Long-term effectiveness of intra-articular injections on patient reported symptoms in persons with knee osteoarthritis

Abstract

Objective

We examined the long-term effectiveness of corticosteroid or hyaluronic acid injections in relieving symptoms among persons with knee osteoarthritis (OA).

Methods

Using Osteoarthritis Initiative (OAI) data, a new-user design was applied to identify participants initiating corticosteroid or hyaluronic acid injection use (n=412). Knee symptoms (pain, stiffness, function) were measured using The Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC). We used marginal structural models adjusting for time-varying confounders to estimate the effect of newly initiated injection use compared to non-users on symptoms over two years of follow-up.

Results

Among 412 participants initiating injections, 77.2% used corticosteroid injections and 22.8% used hyaluronic acid injections. Approximately 18.9% had additional injections use after initiation, but switching between injection types was common. Compared to non-users, on average, participants initiating a corticosteroid injection experienced a worsening of pain (yearly worsening: 1.24 points; 95% confidence interval [95% CI]: 0.82 to 1.66), stiffness (yearly worsening: 0.30 points; 95% CI: 0.10 to 0.49), and physical functioning (yearly worsening: 2.62 points; 95% CI: 0.94 to 4.29) after adjusting for potential confounders with marginal structural models. Participants initiating hyaluronic acid injections did not show improvements of WOMAC subscales (pain: 0.50; 95% CI: −0.11 to 1.11, stiffness: −0.07; 95% CI: −0.38 to 0.24, and functioning: 0.49; 95% CI: −1.34 to 2.32).

Conclusion

Although intra-articular injections may support the effectiveness of reducing symptoms in short-term clinical trials, the initiation of corticosteroid or hyaluronic acid injections did not appear to provide sustained symptoms relief over two years of follow-up for persons with knee OA.

Introduction

Despite a large number of studies, the short-term safety and efficacy of intra-articular injections among persons with knee osteoarthritis (OA) remains inconclusive leading to a lack of consensus across clinical guidelines (1–3). Evidence about the extent to which long-term use of intra-articular injections improves patient outcomes is lacking. Recent systematic reviews and meta-analyses suggest that the effect of US-approved viscosupplement injections can last through 26 weeks but there is no similar evidence for corticosteroid injections for persons with knee OA (4–6). Among patients with milder disease, receiving intra-articular sodium hyaluronate appears to slow joint space narrowing compared to placebo (7). For corticosteroid injections, there is no difference between treatment and placebo groups in joint space changes over two years of follow-up (8). Outside of the short-term clinical trials settings, studies documenting the long-term impact of injections on patient-reported outcomes on knee OA are scarce.

In light of the lack of evidence and conflicting practice recommendations from guidelines, it is somewhat surprising that the use of injections is increasing among Medicare beneficiaries newly diagnosed with knee OA (9). The cost of long-term injection could be substantial (i.e. $1700 to $3700 for viscosupplementation treatments) (10). Given the widespread use of injections and the rising costs of these treatments (10–12), understanding whether the short-term symptom relief extends to long-term effectiveness of repeated injections among persons with knee OA is warranted.

The patient population in the Osteoarthritis Initiative (OAI) provides an opportunity to conduct the effectiveness research of injection use among persons with knee OA living in the community since this sample is more reflective of the general population with OA than those typically recruited in clinical trials. However, statistical inferences from observational studies are often subject to observed and/or unobserved confounding factors. In addition, when time-varying confounders (e.g., symptoms) are themselves affected by the previous treatment (e.g., injection) and thus are intermediates on the causal path way from treatment to outcome, using the standard modeling approach may generate biased estimates of the overall treatment effect (13,14). The aim of the present study was to estimate the effect of intra-articular injections use on changes in patient reported symptoms over a two-year period using marginal structural models to carefully adjust for time-varying confounders and intermediary variables. Through this study, we hoped to provide complementary information on patient reported symptoms over longer periods of follow-up time in a more heterogeneous population compared to clinical trials so that patients and their providers would have more information about what to expect from the long-term use of these treatment options.

Materials and Methods

The University of Massachusetts Institutional Review Board considered this study exempt since publicly available data were used.

Data source

We used publicly available data from the OAI (http://oai.epi-ucsf.org/). The OAI was a longitudinal, multi-center, and prospective cohort study enrolling 4,796 adults aged 45 to 79 years at baseline using four study sites (i.e., Baltimore, MD; Columbus, OH; Pittsburgh, PA; and Pawtucket, RI). The aims of the OAI study were to examine the development and progression of knee OA among adults with symptomatic OA in at least 1 knee or at least 1 established risk factor. Participants had annual follow-up assessments for up to 9 years. Detailed information about the OAI protocol has been described elsewhere (15).

Study sample and design

Figure 1 shows the inclusion/exclusion criteria for our study sample. Only participants with radiographically confirmed knee OA in at least 1 knee at baseline (Kellgren – Lawrence grade ≥ 2) were included (n=2,550). To improve validity of the study, we restricted our analysis to “new users” of knee injections (16). As such, participants who had reported injection use at baseline were not eligible (n=97). In addition, participants indicating no injection use but having missing values for more than half of the follow-up visits over the 9 years were also excluded (n=303). From the remaining group, we identified participants with and without initiation of injection use during the follow-up period. Among initiators, we excluded those reporting use of both injection types (concurrent hyaluronic acid and corticosteroid injection users, n=52), those reporting the first initiation at year 9 because we had no follow-up data after the injections (n=47), and those who reported injection in the affected knee after total knee replacement (n=3). To mimic the study design from clinical trials (8,17), we further excluded participants who did not have symptomatic knee OA at baseline (Western Ontario and McMaster Universities Arthritis Index (WOMAC) pain ⩾2) for non-users since they were considered ineligible for injection use. The final analytic sample included 412 participants initiating injection use and 576 non-users. Among those initiating injection use with available follow-up information for at least one year, 94 initiated hyaluronic acid injections and 318 initiated corticosteroid injections.

Figure 1.

Flowchart of study participants.

Use of Index knee

We used an index knee for the analysis based on: 1) radiographic evidence of OA and 2) the presence of symptoms (e.g. pain) in the same knee. If only one knee had radiographically confirmed OA at baseline, then that knee was used as index knee. If participants had radiographically confirmed OA for both knees, then the knee with higher pain scores at baseline measured by (WOMAC pain subscale was used as the index knee. If pain scores for both knees were equal, then index knee was the one with worse K-L grade. If the aforementioned conditions were the same for the participants, we randomly selected one knee as the index knee.

Assessment of injection use

In OAI, injection use was assessed separately for both knees. Participants were first asked “During the past 6 months, have you had any injections in either of your knees for treatment of arthritis?” For those answering “yes”, two separate questions were posed regarding hyaluronic acid or corticosteroid injections use. For hyaluronic acid injections, participants were asked: “During the past 6 months, have you had an injection of hyaluronic acid (Synvisc or Hyalgan) in either of your knees for treatment of your arthritis?” These injections are given as a series of 3 to 5 weekly injections. To assess corticosteroid injection use, participants were asked: “During the past 6 months, have you had an injection of steroids (cortisone, corticosteroids) in either of your knees for treatment of your arthritis?” For participants whose index knees were censored during the follow-up (e.g., due to death, switching injection, and/or having total knee replacement), we used available information from the other knee to recapture the sample (16 out of 412).

Assessment of OA symptoms

Knee symptoms were evaluated annually using the WOMAC scales (Likert version 3.1) including three subscales: pain (5 items), stiffness (2 items), and physical function (17 items) (18). Each item of the subscale ranged from 0 to 4 (0=none and 4=extreme). Responses to items in each subscale were summed to produce the individual summary score ranging from 0-20 for pain, 0-8 for stiffness, and 0-68 for physical function. Higher WOMAC scores indicate worse symptoms/function. The primary outcome was change in each subscale between baseline visit (one year before the injection), index visit, and one year after the index visit.

Covariates

We considered covariates in two groups: time-invariant (e.g., sociodemographic factors measured at the time of enrollment) and time-dependent (e.g., factors measured annually including clinical characteristics of OA, general health status, body mass index (BMI), and use of medications and biological supplements) (19). Multi-joint symptoms were present if participants had frequent symptoms in at least 2 joints other than the knee (20). Knee malalignment including varus or valgus deformity was measured and recorded using a goniometer. History of knee injuries was present if a prior injury limited the participant’s ability to walk for at least 2 days indicating on any previous visit. A history of having knee surgery was present if participants indicated that they had arthroscopy, ligament repair or meniscectomy on any previous visit.

The 12-item Short-Form (SF-12) health survey was used to assess health status including physical and mental component summary scores (21). Elevated depressive symptoms were considered present if participants had The Centers for Epidemiologic Studies Depression Scale (CES-D) score >16 (22). The Charlson index was used to develop a comorbidity score and was categorized into 0, 1, and ≥2 (23). BMI was calculated from measured height and weight [weight (kg)/height (m)²] and categorized as normal weight (<25); overweight (25 - <30); and obese (≥30) (24). Analgesic use including acetaminophen, non-steroidal anti-inflammatory drugs (NSAIDs), COX-2 selective inhibitors, opioids, and doxycycline was assessed for the previous 30 days. Biological supplements including glucosamine, chondroitin sulfate, Methylsulfonylmethane and S-adenosylmethionine were also assessed.

Statistical analysis

To understand the potential selection bias that may arise due to “lost to follow-up”, we first compared the characteristics of sociodemographic and clinical factors and concurrent pharmacological treatment use at baseline (one year before initiation), index year, and one year after initiation. We also examined the distribution of the outcome variables and ruled out departures from normality. We then developed a series of models to derive crude estimates, an estimate adjusted for baseline covariates, and adjusted for time-varying confounders using generalized estimating equations (GEEs) for continuous outcomes adjusted for within-participant correlation with an unstructured correlation matrix (25).

Given the OAI data structure, we considered that previously measured study outcomes and time-varying confounders may be simultaneously confounders and intermediate variables (Figure 2). As a result, the estimated overall treatment effects would likely be biased using standard regression models (26). To account for time- varying confounders that may lie on the causal pathway from previous treatments to the study outcomes, we used marginal structural models to estimate the overall treatment effects of injection use through inverse probability of treatment weights (13,14).

Figure 2.

Directed acyclic graph (causal diagram) between the initiation of injection use, study outcomes, censoring, and potential time-varying confounders.

The weights were calculated in three steps. First, we estimated time-varying stabilized inverse probability treatment weights separately for hyaluronic acid injection and corticosteroid injection use using non-users as the comparator at the index and follow-up visit. While the numerator was estimated using the conditional probability of observed injection use given the baseline characteristics, the denominator was the predicted probability of observed injection use at the index and follow-up visit conditional on baseline covariates and time-varying confounders (e.g., WOMAC subscale scores measured at the prior visit and the same visit as use of injections). To construct appropriate weights, we also explored the sensitivity of weights to different model specifications at the index visit (Supplementary eTable 1) (27). For three different outcomes, we adjusted for the previously measured WOMAC subscales as a potential confounder. Second, since participants were censored due to death, switching injection (28), and/or having total knee replacement during the follow-up, we estimated and incorporated the inverse probability of censoring weight to account for the potential selection bias due to differential censoring by injections use (13,14). Sociodemographic and clinical factors among participants who were censored by censoring mechanisms were also examined (Supplementary eTable 2). Censoring weights were calculated using similar approach as treatment weights, except that past treatment use was also added into models to estimate the probability of having observed censoring status. Lastly, final weights were then calculated as the products of treatment (including index and follow-up visits) and censoring weights. In addition to checking the distributions of the final weights, we also plotted the log odds of injection use conditional on the covariates to examine if there was an adequate degree of variation given observed values against the predicted injection use (Supplementary eFigure 1) (29). To minimize the impact of potential violations of the positivity assumption, we also truncated the weights at the first and 99th percentile (27).

With the final estimated weights, we used weighted linear models adjusted for baseline covariates to estimate effects of injection use on changes in symptoms with 95% confidence intervals (95% CIs). Under the assumptions of no unmeasured confounding with correct specifications of treatment and outcome models, the beta coefficients from marginal structural models indicated the effects of hyaluronic acid or corticosteroid injection use compared to non-users on yearly changes in WOMAC scores. Minimal clinically important changes for improvements were defined using previous validation studies ranging from −4.6 to −1.2 for WOMAC pain, −1.5 to −0.5 for WOMAC stiffness, and −9.9 to −4.1 for WOMAC physical function (30–33).

Sensitivity analysis

To examine the robustness of findings, we conducted sensitivity analyses to account for missing values of covariates. In particular, we were concerned about the missing data of one of the confounders measured at the index visit. By virtue of how we selected the sample for the current investigation and the OAI protocol, ~ 30% had missing data on K-L grade. Multiple imputation was performed to handle missing data in the context of marginal structural model analyses (34). We applied the Fully Conditional Specification method for imputation of missing data using SAS PROC MI FCS (35). We first used all available information from the covariates (including the outcome variable) as variables in the imputation model to impute the missing values (36). Twenty imputed datasets were created. We then incorporated the imputed values to rebuild the inverse probability treatment weights and fit the outcome models for each imputed dataset. Finally, we combined estimates and generated valid inferences using SAS PROC MIANALYZE to compare results.

Results

Sociodemographic and clinical characteristics of study participants

Table 1 shows sociodemographic and clinical characteristics of study participants at baseline (one year before the injection initiation), index year, and one year after the injection initiation among those remaining uncensored during follow-up. Among participants initiating injection use, the majority had K-L ≥ 3. Both CO injections initiators and non-users had similar distribution of sociodemographic and clinical characteristics at baseline. Men and those with higher household income (e.g. > $50,000) comprised the majority of HA injection initiators relative to non-users. Among HA injection initiators, 33.7% of had K-L grade 4, while 17.4% of non-users had K-L grade 4. During follow-up, the distribution of characteristics was similar over time compared to the distribution at baseline. The proportion of those censored at one year after initiation was 29.8% for HA injection initiators relative to the other groups (e.g. CO: 19.9%; non-users: 3.0%).

Table 1.

Sociodemographic and clinical characteristics among participants with radiographically confirmed knee OA by use of injections.

Characteristics	Baselinea			Index year			One year after injection initiation
	CO	HA	Non-user	CO	HA	Non-user	CO	HA	Non-user
Total n *	318	94	576	318	94	576	257	66	559
Proportions relative to baseline	100	100	100	100	100	100	80.1	70.2	97.0
Injection use (n)	0	0	0	318	94	0	63	15	0
	Percentage
Mean age (years, (SD))	66.9 (9.2)	65.0 (8.6)	64.0 (9.3)	67.9 (9.2)	66.0 (8.6)	65.0 (9.3)	68.5 (9.1)	66.0 (8.8)	65.8 (9.4)
Women	64.2	44.7	55.6	64.2	44.7	55.6	65.4	43.9	55.5
Ethnicity/Race
Non-Hispanic white	77.4	88.3	67.9	77.4	88.3	67.9	75.5	87.9	67.4
Non-Hispanic black	17.9	6.4	30.0	17.9	6.4	30.0	21.0	7.6	30.4
Other	4.7	5.3	2.1	4.7	5.3	2.1	3.5	4.6	2.2
Education
High school or less	18.6	12.9	20.9	18.6	12.9	20.9	20.2	10.8	20.
Some college	27.1	23.7	26.7	27.1	23.7	26.7	28.8	24.6	27.1
College graduate	22.1	20.4	19.2	22.1	20.4	19.2	19.5	18.5	19.8
Graduate school	32.2	43.0	33.3	32.2	43.0	33.3	31.5	46.2	32.9
Income ($)
<25,000	15.5	4.3	17.9	15.5	4.3	17.9	17.1	4.6	18.1
25,000 – 50,000	30.9	22.6	27.8	30.9	22.6	27.8	32.7	16.9	27.6
>50,000	53.6	73.1	54.3	53.6	73.1	54.3	50.2	78.5	54.3
K-L grade
2	40.1	22.1	52.3	35.7	16.3	50.0	38.9	22.2	50.7
3	39.1	44.2	30.3	37.9	45.0	30.8	41.2	46.3	31.2
4	20.8	33.7	17.4	26.4	38.8	19.3	19.9	31.5	18.2
Symptom-related multi-joint OA	55.4	57.5	55.2	55.7	60.6	50.4	56.0	62.1	49.0
History of knee injury	40.6	57.5	49.3	47.2	58.5	50.2	49.4	57.6	51.2
History of knee surgery	28.6	48.9	34.6	32.1	54.3	35.8	35.8	50.0	35.2
Body Mass Index (kg/m2)
<25	12.9	14.9	13.2	12.9	16.0	13.0	12.1	15.2	14.5
25 – <30	40.6	31.9	39.7	40.6	30.9	37.7	40.5	31.8	36.3
≥30	46.5	53.2	47.1	46.5	53.2	49.3	47.5	53.0	49.2
Knee alignment
Normal	20.5	18.9	18.4	17.4	15.9	15.9	18.4	11.9	16.0
Varus	39.3	40.0	41.9	42.9	44.3	45.9	38.2	44.1	46.5
Valgus	40.3	41.1	39.7	39.7	39.8	38.2	43.4	44.1	37.5
CES-D (>16)	11.0	8.9	11.0	14.3	12.2	13.3	13.2	12.9	12.1
Charlson Comorbidity Index
0	65.4	70.2	65.7	62.0	71.0	63.4	56.7	78.5	63.0
1	18.2	22.3	19.3	19.9	20.4	19.9	21.7	15.4	19.9
≥2	16.4	7.5	15.1	18.0	8.6	16.7	21.7	6.2	17.1
	Mean (standard deviation)
WOMAC Pain	5.0 (3.9)	5.1 (3.9)	4.9 (4.1)	6.2 (4.2)	5.7 (3.6)	4.4 (3.9)	5.6 (4.0)	5.2 (3.7)	4.2 (3.8)
WOMAC Stiffness	2.7 (1.7)	2.5(1.7)	2.5 (1.8)	2.8 (1.7)	2.8 (1.6)	2.4 (1.8)	2.7 (1.7)	2.5 (1.6)	2.2 (1.8)
WOMAC Physical Function	16.8 (13.0)	16.7 (11.8)	15.3 (13.0)	19.2 (13.2)	17.1 (11.3)	14.4 (12.8)	17.9 (13.2)	16.9 (11.1)	14.1 (12.8)
KOOS-QoL	54.6 (19.9)	52.2 (18.8)	56.0 (21.7)	48.3 (20.2)	49.0 (20.1)	57.4 (22.4)	50.6 (20.3)	50.5 (19.5)	58.4 (22.0)
SF-12 Physical Component Score	42.8 (9.6)	42.7 (9.2)	44.8 (9.7)	40.8 (9.4)	40.8 (9.9)	44.4 (10.1)	41.1 (9.5)	40.6 (9.3)	44.4 (10.0)
SF-12 Mental Component Score	54.1 (8.4)	55.4 (7.1)	53.8 (8.3)	54.1 (9.5)	55.4 (8.2)	53.4 (9.3)	53.6 (9.1)	55.1 (8.5)	53.2 (9.2)
Joint space width (mm)	4.6 (1.9)	4.2 (2.2)	5.1 (1.6)	4.4 (1.9)	4.0 (2.1)	4.9 (1.7)	4.4 (1.9)	4.3 (2.1)	4.9 (1.6)

Abbreviations: SD, standard deviation; K-L grade, The Kellgren–Lawrence grade; CES-D, Centers for Epidemiologic Studies Depression Scale; WOMAC, The Western Ontario and McMaster Universities Osteoarthritis Index; KOOS-QoL, Knee injury and Osteoarthritis Outcome Score Quality of life subscale.

^aOne year before the index year.

^*Number of participants with missing information:

At baseline: education (4), income (3), KL grade (114), body mass index (1) knee alignment (99), CES-D (63), Charlson Comorbidity Index (5), WOMAC Pain (29), WOMAC Stiffness (27), WOMAC Physical function (3), KOOS-QoL (27), SF-12 Physical Component Score (70), SF-12 Mental Component Score (70), joint space width (234).

At index year: education (4), income (3), KL grade (153), knee alignment (123), CES-D (72), Charlson Comorbidity Index (5), WOMAC Pain (25), WOMAC Stiffness (27), WOMAC Physical function (41), KOOS-QoL (27), SF-12 Physical Component Score (93), SF-12 Mental Component Score (93), joint space width (300).

One year after injection initiation: education (3), income (2), KL grade (145), knee alignment (120), CES-D (80), Charlson Comorbidity Index (6), WOMAC Pain (56), WOMAC Stiffness (56), WOMAC Physical function (71), KOOS-QoL (56), SF-12 Physical Component Score (93), SF-12 Mental Component Score (93), joint space width (292).

Concurrent pharmacological treatment use

NSAIDs were the most commonly reported concurrent-use pharmacological treatments among the study groups at baseline (Table 2). The majority of injection initiators reported analgesic use. Similar to sociodemographic and clinical characteristics, both CO injection initiators and non-users had similar distributions of concurrent pharmacological treatments use at baseline. Among HA injection initiators, 48.9% reported concurrent use of supplements such as glucosamine whereas CO injection initiators and non-users reported 30.8% and 27.1%, respectively. During follow-up, the distribution of concurrent pharmacological treatments use remained similar over time between CO injection initiators and non-users group. However, among HA injection initiators, the use of glucosamine or chondroitin sulfate decreased from baseline to one-year after initiation (e.g. glucosamine: 48.9% to 36.4%; chondroitin sulfate: 44.7% to 30.3%).

Table 2.

Concomitant use of medications and supplements among persons with radiographically confirmed knee OA by the use of injections.

	Baseline			Index year			One year after injection initiation
	CO	HA	Non-user	CO	HA	Non-user	CO	HA	Non-user
	Percentage
Use of analgesics
Acetaminophen	23.6	21.3	16.2	22.0	19.2	12.3	21.4	18.2	14.5
NSAIDsa	35.2	39.4	31.5	37.3	42.6	28.3	38.4	40.9	27.7
COX-2 inhibitors	10.4	11.7	4.2	7.2	12.8	3.1	7.8	7.6	3.4
Opioids	9.4	8.5	8.7	12.9	11.7	6.9	14.8	12.1	8.1
Any use of analgesics	55.0	56.4	44.4	58.5	63.8	39.6	57.6	57.6	38.6
3+	4.7	8.5	1.9	3.5	5.3	1.7	5.8	6.1	2.0
2+	17.9	14.9	12.7	17.3	17.0	8.0	16.3	13.6	10.9
Use of supplements
Glucosamine	30.8	48.9	27.1	34.6	44.7	25.9	26.9	36.4	21.8
Chondroitin sulfate	27.4	44.7	24.8	30.5	41.5	22.9	24.5	30.3	19.5
Methylsulfonylmethane	8.2	14.9	6.9	10.1	20.2	8.3	9.0	19.7	9.3
S-adenosylmethionine	0.3	1.1	0.4	0.9	0	0.5	0.8	0	0.5
Other current prescribed medications
Doxycycline	0	0	0.5	0.3	0	0	0	0	0.2
Vitamin D	0.3	3.2	0.5	0.6	5.3	2.2	2.8	3.0	2.5
Medications for osteoporosis	11.0	7.5	8.0	8.8	7.5	7.1	7.4	4.6	6.1

Abbreviations: NSAIDs, nonsteroidal anti-inflammatory drugs.

^aIncluding self-reported over-the-counter use and current prescriptions such as Aspirin, Ibuprofen, and Salicylate.

^*Number of participants with missing information:

At baseline: NSAIDs (29), Doxycycline (27), Vitamin D (27).

At index year: NSAIDs (31), Doxycycline (26), Vitamin D (31).

One year after injection initiation: NSAIDs (58), Doxycycline (40), Vitamin D (43).

Effects of injections use on knee OA

Table 3 shows average effects of initiating corticosteroid or hyaluronic acid injection use compared to non-users on patient-reported outcomes. After adjusting for potential confounders with marginal structural models, the use of corticosteroid injections did not improve WOMAC subscales compared with non-users. On average, the yearly changes were 1.24 (95% CI: 0.82 to 1.66) for WOMAC pain, 0.30 (95% CI: 0.10 to 0.49) for WOMAC stiffness, and 2.62 (95% CI: 0.94 to 4.29) for WOMAC physical function. While results from sensitivity analyses were qualitatively similar to our main findings, the effect of estimates for WOMAC pain did not meet a priori definitions of minimal clinically important differences.

Table 3.

Estimated effects of injection use compared with non-users on symptoms among persons with radiographically confirmed knee OA.

	Use of CO (β* coefficient (95% CI))	Use of HA (β* coefficient (95% CI))
WOMAC pain subscale
Crudea	1.17 (0.80 to 1.54)	0.77 (0.17 to 1.38)
Baseline covariatesb	3.49 (−2.36 to 9.33)	0.58 (0.07 to 1.09)
Baseline plus time-varying covariatesc	2.11 (0.70 to 3.53)	0.40 (−0.30 to 1.10)
Marginal structural modeld	1.24 (0.82 to 1.66)	0.50 (−0.11 to 1.11)
Sensitivity analysis	0.51 (−0.31 to 1.32)	0.42 (−1.44 to 2.29)
WOMAC stiffness subscale
Crudea	0.23 (0.04 to 0.43)	0.28 (−0.03 to 0.59)
Baseline covariatesb	−0.55 (−1.27 to 0.17)	−2.10 (−2.35 to −1.85)
Baseline plus time-varying covariatesc	0.05 (−1.08 to 1.18)	0.17 (−0.17 to 0.51)
Marginal structural modeld	0.30 (0.10 to 0.49)	−0.07 (−0.38 to 0.24)
Sensitivity analysis	0.14 (−0.21 to 0.50)	−0.43 (−1.54 to 0.67)
WOMAC physical function subscale
Crudea	2.60 (1.36 to 3.84)	1.21 (−0.70 to 3.12)
Baseline covariatesb	−1.73 (−9.50 to 6.05)	1.16 (−0.50 to 2.83)
Baseline plus time varying covariatesc	0.06 (−0.51 to 0.62)	−0.37 (−2.45 to 1.71)
Marginal structural modeld	2.62 (0.94 to 4.29)	0.49 (−1.34 to 2.32)
Sensitivity analysis	1.05 (−1.23 to 3.33)	3.84 (−3.72 to 11.40)

Abbreviations: CO, corticosteroid; CI, confidence interval; GEE, generalized estimating equations; HA, hyaluronic acid; WOMAC, The Western Ontario and McMaster Universities Arthritis Index.

^*A negative β coefficient indicates improvement for the Western WOMAC subscales and worsening for the joint space width.

^aValues derived from GEE models used an unstructured correlation matrix.

^bModels were adjusted for baseline characteristics including age, sex, race/ethnicity, education, income, Kellgren-Lawrence grade, body mass index, history of knee injury, history of knee surgery, Short Form 12 physical and mental component summary scores, WOMAC subscales, and use of analgesics and glucosamine.

^cIn addition to baseline covariates, time-varying confounders including Kellgren-Lawrence grade, history of knee injury and surgery, WOMAC subscales, and Short Form 12 physical and mental component summary scores measured at the same visit as injection use were also adjusted.

^dInverse probability-weighted analyses with truncated weights were used.

While the use of hyaluronic acid injections did not show improvements of WOMAC subscales compared with non-users, the magnitude of effects were relatively smaller compared to corticosteroid injections use. On average, the yearly changes were 0.50 (95% CI: −0.11 to 1.11) for WOMAC pain, −0.07 (95% CI: −0.38 to 0.24) for WOMAC stiffness, and 0.49 (95% CI: −1.34 to 2.32) for WOMAC physical function. The findings from sensitivity analyses remained similar.

Discussion

Using data from the OAI, we did not observe reduced symptoms associated with the initiation of corticosteroid or hyaluronic acid injections compared with non-users among participants with radiographically confirmed knee OA in the 2 years of follow-up after carefully controlling for potential time-varying and time-independent confounders with marginal structural models.

Among participants initiating corticosteroid injection use compared to non-users, our study findings are consistent with a newly updated review (4). However, the study duration included in the review varied (range: 2 weeks - 1 year) and mixed with single and/or multiple injection use. To our knowledge, there are only 2 published trials that are comparable to our study design which assessed the effect of continuous intraarticular corticosteroid use over two years. Our findings are consistent with both studies and demonstrate that the use of corticosteroid injections do not appear to reduce symptoms (8,37). Despite the fact that, more studies with adequate power and proper design may still be needed, our findings do contribute to the growing body of evidence produced using non-experimental study design with advanced analytical techniques.

Our results did not appear to support the notion that initiating hyaluronic acid injections use would be effective at relieving symptoms over longer periods of time (~2 years). Although our findings are not consistent with evidence from some reviews and meta-analysis (5,38), there are some issues that may hamper the comparison. First, the follow-up periods in trials included in the reviews are mostly short-term with only one treatment cycle. As such, the beneficial effects of long-term use remain unclear. In addition, the potential efficacy of hyaluronic acid injections to patients with more severe disease remains unknown since some trials excluded patients with severe knee OA. In our study, we observed a substantial percentage of participants initiating these injections had K-L 4. It may well be that hyaluronic acid injections may be more effective during earlier stages of disease. Furthermore, when comparing results from larger trials with better quality, later updated reviews suggested that the use of hyaluronic acid injections compared to non-users is associated with small but not clinically important improvement in knee symptoms (39,40). Indeed, for changes in symptoms, our results are consistent with the study with a larger sample size over one year of follow-up (41).

This study builds on previous research in several areas. First, we used data from the OAI, a cohort study that enrolled participants with knee OA living in the community and conducted annual assessments with validated patient-reported outcomes and measures of disease progression. As such, we were able to evaluate the effectiveness of injections use in the real-world setting. With detailed information regarding disease severity, we included comparable participants who did not receive injections to reduce confounding by indication. Second, compared to clinical trials (4,5), this longitudinal and non-experimental study enabled us to examine injection use over a longer period of time and to evaluate treatment benefits in a real-world setting. While the typical clinical trial in patients with OA is ~6 months in length, we followed patients for 2 years. Lastly, to address threats to the validity of the study such as time-varying confounders and “lost to follow-up”, we used advanced statistical techniques of inverse probability treatment/censoring weights with marginal structural models (14). We further performed sensitivity analyses using multiple imputation to evaluate the robustness of the main results. Results from sensitivity analyses showed consistent findings.

Several limitations are also acknowledged. No information was available on the formulation of injections as well as dosages used (40). Given the questionnaires used in the OAI, there is a potential for mismatch between the time of injections use and outcome assessments. For example, at annual assessment visits when participants were asked about injections used, they could be in the middle of treatment cycle. On the other hand, misclassification of participants as non-users is possible. In addition, whether the use of injection was due to pain fluctuation or flares (42,43) and thus patients may experience clinically meaningful improvement remains unknown. Confounding by indication is still a possibility despite the comprehensive assessments of disease severity and concurrent treatment use in the OAI. Although sensitivity analyses were conducted due to the concern of missing data on some key covariates (e.g., K-L grade), little guidance exists regarding how to best handle missing data in the context of marginal structural models (34,44). Lastly, in the practice of constructing appropriate weights, there could be a model misspecification and/or violation of positivity assumption (27). However, we carefully constructed weights using an iterative process and graphically examined if there was an adequate degree of variation given observed and predicted injection use. We also truncated weights to reduce the potential impact of violating the positivity assumption (27).

In summary, initiating treatments with either corticosteroid or hyaluronic acid injections was not associated with reduced symptoms compared to non-users over two years in patients with knee OA. In addition to the substantial cost from long-term use (10), patients initiating such treatment options should be advised that the beneficial effects of such treatments may be time-limited. Future research targeting comparative effectiveness of these commonly used injections may be helpful to understand the usefulness of these treatments for patients with knee OA.