Testing and Estimation of structural breaks in Stata
For an overview of xtbreak see xtbreak.
Table of Contents
xtbreak test depvar [indepvars] [if] ,
breakpoints(numlist| datelist [,index| fmt(string)]) options1 options5 options6
breakpoints() specifies the time period of the known structural break.
xtbreak test depvar [indepvars] [if] ,
hypothesis(1|2|3) breaks(#) options1 options2 options3 options4 options5
hypothesis(1\2\3) specifies which hypothesis to test, see hypothesises. breaks(#) sets the number of breaks.
options1 | Description — | — breakconstant | break in constant noconstant | suppresses constant nobreakvariables(varlist1) | variables with no structural break(s) vce(type) | covariance matrix estimator, allowed: ssr, hac, hc and np inverter(type) inverter, default is speed. See options. python use Python to calculated SSRs to improve speed. See details. noreweigh do not reweigh time-unit specific errors by the number of total observations over actual observations for a given time period in order to increase the SSR of segments of unabalanced panels with missing data.
| options2 | Description |
|---|---|
| trimming(real) | minimal segment length |
| error(real) | error margin for partial break model |
| options3 | Description |
|---|---|
| wdmax | Use weighted test statistic instead of unweighted |
| level(#) | set level for critical values |
| options4 | Description |
|---|---|
| sequential | Sequential F-Test to obtain number of breaks |
| options5 | Description |
|---|---|
| nofixedeffects | suppresses fixed effects (only for panel data sets) |
| breakfixedeffects | break in fixed effects |
| csd | add cross-section averages of variables with and without breaks. |
| csa(varlist) | Variables with breaks used to calculate cross-sectional averages |
| csanobreak(varlist) | Variables without breaks used to calculate cross-sectional averages |
| kfactors(varlist) | Known factors, which are constant across the cross-sectional dimension but are affected by structural breaks. Examples are seasonal dummies or other observed common factors such as asset returns and oil prices. |
| nbkfactors(varlist) | same as above but without breaks. |
| options6 | Description |
|---|---|
| skiph2 | skips hypohesis B |
| clevel(#) | specifies level for critical values to detect breaks. |
| strict | strict behaviour of sequential test. Improves speed. |
| maxbreaks(#) | sets maximum number of breaks for sequential test. Improves speed. |
Data has to be xtset before using xtbreak. depvars, indepvars and varlist1, varlist2 may contain time-series operators.
xtbreak test implements multiple tests for structural breaks in time series and panel data models. The number and period of occurrence of structural breaks can be known and unknown. In the case of a known breakpoint xtbreak test can test if the break occurs at a specific point in time. For unknown breaks, xtbreak test implements three different hypothesises. The first is no break against the alternative of s breaks, the second hypothesis is no breaks against a lower and upper limit of breaks. The last hypothesis tests the null of s breaks against the alternative of one more break (s+1).
xtbreak test implements the tests for structural breaks discussed in Bai & Perron (1998, 2003), Karavias, Narayan, Westerlund (2021) and Ditzen, Karavias, Westerlund (2021).
For the remainder we assume the following model:
y(i,t) = sigma0(1) + sigma1(1) z(i,t) + beta0(1,i) + beta1 x(i,t) + e(it) for t = 1,...,T1
y(i,t) = sigma0(2) + sigma1(2) z(i,t) + beta0(1,i) + beta1 x(i,t) + e(it) for t = T1+1,...,T2
...
y(i,t) = sigma0(s) + sigma1(s) z(i,t) + beta0(1,i) + beta1 x(i,t) + e(it) for t = Ts,...,T
where s is the number of the segment/breaks, z(i,t) is a NT1xq matrix containing the variables whose relationship with y breaks. A break in the constant is possible. x(i,t) is a NTxp matrix with variables without a break. sigma0(s), sigma1(s) are the coefficients with structural breaks and T1,…,Ts are the periods of the breakpoints.
In pure time series model breaks in the constant (or deterministics) are possible. In this case sigma0(s) is a constant with a structural break. Fixed effects in panel data models cannot have a break.
xtbreak will automatically determine whether a time series or panel dataset is used.
Assume that the numbers of breaks and their occurence is known. xtbreak test can test the breakpoints. The F-Statistic for the test with s breaks at known dates is:
F(s,q) = dof_adj sigma' R' (R V R')^(-1) R sigma
dof_adj is a degree of freedom adjustment, sigma a matrix containing the coefficient estimates of sigma, R is the conventional matrix of a Wald test and V is a variance-covariance matrix. Under the null F(s,q) is F distributed.
A known break date can be tested with xtbreak test using the option breakpoints(numlist|datelist,[index|fmt()]). numlist|datelist defines the periods of the breaks. If numlist is used, then the option index is required. If datelist is used, then option fmt(format) defines the format in use.
If the number and thus the time period of breaks is unknwon, xtbreak test offers three different hypothesises:
Formally the hypothesis are:
H_0: no break vs. H_1: s breaks at unknown dates
Bai & Perron (1998) suggest to take the supremum of the F-Statistics:
supF(s,q) = sup (l1,..,lq) F(l,q)
where l1, lq are the different sets of possible breakpoints. Essentially the test is the test of a known break after estimation of the breakpoints given a number of breaks. For a discussion of the estimation of the breakpoints see xtbreak estimate .
The supremum F-Test is called in xtbreak test using the options breaks(#) hypothesis(1). breaks(#) sets the number of breaks. Critical values can be found in Bai & Perron (1998, 2003) and are supplied by xtbreak test.
A test of the null hypothesis of no structural change against the alternative that an unknown number of structural breaks have occurred, where this unknown number of breaks lies between s0 and s1 is:
H_0: no break vs. H_1: s0 <= s <= s1 breaks at unknown dates
The so-called double maximum test statistic is:
supF = sup(s0,..s,..,s1) supF(s,q)
where supF(s,q) is as defined above.
Generally speaking, the double maximum test estimates the breakdates for each number of breaks between s0 and s1, calculates the corresponding test statistic and then selects the largest one. Two versions of the double maximum test are available, an unweighted and a weighted test. For the weighted test the test supF(l,q) test statistics are weighted by critical values.
The double maximum test can be used if the options breaks(#) hypothesis(2) are used. Critical values can be found in Bai & Perron (1998, 2003) and are supplied by xtbreak test.
A test of the null hypothesis that, s structural breaks have occurred, against the alternative that s + 1 breaks have occurred is:
H_0: s breaks vs H_1 s+1 breaks
F(s+1\s) = sup(s=1,..,s+1) sup(s0,..s,..,s1) supF(s,q)
The test is essentially comparing the SSR of the model with s breaks to the minimum of the SSR of the model with s+1 breaks.
The F(s+1\s) test is integrated in xtbreak test with the options breaks(#) hypothesis(3). Critical values can be found in Bai & Perron (1998, 2003) and are supplied by xtbreak test.
| Option | Description | ||
|---|---|---|---|
| breakpoints(numlist\datelist [,index\fmt(format)]) | specifies the known breakpoints. Known breakpoints can be set by either the number of observation or by the value of the time identifier. If a numlist is used, option index is required. For example breakpoints(10,index) specifies that the one break occurs at the 10th observation in time. datelist takes a list of dates. For example breakpoints(2010Q1) , fmt(tq) specifies a break in Quarter 1 in 2010. The option fmt() specifies the format and is required if a datelist is used. The format set in breakpoints() and the time identifier needs to be the same. |
||
| breaks(#) | specifies the number of unknown breaks under the alternative. For hypothesis 2, breaks() can take two values, for example breaks(4 6) test for no breaks against 4-6 breaks. If only one value specified, then the lower limit is set to 1. | ||
| hypothesis(1\2\3) | specifies which hypothesis to test. h(1) test for no breaks vs. s breaks, h(2) for no break vs. s0 <= s <= s1 breaks and h(3) for s vs. s+1 breaks. Hypothesis 3 is the default. | ||
| sequential | sequential F-Test to determine number of breaks. The number of breaks is varied from s = 0 to breaks()-1 or floor(1/min_length). | ||
| breakconstant | break in constant. Default is no breaks in deterministics. | ||
| noconstant | suppresses constant. | ||
| nofixedeffects | suppresses individual fixed effects (panel data only). | ||
| breakfixedeffects | break in fixed effects. | ||
| nobreakvariables(varlist1) | defines variables with no structural break(s). varlist1 can contain time series operators. | ||
| vce(type) | covariance matrix estimator, allowed: ssr, hac, hc and np. | ||
| trimming(real) | minimal segment length in percent. The minimal segment length is the minmal time periods between two breaks. The default is 15% (0.15). Critical values are available for %5, 10%, 15%, 20% and 25%. | ||
| error(real) | define error margin for partial break model. | ||
| wdmax | Use weighted test statistic instead of unweighted for the double maximum test (hypotheis 2). | ||
| level(#) | set level for critical values for weighted double maximum test. If a value is choosen for which no critical values exists, xtbreak test will choose the closest level. | ||
| csd | adds cross-section averages of variables with and without breaks. | ||
| csa(varlist) | specify the variables with and without breaks which are added as cross-sectional averages. xtbreak calculates internally the cross-sectional average. | ||
| csanobreak() | same as csa() but for variables without a break. | ||
| kfactors(varlist) | Known factors, which are constant across the cross-sectional dimension but are affected by structural breaks. Examples are seasonal dummies or other observed common factors such as asset returns and oil prices. | ||
| nbkfactors(varlist) | same as above but without breaks. | ||
| inverter(type) | sets the inverter. type can be: speed (invsym), precision, qr (equivalent to precision; qrinv), chol (chol), p (pinv), or lu (luinv). Choice of inverter has implications on speed and precision. For an overview see https://www.stata.com/manuals/m-4solvers.pdf. | ||
| python | use Python to calculated SSRs to improve speed. Requires Stata 16 or later, Python and the following packages: scipy, numpy, pandas and xarray. | ||
| noreweigh | avoids to reweigh time-unit specific errors by the number of total observations over actual observations for a given time period in order to increase the SSR of segments of unabalanced panels with missing data. Results with this options should be used indicative. See also section on Unbalanced Panels. | ||
| skiph2 | Skips Hypothesis 2 (H0: no break vs H1: (0 < s < s_{max}) breaks) when running xtbreak without the estimate or test option. | ||
| cvalue(level) | specifies the level of the critical value to be used to estimate the number of breaks using the sequential test. For example cvalue(0.99) uses the 1\% critical values to determine the number of breaks using the sequential test. See level(#) for further details. | ||
| strict | enforces strict behaviour of the sequential test to determine number of breaks. Sequential test will stop once F(s+1 | s) is not rejected given a rejection of F(s | s-1). Option improves speed in large time series, but should be used with caution. |
| maxbreaks(#) | limits number of breaks when using the sequential test to determine number of breaks. Option improves speed in large time series, but should be used with caution. |
If a panel dataset is used, xtbreak differentiates between four models. The first model is a fixed effects model. A break in the fixed effects is not possible. The second and third models are with a pooled constant (pooled OLS) with and without a break. The last model is a model with neither fixed effects nor a pooled constant.
The following table gives an overview:
| Model | Equation (xtbreak options) |
|---|---|
| Fixed Effects (nobreak) | y(i,t) = a(i) + b1 x(i,t) +s1(s) z(i,t,s) + e(it) |
| Fixed Effects(break) | y(i,t) = a(i,s) + b1 x(i,t) +s1(s) z(i,t,s) + e(it) (breakfixedeffects) |
| Pooled OLS (nobreak) | y(i,t) = b0 + b1 x(i,t) +s1(s) z(i,t,s) + e(it) (nofixedeffects) |
| Pooled OLS (break) | y(i,t) = b1 x(i,t) +s0(s) + s1(s) z(i,t,s) + e(it) (nofixedeffects breakconstant) |
| No FE or POLS | y(i,t) = b1 x(i,t) + s1(s) z(i,t,s) + e(it) (nofixedeffects noconstant) |
where b0 is the pooled constant without break, a(i) the fixed effects, b(1) a coefficient without break, s0(s) a pooled constant with break and s1(s) a coefficient with break.
In the estimation of the breakpoints, cross-sectional averages are not taken into account.
The option python uses Python to calculate the sum of squared residuals (SSRs) necessary to compute the F-Statistics to estimate the dates of breaks and perform tests for an unknown break date. The number of possible SSRs can be very large and computation time consuming. For example, for a model without non-breaking variables, one break (m=1) and a minimal segment length of h=trimming * T, the number of SSRs is:
T (T + 1)/2 − (h − 1)T + (h − 2)(h − 1)/2 − h2m(m + 1)/2,
hence in the order of O(T^2). Using Python improves the speed of calculations.
Python cannot be combined with unbalanced panels. It uses the standard inverter from numpy (linalg.inv), the pseudo-inverse (linalg.pinv) or SVD decomposition (scipy.linalg.svd). Differences between results obtained with and without the Python option may occur for ill-conditioned or (nearly) invertible matrices.
xtbreak checks if the Python and required packages (numpy, scipy, xarray and pandas) are installed. The option python can only be used with Stata 16 or later.
xtbreak allows for unbalanced panels when using panel data. Pure time series data (i.e. data with only one cross-section) with gaps is not allowed. In the case of unbalanced panels, the degree of freedom adjustment for the sup F(s) statistic are adjusted.
While xtbreak kallows for unbalanced data, results should be taken with extra caution. The underlying assumption is that the break dates are the same for all units, including those with gaps in the data. The break date estimation can be biased if data is very unbalanced, that is if a large number of time periods are missing for multiple units. Care is also required if estimated breaks coincide with the start or end of unbalanced panels. We strongly recommend to investigate the SSRs using estat ssr after an estimation with a single break point to identify increases or decreases in the estimated SSRs.
The option noreweigh avoids to reweigh time-individual errors for the calculation of the SSR to artificially increase the SSR of unabalanced sections of the panel. Results with this options should be used indicative.
xtbreak test stores the following in r():
| Scalars | Description |
|---|---|
| r(Wtau) | Value of test statistic. |
| r(p) | p-Value from F distribution. |
| Scalars | Description |
|---|---|
| r(supWtau) | Value of supF statistic (hypothesis 1). |
| r(Dmax) | Value of unweighted double maximum test (hypothesis 2). |
| r(WDmax) | Value of weighted double maximum test (hypothesis 2). |
| r(f) | Value of supremum of supF statistic (hypothesis 3). |
| r(c90) | Critival value at 90%. |
| r(c95) | Critival value at 95%. |
| r(c99) | Critival value at 99%. |
For example we want to find breaks in the US macro dataset supplied in Stata 16. The dataset contains quarterly data on the inflation, GDP gap and the federal funds rate. We load the data in as:
use http://www.stata-press.com/data/r16/usmacro.dta, clear
A simple model to estimate the GDP gap using the federal funds rate and inflation would be:
regress ogap inflation fedfunds
Assume we want to test for a break in Quarter 1 1970:
xtbreak test ogap inflation fedfunds, breakpoint(tq(1970q1))
and if we want to test for a break in 1970q1 and 1990q4:
xtbreak test ogap inflation fedfunds, breakpoint(tq(1970q1) tq(1990q4))
If we want to test if breaks occurs after 10 and 20 periods:
xtbreak test ogap inflation fedfunds, breakpoint(10 20, index)
To test hypothesis 1 with for example 3 breaks:
xtbreak test ogap inflation fedfunds, hypothesis(1) breaks(3)
The default is to assume no break in the constant. To add a break in the constant, the option breakconstant is added:
xtbreak test ogap inflation fedfunds, hypothesis(1) breaks(3) breakconstant}
Hypothesis 2 can be tested with:
xtbreak test ogap inflation fedfunds, hypothesis(2) breaks(3)
The test assumes that under the alternative, there are between 1 and 3 breaks. To test if there are between 2 and 4 breaks under the alternative:
xtbreak test ogap inflation fedfunds, hypothesis(2) breaks(2 4)
To use the weighted double maximum test we use the option wdmax
xtbreak test ogap inflation fedfunds, hypothesis(2) breaks(2 4) wdmax
Hypothesis can be tested using option hypothesis(3):
xtbreak test ogap inflation fedfunds, hypothesis(3) breaks(2)
To change the minimal segment length to 5%:
xtbreak test ogap inflation fedfunds, hypothesis(3) breaks(2) trimming(0.05)
An early version of xtbreak test was presented at the 2020 Swiss User Group meeting (see slides, NOTE The examples are from an early version of xtbreak. Results have changed!) The empircal example was on the question if can we identify structural breaks in the excess deaths in the UK in 2020 due to COVID19? Data from Office of National Statistics (ONS) for weekly deaths in the UK for 2020 was used. The data can be downloaded here.
To test for an unknown breakdate with up to 4 breaks:
xtbreak test d.deaths d.cases , breakconstant breaks(4) hypothesis(2)
We can test if there is a break in weeks 13 and 20 against the hypothesis of no break.
xtbreak test d.deaths d.cases , breakconstant hypothesis(1) breakpoints(13 20, index)
Using a HAC consistent estimator rather than the SSR.
xtbreak test d.deaths d.cases, breakconstant hypothesis(1) breakpoints(13 20, index) vce(hac)
Test for 2 breaks at unknown dates:
xtbreak test d.deaths d.cases , breakconstant breaks(2) hypothesis(1)
Test for 1 vs. 2 breaks:
xtbreak test d.deaths d.cases, breakconstant breaks(1) hypothesis(3)
Andrews, D. W. K. (1993). Tests for Parameter Instability and Structural Change With Unknown Change Point. Econometrica, 61(4), 821–856. link.
Bai, B. Y. J., & Perron, P. (1998). Estimating and Testing Linear Models with Multiple Structural Changes. Econometrica, 66(1), 47–78. link.
Bai, J., & Perron, P. (2003). Computation and analysis of multiple structural change models. Journal of Applied Econometrics, 18(1), 1–22. link.
Ditzen, J., Karavias, Y. & Westerlund, J. (2024) Testing for Multiple Structural Breaks in Panel Data. Journal of Applied Econometrics link
Ditzen, J., Karavias, Y. & Westerlund, J. (2025) Testing and Estimating Structural Breaks in Time Series and Panel Data in Stata. arXiv:2110.14550 [econ.EM]. link.
Karavias, Y, Narayan P. & Westerlund, J. (2021) Structural breaks in Interactive Effects Panels and the Stock Market Reaction to COVID–19. arXiv:2111.03035 [econ.EM]. link
The latest version of the xtbreak package can be obtained by typing in Stata:
net from https://janditzen.github.io/xtbreak/
xtbreak requires Stata 15 or newer.
Email: jan.ditzen@unibz.it
Web: www.jan.ditzen.net
Email: yiannis.karavias@brunel.ac.uk
Web: https://sites.google.com/site/yianniskaravias/
Email: joakim.westerlund@nek.lu.se
Web: https://sites.google.com/site/perjoakimwesterlund/
Ditzen, J., Karavias, Y. & Westerlund, J. (2025) Testing and Estimating Structural Breaks in Time Series and Panel Data in Stata. arXiv:2110.14550 [econ.EM].