Effects of non-pharmaceutical interventions on the death rates from COVID-19 in Canada

Alexander Meek

doi:10.1016/j.ssaho.2023.100449

. 2023 Feb 20;7(1):100449. doi: 10.1016/j.ssaho.2023.100449

Effects of non-pharmaceutical interventions on the death rates from COVID-19 in Canada

Alexander Meek ¹

PMCID: PMC9939403 PMID: 36846438

Abstract

This paper examines the effectiveness of provincially enacted policies to deter the COVID-19 mortality rate during the pre-vaccine rollout pandemic in Canada. The data has been collected from Statistics Canada and other various online resources including the Blavatnik School of Government and provincial statements. Relevant information was collected for individual provinces from 11 March 2020 to 31 Jan 2021. A two-stage least squares methodology was applied by examining, by province, the cumulative number of deaths reported as due to COVID-19 before and after policy implementation. We examine the effect of each of several policies based on a 20 plus day lagged period after being put into place. Our main finding is that workplace closures and strict gathering restrictions are associated with a decrease in mortality rates of COVID-19 in Canada. We also find that the strength of the policies overall is associated with a decrease in mortality rates of COVID-19 in Canada. Using data from the Google Mobility Report we confirm that the announcements of the policies had significant behavioural effects on individuals’ movements. We interpret that social distance policies, specifically the workplace closure and strict gathering policies, have been effective factors decreasing mortality from the Coronavirus epidemic in Canada.

Keywords: Policy COVID-19 epidemic mitigation provincial

1. Introduction

As of January 1, 2022 there have been 2,234,427 confirmed cases of the novel Coronavirus or COVID-19 in Canada and 30,339 deaths reported by the Canadian Government. With additional cases and deaths being reported daily, and the numbers consistently growing, it is important to understand the effects of the policies that different provinces have implemented to mitigate the spread of the virus. COVID-19 has affected many countries and forced governments to enact aggressive policies to combat this pandemic.

The World Health Organization (WHO) declared a “pandemic” on March 11, 2020 (Canadian Public Health Association, 2020). Shortly thereafter, the Canadian Federal government urged Canadians abroad to return home on March 14, 2020; international travel restrictions were implemented March 16, 2020 (The Canadian Press Staff, 2021). An integral part of Canada's fight against COVID-19 has been shouldered by provincial governments initiating policies promoting social distancing (non-pharmaceutical social distance inducing interventions) based on the severity in their provinces. This makes a provincial policy comparison compelling to examine quantitatively. Quebec, Manitoba, and Ontario have been impacted the most by the virus as of the end of our dataset, January 31, 2021. Other provinces have not seen as adverse impacts, specifically New Brunswick, PEI and Newfoundland and Labrador. PEI has yet to confirm a death related to the novel Coronavirus as of January 31, 2021. Data from PEI has been removed from all regressions. Canada had a total of 20,030 deaths as of this date. The graphical representation of the different provinces deaths per 10,000 is shown in Fig. 1 just below.

Fig. 1 — Deaths per 10,000 across provinces over time.

It is key that we understand which policies have been effective. Using the Oxford stringency index, we find that the overall policy strength by provinces was effective in mitigating mortality rates. We then break down the various social distance measures within the index to determine which interventions are more effective than others. We discern which of the various measures that Canadian Provinces adopted were especially effective in mitigating the effects of the pandemic. We then confirm our findings through a behavioural analysis using the Google Mobility Reports showing that people altered their behaviours both immediately and long-term following a policy implementation. The measures enacted by the provinces did not occur simultaneously and there were significant differences in the rate of spread within them. This further allows us to estimate and examine effects of the policies on the virus.

Figs. 2–10 in the appendix show for each province the Oxford Stringency Index over the course of our data against the cumulative deaths per 10,000 population. Below are Fig. 2, Fig. 3 and 8 that represent Alberta, B.C., and Ontario, respectively.

Fig. 2 — Alberta Fig. 3: British Columbia.

These figures indicate trends; we can see declines in the index followed by large increases in deaths and we can also see that following increases in the index death rates slow significantly. This leads us to ask which restrictions being eased creates this increase in the mortality rate.

A similar study conducted for US states by Dave et al. (2020) estimate the effects of stay at home or shelter in place orders (SIPO) on the cases and deaths of COVID-19 in America. They found SIPOs resulted in a decrease in cases but no significance evidence for decreasing the number of deaths. See also Chernozhukov and Paul (2021) for a similar study where they evaluated the impact on the cases and deaths from COVID-19 using the Google Mobility Reports and a variety of state level policies designed to limit spread. Karatayev et al. (2020) uses a theoretical simulation to demonstrate that approaching the pandemic on a county-to-county bases would limit the number of closure days and the effect on people when compared to the provincial approach. A theoretical modelling study by Ogden et al. (2020) predicts that without non-pharmaceutical interventions most Canadians would have contracted COVID-19 in the first wave and overwhelmed the hospital system. Anderson et al. (2020) looks at British Columbia, Canada (in addition to a few other non-Canadian jurisdictions) using survey data. They find that physical distancing reduced contacts between people to 22% of pre-pandemic levels/interactions. This paper contributes by using policy data from the Oxford Stringency Index and statistical behavioural changes from the Google Mobility Report and finding the most effective policies to limit the spread during the pandemic. Perra (2021) conducted a literature review of 348 articles on non-pharmaceutical interventions on COVID-19. Perra (2021) finds that non-pharmaceutical interventions were important and effective at slowing the spread of COVID-19, and that implementing these policies early was the most effective at limiting spread. This paper confirms this finding through a Canadian context with additional data and contributes by specifying impactful critical policies.

We use an empirical strategy based on difference-in-difference regression techniques and instrumental variables to estimate the relationship between these policies and deaths across provinces in Canada. Measures to mitigate the spread of the virus may be brought in at a time when an increase in its growth is anticipated. Thus, one can expect causality running in both directions. This could lead to a positive correlation between the instigation of the measure and the number of confirmed cases and possibly even future deaths from those cases for at least some short to medium term period in the future. This does not mean the increased number of cases was caused by the measure but rather that the measure acted like an announcement of an inevitable and immediate rise in the number of infections. This rise may well have been mitigated by the measure despite the apparent positive relationship. Moreover, the size of any negative coefficient, although indicative that the measure is effective in mitigating the spread of the virus, may be due to heightened awareness of citizens created by the instigation of the measure or other information that enlightened the introduction of the measure and so the total effect of the introduction of a policy may be due to other reasons for changes in citizens’ behaviours.

2. Data and methodology

2.1. Data

The data for this paper is collected from a variety of sources. The COVID-19 statistics (i.e., daily and cumulative confirmed cases of infections and number of deaths determined to be due to COVID-19) come from Statistics Canada reporting tables. The dates that policies are introduced are collected by The Blavatnik School of Government and the University of Oxford. “The Oxford COVID-19 Government Response Tracker (OxCGRT) systematically collects information on several different common policy responses that governments have taken to respond to the pandemic on 17 indicators such as school closures and travel restrictions” (Petherick et al., 2020). This paper's data takes into consideration the measures used in collecting the OxCGRT and includes school closures, workplace closures, gathering restrictions, internal travel restrictions, and the overall stringency index level of each province. The mobility trend data comes from the Google Mobility data. Hale et al., 2021 provides an in-depth analysis of the Oxford Index.

Our data begins March 11, 2020 and goes until Jan 31, 2021. We use this specific data period as provinces were easing restrictions and the first wave was ending. January 31, 2021 was chosen as the end date of our regressions because we want to capture the effect of Government policy on the spread of COVID-19 before the effect of the vaccination roll-out. The vaccine's ability to limit the spread and severity impacts the relationship between policies and number of deaths. Our model would need to be reconsidered. Therefore, this date acts as a natural cut-off point such that we can more directly examine social distance policy measures.

We use a seven-day rolling average of the total number of recovered cases to create the number of active cases. This will help to eliminate sudden jumps in recovered cases due to reporting times (early in the dataset specifically). Taking away the number of recovered cases and the total number of deaths from the total number of cumulative cases we create the number of active cases. To normalize this, we divide by the population per 10,000 in order to create the variable activeper (active cases per 10,000 population).

The Oxford Stringency Index is a weighted average of all measures in the index; it represents the percentage of policies in place out of all potential policies and policy severity the index tracks. It consists of nine policy categories: school closures, workplace closures, public event closures, gathering restrictions, public transit closures, stay at home requirements, restrictions on internal movement, international travel controls and public information campaigns. Each of these variables can take different values depending on how severe the restriction. For our main regressions we use one of non-essential workplace closures, gathering restrictions of 10 people of less, recommend closing or significant route reductions of public transit, or internal travel restrictions such as checkpoints for each regression as dummy variables in addition to the overall Stringency Index value.1 The strictest value of the variable is often the one we choose.

We do not include the cancelling of public events. Some provinces like PEI had no public event cancellations but had some gathering restrictions and some provinces like Ontario cancelled public events throughout the whole dataset and had strict gathering restrictions at times. Therefore, we feel gathering restrictions would be a better measure to use because public events can be classified as large gatherings. A list of all the components and their definitions can be found on the Blavatnik School of Government website. Stay at home orders were also not prevalent through provinces in Canada, only to the degree of recommendations and not enforcement in almost all provinces so we do not include those either. We also exclude public information campaigns as we are focusing on social distance enforcing measures not recommendations. International travelling was controlled on a federal level, not provincial so that is excluded from our analysis as well.

The baseline value for the Google Community Mobility Reports, which is the average number of visitors (in the case of residential areas this would be the average duration) during a standard weekday, is calculated from the median value from “the 5-week period Jan 3-Feb 6, 2020” (Google LLC, 2021). For example, all Mondays following this period present the percentage differences on this Monday compared to the median value of all Mondays in the baseline period. In Canada, this baseline calculation period is very cold and might have a low baseline relative to a later date when reviewing park visitors for example. Our data for the report goes until the end of 2020 and does not cover the first month of 2021 as the rest of our regressions do due to data availability. The baseline calculation is done before COVID-19 hit Canada and therefore our comparisons are versus a standard pre-COVID-19 day. Google recommends removing data for local holidays to separate the effect of COVID-19 restrictions from normal holiday differentials. As a result, for our regressions we remove federal holidays and Christmas week.

2.2. Methodology

In this paper we use both an OLS and two-stage least squares regression technique to estimate the effect of policy changes on the number of deaths (cumulative to each date) per 10,000 population. Provinces enacted legislation at different dates which allows for analysis using a difference in differences regression technique to analyze treatments across provinces.

This paper will use an instrumental variable two-stage least squares regression. This is because of an endogeneity problem that arises with the formula. The number of active cases are expected to affect the number of deaths and also provide an indicator of pandemic severity at a given point in time. There are other unobservable characteristics such as behavioural changes that may be present in the error term that also affect the number of active cases. The other dependent variables in other regressions will also affect the number of active cases. This is the bases of the endogeneity problem. To correct for this, we must find a variable that affects the number of cases but not the number of deaths directly. The variable that we determine meets this criterion is the number of COVID-19 tests reported daily. As a result, in the regression equations above we instrument the number of active cases per 10,000 with the number of tests conducted per 10,000. More tests would lead to more reported cases but testing more would not lead to more deaths caused by COVID-19.

Fernández-Villaverde and Jones (2020) indicate that as a result of different testing availability and parameters, number of deaths reported due to COVID-19 is the most accurate measure to use as an indicator of the spread of the virus. Comparing the experience of British Columbia to that in Alberta provides a convincing example of this claim. Cumulative number of deaths per 10,000 as of May 19, 2020 was 0.307 compared to Alberta's 0.314. However, the confirmed cases in B.C. is roughly one third of Alberta's (6683 in Alberta and 2444 in B.C.). The difference was a direct result of the different amounts of testing done (Cardoso, 2020). To deal with this concern this paper will focus on the measure of deaths as our dependent variable.

The OLS regression for deaths based on the Oxford Stringency Index is as follows:

Y_i,t = β₀ + β₁*X_1i,t + β₂∗ X_2i,t + α_p∗X_3i + e_i,t

(1)

Where Y_i,t is the deaths per 10,000 of province i at time t, X_1i,t is the number of active cases per 10,000 population in province i at time t, X_2i,t is the 20 day lag of the stringency index of province i at time t and X_3i is a dummy variable for each province. This paper will attempt to then break down the various social distance policies in order to determine which policies appear to be more effective. That regression will be as follows. For simplicity, we replace X_2i,t as 20 plus days following a policy implementation. For example if we are talking about non-essential workplace closures, X_2i,t is 20 plus days following the implementation of non-essential workplace closures in province i at time t. Lastly, e is our error term. We will present a separate regression for each of our social distance policies.

Since their effects may not be immediate, we lag policy variables by twenty plus days following the policy implementation. The average number of days between contact with COVID-19 to death varies greatly in the literature. The Government of Canada states the range of incubation is 1–14 days with a 5–6 day median onset. The Canadian Institute for Health Information determined that the average hospital stay was 12.2 days. However, they also state that the average stay of a patient with at least one comorbidity is 19 days compared with 9 days of those who do not (Canadian Institute for Health Information, 2022). Faes et al. (2020) find that the median incubation time is 3–10.4 days, and the median ICU stay is 6.4 days for the working age and 7.6 days for the elderly. They also find that a quarter of the stays for the working age are longer than 13 days and 15.6 days for the 60–80-year-old group in Belgium. Due to the average incubation phase and potential days in the hospital we feel the effect would be most prominent twenty days or more on the number of deaths in the province. Therefore, our policy dummy will be set to one 20 days after the policy is implemented and go back to zero 20 days after it is altered or cancelled until the end of the period, Jan 31, 2021. We also take into effect timing of policies and only utilize policies which lasted at least seven days. We will examine one policy at a time and present both an OLS and 2SLS regression.

The above variables represent the policies used in the regression to analyze the relationship on mitigating the spread of the virus. Provincial dummies are included in the regressions to account for any relevant idiosyncratic provincial factors not included in the set of regressors. They will not be presented with tables but are available upon request. Additional factors are many and include differences in the operation of long-term care residences, degree and effectiveness of testing strategies, et cetera. This compounds the possible endogeneity problem for active cases as well.

A standard method for our 2SLS regression is applied. We need an over-identifying restriction. The first stage of our 2SLS regression will regress the number of active cases per 10,000 on the number of tests and a vector of our OLS regression variables as seen in equation (1). The fitted values will then be used in the second stage of our regression which will be the same as our first stage but now using the fitted values for active cases per 10,000 population.

To provide more evidence on the effectiveness of these policies we run OLS regressions using the Google Mobility Report on the variables we find negative and significant from our 2SLS instrumental variable regressions. These will be simple OLS regressions as shown below:

Y_i,t = β₀ + β₁*X_ji,1-14 + β₂∗ X_ji,15-19 + β₃∗ X_ji,20+ + α_p∗X_i + e_i,t

(2)

Where Y_i,t is a Google Mobility Report indicator as shown in Table 1 for province i at time t. X_ji is social distance policy j (we find negative and significant in our 2SLS regressions) for province i. X_ji,1-14 represents 1–14 days after the policy has been implemented in province i, X_ji,15-19 represents 15–19 days and X_ji,20+ represents 20 plus days (same as our 2SLS regression). Earlier lags are included due to the incubation stage as mentioned above. One can expect immediate behavioural changes following the policy implementation to affect deaths in the longer term. We use these regressions to show an immediate and long-term change in baseline mobility created from the implementation of these policies that would promote social distancing and hence limit spread and prevent deaths.

To check the robustness of the results, some provinces with low case and death totals have been dropped from the regressions to ensure that their inclusion does not affect the qualitative nature of the results. We find that results are robust to these changes. To further check robustness of our results, we ran a negative binomial regression that is identical in variables to our OLS regression and obtain the same significance in our non-essential workplace closures and the 20-day lag of the Oxford Stringency Index. The significance on gathering restrictions less than 10 people is no longer evident. We also ran the negative binomial regression on the number of deaths in province t on day i (to match Dave et al., 2020) and obtain statistically similar results as our negative binomial on cumulative deaths per 10000. Those regression tables available upon request.

A Durbin-Wu-Hausman test for endogeneity was applied to each regression to examine the potential correlation of the number of active cases per 10,000 on the error term in our OLS regression. We found that OLS was not consistent. Therefore, we conclude that there is an endogeneity problem from the correlation between the number of active cases and the error term. As a result, the 2SLS regressions are more econometrically sound. We will focus on results from these regressions for that reason.

3. Results

Table 2, Table 3, Table 4, Table 5, Table 6, Table 7, Table 8, Table 9 provide the results of the OLS and 2SLS IV (two-stage least squares instrumental variable) regressions on deaths per 10,000 for varying contact limiting social distance policies. The Google Mobility Report regressions are in Table 10, Table 11, Table 13 . For reasons discussed in the introduction, inferring causality from regression coefficients is not straightforward. However, in the case of negative and significant coefficients, we proceed with such statements inferring causality recognizing that caution must be applied when considering these interpretations. We find significant support for social distance policies lowering the number of deaths across Canada.

Table 2.

Non-Essential workplace Closure Regressions.

	(OLS)	(2SLS)
	deathsper10000	deathsper10000
activeper	0.0547***	0.141***
activeper	(0.00211)	(0.00439)

20+ days after	−0.173**	−0.531***
Non-essential work closures	(0.0600)	(0.0764)
_cons	0.371***	−0.0833
_cons	(0.0687)	(0.0878)
N	2931	2931

	(1)	(2)	(3)	(4)	(5)	(6)
	retail_and_recreation_percent_ch	grocery_and_pharmacy_percent_cha	parks_percent_change_from_baseli	transit_stations_percent_change_	workplaces_percent_change_from_b	residential_percent_change_from_
1–14 days	−32.04***	−20.37***	−81.62***	−21.55***	−20.95***	10.19***
Since non- essential workplace closures	(1.158)	(1.021)	(6.715)	(1.076)	(1.301)	(0.457)

15–19days	−32.40***	−20.92***	−80.58***	−22.58***	−22.78***	10.54***
Since non- essential workplace closures	(1.877)	(1.654)	(10.90)	(1.741)	(2.108)	(0.741)

20+ days	−20.86***	−11.34***	−31.00***	−15.77***	−16.19***	7.062***
Since non- essential workplace closures	(0.728)	(0.642)	(4.287)	(0.680)	(0.818)	(0.287)

_cons	−18.14***	3.843***	84.72***	−40.10***	−29.17***	8.903***
_cons	(0.826)	(0.728)	(4.748)	(0.765)	(0.927)	(0.326)
N	2562	2546	2354	2484	2562	2562

Variable Name	Brief Description	Mean	Min Value	Max Value
Deathsper10000	cumulative number of deaths per 10,000 population by province	1.222193 (2.223835)	0	11.99604
retail_and_recreation_percent_ch	Percent change from baseline in retail areas such as malls, restaurants, theme parks, etc.	−18.49352 (19.96068)	−90	65
grocery_and_pharmacy_percent_cha	Percent change in the number of visitors to grocery and pharmacy stores from baseline	−2.214408 (15.58345)	−87	69
parks_percent_change_from_baseli	Percentage change in park visitors from baseline	67.02819 (83.51918)	−68	407
transit_stations_percent_change_	Percent change in the number of visitors to public transit areas from baseline (subway, bus, train, etc.)	−42.30322 (17.67749)	−84	48
workplaces_percent_change_from_b	Percent change in workplace visitors from baseline	−30.32091 (18.74058)	−86	29
residential_percent_change_from_	Percentage change in duration in places of residence compared to baseline	9.655196 (6.913145)	−10	36
Independent Variables
Stringency Index	Percent of contact limiting provincial policies in place and public information campaign as defined by the Oxford Index	63.37434 (11.43394)	11.11	90.74
Activeper	Number of active cases per 10,000 population	6.471294 (11.69082)	0	72.43627
20+ Days after School Closings	Provincially declared school closures for all levels	0.3565749 (0.479061)	0	1
20+ Days after Some sectors workplace restrictions	Work closure short of non-essential/workplace regulations (i.e. limited occupations)	0.4507645 (0.4976461)	0	1
20+ Days after Non-Essential workplace closures	Only essential business allowed open unless in a work from home scenario	0.1782875 (0.3828131)	0	1
20+ Days after Lighter gathering restrictions	Gathering size restriction in the range of 11–100 people	0.5250765 (0.4994471)	0	1
20+ Days after Strict gathering restrictions	Gathering size restrictions on 10 people or less	0.3587156 (0.4796969)	0	1
20+ Days after Internal travel restrictions	Internal movement restrictions such as checkpoints and border closing in place	0.4899083 (0.4999746)	0	1
20+ Days after Public transit closures	Closing requirement on some public transportation or Route reductions	0.2672783 (0.4426065)	0	1
Provincial Dummies	Accounting for unidentified provincial effects	N/A	0	1

	(OLS)	(IV 2SLS)
	deathsper10000	deathsper10000
activeper	0.0555***	0.143***
activeper	(0.00211)	(0.00451)

20+ Days	−0.397***	−0.853***
After Travel Restrictions	(0.0736)	(0.0946)

_cons	0.454***	0.0646
_cons	(0.0708)	(0.0906)
N	2931	2931

	(OLS)	(IV 2SLS)
	deathsper10000	deathsper10000
activeper	0.0562***	0.142***
activeper	(0.00215)	(0.00441)

20 day lag of	−0.0106***	−0.0332***
Stringency Index	(0.00235)	(0.00306)

_cons	1.013***	1.947***
_cons	(0.163)	(0.205)
N	2931	2931

	(1)	(2)	(3)	(4)	(5)	(6)
	retail_and_recreation_percent_ch	grocery_and_pharmacy_percent_cha	parks_percent_change_from_baseli	transit_stations_percent_change_	workplaces_percent_change_from_b	residential_percent_change_from_
1–14 days	−23.49***	−14.04***	−98.81***	−18.63***	−10.59***	7.354***
Since 10 people or less gathering restriction	(1.200)	(1.000)	(5.969)	(1.016)	(1.293)	(0.453)

15–19days	−24.05***	−15.37***	−102.9***	−21.01***	−13.05***	8.233***
Since 10 people or less gathering restriction	(1.895)	(1.578)	(9.345)	(1.600)	(2.041)	(0.715)

20+ days	−14.21***	−7.241***	−43.77***	−14.05***	−8.545***	5.336***
Since 10 people or less gathering restriction	(0.754)	(0.628)	(3.856)	(0.643)	(0.812)	(0.285)


_cons	−20.71***	2.276**	86.08***	−41.34***	−31.75***	9.690***
_cons	(0.926)	(0.771)	(4.550)	(0.781)	(0.997)	(0.349)
N	2562	2546	2354	2484	2562	2562

	(1)	(2)	(3)	(4)	(5)	(6)
	retail_and_recreation_percent_ch	grocery_and_pharmacy_percent_cha	parks_percent_change_from_baseli	transit_stations_percent_change_	workplaces_percent_change_from_b	residential_percent_change_from_
Stringency	−1.114***	−0.700***	−1.517***	−0.968***	−0.941***	0.377***
Index	(0.0251)	(0.0223)	(0.156)	(0.0210)	(0.0273)	(0.00989)


_cons	47.51***	45.39***	171.1***	17.75***	26.89***	−13.32***
_cons	(1.791)	(1.591)	(11.10)	(1.498)	(1.945)	(0.706)
N	2562	2546	2354	2484	2562	2562

PERMALINK

Effects of non-pharmaceutical interventions on the death rates from COVID-19 in Canada

Alexander Meek

Abstract

1. Introduction

Fig. 1.

Fig. 2.

Fig. 3.

2. Data and methodology

2.1. Data

2.2. Methodology

3. Results

Table 2.

Table 3.

Table 4.

Table 5.

Table 6.

Table 7.

Table 8.

Table 9.

Table 11.

Table 10.

4. Conclusion

CRediT authorship contribution statement

Declaration of competing interest

Footnotes

Appendix.

Fig. 1.

Fig. 2.

Fig. 3.

Fig. 4.

Fig. 5.

Fig. 6.

Fig. 7.

Fig. 8.

Fig. 9.

Fig. 10.

Table 1.

Table 2.

Table 3.

Table 4.

Table 5.

Table 6.

Table 7.

Table 8.

Table 9.

Table 10.

Table 11.

Table 12.

Table 13.

References

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases