WorksMathematics
Proportion Outcomes in Interrupted Time Series Analysis: Why Both Boundedness and Autocorrelation Matter
Abstract
Background: Proportions are common in interrupted time series analysis (ITSA), but bounded outcomes and serial dependence require distinct modeling considerations. This article provides guidance on selecting and evaluating models for proportion outcomes. Methods: Two published healthcare ITSA studies were reanalyzed. The first examined weekly proportions near the boundaries of the unit interval using segmented linear regression. The second examined monthly proportions without reported assessment or adjustment for autocorrelation. Both were reanalyzed using models appropriate for their outcome distributions and temporal dependence. Results: In the first study, segmented linear regression produced six fitted values below zero, and 35 of 66 confidence intervals extended beyond the permissible outcome range. Fractional logistic regression constrained fitted means to the unit interval while accommodating zeros and ones. In the second study, beta autoregressive regression identified significant first-order serial dependence (rho1 = 0.263, 95% CI 0.093 to 0.432; P = .002). After adjustment, residuals showed no evidence of remaining autocorrelation (Ljung-Box Q12 = 15.59; P = .211). Conclusions: Boundedness and serial dependence are distinct considerations in ITSA of proportions. Accounting for autocorrelation does not ensure predictions remain within the permissible range, and modeling bounded outcomes does not ensure serial dependence is addressed. Investigators should assess both features and select models appropriate for the outcome distribution and temporal dependence.
Introduction
Interrupted time series analysis (ITSA) is widely used in clinical, health services, and public health research to evaluate interventions when randomization is infeasible [1, 2, 3]. ITSA has been argued to have strong internal validity, primarily through its control over regression to the mean [1, 2, 4], and generally has good external validity, particularly when the unit of measurement is at the population level or when results can be generalized to other units, treatments, or settings [2, 5]. In the single-group design, observations collected at regular intervals before and after an intervention are used to estimate the pre-intervention trend and changes in the level and trend of the outcome following the intervention [3]. Valid inference depends not only on specifying these intervention-associated changes but also on using a model appropriate for the outcome and its temporal dependence [6, 7].
Many outcomes analyzed using ITSA are proportions or percentages bounded by zero and one. Examples include the proportion of patients receiving a recommended treatment, prescriptions meeting specified criteria, or procedures complying with a clinical guideline. Nevertheless, proportion outcomes are sometimes analyzed using methods developed for continuous unbounded outcomes. Linear regression, for example, does not constrain the conditional mean to the unit interval and can produce fitted values or confidence limits below zero or above one, particularly when observations lie near either boundary. Conventional autoregressive integrated moving-average (ARIMA) models can address serial dependence but likewise do not inherently constrain fitted values to the permissible range. Thus, accounting for the time-series structure does not by itself address the bounded nature of a proportion outcome.
Serial dependence must also be considered separately. Failure to account adequately for autocorrelation can result in incorrectly estimated standard errors, confidence intervals, and hypothesis tests [8, 9, 10, 11]. Turner et al. found that autocorrelation was acknowledged in only 63% of published ITSA series and that an estimate of the autocorrelation coefficient was reported in only 1% [7]. The data underlying this and related empirical evaluations of published ITSA are publicly available in an interrupted time series repository [12], which includes numerous examples of outcomes expressed as proportions or percentages. Boundedness and serial dependence should therefore be viewed as distinct modeling considerations: addressing one does not necessarily address the other.
Models designed for proportion outcomes provide a natural starting point. Fractional regression constrains the conditional mean to the unit interval and can accommodate observations at zero and one [13]. For continuous proportions strictly between zero and one, beta regression provides a distributionally appropriate alternative [14]. When serial dependence is present, beta autoregressive regression extends beta regression by explicitly modeling that dependence. A beta-AR() model is the autoregressive special case of the broader beta-ARMA framework [15, 16, 17]. Thus, when the dependence structure can be represented adequately by autoregressive terms, beta-AR() regression may be used, whereas more complicated temporal patterns may warrant the broader beta-ARMA model. In either case, the model addresses the bounded outcome and its temporal dependence within a common framework.
The purpose of this teaching article is to provide practical guidance for analyzing proportion outcomes in ITSA by treating boundedness and serial dependence as separate features that should each be considered during model selection and evaluation. Two published healthcare applications are reanalyzed to demonstrate the practical consequences, using data obtained from the publicly available interrupted time series repository described by Turner et al. [12]. The first contains observations at or near both boundaries of the unit interval and illustrates how an unbounded linear model can produce fitted quantities outside the permissible range. The second illustrates how serial dependence can remain unrecognized when autocorrelation is not assessed. Together, the examples demonstrate a practical approach to selecting, diagnosing, and evaluating models for proportion-outcome ITSA.
Example 1: When the Outcome Approaches its Boundaries
Aiken et al. evaluated the implementation of a surgical antibiotic prophylaxis policy at Thika Hospital in Kenya [18]. One of the principal process outcomes was the weekly proportion of patients undergoing clean or clean-contaminated surgery who received antibiotic prophylaxis within 60 minutes before surgery. The policy was implemented in February 2011, with 26 weekly observations before the intervention and 40 afterward. Preoperative prophylaxis was uncommon during most of the pre-intervention period, with a sharp increase immediately before the intervention point, followed by proportions consistently near one during the post-intervention period. The resulting series contains observations at or near both boundaries of the unit interval.
The original study evaluated changes associated with the intervention using segmented linear regression [18]. The mean structure for a single-group ITSA can be written generally as [3, 19]
where is the expected outcome at time , indexes time, indicates the post-intervention period, and denotes the intervention time. The coefficients represent the baseline level (), pre-intervention trend (), immediate change in level following the intervention (), and change in trend following the intervention (). The link function determines how this mean structure is related to the outcome. With ordinary linear regression, the identity link places no restriction on the range of the fitted values. A link appropriate for a proportion outcome can retain the same ITSA mean structure while constraining the fitted conditional mean to the unit interval.
I first reproduced the original ITSA using linear regression. As shown in Figure 1A, the consequences of using an unbounded model are apparent at both ends of the series. Six fitted values were below zero. The lower 95% confidence limit was below zero for all 26 pre-intervention observations, while the upper 95% confidence limit exceeded one for the final nine observations. Overall, the 95% confidence interval extended outside the permissible [0, 1] range at 35 of the 66 time points. These are impossible values for the outcome: a negative proportion cannot occur, nor can more than 100% of eligible patients receive prophylaxis.

(a)

(b)
Figure 1. Weekly proportion of surgical patients receiving antibiotic prophylaxis at Thika Hospital, Kenya, using data from Aiken et al. [18]. Points show observed proportions; solid lines and dotted lines show fitted values and 95% confidence intervals from (A) segmented linear regression and (B) fractional logistic regression. The vertical line marks policy implementation (2011w7).
I next re-estimated the ITSA using fractional logistic regression [13]. Fractional regression uses a logit link for the conditional mean, thereby constraining fitted values to the unit interval. It also permits observed outcomes equal to zero or one, an important feature in this application because both boundary values occur in the series. Standard beta regression, in contrast, assumes a continuous proportion strictly within (0, 1) and therefore cannot be applied directly when observations equal zero or one [14].
The fractional logistic model showed no evidence of a pre-intervention trend (, ), a large immediate increase following implementation of the policy (, ), and no evidence of a change in trend after the intervention (, ). Because these coefficients are expressed on the logit scale, their magnitudes should not be interpreted as percentage-point changes. Figure 1B therefore presents the fitted trajectory on the proportion scale, where the fitted values remain within the permissible range.
Serial dependence was also assessed after fitting the fractional logistic model. The Ljung–Box Portmanteau test showed no evidence of residual autocorrelation through 12 lags (, ) [20]. Similarly, the Cumby–Huizinga test [21], implemented using the community-contributed actest command in Stata [22], showed no evidence of first-order serial correlation (, ). Thus, although the bounded nature of the outcome required attention in this example, there was no evidence that an autoregressive adjustment was also needed.
This example does not imply that linear regression will always produce out-of-range fitted values when the outcome is a proportion. Rather, it demonstrates that an unbounded model provides no protection against such values, and that the problem can become consequential when a series approaches either boundary. It also illustrates why the observed support of the data matters when choosing among models for proportions. When legitimate observations of zero or one occur, fractional regression provides a bounded conditional mean while retaining those observations unchanged; when a continuous proportion lies strictly within , beta regression provides an alternative designed specifically for such outcomes [14].
Respecting the bounds of the outcome addresses only one of the two modeling considerations highlighted in this paper. In this example, assessment of the residuals provided no evidence that serial dependence also required modeling. The next example illustrates the contrasting case: an ITSA in which the observed proportions lie within the open unit interval, but significant autocorrelation is present.
Example 2: When Serial Dependence Is Present
Barber et al. evaluated whether prescribing directives issued by the Veterans Health Administration (VHA) and the U.S. Food and Drug Administration were associated with changes in opioid prescribing practices among VHA providers [23]. The study analyzed monthly proportions of prescriptions meeting specified prescribing criteria from January 2000 through December 2009. The example considered here concerns the directive for controlled-release oxycodone, issued in July 2003. The outcome was the monthly proportion of new controlled-release oxycodone prescriptions that were preceded by a prescription for long-acting morphine or methadone. Unlike the first example, all observed proportions lie strictly within the open unit interval , making beta regression applicable.
The original study analyzed the series using linear ITSA but did not report an assessment or adjustment for autocorrelation [23]. I first reproduced the original ITSA using linear regression (Figure 2A). The fitted trajectory remains within the permissible range of the outcome and appears reasonable by visual inspection. Thus, unlike the first example, the principal modeling problem is not apparent from the fitted trajectory alone.

Figure 2. Monthly proportion of new controlled-release oxycodone prescriptions in the Veterans Health Administration preceded by a prescription for long-acting morphine or methadone, using data from Barber et al. [23]. Points show observed proportions; solid lines and dotted lines show fitted values and 95% confidence intervals from (A) linear regression and (B) beta-AR(1) regression. The vertical line marks the July 2003 prescribing directive.
I next assessed serial dependence before specifying the autoregressive structure. Examination of the residual autocorrelation function (ACF) and partial autocorrelation function (PACF) from the unadjusted linear model indicated positive first-order dependence without a consistent pattern suggesting a higher-order or seasonal structure. The Ljung–Box Portmanteau test rejected the null hypothesis of white noise at lag 1 (, ). Similarly, the Cumby–Huizinga test identified significant serial correlation at lag 1 (, ), whereas the individual tests for lags 2 through 9 were not significant. Taken together, these diagnostics supported an AR(1) specification as a parsimonious starting model.
Because the outcome is a continuous proportion within , the ITSA was then re-estimated using beta-AR(1) regression, which models the bounded outcome and first-order serial dependence jointly [16, 17]. Estimation was performed using the community-contributed betark command in Stata [24]. The beta-AR(1) model estimated a first-order autoregressive parameter of (95% CI 0.093–0.432; ), providing further evidence of first-order serial dependence. The model also estimated a positive pre-intervention trend (, ), an immediate increase following the directive (, ), and a reduction in the post-intervention trend relative to the pre-intervention trend (, ). Because these coefficients are expressed on the logit scale, their magnitudes should not be interpreted as percentage-point changes. Figure 2B presents the fitted beta-AR(1) trajectory on the proportion scale.
The two panels of Figure 2 illustrate an important feature of this example. The fitted trajectories from the linear and beta-AR(1) models are visually similar. Nothing in the linear fit itself clearly reveals the presence of serial dependence, yet the diagnostic assessment and beta-AR(1) model provide evidence of first-order autocorrelation. A fitted trajectory that appears plausible by visual inspection therefore does not eliminate the need to assess autocorrelation.
Selection of the autoregressive structure should not end with model estimation. The residuals should be examined to determine whether the specified dependence structure has adequately accounted for the serial correlation. The Ljung–Box Portmanteau test applied to the standardized conditional residuals from the beta-AR(1) model showed no evidence of remaining serial dependence (, ). Thus, the postestimation diagnostic supported the adequacy of the AR(1) specification for these data. Had residual serial dependence remained, a higher-order beta-AR() specification or the broader beta-ARMA framework would have warranted consideration [15, 16, 17].
This example demonstrates that using a model appropriate for a proportion outcome does not, by itself, address temporal dependence. Beta regression respects the bounded support of a continuous proportion within , but serial dependence must still be assessed and, when present, modeled appropriately. The complete process is therefore to assess the dependence structure, incorporate it into the model, and then examine the residuals to determine whether the selected structure was adequate.
Practical Recommendations for Proportion-Outcome ITSA
Taken together, the two examples illustrate complementary problems in proportion-outcome ITSA. In the first, the series approached and reached the boundaries of the unit interval, and an unbounded linear model produced impossible fitted quantities. In the second, the observed proportions remained within and the fitted linear trajectory appeared reasonable, but significant serial dependence was present. Model selection for proportion-outcome ITSA should therefore consider both the support of the outcome and the temporal dependence among observations.
Table 1 summarizes a practical framework for model selection. The first consideration is whether the observed proportion lies strictly within or includes legitimate observations at zero or one. Continuous proportions strictly within can be modeled using beta regression, whereas fractional regression provides a convenient alternative when boundary observations are present [13, 14]. The second consideration is whether the series exhibits serial dependence. For an interior proportion with autoregressive dependence, beta-AR() regression allows the outcome distribution and dependence structure to be modeled jointly; more complicated dependence may warrant the broader beta-ARMA framework [15, 16, 17].
| Observed outcome | Serial dependence | Modeling consideration |
| Continuous proportion strictly within (0, 1) | No evidence of serial dependence | Beta regression |
| Continuous proportion strictly within (0, 1) | Autoregressive dependence | Beta-AR() regression |
| Continuous proportion strictly within (0, 1) | More complicated temporal dependence | Consider the broader beta-ARMA framework |
| Includes legitimate observations at 0 or 1 | No evidence of serial dependence | Fractional regression or another model that accommodates boundary values |
| Includes legitimate observations at 0 or 1 | Serial dependence present | Use a boundary-capable proportion model together with an appropriate method for addressing the temporal dependence; assess model adequacy after estimation |
Table 1. Practical considerations for selecting a model for proportion-outcome ITSA
The final row of Table 1 is intentionally less prescriptive. The presence of zero or one rules out standard beta regression, but it does not remove the need to address serial dependence. Fractional regression can accommodate the boundary values, and its inference can be made robust to serial dependence using an appropriate covariance estimator. Alternatively, specialized time-series models for proportions with boundary values may be considered when their assumptions and implementation are appropriate. The important point is that changing the standard errors does not alter the conditional mean model, just as choosing a bounded conditional mean does not by itself model serial dependence.
Autocorrelation should be assessed rather than assumed to follow a particular order. A range of complementary diagnostic approaches is available, including examination of residual autocorrelation patterns and formal tests for serial dependence, such as the Ljung–Box Portmanteau [20] and Cumby–Huizinga [21] tests. No single diagnostic should be regarded as definitive; rather, the available evidence should be considered together when identifying a plausible and parsimonious dependence structure. Once a candidate structure has been incorporated into the model, the residuals should be reassessed for remaining serial dependence. If residual autocorrelation persists, a higher-order or otherwise more flexible dependence structure should be considered. Conversely, additional complexity is unnecessary when a simpler model adequately accounts for the observed dependence. This identify–estimate–check sequence is preferable to routinely assuming AR(1) simply because it is commonly used in ITSA.
The distinction between modeling serial dependence and adjusting inference for it is also important. Approaches such as heteroskedasticity- and autocorrelation-consistent (HAC) standard errors can adjust estimated uncertainty in the presence of serial correlation without explicitly modeling the dependence structure. Simulation evidence comparing fractional regression with HAC standard errors to beta regression with autoregressive errors suggests that explicit modeling of the serial dependence can provide better inferential performance under persistent or higher-order autocorrelation, although performance depends on the data-generating conditions [25]. Thus, a covariance correction and an explicit time-series model should not be regarded as interchangeable simply because both address consequences of autocorrelation.
The considerations emphasized here do not replace the broader requirements for specifying a valid ITSA model. Investigators should also consider features such as seasonality, nonlinear underlying trends, time-varying confounders, and concurrent events or co-interventions, as well as whether the intervention effect is expected to be immediate, delayed, gradual, or otherwise time-varying. These features require separate consideration because neither selecting a model appropriate for the outcome distribution nor accounting for period-to-period serial dependence will, by itself, control for them. More comprehensive guidance on these general aspects of ITSA specification is available elsewhere [3, 6, 26].
Finally, graphical displays remain useful but should complement rather than replace diagnostic assessment. In the first example, the consequences of an unbounded model were immediately apparent because fitted values and confidence limits crossed the boundaries of the outcome. In the second, the linear and beta-AR(1) fitted trajectories were visually similar despite significant first-order serial dependence. A graph can therefore reveal some forms of model misspecification while leaving others largely invisible. Assessment of both the support of the outcome and the residual dependence structure should be part of routine model evaluation in proportion-outcome ITSA.
Discussion
The issues considered in this paper are likely to extend beyond the two examples presented here. Turner et al. documented substantial variation in the statistical methods used in published ITSA studies and found that autocorrelation was not consistently assessed or reported [7]. Their accompanying repository contains a large collection of published interrupted time series datasets, including series expressed as proportions and percentages [12]. Although the present paper was not designed to systematically evaluate the analytical methods used across that literature, the availability of these data suggests that the modeling issues demonstrated here can be explored much more broadly.
The repository also provides a useful resource for extending the examples in this teaching article. Readers can use the available proportion and percentage series to examine the range and distribution of an outcome, compare bounded and unbounded fitted models, investigate residual serial dependence, and evaluate alternative dependence structures. Such reanalyses can be used as individual or classroom learning exercises to explore how modeling choices affect fitted values, uncertainty, and interpretation using data from actual published studies rather than simulated examples.
The contribution of this paper is primarily conceptual and practical rather than the introduction of a new statistical method. Beta regression, fractional regression, and beta time-series models are established methods. The central argument is that the distributional characteristics of the outcome and the temporal dependence of the series represent separate dimensions of model specification that should be considered together in proportion-outcome ITSA. This perspective also helps distinguish methods that explicitly model serial dependence from approaches that adjust inference for its consequences. For readers interested in the comparative statistical performance of these approaches, simulation evidence comparing beta regression with autoregressive errors and fractional regression with HAC standard errors is reported elsewhere [25].
Several limitations should be considered. The two applications were used as illustrative examples and were not selected through a systematic review of published proportion-outcome ITSA studies. Accordingly, no claims are made about the prevalence of inappropriate model specification in the published literature. The examples also represent only two configurations of outcome support and temporal dependence and do not encompass the full range of models that may be appropriate for bounded time-series data. In particular, proportions containing boundary values together with complex serial dependence may require methods beyond those demonstrated here. Finally, the applications use single-group ITSA. Although the principles concerning bounded outcomes and temporal dependence are not unique to the single-group design, additional considerations arise in controlled or multiple-group ITSA [3, 19].
These limitations also point to opportunities for further work. The public availability of large collections of published ITSA datasets makes it possible to examine empirically how often proportion outcomes approach their boundaries, how frequently alternative outcome models produce materially different fitted values or inference, and how often serial dependence remains after commonly used specifications. Such work would complement the illustrative purpose of the present paper and could help establish how consequential these modeling choices are in applied ITSA practice.
Conclusions
For proportion outcomes in ITSA, model specification should reflect both the support of the outcome and the temporal dependence among observations. Investigators should select a model appropriate for the observed proportion, assess the series for serial dependence, and evaluate residual dependence after estimation. These steps should be incorporated alongside the broader requirements of good ITSA specification and provide a practical framework for avoiding preventable model misspecification in proportion-outcome ITSA.
Data and code availability
The data used in both illustrative examples are publicly available from the interrupted time series repository described by Turner et al. [12]. The itsa command [3] and the stand-alone betark command [24], used for the analyses presented here, and the actest command [22], used for the Cumby–Huizinga tests, are available from the Statistical Software Components (SSC) archive.
References
- [1]Donald T. Campbell and Julian C. Stanley. Experimental and Quasi-Experimental Designs for Research. Rand McNally, Chicago, 1966.
- [2]William R. Shadish, Thomas D. Cook, and Donald T. Campbell. Experimental and Quasi-Experimental Designs for Generalized Causal Inference. Houghton Mifflin, Boston, 2002.
- [3]Ariel Linden. Conducting interrupted time-series analysis for single- and multiple-group comparisons. Stata Journal, 15(2):480–500, 2015. doi: 10.1177/1536867X1501500208. URL https://doi.org/10.1177/1536867X1501500208.
- [4]Ariel Linden. Estimating the effect of regression to the mean in health management programs. Disease Management and Health Outcomes, 15(1):7–12, 2007.DOI
- [5]Ariel Linden, John Adams, and Nancy Roberts. The generalizability of disease management program results: getting from here to there. Manage Care Interface, 17(7):38–45, 2004.
- [6]James Lopez Bernal, Stephen Soumerai, and Antonio Gasparrini. A methodological framework for model selection in interrupted time series studies. Journal of Clinical Epidemiology, 103:82–91, 2018. doi: 10.1016/j.jclinepi.2018.05.026. URL https://doi.org/10.1016/j.jclinepi.2018.05.026.
- [7]Simon L. Turner, Amalia Karahalios, Andrew B. Forbes, Monica Taljaard, Jeremy M. Grimshaw, Allen C. Cheng, Lisa Bero, and Joanne E. McKenzie. Design characteristics and statistical methods used in interrupted time series studies evaluating public health interventions: a review. Journal of Clinical Epidemiology, 122:1–11, 2020. doi: 10.1016/j.jclinepi.2020.02.006. URL https://doi.org/10.1016/j.jclinepi.2020.02.006.
- [8]Ariel Linden. Adjustment for autocorrelation in multiple-group (controlled) interrupted time series analysis and its effect on power: a simulation study of the Newey-West and Prais-Winsten methods, 2026. URL https://doi.org/10.21203/rs.3.rs-8865851/v1. Preprint, Research Square.
- [9]Ariel Linden. Multiple-group (controlled) interrupted time series analysis with higher-order autoregressive errors: A simulation study comparing Newey–West and Prais–Winsten methods. Journal of Statistical Computation and Simulation, 2026. doi: 10.1080/00949655.2026.2697351.DOI
- [10]Simon L. Turner, Andrew B. Forbes, Amalia Karahalios, Monica Taljaard, and Joanne E. McKenzie. Evaluation of statistical methods used in the analysis of interrupted time series studies: a simulation study. BMC Medical Research Methodology, 21:181, 2021. doi: 10.1186/s12874-021-01364-0. URL https://doi.org/10.1186/s12874-021-01364-0.
- [11]Christian Bottomley, Mildred Ooko, Antonio Gasparrini, and Ruth H. Keogh. In praise of Prais–Winsten: an evaluation of methods used to account for autocorrelation in interrupted time series. Statistics in Medicine, 42(8):1277–1288, 2023. doi: 10.1002/sim.9669. URL https://doi.org/10.1002/sim.9669.
- [12]Simon L. Turner, Elizabeth Korevaar, Amalia Karahalios, Andrew B. Forbes, and Joanne E. McKenzie. Interrupted time series datasets from studies investigating the impact of interventions or exposures in public health and social science: a data note. BMC Research Notes, 18:32, 2025. doi: 10.1186/s13104-024-07055-5.DOI
- [13]Leslie E. Papke and Jeffrey M. Wooldridge. Econometric methods for fractional response variables with an application to 401(k) plan participation rates. Journal of Applied Econometrics, 11(6):619–632, 1996.
- [14]Silvia Ferrari and Francisco Cribari-Neto. Beta regression for modelling rates and proportions. Journal of Applied Statistics, 31(7):799–815, 2004. doi: 10.1080/0266476042000214501.DOI
- [15]Michael A. Benjamin, Robert A. Rigby, and D. Mikis Stasinopoulos. Generalized autoregressive moving average models. Journal of the American Statistical Association, 98(461):214–223, 2003. doi: 10.1198/016214503388619238. URL https://doi.org/10.1198/016214503388619238.
- [16]Andréa V. Rocha and Francisco Cribari-Neto. Beta autoregressive moving average models. TEST, 18(3):529–545, 2009. doi: 10.1007/s11749-009-0136-z.DOI
- [17]Guillermo Ferreira, Jorge I. Figueroa-Zúñiga, and Mário de Castro. Partially linear beta regression model with autoregressive errors. TEST, 24(4):752–775, 2015. doi: 10.1007/s11749-015-0433-7.DOI
- [18]Alexander M. Aiken, Anthony K. Wanyoro, Jonah Mwangi, Francis Juma, Isaac K. Mugoya, and J. Anthony G. Scott. Changing use of surgical antibiotic prophylaxis in thika hospital, kenya: A quality improvement intervention with an interrupted time series design. PLoS ONE, 8(11):e78942, 2013. doi: 10.1371/journal.pone.0078942.DOI
- [19]Ariel Linden. A comprehensive set of postestimation measures to enrich interrupted time-series analysis. Stata Journal, 17(1):73–88, 2017. doi: 10.1177/1536867X1701700105. URL https://doi.org/10.1177/1536867X1701700105.
- [20]Greta M. Ljung and George E. P. Box. On a measure of lack of fit in time series models. Biometrika, 65(2):297–303, 1978. doi: 10.1093/biomet/65.2.297.DOI
- [21]Robert E. Cumby and John Huizinga. Testing the autocorrelation structure of disturbances in ordinary least squares and instrumental variables regressions. Econometrica, 60(1):185–195, 1992. doi: 10.2307/2951684.DOI
- [22]Christopher F. Baum and Mark E. Schaffer. ACTEST: Stata module to perform Cumby-Huizinga general test for autocorrelation in time series, 2013. URL https://ideas.repec.org/c/boc/bocode/s457668.html. Statistical Software Components S457668, Boston College Department of Economics, revised 28 February 2025.
- [23]C. Barber, D. Gagnon, J. Fonda, K. Cho, J. Hermos, and M. Miller. Assessing the impact of prescribing directives on opioid prescribing practices among Veterans Health Administration providers. Pharmacoepidemiology and Drug Safety, 26(1):40–46, 2017. doi: 10.1002/pds.4066. URL https://doi.org/10.1002/pds.4066.
- [24]Ariel Linden. BETARK: Stata module for computing beta regression with autoregressive-corrected errors for proportion outcomes, by joint conditional maximum likelihood, 2026. Statistical Software Components S459773, Boston College Department of Economics.
- [25]A. Linden. Beta regression with autoregressive errors for interrupted time series analysis of proportion and rate outcomes: A simulation study, 2026. URL https://arxiv.org/abs/2607.07914. Preprint.
- [26]James Lopez Bernal, Steven Cummins, and Antonio Gasparrini. Interrupted time series regression for the evaluation of public health interventions: a tutorial. International Journal of Epidemiology, 46(1):348–355, 2017. doi: 10.1093/ije/dyw098.DOI