Citizens and the state during crisis: Public authority, private behaviour and the Covid‐19 pandemic in France

Abstract How do democratic states induce citizens to comply with government directives during times of acute crisis? Focusing on the onset of the Covid‐19 pandemic in France, I argue that the tools states use to activate adherence to public health advice have predictable and variable effects on citizens’ willingness to change their routine private behaviours, both because of variation in their levels of restrictiveness but also because of differences in people's political motivations to comply with them. Using data collected in March 2020, I show that people's reports of changes in their behavioural routines are affected by the signals governments send, how they send them and the level of enforcement. I find that a nationally televised speech by President Macron calling for cooperative behaviour and announcing new restrictions elevated people's willingness to comply. Moreover, while co‐partisanship with the incumbent government increased compliance reports before the President's primetime television address, presidential approval boosted reports of compliance after.


Government Policies
The stringency of government policies measure was collected by researchers at the University of Oxford who collected and coded publicly available information on 17 indicators of government responses. Eight of the policy indicators record information on containment and closure policies, such as school closures and restrictions in movement. Four of the indicators record economic policies, such as income support to citizens or provision of foreign aid. Five of the indicators record health system policies such as the COVID-19 testing regime or emergency investments into healthcare. The data from the 17 indicators are aggregated into a set of four common indices, reporting a number between 1 and 100 to create an overall government response index as well as specific indices measuring responses in the areas of containment and health index; economic support, and stringency (which records the strictness of 'lockdown style' policies that primarily restrict people's behavior).
The analysis utilizes the overall stringency index, which combines information about restrictions on movement, school and business closures, and so on together with information about the strictness of health containment policies. For further details on coding, see https://www.bsg.ox.ac.uk/research/research-projects/coronavirus-government-responsetracker. (also see https://www.bsg.ox.ac.uk/sites/default/files/2020-05/BSG-WP-2020-032-v6.0.pdf for further details on the index).

Individual-Level Data and Measures
To examine individual-level attitudes and behaviors, the analysis relies on panel surveys collected as part of the French National Election Study (Enquête Électorale Française-ENEF). These surveys are based on a panel of nationally representative samples of adults. I used data from three panel waves collected in March 2020, supplemented with data on respondent characteristic recorded in prior years (Brouard 2020  Respondents were asked to indicate agreement or disagreement on a 0 to 10 scale, with "0" meaning "no, not at all" and "10" meaning "yes, absolutely".

Presidential Approval
Based on question asking respondents to rate how satisfied they were with the performance of the President on a scale from 0 to 10, with 0 indicating "absolutely not satisfied" and 10 indicating "absolutely satisfied".

LREM Partisanship
Based on a series of questions asking respondents to indicate whether there is a party they feel close to or less distant from, and then to indicate which party that is.

Extreme Left
Based on a question asking respondents: "On a scale of 0 to 10, where 0 corresponds to the left and 10 corresponds to the right, where would you say you are?" Respondents were recoded so that 0-2 were coded as extreme left.

Extreme Right
Based on a question asking respondents: "On a scale of 0 to 10, where 0 corresponds to the left and 10 corresponds to the right, where would you say you are?" Respondents were recoded so that 8-10 were coded as extreme right.
Health Risk Acceptance Based on survey question that asked: "Is it easy or difficult for you to accept or take risks that … affect your health"?

Personal Health
Based on survey question that asked: "Generally speaking, you would say that your health is .

Measuring Compliance Behaviors
Because the measure of compliance is based on people's reported behavior rather than their observed behavior, validity and reliability concerns cannot be completely ruled out.
However, this shortcoming is mitigated in several ways. First, asking about specific health and social behaviors is simple and easy to understand. Second, the survey asks respondents about recent behavior during moments of acute crisis; as such, the context is vivid and the question should not require much thought. Third, given that there is widespread public evidence that the French public did indeed comply with government demands, high levels of reported compliance provide meaningful face validity that the question indeed measures compliance.
Factor analyses showed that these items also loaded highly on a single factor.
Specifically, I conducted factor analyses for these behaviors in the two waves separately. In both waves, only one factor achieved an Eigenvalue greater than 1 (wave 1: 4.32; wave 2: 3.13). I combined these items into a scale that averaged responses across the items. The combined scale (also ranging from 0 to 10) had very high reliability (Cronbach's alpha was 0.89 and 0.77 for reported behaviors in the March 16 and March 24 waves, respectively).

Entropy Balancing
In addition to using the population weighted sample for the multiple regression models, the estimations also include a so-called balanced specification. This procedure uses entropy balancing (Hainmueller 2012) to reduce any imbalance between respondents interviewed before or after the speech by President Macron. Table A.3 shows the descriptive statistics of for the outcome and control variables used in the analysis. It also separates respondents by whether they were surveyed before or after 8pm on March 16 and provides information about the sample's potential imbalance before and after the speech. This imbalance appears small for most variables (including age, education, income, employment, and living with a partner), and slightly larger for sex, which may bias estimates if not corrected. I therefore also use a specification that weights observations through entropy balancing (so called Balance specification), which helps to increase internal validity by weighting the distribution of covariates among non-treated respondents to make it mimic the first and second moment of the equivalent distribution among treated respondents.

Appendix B. Supplemental Analyses and Robustness Tests
To establish the robustness of the findings, this section examines potential challenges to the inferences. These relate to the data, measures, and estimation methods.

Estimations of Individual Behavioral Items
To confirm that the effects of political and policy attitudes are consistent across the range of behaviors that make up the compliance index, I estimated a set of models for each compliant behavior separately. These are shown in Table B.1. Even though the impact varies, the conclusions reported in the paper are consistently supported across the different behaviors.

Placebo Tests
I also undertook several placebo tests to establish that the relationship between individual attitudes and characteristics and subsequent reports of voluntary compliance is not simply part of a broader syndrome of behavioral adjustments or survey responses during times of crisis. If exposure to the pandemic and government announcements change reports of behaviors not explicitly related to the restrictions enacted by governments, it would be misleading to interpret the findings above as the effect of government policies on voluntary compliance.
To examine this possibility, I estimated identical models (in terms of explanatory variables) for placebo outcomes related to changes in other kinds of behaviors in the wake of the threat posed by the pandemic. These include stocking up on provisions, drinking more alcohol, and cleaning the home more frequently than before the crisis. Results reported in Table B.2. show that the political variables had few and inconsistent effects -while presidential approval reduces the odds of cleaning the home, LREM partisanship reduces the propensity to drink more alcohol than before the crisis. The most notable and plausible impact is found for perceptions of the virus as a public health threat, which increased people's propensity to buy provisions and clean the home.  Results demonstrate that -despite the novelty of LREM -levels of partisanship were quite stable, especially given the upheaval in the party system. They also compare favorably to the level of partisan identification in other democracies and individual-level attachments to LREM over successive panel waves were both stable and similar to attachments to other political parties.

Do attachments to other parties matter?
Given the comparatively fragmented nature of the French party system, one additional question is whether attachment to a party other than president's party and the parliamentary majority has the opposite effect to being a government co-partisan. This question is especially relevant in light of anecdotal evidence that compliance with government restrictions during the pandemic has become a partisan issue in a number of democracies, most prominently in the U.S. Specifically, I therefore estimated models identical to those reported in Table 1, but with partisanship for each of the main opposition parties as additional independent variables. Because model results are essentially identical across specifications, I plot the coefficients from the balanced estimation model in Figure B.1. below. They show that attachments to these parties had no systematic effects on reported compliance.

Figure B.1. Effects of Partisanship on Compliance Behaviors
Coefficients are unstandardized coefficients from balanced logistic regression. After Macron Speech