Cost-Effectiveness of Health Coaching and Financial Incentives to Promote Physical Activity after Total Knee Replacement

Abstract

Objective:

We evaluated the cost-effectiveness of telephonic health coaching and financial incentives (THC+FI) to promote physical activity in total knee replacement recipients.

Design:

We used the Osteoarthritis Policy Model, a computer simulation of knee osteoarthritis, to evaluate the cost-effectiveness of THC+FI compared to usual care. We derived transition probabilities, utilities, and costs from trial data. We conducted lifetime analyses from the healthcare perspective and discounted all cost-effectiveness outcomes by 3% annually. The primary outcome was the incremental cost-effectiveness ratio (ICER), defined as the ratio of the differences in costs and quality-adjusted life years (QALYs) between strategies. We considered ICERs <$100,000/QALY to be cost-effective. We conducted one-way sensitivity analyses that varied parameters across their 95% confidence intervals and limited the efficacy of THC+FI to one year or to nine months. We also conducted a probabilistic sensitivity analysis (PSA), simultaneously varying cost, utilities, and transition probabilities.

Results:

THC+FI had an ICER of $57,200/QALY in the base case and an ICER below $100,000/QALY in most deterministic sensitivity analyses. THC+FI cost-effectiveness depended on assumptions about long-term efficacy; when efficacy was limited to one year or to nine months, the ICER was $93,300/QALY or $121,800/QALY, respectively. In the PSA, THC+FI had an ICER below $100,000/QALY in 70% of iterations.

Conclusions:

Based on currently available information, THC+FI might be a cost-effective alternative to usual care. However, the uncertainty surrounding this choice is considerable, and further research to reduce this uncertainty may be economically justified.

INTRODUCTION

Knee osteoarthritis (OA) patients are highly sedentary and inactive¹ and could benefit substantially from the increases in quality of life associated with physical activity^2–7. Total knee replacement (TKR) is a widely-used surgery that substantially decreases pain and improves function in knee OA patients⁸. Half of knee OA patients in the United States eventually undergo TKR⁹. However, TKR alone does not lead to increases in physical activity in the majority of recipients^10–13.

The Study of Physical Activity Rewards after Knee Surgery (SPARKS) was a randomized controlled trial that evaluated two interventions to promote physical activity among TKR recipients: telephonic health coaching (THC) and financial incentives (FI)¹⁴. The combination of THC+FI was the most successful at increasing activity: six months after TKR, daily step count increased from baseline by 1808 steps in the THC+FI arm compared to 680 steps in the control arm, 274 steps in the THC arm, and 826 steps in the FI arm. Weekly minutes of moderate-to-vigorous physical activity (MVPA) increased by 39 minutes in the THC+FI arm compared to 14 minutes in the control and THC arms and 16 minutes in the FI arm¹⁴.

The SPARKS trial showed that THC+FI is effective, but the long-term value of THC+FI is unknown. Expenses and time costs preclude randomized controlled trials from effectively evaluating long-term value. In contrast, model-based analyses can project trial results beyond the duration of the trial and determine the conditions under which an intervention would be cost-effective. We conducted a model-based analysis to examine the cost-effectiveness of THC+FI by using assumptions regarding efficacy duration to project the results of the trial over participants’ lifetimes. In addition, we evaluated the value of future research on THC+FI outcomes. Together, the cost-effectiveness and value of information analyses will inform decision-making about current treatment options post-TKR and the prioritization of outcomes in future studies of THC+FI.

METHOD

Analytic Overview

This cost-effectiveness analysis is based on the SPARKS trial, which is reported in detail elsewhere¹⁴. SPARKS was pre-registered at https://ClinicalTrials.gov (NCT01970631).

SPARKS included four arms: attention control, THC, FI, and THC+FI. THC consisted of calls from trained health coaches, and FI consisted of incentives for completing activity logs and for increasing activity. Intervention details are reported with the main trial results¹⁴. As the THC and FI alone interventions were no more effective than the control but carried additional costs, we did not include them in these analyses. Physical activity was measured at baseline, six months (the end of the main intervention), and nine months post-TKR.

We used the Osteoarthritis Policy (OAPol) Model, a validated microsimulation model of knee OA natural history and treatment^{9, 15–20}, to project clinical, economic, and quality-of-life effects under a range of different treatment effectiveness and cost assumptions for subjects receiving either usual care (attention control arm) or THC+FI. We measured cost-effectiveness with the incremental cost-effectiveness ratio (ICER), the ratio of difference in cost and difference in quality-adjusted life expectancy (QALE) between treatment strategies. We labeled a strategy as cost-effective if it conferred one quality-adjusted life year (QALY) gained for less than $100,000²¹. We discounted costs (2016 USD) and QALYs at 3% annually. We conducted the analysis from the healthcare sector perspective, which, per established convention²², includes direct medical costs to payers and patients. Appendix Table 1 lists the health and costs effects considered in the analyses.

To inform priority setting for future research, we also conducted a value of information (VOI) analysis by finding the expected value of partial perfect information (EVPPI) for key model parameters²³. This analysis provides an estimate of the monetary value of knowing the exact value for a given model parameter.

Model Overview

The OAPol model is a state-transition, Monte Carlo microsimulation model of knee OA natural history and treatment^{9, 15–20}. The model generates a cohort of knee OA patients one at a time based on demographic and clinical characteristics, including age, BMI, comorbidities, knee OA pain, and knee OA structural severity. Each year, subjects in the model transition between health states that are associated with a cost and a quality-of-life (QoL) utility, a value from 0-1 that reflects health-related QoL, with 1 representing perfect health. At the end of a subject’s life, the model aggregates the time spent in each state and the associated health-related QoL effects and costs.

OAPol includes three activity states: active, active for half the year, and inactive (Figure 1). In this analysis, all subjects underwent TKR in the first year and received either no physical activity intervention (usual care) or THC+FI. That first year, subjects could be inactive, active for half the year, active for the full year, or could die. In subsequent years, subjects in the active state could remain active, transition to active for half the year, or transition to inactive. Subjects in the active for half the year state became inactive in the following year, and inactive subjects remained inactive. We assumed that inactive subjects remained inactive because studies of TKR and physical activity have generally found that TKR alone does not lead to greater physical activity^10–13.

Figure 1. Structure of Physical Activity in the OAPol Model.

Possible transitions between health states are indicated by arrows. Subjects can only become active in the first year when all subjects receive TKR and either no intervention (usual care) or THC+FI. In subsequent years, they may remain active or transition back to the inactive state, either by becoming fully inactive or by becoming active for half the year. If a subject becomes inactive after the first year, they cannot become active again. Abbreviations: THC+FI = telephonic health coaching + financial incentives; TKR = total knee replacement.

Model Input Parameters

Model Cohort

The model cohort was based on SPARKS participants’ characteristics. Mean age was 65 (standard deviation (SD) 8). Fifty-seven percent were female, mean BMI was 31.1 kg/m² (SD 5.8), and Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) Pain²⁴ was 53 (SD 18) (WOMAC Pain is out of 100, 100 = highest pain).

Activity Thresholds

Transition probabilities among activity states in OAPol were derived from SPARKS participants. To be classified as active, participants had to meet a threshold of 60 minutes per week of activity at an intensity of at least 66 steps/minute. This activity threshold was selected as a walking speed of 2.5km/hr (41m/min) in a population of older adults was found to be equivalent to 3 metabolic equivalents of task²⁵, the threshold for moderate physical activity²⁶. Assuming a step length of 0.62 meters²⁷, this walking speed is equivalent to 66 steps/minute.

Sixty minutes per week is lower than the 150 minutes of moderate activity recommended by the 2008 Physical Activity Guidelines for Americans²⁸. We opted for the lower threshold because the recommended guidelines may be too restrictive for older adults. Moreover, QoL benefits occur at lower levels of activity⁷. We varied the threshold to qualify as active in sensitivity analyses reported in the appendix.

In the base case, we assumed that all model subjects were inactive prior to TKR. We made this assumption because if THC+FI were to be included in clinical care, it would likely not be offered to patients who were already active. The SPARKS trial did not include high activity as an exclusion criterion because we had incorrectly anticipated that the pre-TKR population would be almost entirely inactive. To test the effect of the model assumption, we conducted a sensitivity analysis for a cohort with the baseline activity of SPARKS participants (28% active).

Transition Probabilities

As the OAPol model is run on an annual cycle, we extrapolated annual transition probabilities between the three activity states from probabilities derived from the SPARKS trial. We used two probabilities from the SPARKS trial: 1) the probability of becoming active at six months post-TKR (end of main intervention) and 2) the probability that active subjects at six months would become inactive at nine months. Base case annual transition probabilities are shown in Table 1 and derivation details for the conversion to annual transition probabilities are in the appendix.

Table 1.

Model Input Parameters

Parameter				Estimate		Source
Age: Mean (SD)				65 (8)
Percentage Female				57%
BMI: Mean (SD)				31.1 (5.8)		Losina et al., 201714
WOMAC Pain: Mean (SD)a,b
Year 1				53 (18)
Increase in Years 2+c				2 (10)
Quality-of-Life Utilities
		WOMAC Pain (0-100)a				Osteoarthritis Initiative30
Age	0	1-15	16-40	41-70	71-100
0 Comorbidities (Nonobese/Obesed)
25-44	0.865/0.845	0.840/0.820	0.781/0.761	0.699/0.679	0.609/0.589	Brazier et al., 200431
45-54	0.841/0.830	0.816/0.806	0.780/0.769	0.714/0.703	0.656/0.645
55-64	0.847/0.836	0.822/0.812	0.786/0.775	0.720/0.709	0.662/0.651
65-74	0.871/0.860	0.846/0.835	0.810/0.799	0.744/0.733	0.685/0.675
75+	0.854/0.843	0.829/0.818	0.793/0.782	0.727/0.716	0.669/0.658
1 Comorbidity (Nonobese/Obesed)
25-44	0.845/0.825	0.820/0.800	0.761/0.741	0.679/0.659	0.589/0.569
45-54	0.818/0.807	0.791/0.780	0.755/0.744	0.679/0.668	0.645/0.634
55-64	0.824/0.813	0.797/0.786	0.761/0.750	0.685/0.674	0.651/0.640
65-74	0.848/0.837	0.821/0.810	0.785/0.774	0.708/0.698	0.674/0.664
75+	0.831/0.820	0.804/0.793	0.768/0.757	0.692/0.681	0.658/0.647
2+ Comorbidities (Nonobese/Obesed)
25-44	0.825/0.805	0.800/0.780	0.741/0.721	0.659/0.639	0.569/0.549
45-54	0.806/0.795	0.794/0.783	0.732/0.721	0.635/0.624	0.500/0.489
55-64	0.812/0.801	0.800/0.789	0.738/0.727	0.641/0.630	0.506/0.495
65-74	0.836/0.825	0.824/0.813	0.762/0.751	0.665/0.654	0.530/0.519
75+	0.819/0.808	0.807/0.796	0.745/0.734	0.648/0.637	0.513/0.502
Increase in QoL Utility if Active						Unpublished SPARKS data
Active for Half Year			1.7%
Active for Full Year			3.3%
Annual Non-OA Medical Cost
Age	0-1 Comorbidity		2-3 Comorbidities	4+ Comorbidities		See footnote for sourcee
25-34	$1,290		$6,813	$13,110
25-44	$1,800		$7,323	$13,110
45-49	$2,435		$7,462	$13,110
50-54	$2,435		$7,462	$13,110
55-59	$3,191		$8,050	$13,484
60-64	$3,888		$8,747	$14,181
65-69	$4,179		$9,075	$14,179
70-74	$4,867		$9,764	$14,867
75-79	$5,690		$10,586	$15,690
80+	$7,478		$12,374	$17,478
Transition Probabilities
Year	Starting State		Transitioning to	Usual Care		THC+FI
First Year	Inactive		Active for the Full Year	5%		19%
			Active for Half Year	16%		19%
			Inactive	79%		62%
Subsequent Years	Active for Full Year		Active for the Full Year	6%		24%
			Active for Half Year	19%		25%
			Inactive	75%		51%
Subsequent Years	Inactive or Active for Half Year		Active for the Full Year	0%		0%
			Active for Half Year	0%		0%
			Inactive	100%		100%

Abbreviations: SD = standard deviation; BMI = body mass index; WOMAC = Western Ontario and McMaster Universities Osteoarthritis Index; THC+FI = Telephonic Health Coaching + Financial Incentives

^aWOMAC pain is measured out of 100, 100 is the worst pain

^bThis is the average pain trajectory without TKR; pain decreases for subjects who undergo TKR. See technical appendix of Katz et al., 2016 for details on TKR intervention.¹⁹

^cAssumption

^dObese is defined as a BMI ≥30kg/m²

^ePope et al. (2004),⁴⁵ NHANES 2009-2010,⁴⁶ MCBS 2009,⁴⁷ US Census Bureau 2014 Population Estimates⁴⁸; converted to 2016 dollars using PHC³² and PCE³³. The methodology for these derivations is described in a previous publication on lifetime costs of knee OA.⁹ Since that publication, our cost derivations have been updated to include the most recent releases of data sources and the inflation methodology has been updated.

Quality-of-Life Utility

We mapped SPARKS participants’ responses to the EQ-5D to QoL utilities²⁹ and calculated the effect of physical activity on QoL utility in a regression model adjusting for age, sex, obesity, pain, and comorbidities. With the base case activity threshold, active subjects had an average utility of 0.809 (standard error (SE) 0.009) while inactive subjects had an average utility of 0.783 (SE 0.007) (p=0.013). This corresponds to an increase of 3.3% for active participants. We applied the percentage increase to base utilities that were derived from the Osteoarthritis Initiative (OAI)^{30, 31} (Table 1). We applied a 1.7% increase for subjects who were active for half of a year and a 3.3% increase for subjects who were active for an entire year. Inactive subjects had no change in utility.

We implemented a nine-month efficacy limit in the annual model by creating a new health state: active for nine months. In the first year, all subjects who would have been active for either nine months or the full year instead were considered active for only nine months. They received an increase of 2.48% to their QoL utility, which is 75% of the increase associated with a full year of activity. In subsequent years, all subjects were inactive and had no increase in utility.

Cost

The total cost of THC+FI was $287 per person. This cost comprised the Fitbit Zip ($51, discounted from the retail price of $59.95), average incentive payments to THC+FI participants ($136), and the average cost of health coaching for THC+FI participants ($101). The Fitbit Zip price and the average incentive payments were calculated from trial payment data. Health coaching costs were calculated from logs that health coaches completed during the trial. The coaches logged an average of 4 hours and 2 minutes (SD = 1 hour 45 minutes) with each THC+FI participant. To calculate the per-person cost of coaching, we multiplied this by the hourly salary for health coaches (including an additional 30% for fringe benefits) which was $25/hour. This amounts to a per-person health coaching cost of $101. We assumed that each participant would incur the cost of one Fitbit, as the trial duration was nine months, which is within the one-year Fitbit warranty. Other costs in OAPol included the cost of TKR, revision TKR if necessary, and non-OA medical costs. Derivations for these costs can be found in a previous publication⁹.

Costs were adjusted for inflation to 2016 USD per the guidelines of the Second Panel on Cost-Effectiveness in Health and Medicine²². All medical costs were inflated from their year of origin to 2015 USD (the most recent year available) using the personal health care (PHC) expenditure deflator³². They were then inflated to 2016 USD with the personal consumption expenditure (PCE) price index³³. Non-medical costs were inflated to 2016 USD using the Consumer Price Index for All Urban Consumers (CPI-U)³⁴. As the trial ran from November 2013 to January 2016, and the incentive amounts were not changed during these years, we assumed that the costs of the incentives were in 2014 USD and inflated to 2016 USD with the CPI-U. As the retail price of the Fitbit Zip ($59.95) has not changed since 2014, we did not inflate the discounted price of $51 paid during the trial. Health coaching costs were derived using the 2016 salaries for health coaches, so also were not inflated.

Sensitivity Analyses

In accordance with established guidelines^{22, 35}, we conducted both deterministic one-way sensitivity analyses and a probabilistic sensitivity analysis (PSA). The one-way sensitivity analyses allowed us to test key assumptions in the model as well as the effect of varying each parameter to its lower and upper bounds. The PSA considered parameter uncertainty simultaneously for all parameters and allowed us to quantify the uncertainty around the decision of whether to pay for THC+FI.

Deterministic Sensitivity Analyses

The one-way sensitivity analyses varied three model parameters: THC+FI cost, the probability of becoming active post-TKR, and the probability of active subjects becoming inactive. The parameters were tested at the low and high ends of their 95% confidence intervals (CI) as is recommended by uncertainty analysis guidelines³⁵ (Appendix Tables 2 & 3).

We also varied a model assumption. In the base case, we assumed that THC+FI efficacy was governed solely by the transition probabilities derived from the SPARKS trial. In sensitivity analyses, we limited the duration of efficacy such that subjects could only be active for one year or for nine months. These durations were selected because the last activity measurement in SPARKS was at nine months.

Impact of Baseline Activity Levels

In an additional sensitivity analysis, we evaluated the cost-effectiveness of THC+FI using a cohort with the distribution of active and inactive participants from SPARKS. We assumed that both the active and inactive subjects would participate in THC+FI, but that only inactive participants who became active would receive an increase in QoL utility. Active participants who remained active had no change in QoL utility. This sensitivity analysis reflects the conditions of the SPARKS trial, while the base case analysis reflects the conditions under which we expect the THC+FI intervention would be implemented.

Probabilistic Sensitivity Analysis

We conducted a PSA with 1000 iterations. We varied the model parameters that define the cohort characteristics as well as the efficacy and cost of usual care and THC+FI. Due to computation limitations, we did not vary the other model parameters (e.g., TKR cost), which are unlikely to affect the cost-effectiveness of THC+FI. In each iteration, a value for each parameter was drawn from its respective distribution, which incorporates the variability in the estimates from the SPARKS trial (Table 2). All other parameters were set to base case values. We report the results as a cost-effectiveness acceptability curve, showing the probability that each treatment strategy is the cost-effective option at different levels of willingness-to-pay.

Table 2.

Probabilistic Sensitivity Analysis Distributions

	Usual Care	THC+FI
Age*	Normal (65, 0.56)a	Normal (65, 0.56)
BMI*	Normal (31.1, 0.41)	Normal (31.1, 0.41)
Percentage Female*	Beta (115, 87)b	Beta (115, 87)
WOMAC Pain*	Normal (53, 1.27)	Normal (53, 1.27)
Quality-of-Life Utility
Active	Normal (0.809, 0.009)	Normal (0.809, 0.009)
Inactive	Normal (0.783, 0.007)	Normal (0.783, 0.007)
Probability of Becoming Active	Beta (6, 23)	Beta (11, 18)
Probability of an Active Subject Becoming Inactive at Nine Months	Beta (3, 3)	Beta (3, 7)
FI Cost	$0	Normal ($135, $10)
THC Cost	$0	Normal ($101, $6)

Abbreviations: FI = Financial Incentives; THC = Telephonic Health Coaching; THC+FI = Telephonic Health Coaching + Financial Incentives

^*Due to computational limitations, these parameters were not varied in the EVPPI analysis

^aNormal distributions are represented as Normal (Mean, Standard Error)

^bBeta distributions are represented as Beta (α,β) where α is the number of events and β is the number of non-events

In the PSA, we assumed that a change in the probability of becoming active post-TKR would change the cost of incentive payments; for example, if more subjects become active, more incentives would be paid out. We assumed that the percentage change from the base case probability of becoming active post-TKR would result in an equivalent percentage change in the amount spent on incentives.

Value of Information Analysis

VOI analysis is a method of quantifying the value of obtaining additional information to inform a decision²³. Future research is valuable, as it reduces the uncertainty around a decision. However, future research comes at a cost; it requires resources, time, and funding. By explicitly quantifying the value of additional information, VOI analysis allows the decision-maker to weigh the value of acting optimally using current information against the value of delaying the decision in anticipation of better information at some later date. We calculated the expected value of partial perfect information (EVPPI), which is an estimate of the value of eliminating the uncertainty around a given model parameter. EVPPI identifies the parameters for which additional information would be most valuable, and this information can inform the design of subsequent studies. We calculated EVPPI as the difference between the expected net monetary benefit [(ΔQALYs × Willingness-to-Pay (WTP) Threshold) − ΔCost] with perfect information about the parameter and the expected net monetary benefit with current information. Details of EVPPI calculations are included in other publications²³.

Due to computational limitations, our EVPPI analysis included variation in four model parameters: 1) THC+FI cost, 2) the percentage increase in utility for active participants, 3) the probability of becoming active post-TKR, and 4) the probability of active subjects becoming inactive (Table 2). We estimated the per-person EVPPI for the latter three parameters, which are the parameters that are likely to be investigated in future trials. EVPPI analyses consisted of 200 values drawn from the distribution for the parameter of interest (outer loops). Each outer loop was evaluated with 10,000 iterations of values drawn from the distributions for all other parameters (inner loops). Due to the computationally intense nature of EVPPI analyses, these analyses were implemented with a Markov cohort model built in TreeAge Pro 2017³⁶. The cohort model uses the same structure as the OAPol model but without simulating individual subjects.

RESULTS

Base Case Analysis

Base case results, including the percentage of subjects who were active each year, cost, QALE, and ICERs, are presented in Table 3. Peak activity occurred six months post-TKR and by the end of the third year, nearly all subjects returned to being inactive. THC+FI was cost-effective (ICER <$ 100,000/QALY) with an ICER of $57,200/QALY.

Table 3.

Base Case Results

Percentage of Model Subjects Active for Full Year
	Year 1	Year 2	Year 3
Usual Care	5.18%	0.32%	0.02%
THC+FI	18.60%	4.47%	1.08%
Cost-Effectiveness Results
	Cost	QALE	ICER
Usual Care	$140,700	9.783	--
THC+FI	$141,000	9.788	$57,200a

Abbreviations: THC+FI = Telephonic Health Coaching + Financial Incentives

^aCalculations may not match exactly due to rounding

Sensitivity Analyses

THC+FI was cost-effective at a threshold of $100,000/QALY across almost all plausible parameter values (Figure 2). THC+FI was most cost-effective when evaluated using the low end of the 95% confidence interval for the probability of active subjects becoming inactive (ICER = $4,700/QALY). Our assumptions about long-term efficacy were the only parameter that produced an ICER greater than $100,000/QALY. THC+FI remained cost effective when efficacy was limited to one year (ICER = $93,300/QALY). However, when efficacy was further limited to the nine-month trial endpoint, THC+FI was no longer cost-effective, with an ICER of $121,800/QALY. When the modeled cohort had the baseline distribution of active and inactive participants in SPARKS (28% active, 72% inactive), THC+FI remained cost-effective, but the ICER increased to $74,200/QALY. In the PSA, THC+FI was the cost-effective treatment option in 70% of iterations (Figure 3).

Figure 2. One-way Sensitivity Analysis of THC+FI Compared to Usual Care.

This figure illustrates the incremental cost-effectiveness ratio (ICER) estimated for THC+FI under a variety of conditions. In each analysis, all parameters were held at base case values except for the parameter listed on the vertical axis, which was varied according to the values listed. The left end of each bar reports the ICER when the parameter of interest is set to its most favorable value. The right end of each bar reports the ICER when the parameter is set to its least favorable value. The black vertical line shows the base case ICER. Abbreviations: THC+FI = telephonic health coaching + financial incentives; ICER = incremental cost-effectiveness ratio; CI = confidence interval.

Figure 3. Cost-Effectiveness Acceptability Curve.

The curves show the percentage of simulations (out of 1000) in which a given strategy was the cost-effective treatment option at a given willingness-to-pay threshold. Each of the 1000 simulations independently sampled model input parameters from specified distributions for cohort characteristics, quality-of-life utility, transition probabilities, and THC+FI cost (Table 2). Abbreviations: THC+FI = telephonic health coaching + financial incentives.

Value of Information Analysis

At a threshold of $100,000/QALY, the per-person EVPPI was $6.08 for the probability of becoming active post-TKR, $15.10 for the utility increase associated with being active, and $35.17 for the probability of active subjects becoming inactive. Figure 4 shows EVPPI at varying levels of society’s willingness to pay for an additional QALY.

Figure 4. Expected Value of Partial Perfect Information.

The per-person EVPPI for three key model parameters is shown at different levels of society’s willingness to pay for an additional QALY. Abbreviations: EVPPI = expected value of partial perfect information; THC+FI = telephonic health coaching + financial incentives; QALY = quality-adjusted life year.

DISCUSSION

We used a computer simulation model of knee OA to evaluate the cost-effectiveness of using THC+FI to promote physical activity after knee replacement. Our analyses suggest that THC+FI is potentially cost-effective (ICER <$100,000/QALY) with a base case ICER of $57,200/QALY. However, this result is sensitive to assumptions about long-term efficacy, and further research to reduce uncertainty around this parameter is warranted. THC+FI was cost-effective in all deterministic sensitivity analyses unless we assumed that benefits from THC+FI did not extend past the final trial timepoint. In the PSA, THC+FI was cost-effective in 70% of iterations.

Cost-effectiveness analyses of programs to increase physical activity have found that the programs are generally cost-effective, though with substantial heterogeneity in ICERs^37–41. An analysis of seven interventions to promote activity found that all were cost-effective, with ICERs ranging from $14,000/QALY to $69,000/QALY⁴². A 2013 study found that financial incentives to promote workplace physical activity were also highly cost-effective⁴³. Our results fit with these findings and suggest that THC+FI has an ICER within the range of other physical activity programs. Our VOI results are also consistent with a VOI analysis for another physical activity intervention. Singh and colleagues found that for a brief pedometer intervention, intervention effects (i.e., changes in activity) were the parameter group with the highest EVPPI, and they concluded that further research on the efficacy of the intervention is likely warranted⁴⁰.

This analysis demonstrates how VOI can guide the planning of subsequent studies. Conducting an EVPPI assessment on a given parameter allows us to estimate what society would be willing to pay to eliminate the uncertainty surrounding that parameter. By conducting EVPPI assessments on a range of different parameters, we can identify which uncertainties are exerting the greatest impact on our decision and can use this information to inform where we might focus our future research efforts. Our results suggest that subsequent trials may be warranted and should focus on better estimating the rate at which active participants revert to inactivity. Extending follow-up beyond three months after the THC+FI program ends and increasing the sample size will reduce the uncertainty around this parameter.

We note several limitations. First, assumptions were necessary to project results beyond the duration of the trial. We assumed that active subjects would become inactive at a constant rate after the trial ended. However, while THC and FI interventions tapered between 6 and 9 months, they did not end completely: there were monthly coaching calls and the possibility of $50 in FI at the 9-month time point. Thus, the true rate of becoming inactive after THC+FI end may be higher than our estimate. To address this, we included a sensitivity analysis that evaluated the cost-effectiveness of THC+FI when all subjects became inactive after nine months. Second, we assumed that in the base case, the THC+FI program would include only inactive TKR recipients. We tested this assumption in sensitivity analyses, and THC+FI remained cost-effective when active participants were allowed to participate in proportions similar to those observed in SPARKS. Third, due to computations limitations, the PSA and EVPPI analyses included variations in a select number of parameters.

Other limitations arise from our assumptions regarding QoL utilities and costs. We estimated the benefit from becoming active as a percentage increase in QoL utility. As utilities have interval scale properties²², applying an absolute increase may have been more appropriate. However, the percentage increase results in a more conservative estimate of the benefit from activity. With the percentage increase, the average QoL increase for model subjects who were active for a full year is 0.023 QALYs. Had we applied an absolute increase, the increase would have been 0.026 QALYs. As sample size limitations prevented us from using baseline utilities from SPARKS, the utilities were derived from the Osteoarthritis Initiative, a large longitudinal OA cohort.

We acknowledge that the 0.005 difference in QALE between usual care and THC+FI is small. This is consistent with analyses of other activity interventions^{40, 44}. In addition, as we did not incorporate reductions in comorbidities and mortality associated with increased physical activity, we may have underestimated the health benefits received by active participants. We did not include time costs for administering the FI program. However, as FI reimbursement uses an automated computer program, the per-person cost of this process is likely negligible. We did not include analysis from the societal perspective, as we had limited data on indirect costs.

Finally, our conclusions are limited by the uncertainty around key parameters of the THC+FI intervention. Our transition probabilities were derived from small sample sizes. For example, the probability of active subjects becoming inactive is derived from data on six subjects in the usual care arm and 10 subjects in the THC+FI arm. The uncertainty around these parameters is reflected in the PSA, where THC+FI was cost-effective in only 70% of iterations.

Physical activity is associated with meaningful increases in quality of life; consequently, interventions to promote physical activity after TKR have the potential to improve the value of TKR. If a decision needed to be made today, the evidence currently points to THC+FI as a cost-effective alternative to usual care. However, the uncertainties surrounding this choice are great enough that, if the decision can be deferred, it might be preferable to acquire better data on the sustainability of THC+FI’s impact on participant activity levels.

Appendix Part I. Model Inputs

Impact Inventory

We conducted the analysis from the healthcare sector perspective. The components included in the analysis are shown in Appendix Table 1.

Appendix Table 1.

Impact Inventory

Formal Health Care Sector
Health	Health Outcomes (effects)
	Longevity Effects	√a
	HRQoL Effects	√
	Adverse events from treatment	√
	Medical Costs
	Paid for by third party payers	√
	Paid for by patients out of pocket	√
	Future related medical costs	√
	Future unrelated medical costs	√
	Treatment time costs	NA
	Transportation costs	NA
Non-Health Care Sector Costs
Productivity	Labor market earnings lost	NA
	Cost of care-giving	NA

Abbreviations: HRQoL = Health-Related Quality of Life;

^athe overall model includes longevity, but physical activity does not affect longevity

Transition Probability Derivations

Appendix Table 2 shows the two SPARKS results that informed the model transition probabilities. The first is the probability that an inactive subject at baseline becomes active at six months. The second is the probability that an active subject at six months becomes inactive at nine months.

Appendix Table 2.

SPARKS Results Used to Derive Annual Transition Probabilities

	Probability of being inactive at baseline and active at six months post-TKR (95% CI)	Probability of being active at six months and inactive at nine months post-TKR (95% CI)
Usual Care	21% (6% - 35%)	50% (10% - 90%)
THC+FI	38% (20% - 56%)	30% (2% - 58%)

Abbreviations: TKR = Total Knee Replacement; THC+FI = Telephonic Health Coaching + Financial Incentives

We calculated the model transition probabilities as follows. Usual care is used for example calculations:

We assumed that the trial probability of becoming inactive between six and nine months is constant until all subjects become inactive. We calculated the six- and twelve-month probabilities of becoming inactive by converting the trial probability of becoming inactive from six to nine months to a rate, and then back to a six- or twelve-month probability. For usual care, the six-month probability of becoming inactive was 75% and the twelve-month probability of becoming inactive was 94%.

First Year

1. The probability of being inactive is equal to 1 – the probability of becoming active at six months, or 1 – 21% = 79%.

2. The probability of being active for only half the year is equal to the probability of becoming active at six months multiplied by the six-month probability of becoming inactive, or 21% × 75% = 16%.

3. The probability of being active for the entire year is equal to the probability of becoming active at six months minus the probability of being active for half the year. For usual care, this is 21% - 16% = 5%.

Subsequent Years

For subjects starting in the active state:

1. The probability of becoming inactive for the entire year is equal to the 6-month probability of becoming inactive. In other words, model subjects who became inactive within 6 months were considered inactive for the entire year. In usual care, this is 75% of subjects.

2. The probability of being active for half the year is equal to the 12-month probability of failure minus the six-month probability of failure. This is the probability of being inactive at 12 months but being active at six months. For usual care subjects, this equals 94% - 75% = 19%.

3. The probability of being active for the entire year is equal to 1 minus the 12-month probability of failure, or 1 – 94% = 6% for usual care subjects.

All subjects starting in the half year active state transition to the inactive state in the next year, and all subjects in the inactive state remain inactive.

Table 1 in the main text shows the base case transition probabilities. Appendix Table 3 shows the transition probabilities when the probability of becoming active at six months was varied from the low to high ends of the 95% confidence interval (CI) and when the probability of an active subject at six months becoming inactive at nine months was varied from the low to high ends of the 95% CI. These values were used for the one-way sensitivity analysis.

Appendix Table 3.

95% Confidence Intervals for Transition Probabilities

Year	Starting State	Transitioning to	Usual Care	THC+FI
Varying Probability of Becoming Active at Six Months
First Year	Inactive	Active for the Full Year	1% – 9%	10% – 27%
		Active for Half Year	4% – 27%	10% – 28%
		Inactive	94% – 65%	80% – 44%
Subsequent Years	Active for Full Year	Active for the Full Year	6%	24%
		Active for Half Year	19%	25%
		Inactive	75%	51%
Subsequent Years	Inactive or Active for Half Year	Active for the Full Year	0%	0%
		Active for Half Year	0%	0%
		Inactive	100%	100%
Varying Probability of Active Subjects at Six Months Becoming Inactive at Nine Months
First Year	Inactive	Active for the Full Year	17% – 0%	37% – 7%
		Active for Half Year	4% – 20%	1% – 31%
		Inactive	79% – 79%	62% – 62%
Subsequent Years	Active for Full Year	Active for the Full Year	66% – 0%	94% – 3%
		Active for Half Year	15% – 1%	3% – 14%
		Inactive	19% – 99%	3% – 83%
Subsequent Years	Inactive or Active for Half Year	Active for the Full Year	0%	0%
		Active for Half Year	0%	0%
		Inactive	100%	100%

Abbreviations: THC+FI = Telephonic Health Coaching + Financial Incentives

Appendix Part II. Sensitivity Analysis Varying Activity Threshold

Methods

We evaluated the cost-effectiveness of THC+FI using alternative thresholds for subjects to qualify as active. We tested a low threshold, 45 minutes of activity per week at an intensity ≥50 steps per minute, and a high threshold, 75 minutes per week at an intensity ≥100 steps per minute.

QoL utility, the probability of becoming active at six months, and the probability that an active subject would become inactive at nine months changed with the new activity thresholds. At the low activity threshold, more subjects became active and they stayed active longer. The probability of becoming active was 29% for usual care and 41% for THC+FI. The probability of an active subject becoming inactive at nine months was 71% for usual care and 22% for THC+FI. However, active subjects received a smaller increase in quality of life (0.8%).

At the high threshold, fewer subjects became active and they reverted to inactivity quickly, but the percent increase in QoL was greater (4.3%). The probability of becoming active was 10% for usual care and 22% for THC+FI. The probability of an active subject becoming inactive at nine months was 50% for usual care and 75% for THC+FI.

Results

With the alternative physical activity thresholds, THC+FI was less cost-effective than the base case analysis. With the high activity threshold, THC+FI had an ICER of $164,400/QALY and with the low activity threshold, THC+FI had an ICER of $96,800/QALY. This is likely because the high activity threshold was more restrictive and did not account for the activity benefits received by participants at lower levels of activity. In contrast, the low activity threshold was low enough that the QoL utility difference between active and inactive participants (0.8%) did not capture benefits that participants received at higher levels of activity.