- Variable Measurement and Data Sources
This study created 30 provincial balanced panel data of China from 2014 to 2017. The main variables and data sources were as follows.
The numbers of physicians, registered nurses, pharmacists, clinical microbiologists per 1000 populations were selected as independent variables. The number of clinical microbiologists was estimated by technicians.
The number of daily visits of physicians in hospitals was used as moderation variable since the increase of daily visits of physicians indicates that the medical professionals need to work longer hours or take fewer breaks to meet medical demand. So it has become one of the factors to be considered in measuring medical professionals’ workload [13,18].
The number of tertiary hospitals was used as the control variable. Due to the incomplete establishment of hierarchical diagnosis and treatment system in China, patients tend to choose large general hospitals which with high qualitymedical resources. Large hospitals are usually overcrowded and concentrate most cases of pseudomonas aeruginosa infections. CRPA rates, diagnostic proficiency, diagnostic availability and competence with infection control practices in every province are also hospital-specific and are influenced by the number of large hospitals.
The rates of CRPA were obtained from the China Antimicrobial Resistance Surveillance System (CARSS) 2014–2017.The independent variables, moderation variable and control variable were obtained from China Health and Family Planning Statistical Yearbooks 2015- 2018[20,21,22, 23].
The summary statistics of the above variables were shown in Table 1. The number of registered nurses per 1000 people was the highest, followed by physicians, clinical microbiologists and pharmacists.
Table 1. Variable definitions and summary statistics
Variable
|
Definition
|
N
|
Mean
|
SD
|
Min
|
Max
|
CRPA
|
Rate of carbapenem-resistant Pseudomonas aeruginosa
|
120
|
21.445
|
6.514
|
8.7
|
36.4
|
PHY
|
The number of physicians per 1000 populations
|
120
|
1.959
|
0.476
|
1.3
|
4.1
|
NUR
|
The number of registered nurses per 1000 populations
|
120
|
2.524
|
0.525
|
1.65
|
4.8
|
PHA
|
The number of pharmacists per 1000 populations
|
120
|
0.325
|
0.083
|
0.186
|
0.648
|
MCB
|
The number of clinical microbiologists per 1000 populations
|
120
|
0.335
|
0.065
|
0.236
|
0.587
|
WORKL
|
The number of daily visits of physicians in hospitals
|
120
|
7.170
|
2.520
|
3.7
|
15.20
|
HOS
|
The number of tertiary hospitals
|
120
|
71.825
|
40.757
|
7
|
170
|
SD: standard deviation.
Statistical Analysis
Hierarchical regression analysis was used to test the moderation effect. The following equation was constructed to test the relationship between the number of medical professionals and CRPA rate firstly:
Secondly, in order to test the effect of the number of medical professionals on CRPA rate under the moderation of workload, it was gradually incorporated into the workload (equation (2)) and the interaction term between workload and the number of medical professionals (equation (3)). To avoid multi-collinearity effects, variables in the interaction terms were mean-centered in equation (3).If the regression coefficients of interaction term were statistically significant, the moderation effect of workload was significant [24]:
Where i (i = 1, 2, 3,…, 29, 30) indicated province, and t (t = 1, 2, 3, 4) indicated year, μi was an unobservable regional effect, εit was a random error term. STAFF indicated one type of medical professionals as physicians, registered nurses, pharmacists, clinical technicians. Each variable was taken in logarithmic form to eliminate the influence of heteroscedasticity.
Panel data model is a quantitative method for longitudinal data, which could increases the estimator precision by increasing observations and obtain more dynamic information than a single cross-sectional data [25]. The pooled ordinary least squares (OLS) model, fixed-effect (FE) model and random-effect (RE) model are often used to estimate panel data. Because OLS does not control for the fixed effect of provinces, the estimation caused by endogenous problems may be biased. FE model or RE model is further applied for estimation. RE model is relatively more effective, but exogenous variables and individual effects are not required to be correlated, while FE model, although there is no requirement between the exogenous variables and individual effects, consumes more degrees of freedom [26]. Hausman test can determine which model was more appropriate for result estimation. As shown in Table 2, every test rejected the null hypothesis, indicating that FE model was more reasonable (Table2).
Table 2. Panel data model estimation
Variable
|
FE
|
RE
|
FE
|
RE
|
FE
|
RE
|
FE
|
RE
|
ln PHY
|
-0.888***
|
-0.706***
|
|
|
|
|
|
|
ln NUR
|
|
|
-0.757***
|
-0.833***
|
|
|
|
|
ln PHA
|
|
|
|
|
-1.126***
|
-0.731***
|
|
|
ln MCB
|
|
|
|
|
|
|
-0.785***
|
-0.728***
|
ln WORKL
|
0.001
|
0.524***
|
0.181
|
0.537***
|
0.398
|
0.678***
|
0.296
|
0.486***
|
ln HOS
|
-0.418***
|
0.046
|
-0.316**
|
0.094
|
-0.443***
|
0.021
|
-0.433***
|
0.039
|
Constant
|
5.289***
|
2.285***
|
4.646***
|
2.359***
|
2.764**
|
0.790
|
3.345***
|
1.119**
|
Hausman test value
|
33.30***
|
24.72***
|
33.45***
|
34.28***
|
P
|
0.000
|
0.0001
|
0.000
|
0.000
|
FE: Fixed-effect (FE) model; RE: random-effect (RE) model;
*** p<0.01, ** p<0.05, * p<0.1
STATA(version 13.0, Stata Corp., College Station, Texas, USA) was used for the statistical analysis.