We appreciate the commentary by Wiebe1 that carefully considered our study2 and suggested ideas for more exploration and new directions. We aim to extend the conversation on a few topics—specifically chronic and acute aspects of violence, within-versus cross-community comparisons, interpretation of violence residuals, and asymmetrical relations for increases and declines in acute violence.
We agree with Wiebe that, by focusing on within-place changes in community violence, we miss the possibility that community violence in nearby places may also impact residents (eAppendix 1; https://links.lww.com/EDE/B372 maps the included census places, and there are certainly clusters of places around the metropolitan areas). Wiebe’s comment highlights the crux of the challenge, as we see it, in quantifying the health effects of community violence. The strong correlation of community violence with economic, social, and physical features of communities creates problems of structural confounding, which led us to limit our analysis to within-place comparisons to control by design for time-invariant community-level confounding variables. Strong cyclical patterns in violence and health outcomes within places over time, for example due to seasonality, also need to be addressed to avoid associations that are an artifact of shared temporal autocorrelation. After processing out predictable temporal patterning, we are left examining deviations from what would be normative or expected within a place given the secular, seasonal, and other patterns in time. Ultimately, this approach provides greater control for important sources of confounding than between-place comparisons.
However, our intuition tells us that to understand the full impact of violence; we must also try to understand how different overall rates of violence affect health. This is the day-in and day-out burden of violence that constitutes a chronic stressor in our communities. Wiebe clearly shares this perspective, noting that the overall burden of violence would be expected to have a larger health impact than violence spikes. In our work in California, the overall level of violence tends to be relatively constant over shorter time periods within place. Thus, within-place comparisons either focus on relatively small overall violence differences, or they extend across enough time that overall violence rates have changed more substantially. However, as within-place comparisons extend across time, they face confounding problems similar to between-place comparisons; many other characteristics interrelated with violence may have changed within a place, structurally confounding the comparison.
We think the methodologic challenges of estimating the health effects of community violence need to be taken seriously, and we have done some work on how to improve research on this topic given the inherent challenges. For example, when examining overall levels of violence, we have been careful to estimate parameters that compare values across the range actually observed within a given community to avoid extrapolation.3 When using individual-level data with a rich set of individual confounders, we think that leveraging matching approaches is worthwhile for within- or cross-community comparisons. These approaches can avoid extrapolation and are well suited to situations of strong confounding.4,5 Although these and other quasi-experimental approaches would not always exclude residual confounding as an alternative explanation, they have potential to reduce bias in estimates of the health effects of violence. If we accept this as worthwhile despite imperfection, we agree with Wiebe that identification of modifiers of the effects of violence, including underlying levels of disease, will be important to fully understand the relationships and potentially to target interventions.
Importantly, we see value in rigorously quantifying the effects of programs that aim to prevent or reduce community violence. When such programs are effective, we should then determine whether there were corresponding health improvements. This work will bring us far closer to identification of effective tools to reduce community violence and improve community health. If opportunities arise that allow for randomized designs, such as stepped wedge implementation of the same program in different settings, this will improve the strength of the evidence. However, we think it is still important to do rigorous evaluation of violence prevention programs that cities initiate independently given that this is how the vast majority of violence prevention activities occur.
Returning to some specifics in our article, Wiebe raises the question of whether deviations from expected levels of violence are capturing changes in violence. To address this question, it is helpful to think about the problems that could arise if we were to simply examine changes in violence related to changes in health outcomes from consecutive time points. For example, in the case of seasonal patterning in the exposure and the outcome, there could be a predictable increase in exposure at the beginning of summer months that is mirrored in the outcome but is owing to effects of weather, temperature, or changing activities of the population rather than changes in the exposure itself. The varied kinds of temporal autocorrelation, including trends, cycles, and regression to the mean, are the very reason that time-series procedures were developed to remove predictable patterns. From this perspective, we can view residuals as changes in the sense that they capture the deviation of the observed value from the value that would have been predictable, but not literally as changes between adjacent time points.
Finally, we agree that we might anticipate spikes in violence to be harmful for health, while times of unusually low violence may not be correspondingly salutary. This was the motivation for focusing in the paper on the spikes. We also examined continuous residuals to ensure that we had not diminished our ability to detect associations by removing most of the variability in the exposure when limiting the exposure definition to binary spikes.
In conclusion, we appreciate this opportunity to extend the dialogue on how to improve and extend work that examines the health effects of community violence. These same methodologic challenges apply to other community determinants of health, and continued development of novel ways to handle these challenges has potential to improve this area of work.
1. Wiebe D. Preventing community violence to prevent communities’ disease events: an efficient approach with short- and long-term potential. Epidemiology. 2018:29:692694.
2. Ahern J, Matthay EC, Goin DE, Farkas K, Rudolph KE. Acute changes in community violence and increases in hospital visits and deaths from stress-responsive diseases. Epidemiology. 2018;29:684691.
3. Matthay EC, Farkas K, Skeem J, Ahern J. Exposure to community violence and self-harm in California: a multi-level, population-based, case–control study. Epidemiology. 2018:29:697706.
4. Colson KE, Rudolph KE, Zimmerman SC, et al. Optimizing matching and analysis combinations for estimating causal effects. Sci Rep. 2016;6:23222.
5. Oakes JM, Johnson PJ. Oakes JM, Kaufman JS. Propensity score matching for social epidemiology. In: Methods in Social Epidemiology, pp. 2006.San Francisco, CA: Josey-Bass, 370392.