SARS-CoV-2 outbreaks on Danish mink farms and mitigating public health interventions

Abstract Background First severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infections on Danish mink farms were reported in June 2020 and thereupon spread geographically. We provide population-level evidence on excess human incidence rates in Danish municipalities affected by disease outbreaks on mink farms and evaluate the effectiveness of two non-pharmaceutical interventions, i.e. culling of infected mink and local lockdowns. Methods We use information on SARS-CoV-2 outbreaks on mink farms in 94 Danish municipalities together with data on human SARS-CoV-2 cases and tested persons in Weeks 24–51 of 2020. Difference-in-difference estimation and panel event studies for weekly human incidence rates are applied to (i) identify epidemiological trends of human SARS-CoV-2 infections associated with disease outbreaks on mink farms, and (ii) quantify the mitigating effects from the two non-pharmaceutical interventions. Results SARS-CoV-2 outbreaks on mink farms in a municipality associate with an increase in weekly human incidence rates by about 75%; spatial spillover effects to neighbouring municipalities are also observed. Local lockdowns reduce human incidence rates, while culling of mink appears to be more effective in combination with a lockdown. The temporal lag between an outbreak on a mink farm and a significant increase in human incidence rates is estimated to be 1–3 weeks; lockdowns and culling of mink neutralize this effect 4–8 weeks after the initial outbreak. Conclusions SARS-CoV-2 infections among farmed mink in Denmark significantly link to local human infection trends. Strict animal and human disease surveillance in regions with mink farming should be pursued internationally to mitigate future epidemic developments.


Difference-in-Difference estimation (DiD
where , denotes the human SARS-CoV-2 incidence rates for municipality i in calendar week t and , is the treatment indicator for SARS-CoV-2 infection outbreaks on Danish mink farms. In the baseline specification, this variable is an absorbing treatment indicator, which takes values of one for municipality i from week t onwards for which the first infection on a mink farm in this municipality is reported; it is zero before that date and for municipalities not affected by a disease outbreak in week t. The parameter tests for the link from SARS-CoV-2 infections on mink farms to SARS-CoV-2 infections in the human population of municipality i; the parameter similarly tests for the link between SARS-CoV-2 infections on mink farms and SARS-CoV-2 infections in the human population of neighbouring municipalities to municipality i. The underlying spatial treatment indicator , takes values of one, if a municipality j  i is located in the spatial neighbourhood of municipality i experiencing a disease outbreak on a mink farm. Two municipalities i and j are classified as geographical neighbours if the centroids of these municipalities lie within a 50km radius. From , we exclude all municipalities that are included in , to avoid a double counting of treatment effects. Two factors of mitigating measures (non-pharmaceutical interventions) in public health policy are included.
First, the lockdown in seven severely affected municipalities in Northern Jutland is measured by a binary dummy variable. The variable , accounts for the potential mitigating effect of the local lockdown in seven municipalities in Northern Jutland from calendar week 45 (i.e., it takes values of one for seven municipalities during the weeks 45 to 49 and is zero otherwise, see Panel A in Figure 1 in the main manuscript for a visualization of locked down municipalities). The local lockdown was motivated by the ambition to curb human SARS-CoV-2 infections in municipalities particularly affected by SARS-CoV-2 infections in farmed mink, why the parameter 1 tests for the effectiveness of this mitigating measure. The variable , covers geographical neighbours of lockdown municipalities (again, based on a 50 km radius) and the coefficient accordingly tests for the presence of spatial spillovers for this mitigating measure.
Second, a political decision to cull mink on infected mink farms and at later stages all mink farms was decided by the Danish Government as a further mitigating measure next to the local lockdowns. The variable , is a binary dummy that indicates whether a farm with infected mink has been subject to culling in a given municipality i and week t. To account for transmission lags from culling of mink to human incidence rates, we accumulate the number of culled mink on infected mink farms over a period of three weeks for the definition of , . To capture potential spatial spillover effects from culling, , measures whether a municipality is a neighbor to a municipality subject to culling ( , 1 or not. The parameters and therefore tests for the effectiveness of a public health policy stressing culling of mink to curb human SARS-CoV-2 infections, directly or through spatial spillovers. and periphery) based on regional demographics like population density and social structure. Finally, in equation (A.1), denote week-fixed effects common to all municipalities (to cover cyclical trends in human SARS-CoV-2 infections in Denmark), and are municipality-fixed effects controlling for unobservable regional heterogeneity that may confound outcomes beyond the level of region types (e.g. to cover differences in infection levels in regions with external borders); , is the error term.

Alternative treatment indicators.
We focus on estimating the coefficient which captures the effect of mink farm infections on subsequent human SARS-CoV-2 incidence rates. This treatment indicator , is constructed in three different ways to assess the robustness of the obtained results. By default, it is a binary absorbing indicator specified as , , i.e. the dummy takes values of one from calendar week t onwards when the first SARS-CoV-2 mink farm infection was reported in municipality i and stays at this value for treated municipalities during the remainder sample period. This is our default specification.
One may conjure that the two-way infection dynamics between animals and humans associated with a disease outbreak on a mink farm phases out over time and that treatment effects are thus only transitory in nature. As an alternative specification we therefore define the binary treatment indicator , in a non-absorbing way, i.e. that it reverts back to zero for the affected municipality n weeks after the first SARS-

CoV-2 mink farm infection . A new infection on another farm in municipality i during this
n-week period extend the treatment period subsequently. We perform sensitivity analyses by considering alternatives as n=3 or n=4 weeks, which appear reasonable if treatment effects are assumed to be static with an incubation time of approx. 1-2 weeks and an equally lengthened infection duration. We also check the sensitivity of the empirical results with regard to alternative data sources: While our default absorbing treatment indicator is defined on the basis on published data obtained from the home page of the Danish Veterinary and Food Administration, we also use updated process data from the same administration to define treatment indicators on a rolling basis. Process data has the advantage that it accounts for near-time adjustments in the reporting of SARS-CoV-2 infections on mink farms and culling of mink, while it may come at the cost of being preliminary. Taken together, we believe that utilizing all available data from the Danish Veterinary and Food Administration offers the best way to arrive at robust results.
Finally, rather than using a binary flag indicator, a third continuous specification of , counts the stock of SARS-CoV-2 infected farms in municipality i by accumulating newly reported mink farm infections in each municipality until week t. In analogy to the binary treatment indicator, we either specify a nonreverting stock by summing over all infected farms and weeks. Alternative, we accumulate infections over n=3 or n=4 weeks. Different ways to calculate cumulative stocks are presented in Figure A2.

Panel event study (PES).
A panel event study is applied as a complementary fully flexible estimation approach to allow treatment effects to vary by week rather than treatment having an absorbing or nonabsorbing static structure with underlying assumptions as in the DiD approach. [3][4][5][6] By considering timeheterogeneity in the estimated coefficient of the treatment indicator The index , … , denotes the maximum number of periods before first treatment (-N) and after first treatment (M) in municipality i. The inclusion of several periods before first treatment is observed allows us to test for early anticipation effects in the outcome variable prior to first treatment. If such effects are significant and positive, a rise in the incidence rate in municipality i is likely driven by other latent factors rather than our treatment in focus. In the absence of early anticipation effects, however, significant effects arriving with the treatment start in the included M periods after the first treatment, can be taken as evidence for an effect of disease outbreaks on mink farms on the epidemiological trend in affected municipalities.
Plotting the weekly treatment effects ( allows us to time the phasing-in of this transmission channel and identify potential intervening effects associated with the local lockdown and culling of mink. Estimated effects beyond this time interval are accumulated in a single coefficient shown in the first pre-and last posttreatment period, which is typically referred to as binning. 6 The treatment indicator , for the last pretreatment observation ( 1) is omitted to capture the baseline difference between treated and non-treated municipalities. Moreover, as an extension to the above specification we also include neighbouring regions in the treatment group to identify not only direct but also indirect treatment effects.     Notes: The permanent cumulative stock of infected farms adds the number of infected farms by week to the total cumulative sum for each municipality. The cumulative stock calculated on a rolling basis shown in Panel B accumulates the number of infected farms over the last four weeks building on the assumption of temporally diminishing effects from a given SARS-CoV-2 outbreak on a mink farm in the light of public health interventions. Alternative specifications use a three week period; results can be obtained upon request from the authors or generated through the replication files linked to this paper.

B. Descriptive statistics and additional estimation results
Source: Fødevarestyrelsen, Smittede mink farme uge for uge (in Danish), Smittede minkfarme uge for uge (foedevarestyrelsen.dk), Retrieved April 8 2021.  Notes: 95% confidence intervals in brackets. * = variables defined on a three week rolling basis; # = variables defined on a four weeks rolling basis; $ = variables defined on a permanent basis; neighbour values for the variable "Infected farms" are measured in the same dimension as the underlying variable in each column. Notes: CI = Confidence Interval; dotted vertical lines indicate the last pre-treatment observation (baseline); dynamic treatment effects obtained from panel event study (PES) controlling for the stock of infectious individuals in municipality and spatial neighbourhood; number of PCR tested persons, (lagged) temperature, workplace mobility, region-type specific trends and region-and time-fixed effects. The last pre-treatment observation is omitted to capture the baseline difference between treated and nontreated municipalities. Treated municipalities are restricted to those belonging to the top-5 percentile of SARS-CoV-2 outbreaks on mink farms per municipality.