between departments. In the second approach, a separate regression could have been run for each department, but this could lead to noisy (overfit) estimates. A hierarchical regression was instead implemented to avoid these issues. In this hierarchical regression, each fire department has its own regression coefficients, but they are informed by the whole dataset. This method leads to more stable baseline estimates and allows for missing data. In this work, a Bayesian formulation of this hierarchical model was used. At a high level, a Bayesian model uses Bayes’ theorem to update the probability estimates for a parameter based on evidence/data. Bayesian models have been used in fire research; for example, Buffington et al. (2021) used hierarchical models to predict residential fires in the United States, and Pimont et al. (2020) used hierarchical models to predict wildfires in southern France. In a Bayesian regression, one sets “prior” distributions on the coefficients of the regression. Prior distributions represent one’s belief about what the regression coefficients may be before incorporating any data. The goal of the Bayesian regression is to obtain posterior distributions for the regression coefficients. The posteriors are distributions that represent what is known about the regression parameters after observing data; they are “updated” versions of the priors. A posterior is proportional to the product of the prior distribution and the likelihood. The likelihood is defined as the probability of observing the data given the parameters (coefficients) of the model. To obtain posteriors, one must use “Bayesian inference”. Bayesian regressions sometimes have analytical solutions to the posteriors; however, this is not the case here. When analytical solutions are not available, numerical methods must be used. Markov Chain Monte Carlo (MCMC) algorithms are a common approach to sample from posterior distributions when analytical solutions are not available. This was the approach used in this paper. The Metropolis-Hastings algorithm is an example of a simple, widely used MCMC algorithm. In this algorithm, a “chain” is first initialized; then, using a transition mechanism, the chain moves from the current point to the next. An acceptance criterion is then used to either accept or reject the new point. This process is repeated to generate a sequence of samples, which together converge to the desired posterior distribution. The MCMC algorithm used in this paper was NUTS (No-U-Turn Sampler) (Hoffman and Gelman, 2014), as opposed to Metropolis-Hastings. This algorithm is similar to Metropolis-Hastings, but it selects proposed values more efficiently, which reduces the number of samples needed and reduces computational cost.

The package PyMC (Salvatier et al., 2016) was used to implement the hierarchical regression model. PyMC is a Python package used for Bayesian inference. Various MCMC algorithms are implemented in this package, but the default algorithm is NUTS, which was used here. In PyMC, users build their models using the probability distributions available in the package. This makes PyMC adaptable to a wide variety of models, including hierarchical regressions.

Several regression models may be used to predict count data (here, incident counts). A common choice is a Poisson regression. However, it is restrictive in that it assumes that the conditional mean and variance are equal, which is rarely true in practice. Instead, a (hierarchical) negative binomial regression was used, which relaxes this assumption by including a dispersion parameter. The negative binomial model used in this work is described below:

(1)(yij|μij,αj)∼NB(μij,αj)

(2)μij=exp(Xijβj)

(3)(βj|θ,σ12,…,σk2 )∼N(θ,diag[σ12,…,σk2])

Here, yij is the number of incidents for department j on date i, and αj is the dispersion parameter for department j. Note that the parametrization of the negative binomial is not standard; the parametrization here is the one used by PyMC. The analysis is limited to the onset of the pandemic (February to May 2020), so the only relevant months to calculate the baseline were February to May 2019 (four months). One-hot encoding was used to create the design matrices of each fire department. The design matrix Xj for department j is a n-by-10 matrix: one column for an intercept, 6 for the days of the week, 3 for the months, and n for the number of incidents for each fire department. For a given row, if all three columns corresponding to the month are 0, this means that the month is February. Likewise, if all columns corresponding to the day of the week are 0, this means the day is Monday. Therefore, the intercept represents the number of incidents for Mondays in February. Each date was encoded by the corresponding day of the week and month. The vector Xij is the row for date i of the design matrix for department j. βj is a length-ten vector containing the coefficients for the intercept and months/day of the week. The length-ten vector of means θ corresponds to all (ten) coefficients, and diag[σ12,…,σk2] is the diagonal covariance matrix; k describes the coefficient number (intercept, month/day of week). This model is said to be “partially pooled” because of the assumption that the β coefficients of all departments come from the same normal distribution as shown in equation (3). Priors were defined for all unknown parameters: βj, θ, σ12,…,σk2 and are specified in the supplementary material.

This negative binomial model was fit on 2019 (pre-pandemic) data. Three isolated models were run to calculate three separate baselines: two for EMS incidents (including and not including COVID-19-related calls) and one for fire incidents. Figure 3 shows the process of fitting the negative binomial model to 2019 data and using the resulting baseline for the EMS incidents of one fire department. Figure 3(a) shows the data for 2019 (red dots), and the baseline fit to these data (black line). The blue lines show uncertainty around this baseline. Then, in Figure 3(b), the baseline fit on 2019 data is compared to 2020 data. For this department, the observed 2020 data is clearly inconsistent with the baseline model of no change because the incident counts are clearly lower than the baseline calculated from 2019 data. The baseline was then used in equation (4) to calculate the percentage difference between 2020 and 2019 incident counts, which for this department would clearly be negative (for their EMS incidents).

3.5 Calculating the percentage change in incidents

For each department and each incident type (EMS, EMS not including COVID-19 calls, and fire), the number of incidents from February to May 2020 was compared to the baseline calculated using the baseline model run on the 2019 data. The percentage difference between the predicted counts for 2020 was calculated using equation (4).

(4)% difference=actual count (pandemic)−baseline count (pre−pandemic)baseline count (pre−pandemic)×100 

Once the percentage change in EMS and fire incidents was obtained for all fire departments, a 7-day moving average was calculated, and the weekly change was taken to be the percentage change on Mondays.

3.6 Matching Google mobility data to incident data

The incident data were matched to the Google mobility data using the FIPS code corresponding to the county of each fire department. FIPS stands for Federal Information Processing Standard Code. It is a code used to define geographic areas in the United States. The code comprises five digits: the first two digits specify the state, and the following three specify the county. Figure 4 shows the two datasets (EMS not including COVID-19-related calls and fire incidents) obtained for one fire department. The black dashed and dotted lines are the percentage changes in fire and EMS incidents not including COVID-19-related calls. The colored lines are the percentage changes for the various Google mobility categories. For most mobility categories – grocery/pharmacy, retail, workplace – the pandemic caused people to spend less time in those locations (i.e. the percentage change in Figure 4 is negative). However, the pandemic caused people to spend more time at home (residential category) and parks. Similar mobility trends were observed for all fire departments. For this department, fire incidents dropped by at most 50% during the pandemic and EMS incidents by 95%. This department was among those having experienced a very large drop in incidents – as shown in Figure 2, the drop in EMS calls was less pronounced at the aggregated level.

3.7 Choosing a mobility category

The last step was to choose the mobility categories to include in the final model. As a reminder, the final model explores the impact of mobility on incident counts. It is evident from Figure 4 that the mobility types are correlated. Further, Figure 5 shows a correlation matrix for the various mobility categories for the counties relevant to this work. Most correlation coefficients are high in magnitude (close to 1 or -1), indicating that the mobility categories are highly correlated. For example, when residential mobility increases (people stay at home), workplace mobility also tends to decrease (people do not go to work because they stay home). This poses issues when running a regression because adding more mobility categories to the model does not add much new information since mobility categories are related. Principal component analysis (PCA) could have been used to generate covariates; however, for simplicity and interpretability, a single category, the residential mobility category, was used in the final model. Despite providing less information, using only residential mobility makes the goal of the model easy to interpret as “understanding the impact of people staying at home on EMS and fire incident volume”.

4. Model

To examine the relationship between residential mobility and the change in the number of EMS and fire incidents, a (Bayesian) Gaussian process hierarchical regression was used. This model is very similar in concept to the hierarchical regression run to calculate incident baselines in section 3.4. In this analysis, the goal was to run a regression model to establish the relationship between the percentage change in mobility and the percentage change in EMS and fire incidents. Again, the advantage of using a hierarchical model is that each fire department has its own regression coefficients, but they are informed by the whole dataset. A key difference with the model of section 3.4. is that a linear regression model was used rather than a negative binomial model. In a linear regression model, the dependent variable (here, the percentage change in incidents) is assumed to be a linear function of the independent variables (here, the percentage change in mobility). In addition, in a linear regression, the residuals (the difference between the predicted and true values) conditioned on the parameters of the model are assumed to be normally distributed. Linear regression models are appropriate for continuous data that meet this assumption. In a linear regression, the residuals are often assumed to be independent. However, in some scenarios, this assumption is too restrictive. For example, in the context of this model, if the model overpredicts incidents one week, it may be more likely to overpredict the following week as well. A Gaussian process was used to allow for the possibility that residuals may be temporally correlated. In this framework, the temporal correlations are defined using a covariance function (which is described below). The function takes in two different times and outputs a value for the correlation, which is assumed to be stronger for times that are close together and weaker for times that are far apart. However, the parameters that quantify this relationship are learned from the data. Spatial correlations between the fire departments were not taken into account. They were assumed to be negligible, given that the fire departments in the dataset are not in close proximity. The model was run separately on EMS, EMS not including COVID-19-related calls, and fire incidents. Again, this model was Bayesian, and an MCMC method was employed to sample from the posteriors of the parameters of the model. The same Python package, PyMC, and the NUTS algorithm were used. Formally, the model was specified as shown below:

(5)(yj|βj,Kj)∼N(Xjβj|Kj)

(6)(βj|μ,Σ)∼N(μ,Σ)

(7)bj∼TruncNorm(μb,σb2)

where j is the fire department, and i is the coefficient number (there are two coefficients, one for the intercept and one for residential mobility). yj is the percentage change in EMS or fire incidents for department j; it is an nj-length vector where nj is the number of observations for department j. Xj is the nj-by-two design matrix for department j, consisting of one column for the intercept and one column for the percentage change in residential mobility. μ is a length-two vector for the two coefficients, and Σ is a two-by-two diagonal matrix. This model is “partially pooled” because the β coefficients of all departments are assumed to come from the same normal distribution, as shown in equation (6). Kj is an nj-by-nj department-specific covariance matrix whose elements are the evaluation of a squared exponential covariance function:

Kj(t,t′)=τj2⁡exp[(t−t′)22bj2]

Here, t denotes the day of the year (given that a weekly average was used, the day of the year can only correspond to a Monday). bj is often named the bandwidth or length scale and controls how smooth the covariance function is. τj2 the signal variance. A noise term of 1E−6 was also added to the diagonal of the covariance matrices. Weakly informative priors were used for the parameters; they are shown in the supplementary material.

5. Results and discussion

5.1 Relationship between residential mobility and fire incidents

This section examines the relationship between the increase in residential mobility due to the pandemic and the change in fire incidents using the Gaussian process hierarchical framework. The outputs of interest from the model described in section 4 were the posterior distributions for the regression parameters. Again, the posteriors are distributions resulting from the MCMC sampling process, which represent what is known about the regression parameters after observing data. Posteriors describe uncertainty about the parameters in that they are distributions rather than single-point estimates. To summarize the characteristics of a posterior, one may use the posterior mean (the mean of all posterior samples drawn using the MCMC algorithm), and a 95% credible interval (CI). A 95% credible interval can be calculated by finding the 2.5th and 97.5th percentiles of the posterior samples. A regression coefficient may be said to be significantly positive or negative if its 95% credible interval does not contain 0.

In the regression run, several posteriors were of particular interest. First, the “high-level” posteriors were of interest, meaning those common to all fire departments. These were the posteriors for μ, and especially the second element of μ, μ1, which represents the overall (across all fire departments) effect of residential mobility on the percentage change in fire incidents. The elements of Σ show the variance around the mean μ. The higher the variance, the more heterogeneity there is across fire departments. Second, the fire-department-specific posteriors were also of interest. These were specifically those for β1 (the second element of β), which represent the effect of residential mobility on the percentage change in fire incidents for each department. There is a posterior distribution for β1 for each fire department, while there is a single distribution for μ1.

Figure 6 shows the posterior distributions for μ and Σ as defined in equation (6) for fire incidents. The posterior mean for μ1 was −0.46 and the 95% credible interval was [−0.69, −0.22]. The 95% CI for the posterior of μ1 does not contain 0. It only covers negative values, which means that, on average, increasing residential mobility is associated with a significant drop in fire incidents. In summary, these findings mean that overall, across fire departments, a 1% increase in residential mobility was associated with a 0.46% drop in fire call volume.

Of 56 departments, only 12 departments (21% of departments) had significantly negative β1 coefficients, meaning that the 95% CI for β1 only included negative values. This means that fire call volume dropped significantly with increased time spent at home for 21% of the fire departments analyzed. However, the posterior mean was negative for 50 out of 56 departments. This suggests that overall, when residential mobility increases, fire incident counts also tend to drop for these departments, though not enough to meet the chosen significance criteria. No posteriors for β1 were significantly positive, meaning that no departments saw a significant increase in fire call volume with increased time spent at home.

The difference in the number of incidents between 2019 and 2020 was investigated for various incident categories to qualitatively explain the drop in the overall number of fire incidents. To the authors’ knowledge, no other studies have examined the frequency of detailed incident types over the pandemic other than Suzuki and Manzello (2022a), who explored the effect of the pandemic on cooking fires. The two variables in the NFORS dataset that can help determine incident category beyond EMS and fire types are the “description type” and “comments” variables. The description type variable is a more detailed version of the category variable, which can include keywords such as “brush fire” or “structure fire”, for example. However, departments use different category labels for this variable, so it is difficult to analyze these data across all fire departments. Furthermore, the categories are not necessarily very granular. For example, few departments have a cooking fire category despite cooking fires being the most frequent type of fire. Instead, most departments classify a cooking fire as a fire alarm incident or structure fire. Comments, on the other hand, contain more information about incidents. In the literature, fire report comments have successfully been parsed by Mirończuk (2020) to extract useful incident data. A comment for a cooking fire incident would likely contain keywords such as “cook” or “stove”. Using the comments was preferred because of the level of granularity that they can provide. An incident was assumed to belong to a specific category if any keyword associated with the category was found in the comments. The keywords associated with each chosen category are shown in Table 1.

In the analysis, fire departments were considered separately. For each department, the number of incidents for each category was calculated for 2019 and 2020. The numbers in 2019 and 2020 were divided by the number of days that had data in 2019 and 2020, respectively. This ensured that the counts did not erroneously seem lower due to missing days in the data. Then, for each category, the percentage of fire departments for which the 2020 counts were higher than the 2019 counts was calculated. Note that for each category, fire departments with fewer than 30 counts across 2019 and 2020 combined were removed from the analysis. These departments typically had insufficient details in their comments.

As shown in Figure 7, over 50% of departments had more vehicle accidents, structure fires, brush fires, and electrical/appliance fires in 2019 than in 2020. 50% of departments or more had more cooking fires and false alarms in 2020 than in 2019. The drop in vehicle accidents/fires, structure fires, brush fires, and electrical/appliance fires for over 50% of departments in 2020 may help explain why fire incidents, on average, dropped with increasing residential mobility.

5.2 Relationship between residential mobility and EMS incidents

The analysis above was repeated for incidents classified as “EMS” that were not due to COVID-19 cases and for all “EMS” incidents, as explained in section 3.3. Figure 8 shows the posteriors for μ and Σ for both cases. For μ1, both posteriors are significantly negative (their 95% CI does not contain 0). The posterior for all EMS calls is less negative than when COVID-19 EMS incidents are removed. This is not surprising because the number of calls in any given week is lower if COVID-19 incidents are removed. These findings mean that on average, across all departments, when residential mobility increases (i.e. time spent at home increases), both the total number of EMS incidents and the total number of EMS incidents minus the number of EMS incidents directly due to COVID-19 tend to drop. Both posteriors for μ1 are smaller than for fire incidents (see Figure 6), which indicates that EMS incidents dropped more significantly with residential mobility than fire incidents. For EMS incidents not due to COVID-19 cases, the posterior mean for μ1 was −1.43, and the 95% CI was [−1.68 −1.17]. This means that overall, a 1% increase in residential mobility was associated with a 1.43% drop in EMS call volume (not including COVID-19 calls).

Out of the 53 fire departments with EMS data, 47 (89% of departments) had significantly negative posteriors for β1, meaning that their EMS incidents (excluding COVID-19 EMS incidents) significantly dropped with increasing residential mobility. All 53 fire departments had negative posterior means.

The number of incidents for a few categories of EMS calls (not including COVID-19-related calls) for 2019 and 2020 was compared for each department as previously done with fire incidents. The results are shown in Figure 9. As explained for fire incidents in section 5.1, the number of incidents for each category was estimated using incident comments. The keywords used for each category are shown in Table 2. For most departments, the number of vehicle accident, cardiac, breathing, and trauma incident calls dropped, which likely explains the fact that overall EMS calls decreased with increasing residential mobility. Only the number of overdose/alcohol poisoning-related incidents increased for most departments. This result is consistent with the findings of Friedman et al. (2021).

5.3 Discussion

The COVID-19 pandemic presents an opportunity to draw lessons on public health strategies. For example, Hertelendy et al. (2023) discuss why we are “not ready” for the next pandemic and describe actions that can make us better prepared in the future. Hertelendy et al.’s research highlights the importance of developing resilient public health systems, tailoring emergency measures to communities, and harnessing the untapped power of communities. Their research highlights the need to use new technologies and data sources to achieve these goals and to use the available resources to their maximum potential. The present work showed that mobility data can serve as a tool that can be used for predicting incident call volumes, either ahead of time if expected mobility trends are known or almost instantaneously using real-time mobility data. This tool could be used to increase the preparedness and effectiveness of fire departments and emergency services in managing future public health crises that may disrupt mobility patterns. For the NFORS departments part of this study, overall, the increase in the amount of time that people spent at home because of the pandemic was associated with a drop in both fire and EMS call volume. In this instance, where call volumes are predicted to be lower than usual, fire departments may use this information to allocate staff to tasks more effectively. Beyond first responder duties, firefighters often engage in public education programs. These public education programs may be used to educate the public about the ongoing health crisis. For example, for the COVID-19 pandemic, these programs could include educating the public about testing and vaccination campaigns, available health services, mental health resources, and community engagement opportunities. Additional time may also be dedicated to training firefighters to enhance their ability to handle an ongoing public health crisis. Although incident volume dropped during the pandemic for these NFORS departments, NFORS departments are not necessarily a representative sample of all fire departments in the United States. Other places, such as New York City, experienced a surge in EMS calls due to the pandemic (Prezant et al., 2020). For such departments, increased time spent at home is likely positively correlated with call volume rather than negatively correlated. In this case, mobility data could be used to predict surges in call volume and secure additional resources as needed.

6. Conclusion

The purpose of this study was to determine the relationship between residential mobility and fire and EMS call volume to help inform resource allocation decisions for NFORS users. This was accomplished using a Gaussian process hierarchical model that accounted for temporal correlation in the residuals.

The posteriors for the overall (across fire departments) relationship between residential mobility and incident call volume showed a significant negative relationship between time spent at home and both fire and EMS call volumes. Overall, a 1% increase in residential mobility caused a 1.43% and 0.46% drop in EMS (not considering COVID-19-related calls) and fire call volume, respectively. Department-specific posteriors showed that 89% and 21% of departments had a significant negative relationship between residential mobility and EMS (non-COVID related) and fire call volumes, respectively. For fire calls, those results appeared to be driven by the decrease in vehicle fire/accidents, electrical/appliance, brush, and structure fire calls. For EMS calls, the results appeared to be driven by the decrease in vehicle accidents, cardiac, breathing, and trauma calls. These results could help fire departments more effectively allocate resources in case of a future pandemic or public health crisis that has the potential to significantly disrupt human mobility. Analyzing mobility data can provide a convenient method to quickly predict call volumes. Based on these predictions, fire departments may decide to adjust their resource allocation strategies. For instance, if call volumes are low, fire departments may decide to allocate extra firefighter time towards public education initiatives about the ongoing health crisis and provide additional training to firefighters to enhance their ability to handle such crises.

The are some limitations associated with the methods and data used in this paper. From a data perspective, while 23 states are represented in the NFORS dataset, the list of NFORS fire departments is not a random sample of all departments in the United States. Therefore, this analysis may not be representative of all fire departments in the United States. Furthermore, the text comment analysis approach used to estimate the number of incidents belonging to various categories was simple and may have led to over or underestimated numbers. Future work exploring incident comments may benefit from using more advanced natural language processing (NLP) algorithms. It should also be noted that the negative relationship between residential mobility and fire incidents found for some departments may partially be due to the fact that some departments categorize vehicle accidents (which dropped during the pandemic) as fire incidents rather than EMS incidents. Enforcing more uniform data reporting standards across departments would make analyzing the data of multiple fire departments easier. Lastly, this analysis did not consider the impact of the COVID-19 pandemic on the spatial distribution of incidents. Considering this aspect could help fire departments better allocate resources from a geographic standpoint should another pandemic or public health crisis occur.

Figures

Figure 1

Bar plot showing the average daily number of incidents for the 56 departments used in the analysis ranked from largest to smallest

Figure 2

Plot showing the daily total number of EMS incidents for all departments combined and the daily number of COVID-19-related EMS incidents, as well as the difference between the two

Figure 3

Plot illustrating the procedure for expressing EMS incident counts as a change relative to the baseline for one fire department

Figure 4

Plot showing the data derived from the data processing steps for one department for fire incidents and for EMS incidents (not including COVID-19-related calls)

Figure 5

Correlation matrix showing the correlation coefficient between mobility categories

Figure 6

Plot showing the posterior distributions for μ and Σ as defined in equation (6) for fire incidents

Figure 7

Bar plot showing the percentage of fire departments having higher counts in 2020 compared to 2019 for each fire category

Figure 8

Plot showing the posterior distributions for μ and Σ as defined in equation (6) for EMS incidents for the cases where all EMS incidents were considered and where only EMS incidents not related to COVID-19 cases were considered

Figure 9

Bar plot showing the percentage of fire departments having higher counts in 2020 compared to 2019 for each EMS category

Table 1

Keywords used to find incidents corresponding to each fire category

Fire type	Keywords
cooking fires	stove, cook, kitchen, food, oven, microwave
vehicle accidents/fires	veh, scooter, car crash, vehicle crash, car accident, motorcycle accident, traffic, motor, SUV, sedan, truck
structure fires	building fire, structure fire, structure, house fire, apartment fire, residential fire, home fire, high-rise fire
false alarms	no smoke, false alarm, no fire, no flame, disregard
brush fires	vegetation, brush, forest, fireworks, tree, outdoor fire, park
electrical/appliance fires	transformer, washer, laundry, dryer, electrical, wire, appliance, radiator, lamp, light, wiring, heater

Source(s): Authors’ own work

Table 2

Keywords used to find incidents corresponding to each EMS category

EMS incident type	Keywords
vehicle accidents	veh, scooter, car crash, vehicle crash, car accident, motorcycle accident, traffic, motor, SUV, sedan, truck
cardiac	cardi, heart, chest pain
breathing	breathe, respi
trauma	trauma, injur
overdose/alcohol	overdose, intox, alcohol

Source(s): Authors’ own work

Supplementary Material

The supplementary material for this article can be found online.