Study population
We used data from the national panel of French self-employed GPs, set up in 2018 [13].
GPs were randomly selected from a French exhaustive database of health professionals as of January 1, 2018 and who have signed an informed consent at inclusion (in 2019) to answer six future cross-sectional surveys (one every 9 months). Sampling was stratified for gender, age, workload (annual number of office and home visits; in terciles) and practice location in low GP density municipalities. The panel is representative of GPs practicing in France (excluding Mayotte). GPs planning to retire or to move before the end of data collection, those exclusively practicing alternative medicine as well as those with few gatekeeping duties (fewer than 200 patients) were excluded. The sample benefits from the French “public statistics” label of the National Authority for Statistical Information (Conseil National de l’Information Statistique).
Procedure and questionnaire
At inclusion in 2019, professional investigators contacted GPs to ask them to participate, obtain their consent, and verify inclusion criteria; they then conducted the inclusion interview, collecting information about GPs’ professional characteristics, using computer-assisted telephone interview (CATI) software.
A special Covid-19 cross-sectional survey was decided in March 2020 in order to study GP practices in the face of the pandemic. In line with the prior systematic collection of longitudinal data planned in 2019, an ad-hoc committee was formed at the beginning of the first lockdown in France in order to elaborate short-form surveys focusing on the GPs’ response to the COVID-19 crisis. The committee was comprised of various public health scientists (epidemiologists, sociologists, and health economists), already trained in surveys among GPs. Since there was no past experience of lockdowns, the questionnaire was not based on standardised questions, although it was inspired by literature dealing with the H1N1 crisis [14]. We pilot-tested the questionnaire for clarity, length, and face validity among 6 GPs and modified the wording of several questions found to be unclear.
The web-survey took place between April 9 and April 21 2020; 1191 GPs out of 2761 contacted (43.1%) responded. Since the eligibility criteria were verified at inclusion (in 2019), all the remaining GPs were eligible to participate in the survey in 2020.
We exploit the part of the questionnaire focused on the impact of the lockdown on GPs’ activity (the week before the survey compared to a “typical” week before the pandemic). A “don’t know” answer was also included in each item of the questionnaire.
We used an indicator variable of the intensity of the Covid-19 pandemic at the département level (France is divided into 18 regions that are further subdivided into 101 départements; each département belongs to only one region). This indicator was constructed by the Directorate for Research, Studies, Assessment and Statistics (DREES, French Ministry of Health) from National Institute of Statistics and Economic Studies (Insee) Covid-19 mortality data collected between March 1 and April 20 2020 [15]. We used a dummy variable to isolate the most affected départements, where an average change in excess mortality rate was 110.5%.
In addition, we used a set of five dummy variables regarding motivations for the choice of the current practice location from Round 1 of the survey that took place from October 2018 to April 2019. The GPs had to choose between the items related to 1) healthcare services available in the area, 2) the possibility of creating or joining a group practice, 3) the search for an area with low GP density, 4) available amenities for the GPs’ families, or 5) a previous experience as an intern, locum or associate in the area; it was possible to choose several items simultaneously.
Finally, we exploited a previous edition of the survey (Round 3 of the third national panel of French self-employed GPs, December 2015 to March 2016) to construct a dummy variable indicating the presence of a MGP in a département in 2013.
Statistical analysis
To correct for possible systematic non-response bias in our subsample, we used weights to match the nationwide GP population for the four main stratification variables: age, gender, workload and GP density. When missing data occurred for some variables, they were included in the analysis and treated as a separate ‘missing’ category.
We defined a set of dependent variables regarding chronic care management (see Additional file 1 for a detailed description of the questionnaire): (1) a dummy variable reporting estimated variation in the number of weekly visits related to complications of chronic diseases (“Over the past week, what was the change in the frequency of visits related to complications of previously stable chronic diseases, compared to a typical week before the epidemic of Covid-19?”), as well as (2) a dummy variable indicating whether the GP makes a proactive effort to contact her chronic patients herself (“To address the current care needs of your most at-risk chronic patients, do you take an active approach to contacting them (by phone or other means of communication)?”).
We use probit regressions to estimate the following model:
$${Y}_i={\alpha}_i+{\beta}_1\ {MGP}_i+{\beta}_2\ Covid19\ {indicators}_i+{\beta}_3\ {Control\ variables}_i+{\varepsilon}_i$$
where
-
Yi is one of the dependent variables described above,
-
MGPi is a dummy variable indicating practicing in MGP,
-
Covid-19 indicatorsi include the indicator variable of the intensity of the pandemic presented above, as well as GPs’ perceptions regarding the severity of Covid-19 (from 0 “not at all severe” to 10 “extremely severe”) and their estimation of the percentage of the French population that would be contaminated by Covid-19 by the end of 2020 (less than 50, 50 to 75%, more than 75% of the population),
-
and the control variables include GP i’ personal and professional characteristics: gender, age (in tertiles), workload (in tertiles), as well as a dummy variable indicating whether the practice is located in the area with the lowest (first decile) GP density in 2018 or not.
To address the possible endogeneity of choice of practicing in MGP, we estimate the bivariate probit model, since it is suited to simultaneously estimating both equations with binary outcome variables:
$$\left\{\begin{array}{c} MG{P}_i={\alpha}_i+{\gamma}_1\ {women}_i+{\gamma}_2\ {age}_i+{\gamma}_3\ {workload}_i+{\gamma}_4\ lowest\ GP\ {density}_i+\\ {}{\gamma}_5\ Pioneer\ d{\acute{e}}{partement}_i+{\delta}_i\ {Motivation}_i+{\varepsilon}_{i1}\\ {}{Y}_i={\alpha}_i+{\beta}_1\ {MGP}_i+{\beta}_2\ Covid19\ {indicators}_i+{\beta}_3\ {Control\ variables}_i+{\varepsilon}_{i2}\end{array}\right.$$
where the equation related to practicing in MGP contains several variables that might influence the GPs’ choice, but are unlikely to have an impact on chronic care management strategies during the sanitary crisis:
-
Pioneer départementi is a dummy variable indicating practices located in a département that had adopted MGP early (before 2013),
-
Motivationi is a set of dummy variables regarding the selection criteria prior to the choice of the current practice location (healthcare services available, possibility of creating or joining a group practice, search for an area with low GP density, available amenities, or previous medical studies (a previous experience, e.g., internship, in the area)).
These two sets of instruments relate to the past behaviour of GPs or the behaviour of peers in the same département before 2013 (we use the term ‘instrument’ to emphasise the fact that these variables allow the identification of the model, even if the estimated model is not using a two-step procedure). This strengthens the fulfilment of the exclusion condition. As bivariate probit models do not allow testing for overidentification restrictions, we follow the procedure described in Wooldridge [16]. Using a linear probability model, we calculate the fitted value for MGPi. Next, we estimate the model described above by a Two-Stage Least Squares using the fitted value for MGPi. Having this adapted framework at our disposal, we were able to test the statistical properties of the instrumental variables using the Sargan overidentification test (The instruments are considered valid when they are correlated with the explanatory variables and uncorrelated with the error term. An instrument is considered weak when there is significant correlation between the instrument and the explanatory variable, but with a low value for the correlation coefficient). For both models, we calculate the marginal effects (on predicted probabilities) for GPs practicing outside MGP, outside the most affected départements, male, youngest, with lowest workload, those not practicing in lowest GP density areas.
All analyses were conducted with Stata 14 (StataCorp. College Station, Texas).