Rating the digital help: electronic medical records, software providers, and physicians

Abstract

To separate the effects of physicians’ characteristics on the perceived productivity of EMRs from the effects of limitations on usability inherent in EMR design, a multivariate regression model is used to estimate the factors influencing physicians’ rankings of five attributes of their EMRs, namely; ease of use and reliability; the EMRs effect on physician and staff productivity and the EMRs performance vs. vendor’s promises. We divide the factors influencing the rankings into three groups: physician characteristics, EMR characteristics and practice characteristics (type of practice, size, and location). The data are from approximately 1800 practicing physicians in Arizona. Physician’s characteristics influence perceived ease of use and physicians’ productivity, but not staff productivity, reliability or vendors’ promised performance. Practice type and EMR characteristics affect perceived productivity, reliability and performance versus vendors’ promises. Vendor-specific effects are highly correlated across all five attributes and are always jointly significant. EMR characteristics are the most significant influence on physicians’ perceptions of the EMRs effect on their productivity and that of their staff. Physicians’ characteristics (particularly age) have a small but significant influence on perceived productivity.

Appendix

Appendix with ordered logistic regression results for “Rating the Digital Help: Electronic Medical Records, Software Providers, and Physicians”

An ordered logit regression accounts for the interval nature of outcomes (namely, the dependent variable only takes the integer values 1, 2, 3, 4, or 5) by calculating the likelihood that individual i’s index function ($X_i \beta )$ falls in ordered predictive buckets, where $\beta$ are the empirically calculated weights for the physician-EMR-practice variable vector $X_i$. The constructed buckets are one of five intervals of a logit distribution function, the transition to each bucket having a distinct intercept value ($\alpha _i )$. Hence the likelihood to be maximized, by choosing the slope coefficients (the $\beta )$ for data falling in one of j intervals, $L\left( {\alpha _j +X_i \beta } \right) -L(\alpha _{j-1} +X_i \beta )$ (with $z_{i,j}$ is an indicator function equal to one when the ith observation falls in the jth interval), is

$$\begin{aligned} \mathop \sum \limits _j \mathop \sum \limits _i z_{i,j} \hbox {ln}\left( {L\left( {\alpha _j +X_i \beta } \right) -L\left( {\alpha _{j-1} +X_i \beta } \right) } \right) \end{aligned}$$

A positive coefficient in Table 5 represents the effect of a unit increase in the respective variable on the log-odds ratio of moving up one more productivity index value (Tables 6, 7).

Table 5 Ordered logit models of perceived EMR productivity (1-awful to 5-outstanding) for all physicians practicing in Arizona

Table 6 Correlations based on Table 5 ordered logit vendor coefficients

Table 7 Ordered logit regression models of perceived EMR productivity (1-awful to 5-outstanding, estimating the log odds effects), for primary doctors only