The role of waiting times in GP referral decisions for suspected cancer.

Whilst NICE guidelines provide recommendations on symptoms which warrant an urgent referral for suspected cancer, signs such as weight loss and fatigue are not always distinguishable from other less serious conditions. GPs must therefore find a balance between referring all patients and inducing unnecessary anxiety and potential overtreatment through further testing [20]. GPs may have different thresholds for referring to hospital when symptoms are unclear, and this may impact outcomes [21]. Thresholds for cancer referral have been found to vary significantly between general practices despite the presence of these guidelines, as has the accuracy of referral decisions [22]. Research has also found compliance with the NICE referral guidelines to be poor even in situations where patients present to their GP with “red flag” features including breast lumps and rectal bleeding [23]. Most patients subsequently diagnosed with cancer see their GP multiple times before being referred to specialist hospital services [24, 25], providing further evidence that accurately spotting cancer is difficult even with the aid of referral guidelines.

We hypothesise that whilst GPs will always refer patients whom they suspect to have cancer, local hospital waiting times may influence referral decisions on the margin for patients whom the GP believes cancer is unlikely, i.e., for patients where the diagnosis is particularly unclear. The direction of the effect could therefore go either way, driven by two competing possible response mechanisms.

Amongst patients for whom the GP does not strongly suspect cancer but an alternative diagnosis is also not obvious, reduced waiting times on the suspected cancer pathway may lead GPs to refer more patients with the hope of reducing the period of uncertainty for the patient and ruling out cancer. This would mirror previous findings on the GP demand response to shorter waiting times for non-urgent and elective surgery [26, 27], where referral volumes were reported to rise as waiting times fell.

Alternatively, shorter waiting times on the suspected cancer pathway may act as a safety net for GPs who do not suspect cancer, allowing them more scope to explore other alternatives first in the knowledge that subsequent waiting times will be short if these explorations do not result in an alternative diagnosis. Anxiety around cancer is very high compared to other conditions [28], and so GPs may prefer not to raise the possibility of cancer with patients unnecessarily when they believe it to be unlikely. However, when waiting times are long, GPs may decide it is necessary to start the referral process early just in case, and so be more inclined to refer patients whom they do not believe have cancer onto this pathway as a precaution.

A priori, we do not know which of these potential mechanisms will dominate.

Receipt of waiting times information by GPs.

Waiting times can only play a role in referral decisions if GPs have access to information on waiting times at the local hospitals to which they can refer. This information would have reached GPs in at least two ways during the period we study. Hospital performance against the national waiting times standards was publicly reported on a monthly basis [29], with a three month delay. GPs were therefore able to access this performance data easily, and could follow trends over time at each hospital. Secondly, GPs could observe the waiting times of their own patients directly once they were referred. The average general practice in England made the equivalent of approximately 65 annual referrals for suspected cancer per full time equivalent GP [30]. GPs therefore repeatedly observed the waiting times experienced by their own patients throughout the year, providing direct personal experience of whether these were changing over time.

Sample.

Our sample consisted of 6,667 GP practices referring to 185 hospitals in England, between the financial years 2012/13 and 2017/18. We obtained annual data on GP practices from the 2018/19 update of Public Health Profiles provided by Public Health England [31]. Only practices with a list size of over 1,000 registered patients in 2018/19, and practices which are in the GP Patient Survey (a patient experience survey of services provided by the GP practice) are included in the Public Health Profiles data.

We obtained hospital level data on waiting times from NHS England [32]. We used Hospital Episode Statistics (HES) data on two-week wait appointments to link practices to local hospitals. As data from HES are available up to the financial year 2017/18 we included up to this year of GP practice level data from Public Health Profiles in our analysis. The final sample included 37,556 practice-year observations. We do not have data for all periods for all practices due to practice openings and closures during the period of interest. There were also 8,208 missing observations on practices when they did not refer any two-week wait patients in HES in a given year. We carried out a sensitivity analysis restricting the analysis to practices with observations in every year.

Ethical approval was not required for this study as data obtained were publicly available, published data on English GP practices and hospitals. HES data received for the purpose of linking the practice and hospital level data had been de-identified.

Variables.

Outcome variable. We examined the practice-level count of urgent GP referrals for suspected cancer. Urgent GP referrals were assigned to the financial year in which that patient was referred. It is possible that an individual patient could have appeared more than once, but only if they were referred for suspected cancer on multiple occasions by their GP.

Variable of interest. Our variable of interest was local hospital waiting times. Average length of waiting times for suspected cancer are not published at the practice or provider level. Therefore, as a measure of waiting times performance, we used hospital breaches of the two-week wait target, as a proportion of total suspected cancer patients referred [32]. This measure was calculated retrospectively, based on how long patients who were seen for an urgent referral had waited since the time of referral.

Waiting time breaches as a proportion of total patients referred are available at the hospital level. Many practices do not exclusively refer all of their patients to a single hospital, although there are some that do. For each general practice we therefore created a weighted local hospital waiting times measure. The measure was weighted by the observed shares of urgent suspected cancer referrals from the practice to each hospital taken from HES. We used the average shares over time for each practice to avoid changes in hospital referral destination as a potential response to waiting times.

The calculation of the local waiting times indicator (LW_pt) is outlined in Eq 1, such that: Eq 1

In which,

p = practice

h = hospital

n = number of hospitals to which the practice refers suspected cancer patients

W_ht = breaches of the two week wait rule as a proportion of the total suspected cancer patients referred, in a given hospital-year

R^HES_ph = total (attended) referrals from practice p to hospital h

R^HES_p = total (attended) referrals made by practice p

Covariates. We included several covariates which may be correlated with local hospital waiting times and GP referral activity. We controlled for: the size of the population registered with the GP practice (list size); characteristics of the registered practice population’s age (proportion aged under 18, proportion aged 65+), health status (proportion with a long-standing health condition) and employment status (proportion unemployed). We also controlled for patient-reported access to the practice from the GP Patient Survey (proportion reporting good experience of making an appointment, proportion satisfied with phone access), and clinical quality (proportion Quality and Outcomes Framework (QOF) points achieved across all domains as a proportion of all achievable points). The QOF measures practice performance against a number of quality indicators [33].

GP demand analysis.

We explored a general model of the form expressed in Eq 2: Eq 2

Where R^PHP_pt is total referrals (to any hospital) from GP practice p in year t. The PHP superscript indicates that this practice-level measure of two-week wait referrals was taken from the Public Health Profiles data (Public Health England, 2020). LW_pt is the local average waiting time faced by practice p in year t (see the section “Variable of interest” for calculation). X_pt are the set of practice-level, time-varying covariates outlined in the section “Covariates”. τ_t denotes time fixed effects and ε_pt is the idiosyncratic error term.

We first estimated Eq 3 using pooled OLS: Eq 3

Next, we estimated between-practice effects in Eq 4 using OLS. This focuses on the cross-sectional association between waiting times and referrals by regressing the mean of the outcome variable, across all time periods, for each practice, on the mean values of each of the explanatory variables: Eq 4

We then estimated the within-practice effects using a Dummy Variable Ordinary Least Squares model: Eq 5

Eq 5 is an altered version of Eq 3 which now includes fixed effects (or dummy variables) for each practice (α_p). These control for factors that are not measured but are constant over time for each practice (such as location).

By utilising these three regression models, we can identify whether associations we find are driven by between-practice (cross-sectional) or within-practice (over time) variation. As the fixed effects analysis controls for unobserved characteristics of a practice which are constant over time, this is the analysis least vulnerable to unobserved confounding.

We used inverse hyperbolic sine transformations of all variables in all models. Using inverse hyperbolic sine transformations is recommended to transform right-skewed variables which include zeros, approximating a normal distribution and reducing the effect of outliers [34]. The inverse hyperbolic sine transformation of a variable approximates the natural logarithm of that variable, while not losing observations which are zero. Transforming these variables allowed us to interpret the coefficients as elasticities. In this instance, these will measure the elasticity of demand with respect to waiting times, i.e. the responsiveness of GP referral demand to changes in hospital waiting times. An elasticity estimate of -1 would indicate that a one percent increase in waiting times breaches would lead to a one percent reduction in referral volumes. An elasticity estimate of between -1 and 0 would indicate that GP demand is inelastic with respect to waiting times, as a one percent increase in waiting times breaches would lead to a reduction in referrals of less than one percent.

Robustness checks.

Testing the assumption that GP practices are price-takers. A key assumption in our analysis is that a single GP practice cannot influence local hospital waiting times (i.e. that GP practices are ‘price-takers’ in terms of the local waiting times they experience). To confirm the validity of this assumption we created an index which measures the degree to which this assumption is held, which we call the ‘concentration of referrals index’.

This index measures the concentration of referrals going from a practice to any single hospital, weighted by the concentration of referrals that a hospital receives from that practice. The proportion of a hospital’s referrals that come from a single practice will reflect the degree to which a practice is a price-taker for that particular hospital, as a practice is less likely to influence the hospital’s waiting times if it represents a lower proportion of its activity. However, as we are concerned with whether the practice is a price taker for the overall local waiting time it experiences, we also need to consider whether the referrals that a particular practice makes to a particular hospital are also a large proportion of the practice’s total referral volumes.

For each practice, in each year, we calculated the concentration of referrals index as outlined in Eq 6 using data from HES. For each practice-year, we multiplied the share of a hospital’s referrals which come from that practice in that year with the share of a practice’s referrals going to that hospital. We repeated this for each hospital and created the concentration of referrals index as the sum of these values in each practice-year: Eq 6

The index will equal 1 if a practice sends all of its activity to a hospital where it is the only referrer in that year. More generally, the higher the concentration of referrals index, the more influence the practice has over its local hospital waiting times in that year.

We then examined the average concentration of referrals index for each practice across all years in our sample, to evaluate whether the price-taker assumption holds for each practice in general.

Eq 7

Further sensitivity checks on the GP demand analysis. We repeated estimation of Eqs 3–5 using a Poisson regression on the untransformed data, to allow for the count nature of the outcome variable. For the Poisson regressions, we used the GP practice’s registered list size as the exposure variable.

We then examined whether the effect we find persists when we modelled the outcome as a rate. To do this we repeated the estimation of Eqs 3–5, where the outcome variable is the referral rate (urgent referrals as a proportion of the practice population), and population size is not included as a covariate.

We also repeated the main analysis (Eqs 3–7) on a balanced panel, removing practices which do not appear in all six years of data.

Supplementary outcomes analysis.

Finally, we conducted some further analysis to examine the validity of the local waiting times measure. Our concern over the local waiting times measure is twofold: (1) we used waiting time breaches as a proxy for actual waiting times, and (2) our calculation which used weighted values of local hospital waiting times may not accurately reflect average waiting times for the specific patients of each practice.

To check that our measure of local waiting times was valid, we also examined the association between our measure of local hospital waiting times and patient outcomes. If there was an association between waiting times and practice-level patient outcomes, this provides reassurance that our measure of local hospital waiting times is valid.

Specifically, we hypothesise that longer waiting times will result in an increase in diagnoses via emergency presentation. Emergency presentations of cancer tend to be later stage and result in worse outcomes [35]. We therefore proxied patient outcomes by comparing the number of diagnoses via urgent referral (the appropriate pathway) and diagnoses via emergency presentation. These are two of eight possible routes to diagnosis (see S1 Appendix). We carried out fixed effects analysis following Eq 5, but using these two variables (urgent referral diagnoses and diagnoses via emergency presentation) as outcomes. To control for the number of diagnoses via any route, we included the inverse hyperbolic sine transformed count of total cancer diagnoses at the practice level as a covariate.

Testing the assumption that GP practices are price takers.

Table 5 provides descriptive statistics on the concentration of referrals indices for each practice-year (N*T values), and for each practice on average across all years (N average values). There were a few cases where low numbers of referrals between practice and hospital in a given year create a high concentration of referrals index in a practice-year. For example, the maximum value of 1 (row 1, Table 4) was due to a single practice making only one referral to a hospital which only received one urgent referral in that year. Averaging over all years for each practice (row 2, Table 4) the mean concentration of referrals index across all practices was 0.014. We can therefore conclude that practices did not consistently dominate a hospital’s referrals and were not dominated by certain local hospitals over our period of interest.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Summary statistics on the indices of concentration of referrals between practices and hospitals.

https://doi.org/10.1371/journal.pone.0294061.t005

Further sensitivity checks on the GP demand analysis.

Sensitivity analyses of the GP demand analysis using Poisson regressions, and where the outcome variable is modelled as the referral rate as a proportion of the practice population (S2 and S3 Appendices) supported the main analysis.

Both sets of results exhibited the same pattern as the main analysis. There was a statistically significant positive association between local hospital waiting times and GP referrals for pooled OLS and the between estimator, and a smaller and statistically insignificant positive association for the fixed effects analyses.

Repeating the analysis using a balanced panel supports the main results in terms of the sign, significance and magnitude of the associations (S4 Appendix). Unadjusted estimates are presented in S5 Appendix.