Depression is a leading cause of disability worldwide (James et al., 2018). Consequently, it is vital for researchers to identify causes of depression. Perfectionistic concerns and strivings are two dimensions of perfectionism. Perfectionistic concerns are socially prescribed and encompass beliefs that others require perfection of the self, doubts that one can live up to the standards, and self-criticism when one allegedly fails others' expectations. Perfectionistic strivings, on the other hand, are self-oriented and include high personal standards and lofty goals (Frost et al., 1993; Smith et al., 2021). Studies have shown a positive association between depressive symptoms and perfectionism, where associations with perfectionistic concerns tend to be stronger than associations with perfectionistic strivings (Limburg et al., 2017).

According to the vulnerability model, some personality traits, including perfectionism, cause people to think, feel, and behave in ways that increase the risk for depressive symptoms (Hewitt et al., 1996; Graham et al., 2010; Sherry et al., 2013, 2014). Contrarily, according to the complication model, depressive symptoms may cause changes in personality, including perfectionism (Bagby et al., 2008). The reciprocal relations model incorporates both the vulnerability and complication models and suggests dynamic and bidirectional effects between perfectionism and depressive symptoms (McGrath et al., 2012).

Smith et al. (2021) conducted a meta-analytic cross-lagged panel analysis and found a significant positive effect of initial perfectionism on subsequent depression when adjusting for initial depression. For perfectionistic concerns, but not for perfectionistic strivings, Smith et al. also found a reversed effect of initial depression on subsequent perfectionism when adjusting for initial perfectionism. Smith et al. concluded that individuals with elevated perfectionistic concerns appear to be trapped in a vicious loop with ever-increasing levels of depression and perfectionistic concerns.

However, it is known that cross-lagged effects while adjusting for a prior measurement of the outcome variable may be spurious due to correlations with residuals and regression to the mean (Glymour et al., 2005; Eriksson and Häggström, 2014; Castro-Schilo and Grimm, 2018; Sorjonen et al., 2019). As an example, picture individuals with the same initial depression score but with different initial perfectionism scores. Due to the positive association between perfectionism and depression, we should suspect that individuals with a high initial perfectionism score have received a low initial depression score compared with their true degree of depression, i.e., they have received a negative residual in the measurement of depression. Contrarily, individuals with a low initial perfectionism score have probably received a high initial depression score compared with their true degree of depression, i.e. they have received a positive residual in the measurement of depression. However, residuals tend to regress toward a mean value of zero between measurements. Consequently, we should expect a more positive, but spurious, change in the depression score to a subsequent measurement for those with a high initial perfectionism score compared with those with the same initial depression score but with a lower initial perfectionism score. Furthermore, as regression toward the mean is independent of the direction of time, if the effect is spurious, we should also expect a positive effect of initial perfectionism on the initial depression score when adjusting for the subsequent depression score.

To elaborate and exemplify, imagine that data are generated as in Figure 1A. Here, individuals' general/true levels of perfectionism (gP) and depression (gD) are affected by their general level of negativity (gNeg.) but have no direct effects on each other. Perfectionism and depression measured at two occasions (P₁, P₂, D₁, and D₂, respectively) are affected by general levels but have, again, no direct effects on each other. We generated data (N = 10,000) with the model in Figure 1A. Although perfectionism and depression had no direct effects on each other, P₁ had a positive effect on D₂ when adjusting for D₁ (β = 0.177, p < 0.001, this effect corresponds to −0.221 × −0.800 in Figure 1B). However, if adjusting for the D₁-gD difference/residual in addition to D₁, P₁ had no effect on D₂ (β = 0.000, p = 1, Figure 1B). This means that the effect of P₁ on D₂ when adjusting for D₁ was fully accounted for by a negative association between P₁ and the D₁-gD residual in combination with a negative association between the D₁-gD residual and D₂.

It has been proposed that one of the problems with cross-lagged panel models (CLPM) is an inability to separate within-individual effects from between-individual differences (Hamaker et al., 2015). The random-intercept cross-lagged panel model (RI-CLPM) is an extension of the traditional CLPM, where longitudinal scores are regressed on stable trait-like latent variables. Then, autoregressive and cross-lagged effects are estimated between within-individual residuals not accounted for by the stable trait-like levels (Hamaker et al., 2015; Mulder and Hamaker, 2021). It has been proposed that such within-individual effects are better estimates of causality than between-individual differences in CLPM (Usami et al., 2019). However, RI-CLPM is not able to account for time-varying confounders (Rohrer and Murayama, 2023; Murayama and Gfrörer, 2024). This means that RI-CLPM is susceptible to similar spurious findings as CLPM and do, consequently, not allow strong causal inference. We have suggested that causal conclusions based on findings by the RI-CLPM can be scrutinized, similarly as causal conclusions based on the CLPM, by fitting alternative models to data (Sorjonen et al., 2023d, 2025b; Sorjonen and Melin, 2025). It should be noted that the meta-analytic data analyzed by Smith et al., and reanalyzed by us, included two waves of measurement and could, consequently, not be analyzed with the RI-CLPM, which requires data from at least three waves of measurement.

The objective of the present study was to reanalyze the meta-analytic associations used by Smith et al. (2021) in order to evaluate if the prospective effects between perfectionism and depression may have been spurious due to correlations with residuals and regression to the mean rather than, as suggested by Smith et al., truly increasing.

See Smith et al. (2021) for a more comprehensive description of selection of studies, sample characteristics, etc. In short, Smith et al. extracted autoregressive, concurrent, and cross-lagged zero-order correlations between perfectionism and depressive symptoms measured at two occasions from 67 studies with data from 77 samples [total N = 20,583, mean age = 25.8 (SD = 11.6), average percentage female participants = 65.4% (SD = 25.4)].

In turn, we extracted the 2 × 6 meta-analytic zero-order correlations between depressive symptoms and perfectionistic concerns and strivings, respectively, measured at two occasions, from Smith et al. We simulated two datasets with these correlations between variables and with sample sizes N = 16,131 and N = 11,494 for perfectionistic concerns and strivings, respectively. These sample sizes corresponded to total sample sizes in Smith et al. We fitted various models (Figure 2) to data in order to discriminate between truly increasing and spurious prospective effects:

Simulations and analyses were conducted with R 4.1.3 statistical software (R Core Team, 2025) employing the MASS (Venables and Ripley, 2002) and lavaan (Rosseel, 2012) packages. Analytic script, which also generates the simulated data, is available at the Open Science Framework at https://osf.io/6mr9b/.

Correlations and standardized regression effects in the different models are presented in Figure 2. Initial perfectionism had a positive effect on subsequent depressive symptoms while adjusting for initial depressive symptoms, and vice versa (Figures 2A, F). This suggested, as already shown by Smith et al. (2021), a more positive subsequent change (i.e., a larger increase or a smaller decrease) in depressive symptoms for those with high, compared with low, initial perfectionism but with the same initial degree of depressive symptoms (e.g., D₂ = 0.12 × PC₁ + 0.58 × D₁ + e, Figure 3A), and vice versa (e.g., PC₂ = 0.15 × D₁ + 0.73 × PC₁ + e, Figure 3D).

However, the effects of initial perfectionism on initial depressive symptoms while adjusting for subsequent depressive symptoms (Figures 2B, G), and vice versa (Figure 2D), were also positive (except for the effect of initial depressive symptoms on initial perfectionistic strivings, Figure 2I). This means that among individuals with the same subsequent degree of depressive symptoms, those with high initial perfectionism had tended to have higher initial degree of depressive symptoms and, consequently, to have experienced a more negative subsequent change (i.e., a larger decrease or a smaller increase) in depressive symptoms compared with those with the same subsequent degree of depressive symptoms but with lower initial perfectionism (e.g., D₁ = 0.25 × PC₁ + 0.54 × D₂ + e, Figure 3B), and vice versa (e.g., PC₁ = 0.09 × D₁ + 0.76 × PC₂ + e, Figure 3E).

Moreover, the negative crude effect of initial perfectionism on the subsequent depressive symptoms—initial depressive symptoms difference (Figures 2C, H) suggested a more negative subsequent change (i.e., a larger decrease or a smaller increase) in depressive symptoms for those with high, compared with low, initial perfectionism (e.g., D₂-D₁ = −0.08 × PC₁ + e, Figure 3C). Contrarily, a positive effect of initial depressive symptoms on the subsequent perfectionism—initial perfectionism difference (Figures 2E, J) suggested a more positive subsequent change (i.e., a larger increase or a smaller decrease) in perfectionism for those with high, compared with low, initial degree of depressive symptoms (e.g., PC₂-PC₁ = 0.05 × D₁ + e, Figure 3F).

In summary, effects of perfectionism on depressive symptoms agreed better with a hypothesis of spuriousness than with a hypothesis of true reciprocal effects (compare effects in Figure 2 with predictions in Table 1). Effects of depressive symptoms on perfectionism were more ambiguous.

As already shown by Smith et al. (2021), in the present reanalyses we found a meta-analytic positive effect of initial perfectionism on subsequent depressive symptoms while adjusting for initial depressive symptoms, and vice versa. This could be seen to indicate, as suggested by Smith et al., that perfectionism and depressive symptoms had reciprocal increasing effects on each other. However, we also found a positive effect of initial perfectionism on initial depressive symptoms while adjusting for subsequent depressive symptoms, and a negative effect of initial perfectionism on the subsequent depressive symptoms—initial depressive symptoms difference. These findings suggested, contrarily, a negative effect of perfectionism on subsequent change in depressive symptoms. These contradictory findings indicated that Smith al.s' meta-analytically estimated prospective effect of perfectionism on depressive symptoms may have been spurious due to correlations with residuals and regression to the mean rather than truly increasing. Consequently, the present findings call the vulnerability model of perfectionism and depressive symptoms into question.

Effects of depressive symptoms on perfectionism were more ambiguous. A positive effect of initial depressive symptoms on initial perfectionistic concerns while adjusting for subsequent perfectionistic concerns agreed with a hypothesis of spuriousness. On the other hand, positive effects of initial depressive symptoms on the subsequent perfectionism (both concerns and strivings)—initial perfectionism differences agreed with a hypothesis of a truly increasing effect, i.e., with the complication model. This positive effect was due to a stronger positive meta-analytic correlation between initial depressive symptoms and subsequent perfectionism (r = 0.49 and r = 0.19 for perfectionistic concerns and strivings, respectively) compared with initial perfectionism (r = 0.46 and r = 0.15 for perfectionistic concerns and strivings, respectively). Such a temporal strengthening of the association is what can be expected if depressive symptoms had a truly increasing prospective effect on perfectionism. However, as always, caution is advised when/if inferring causality from correlational data.

We have conducted several reanalyses of meta-analyses employing cross-lagged panel analyses and found that most effects may have been spurious, possibly due to correlations with residuals and regression to the mean, rather than truly increasing or decreasing (Sorjonen et al., 2022, 2023a,b,c, 2024, 2025a; Sorjonen and Melin, 2023, 2024a,b). A recurring message in these studies is that cross-lagged effects while adjusting for a prior measurement of the outcome variable usually do not prove anything over and above a cross-sectional association combined with less than perfect reliability in measurements. This limitation of cross-lagged effects, meta-analytically estimated or not, is important for researchers to bear in mind in order not to overinterpret findings, something that appears to have happened to Smith et al. (2021). The continued output of studies with uncritical use of cross-lagged panel analyses indicates that this knowledge, although far from new, is largely missing in the research community. Hence, the limitation of cross-lagged panel analyses is worth repeating. We recommend researchers to employ analyses with a reversed treatment of time, as we have done here, in order to discriminate between true (or, more correctly, “not yet disproven”) and spurious prospective effects.

The present reanalyses suffered from some of the same limitations as the challenged meta-analysis by Smith et al. (2021). For example, 79.4% of the included samples were American, British, or Canadian. Consequently, it remains an open question if the present main finding, that prospective effects of perfectionism on depressive symptoms appear to be spurious due to correlations with residuals and regression to the mean, is generalizable to other cultural contexts.

Measures of perfectionism and depressive symptoms in the studies included in the meta-analysis by Smith et al., and consequently in the present reanalyses, might not always have been optimal. Furthermore, in the present reanalyses we did not consider possible moderating effects of the sex and age composition of the samples, time lag between measurements, etc. However, it is important to note that such factors were constant across the analyzed models and could not, consequently, explain why some models indicated increasing and others decreasing effects between perfectionism and depressive symptoms.

We do not claim to have proven, once and for all, that perfectionism and depressive symptoms have no genuine prospective effects on each other. Correlations in observational (i.e., non-experimental) data can probably never be used to prove causality nor lack thereof. We do claim, however, that a conclusion of spuriousness agrees better with the divergent findings of both increasing and decreasing effects compared with a conclusion of causality. The present findings suggest that the meta-analytic data analyzed by Smith et al. (2021), and reanalyzed by us here, may have been generated without any direct effects between perfectionism and symptoms of depression, e.g., as in the model in Figure 1A.

The present reanalyses found that a meta-analytic prospective effect of perfectionism on depressive symptoms may have been spurious, possibly due to correlations with residuals and regression to the mean, rather than truly increasing. Hence, the conclusion by Smith et al. (2021), that individuals with elevated perfectionistic concerns appear to be trapped in a vicious loop with ever-increasing levels of depression and perfectionistic concerns, can be called into question. Furthermore, the present findings challenge the vulnerability model which claims an increased risk for depression for those with high levels of perfectionism. It is important for researchers to be aware of the limitations of cross-lagged panel analyses, and of correlations in general, in order not to overinterpret findings.